Corruption: A barrier preventing public spending efforts from improving student outcomes

(1)

Corruption: A barrier preventing public

spending efforts from improving student

outcomes

By Carlota Cores

Supervised by: Andro Rilović Faculty of Economics and Business Student number: 10464034

29th of June, 2015

Abstract

Recent empirical research has highlighted the importance of efficiency regarding national education spending in order to improve student outcomes. The majority of studies linking corruption and education mostly conclude an adverse effect of corruption on educational performance and attainment. The aim of this study is to assess the effectiveness of secondary education expenditure in improving educational results under the presence of corruption. Student outcomes are approximated by PISA scores. Government expenditure on secondary education is found to positively impact PISA scores, whilst public education spending is estimated to be less effective in improving student results in more corrupt countries. These findings emphasize the need for transparency in the public sector if improved national student outcomes are to be achieved.

(2)

2 Statement of Originality

This document is written by Carlota Cores who declares to take full responsibility for the contents of this document. I declare that the text and the work presented in this document is original and that no sources other than those mentioned in the text and its references have been used in creating it. The Faculty of Economics and Business is responsible solely for the supervision of completion of the work, not for the contents.

(3)

3 Table of Contents

1. Introduction 4

2. Related Literature 5

2.1 The importance of education 5

2.2 Efficiency: The link between education expenditure and successful outcomes 6

2.3 How corruption stands in the way 7

3. Empirical approach 9

3.1 The model 9

3.2 Data in context

4. Results 19

4.1 Discussion of main results 19

4.2 Empirical robustness 22

5. Conclusion, limitations and policy advice 26

Appendix A 28

Appendix B 29

(4)

4 1 Introduction

The recent Great Recession has encouraged governments around the world to take action and improve their policies. Leaving more than 46 million people unemployed in countries belonging to the Organisation for Economic Co-operation and Development (OECD), the importance of education for long-term growth has been emphasized. The European Commission stresses the need for valuing education and training and has set it as one of their quantifiable targets in the Europe 2020 strategy (European Commission, 2010). The priorities mentioned by the Union are the development of a knowledge-based economy, which is more resource efficient and creates employment. Nevertheless, stronger economic governance is crucial.

Education is an essential component of economic progress. Human Capital Theory has shown the positive impact of human capital development, especially through education and training, on economic growth (Becker, 1964; Barro, 1991; Hanushek, 1995). However, empirical research has also demonstrated that the relationship between resources and achievement is considerably weak and ineffective (Tan & Mingat, 1992). If the explanation for a quality shortage in education is solely due to a lack of resources, the obvious policy recommendation would be to increase expenditure in this area (Pritchett and Filmer, 1999). However, public budget constraints accentuate the need for efficiency regarding education expenditure.

Despite the increased transparency achieved by governments in recent decades, corruption is still damaging the financial system by decreasing government budgets and lowering investment. The costs of corruption are estimated to be more than 5% of the world’s Gross Domestic Product (World Economic Forum, 2012). The fight against corruption can be hindered by government officials and influential authorities who abuse public office for personal benefit. Transparency International’s 2014 report revealed that corruption scores worsened in Italy, France and Spain where scandals have involved serving presidents, members of the royal family and several politicians as well as powerful business people. Protests took place in Romania, Hungary, Spain, Bulgaria, Slovakia and Czech Republic among others. One of the many reasons why poor governance is a matter of concern, is because it affects the quality of public service delivery through distorted government financing

(5)

5 and resource misallocation (Fiszbein, Ringold & Rogers, 2011). This has negative implications for secondary education quality, since it is a predominantly public service in most countries.

Extensive research has been focusing either on educational expenditure efficacy or corruption as two separate fields. However, the main objective of this research is to investigate to what the extent to which there exists a relationship between the two. Following a similar methodology as Rajkumar and Swaroop (2007) in their research ‘’Public spending and outcomes: Does governance matter?’’, the relationship between corruption and efficiency of public expenditure in secondary education will be examined. In this study, spending efficiency is defined as the degree to which a higher public spending results in improved educational outcomes. The hypothesis this research will test is whether efficiency can be obstructed by corruption in such a way that a higher expenditure on education will not result in the desired outcomes. An interaction term between corruption and government expenditure is included in the empirical model in order to examine such a link. The coefficient of this interaction term represents the difference per expenditure on student outcomes between corrupt and non-corrupt nations. Rajkumar and Swaroop’s (2007) analysis focused on primary education and it included data on 57 countries covering the years 1990, 1997 and 2003. This study focuses on secondary education and it analyses 24 OECD member countries and partners in 2006, 2009 and 2012.

The following section contains an overview of the literature in the topics of educational spending efficiency and corruption. A description of the data and the empirical model is provided in Section 3. Section 4 discusses the results of the empirical research and Section 5 contains the conclusion and limitations of the research findings as well as policy implications.

2 Related literature

2.1 The importance of education

Education is the foundation of all economies and extensive research has proved so. Human Capital Theory is founded on the basis that human beings’ skills can be shaped and improved by education. A well-educated workforce leads to higher productivity, facilitates adaptation to new technologies and it has a positive impact on social outcomes (Barro and Lee, 2001). Adam Smith (1812) already stated in his publication ''The Wealth of Nations'' that all valuable skills acquired by a country's citizens were a part of its capital. The inherent capabilities developed

(6)

6 by human capital investment are represented by wages in the labor market (Schultz, 1961, p.8-13). These wages are positively affected by increasing years of schooling and experience (Becker, 1964; Jacob Mincer, 1974). The return on investment in human capital is at least equal to the return on physical investment. Concretely, one tenth to one half of the unexplained increase in US national income could be attributed to the yield to education (Schultz, 1961). The belief that enlarging education facilitates economic growth was also pointed out in several growth models. The main findings suggest that the initial level of human capital helps to explain the rate of income growth (Barro, 1991; Romer, 1990). Furthermore, the extended version of the Solow growth model formulated by Mankiw, Romer and Weil (1990, p.407) includes the accumulation of human capital as an important determinant of income per capita, besides raw labor and physical capital.

2.2 Efficiency: the link between education expenditure and successful outcomes

In order to achieve a successful national educational system that develops human capital (i.e. a skilled workforce), investment by public or private institutions, or a mix of both, is needed. However, for the purpose of this dissertation, only public monetary investment will be considered. From 1870 onwards, public spending has been on the rise (Tanzi & Schuknecht, 1997). Nonetheless, the evidence on whether public expenditure leads to better outcomes is mixed. Some studies find that higher expenditure on education above a certain level does not necessarily lead to welfare improvements (Tanzi & Schuknecht, 1997; Tan & Mingat, 1992). Whereas other authors show that public spending does improve human development indicators (Verhoeven, Gupta & Tiongson, 1999; Gallagher, 1993). Considering that government resources are not unlimited and that the global economic crisis has put substantial pressure on governmental budgets, the key to successful outcomes lies in how efficient governments are at spending taxpayer's money. It is not the disparity between national budgets that determines an education system’s performance but the efficiency of national governments in allocating such budgets (Hanushek, 1995). An education system is considered to be inefficient if it would either achieve the same results with less resources or better results with the same resources (Dolton, Marcenaro-Gutiérrez & Still, 2014). Moreover, if the efficiency of the system is improved, it might lead to government savings which can be reinvested in education or higher student achievement at no extra cost to the public authorities.

(7)

7 Education is treated in most literature as a production process, consisting of inputs (expenditure on academic staff, facilities or students) and outputs (number of graduates, enrollments, citations or published articles). A main conclusion of the literature written on the input-output relationship is that schools around the globe follow considerably inefficient spending policies (Hanushek, 1995). Research on public spending efficiency has been mostly carried out by the use of non-parametric methods which adopt the production function approach. Data Envelopment Analysis (DEA) is the most common one. It consists of comparing decision-making units (DMUs), which are the units under examination, with a benchmark or ''best producer''. DEA is used to minimize educational inputs while maximizing outputs in analyses regarding education systems. Different studies follow similar approaches. Firstly, countries are ranked according to how efficient their national education systems are. Subsequently, an econometric regression on the efficiency scores is performed (Alfonso, Schuknecht & Tanzi, 2010; Aubyn, Garcia & Pais, 2008; Dolton et al., 2014). Alfonso et. al (2010) find that the only variables contributing to higher efficiency of spending are competence of civil servants, GDP per capita and property rights. Dolton et.al (2014) rank several OECD countries' spending effectiveness in primary and secondary education by using 63 different inputs and approximating student outcomes by PISA test scores. They find that the only two components contributing to education outcomes are teacher salaries and class sizes. They conclude that the most efficient systems are Finland, Korea and Czech Republic. These countries achieve the highest rankings due to the efficient combination of their inputs. Nevertheless, they stress that there is no fixed combination of teacher wages and class sizes which guarantees success but instead, that the achievement of successful outcomes depends on the context of each country (Dolton et al., 2014, p.8).

2.3 How political corruption stands in the way

In a period where the global financial crisis has hit economies around the world, citizens are demanding public transparency. Most European residents believe that their political parties are corrupt or highly corrupt (OECD, 2010). Initiatives to fight corruption include important organizations such as the International Monetary Fund, The World Trade Organization, the United Nations and the European Union (OECD, 2011). Findings by the World Bank suggest that the developed economies are less corrupt than emerging economies, however in some of

(8)

8 those emerging economies corruption is less prevalent than in several member countries of the OECD (Kaufmann, Kraay & Mastruzzi, 2007).

The existing research linking corruption and education mostly concludes negative effects of corruption on educational performance and human capital formation. Considering that human capital is one of the main determinants of economic growth, corruption is ultimately damaging national growth rates (Mauro, 1995; Dreher & Herzfeld, 2005). Mo (2001) estimates that corruption reduces average schooling by 0.25 years whereas Gupta, Davoodi and Tiongson (2000) find that corruption increases student dropout rates. Improvements regarding education outcomes will not necessarily come from a higher public spending but mainly from an enhancement of financial accountability and policy transparency. Corruption does not affect the total amount of government spending, but it affects the composition of the public budget instead. Studies have shown that corrupt countries invest lower amounts in the education and health sectors (Mauro, 1998; Gupta, Davoodi and Alonso-Terme, 2002). This result is attributed to the fact that corrupt officials will be inclined to choose inputs produced by oligopolistic suppliers and whose value is hard to be tracked. Simultaneously, a lower investment in the social sectors can lead to an accentuation of income inequalities (Gupta et. al, 2002). When examining the link between national corruption and income inequality a careful analysis must be carried out (preferably an instrumental variable regression) due to reverse causality. This is, corruption may cause inequalities but inequalities are also likely to cause higher corruption (Jong-Sung & Khagram, 2005). A possible explanation is that in countries suffering from high income inequalities the most powerful sector will have higher opportunities to engage in corruption whereas the poor sectors are unprotected against extortion. The effect of corruption on inequality could be alleviated by a higher public expenditure efficiency. However, corruption per se distorts budget equilibrium, therefore decreasing public spending effectiveness (Delavallade, 2006). This spending inefficiency is explained by the fact corrupt agents will not deliberately choose to invest in the most efficient sectors but in the ones where the collection of bribes is easier (Murphy, Shleifer & Vishny, 1993). Consequently, the ambiguity found on the effect of spending on outcomes could be simply attributed to inefficiencies in public spending. These inefficiencies could be a result of different factors including corruption (Rajkumar & Swaroop, 2008, p.98).

The aim of this research is to examine the effectiveness of public expenditure on improving secondary educational outcomes under the presence of corruption. This link has

(9)

9 already been investigated by Rajkumar and Swaroop (2008). These authors estimate a multiple regression model in which they approximate primary educational outcomes by using dropout rates. They capture the impact that corruption has on outcomes by the inclusion of an interaction term, which combines a governance indicator with the primary education expenditure variable. The authors find that expenditure does not have a significant impact on dropout rates but that improved governance reinforces the link between spending and primary education outcomes. This research examines whether this result is consistent regarding secondary education outcomes. Nonetheless, somewhat different variable proxies are used. Two plausible limitations of Rajkumar and Swaroop’s (2008) research are the use of dropout rates as a proxy for education outcomes as well as the incorporation of a continuous corruption variable in their empirical model. Consequently, standardized PISA test scores and a binary corruption variable will be used in this study. A concrete empirical specification and a full description of the data chosen will be provided in the following section. Note that the terms ‘efficiency’ and ‘effectiveness’ will be used interchangeably throughout this research.

3 Empirical approach

3.1 The model

The purpose of this empirical research is to assess the effect that corruption has on public spending efficiency. A method similar to the one followed by Rajkumar and Swaroop (2008) in their paper ''Public spending and outcomes: Does governance matter?'' will be used. The differences in methodology between both studies will be explained after the empirical model is stated. The hypothesis under examination is that in countries where corruption is higher, governments are less efficient at improving education outcomes through increased expenditure. In order to test such hypothesis, a pooled nonlinear regression model will be estimated. Panel data consisting of observations of the same 24 countries covering three different time periods is used. Concretely, the data belongs to the years 2006, 2009 and 2012.

(10)

10

The specification chosen is the following

𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑖𝑖,𝑡𝑡 = 𝛽𝛽0+ 𝛽𝛽1ln (𝑒𝑒𝑒𝑒𝑒𝑒𝑖𝑖,𝑡𝑡) + 𝛽𝛽2𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝑖𝑖 ∗ ln (𝑒𝑒𝑒𝑒𝑒𝑒𝑖𝑖,𝑡𝑡) + 𝛽𝛽3𝑋𝑋𝑖𝑖,𝑡𝑡+ 𝜀𝜀𝑖𝑖,𝑡𝑡

𝑋𝑋𝑖𝑖,𝑡𝑡 = �ln (𝐺𝐺𝐺𝐺𝑃𝑃𝑖𝑖,𝑡𝑡�; 𝐹𝐹𝐹𝐹𝐹𝐹𝑖𝑖,𝑡𝑡; 𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑖𝑖,𝑡𝑡; 𝐷𝐷𝑒𝑒𝑒𝑒𝑖𝑖,𝑡𝑡; 𝑈𝑈𝐶𝐶𝑈𝑈𝑃𝑃𝐺𝐺𝑖𝑖,𝑡𝑡]

where 𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃_{𝑖𝑖,𝑡𝑡} represents the Programme for International Student Assessment (PISA) scores and ln�𝑒𝑒𝑒𝑒𝑒𝑒_{𝑖𝑖,𝑡𝑡}� the natural logarithm of the secondary education expenditure of country 𝑠𝑠 at time 𝑠𝑠. 𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶_𝑖𝑖 is a binary variable representing corruption. The variable takes the value of 0 if the country is considered to be less corrupt and 1 if the country is considered to be corrupt. A country is considered to be corrupt if it scores a punctuation in the corruption index which is lower than the 25th percentile of the total countries’ index scores.

𝑋𝑋𝑖𝑖,𝑡𝑡 represents the vector including the control variables. It comprises of the natural logarithm of the Gross National Product (ln (𝐺𝐺𝐺𝐺𝑃𝑃_{𝑖𝑖,𝑡𝑡})), the student-to-teacher ratio (𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠_{𝑖𝑖,𝑡𝑡}), female education (𝐹𝐹𝐹𝐹𝐹𝐹_{𝑖𝑖,𝑡𝑡}), the age dependency ratio (𝐷𝐷𝑒𝑒𝑒𝑒_{𝑖𝑖,𝑡𝑡}) and the degree of urbanization of a country (𝑈𝑈𝐶𝐶𝑈𝑈𝑃𝑃𝐺𝐺_{𝑖𝑖,𝑡𝑡}).

As previously mentioned, the literature mostly makes use of Data Envelopment Analysis in order to evaluate spending efficiency among different countries. This method was not considered in this study due to one main reason. DEA is a deterministic approach which is sensitive to measurement error (Steering Committee for the Review of Commonwealth/State Service Provision, 1997). If the educational inputs of one country are understated or the outputs overstated, then that country can evolve into an outlier and will significantly deteriorate the shape of the production frontier. This would lead to a decrease in the efficiency scores of other countries. In this research, publicly available databases are used. Despite their accuracy, they comprise large amounts of statistical measures worldwide, and as a result, the data provided can be subject to error. Therefore, an econometric regression is used instead. This method contains error terms in the estimation which decrease the impact of outliers. A limitation of only using a regression is that specific efficiency scores cannot be calculated for each of the countries in our sample. Instead, for the purpose of this study, effectiveness of spending is simply defined as to the degree to which public spending results in improved outcomes.

As in Rajkumar and Swaroop’s (2008) research, the impact of corruption on effectiveness of spending is analyzed through the inclusion of an interaction term. This term

(11)

11 interacts a corruption binary variable with national levels of government expenditure. The coefficient of this interaction term will enable us to see whether the presence of corruption in a country interacts with public expenditure in education in such a way, that the impact on student outcomes of a percentage change in the level of expenditure, depends on the presence of corruption in the public sector. It is expected that the coefficient of the interaction term (𝛽𝛽2) is negative. That is, the amount of money spent in secondary education is expected to be less effective in yielding the desired outcomes in more corrupt countries (𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶=1) than in less corrupt ones (𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶=0). In addition to the interaction term, Rajkumar and Swaroop (2008) include corruption as a variable on its own to see the effect of corruption levels on education outcomes. However, in this paper, only expenditure and the interaction term between expenditure and corruption is included. That is, a corruption variable on its own is excluded from our empirical model. The reason is that the matter of interest in this study is to see whether corruption affects educational outcomes through expenditure and not the direct relationship between corruption and outcomes. This is done under the assumption that that under zero corruption levels, the effect of expenditure on test scores should be ‘’theoretically’’ the same for all countries (Stock & Watson, 2012, p.322). This is a reasonable assumption to make given that the control variables in our model are adjusting for other cross-country differences.

A linear-log specification is chosen because it represents a better fit to the data collected. Logarithms convert variables into percentage changes, making some relationships easier to interpret (Stock & Watson, 2012, p.309). For instance, the expenditure-scores relationship can be interpreted as if an increase in secondary education expenditure of 1% of GDP leads to an increase in PISA scores of 0.01𝛽𝛽₁. This percentage interpretation makes it easy to compare effects among countries. In addition, increases in Gross National Product can also be more easily explained if expressed in percentage terms. The adjusted R-squared can only be used for comparisons between linear and nonlinear models if the dependent variable is the same for both models. Consequently, the adjusted R-squared can be used to compare the linear regression and the linear-log regression (Stock & Watson, 2012, p.313). In Appendix B, a comparison of both regressions and their adjusted R-squared can be found in Tables 4 and 5.

(12)

12 3.2 Data in context

In this research, data comprising the period 2003 to 2012 was mostly gathered from the United Nations Educational, Scientific and Cultural Organization (UNESCO) and the World Bank databases. The time period is selected due to data availability constraints which will be explained subsequently. The cross-section of countries chosen for our sample is limited to the countries and partners belonging to The Organization for Economic Co-operation and Development (OECD). Unfortunately, several OECD members and partners had missing data on the independent variables of our model. Likewise, some countries did not participate in the PISA tests for all of the three years examined in this research. These countries -including Australia, Canada, Greece, Turkey, Croatia, Russian Federation, Thailand and Qatar among others- were consequently taken out of the sample. The remaining sample after removing missing observations comprises 24 countries, of which 15 are OECD members and the rest are partners. Fortunately, the remaining countries constitute a heterogeneous group of countries, which makes them an interesting sample for our analysis. A complete list of the countries employed in this paper can be found in Appendix A, Table 1.

Below a full description of the data can be found. The explanation of the variables chosen, in combination with the literature discussed in the preceding section, will enable a straightforward interpretation of the results. A summary of the descriptive statistics is exhibited in Table 1.

PISA scores: Educational outcomes are represented by PISA test scores. They ultimately

depict how high the quality of an education system is. An education system’s quality is something hard to measure quantitatively. Especially, because the benefits to improvements in an education system are only revealed years after the investment is made. Early studies attempted to measure educational attainment by using enrollment and literacy rates, however these measures, although publicly available, are not adequate measures of educational quality (Barro and Lee, 2001). It is suggested that a better measurement is educational attainment (Barro and Lee, 2001). This indicator represents the number of students who enter and complete a specific education level. Rajkumar and Swaroop (2008) use dropout rates as a proxy for primary education quality. However, regarding secondary education, high dropout rates can be a misleading indicator since they do not necessarily indicate bad quality and outcomes. High dropout rates at this educational level can be a feature of a high quality

(13)

13 education system, which nevertheless fails to motivate all types of students. For instance, if the education method is too theoretical, those students who prefer a more practical and interactive way of learning would be incentivized to drop-out. Another reason for high dropout rates can be attributed to high wages in the labor market for unskilled occupations. An example of this was the housing-market boom, which increased high school dropout rates due to the high salaries offered in the construction sector.

The Programme for International Student Assessment (PISA) is an international educational assessment created by the OECD. The survey takes place every three years and its main aim is to test the knowledge and skills of secondary education students from OECD member countries and partners. Students’ skills are evaluated in three main areas: mathematics, science and reading. The standardization of the PISA tests makes the scores the best available proxy for educational quality. All the 15 year-old students participating in PISA will take exactly the same tests based on which, each country’s educational system will be evaluated. However, there are limitations to approximating educational quality with standardized scores. The fact that one country scores better than another in PISA, does not always imply that its education system will be more efficient (OECD, 2013, p.265). It is important to consider differences in participant countries such as policies on school entry, selection or promotion which might lead to students being assigned to different grades or tracks across different education systems. Furthermore the students’ inherent capabilities or social background are not included. Consequently, this background will be ''controlled'' for with the control variables explained subsequently. The PISA scores used in this research belong to the years 2006, 2009 and 2012. The average score of each country is computed by taking into account the total scores that the students achieved in the areas of mathematics, science and reading. These specific years are chosen due to data constraints regarding the independent variables in the model.

Government expenditure on secondary education (as a percentage of GDP): Time series data

from 2003 to 2011 is taken from the UNESCO Institute for Statistics (UIS) database. As mentioned in the previous section, education can be seen as a pedagogical process and reforms made in an educational system today will not influence student outcomes until some years have passed (Kirk, 2014). That is the reason why cost efficiency must be examined over time. Hence, data on secondary education expenditure is selected three years before each PISA test

(14)

14 was taken. The time frame of three years is chosen because the average lower secondary education age in the countries in our sample is 11.51. The starting age for each of the countries can be found in Figure 1. Due to the fact that students take the PISA tests at the age of 15, there is a time period of three to four years in between the time the students enroll in secondary education and they take the PISA test. Reforms (through spending) will affect lower secondary education from the time the students enroll in this level. Therefore, three years before the test is taken, is considered an adequate amount of time for expenditure (and educational reforms) to affect student outcomes. An average of the public secondary education spending during those three years is computed. Hence, if the PISA test is taken by the students in 2006, the expenditure is calculated by taking the average of the spending in secondary education in the years 2003, 2004 and 2005. For 2009 scores, the average expenditure of the years 2006, 2007 and 2008 is calculated, and so on. An average is calculated instead of applying a lagged variable, because investment in education has an ‘’accumulated’’ effect on student outcomes. In other words, it is not the expenditure realized in 2003 that will affect the scores in 2006 but instead, the aggregate average expenditure of the years preceding 2006. Concerning previous research, mixed evidence about the significance of the effect of public expenditure on education outcomes has been presented. For instance, Tan & Mingat (1992) find that the relationship between expenditure and outcomes is weak whereas Gallagher (1993) finds a significant link. Nevertheless, the coefficient is expected to be positive even if insignificant.

(15)

15

Corruption: The Corruption Perceptions Index (CPI) is used to approximate corruption levels

across our sample of OECD countries and partners. It is a widely recognized index, which is published by Transparency International. This aggregate measure includes corruption involving public officials and focuses on surveys which question the abuse of public power. The CPI aggregates data from 12 different sources in just one score, which ranks the countries from least to most corrupt. The scores, as published by the organization, award the most corrupt countries the lowest scores (a minimum of 0) whereas the less corrupt ones receive the highest scores (a maximum of 10). The fact that the CPI comprises of several corruption indicators reduces measurement error. As stated by Kauffman, Kraay and Zoido-Lobatón (2000), corruption indices are informative and reliable since they aggregate many different sources that strongly agree about the level of corruption prevailing in each country. A benefit of using the CPI is that Transparency International only chooses data sources which are comparable among countries over time. A limitation of this index is that it is based on perceptions, not on quantifiable corruption experiences. Nevertheless, the scores correlate adequately with bribery rates accounted by Transparency International’s Global Corruption Barometer, which verifies the index’s credibility.

Contrastingly to the analysis by Rajkumar and Swaroop (2008), in which they treat corruption as a continuous variable, corruption is treated in this study as a binary variable. Findings by Kauffman et.al (2000) suggest that due to the imprecision of continuous measures of corruption, corruption indices should be used to classify countries into three groups: the most corrupt, the less corrupt and those with a moderate corruption level. In this study, for simplicity and due to the small sample available, the corruption index is used to divide countries between two main categories: most corrupt and less corrupt countries. As previously mentioned, the bound distinguishing both types of countries is set by the 25th percentile of the corruption index scores. The reason for choosing this percentile as opposed to, for instance, the median, is that it gives more weight to the really corrupt countries in the sample. For example, when choosing the median as a threshold Spain and Chile are classified as less corrupt countries whereas Hungary and Czech Republic fall into the most corrupt category. Nevertheless, these countries do not differ much when it comes to their level of corruption. When the 25th percentile is used as a lower bound, those countries fall into the same category of ‘less corrupt’ which includes nations with medium and low levels of corruption. Countries classified as ‘more corrupt’ are those exceeding a score of 4.06667 (the 25th percentile). The

(16)

16 corruption index percentile table can be found in Appendix B, Table 1. As mentioned in the previous section, research has shown that corruption has a negative impact on education attainment and public expenditure on education (Mo, 2001; Gupta et al., 2000).

Control variables are not the variables of interest of this study, however they are included in order to hold constant other factors which could lead the causal effect of interest to suffer from omitted variable bias (Stock and Watson, 2012, p.272). In the case of this research, omitted variables leading to a bias in the results could be those related to economic differences among countries. These student background differences could explain divergences in their test scores. Examples are parental background and the quality of the education system among others. The five control variables described below are chosen on the grounds of the importance given by the literature to the impact of student background in educational outcomes. The average of all the control variables over the years the expenditure was made is computed. That is, three years before the PISA test was taken.

Gross National Product per capita (GNI): Data on per capita GNI was collected from the

World Bank database from the period 2003 until 2011. GNI is defined by the World Bank as the sum of value added by all producers residing in a country, including taxes (and excluding subsidies) not included in the output estimate, plus net receipts of employee compensation and property income from abroad. It is calculated by converting the GNI per capita (in each respective foreign currency) to international dollars applying purchasing power parity. This variable is chosen to adjust for welfare differences between countries. It represents a country's total production and therefore it is a good indicator of its economic performance. As opposed to Gross Domestic Product, GNI does not only incorporate the income accrued by a country’s citizens but it also takes into account international income flows from abroad and excludes income generated in the country but repatriated overseas. Consequently, as stated in the United Nations 2010 Development Report (United Nations, 2010, p.15), GNI is a more accurate measure of a country’s welfare. In the context of the macroeconomic analysis being performed in this study, GNI accounts for family income differences of the students being tested in PISA.

Percentage of female population older than 25 with completed secondary schooling: Student

performance can be partly determined by background characteristics. Particularly, parental education can be an important factor influencing student outcomes. For this reason, a variable

(17)

17 controlling for such characteristics is included in the empirical model. The percentage of female population (older than 25) with completed secondary schooling is the best proxy available to depict the educational background of the 15-year-old students taking the PISA tests. The female percentage is chosen as opposed to the total percentage because it represents the maternal education of the students. The World Bank database (concretely the Barro and Lee dataset) included this indicator regarding primary, secondary and tertiary education. Maternal secondary completion is the best proxy to account for the importance given by the students to the successful completion of their own secondary education, which consequently, will have an impact on their results. Data regarding this indicator was retrieved for the period 2003 until 2011.

Student to teacher ratio (secondary education): Student-to-teacher ratios are frequently used as

a proxy for the quality of different education systems. If the number of students who are supervised by one teacher is too large, it is probable that just one teacher will not be able to give sufficient attention to all the individual students and, consequently, contribute less to their learning experience. Another control variable for education quality could have been repetition or dropout rates, however the most data was available for pupil-to-teacher ratios. Frequency of teacher absence is an adequate indicator of school quality (OECD, 2003, p. 397; Fiszbein et al, 2011). If the amount of times a class has to be cancelled or replaced by another teacher is high, quality standards will decline. A reason for this is that the replacement teacher may not exert all the functions of the designated teacher regarding the scheduled tasks (OECD, 2003). Unfortunately, data on teacher absence frequency was only available for upper secondary education (which would not represent an appropriate explanatory variable of lower secondary education results) and for a limited amount of countries in our sample. Consequently, student-to-teacher ratios was the quality measure selected. If one single observation was missing and it was placed in between two other observations, the average between the previous observation and the following one was computed in order to replace the missing observation. The data was collected from the UNESCO Institute of Statistics (UIS) database on education.

Dependency ratio: A variable that is widely used in the literature when trying to explain

educational outcomes is the dependency ratio. The age dependency ratio is the percentage amount of people in a country who are younger than 15 or older than 64 years old (the

(18)

18 ‘dependents’) with respect to the total working-age population. This is an important indicator affecting educational outcomes because higher dependency ratios can decrease the amount of public investment directed towards the school-age population (Behrman, Duryea &Szekely, 1993). Dependency ratios computed by the World Bank from the years 2003 until 2011 are included. Data are provided in the database as the percentage of dependent citizens per 100 working-age population.

Degree of urbanization: This indicator is defined by the World Bank as the amount of people

living in urban areas (as a percentage of the total population). Also included by Rajkumar and Swaroop (2008), this control variable is meant to captivate the fact that in urban areas access to education is frequently more widespread. Furthermore, better facilities are likely to be constructed in the urban versus the rural areas of a country. Consequently, a higher degree of urbanization in a country could potentially explain cross-country differences regarding educational outcomes. The data gathered regarding this indicator corresponds to the years 2003 until 2011.

Table 1

Descriptive statistics

Variables Observation Mean Std. Dev. Min Max

PISA 72 475.21 42.41 381.11 552.85 exp 72 1.95 0.49 0.83 2.88 CORR 72 6.02 1.98 2.90 9.67 GNI 72 24450.74 12819.30 7503.33 63003.33 FEM 72 37.10 16.12 9.63 71.61 sttratio 72 13.03 4.31 7.21 26.46 Dep 72 48.23 5.10 37.87 61.01 URBAN 72 73.83 12.00 50.04 94.41

(19)

19 4 Results

4.1 Discussion of main results

Pooled analysis has become central in research concerning economic policy and performance (Podestà, 2002). Consequently, a panel data pooled regression estimated by Ordinary Least Squares (OLS) is run in Stata 13.

The results of the empirical model are presented in Table 2. The high R-squared implies that 74.44% of the variation in education outcomes can be explained by the model. Such a high value could also be an indicator of multicollinearity. Imperfect multicollinearity emanates when one of the explanatory variables is highly correlated with another explanatory variable (Stock & Watson, 2012, p.241). This problem leads to an imprecise estimation of the regression coefficients, highly inflating coefficients’ standard errors. The correlations among the variables in the model can be found in Table 2, Appendix B. It can be seen that some variables are highly correlated, such as female completion and the dependency ratio, or the degree of urbanization and corruption. In order to detect multicollinearity, a Variance Inflation Factor (VIF) is computed in Stata. The VIF measure reports how much of one independent variable’s variability is explained by other independent variable’s variability due to high correlations between them. The rule of thumb is that values higher than 20 should be further investigated (as cited in Greene, 2003, p.56). Fortunately, our model does not seem to suffer from high multicollinearity since all the VIF values are below the threshold. These values can be found in Table 3, Appendix B. In addition, a further analysis of multicollinearity can be done by testing different model specifications with the same data, for instance, by dropping some control variables. When doing so, the coefficients and signs of our variables of interest do not change in a drastic manner. This is a positive indication which consequently means that the model is not likely to be misspecified. The different specifications can be found in Table 3.

(20)

20 Table 2: OLS multiple regression, factors affecting educational outcomes

Dependent variable: PISA test scores Independent variables

Education spending (ln of share of GDP) 40.652 (3.03) Corruption dummy x education spending (ln of share of GDP) -40.427 (-2.33)

GNI per capita (ln) 32.919 (4.00)

Female education 0.584 (2.27) Student-to-teacher ratio -1.929 (-2.79) Dependency ratio 0.057 (0.06) Urbanization degree 0.196 (0.65) Constant 112.786 (1.14) R-squared 0.7444 Observations 72

Note: (White heteroskedasticity-corrected t-statistics in parenthesis)

Table 3: OLS multiple regressions, factors affecting education outcomes Dependent variable: PISA test scores

Independent variables EQ(1) EQ(2)

Education spending (ln of share of

GDP) 40.73(2.93) 77.63(5.91)

Corruption dummy x education spending (ln of

share of GDP) -39.64(-2.14) -88.20(-5.20)

GNI per capita (ln) 32.03(4.09)

Female education 0.69(3.38) Student-to-teacher ratio -2.37(-4.18) Dependency ratio Urbanization degree 0.33(1.03) Constant 166.18(2.22) 387.88(13.38) R-squared 0.7069 0.6104 Observations 72 72

(21)

21 The results reported make use of Eicker-Huber-White heteroskedasticity-robust standard errors as recommended by Stock and Watson (2012, p.201). Heteroskedasticity-corrected standard errors are valid no matter if the error term is heteroskedastic or homoskedastic, since robust standard errors will just become standard OLS in the case of homoskedasticity. The results obtained and reported in Table 3 provide several insights. The coefficients of our main explanatory variables are found to be significant and they have the expected signs. As stated previously, literature provides mixed evidence with respect to the effect of public expenditure on outcomes. In this model, it is estimated that if secondary education expenditure increases by 1% of GDP, PISA scores increase by approximately 0.407 points. This is consistent with Gallagher’s (1993) findings, that show that education spending has a positive impact on attainment indicators when controlling for education quality and efficiency. Controlling for these two factors is exactly what it is done when the student-to-teacher ratios (quality) and the relation between the interaction term and outcomes (efficiency) were included in the model. Contrastingly, Rajkumar & Swaroop (2007) as well as several other authors found insignificant effects of spending on outcomes. Therefore this result must be examined with caution. The interaction term is also significant and positively related to PISA scores. Its coefficient indicates the difference per expenditure on PISA scores between more and less corrupt nations. Specifically, in corrupt countries each increase in education spending of 1% of GDP increases PISA scores by 0.404 points less than in countries with lower corruption levels. This is consistent with Rajkumar and Swaroop’s (2008) conclusions. This result means that lower levels of corruption lead to a more efficient public spending in secondary education. Public expenditure in education will have a higher effect on educational results in those countries with better governance. This finding makes sense since automatic promotion or spending in items which are likely to yield the highest bribes will be more usual in those countries where a corrupt governance prevails.

From our control variables, Gross National Product per capita (GNI), the amount of female population with completed secondary education (FEM) and student-to-teacher ratios (sttratio) appear to explain variations in PISA scores. A 1% increase in GNI per capita leads to an increase of approximately 0.33 points in test scores. Intuitively, students from wealthier nations will have an easier access to quality education than those belonging to poorer countries. This is consistent with other findings that estimate that a higher economic welfare has a positive effect on educational attainment. Tan and Mingat (1992) find that Gross

(22)

22 National Product (GNP) is positively related to educational attainment whereas Rajkumar and Swaroop (2007) find a negative effect of GDP on dropout rates. In addition, the FEM variable and sttratio have the expected signs. Higher parental attainment has a positive effect on student outcomes and increasing student-to-teacher ratios or class sizes, will lead to lower student results. The positive effect of higher family attainment on student results is explained by extensive theory pointing out that the background characteristics of the students will impact their educational performance. In addition, the higher the amount of students being supervised by one teacher, the lower the individual impact the teacher will exert in each of them. This could lead to a lower quality of teaching and in turn, lower results. The other two control variables in our model, namely the dependency ratio and the degree of urbanization, are not significant. The former has a surprising sign while the latter has the expected sign. Nonetheless, their coefficients do not provide a useful interpretation due to their non-significance.

4.2 Empirical robustness

The main limitation concerning a pooled OLS regression is that it does not account for individual differences (heterogeneity) among countries. If those characteristics are not controlled for, they might lead to bias and inconsistency in the parameters’ estimation (Hausman & Taylor, 1981, p.1377). In order to check the robustness of our previous results, a panel data model is estimated. Our data is considered to be a balanced panel since all the variables are observed for each country at each time period. The main benefit of panel data models is their ability to control for observed and unobserved entity and time-specific characteristics which might be correlated with the other explanatory variables in the model (Hausman & Taylor, 1981, p.1377). The two most widely used panel data techniques are the fixed and the random effects models. The fixed effects model allows controlling for individual factors that are country or time-specific and might affect the predictors in the model (Torres-Reyna, 2008). For example, this type of model allows to control for cross-country differences that do not change over time, such as cultural factors, and also for variables that differ over time but not across countries, such as international agreements. Another model for estimating panel data is the random effects model. The main difference between the random and fixed effects model is that the former assumes that the variation across countries is random and uncorrelated with the independent variable of the model (Torres-Reyna, 2008, p.25). In order

(23)

23 to decide between both models, a Hausman test is performed (Greene, 2008, pp.208-209). The null hypothesis tests whether a random effects model is preferred versus a fixed effects model (the alternative hypothesis). The result of the Hausman test statistic is 109.59. A value of p=0.000 is obtained, which means that the null hypothesis is rejected at the 5% significance level. Consequently, the fixed effects model is preferred. Within the fixed effects model, mainly two types of regressions can be run: entity and time-fixed effects regressions (Stock & Watson, 2012, p.400). Both are run in Stata 13 but only the results from the time-fixed effects are reported. This is due to the fact that the entity-fixed effects model did not perform well. A possible reason for a worse performance of the entity-fixed effects model is that the sample of countries is too small and the control variables included in the specification were too correlated with the country-fixed effects. In other words, the control variables already accounted for differences across countries. Time-fixed effects models control for variables that are constant across entities but change over time (Stock & Watson, 2012, p.400). These effects were not taken into account with the control variables chosen and they might affect PISA scores. An explanation for a better performance of this model is attributed to the fact that all the countries in our sample belong to the same organization, namely the OECD. Consequently, they can all be affected by international agreements or advised educational reforms implemented by the OECD, which do not differ across countries but change overtime. These reforms can improve education outcomes in all countries. Examples are the policy frameworks developed by the OECD such as ''Lifelong learning for all'' (Levin, 2003). The results from the time-fixed effects regression model can be found in Table 4.

(24)

24 Table 4: Time fixed effects regression, factors affecting educational outcomes

Note: (White heteroskedasticity-corrected t-statistics in parenthesis)

The output of the time fixed effects model implies that our previous results hold even when controlling for year-fixed effects. The R-squared indicates that 75.88% of the variation in education outcomes is explained by the explanatory variables included. The same signs as in the pooled regression (and very similar t-statistics) are found. In addition, the interaction term seems to be even more significant when controlling for time differences between nations. An increase in secondary education expenditure of 1% of GDP leads to an increase of 0.392 points in PISA scores. Furthermore, if more corrupt nations increase their spending by 1% of GDP the impact on test scores will be 0.396 points lower than countries with less corruption. On the one hand, this can be the case if higher investment is directed towards practices such as hiring more teachers on the basis of favors as opposed to merit, consequently lowering the efficiency of spending. On the other hand, as stated by Heyneman (2004, p.638), ''if the public does not trust the education system to be fair more will be sacrificed than economic growth.'' If it is widely believed in a country that the most influential people acquired their status through privilege rather than merit, the school system will also suffer. The reason is that such belief

(25)

25 will discourage students from working hard to achieve their goals, since only the privileged ones will gain access to the best positions later in life.

In addition, another robustness check is performed. There are possibilities of measurement errors in corruption indices, since these are mostly based on perceptions. Measurement errors in the independent variable can lead to bias and inconsistency in the OLS estimator (Stock & Watson, 2012, p.361). For this reason, a robustness check with a different corruption index is performed. This measure is a ''corruption control'' measure gathered from the World Bank Worldwide Governance Indicators (WGI) and developed by Kauffman, Kraay and Mastruzzi (2008). It ranks countries from a score of -2.5 (most corrupt) to 2.5 (less corrupt). As with the CPI, the 25thpercentile is computed and used as a benchmark to differentiate among more and less corrupt countries. A robust regression estimated by OLS is run in Stata 13. Again, the consistency of the results is verified. The same signs are obtained as well as similar significance levels. The results of this regression can be found in Table 5.

Table 5: OLS multiple regression with WGI corruption index Dependent variable: PISA test scores

Independent variables

(26)

26 5 Conclusion and policy advice

There is a large amount of empirical literature examining the importance of efficiency when it comes to public spending in education. Extensive research has also investigated the negative effects of corruption on several aspects of economic growth, including human capital development. However, research is lacking when it comes to examining the impact of corruption on the efficiency of government spending. This study aimed to investigate to what extent the efficiency of government spending in secondary education is diminished under the presence of corruption. Public education spending efficiency was defined as the degree to which expenditure leads to the desired student outcomes (approximated by PISA test scores). The posed research question was investigated by the incorporation of an interaction term between expenditure and a corruption binary variable in the main empirical model. This term enabled us to see the different per expenditure on student outcomes between corrupt and less corrupt nations.

In order to answer the research question, data was gathered from a cross-section of OECD members and partners covering the years 2006, 2009 and 2012. A linear-log OLS model was run in Stata 13. More corrupt countries were found to be less effective at improving secondary education results through increased public spending. These findings are consistent with existing literature (see Rajkumar & Swaroop, 2008). The current empirical literature analyzing the effect of an increased expenditure on educational results is controversial, nevertheless this study suggests a positive and significant relationship between secondary education expenditure and outcomes. To verify the main results two robustness checks were performed. Firstly, a time fixed effects panel data model was run to corroborate the results found by the OLS regression even when controlling for year fixed effects. Secondly, Transparency International's corruption index was replaced by Kauffman et. al’s (2008) ''control of corruption'' indicator in order to correct for possible measurement error of the corruption indicator that could lead to endogeneity in the model.

Although the findings appear to be robust, the estimates must be interpreted with caution due to three main limitations being faced in this study. Firstly, a sample of only 24 countries with three observations (years) per country is fairly small. Small samples due to missing data can lead to sample selection bias (Stock & Watson, 2012, p.365). This could be the case when countries with missing data were removed from our initial sample. Secondly,

(27)

27 simultaneous causality is always a matter of concern when the effect of corruption on educational attainment is examined. Simultaneous causality arises when the causal relationship between the dependent and the independent variables is reciprocal. Applied to the current study, corruption can lead to a lower educational quality, but also an environment with low education levels can incentivize corruption (Gupta et.al, 2000, p,126). A way to address this problem would be through an instrumental variable (IV) regression. An IV regression consists of choosing an ''instrument'' which is correlated with corruption but does not affect PISA scores. Such empirical analysis is recommended for further research. Thirdly, there is a possibility that variables which explain variations in the dependent variable are not included in the regression (Stock & Watson, 2012, p.358). This exclusion of relevant variables can lead to omitted variable bias. This issue is partly addressed when the time-fixed effects model is used. However, with a larger sample including many countries and years, a better way to avoid omitted variables would be a fixed effects model that controls for country-specific as well as time-fixed effects.

The results obtained have implications for economic policy and emphasize the need for improved governance and transparency. Corruption seems to be posing an obstacle between public policy (specifically government education spending) and the desired educational outcomes. Therefore, more vigorous and sustained campaigns against corruption are needed if higher economic growth is to be achieved through human capital development. Corruption is likely to be a barrier to public spending efficiency in public sectors other than education. Examples are the health, culture or environmental protection sectors. Further research investigating the effect of corruption on spending efficiency regarding other categories of public spending is advised. Moreover, the use of more developed techniques to compute country efficiency scores is desirable. As previously mentioned, Data Envelopment Analysis has been used in other research to compute the spending efficiency scores in education for cross-sections of countries. A suggestion for further research is an examination of the effect of corruption on such scores by performing a regression that includes corruption as one of the explanatory variables. This would enable an examination of the direct effect of corruption on education spending efficiency.

(28)

28 Appendix A

Table 1: List of countries

OECD Members OECD Partners

Austria Brazil

Chile Bulgaria

Czech Republic Colombia

Finland Latvia France Lithuania Hungary Romania Japan Uruguay Luxembourg Mexico Netherlands New Zealand Poland Portugal Slovak Republic Slovenia Spain United States

(29)

29 Appendix B

Table 1: Corruption Index percentiles

CORR 1% 2.9 5% 3.466667 10% 3.566667 25% 4.066667 50% 6.066667 75% 7.416667 90% 8.866667 95% 9.4 99% 9.666667 Table 2: Correlations

PISA exp GNI CORR FEM sttratio Dep URBAN

PISA 1 exp 0.5070 1.0000 CORR 0.6377 0.2997 1.0000 GNI 0.6215 0.2796 0.7492 1.0000 FEM 0.2609 0.0720 -0.3072 -0.0771 1.0000 sttratio -0.6113 -0.4128 -0.1897 -0.3746 -0.2700 1.0000 Dep -0.2869 -0.1587 0.1945 -0.0095 -0.7056 0.3732 1.0000 URBAN -0.0300 -0.0872 0.4872 0.2851 -0.4708 0.3399 0.6255 1.0000

(30)

30 Table 3: Variance Inflation Factor (VIF)

Variable VIF 1/VIF

Dep ln(GNI) URBAN FEM sttratio corr_x_ln(exp) ln(exp) Mean VIF 2.81 2.51 2.17 2.03 1.87 1.69 1.43 2.08 0.3557 0.3982 0.4599 0.4918 0.5334 0.5924 0.6980

Table 4: Linear OLS multiple regression

Dependent variable PISA test scores

Independent variables

Education spending (share of GDP) 25.251 (3.42) Corruption dummy x education spending ( share of GDP) -19.168 (-3.80)

GNI per capita 0.001 (2.97)

Female education 0.545 (1.97) Student-to-teacher ratio -2.089 (-2.84) Dependency -0.196 (-0.20) Urbanization degree 0.181 (0.56) Constant 415.868 (8.90) R-squared 0.7265 Adjusted R-squared 0.6966 Observations 72

(31)

31 Table 5: Linear-log OLS multiple regression

Female education 0.584 (2.27) Student-to-teacher ratio -1.929 (-2.79) Dependency 0.057 (-0.06) Urbanization degree 0.196 (0.65) Constant 112.786 (1.14) R-squared 0.7444 Adjusted R-squared 0.7164 Observations 72

(32)

32 Reference list

Alfonso, A., Schuknecht, L., & Tanzi, V. (2010). Public sector efficiency: evidence for new EU member states and emerging markets. Applied Economics, 42(17), 2147-2164.

Aubyn, M. S., Garcia, F., & Pais, J. (2009). Study on the efficiency and effectiveness of public spending on tertiary education. European Commission Economic Papers.

Barro, R. J. (1991). Economic growth in a cross section of countries.

Quarterly Journal of Economics, 106(2), 407-443.

Barro, R. J., & Lee, J. W. (2001). International data on educational attainment: updates and implications. Oxford Economic Papers, 53(3), 541-563.

Becker, G.S. (1964) .“Human Capital Theory”. New York: Columbia University Press for the

National Bureau of Economic Research.

Behrman, J., Duryea, S., & Székely, M. (1999). Aging and economic opportunities

major world regions around the turn of the century. The Office of the Chief Economist

(OCE).

Delavallade, C. (2006), Corruption and Distribution of Public Spending in Developing Countries. Journal of Economics and Finance, 30, 222-239.

Dolton, P., Marcenaro-Gutiérrez, O., & Still, A. (2014). The Efficiency Index: Which education systems deliver the best value for money? (1st ed.). London: GEMS

Education Solutions.

Dreher, A., & Herzfeld, T. (2005). The economic costs of corruption: A survey and new evidence. Available at: http://ssrn.com/abstract=734184

European Commission. (2010). Europe 2020: A Strategy for Smart, Sustainable and Inclusive Growth: Communication from the Commission. Publications Office. Fiszbein, A., Ringold, D., & Rogers, F. H. (2011). Making services work: Indicators,

assessments, and benchmarking of the quality and governance of public service delivery in the human development sectors (Working paper). Office of the Chief

Economist (OCE).

Gallagher, M. (1993). A public choice theory of budgets: implications for education in less developed countries. Comparative Education Review, 37(2), 90-106. Greene, W. (2003). Econometric analysis (5th ed.). Upper Saddle River, N.J.: Prentice

(33)

33 Gupta, S., Davoodi, H. R., & Tiongson, E. (2000). Corruption and the provision of

health care and education services. International Monetary Fund.

Gupta, S., Davoodi, H., & Alonso-Terme, R. (2002). Does corruption affect income inequality and poverty?. Economics of governance, 3(1), 23-45.

Hanushek, E. A. (1995). Interpreting recent research on schooling in developing countries. The World Bank research observer, 10(2), 227-246.

Hausman, J. A., & Taylor, W. E. (1981). Panel data and unobservable individual effects. Econometrica: Journal of the Econometric Society, 42(6), 1377-1398. Heyneman, S. P. (2004). Education and corruption. International Journal of

Educational Development, 24(6), 637-648.

Jong-Sung, Y., & Khagram, S. (2005). A comparative study of inequality and corruption. American Sociological Review, 70(1), 136-157.

Kaufmann, D., Kraay, A., & Zoido-Lobatón, P. (2000). Governance matters. Finance

Dev, 37(2), 10.

Kaufmann, D., Kraay, A., & Mastruzzi, M. (2008). Governance Matters VII: World Governance Indicators for 1996-2007. World Bank Institute and World Bank

Development Research group.

Levin, B. (2003). Approaches to equity in policy for lifelong learning. Education and

Training Policy Division, OECD. Paris: OECD.

Mankiw, N. G., Romer, D., & Weil, D. N. (1990). A contribution to the empirics of economic growth. National Bureau of Economic Research, 407-437.

Mauro, P. (1995). Corruption and growth. The quarterly journal of economics, 110(3), 681 -712.

Mauro, P. (1998). Corruption and the composition of government expenditure. Journal of

Public economics, 69(2), 263-279.

Mincer, J. (1974). Schooling, Experience, and Earnings. New York: Columbia University Press.

Mo, P. H. (2001). Corruption and economic growth. Journal of comparative

economics, 29(1), 66-79.

Murphy, K. M., Shleifer, A., & Vishny, R. W. (1993). Why is rent-seeking so costly to growth?. The American Economic Review, 83(2), 409-414.

(34)

34 Organisation of Economic Co-operation and Development. (2003). Indicator D7: Teacher

Supply and Demand. In Education at a glance: OECD indicators 2003. 396-398. Organisation of Economic Co-operation and Development. (2010). Transparency and

Integrity in Political Finance.

Organisation of Economic Co-operation and Development. (2014). PISA 2012 Results: What

Students Know and Can Do – Student Performance in Mathematics, Reading and Science. OECD Publishing.

Podestà, F. (2002). Recent developments in quantitative comparative methodology: the case of pooled time series cross-section analysis. DSS Papers Soc, 3(2), 5-44. Pritchett, L., & Filmer, D. (1999). What education production functions really show: a

positive theory of education expenditures. Economics of Education review, 18(2), 223-239.

Romer, P. M. (1990). Human capital and growth: Theory and evidence. Carnegie-rochester

Conference Series on Public Policy, 32, 251-286.

Schultz, T. W. (1961). Investment in human capital. The American economic review, 51(1), 1-17.

Smith, A. (1812). An Inquiry into the nature and causes of the wealth of nations.

Stock, J., & Watson, M. (2012). Introduction to econometrics. Boston: Pearson/Addison Wesley.

Steering Committee for the Review of Commonwealth/State Service Provision. (1997). Data Envelopment Analysis: A technique for measuring the efficiency of government service delivery. AGPS, 21-23.

Tan, J. P., & Mingat, A. (1992). Education in Asia: A comparative study of cost and

financing. World Bank Publications, 66(4).

Tanzi, V., & Schuknecht, L. (1997). Reconsidering the fiscal role of government: the international perspective. The American Economic Review, 87(2), 164-168. Torres-Reyna, O. (2007). Panel Data Analysis Fixed and Random Effects using Stata

(1st ed.). Princeton University. Retrieved from http://www.princeton.edu/~otorres/Panel101.pdf

United Nations Development Programme. (2010). Human Development Report 2010: The real wealth of nations ; pathways to human development. United Nations

(35)

35 Verhoeven, M., Gupta, M. S., & Tiongson, M. E. (1999). Does higher government

spending buy better results in education and health care?. International Monetary

Fund.

World Economic Forum. (2012).Global Agenda Councils 2011-2012 report. Retrieved