Skip to main content


We analyze the determinants of corruption in Russia using law enforcement data on corruption incidents for a panel of 79 Russian regions for the period 2004–13. We find that the relative salaries of bureaucrats determine corruption levels: corruption declines as relative salaries rise, yet at strongly diminishing rates. Furthermore, we show that even very limited media freedom helps to curtail corruption. Other important determinants are the strength of law enforcement, education levels, and unemployment rates.

Online enhancement:   supplementary data.

1. Introduction

Low salary levels of public officials have long been regarded as one of the root causes of corruption among public officials (see, among others, Palmier 1985; Mauro 1997; World Bank 1997; Kaufmann 1997). Underpaid civil servants seeking to make ends meet or to achieve an income comparable to that of their peers in the private sector may be tempted to accept bribes in exchange for favors, such as government contracts, nonprosecution, easier licensing, and so forth.1 Increasing the remuneration of public officials has therefore been a key element of anticorruption strategies in many countries.2 The prime example of the success of this strategy is Singapore, which became one of the world’s least corrupt countries after increasing public officials’ salaries significantly and introducing complementary measures (Quah 2001).

Most theoretical contributions have supported the negative relationship between the relative salaries of public officials and the level of corruption. Starting with Becker and Stigler (1974), they have made efficiency-wage types of arguments—better-paid civil servants have more to lose (Cadot 1987; Andvig and Moene 1990; Bond 2008) and greater motivation and loyalty, as they are treated more fairly (Akerlof 1982; Akerlof and Yellen 1986). Bond (2008) suggests that high pay attracts honest individuals, thus improving the pool of candidates for public positions.

Yet Besley and McLaren (1993) show that raising civil service pay may actually increase corruption under certain conditions. Superauditors, responsible for the detection and prosecution of corruption, may be corruptible themselves. When the civil service is better paid, the superauditors can extract higher rents when agreeing not to prosecute corrupt officials and thus may be more inclined to engage in this type of corruption. This may lead to a lower probability of detection and higher overall corruption levels. Sosa (2004) shows that higher salaries can lead to more corruption if higher income reduces risk aversion sufficiently (and if penalties are not too high). It has long been recognized in labor economics that increased wealth may erode work incentives. For instance, Thiele and Wambach (1999) show in a principal-agent model that a wealthier agent will create a smaller surplus for the principal if the absolute risk aversion of the agent is relatively high compared to the degree of absolute prudence of the principal. Gneezy and Rustichini (2000) provide empirical evidence for a nonmonotonic relationship between wages and productivity. Since the opportunity-cost effect and the risk-attitude effect of higher wages are not mutually exclusive, the net effect is not clear a priori and may change sign over the relevant range. Mookherjee (1997) provides a model in which the wage-effort relationship has an inverted U shape; for our context his reasoning suggests a U-shaped relationship between the relative salaries of public officials and the level of corruption. Thus, given the number of different theoretical predictions, economic theory does not provide clear guidance on what the relationship between salary levels of public officials and corruption levels may look like, which makes the question essentially an empirical issue.

Yet there is no clear consensus in the empirical literature either. Van Rijckeghem and Weder (2001) find a significant negative relationship in a panel data set comprising 31 developing countries for the period 1982–94. This relationship, however, disappears when the authors look at within-country variations (using fixed-effects regressions). They use manufacturing wages as the reference remuneration and perceptions of corruption assembled in the International Country Risk Guide as a measure for corruption. Other cross-national studies such as Treisman (2000), Rauch and Evans (2000), and Pellegrini and Gerlagh (2008) find no significant effect. While insightful, these studies may potentially suffer from unobserved heterogeneity—countries may differ in dimensions not (sufficiently) controlled for that affect corruption and are correlated with the variables of interest, such as the quality of institutions, general attitudes, customs, and traditions. Moreover, for want of better internationally comparable data, they use perceptions of corruption as the measure of corruption. Perception data have been widely criticized for lack of validity (Knack and Keefer 1995; Golden and Picci 2005; Seligson 2006, Liu 2012). Perceptions of what constitutes corruption and how severe it is, whether they are held by experts or by the population at large, strongly vary across countries and are influenced by culture, traditional norms, and individual attitudes (Bertrand and Mullainathan 2001). Unsurprisingly, they have been shown to reflect the true extent of corruption very inaccurately (Mocan 2008; Donchev and Ujhelyi 2014). In a natural experiment, Olken (2009) finds that Indonesian villagers’ perceptions of corruption in a road construction project are only weakly correlated with more objective measures of missing expenditures (especially for material inputs). These findings cast doubt on the validity of perception-based studies of the determinants of corruption.

Recently, a small literature has emerged that analyzes the determinants of corruption at the national or subnational level using law enforcement data. This approach has two distinct advantages: First, studies using within-country variation suffer much less from unobserved heterogeneity and thus from omitted-variable biases, as unobserved determinants of corruption like institutions, tradition, histories, and so forth, are much more similar, if not equal. Second, law enforcement data are much more reliable, as they do not suffer from perception biases. Goel and Rich (1989) use conviction rates of corrupt officials at federal, state, and local levels in the United States and find that differences between the salaries of public officials and of middle-grade accountants affect the conviction rate. Goel and Nelson (1998) employ a US cross-state data set of convictions among public officials and find that high salaries reduce corruption. Glaeser and Saks (2006) use federal corruption conviction rates in the 50 US states and find that states whose populations are richer and better educated are less corrupt, as are those with lower levels of inequality and racial dissimilarity. They do not include a measure of civil servants’ relative salary. Contrary to Goel and Rich (1989) and Goel and Nelson (1998), Karahan, Razzolini, and Shughart (2006) find that corruption among the supervisors responsible for the governance of 82 counties in Mississippi is significantly positively correlated with their remuneration. Finally, Alt and Lassen (2010) find inconclusive support for a (negative) relationship between public officials’ pay and corruption across American states; a significant negative relationship is found only in the absence of fixed effects in the regression model.

While most of the literature focuses on the United States, Di Tella and Schargrodsky (2003) analyze corruption in public procurement in Argentina and find that, conditional on monitoring effort, high civil service wages decrease corruption. Del Monte and Papagni (2007) analyze the determinants of corruption in Italy and find that government consumption, level of development, and political culture influence corruption. Dong and Torgler (2013) use province-level data on registered cases of corruption in China to show that anticorruption campaigns, trade openness, media access, fiscal decentralization, and higher salaries for government employees decrease corruption, while social heterogeneity, natural resources, and regulation increase it.

Our paper contributes to this literature but differs in important aspects. First, we investigate the determinants of corruption in the Russian Federation using law enforcement data. Russia is a particularly interesting case: It is the ninth-largest country in the world and a former superpower, and it is geographically and socioeconomically very diverse. Corruption is rampant. In the recent Corruption Perception Index (CPI) compiled by Transparency International (TI), Russia ranks 136 out of 175 countries surveyed in 2014.3 Its federal structure and high geographical differentiation allow for a sound econometric analysis at the subnational level. Despite this, very little has been written on corruption in Russia from an economic perspective. Dininio and Orttung (2005) use experience-based corruption data to measure corruption but do not include the relative salary of public officials in their regressions. Given their small sample size of 40 observations in a cross-sectional approach (for 2002) and the missing variable of interest, their findings are of limited use for our purposes.

Second, we investigate whether a nonlinear relationship between public officials’ relative salaries and corruption exists. Even though the notion of a U-shaped relationship between public wages and corruption has been suggested in the theoretical literature (as discussed above; see Besley and McLaren 1993; Mookherjee 1997; Thiele and Wambach 1999; Sosa 2004), empirical studies have included relative wages of public officials only as a linear term, which disregards the possibility of a sublinear or even nonmonotonic influence of relative wages on corruption levels. We allow for such a possibility. Not only are our data conducive to a panel analysis; we are also able to use measures of relative salary that are much more precise than the broad measures used in cross-country analyses. We compare public officials’ remuneration to salaries in business counseling—an occupation that requires a skill set very similar to that of public officials.

Third, we analyze whether, in a country that has no free press by international standards, even relative freedom of the press (hereafter, press freedom) affects the level of corruption. Press freedom has been shown to reduce corruption in cross-country analyses (Brunetti and Weder 2003; Chowdhury 2004; Freille, Haque, and Kneller 2007), and information provided by newspaper campaigns was found to reduce leakage of funds from government programs in Uganda (Reinikka and Svensson 2005). We analyze whether variations in relative press freedom in a country have the potential to reduce corruption, an issue that has not been analyzed so far and can be studied only in a semiauthoritarian country like Russia where restrictions on press freedom can be differently enforced across regions and over time.

We use a panel data set covering 79 Russian regions and the period 2004–13 that we assembled from various government sources. As a primary measure of corruption we use corruption incidents registered by the police in a Russian state or region, data that were recently made available to us by the Russian Ministry of the Interior. We argue that—unlike in the United States—convictions may be open to political influence and thus less reliable as a measure than registration of incidents with the police. However, we use conviction rates as an alternative measure of corruption.

We find that corruption is negatively related to the level of civil servants’ salaries and that this influence is strongly sublinear. We cannot prove in a strict sense that the wage-corruption nexus is nonmonotonic, but we cannot exclude this possibility either. Our results are very robust with respect to the corruption measure, the salary concept, and the inclusion of different sets of controls. Using data from the Glasnost Defense Foundation (GDF), an independent nongovernmental organization (NGO), we find that even limited press freedom reduces corruption compared to a completely unfree press. Moreover, corruption is determined by the quality of law enforcement, opportunity costs in the labor market, and the average education level of the population.

Our paper proceeds as follows. In Section 2, we introduce our data. Section 3 presents our main findings from the panel regressions. Section 4 reports results using corruption experiences as a different measure of corruption, which corroborates our previous findings. Section 5 reports several robustness checks; notably, we use conviction rates as an alternative measure of corruption, and we examine alternative salary concepts. Section 6 concludes.

2. The Data

We investigate corruption in 79 regions in Russia in the period 2004–13. The time period was chosen because there are no law enforcement data available for earlier years and because Russia became involved in the Ukrainian conflict and annexed Crimea in early 2014, which triggered international sanctions and Russian countersanctions and led to a recession in 2014 (Doronina 2014).

By 2004, the first year of our analysis, Russia had fully recovered from the financial crisis of 1998 and had established a market economy. In the first part of the period examined, the economy grew by around 7 percent per year; it contracted by 7 percent in 2009, following the global financial crisis, and recovered in the following year, experiencing a modest growth of around 4 percent between 2010 and 2013.4 The political situation was stable, as federal power was firmly in the hands of a political elite concentrated around the Russian leader Vladimir Putin.5,6 Throughout the period, corruption was a major impediment to economic growth, but no radical steps were taken to reduce it except for an anticorruption campaign initiated by then-president Medvedev, comprising the National Anti-Corruption Plan introduced in 2009 and the Anticorruption Strategy in 2010. Even though the campaign was ambitious, there was almost no progress in the fight against corruption, as Medvedev later acknowledged (Lisova and Kostenko 2011).7

The 79 regions studied account for 99.2 percent of the population.8 They exhibit significant socioeconomic heterogeneity but are homogeneous in terms of official language, legislation, taxation system, and business regulations. Some regions have the status of a republic; while they have a legal status similar to the others, they have their own constitution in addition to the national one and are de facto more independent. They have a strong non-Russian minority or even a majority with its own official regional language, receive larger transfers from the center, and show stronger support for Russia’s ruling party than the average region (Jarocińska 2010).

2.1. Measuring Corruption

As corruption is clandestine and illegal, it cannot be measured accurately; however, the empirical literature suggests two possible proxies: data on the perceptions of corruption and law enforcement data. Indexes of the perception of corruption are widely used, especially in cross-national analyses, as they are readily available. Yet they have been strongly criticized for being subjective and biased (see, among others, Knack and Keefer 1995; Bertrand and Mullainathan 2001; Seligson 2006; Mocan 2008; Olken 2009; Donchev and Ujhelyi 2011). Criminal data reported by law enforcement agencies are a more objective measure of corruption, as they are not based on perceptions but on convictions or other measures of legal action. However, their quality depends on the quality of law enforcement. For this reason, they are appropriate only in within-country studies of countries where the police force is under central authority and where the legal system, operating procedures, and determination of law enforcement are the same across all units of observation.9 These proxies have been used for the 50 US states (Glaeser and Saks 2006; Goel and Nelson 2011; Alt and Lassen 2014); we use similar data for the 79 Russian regions. In Russia, the police are under central control, and thus operating procedures and determination are similar across regions, yet we control for differences in the resources and efficiency of law enforcement (as discussed below).

Our primary corruption measure (CORR) is the number of registered incidents of bribe acceptance in the region per 10,000 public officials. Russian Criminal Code (art. 290) defines bribery as the acceptance of money, securities, or other valuables by a public official (personally or through an intermediary) for his or her performance (action or inaction) for the benefit of a giver or an affiliated person, if such action implies that the public official exploits his or her position or authority or installs patronage. This is in line with the conventional definition of corruption as “the abuse of public power for private gain” (World Bank 1997, p. 102). Bribery necessarily implies that a public official is involved as a recipient. To account for the potential scope of bribe taking, we use the number of public officials as the denominator of CORR.

When we juxtapose our measure of corruption aggregated at the country level with the CPI compiled by TI for Russia, we observe similar trends (see Figure 1), which indicates that the measures may not be too dissimilar. Corruption levels increased until 2009 and decreased afterward.

Figure 1.
Figure 1.

Dynamics of corruption in Russia (percentage of the value in 2004)

Our primary corruption measure, CORR, differs from the conviction rate used by Glaeser and Saks (2006) and others in three important aspects. First, conviction is a product of completed judicial proceedings, while registering a crime is only the first stage of criminal prosecution. In an environment in which the judicial system may be corrupted as well (or politicized), conviction rates may be distorted much more than primary criminal statistics, such as the incidence rate that we use. While the registration decision may be in principle corrupted as well, we argue that this is much less of an issue, as police officers face the risk of being accused and prosecuted for failing to register a crime if the person reporting complains to a higher official. It is much easier to protract and effectively sabotage an investigation in exchange for a bribe through nontransparent and complicated investigation procedures than not to register a crime, which is an offense that is easy to prove.10 Second, convictions refer to the number of convicted individuals independent of the number of cases and not the severity or number of crimes they have committed, whereas each registered incident corresponds to one detected criminal act. The number of incidents arguably portrays the corruption landscape better, as it measures the frequency of corruption.11 Third, the time span required to register an incident is much shorter than that to complete criminal proceedings with a conviction: the legislation relating to criminal procedure requires the registration of incidents of corruption within 3 days of the moment of detection or notification. This allows a much timelier fit of the corruption measure, as an endogenous variable, with the explanatory variables—notably, relative-salary levels, which change over time. Data about incidents are more responsive than conviction rates to changes in actual corruption levels.

For these reasons, we think that the incident rate is better suited to a panel data analysis. Nevertheless, we use conviction rates as a secondary corruption measure in Section 5.1 to investigate if our results are affected by choice of corruption measure. We obtained both data sets from the Ministry of the Interior for the period 2004–13. Figure 2 depicts the distribution of incidents (for the pooled sample).

Figure 2.
Figure 2.

Distribution of incidents of corruption per region-year

2.2. Relative Salary

Russia’s bureaucracy has three tiers—federal, regional, and municipal—with the wages of the federal level set centrally and regional and municipal wages set at the regional level. The law relating to the civil service states as one of its principles that remuneration should be comparable across all levels of civil service (Federal Law No. N79-FZ [July 27, 2004], ch. 10). This principle governs the setting of the basic, guaranteed part of remuneration, such as pay for the position, grade of service, or seniority, but only governs to a much lesser degree the payment of bonuses. Yet there is no clear guidance for the salary level, at either the federal or the regional level. Gimpelson and Lukiyanova (2009) argue that wages for public officials are adjusted relatively spontaneously in response to changes in expected revenues and as a consequence of a bureaucratic bargaining process rather than with the goal of maintaining a certain wage differential vis-à-vis the private sector (see also Gimpelson and Kapeliushnikov 2011).

Figure 3 illustrates this point. It displays the recent dynamics in the aggregate relative salaries of public officials, using an average salary in the economy as the reference point, and shows that the ratio of public officials’ average salary to the average salary in the economy has fluctuated substantially over the years. (Of course, there is substantial variation across regions.) After the economic recovery following the crisis in 1998, Russia witnessed significant salary increases in nearly all industries of the private sector, while the public sector lagged behind. After the presidential election in 2004, civil service salaries were increased starting on January 1, 2005. Although it was not officially announced as an anticorruption measure, fighting corruption was the most commonly named reason for this increase in salary (Petrova 2004). The economic downturn after the global financial crisis of 2008 made the public sector financially more attractive; however, Russia’s fast recovery and the accompanying growth of salaries in the private sector again reduced the salary ratio between the public and private sectors.

Figure 3.
Figure 3.

Dynamics of the ratio of public officials’ salaries to average private-sector salaries in Russia.

Our relative-salary variable (RELSAL) is constructed as the ratio of the average monthly salary of a public official in the region12 to the average monthly salary in a comparable sector, lagged for 1 period. The data on public officials include civil servants with executive and legislative functions at the federal, regional, and local levels. These functions include taxation, budget execution, budget and taxation supervision, customs, procurement, inventory and estate management, socioeconomic planning, and governance. The numerator of the ratio reflects the legal earnings of the potential bribe takers.

The denominator of RESAL reflects the opportunity costs of being a civil servant (or the remuneration level that a civil servant may deem adequate) and thus indicates the temptation to accept bribes. While per capita gross domestic product (GDP) and average per capita income have been frequently used as a comparison (Treisman 2000; Pellegrini and Gerlagh 2008; Goel and Nelson 1998; Karahan, Razzolini, and Shughart 2006), they are extremely aggregate measures and can be misleading, as they do not portray the remuneration of the group of people to whose income civil servants compare their own. Moreover, the GDP per capita strongly depends on the sector composition of the regional economy, and therefore differences in relative wages constructed with regional GDP per capita as the denominator may not reflect differences in opportunity costs but rather differences in the sector composition of the regions. Civil servants may compare themselves to workers in the private sector whose occupations require comparable skills. As a solution, Van Rijckeghem and Weder (2001) propose using wages in the manufacturing sector, but they admit that their measure does not match the level of skill of the government employees.13

We argue that a natural alternative wage for public officials is the salary level in business counseling. The skill requirements and level of responsibilities are comparable, so counseling may be a realistic alternative for public officials. Goel and Rich (1989) use a similar strategy by comparing public officials’ pay to that of middle-grade accountants (as a measure of the average salary of white-collar professionals in the private sector), assuming that middle-grade accountants may have skill sets similar to those of the relevant public officials. While we consider their measure to be preferable to average manufacturing wages or GDP per capita, we think that in the Russian context salaries in business counseling are more appropriate because the job requirements and responsibilities are more comparable to those of public officials, who make decisions on a regular basis such as allocating procurement contracts or issuing licenses to businesses, especially if they are able to extort bribes. Especially for well-connected people—such as former bureaucrats—counseling may be an attractive alternative employment, as it involves helping businesses receive required licenses or contracts.14 According to the Russian Federal State Statistics Service (FSSS), business counseling incorporates financial management counseling, the development of accounting and controlling systems, human resources and marketing counseling, consulting about organizational planning, assessment of tangible and intangible property, public relations services, project leadership (management and supervision of resource allocation, quality control, and reporting), services for resolving industrial disputes, and other services pertaining to business operations. Therefore, business counseling resembles the executive and legislative functions of a government, since both sectors provide services to the same commercial entities in a region. Indeed, public salaries are within the same range as salaries in counseling: the average value of RELSAL is 1.03.15

The variable RELSAL is lagged 1 period. Our data are annual averages, which means that this year’s relative salary is the average salary of public officials divided by the average salary in business counseling, both averaged over all occupations in that group and the entire year. Part of the year is in the future, at the time when an official decides whether to engage in corrupt activities—thus, contemporaneous relative salary is unobservable; past relative salaries, in contrast, are well known. Moreover, some corrupt activities may take time to arrange or to be revealed, so the decision to engage in corruption may have been taken in the year before the incident was registered.

Data on the complementary components of civil service pay as the stability of employment and state pensions is not available. But on average, state employment is a very steady field, as 44 percent of civil servants have more than 10 years’ experience in municipal or state service (Statistical Newsletter 2005). Pensions, on the other hand, might be not relevant for reasons of uncertain future, ongoing pension reform, and high inflation and since pensions are not affected by penalties for bribery.

2.3. Other Determinants of Corruption

The primary concern with the criminal data on bribery is to control for the effectiveness of law enforcement. More effective law enforcement may deter corrupt activities (Becker and Stigler 1974), thereby reducing actual corruption levels. At the same time, it may detect a larger share of existing corruption, leading, ceteris paribus, to more registered incidents of corruption.16 While the deterrence effect may take some time to materialize, the second effect may be faster. To capture both effects, we introduce several measures of law enforcement. The first two measures pertain to the Ministry of the Interior (Ministerstvo Vnutrennikh Del, or MVD) of the Russian Federation. The resolution rate of major criminal offenses (SOLVE) reflects the efficiency of the agency in solving registered crimes. Yet it is conceivable that the resolution rate is high because the share of crimes that are registered is low. The resources available to law enforcement, normalized by population, serve as an additional proxy for the ability of the agency to fight crime. We use data on budgetary expenditures on law enforcement and security per capita (ENFORCE). Last, we use the number of employees in the federal courts per 10,000 population (JUDGE).17 Judicial employment is a proxy for effective prosecution in court and does not directly relate to our main dependent variable but to our alternative dependent variable, the number of convictions per 10,000 public officials. Nevertheless, a well-resourced judicial system may indicate a higher probability of conviction once the crime is registered and thus may create a significant deterrence effect.

Government size can be another determinant of corruption, as larger public budgets, especially for public procurement of nonstandard equipment, will increase opportunities for corruption (Shleifer and Vishny 1993). We use the logarithm of per capita real budget expenditures, excluding expenditures for law enforcement, the judicial system, and the total wage bill of public officials (GOV EXPEN), as a proxy for rents to be acquired through corrupt behavior (for instance, misappropriation of funds in exchange for kickbacks).

Natural resource rents often create a corrupt environment (Leite and Weidmann 1999; Kronenberg 2004; Goldberg, Wibbels, and Mvukiyehe 2008; Vicente 2010; Bhattacharyya and Hodler 2010; Van der Ploeg 2011). Firms in the extractive industries operate in noncompetitive market environments with high entry barriers and intensive regulations; as they are not footloose, they cannot escape government extortion. Given the regulatory environment, the high resource rents, and thus the high (potential) profits, bureaucrats and the firms’ executives may face strong temptations to engage in corruption. Gaddy and Ickes (2005) and Bradshaw (2006) suggest that corruption in Russia is particularly pronounced in the oil and gas sector; thus, resource-rich regions could be expected to have more corruption. We use the lagged value of extracted oil and gas in constant prices in a region divided by total population to capture the size of the natural resource rents in the region (OIL & GAS).

We also control for the socioeconomic profile of the regions. We include average income, which has frequently been used in corruption studies and found to be negatively correlated with corruption levels (for example, Meier and Holbrook 1992; Goel and Nelson 1998; Adsera, Boix, and Payne 2003; Glaeser and Saks 2006; Alt and Lassen 2014); our measure is the logarithm of income per capita in constant rubles (INC). We also include income inequality measured by the regional Gini coefficient (GINI), which has been found to be a strong determinant of corruption (see Gyimah-Brempong [2002] for Africa, You and Khagram [2005] for a cross-section of countries, Glaeser and Saks [2006] for the United States, and Dong and Torgler [2013] for China). Education should be negatively associated with corruption levels, as it empowers people to resist the demands of corrupt officials (Knack, Kugler, and Manning 2003; Glaeser and Saks 2006; Alt and Lassen 2014). We include the percentage of the economically active population with college, university, or other higher education (EDU). Data on education are composed on the basis of quarterly representative surveys by FSSS.18

The opportunity costs of corruption are measured by our variable of interest, RESAL, the probability of detection (captured by the law enforcement variables described above), and the ease of finding a different occupation once the public official has been convicted and fired (Van Rijckeghem and Weder 2001). We measure this third influence by the average number of months needed to find a job for all unemployed individuals in the previous period (UNEMPTIME).19 Our measure thus reflects the personal costs of being unemployed, which depend on labor market conditions known to the public official at the time the incident is reported.

Media freedom has been found in other contexts to curb corruption effectively through monitoring, information transmission, and exposure of corrupt officials (Brunetti and Weder 2003; Chowdhury 2004). While it is typically very challenging to find subnational data on press freedom in countries in which the media are not free, either because there is no regional variation or because the government disallows such analyses, we obtained data from the GDF, which conducted expert surveys across Russian regions in 2006, 2008, and 2010 on this topic.20 The GDF classifies the freedom of the print media and the electronic media in every region as completely free, relatively free, relatively not free, or not free. Since there are no completely free regions and it is not possible to quantify the differences between the remaining three categories, we construct a dichotomous variable that equals one if the media are relatively free, and zero otherwise, for the years available (FREE PRESS). These surveys were discontinued by GDF, as they stopped accepting foreign funding in order to avoid being classified as a foreign-sponsored NGO. The Russian parliament passed a law (Federal Law No. 121 [June 29, 2012]) that labeled Russia-based NGOs with foreign funding as foreign agents (inostrannie agenti), obliged them to use stricter reporting standards and undergo closer scrutiny, and made state intervention or interruption of the activities of these NGOs possible.21

In addition, we construct an extended variable measuring media freedom for earlier years by utilizing data from the Carnegie Foundation index of democratization of Russian regions, which was constructed as the average for the years 2000–2004 and includes a component measuring the independence of the media. It ranges from 1 (unfree) to 5 (free).22 For our binary variable of media freedom, we assign the value one to the regions that scored in the top two categories (4 or 5) in the Carnegie Foundation index. Because FREE PRESS is based on incomplete data, we investigate its effect on corruption in a separate set of regressions that cover the period for which data on press freedom are available (see Section 3.2).

We control for the existence of regional anticorruption laws. In 2009 a federal anticorruption law was introduced; it established legislative principles and instruments for fighting corruption, such as reviewing legislation for its potential to create entry points for corruption, creating awareness of and intolerance for corruption (for example, via advertising campaigns), and strengthening requirements for candidates for official positions.23 As the implementation of these measures at the regional level required the introduction of similar regional laws, such laws were subsequently established across the regions at different points in time. Some regions had already introduced anticorruption laws before the federal initiative. We control for anticorruption legislation that may significantly improve the conditions for detecting and prosecuting illegal actions by including a dummy variable ANTICORR LAW, which equals one if a region has an anticorruption law (for at least half a year).

We include a measure of urbanization (URBAN) because urban areas may facilitate criminal activities in general and corruption in particular, as suggested by Glaeser and Sacerdote (1999). The variable URBAN is defined as the share of the population living in urban areas; an area qualifies as urban if it is a settlement with at least 12,000 residents and 85 percent of the working population is employed in nonagricultural sectors. However, Becker, Mendelsohn, and Benderskaya (2012, pp. 19–20) warn that these criteria for urban areas in Russia are often applied arbitrarily and might even be politically determined. Therefore, we also control for infrastructure provision (which is better in urban areas) by including landline telephone density (TEL) in the region.

As standard controls we also include the log of population (POP) to capture scale effects and year fixed effects, which may capture common time trends such as changing priorities in law enforcement. The data are summarized in Table 1.

Table 1.

Variable Descriptions and Summary Statistics

View Table Image: 1 | 2

3. Results

3.1. Fixed-Effects Estimation

We estimate OLS regressions with robust standard errors clustered at the regional level, including a full set of region and time fixed effects. The dependent variable is the number of bribery incidents registered by the police per 10,000 public officials. Results are reported in Table 2.24

Table 2.

Fixed-Effects Estimates of Registered Corruption Incidents per 10,000 Public Officials

The most parsimonious specification, model 1, shows a strong negative effect of RELSAL, which suggests that a high relative salary for public officials reduces corruption significantly and strongly. This result corroborates the conventional wisdom. The inclusion of a nonlinear term for relative-salary levels in the parsimonious model increases in absolute value the negative linear effect of relative salaries on corruption, whereas the squared term is strongly positive and significant, which indicates that corruption declines significantly with an increasing relative salary for public officials, but at diminishing rates (model 2).

Our favorite specification is model 3, with the full set of control variables; results for all variables in the more parsimonious model 2 are unaffected by the inclusion of the control variables. The relative-salary level of public officials exerts a strong and highly significant negative effect on the level of corruption, which again is sublinear. For the quadratic specification of RELSAL (RELSALSQ), we calculate the turning point beyond which corruption increases again, which is a relative-salary level of 1.76. This is still within the sample range, as 6 percent of all observations are to the right of this turning point. Taken at face value, this would indicate that at very high relative-salary levels, a further increase in the salary levels of bureaucrats would actually increase corruption levels, which is consistent with the idea of Besley and McLaren (1993). Yet there may be too few observations to conclude that the nexus of corruption and relative salary is non-monotonic.25 What is clear, however, is that rising relative salaries of bureaucrats have strongly diminishing returns in terms of reduced corruption levels. In all specifications the nonlinear term is highly significant, and specifications including such a term consistently outperform specifications with a linear term only, as evaluated by the Akaike information criterion (AIC), the Bayesian information criterion (BIC), or the adjusted R2 statistic.

The effectiveness of law enforcement is measured by three variables: the annual resolution rate of serious criminal offenses by the police, the resources for law enforcement as measured by per capita expenditures for law enforcement, and the number of judicial staff and attorneys per 10,000 people. The resolution rate (SOLVE) is a very strong predictor of registered corruption: a 1-standard-deviation increase in the resolution rate (that is, an increase of 11 percentage points) increases registered corruption cases by one-fifth of a standard deviation (six cases per 10,000 public officials). The resolution rate measures the efficiency of law enforcement; it may thus proxy for the ability to detect clandestine illegal activities such as corruption. At the same time, the incentive to report corruption may increase with a higher resolution rate. The resources available to law enforcement authorities (ENFORCE) may proxy for the ability of law enforcement agencies to launch effective investigations; they may be most visible to the public and the potential bribe takers and as such may exert a deterrence effect. Indeed, the estimate for ENFORCE is negative, but it does not reach usual significance levels. The number of judicial personnel (JUDGE) is not significant in the models with standard controls.

Changes in urbanization and infrastructure levels do not affect the number of corruption incidents in the fixed-effects regression, possibly because such changes occur slowly from year to year. Coefficients for income, inequality, and government expenditures are insignificant as well. Increased oil and gas production does not affect corruption, probably because most rents do not accrue to the regions but are transferred to Moscow.26 The number of corruption incidents decreases significantly with increasing education levels in the population, as measured by the share of the economically active population with tertiary education. This may indicate that people with more education may be more likely to resist demands for bribes but also that better-educated people may hold their bureaucrats more accountable. This result is in line with the literature (Knack, Kugler, and Manning 2003; Glaeser and Saks 2006; Alt and Lassen 2014). An increase in population increases corruption; this may be a simple scale effect, as more inhabitants imply more potential bribers, and more populous regions may also have more complicated bureaucratic hierarchies, which may be prone to more corruption. The introduction of a regional anticorruption law significantly increases reported corruption incidents. These laws raise social awareness and provide additional opportunities for reporting corruption (for instance, through the establishment of anticorruption commissions), but they may not necessarily deter corruption, as argued by Batory (2012)—the target group of the laws may not be willing to comply, especially if the law receives little support from society. Our finding is in line with Goel and Nelson (2014), who find awareness of the whistle-blowers’ law in the United States to be positively correlated with conviction rates for corruption.

An increase in the average unemployment time significantly reduces corruption. Other things equal, a deteriorating labor market increases the costs of corruption and thus deters bureaucrats from engaging in corrupt activities.27 This mirrors our central result for the relative salary of bureaucrats: increased opportunity costs of corruption reduce corruption at the margin. These results are in the spirit of the literature on crime and punishment as pioneered by Gary Becker and George Stigler (Becker 1968; Stigler 1970; Becker and Stigler 1974).

3.2. Freedom of Press

In this section, we present the results for the effects of—partial—press freedom on corruption in Russia using data generated by GDF. These data are available only for the years 2006, 2008, and 2010 (see Section 2.3). To deal with this data issue, we follow different approaches. Table 3 presents the results. Regression (1) shows fixed-effects estimates for the years for which GDF data are available; regression (2) includes data from the Carnegie Foundation for the year 2004. In regression (3), we linearly interpolate by averaging the data for press freedom for the missing years 2005, 2007, and 2009 using the data in regression (2). Finally, regression (4) extends the data set to 2013 by using the data for 2010 for the subsequent years, as media freedom has been stable since 2010.28 All regressions include the control variables as in our baseline specification (Table 2, model 3); for brevity we report only the variables of interest.

Table 3.

The Effect of Freedom of the Press on Incidents of Corruption

All specifications provide similar pictures. First, even relative press freedom curtails corruption significantly and substantially. This corroborates earlier findings made in cross-country analyses (Brunetti and Weder 2003; Chowdhury 2004; Freille, Haque, and Kneller 2007) and evidence from a case study (Reinikka and Svensson 2005), in a novel context. We are able to show that even in a semiauthoritarian regime like Russia, there are differences in the degree of (relative) press freedom across regions, and these differences affect regional corruption levels significantly. All estimates are significantly positive and of similar magnitude. Second, our result for the effect of relative salaries for public officials on the corruption level remains unaffected by the inclusion of the variable for media freedom. Other results are likewise largely unaffected by the inclusion of the media variable and are available on request.

4. Experiences of Corruption

We consider law enforcement data to be the best measure of corruption. Yet if the legal system is systematically biased (especially if this bias is regionally different), convictions for corruption and, to a lesser extent, registered incidents of corruption will give a biased picture. In that case, the population’s experiences of corruption will be a better measure. To analyze whether data on such experiences portray a different picture, we analyze the available survey data on experiences of corruption on a regional level. As there are only two surveys (2002 and 2010), we employ a difference-in-differences approach and relate the change in corruption to the changes in relative salaries and the control variables.

In 2002, TI and the Information for Democracy Foundation (INDEM) surveyed 5,666 citizens and 1,838 representatives of small and medium enterprises in 40 Russian regions.29 These regions are home to 73 percent of the population and generally representative of the Russian Federation; however, they exclude the regions with populations of predominantly non-Russian ethnicity in the North Caucasus (Chechnya, Dagestan, and Ingushetia). The survey asked about perceptions and experiences of corruption. As we are skeptical about the accuracy of perceptions of corruption (see Section 2.1), we use only the results for experiences of everyday corruption.30 The survey calculates the amount of everyday corruption payments as a share of regional GDP, which is used as the measure of corruption in the region.

The second survey was conducted in 2010 by the Ministry of Economic Development of the Russian Federation and the Public Opinion Foundation by order of the president of the Russian Federation (Order No. 670 [March 14, 2010]) (Russian Federation 2011). It utilized the same methodology as the 2002 TI and INDEM survey and covered 70 regions. This allows us to analyze the dynamics of everyday petty corruption in the 40 regions included in both surveys. We use the following protocol: First, the 2002 value of petty corruption in a region as a share of regional GDP is subtracted from the 2010 value. Second, this regional difference is subtracted from the difference for the whole country during this period. The changes in everyday corruption (ΔCORR_EXP) are subsequently regressed on the changes in relative salaries and relative salaries squared and on a set of controls.31 This approach allows us to address the issue of omitted-variable bias stemming from unobserved—time-invariant—heterogeneity to obtain robust results. Descriptions of the data and descriptive statistics are given in Appendix B.

Table 4 provides the results. Model 1 contains the baseline regression for the full set of 40 regions. We include a dummy for republics, as the four regions in our sample with the status of an independent republic have a different political situation—they are strongly authoritarian, have a larger non-Russian population, support the ruling party more strongly, and enjoy more independence than the other regions.32 We expect the self-reported measure of experienced corruption to be biased in republics, as the population will be afraid of stating any corruption experiences. This concern is expressed by Dininio and Orttung (2005, p. 518), who analyzed corruption in the same 40 regions in 2002, and it may affect our analysis of the change in corruption over time as these republics have become even more authoritarian. Instead of including the dummy REPUBLIC, Dininio and Orttung (2005) exclude the republics of Baskortostan and Tatarstan as the main potential outliers from their regression model; we do the same in model 4. In model 5, we also omit the region Primorskii Krai, as suggested by Dininio and Orttung (2005), as it is infamous for widespread organized crime.

Table 4.

Determinants of Corruption Regressed on Measured Experiences of Corruption

Model 2 includes the variable ΔFREE PRESS, which captures the change in press freedom from unfree to relatively free by taking the value of 1 or from relatively free to unfree by taking the value of −1 (and 0 otherwise). For 2002, we use the regional data of the Carnegie Foundation, and for 2010, we use the GDF report from 2010 in the same way as it is done for the interpolation of FREE PRESS in Table 3, columns 3 and 4. The variable ΔFREE PRESS measures only the difference in status between 2002 and 2010 but not when the change took place. Since the effect of a change in press freedom on corruption dynamics may depend on the duration for which a region experienced improving or deteriorating press freedom, we use a second variable, ΔLASTING FREE PRESS, which measures the number of years in our 9-year period in which the press was relatively free. Models 3–5 use this variable instead of ΔFREE PRESS.

The results are in line with our previous findings. We find that experiences of petty corruption also decrease with relative salaries of public officials, and the effect is strongly diminishing. The calculated turning point has the same order of magnitude as before. Longer unemployment spells reduce corruption. While ΔFREE PRESS does not reach usual levels of significance, because of the reasons outlined above and possibly because of the small sample size, our alternative variable ΔLASTING FREE PRESS is significantly negative in all models, which indicates that even an only relatively free press may reduce the extent of (petty) corruption. The small number of observations may make these results illustrative and not entirely convincing by themselves, yet they corroborate our previous results with a different measure of corruption and thus reinforce our previous findings.

5. Robustness Checks

5.1. Conviction Rates as Endogenous Variable

We performed a number of robustness checks, the most important of which are reported below. The literature using law enforcement data almost exclusively relies on conviction rates.33 We argued in Section 2.1 that incidents of corruption are a better measure than convictions for corruption for a number of reasons, at least in the Russian context. In this section we examine whether the two measures of corruption are in accordance. We estimate our preferred specification, model 3 in Table 2, for two conviction rates as alternative endogenous variables. The term CONVICTION—ALL measures all convictions for bribery regardless of the severity of the offense, while CONVICTION—MAJOR measures the rate of conviction for crimes of corruption punishable by 5 years or more of imprisonment.34 Results of the fixed-effects regression with robust standard errors clustered at the regional level are reported in Table 5. For comparison, we repeat our preferred model for incidents of corruption in column 1.

Table 5.

Conviction Rates as Measure of Corruption

For our variable of interest the same picture emerges. Relative salaries have a significant negative effect on corruption; this effect is again nonlinear, with the calculated turning point at somewhat lower relative-salary levels (about 1.34 for all convictions and 1.20 for major convictions). The control variables display roughly similar features, although the significance levels differ; detailed results are available on request.

5.2. Alternative Salary Concepts

To test how robust our results are with respect to our relative-salary measure, we employ two alternative concepts for the wage to which public officials might compare their salaries, one of which has been used in the literature. We argued in Section 2.2 that most of the reference salaries used in the literature are too broad and do not reflect the specific skill set of public officials in executive or legislative functions.35 We therefore use salaries in business counseling as the reference, as the skills required for this occupation come closest to those of public officials. In this section, we investigate whether our results are driven by this particular choice of reference category.

As an alternative reference salary, we use the salary of white-collar workers since their job responsibilities and their education are broadly comparable to those of public officials. For the same reasons of data availability as in Van Rijckeghem and Weder (2001), we use data from the manufacturing sector, which is substantial in all regions. Since the civil service has different hierarchy levels (ranks) and the composition of these ranks differs across regions, we construct a reference white-collar salary for each region reflecting its specific composition. We match hierarchy levels in civil service with the corresponding levels of white-collar jobs in manufacturing and calculate the reference salary as a weighted average using the white-collar manufacturing wages for each level and the share of civil servants at the corresponding level in the civil service of that region. Details of the measure’s construction are provided in Appendix A. As a second alternative measure, we use the average salary in the economy to investigate whether a broader measure would reveal similar corruption dynamics.

Table 6 shows our results. Regression (1) is our baseline model, with business counseling as the reference salary, and regression (2) presents the results for the alternative reference salary of white-collar workers in manufacturing. We find a very similar nonlinear relationship for this new measure of relative salaries. Regression (3) reports estimates for the average salary in the region as the reference salary. The results are no longer significant for this broader concept of relative salary, but they still exhibit the same pattern, which is not surprising given that the reference category is quite imprecise.

Table 6.

Alternative Relative-Salary Concepts

Overall, the results show that our findings are not limited to the relative-salary concept that we use. Our results are more pronounced if we use a reference salary that requires a comparable skill set.

5.3. Placebo Test

In this section, we investigate whether a low relative salary is indicative of a crime-ridden environment in general or a situation that is specifically conducive to corruption. If relative salaries are negatively correlated with crime levels in general, our results would not explain the occurrence of corruption as such, but corruption as an endogenous variable would only be indicative of a crime-intensive overall environment. Such a situation could occur, for instance, if low relative salaries—including those of the law enforcement community—erode the motivation and zeal of the police force to fight crime.

To test for such a possibility, we estimate our baseline specification with different endogenous variables capturing important aspects of the overall crime environment. We use statistics for four types of crimes not related to corruption: the number of all crimes excluding incidents of corruption (CRIME − CORR), the number of murders and attempted murder (MURDERS), the number of drug-related crimes (DRUGS), and the number of registered thefts (THEFTS), all per 10,000 public officials. The results are reported in Table 7.

Table 7.

Correlation of Salaries and Various Types of Crime

All crime variables are uncorrelated with relative-salary levels, except for corruption. Moreover, changes in the time needed to find a job if unemployed and the introduction of anticorruption laws affect the level of corruption but not of any other crime.36 Thus, we can exclude the possibility that our previous results are only indicative of a crime-intensive environment and conclude that low relative salaries for public officials are conducive to corruption in the public sector.

5.4. Additional Robustness Checks

We analyze whether ethnic heterogeneity has affected corruption levels in Russia. Ethnic fragmentation has been shown to negatively affect economic policy, increase favoritism, and reduce economic growth (Easterly and Levine 1997; Alesina et al. 2003; Alesina and La Ferrara 2005; Franck and Rainer 2012). Olken (2006) finds for Indonesia that corruption is greater in ethnically divided villages. This corroborates earlier findings of cross-country studies based on perceptions of corruption (Mauro 1995; La Porta et al. 1999). We use the index of ethnic fragmentation and the index of ethnic polarization in a region to study the effect of ethnic heterogeneity. To construct both indices, we follow the methodology of Montalvo and Reynol-Querol (2005) and use the data from population censuses in 2002 and 2010 (we interpolate the values for years in between).37 These variables have very little variation over time and thus should be captured by fixed effects, yet we still estimate a pooled OLS and a between-effects panel model. In none of the specifications do the variables reach usual significance levels.

In one specification, we exclude the Caucasian regions from our preferred model, as corruption levels and institutional quality may arguably be very different there (Dobler 2011). In other specifications, we exclude regions with the status of a republic or insert a dummy controlling for them. In all cases, the results do not significantly change. We also use Cook’s distance Di (Cook 1977) to identify potential outliers and exclude all observations with a distance larger than 4/n, with n denoting the number of observations. We identify 39 outliers out of 790 observations for which Di > .005. Excluding them does not change the results in any significant way. (Results are available on request.)

Another concern may be that if police and bureaucrats’ salaries are highly correlated, our variable of interest, RELSAL, may capture police efficiency rather than the incentive of public officials to engage in corruption.38 This would be the case if higher police salaries attract more able personnel or enhance the loyalty and determination of the existing police force. To investigate such a possibility, we include police salary and police salary squared, normalized as in our variable of interest in our preferred specification (Table 2, model 3). Police salary has no significant influence on incidents of corruption, and our variable of interest exerts the same influence as before. That makes us confident that we can exclude the possibility that our results are driven by police efficiency.

6. Concluding Remarks

The main finding of this paper is a strong causal nonlinear relationship between the relative salary of public officials and the number of incidents of corruption registered by the police or resulting in conviction by the courts. Increases in salary reduce corruption significantly, but with strongly diminishing returns. We are able to show that this relationship is very robust with respect to the corruption measure used—incidents, convictions, or experiences of bribery—and to the relative-salary measure used. Essentially, we prove an opportunity cost argument in the spirit of Becker and Stigler with regard to the relative salary of public officials; this is corroborated by the finding that unfavorable labor market conditions, as measured by the average time needed to find a new job, also significantly reduce corruption. The duration of unemployment codetermines the opportunity costs of being corrupt.

In addition to our focus on the Russian Federation, our findings are novel in three dimensions. First, we are able to show in a fixed-effects setting that relative salaries of public officials matter for corruption levels. Previous cross-country studies yielded inconclusive results once they controlled for unobserved heterogeneity through country fixed effects. A second novel result is that increasing relative salaries of public officials may lose its effectiveness; it is highly efficient at low and medium relative-salary levels but may be ineffective, or even counterproductive, at high and very high relative-salary levels. Third, we are the first to show that in a country that restricts freedom of the media substantially and lacks democratic accountability at the national level such as Russia—which ranks 148 out of 180 in the World Press Freedom Index (Reporters without Borders 2014)—variations in relative freedom of the press contribute significantly to explaining regional differences in corruption levels. In other words, even partial freedom of the media in a generally unfree environment leads to substantially lower corruption levels.

Our results have important implications for policy reforms. While increasing public officials’ salaries may be an integral part of an anticorruption reform, it is no silver bullet. Salary adjustments need to be carefully designed, with the proper reference salary identified, to achieve the efficient level. Increasing salaries is costly, and the gains in terms of reduced corruption may increasingly dwindle, if not disappear. In contrast, granting more freedom to the press, even if only partially, unambiguously reduces corruption.

Appendix A Construction of the Relative-Salary Measures

The FSSS publishes the composition of the civil service’s executive branch by rank for all Russian regions for the end of the years 2005 and 2011. The four ranks are top administrators, advisers, specialists, and supporting specialists. As data are available only for 2 years, we interpolate the composition of the ranks in the public service for 2006–10 as a linear projection, assuming that the composition did not experience a dramatic shock in any region. We have no information of such events.

The FSSS database provides data on average monthly salaries in the manufacturing sector for the following categories of employees: (a) executive managers and head administrators, (b) high-level specialists, and (c) middle-level specialists (other categories are [d] employees with supportive functions: serving, keeping records, and paperwork, [e] workers in housing sectors, trade, and similar sectors, [f] qualified workers in construction, transport, mining, and manufacturing, [g] qualified workers in mechanics and machinery operators, and [h] unqualified workers). Data are available for 2005, 2007, 2009, and 2011; we interpolate for the missing years.39

The average salary of public officials in a region is calculated as the weighted sum of the salaries of public officials in the executive branch (top administrators, advisers, specialists, and supporting specialists) using regional employment shares as weights.

The reference white-collar workers’ salary is calculated as the weighted sum of incomes in the manufacturing sector of levels a–c above (top managers to middle-level specialists), using the weights of the corresponding levels in the civil service. Top administrators and advisers correspond to executive managers and head administrators in the manufacturing sector, specialists in the civil service correspond to high-level specialists in manufacturing, and supporting specialists in the civil service correspond to middle-level specialists in manufacturing. Categories d–h in the manufacturing sector are not considered, as they are not white-collar workers.40 This white-collar salary represents the average alternative salary that a public official would receive in the manufacturing sector.

Figure A1 illustrates how we constructed the alternative salary for the Saratov region for 2005. The numerator of the relative-salary measure is the same as in our primary variable RELSAL, but the denominator is an alternative white-collar salary. Matching shades illustrate the weighting of the alternative salary with respect to the composition of the civil service.

Figure A1.
Figure A1.

Construction of the alternative relative-salary measure (white-collar workers) for the Saratov region, 2005

Appendix B Data for Experiences of Corruption

Table B1.

Experiences of Corruption: Variable Descriptions and Summary Statistics

Appendix C Variables and Sources (in Russian)

ANTICORR LAW. ConsultantPlus (legal database) (

CONVICTION—ALL. Ministry of the Interior. Available from the authors on request.

CONVICTION—MAJOR. Ministry of the Interior. Available from the authors on request.

CORR. Available from the authors on request. For 2009–15: Ministry of the Interior, Crime in the Region (

ΔCORR_EXP. Ministry of Economic Development, The State of Domestic Corruption in the Russian Federation (

CRIME − CORR. Federal State Statistics Service, Regions of Russia: Socioeconomic Indicators (

DRUGS. Federal State Statistics Service, Regions of Russia: Socioeconomic Indicators (

EDU, ΔEDU. Federal State Statistics Service, Regions of Russia: Socioeconomic Indicators (

ENFORCE, ΔENFORCE. Federal Treasury, The Consolidated Budgets of the Subjects of the Russian Federation and of the Budgets of Territorial Governments’ Extrabudgetary Funds (

FREE PRESS, ΔFREE PRESS. Glasnost Defense Foundation, Monitoring: Glasnost Maps (

GINI, ΔGINI. Federal State Statistics Service, Unified Interagency Informational Statistical System, The Gini Coefficient (Income Concentration Index) (

GOV EXPEN, ΔGOV EXPEN. Federal Treasury, The Consolidated Budgets of the Subjects of the Russian Federation and of the Budgets of Territorial Governments’ Extrabudgetary Funds (

INC, ΔINC. Federal State Statistics Service, Unified Interagency Informational Statistical System, Monetary Income (Average per Capita) (

JUDGE. Federal State Statistics Service, Central Statistical Database (

ΔLASTING FREE PRESS. Glasnost Defense Foundation, Monitoring: Glasnost Maps (

MURDERS. Federal State Statistics Service, Regions of Russia: Socioeconomic Indicators (

OIL & GAS, ΔOIL & GAS. Federal State Statistics Service, Regions of Russia: Socioeconomic Indicators (

POP, ΔPOP. Federal State Statistics Service, Unified Interagency Informational Statistical System, The Resident Population as of January 1 (

RELSAL, ΔRELSAL. Federal State Statistics Service, Unified Interagency Informational Statistical System, The Average Monthly Nominal Wage per Employee for the Full Range of Organizations (

RELSAL_WC. Federal State Statistics Service, Unified Interagency Informational Statistical System, Average Paid Salary of Employees in Organizations by Professional Groups (

REPUBLIC. Russian Constiution, Chapter 3: The Federal Structure (

SOLVE. Ministry of the Interior, State Crime (https://xn--b1aew.xn--p1ai/folder/101762).

TEL, ΔTEL. Federal State Statistics Service, Regions of Russia: Socioeconomic Indicators (

THEFTS. Federal State Statistics Service, Regions of Russia: Socioeconomic Indicators (

UNEMPTIME. Federal State Statistics Service, Regions of Russia: Socioeconomic Indicators (

ΔUNEMP. Federal State Statistics Service, Regions of Russia: Socioeconomic Indicators (

URBAN, ΔURBAN. Federal State Statistics Service, Unified Interagency Informational Statistical System, The Resident Population as of January 1 (


We are indebted to Sam Peltzman and an anonymous referee for advice; we are also grateful to Maryana Antipova, Bernd Fitzenberger, Krisztina Kis-Katos, and Judith Müller for helpful comments and to Jakob Schulze for significant input. The usual disclaimer applies.

1. Governments lacking adequate funds to pay their employees may rely on civil servants to supplement their incomes through corrupt activities (Besley and McLaren 1993). McLeod (2008) argues that civil servants in Indonesia under President Soeharto were deliberately underpaid and expected to increase their income through corruption, which made them a part of a corrupt system that benefited the ruling family.

2. For instance, Peru, Argentina, Georgia, Nepal, and Ghana (Tanzi 1998; Transparency International Georgia 2011; Regmee 2001; Chand and Moene 1999).

3. See Transparency International, Corruption Perceptions Index, 2014: Results ( The scores range from 0 (most corrupt) to 100 (not corrupt); Russia’s score is 27.

4. World Bank, Data: GDP Growth (Annual %) (

5. Even mass social protests against electoral fraud during the parliamentary election in December 2011 were widely regarded as having failed to influence the current political situation (Volkov 2012).

6. Putin, having served as prime minister, became interim president on December 31, 1999, after Boris Yeltsin’s resignation, and was elected president on March 26, 2000, for his first term. He was reelected on March 14, 2004, for his second term, which ended on May 7, 2008. As the Russian constitution prohibits three consecutive terms, Dmitry Medvedev served as president from May 7, 2008, to May 7, 2012, during which time Putin served as prime minister. Putin was reelected as president on March 24, 2012, and his third term will expire on May 7, 2018. The current prime minister is Medvedev.

7. For more information about the anticorruption campaign and its results under the Medvedev presidency, see Ninenko (2012).

8. The Russian Federation currently has 83 administrative regions. Chechenya is omitted because of the ongoing military conflict there; three autonomous districts are aggregated with neighboring regions for the official statistics because of data limitations (the autonomous district Nenets is combined with the Arkhangelsk region, and the districts Khanty-Mansi and Yamalo-Nenets are combined with the Tyumen region). The 83 regions consist of 21 republics, nine krai, 46 oblast, two cities of federal importance, one autonomous oblast, and four autonomous districts (okrug). In 2004, there were 89 regions; six small autonomous okrug were merged with the neighboring larger jurisdictions in 2005 (one), 2007 (three), and 2008 (two). As these districts are very small (often with fewer than 100,000 inhabitants) and the mergers were already planned in 2004, the Federal State Statistics Service did not provide separate data for the soon-to-be-merged districts. We therefore use the current delineation of regions.

9. For this reason, Glaeser and Saks (2006) use data from the Federal Bureau of Investigation rather than from local or state police.

10. Before 2003, the police were paid bonuses solely based on their resolution rate, which might have created an incentive not to register crimes that are difficult to solve. This practice was discontinued after the Ministry of the Interior initiated a police reform introducing a different motivational mechanism (Russian Federation 2002).

11. Moreover, conviction data are often quite noisy (Alt and Lassen 2014), not least because one case can lead to more convictions, making the data lumpy. A potential drawback of data on incidents is that some of the registered incidents of corruption may be unfounded, which might make the data imprecise.

12. The measure includes all paid salaries and various monetary and nonmonetary compensations for labor, bonuses, vacation pay, and so forth.

13. In cross-country studies, data availability often leaves little choice in selecting a comparable remuneration, which casts some doubt on the explanatory power of such studies.

14. A prime example may be Roman Putin, a first cousin of Vladimir Putin, who opened a consulting agency to promote foreign businesses in Russia, See Roman Putin Consulting (

15. We interpolate the data on salaries in business counseling for two missing data points (for Khakassia Republic for 2009 and 2010).

16. People may be more inclined to report corruption if they are convinced that the police will be effective in building a case against the perpetrators.

17. The category of federal courts includes district courts, which deal with criminal offenses such as corruption as a court of first instance, and arbitral courts, which act as courts of second instance for criminal offenses. It does not include magistrates and constitutional courts, which are irrelevant to our analysis.

18. Income per capita, education, and inequality may be affected by corruption as well, so these variables may be endogenous. Gundlach and Paldam (2009) find that long-run causality runs from high income to low corruption levels and not conversely. Still, we cannot exclude the possibility of endogenous regressors, which would make their point estimates less reliable, especially in pooled ordinary least squares (OLS) regressions. In fixed-effects regressions using annual data, this is much less of an issue, as the time for feedback effects is comparatively short. As the estimates for our variable of interest are quite similar in pooled OLS and fixed-effects regressions, and as the omission of income per capita, inequality, and education from the regression model does not change the estimates of our variable of interest in any significant way, we believe that this is not a major issue in our context. Except for Glaeser and Saks (2006), no other study using conviction rates instruments for these variables.

19. Ideally we would like to include the unemployment rate for individuals with a comparable skill set, but these data are unavailable.

20. For data and descriptions, see Glasnost Defense Foundation, Monitoring: Glasnost Maps ( (in Russian).

21. Federal Law No. 121 de facto forced many Russian nongovernmental organizations (NGOs) to decline foreign funding to avoid being classified as foreign agents by Russian authorities. The law was generally seen as a measure against politically active NGOs and activists who criticized the Russian government; see, for example, BBC (2012).

22. For data and description, see Independent Institute for Social Policy, Social Atlas of Russian Regions: Integrated Indexes ( (in Russian).

23. Most regional laws provide guidelines for establishing anticorruption commissions as well.

24. We also estimated pooled OLS specifications with the same set of explanatory variables and, in one specification, with dummies for seven of the eight federal administrative districts. Results are remarkably similar to the fixed-effects specifications; they are available on request.

25. We are grateful to the editor for pointing this out.

26. Bradshaw (2006) argues that profits are transferred to the company headquarters, which are located mostly in Moscow and, for Gazprom, in St. Petersburg; taxes accrue to the center and are only partially and nontransparently redistributed to the resource-rich regions. Part of the rents may be retained by the regions through inefficient and costly regional inputs, and others may take the form of corruption; yet since the licensing procedures are likewise nontransparent and involve many government bodies, the regional incidence of resource-related corruption remains unclear (Gaddy and Ickes 2005; Bradshaw 2006).

27. The expected costs of corruption are the probability of detection and punishment, captured by the law enforcement variables described above, multiplied by the punishment according to the criminal code, which is equal across regions; a potential wage discount in the private sector, as measured by our variable of interest, RELSAL; and the costs of finding alternative employment after being removed from office.

28. Freedom House ranked media freedom in Russia consistently at 80–81 points (out of 100, with 100 indicating absolutely no media freedom) for the period 2010–13, whereas it reported deteriorating overall scores for press freedom in the preceding years. See Freedom House, Russia: Freedom of the Press 2010 (, with links to the reports for preceding and following years.

29. For the data, see Information for Democracy Foundation, Regional Indexes of Corruption ( (in Russian).

30. The 2002 survey also asked businessmen for their experiences of corruption, but since the 2010 survey did not, we focus on petty corruption only.

31. Controls are not exactly the same as in the previous sections, as not all data are available for 2002.

32. The republics in our sample are Bashkortostan, Tatarstan, Udmurtskaya, and Karelia.

33. Conviction rates are used by Goel and Rich (1989), Goel and Nelson (1998), Glaeser and Saks (2006), Karahan, Razzolini, and Shughart (2006), and Alt and Lassen (2014) for the United States and by Di Tella and Schargrodsky (2003) for Argentina; exceptions are Del Monte and Papagni (2007) and Dong and Torgler (2013), who use corruption incidents reported to the police for Italy and China, respectively.

34. We obtained data on all convictions for 2006–12 and on major convictions for 2004–12.

35. Goel and Nelson (1998) and Karahan, Razzolini, and Shughart (2006) use per capita income, Dong and Torgler (2013) use average salary in the economy, and Van Rijckeghem and Weder (2001) focus on the average salary in the manufacturing sector for the reason of comparability across countries. Goel and Rich (1989) use middle-grade accountants’ salaries as the reference.

36. We find similar results for pooled OLS regressions, which are available on request.

37. Census data can be accessed at the official Russian census database (

38. We are grateful to the editor for raising this point.

39. Federal State Statistics Service, Unified Interagency Informational Statistical System, Average Paid Salary of Employees in Organizations by Professional Groups ( (in Russian).

40. The salary of head administrators and executive managers in the manufacturing sector is multiplied by the share of top administrators and advisers in the civil service, the salaries of high-level specialists in manufacturing by the share of specialists in the civil service, and the salary of middle-level specialists in manufacturing by the share of supporting specialists in the civil service. The weighted salaries are then summed to form the reference white-collar salary in the region.