THE DETERMINANTS OF CORRUPTION: CROSS-COUNTRY-PANEL-DATA ANALYSIS

Transcription

1 bs_bs_banner The Developing Economies 50, no. 4 (December 2012): THE DETERMINANTS OF CORRUPTION: CROSS-COUNTRY-PANEL-DATA ANALYSIS Nasr G. ElBAHNASAWY 1 and Charles F. REVIER 2 1 Department of Economics, Law School, Mansoura University, Mansoura, Egypt, and Department of Economics, Kent State University, Kent, Ohio, USA; and 2 Department of Economics, Colorado State University, Fort Collins, Colorado, USA First version received February 2011; final version accepted April 2012 This study explores the determinants of corruption, utilizing the Hausman and Taylor s technique to estimate a random effects model that incorporates both the effects of corruption determinants that vary over time and those that are time-invariant, and using a larger panel dataset and a comprehensive set of corruption determinants. The first interesting result is that perception of strong support for rule of law is strongly correlated with reduced corruption, suggesting that a better quality of law enforcement reduces corruption. Rich countries have lower corruption, and the perception of free expression and accountability strongly decreases corruption, indicating that providing greater opportunities for citizens to participate in selecting their government, more freedom of expression, and free media are effective ways of curbing corruption. Conversely, natural resource abundance, country population size, country s dominant religious tradition, ethnic fractionalization, and political stability are unimportant determinants of corruption, while previous research has suggested they are important. Keywords: Corruption; Bribery; Rule of law; Voice and accountability; Governance indicators JEL classification: D72, D73, H11, K42, O11 I. INTRODUCTION Corruption, which the World Bank and the International Monetary Fund (IMF) have defined as the abuse of public office for private gains, has received particular attention in recent years from both policymakers and researchers. Several studies have explored some of the harmful effects of corruption, finding that corruption distorts the allocation of resources, reduces the productivity of public expenditures, and slows economic growth. 1 The World Bank (2001) has identified corruption as a global challenge and the single greatest obstacle to economic and social development. Recently, the World Bank (2004) estimated that more than US$1 trillion is paid in bribes each year. Likewise, the 1 For example, see Mauro (1995, 1996) and Wei (1999). doi: /j x

2 312 the developing economies African Union has estimated the cost of corruption in Africa at US$148 billion annually, which is equivalent to 25% of Africa s GDP. One of the most important problems in empirical studies on corruption is the paucity of high-quality data, both across countries and, more severely, over time. This scarcity is not hard to understand since corruption is a secret activity by nature, occurring away from the glare of publicity, which makes it difficult to measure empirically. Nevertheless, several measures have been attempted, and since the early 1980s a number of organizations have published various crosscountry data sets of measures of the perception of corruption, derived from survey questionnaires sent to networks of correspondents around the world. One such measure is the Corruption Perceptions Index (CPI), issued annually since 1995 by Transparency International, an international nongovernmental organization fighting corruption worldwide. This index is based on surveys of business people, local citizens, and experts in each country regarding their perceptions of both the frequency of bribes and the total value of bribes paid. Each of the CPI sources must provide a ranking of nations. The CPI scores countries on a one-toten scale, which we have adjusted so that 10 represents the most corrupt and 1 represents the least corrupt. 2 An alternative measure of perceptions of corruption is the Control of Corruption (CC) index from the World Bank s Worldwide Governance Indicators (WGI) database, which measures perceptions of the extent to which public power is exercised for private gain, including both petty and grand forms of corruption, as well as capture of the state by elites and private interests. 3 The CC is a weighted average of the underlying individual indicators, where greater weights are assigned to sources with higher correlations with each other. When a new CC indicator comes out, adjustments are made to the scaling of the CC indicators in earlier years, to make the over-time comparison of the CC measure more robust. The CC scores lie between -2.5 and 2.5, which we have adjusted so that 2.5 indicates the most corrupt and -2.5 denotes the least corrupt. Much of the past research on the determinants of corruption has concentrated on particular categories of determinants, such as political institutions (Lederman, Loayza, and Soares 2005) or sociocultural factors such as proportion of females in the labor market and political participation (Swamy et al. 2001; Shaw 2005), ethnic fragmentation, origin of the legal system, and religious affiliation (La Porta et al. 1997, 1999). Our aim in this paper is to obviate any possible omitted-variable bias by including a comprehensive set of potential determinants economic, political, and sociocultural including some that have not previously been 2 For more details, see Lambsdorff (2005). Also see 3 Kaufmann, Kraay, and Mastruzzi (2007, p. 4). For more details, see Kaufmann, Kraay, and Mastruzzi (2006, 2007). Also see

3 THE DETERMINANTS OF CORRUPTION 313 explored. We test the significance of these variables using the Hausman and Taylor (1981) technique to estimate a random effects model, which allows us to obtain consistent estimates of the parameters. In this approach, some regressors are assumed to be correlated with the unobserved country-specific effects and hence endogenous, while others are independent of the random individual effects and therefore exogenous. This approach has not previously been utilized in the corruption determinants literature. Our richer model specification incorporates both variables that vary over time in each country and those that are time-invariant, and employs a larger panel dataset for 159 countries for the years 1998, 2000, 2002, 2003, 2004, and 2005 to take full advantage of both the variation across countries and the variation over time. Because the main determinants of corruption in previous research remain widely debated and no real consensus has emerged on the true determinants of corruption, this study uses two different indexes of perceived corruption, the CPI and the CC, to check the robustness of the results with alternative corruption measures. The findings of this paper support some of the past research on contributors to corruption including low per capita incomes. In addition, a new finding here is that the perception of a stronger rule of law predicts a reduced level of corruption. Furthermore, the perception of strong democratic institutions and accountability lowers corruption. This suggests that the arsenal of anticorruption policies should include stronger and more visible law enforcement, as well as greater scope for citizens to participate in selecting their government, more freedom of expression, free media, and, generally, faster and more sustained democratization. Moreover, our work suggests that natural resource abundance, the population size of a country, religious affiliation, female participation in the labor force and in politics, ethnic division, and political stability are unimportant in determining corruption, although most of the previous research has suggested that they are important determinants of corruption. The following section surveys possible corruption determinants. Section III presents our data sources and empirical strategy. The results of our analysis are described in Section IV. Section V then concludes and discusses the policy recommendations and implications of the results. II. THE BROAD SCOPE OF CORRUPTION DETERMINANTS In some cases, only a single person is involved in a corrupt act, as when a government official chooses to award a lucrative government contract to a business because he has a significant financial interest in that business. In such cases, the corrupt action will occur when the official appraises the benefits of the action to exceed the costs. Most corrupt actions, however, involve the payment of a bribe by a private agent to a public official in exchange for a government decision favorable

4 314 the developing economies to the private agent. The quantity and magnitude of such bribes will be the result of the interaction between the supply of bribes by private agents and the demand for bribes by public officials. On the supply side, private agents reach a decision to offer a bribe through a comparison of the potential benefits of supernormal profits (economic rents) achievable by the bribe with the costs, including both the cost of the bribe itself and the psychic costs determined by the individual s values and tastes, as well as the likely penalties and social censure if the bribe is discovered. The latter will depend upon the probability of being caught, the magnitude of the criminal penalties, and the attitudes and values of others regarding corrupt acts. On the demand side, public officials reach a decision to accept a bribe by comparing the benefits of the additional income with the psychic costs and the costs of the criminal penalties and social censure if exposed. The determinants of corruption will therefore be those economic, sociocultural, and political factors that affect either the benefits or the costs, including psychic costs, on the supply side or on the demand side. Researchers have suggested a broad range of such factors. A number of recent studies have investigated the determinants of corruption, as the recent availability of corruption measures has helped to quantify the extent of corruption and has allowed international comparisons. 4 Nevertheless, no real consensus has emerged on the underlying causes of corruption. Moreover, some economic aspects of the corruption issue remain unexamined, and some possible determinants of corruption have so far received scant attention. A. Economic Determinants There are a number of economic variables that could potentially have an impact on the benefits or costs of paying or receiving bribes. It has been suggested that the wage rate of civil servants relative to that in the private sector has an impact on the incidence of corruption (Van Rijckeghem and Weder 1997; Tanzi 1998; Lindbeck 1998). Relatively low wages in the public sector will make the benefit of a given bribe seem greater and the cost of losing the government job if the bribe is discovered seem less. Various authors have suggested that the level or stage of economic development, as indicated by per capita GDP, will have an impact on corruption (Myrdal 1968; Huntington 1968; Treisman 2000; Knack and Azfar 2003; Lederman, Loayza, and Soares 2005; Serra 2006). A richer country will be able to devote more resources to the detection and prevention of corruption. In addition, the increased education and literacy that development brings will increase the likelihood that an act of corruption will be discovered and punished. 4 For examples of literature surveys on corruption determinants, see Jain (2001), Aidt (2003), Lambsdorff (2006), Pellegrini and Gerlagh (2008).

5 Ades and Di Tella (1999), Treisman (2000), and Knack and Azfar (2003) have suggested that greater openness to foreign trade will be associated with less corruption. As suggested by the trade literature (Bhagwati 1982), trade restrictions, such as tariffs and quotas, provide public officials with higher discretionary power which may result in a considerable amount of rents and rent-seeking activities. Greater foreign competition reduces the potential for economic rents and therefore the potential gain from bribes. Similarly, Leite and Weidmann (1999) argued that the extent of natural resource abundance is an important determinant of a country s level of corruption since natural resources create opportunities for rent-seeking behavior. It also seems likely that rent-seeking opportunities will be more abundant the greater the size of the public sector (Fisman and Gatti 2000). If a significant proportion of goods and services are purchased by government, there are many more government contracts for which bribes might be offered. B. Sociocultural Determinants THE DETERMINANTS OF CORRUPTION 315 Several authors (Fisman and Gatti 2000; Treisman 1999) have found evidence that corruption is less severe in small countries than it is in large countries. In large countries, citizens may be more tempted to bribe officials for public services, to avoid the longer queues that result from the lower ratio of government officials to citizens. On the other hand, in small countries, officials and citizens are more likely to have family ties or other personal relations which makes the supply of and the demand for bribes more costly because of the higher social censure, since everybody knows everybody in a small country. Knack and Azfar (2000), however, argue that this relationship between country size and corruption is an artifact of sample selection. Most available corruption indicators provide ratings for those countries in which multinational investors have the most interest almost all large nations, but among small nations only those that are well governed. Therefore, the relationship between country size and corruption becomes insignificant for an alternative corruption indicator that covers all World Bank borrowers regardless of country size. Thus, the significance of a country s population size remains controversial. For a given population size, the extent of a country s urbanization may have some bearing on the prevalence of corruption. An urban population is likely to be more knowledgeable and informed about political affairs and the operation of the government bureaucracy, and more critical of questionable behavior by public officials. Therefore, the greater the percentage of the population that is rural, the more likely it is that corrupt acts will occur. Some recent empirical studies (La Porta et al. 1997, 1999; Treisman 2000; Serra 2006) claimed that a country s dominant religious tradition may impact corruption. Protestantism s role as a dissenting religion may have contributed to a more

6 316 the developing economies adversarial view of the relationship between religion and the state, leading to less tolerance for corruption, whereas religion and the state are usually more closely related in establishment religions such as Islam and Catholicism. Moreover, the hierarchical nature of Catholicism and Islam may have led to greater respect by the faithful for social hierarchies and greater tolerance for the government s abuse of public power, whereas the individualistic emphasis of Protestantism may have resulted in less respect and less tolerance. In a different vein, Swamy et al. (2001) investigated the differential incidence of corruption by gender. Their cross-country data for 93 countries showed that corruption is less severe where women hold a larger share of parliamentary seats and where they have a larger share of the labor force. They also argue that micro data from the World Values Surveys and an enterprise survey in Georgia show that women are less engaged in bribery and are less likely to regard bribe-taking as excusable. Conversely, Shaw (2005) found that women in Ukraine tend to have a higher propensity to bribe to enter an educational institution and to bribe on exams. Thus, the relationship between gender and corruption remains unsettled. La Porta et al. (1999) and Lederman, Loayza, and Soares (2005) found that the degree of ethnic fragmentation of the country also impacts corruption and other aspects of government performance. Higher ethnolinguistic fractionalization and the prevalence of strong family ties, together with a lack of national identity and the absence of accountability of government officials, may lead people in positions of power to favor friends, relatives, or other members of their own ethnic group, at the cost of a lower quality of public services and greater corruption. Another factor potentially impacting the incidence of corruption for La Porta et al. (1999) is the origin of a country s commercial code and legal system whether English common law, socialist, French, German, or Scandinavian. These legal systems differ in the degree to which they emphasize protection of the power of the state versus protection of individual citizens from the abuse of power by public officials. The English system of common law originated from efforts to restrain the power of the sovereign and to protect citizen rights. The socialist or Communist legal system, on the other hand, placed a heavy emphasis on establishing legal institutions to maintain the power of the state and its ability to extract resources. The civil law systems developed in continental Europe were somewhere in between these extremes. Treisman (2000) and Serra (2006) emphasized the importance of the presence or absence of the common law tradition. An alternative would be to investigate the presence or absence of the socialist/communist origin of the legal system. It would seem plausible that such a system might be associated with greater corruption due to the more limited protections for individuals from the corrupt acts of government officials.

7 C. Political Determinants To the extent that public attitudes influence a public official s assessment of the social stigma attached to a corrupt act, public perceptions about the quality of governance may have an impact on the level of corruption. Recent efforts at the World Bank, described by Kaufmann, Kraay, and Mastruzzi (2007), to construct a set of aggregate Worldwide Governance Indicators (WGI), provide a source of good-quality data on these public perceptions. One of the indexes in the WGI set is the Political Stability and Absence of Violence index, which measures perceptions of the likelihood that the government will be destabilized or overthrown by unconstitutional or violent means, including domestic violence and terrorism (Kaufmann, Kraay, and Mastruzzi 2007, p.3). Various authors (Leite and Weidmann 1999; Lederman, Loayza, and Soares 2005; Serra 2006) have emphasized the significance of political stability in influencing the prevalence of corruption. As Treisman (2000) pointed out, however, the direction of the effect is uncertain. On the one hand, if political instability reduces the likelihood that an official will remain in office for more than a short time, this may reduce the perceived costs of a bribe, and therefore lead the official to act more opportunistically. The potential loss of the job if a corrupt public official is exposed may be of little consequence if the job will not last long anyway. On the other hand, the extended tenure arising from political stability may give a public official more time to build long-term relationships with potential suppliers of bribes and to establish a reputation as one who can be bought and who will keep his side of the bargain. Another index from the WGI is the Voice and Accountability index, which measures perceptions concerning the extent to which a country s citizens are able to participate in selecting their government, as well as freedom of expression, freedom of association, and free media (Kaufmann, Kraay, and Mastruzzi 2007, p. 3). A public official in a country whose citizens believe these freedoms are strong and well protected is likely to feel that a corrupt act will be quickly discovered and punished. The WGI Rule of Law index measures public perceptions of the extent to which agents have confidence in and abide by the rules of society, and in particular the quality of contract enforcement, the police, and the courts, as well as the likelihood of crime and violence (Kaufmann, Kraay, and Mastruzzi 2007, p.4). Since better law and order increases the probability of identifying and punishing illicit rent appropriations, a perception that the rule of law is strong lowers the incentives to behave dishonestly. III. THE DETERMINANTS OF CORRUPTION 317 DATA AND EMPIRICAL METHODOLOGY The particular measures of the explanatory variables suggested above came from a variety of sources. Data on GDP per capita, adjusted for purchasing power parity

8 318 the developing economies (PPP) and measured in constant purchasing power (in thousands of 2005 international dollars), came from the World Bank s World Development Indicators (WDI 2009). For the relative public sector wage, we used cross-national data on the ratio of average government wage to per capita GDP from the World Bank s Statistical Information Management Analysis (SIMA) database. The data in this source were five-year averages in each country for the years and for the years It seems unlikely that this government-wage-to-gdp ratio changed significantly over the period of our sample, and therefore, for most countries, we used the average as the value in each of our observations. This variable is therefore time-invariant for each country. 5 The openness to international trade is gauged by the sum of exports and imports of goods and services as a percentage of GDP, from the World Bank s WDI. 6 Following Ades and Di Tella (1999), our proxy for natural resource endowment is the sum of net exports of fuel, ores and metals, and food, measured as a percentage of merchandise exports. Data on net exports of fuel, ores and metals, and food are from the World Bank s WDI. The data on general government final consumption expenditures as a percentage of GDP also came from the World Bank s WDI. The World Bank s WDI was also the source of data on country populations (in millions) and the percentage of each country s population that was rural. Data on religious affiliation, measured as percentages of the population who were Protestant, Roman Catholic, or Muslim, were obtained from La Porta et al. (1999). Although in principle these percentages could vary from year to year, such changes should be slight for the eight-year period being explored here. Due to limitations in the availability of the data, the percentages are assumed to remain constant over the period of this study. They are therefore time-invariant explanatory variables. For most countries these statistics are for the year For newer countries, the data come from the years The data on the percentage of women in the labor force and the percentage of parliamentary seats held by women were obtained from the World Bank s WDI. We used the index of ethnic fractionalization from Alesina et al. (2003), which is a measure of the probability that two randomly chosen individuals in a country will 5 For some countries for which there were no data for the average, we used the average for instead, if available. Moreover, in this source, the values listed for Kenya and Nepal for both five-year averages were zero, and we therefore omitted those countries from the sample. This measure may fail to reflect the relative overall standard of living of government employees if in-kind benefits (such as housing or trips abroad) are a significant part of the compensation of these employees. 6 However, it would be preferable to use a direct measure of trade protectionism rather than trade volumes. Ideally, this could be provided by a weighted-average tariff rate, including all import charges, calculated by weighting each tariff by the fraction of world trade in that import category, but there were insufficient data for this measure.

9 THE DETERMINANTS OF CORRUPTION 319 be from separate ethnic groups. Their new index is available for almost twice as many countries as previous indices, and uses a broader classification of ethnic groups. It is based on data from the early to mid-1990s, and its authors argue that it should be reasonably stable over time horizons up to 30 years. This is another of our time-invariant explanatory variables. The dummy variable which is equal to 1 for only those countries whose legal system has a socialist/communist origin is from La Porta et al. (1999), and this variable is obviously also time-invariant. Finally, the World Bank s WGI indexes, Political Stability and Absence of Violence, Voice and Accountability, and Rule of Law, were obtained from the World Bank Governance Indicators Database, presented in Kaufmann, Kraay, and Mastruzzi (2007). All scores lie between -2.5 and 2.5, with higher scores corresponding to a higher quality of governance. 7 We thus assembled a panel data set consisting of observations from 150 countries for the years in the CPI models, but for the years 1998, 2000, 2002, 2003, 2004, and 2005 in the CC models, since the CC measure is unavailable for the years 1999 and Because of greater information set in panel data, the panel analysis is a good direction to address the multicollinearity problem, and it allows dealing with the endogeneity and measurement error of various variables. It also enables us to account for heterogeneity by including time-invariant variables. 9 Our panel is unbalanced since observations for some of the variables in some years are missing for some countries. We initially used the CPI as the dependent variable in two different versions of our model, but we then estimated the same regressions using the CC measure to check the robustness of the results to an alternative choice of corruption measure. Since we are combining a large number of variables from a variety of sources, we have many more observations for some of the explanatory variables than for others. There were a total of 628 observations for the CPI measure of corruption and for 13 of the explanatory variables, but only 367 observations for the remaining four explanatory variables. Consequently, we estimated two different specifications, one with the more restricted set of 13 explanatory variables for the full sample, and one with the full set of 17 variables for the more limited sample. In the first specification, the 13 explanatory variables are: GDP per capita; total population; the percentage of the population that is rural; the percentages of the population who are Protestant, Catholic, and Muslim; the percentage of seats held by women in the national parliament; the percentage of the labor force that is female; ethnic fractionalization; the socialist/communist legal system origin dummy; and 7 See Kaufmann, Kraay, and Mastruzzi (2006, 2007). Additional information is provided at: info.worldbank.org/governance/wgi/ 8 Descriptive statistics for all variables and their correlation matrix are presented in the Appendix. 9 For the benefits of the panel data approach, see Baltagi (2005).

10 320 the developing economies the WGI indexes for Political Stability and Absence of Violence, Voice and Accountability, and Rule of Law. Adding dummy variables for the years 2000, 2002, 2003, 2004, and 2005 brought the total number of explanatory variables in this restricted specification to 18. The second specification is the unrestricted specification that adds to the previous set of explanatory variables: the ratio of average government wage to per capita GDP; openness to trade; the endowment of natural resources; and general government final consumption expenditure. Again, adding dummy variables for years brought the total number of explanatory variables in the unrestricted specification to 22. The two most widely used techniques for the analysis of a panel data set are the fixed effects model and the random effects model. The fixed effects model assumes that the unobservable country-specific effects are fixed parameters to be estimated along with the coefficients of the model. In the random effects model (sometimes called the error component model), the unobservable country-specific effect is assumed to be a random disturbance that is distributed independently of the idiosyncratic or remainder disturbance that varies over time as well as across countries. A limitation of the random effects model is that its parameter estimates will be consistent only if the unobserved country-specific component of the disturbance is uncorrelated with the model s explanatory variables. In the fixed effects model, on the other hand, the estimators are consistent regardless of whether the country-specific effects are correlated with the regressors. However, the fixed effects model has at least two limitations of its own. First, since that approach involves estimating the country-specific effect for each country in the panel, the loss of degrees of freedom can be severe. Second, the fixed effects estimator is incapable of estimating the coefficients of any explanatory variables which vary across countries but are invariant over time, since these are not identified in that model. In examining the determinants of corruption, we were anxious to explore the effects of time-invariant variables as well as those that vary over time in each country. Therefore, we initially attempted to estimate the model coefficients using the random effects model. Once those estimates were obtained, however, Hausman s (1978) specification test was used to test the null hypothesis of no correlation between the regressors and the individual country-specific random effects. With the CPI as the dependent variable, in the restricted model with 18 explanatory variables and 628 observations, the value of the Hausman test s chi-squared statistic with 13 degrees of freedom 10 was 97.3, rejecting the null hypothesis even at the 1% significance level. Similarly, for the unrestricted model with 22 explanatory 10 The number of degrees of freedom is less than the number of explanatory variables in the random-effects equation because there were five time-invariant variables in the equation.

11 THE DETERMINANTS OF CORRUPTION 321 variables and 367 observations, the chi-squared statistic with 16 degrees of freedom 11 was 41.5, again rejecting the null even at the 1% level. Thus we were forced to abandon the random effects estimator as inconsistent. But Hausman and Taylor (1981) have suggested an alternative approach for estimating the random effects model. In their approach, some of the regressors are assumed to be independent of the random individual effects, but others are correlated with this country-specific component of the disturbance and thus endogenous. Consistent estimators are obtained by using the instrumental variables technique. Efficiency is increased by taking advantage of the estimated variancecovariance matrix of the disturbance term to conduct feasible Generalized Least Squares estimation of the equation. The instrumental variables consist of the within-country averages over time of the exogenous time-varying variables, the deviations from within-country means for all of the time-varying variables, and the exogenous time-invariant variables. In our case, the equation estimated by the Hausman-Taylor technique is of the form: CRPTN = X β + X β + Z α + μ + ν, (1) it 1it 1 2it 2 i i it where i = 1,...,N (N denotes the total number of countries); t = 1,...,T (T denotes the total number of years); CRPTN it denotes corruption in country i in year t; X 1it is the vector of exogenous time-varying explanatory variables for country i in year t; X 2it is the vector of endogenous time-varying explanatory variables for country i in year t; Z i is the vector of time-invariant explanatory variables (here taken to be uncorrelated with the individual random effect m i ); m i denotes the unobservable, time-invariant, country-specific random effect; and n it denotes the remainder of the disturbance. The m i are assumed to be independent and identically distributed (i.i.d.) random variables with mean 0 and variance s m2. The n it are i.i.d. with mean 0 and variance s n2, and the m i are independent of the n it. Moreover, X 1it, X 2it, and Z i are independent of the n it for all i and all t. One difficulty with the Hausman-Taylor approach is that it is not always obvious which of the regressors should be viewed as correlated with the unobserved country-specific effects, m i, and which should be considered truly exogenous. Our strategy was to examine the statistical significance of the difference between the fixed-effects and random-effects coefficients for the individual explanatory variables, which is calculated in the process of constructing the Hausman test statistic. If the difference was significant at a reasonable significance level, then that variable was taken to be endogenous in the Hausman-Taylor procedure. The logic here is that the fixed-effects estimators are consistent no matter what, so when the 11 The unrestricted random-effects equation included six time-invariant explanatory variables.

12 322 the developing economies random-effects estimators are significantly different, that must be because those variables were correlated with the disturbance term and therefore endogenous. This is essentially the same logic that underlies the Hausman test statistic itself. Potentially, some of the time-invariant variables could also be correlated with m i, and there is no comparable way of identifying endogenous variables among these. But for our time-invariant variables government-wage-to-gdp-per-capita ratio, population percentages who were Protestant, Catholic, or Muslim, ethnic fractionalization, and the socialist-communist legal origin dummy endogeneity seemed unlikely. IV. RESULTS Table 1 shows the results of our Hausman-Taylor estimation of four different specifications of the panel model with cross-section random effects. Columns (1) and (2) show the results when the CPI is used as the measure of corruption to estimate the two specifications discussed above. In the restricted equation of column (1), time-varying explanatory variables treated as endogenous were GDP per capita, percentage of the population that is rural, percentage of parliamentary seats held by women, and the WGI indexes for Voice and Accountability and Rule of Law. Hence, time-varying explanatory variables treated as exogenous were total population, percentage of females in the labor force, WGI index for Political Stability and Absence of Violence. In the unrestricted equation of column (2), the explanatory variables treated as endogenous were percentage of the population that is rural, percentage of females in the labor force, and the WGI indexes for Voice and Accountability and Rule of Law, while the explanatory variables treated as exogenous were GDP per capita, openness to trade, natural resource endowment, general government final consumption expenditure, total population, percentage of parliamentary seats held by women, and the WGI index for Political Stability and Absence of Violence. On the other hand, columns (3) and (4) show the results of those two specifications using the CC measure of corruption. In the restricted equation of column (3), explanatory variables treated as endogenous were GDP per capita, percentage rural, percentage of parliamentary seats held by women, and the WGI index for Rule of Law. The explanatory variables treated as exogenous in this equation were total population, percentage of females in the labor force, and the WGI indexes for Political Stability and Absence of Violence and Voice and Accountability. In the unrestricted equation of column (4), endogenous explanatory variables were GDP per capita, general government final consumption expenditure, percentage of parliamentary seats held by women, and the WGI indexes for Voice and Accountability and Rule of Law, while exogenous explanatory variables were openness to trade, natural resource endowment, total population, percentage rural, percentage

13 THE DETERMINANTS OF CORRUPTION 323 TABLE 1 Hausman-Taylor Estimates of the Panel Models with Cross-Section Random Effects CPI CC (1) (2) (3) (4) Intercept *** *** (1.2567) (1.3612) (0.2937) (0.3397) Government wage / GDP per capita (0.0930) (0.0265) GDP per capita *** *** *** (0.0144) (0.0193) (0.0041) (0.0065) Openness to trade *** (0.0019) (0.0006) Natural resource endowment (0.0030) (0.0008) Government consumption expenditure (0.0126) (0.0035) Total population (0.0013) (0.0008) (0.0003) (0.0003) Percent rural (0.0136) (0.0135) (0.0035) (0.0027) Percent Protestant ** (0.0123) (0.0091) (0.0029) (0.0031) Percent Catholic (0.0082) (0.0054) (0.0019) (0.0019) Percent Muslim (0.0089) (0.0065) (0.0020) (0.0020) Percentage of seats held by women * (0.0053) (0.0069) (0.0014) (0.0020) Percentage of females in labor force *** (0.0195) (0.0257) (0.0048) (0.0057) Ethnic fractionalization * (0.9443) (0.6802) (0.2258) (0.2135) Socialist/Communist legal origin *** * (0.5889) (0.4613) (0.1415) (0.1499) Political stability & absence of violence * (0.0693) (0.0894) (0.0174) (0.0222) Voice and accountability ** ** *** *** (0.0982) (0.1143) (0.0256) (0.0303) Rule of law *** *** *** *** (0.1331) (0.1825) (0.0313) (0.0428) Year 2000 dummy ** (0.0550) (0.0626) (0.0147) (0.0166) Year 2002 dummy ** * (0.0612) (0.0685) (0.0159) (0.0179) Year 2003 dummy *** (0.0634) (0.0680) (0.0166) (0.0184) Year 2004 dummy *** *** (0.0699) (0.0761) (0.0181) (0.0208) Year 2005 dummy ** *** (0.0768) (0.0876) (0.0200) (0.0247) S.E. of country random effects S.E. of idiosyncratic disturbance Hausman test c 132 = c 122 = 3.00 c 132 = c 162 = 6.74 Total panel observations ***, **, and * represent statistical significance at the 1%, 5%, and 10% levels, respectively. Standard errors are in parentheses.

14 324 the developing economies of females in the labor force, and the WGI index for Political Stability and Absence of Violence. 12 In all four equations, the Hausman specification test fails to reject the null hypothesis of no remaining correlation between the regressors and the countryspecific random effects even at the 10% significance level. A. Economic Determinants Relative government wages do not impact perceived corruption in any of our regressions. Strikingly, per capita GDP is statistically significant at the 1% level in all specifications, except in (1) where it is statistically insignificant at the 10% level. Thus there is less perceived corruption in rich countries than in poor countries. Holding everything else constant, a one-standard-deviation increase in real GDP per capita (an increase of $11,797 international dollars per capita) reduces the index of perceived corruption by 0.6 in the CPI regression and by 0.3 in the CC regressions. Openness to trade reduces perceived corruption significantly only in the CPI specification, and even there the magnitude of the effect is small (each onestandard-deviation increase in the ratio of the sum of exports and imports to GDP (a 48.7 percentage-point increase) reduces perceived corruption by 0.3 of a point). On the other hand, natural resource abundance does not impact corruption significantly in either specification. Similarly, government consumption expenditure does not have a statistically significant impact on corruption in any regression. B. Sociocultural Determinants The results provide strong evidence that a country s population size does not seem to influence perceived corruption. The population coefficient is statistically insignificant even at the 10% level in all regressions. Likewise, the proportion of rural population does not affect perceived corruption, since the effect is statistically insignificant at the 10% level in all specifications. Another interesting result is that a country s dominant religious tradition does not impact corruption. None of the population percentages affiliated with the dominant religious traditions (Protestantism, Catholicism, and Islam) is statistically significant in our regressions even at the 10% level, except the coefficient on 12 In the restricted equations of columns (1) and (3), time-invariant explanatory variables treated as exogenous were population percentages who were Protestant, Catholic, or Muslim, ethnic fractionalization, and the socialist/communist legal origin dummy, while in the unrestricted equations of columns (2) and (4), time-invariant variables were government-wage-to-gdp-per-capita ratio, population percentages who were Protestant, Catholic, or Muslim, ethnic fractionalization, and the socialist/communist legal origin dummy.

15 Protestantism which is statistically significant at the 5% level only in specification (1). In this specification, a higher percentage of the population professing Protestantism lowers perceived corruption. The proportion of seats held by women in the national parliament is inversely associated with perceived corruption only in equation (4) where this effect is statistically significant at the 10% level. In this specification, holding everything else constant, each one-standard-deviation increase in the proportion of seats held by women in the parliament (a 9.23 percentage-point increase) reduces perceived corruption by 0.03 of a point. Similarly, a higher proportion of females in the labor force lowers perceived corruption only in equation (2) where the coefficient is statistically significant at the 1% level. In this equation, each one-standarddeviation increase in the proportion of females in the labor force (an 8.63 percentage-point increase) reduces perceived corruption by 0.78 of a point. Ethnic fractionalization does not seem to affect corruption. The ethnic fractionalization coefficient is statistically insignificant at the 10% level in all equations, except in equation (1) where it is statistically significant at the 10% level with a positive sign, indicating that ethnic fractionalization is positively correlated with perceived corruption. The mean corruption level in countries with a socialist/ Communist origin of the legal system is statistically significantly higher by about 1.6 points in the unrestricted CPI regression and about 0.3 of a point in the restricted CC regression, but the effect is statistically insignificant in the other two specifications. C. Political Determinants THE DETERMINANTS OF CORRUPTION 325 Political stability has no significant impact on corruption in any of the specifications except specification (2), where it is statistically significant at the 10% level. In this specification, greater political stability increases perceived corruption. Remarkably, a stronger perception of voice and accountability lowers perceived corruption in all equations. When the perception of voice and accountability rises in a country by one standard deviation (a one-point increase in the voice-andaccountability index), the country s perceived corruption drops by 0.23 to 0.26 of a point in the CPI equations and by 0.12 to 0.13 of a point in the CC equations. This result is statistically significant at the 1% level in the CC regressions, and at the 5% level in the CPI regressions. The most striking result in Table 1 is the strong impact of the rule of law on corruption. Holding everything else constant, when the perception about the strength of the rule of law in a country increases by one standard deviation (a 1.01-point increase in the rule-of-law index), the country s perceived corruption drops by 0.64 to 0.90 of a point in the CPI specifications and by 0.39 to 0.52 of a point in the CC specifications. This result is statistically highly significant in all specifications in both the CPI and CC regressions.

16 326 the developing economies The year 2000 dummy is insignificant, except in equation (4) where it is significant at the 5% level. Likewise, the year 2002 dummy is significant only in equation (2) (at the 5% level) and equation (4) (at the 10% level). The year 2003 effect is statistically significant merely in equation (2), while the 2004 and 2005 year effects are statistically highly significant in equations (2) and (3). V. CONCLUDING REMARKS Although a number of recent studies have examined the determinants of corruption, many of them have focused on a specific set of determinants and no real consensus has emerged on the underlying determinants of corruption. This study has explored the determinants of corruption using larger panel data with a greater information set to take full advantage of the variation across countries as well as the variation over time. The study uses a large array of explanatory variables that may influence corruption, to obviate any possible omitted-variable bias. This comprehensive set of corruption determinants includes economic, sociocultural, and political variables. We analyzed this dataset using Hausman and Taylor s (1981) technique to estimate a random effects model, which has not previously been utilized in the past research on the determinants of corruption. In this approach, some regressors are assumed to be correlated with the unobserved country-specific effects and thus endogenous, while others are independent of the random individual effects and hence exogenous. Our richer model specification incorporates both the effects of corruption determinants that vary over time in each country and those that are time-invariant. By using the Hausman and Taylor (1981) estimator for this model, consistent estimates of the parameters are obtained. Since the underlying determinants of corruption in previous research remain poorly understood and widely debated, this study uses two different indexes of perceived corruption, the CPI and the CC, to check the robustness of the results with alternative corruption measures. This work improves our understanding of the determinants of corruption and suggests a number of ways to reduce corruption by identifying its true determinants. The first notable result indicates that the government wage level does not affect perceived corruption, which contradicts the Tanzi (1998) and Lindbeck (1998) suggestions. Another interesting finding is that rich countries are perceived to have lower corruption than poor countries, all else being equal. In addition, the degree of openness to international trade lowers corruption level in the CPI regression, although the size of this effect is small. Hence, the relationship between the degree of openness to international trade and corruption is sensitive to the choice of corruption measure. On the other hand, natural resource abundance does not significantly impact corruption in any specification, which contradicts the Leite and Weidmann (1999)

17 THE DETERMINANTS OF CORRUPTION 327 argument that natural resource abundance is an important determinant of a country s level of corruption. This result is robust. Likewise, the government final consumption expenditure does not significantly impact corruption in any of our regressions. We also found that country population size does not seem to impact perceived corruption. This robust result differs from the findings of Fisman and Gatti (2000) and Treisman (1999). Another robust result is that a larger percentage of rural population does not influence perceived corruption. Additionally, a high level of affiliation with any of the dominant religious traditions (Protestantism, Catholicism, and Islam) does not seem to be correlated with perceived corruption. This robust finding is inconsistent with Treisman (2000) and Serra (2006). We found little evidence that a higher proportion of seats held by women in the national parliament is associated with lower corruption levels, which contradicts the finding of Swamy et al. (2001) in their within-country data from Georgia. Similarly, little evidence is found that the proportion of females in the labor force is associated with corruption level. Another prominent result is that ethnic fractionalization does not significantly impact corruption, contradicting Fearon and Laitin (1996) and La Porta et al. (1999) who suggested that ethnic fractionalization lowers the quality of the government and hence may increase corruption. On the other hand, this study provides some evidence that the mean corruption level in countries with a legal system of socialist/communist origin may be statistically significantly higher, although the result is not robust across specifications. A perception of political stability and absence of violence does not influence corruption. This robust result is contrary to Lederman, Loayza, and Soares (2005) and Serra (2006) who found that political stability lowers corruption. On the other hand, the level of voice and accountability, which has not previously been investigated empirically as a corruption determinant to our knowledge, lowers perceived corruption in all regressions. Thus, a potential approach to curbing corruption may include the development of greater opportunities for citizens to participate in selecting their government, more freedom of expression, and freer media. One of the most significant robust results in this analysis indicates that the perception of the rule of law strongly impacts corruption and that a higher level of law and order (a better quality of law enforcement) is correlated with lower corruption. Strikingly, this result is statistically significant at the 1% level in all specifications. Indeed, the perception of the rule of law is among the most important determinants of corruption across countries and over time. Moreover, variables usually found to be highly important in determining corruption, such as income per capita, lose their relative importance after accounting for perceptions of the rule of