Is Corruption Contagious? An Econometric Analysis

In this paper, I analyze how corruption in one country may be affected by its neighbors' corruption. It seeks to explain why corruption is perpetuating in large geographical areas populated by developing countries despite anticorruption efforts made in the single country. In our empirical approach, we capture the spatial dependency by regional corruption. Three main techniques are used: spatial autocorrelation tests, GMM and three stage least squares. Our results show that, a lower regional corruption (as measured by the average of the level of corruption in one country's neighbors) is associated with a lower level of national corruption. Among the potential mechanisms explaining this correlation, the level of economic development (GDP per capita) seems to be the most important. Foreign aid and trade openness show less clear results. Noneconomic mechanisms such as cross-country contagion processes of voice expressions and demands on accountability are other possible transmission mechanisms.

This paper aims to study the contagion effects of corruption, captured through spatial interactions between countries. To my knowledge this is a set of mechanisms that so far has not received much attention in empirical research of corruption 2 . I inquire whether spatial interdependency contributes to understand the persistence of corruption. Here, I assume that the persistence of corruption is developing countries can be explained by the contagion phenomenon. If such an effect exists it should to be based on interactions between individuals (Manski, 2000) of two different countries.
The contagion effect may mainly take place at the bribe payers' end, since the migration at the main bribe receivers' end, among the public employees, naturally is quite limited 3 .
Here spillover mechanisms have to be more indirect: being educated at the same regional university centers, the same American, French or British university, or participation to the same conferences; for example customs officers from the same region may participate to an anticorruption conference. Miguel (2006, 2007) consider that diplomats' behaviors in violating parking could be "interpreted as an indication of their home country's cultural tolerance for corruption rather than their own personal values". Bribe payers such as businessmen may physically move into a neighbor country and pay bribes there.
Besides this increase in neighborhood opportunities, common language and culture shared by different people are sources of commercial relations (Moreno and Trehan, 1997).
Another possible link is colonization history. But, it is also possible that people from nearby countries may have different social traditions, norms and trust vis-à-vis malfeasance, crime or 1 A prominent question in the literature has been how to define corruption. It is obvious that there is no unique definition of corruption. Different propositions, ranging from "too restrictive" to "too wide" (Amundsen, and al, 2000) share a common denominator which can be expressed as follows: "the abuse of public authority or position for private gains". 2 Our study is related to the empirical analysis of growth where Ades and Chua (1997) and others authors have documented a relationship between regional political instability and growth 2 . Moreno and Trehan (1997) documented neighborhood growth rate effects. Unlike these authors, I examine if regional dependency will have an impact on the level of corruption in a given country. Seldadyo (2008) has examined the space and governance nexus and concludes that spatial interactions matter. In this paper I will concentrate on corruption issues. 3 An outline of theoretical possibilities here can be found in Andvig (2006). corruption. Proximity induces countries to know about each other or to be influenced not only by economic or political arrangements but also institutional contexts in a given country. Individuals then decide accordingly to corrupt or to be corrupted by individuals from a nearby country, keeping in mind their own culture or that of the closest countries. And because agents of cross-border corruption are capable of doing business under certain constraints everywhere in the world, it is almost impossible to hold them accountable anywhere (Johnston, 1997).
The present study is related to the empirical literature of determinants of corruption.
Existing econometric studies have looked at cross-section econometrics where country observations have been considered statistically independent (Treisman, 2000;Straub, 2000;Frechette, 2001;Seldadyo and de Haan, 2005;Serra, 2006). However, cross-section econometrics (such as ordinary least squares) of corruption which are based on the assumption of stochastic independence across countries are not efficient (Anselin, 2003). Thus, in order to take into account the spatial effects, two methodologies (which I describe in details below) are applied, assuming that countries' observations are not independent: the spatial correlations tests and the Generalized Method of Moments with clusters. The former make it possible to detect the space dependence of a given variable, which is corruption in this study. The latter deals with the problem of endogeneity provided that many variables are plausibly endogenous. A lot of right-side variables in the corruption equation are determined at the same time as corruption and interact with it (Treisman, 2000;Lambsdorff, 1999a). Indeed, they should all be considered endogenous. This is the case for the variables of economic development measured by per capita income (Mauro, 1995;Hall and Jones, 1999;Treisman, 1998), trade policy (Ades and Di Tella (1995;Wei, 2003;Neeman et al., 2003;Laffont andN'Guessan, 1999), decentralization (Fishman andGatti, 1999;de Mello and Barenstein, 2001).
The remainder of this paper is organized as follows. Section I presents some preliminary tests and selected variables. In section II and III the econometric approach and results are discussed. The robustness of the results is tested in section IV. Conclusions and policy implications are finally presented.

Section I -Empirical analysis of corruption: first diagnosis
Empirical analyses are carried out on data covering four biannual periods (1996, 1998, 2000 and 2002) (so as to match the temporal coverage of corruption variables) and 120 developing and developed countries (see appendix A1 for the list of these countries). The preliminary tests rely on the estimates of spatial correlations tests. Before its presentations, we describe the variable of corruption -that is, the national corruption of a single country.

1.1-Corruption measurement issues 4
Corruption is difficult to measure because its activities are developed secretly. At macroeconomic level, only perception indexes are available 5 . The corruption data we use in this study come from the governance database of Kaufmann, Kraay and Mastruzzi (KKM, 2003). This database includes five other indicators 6 of governance apart from that of corruption Our estimates are based on data from 1996, 1998, 2000 and 2002 7 . For the purpose of the analysis, we rescaled this variable from 0 (lowest corruption) to 10 (highest corruption) 8 . These data have some good characteristics which justify our preference for them and because they fill some gaps of the existing indicators found elsewhere in the literature. First, unlike the corruption variable of International Country Risk Guide (ICRG) indicator the index of KKM is a composite one and is thus more precise and informative than any individual indicator. Second, it is a product of an unobserved component model which allows efficient computation of data coming from different individual sources. That is, it ensures comparability and clustering of individual measures. Third, the data from KKM (2003) include continuous variables, getting us rid of multinomial models regressions 9 .
"Control of Corruption" measures perceptions of corruption, conventionally defined as "the exercise of public power for private gains". Despite this straightforward focus, the par-4 Data sources for all other variables employed in the analysis are detailed in appendix A2. Summary statistics are provided in appendix A3. 5 For discussions on use (abuse) and critics of governance and corruption indicators see Williams and Siddique (2007) ;Seligson, 2006;Kurt and Schrank, 2007;Arndt and Oman, 2006;Kaufmann et al, 2007). 6 There are (i) Voice and Accountability; (ii) Government effectiveness, (iii) Regulatory Quality (iv) "Rule of Law" (v) "Political Stability and Absence of Violence". 7 This database is available for more recent years, the latest year being 2007. As part of our future research agenda, the analyses undertaken here could be extended. 8 We rescale the indicator on a 0-10 basis according to the following formula: The corruption variable of Transparency International (TI) is also a composite one but not continous. Another limit of this variable is the aggregation methodology, which change each year, making it difficult year to year comparison. Finally, the sample coverage of TI's variable is lower that the KKM's. ticular aspect of corruption measured by the various sources differs somewhat, ranging from the frequency of "additional payments to get things done," to the effects of corruption on the business environment, to measuring "grand corruption" in the political arena or in the tendency of elites to engage in "state capture". The presence of corruption is often a manifestation of the lack of respect of both the corrupter (typically a private citizen or firm) and the corrupted (typically a public official or politician) for the rules which govern their interactions, and hence represents a failure of governance (KKM, 2003). Table 1 summarizes the main characteristics of corruption data for selected groups of countries covering the study's periods 10 . On average, all developing countries seem to be more significantly affected by corruption than OECD countries. Thus, these figures confirm the idea according to which developing countries have higher levels of corruption than developed countries. Nevertheless, it proved difficult to establish significant corruption differences between sub-Saharan Africa, Latin America and Asia. Another characteristic of this variable is the high correlation (a correlation coefficient of at least 0.94) between values of different periods.

1.2-Measure of regional corruption
The spatial autocorrelation is based on the concept of contiguity: two units are assumed to be contiguous if they share a common border. I first compute the regional corruption based on the neighborhood matrix 11 and then present the tests and results.
10 Figures indicate the unweighted average level of corruption. In brackets, we have the number of countries. 11 Knowing that we have a sample of 120 countries, the neighborhood matrix is 120X120.
To measure the contagion effect, we proceed as follow. For a country i, we calculate the simple average of corruption index values of its neighbors as defined by Ades and Chua (1997). To calculate regional corruption values for Togo for example, I sum corruption values of its neighbors (Benin, Ghana and Burkina Faso) and this sum is then divided by three, the number of Togo's neighbors.
The index values range as previously rescaled between 0 and 10 with bigger values associated with more corruption.
In table 2, we present figures of the unweighted regional corruption. As one could expect, African countries have on average the highest corruption concentration at the regional level. This is the case for all periods and confirms at a first glance high level of corruption in Africa.
Nonetheless this preliminary result which is specially based on descriptive statistics needs to be strengthened through some rigorous empirical analyses. In the next subsections, I use two techniques to quantify the contagion effect of corruption: the tests of spatial autocorrelation and the Generalized Moments Method (GMM).

1.3-Tests of spatial autocorrelation
The global autocorrelation and the Moran Diagram make it possible to detect the space dependence of a given variable, which is corruption in this study.
The global correlation highlights the global spatial dependence at the level of single countries. This correlation is based, for a given variable, on the index of Moran which is calculated as follows: where represents the weighting coefficient and is defined for two countries i a by: Table 3 presents the coefficients of global autocorrelation calculated on a sample of 122 countries (developed and developing) over the period 1996-2002. Calculations are based on corruption data, as described in the previous paragraphs. Estimations based on corruption data of KKM(2003) All the coefficients in Table 3 are positive and significant at 1% confidence level. Hence, neighborhood contries tend on average to share similar behaviors vis-à-vis corruption.
The correlation between national corruption and corruption in neighboring countries can also be detected through the diagrams of Moran (Anselin, 2003). On these diagrams (Figure 2), national corruption is represented in X-axis and the regional corruption in Y-axis. The scatter plots are represented by the coordinates of each country of the sample. The diagram of Moran presents two advantages: (i) the adjustment line (average relationship) makes it possible to consider spatial correlation between national corruption and regional corruption and (ii) the scatter plots contribute to appreciate the geographical distribution of the spatial correlation of corruption. There is a positive autocorrelation when there is a clustering of countries having similar levels of corruption (quadrants I and III). When there is a grouping of countries with dissimilar levels of corruption, one speaks about negative autocorrelation (quadrants II and IV).
The diagrams of Moran show a positive relation between national corruption and regional corruption. Moreover, these diagrams support the assumption of a bipolar distribution of the levels of corruption: a clustering of the majority of developed countries (Sweden, Finland, Denmark, France, Norway, Switzerland, etc) whose levels of corruption are low (quadrant I) and a grouping of the developing countries with high levels of corruption (quadrant III).

2.1-Motivations
As I intend to analyze the contagion effect of corruption, by looking at the spatial interactions, my question is what the better econometric approach to use is. To deal with the weaknesses of the existing literature, I argue that the generalized moment method (GMM) is more appropriate.
Using the GMM approach allows me to deal with the problem of endogeneity as many variables seem to be endogenous. This method is more efficient than the simple instrumental variables analysis (two stage least squares). As we have outlined above, almost all the variables under consideration are affected by issues of simultaneity and endogeneity. Nevertheless, we focus on three of them: (i) the per capita GDP, (ii) trade openness and (iii) regional corruption that we consider all endogenous. As previously mentioned, some variables like fractionalization, colonial tradition and natural endowment could be considered as "mostly exogenous" (Treisman, 1998). The level of development can affect corruption, and the reverse is also true, i.e. corruption reduces growth (Mauro, 1995). On the other hand, trade reduces corruption, but it is also likely that corrupt bureaucrats create opportunities for themselves to extract rents by generating barriers to trade (Treisman, 1998).
The third variable has to my knowledge not been studied in any econometric research on corruption before. Interactions between regional corruption and national corruption might be endogenous (cf. Manski, 2000). Yet, the challenge here is to find some good and valid instruments for regional spillovers. Referring to the empirical studies, we chose population, log of infant mortality, life expectancy and land area and distance from the equator as instruments. A good instrument is the one that is highly correlated to the instrumented variable and slightly if not at all to the unobserved components (errors) of the explained variable. If we consider the instrument candidates, their relationship to GDP is well established in theory. As far as trade is concerned, Di Tella et al. (1999) used population and land area as instruments. Even though no theoretical explanation may support our instruments choice, we rely upon exogeneity tests. More specifically, Hansen tests are provided so as to test if the chosen instruments are exogenous.
In addition, this method makes it possible to take into account the correlation between countries. Unlike much of previous empirical analysis, my study goes one step further and will help to get more insights on determinants of corruption.
As data are only available on four periods, one may question the most appropriate method to employ. Among the studies that have discussed the time dimension, Treisman (1998) estimated separately two equations for the years 1980 and 1990 respectively. So did Ades and Di Tella (1999). In this study, we pool the data for the four periods.
Furthermore, the fixed effects approach presents some limits for analyzing corruption. The main reason why fixed effect models are not appropriate is due to the low variability of corruption level over time. Indeed, only a smaller proportion of the variance of the corruption variable is explained. As a consequence, the standard errors tend to be larger and the coefficients estimates are biased toward zero (Woodridge, 2002). On the other hand, using the standard fixed effects method prevents from estimating the importance of the time-invariant variables which, according to Treisman (1998), seems important in the explanation of corruption 12 . Even though this approach could account for the bias due to the omission of these variables, it is not possible to get further empirical insights on how they contribute to explain the contagion effect which is my concern. Common cultural factors -such as language, ethnicity or the same colonial history -are time-invariant and, as previously argued, contribute to the expansion of corruption and spatial spillovers. WC represents the regional corruption as calculated in previous subsection using W as the spatial matrix weight 13 . It is important to point out that the coefficient 1 β captures the magnitude of spatial interactions among countries (Brueckner, 2003).

2.2-The model specification and the variables of control
Including other determinants of corruption is necessary in order to avoid the bias of omitted variables in the corruption equation. Such determinants may contribute to explain the interactions between countries and thereby the spatial spread of corruption.
12 Using Hausman-Taylor method could help one estimate the magnitude of time invariant-variables. However, in the specific case of corruption dependent variable, this method is not efficient. First, it could hardly overcome the problem of low within variation of corruption. Second, this method, as documented in the literature, is sensitive to the choice of the time-invariant endogenous variables. 13 In this regard, it is important to note that our analysis is consistent with the empirical analysis of spatial interactions. a) Level of development captured by the gross domestic product (GDP) per capita: a higher level of income is associated with a lower corruption (Scott, 1969;Paldam, 2002;Mbaku, 1999;Treisman, 1998;Tanzi, 2000) 14 b) Trade openness: greater (natural) openness is associated with lower corruption (in the context of perfect competition) (Ades and Di Tella, 1996, Nguesan and Laffont, 1999Wei, 2000). As regard rents, the countries that export minerals and fuel are assumed to be more corrupt.
c) Voice and accountability: used to capture the political factors. However the theory is ambiguous on the effect of political factors on corruption (LaPalombara, 1994;Rose-Ackerman, 1999;Huntington, 1996). d) Foreign aid/GNI: the relationship between foreign aid and corruption could be either positive or negative (Knack, 2003;Braütigam, 2004). e) Social factors: some social attributes can generate corruption or constrain it. They are captured by the ethnic and religious fractionalization and the proportion of protestants in 1980.
f) Legal origin: legal origin affects corruption (La Porta et al., 1998). We consider two variables here, namely the British and the French legal origin.

3.1-The neighborhood corruption effect
Before presenting the main results, it is worth noting that apart from regional corruption, all the variables in any econometric specification are defined at the national single country level.
The first column of Table 4 includes only the average of neighbors' corruption. This preliminary specification allows us to evaluate the simple correlation between national and regional levels of corruption. However, this coefficient may be biased because of omitted variables that may explain corruption. I therefore introduce other determinants of corruption in column 2 and thereaf- ter. Different specifications are tested so as to examine the channels by which the contagion effect operates 15 .
Our results strongly support the hypothesis that the perception of corruption in one country is affected by the corruption in its neighbor countries as measured by the regional average of corruption. Indeed, the higher the neighbors' corruption, the higher it is in one single country. 14 Kaufmann and Kraay (2002) have reached the opposite result wherein in their econometric specification Corruption causes lower GDP -strong effect while increased GDP increase corruption -but only slightly so in any rough correlation between them high GDP goes together with low corruption level. 15 For example, to conclude if the log of real GDP per capita is a transmission channel of the contagion effect, one will have to compare the coefficient of regional corruption in columns (1) and (2) and test if the two coefficients are significantly different.
Hence, positive spatial interactions of corrupt behaviors are confirmed. How do these interactions occur? In the next paragraphs, I analyze the main channels through which the regional corruption affects the level of corruption in a given country.

3.2-The other factors of corruption
Theoretically, I expect a negative sign for this coefficient, which means that corruption decreases with the level of development. Globally, our results support this hypothesis (Treisman, 1998). In addition, this variable seems to be an important channel through which the regional corruption exerts its effect on one country's corruption as we observe a great decrease of its coefficient once the income per capita variable is included. In column (2), the magnitude of the coefficient of regional corruption variable is reduced up to 62% of the one in column (1).
In column 4 of Table 4, I include, in addition to the real GDP variable, three other variables: trade openness, mineral and fuels rents and voice and accountability. In the following lines, I comment the specific effect of each variable.
Let us first consider trade openness. The trade variable is captured in the literature of the share of imports to GDP. Di Tella (1995, 1997), Treisman (1998) and Laffont and N'Guessan (1999) among others used the ratio of imports in GDP to capture the competition and rents in the national market of countries.
However, using this indicator as a measure of competition is questionable. Lambsdorff (1999) argues that the ratio of imports to GDP is a distorted indicator of competition. According to him, this variable depends on country size. The larger countries may compensate their low import ratio by increasing competition in the national market. To overcome this distortion, I use the sum of imports and exports of goods and services normalized by the size of the economy (X+M)/GDP. As expected, we have a negative effect from this variable, that is, the more a country is opened to trade, the lower is its perceived rate of corruption. However, this effect seems to be sensitive to the introduction of other variables such as GDP and regional corruption. The last two columns ( (9) and (10)) confirm this assumption when I remove the GDP and the regional corruption variables. That is to say, trade openness is a potential channel through which the contagion effect of corruption is transmitted.
Countries that have high rates of raw material exports are expected to have high corruption levels. I used the export of natural resources (fuels, minerals) which may capture countries' endowment in natural resources (Treisman, 1999). Globally speaking, our results confirm this expectation.
As far as the foreign aid variable is concerned, my regressions show that its effect on corruption is negative but weak.
The expected theoretical relationship between corruption and political factors is ambiguous.
Many factors characterize different political systems: the characteristics of the political regime, press freedom, etc. Knowing that the correlation between these factors is quite high, it is inappropriate to introduce them individually in the corruption index. In consequence, to capture the effect of all these factors as much as possible, we used the variable accountability and voice which is from the database of Kaufmann, Kraay and Mastruzzi (KKM, 2003) as previously described. The results support the hypothesis that accountability and voice reduce national corruption. More importantly, my results suggest that accountability and voice contribute to explain the contagion effect of neighborhood corruption.
In columns (6) to (7) of Table 4, we introduce variables which capture social characteristics: ethnic fractionalization, religious fractionalization, British legal origin and French legal origin.
They are prospective channels through which regional corruption exerts its effects. This assumption is only weakly supported by our data, however. While ethnic fragmentation increases corruption, religious fragmentation on the contrary reduces it. As far as ethnic fragmentation is concerned, some authors explain this result by the fact that ethnic division slows down the level of development (Treisman, 1999;La Porta, 1997;Islam and Montenegro, 2002). Nevertheless, this result is not robust (when we drop the GDP variable from the regression, the effect becomes nonsignificant). But, in his empirical study, Treisman has found that the sign of this variable changes from positive to negative after controlling for GDP per capita. The author explains this result by the fact that ethnic division slows down development and therefore indirectly increases the level of corruption. In theory, the relationship between ethnic fragmentation and corruption could be positive or negative (Isham and Montenegro, 2002). In the same sense, we argue that if there is at least a viable institution which allows individuals to denounce power abuse, the relation should be positive.

3.3-Using an alternative measure of regional corruption
In addition to the previous measure of regional corruption, I used an alternative measure which is a weighted average of corruption. The existing literature debates what "neighbors" are exactly. Geographic distance may be a poor measure of economic distance for other countries not belonging to the neighborhood of a given country (Conley and Ligon, 2002). With globalization and the rapid development of new information and communication technologies, how more close is Côte d'Ivoire to Ghana or to China? I therefore assume, as do Moreno and Trehan (1997), that each country belongs to the neighborhood of every other country.
Clearly, for a given country the regional corruption is calculated as the weighted average of the levels of corruption of all countries in the world: the weight is the inverse of the distance separating the capital of this country from the capital of the others. Put differently, the relative importance of each country in a particular neighborhood varies inversely with its distance from the country whose neighborhood it is. Thus, for a given country, its neighbors' corruption is weighted more heavily than that of the remaining countries: is the distance between the capitals of two countries.
Statistics are reported below. Overall mean 6.07 3.78 4.08 3.90 The obtained figures represent the average corruption of the neighborhood of each country.
Sub-Saharan countries show higher neighborhood corruption than OECD countries and the sample overall average. However, the difference between Latin America and other developing countries seems not to be highly significant.
The alternative measure of regional corruption allows me to look at specific countries.
Four African countries (Nigeria, Zimbabwe, Tanzania, Zambia) are among the ten highest corruption-neighborhood locations. The ten countries with the lowest neighborhood corruption are developed countries (table 6). In this paragraph, we will discuss our previous results in regard to those stemming from the alternative measure of spatial corruption.
Looking at Table 7, the impact of spatial corruption on domestic corruption appears to be reinforced.
This result somehow contributes to explain why corruption persists at a high level in some groups of countries or at low level in others. Let us consider the case of African countries which are assumed to have a higher level of corruption than the other developing countries. These African countries are also supposed to share similar institutional or economic characteristics. Because of the contiguity and the spillovers effects, when the level of corruption increases in a given country (for example Togo), the consequence is an increase in corruption in a nearby country (for example Benin). As far as the two countries exhibit persistent increases in corruption, they will be locked into a vicious circle where corruption feeds corruption. In a symmetric way, any reduction in corruption in a single country, for example South Africa, would be associated with reduced corruption in neighboring countries such as Botswana or Zimbabwe.
What would happen if developed countries interact with developing ones? What will happen if a country like Senegal -where corruption is perceived to be high -interacts with France, a supposed low-corruption country? According to our econometric results, if the level of corruption decreases in France, the corruption level in Senegal will also decrease as a consequence. However, it is important to note that the direction of causality does not seem obvious. Will the level of corruption in France increase if the corruption level in Senegal increases? The answer may be yes. But the final level of corruption resulting from the interactions of these two countries may probably depend on some factors like the negotiation power or the influence of France.
The remaining results don't fundamentally change. That is to say, the levels of development, trade and political factors are the main channels of impact. 687 For a given specification, bold figures refer to variables. Student statistics are in parenthesis. *** means that the variable is significant at 1% level; ** means the variables is significant at 5% level; * means the variable is significant at 10% level.

Section IV-Checking the robustness of our results
The GMM used in the previous section prevents me to take into account the dynamic persistence of corruption because I could not introduce the lagged values of the dependent variable (because convergence conditions could not be met). One may question whether my analysis is sufficient to confirm definitely the spatial contagion effect. In order to clear up any ambiguity, I therefore use the feasible generalized three-stage least square results (3SLS) where data are stacked by periods. Unlike the two-stage least square (2SLS), the 3SLS method takes into account the correlation between the different equations of the system under consideration. By allowing for a correlation between error terms, I argue that the persistence idea in the perception of corruption over time is taken into account. With respect to spatial interactions, a shock in the error term in any country will be transmitted to all other countries through the multiplier effect (Anselin, 2002).
Tables 8 and 9 present 3SLS results for both the simple regional corruption and the weighted regional corruption. Most of our previous results are confirmed after I corrected for the dynamic correlations in the error terms. Specifically, regional corruption affects national corruption. Clearly, the reduction in the level of corruption in a country's neighbors is associated with a lower level of corruption in this country. Besides this result, the level of development as measured by per capita income is inversely related to the spread of corruption.  For a given specification, bold figures refer to variables. Student statistics are in parenthesis. *** means that the variable is significant at 1% level; ** means the variables is significant at 5% level; * means the variable is significant at 10% level.

Conclusion
"Governance matters", this phrase summarizes the new direction given to economic policies in developing countries since 1980. Combating corruption has become a priority because developing countries are more corrupted than rich ones. In this line, researchers are interested in investigating the factors of corruption in theoretical as well as in empirical studies so as to help policy makers. At the same time, in most of the studies the persistence of corruption has hardly been studied. In the present study, contrary to much of previous empirical analyses, I allow for the correlation between countries. I also assume that this persistence in developing countries could be explained by the contagion effect captured by the neighborhood corruption.
My results demonstrate a cross-country spillover mechanism where a lower regional corruption, as measured by the average of the level of corruption in one country's neighbors, is associated with a lower level of national corruption. Among the operating channels, the level of development seems to be the most important one. While foreign aid and trade exhibit similar results, voice and accountability are other possible interaction mechanisms. These results are robust to the temporal error correlation in corruption perception in each country. As such, my results may help to get new insights into the determinants of corruption and its spatial effect.
The results of this paper may point to certain policy recommendations. I argue that since national corruption is affected by neighborhood corruption, countries localized in either close or distant geographical areas should coordinate their individual anti-corruption policies. In this regard, our study highlights and supports the importance of existing and new regional or international agreements on combating corruption 16 . To fight the plague, these actions should be reinforced and made effective through the definition and the application of punishment rules, the possibility to launch international prosecutions against people if they try to evade arrest for corruption at home, etc.
Further, based on the fact that most existing conventions and treaties against corruption took root in developed countries and even though they expand positively their activities in developing ones, new bilateral and global policies that are initiated by and among the developing countries should be encouraged so as to involve them more actively in the combat. If countries do not succeed in collaborating, corruption in whole regions may settle on high equilibria levels or even expand. 16 Among others, we can cite the OECD Anti-bribery Convention (1996), the Inter-American Convention Against Corruption (1997), the European Action on Corruption in the Private Sector (1998)