MSc Economics
Development Economics Track
Master Thesis
Effectiveness of Anti-Corruption Agencies in Curing
Corruption: Empirical Evidence from Serbia
Miloš Drezga
11642289
August 2018
Supervisor:
Statement of Originality
This document is written by Student Miloš Drezga who declares to take full responsibility for the contents of this document.
I declare that the text and the work presented in this document are original and that no sources other than those mentioned in the text and its references have been used in creating it.
The Faculty of Economics and Business is responsible solely for the supervision of completion of the work, not for the contents.
Abstract
Establishing independent Anti-Corruption Agencies (ACAs) has become a go-to policy response go-to widespread corruption in many countries across the globe. Although the merits and challenges of such a policy have been extensively discussed, there is no empirical evaluation of the effectiveness of the intervention in fighting corruption. Employing a quasi-experimental design, this paper presents first empirical evidence on the efficacy of an ACA in Serbia, a small European transition economy, using firm-level micro-data. I find that the intervention had no effect on corruption. I hypothesize that the reason no effect was detected is due to the lack of adequate operating environment. The absence of genuine political will to suppress corruption reflected in explicit and implicit forms of sabotage of the Agency committed by various state institutions whose support the Agency required to cure one of the burning issues of the Serbian economy.
Contents
1 Introduction ... 1
2 Background and Context: Anti-Corruption Agency of Serbia (ACAS) .. 5
3 Data ... 9
3.1 World Bank’s Enterprise Surveys ... 9
3.2 Comparisons and Descriptive Statistics ... 10
3.3 Missing Data ... 17 4 Methodological Considerations ... 18 5 Results ... 22 5.1 Inter-regional Analysis ... 22 5.2 Inter-country Comparison ... 28 6 Conclusion ... 33 References ... 34 Appendix ... 37
1 Introduction
Corruption, a widely discussed topic in economic literature, is often defined as the ‘misuse of public office for private gain’ (Svensson, 2005, p. 20). Common examples of such behavior are bribery, kickbacks in government procurement, and embezzlement of government funds (Svensson, 2005, p. 20). The interest this topic sparks stems from the idea that these irregularities in public life act as a distortive force for the economy. From a theoretical perspective, corruption leads to serious misallocations of resources. First, it causes management issues regarding public funds, affecting both their collection and spending/investment, which hinders the government’s ability to carry out policies aimed at socio-economic development. Second, corruption results in misallocation of entrepreneurial talent, skills, and capital to purposes that are less socially beneficial, such as specializing in rent-seeking activities to ensure preferential treatment, or operating largely in the informal sector to reduce demand for governmental services (Svensson, 2005).
Empirical literature confirms these theoretical predictions. Mauro (1995) presents a pioneer cross-country study on the effect of corruption on economic growth, making use of Business International indices on institutional quality. Utilizing the instrumental variables methodology with ethnolinguistic fractionalization as the instrument to address potential reverse causality issues, Mauro (1995) finds that corruption significantly reduces economic growth. This influential study started the wave of empirical work on corruption and proposed general guidelines regarding methodology used to investigate a complex, multifaceted issue that corruption is. Modeled after Mauro’s work, more recent literature utilizes newly available data, analyzes this relationship in specific sets of countries to check the robustness of findings from large cross-country exercises, and considers other socio-economic development indicators. There seems to be a general consensus in the literature in that corruption bears unfavorable effects on indicators such as economic growth, investment, income inequality, and poverty (see among others, Aidt, Dutta, & Sena, 2008; Damijan & Damijan, 2017; Gupta, Davoodi, & Alonso-Terme, 2002; Gyimah-Brempong, 2002). For a more detailed overview of this literature, see Svensson (2005).
While indicative of the corruption’s omnipresence and its negative consequences for the economy in general, the aforementioned empirical work does not provide evidence of corruption’s ‘every day’ effects on the private sector that acts as the engine of economic development in many countries across the globe. Understanding these micro-level effects is crucial in designing adequate anti-corruption policies that would limit the corruption’s economy-wide adverse effects. Empirical literature using firm-level data provides valuable insights in this regard. Fisman & Svensson (2007) make use of a Ugandan firm survey that contains data on estimated bribe payments to
model the relationship between corruption, taxation and firm growth. Their results show that, although both taxation and bribe payments are negatively correlated with firm growth, a 1 percentage point increase in bribe rate corresponds to a 3 percentage point decrease in firm growth expressed in form of sales, making the magnitude of the effect of bribery 3 times higher than that of taxation. This evidence captures the motivation behind the decision of entrepreneurs to operate within the informal economy, thus minimizing their demand for public services and exposure to corruption. Fisman (2001) studies the importance of corruption in generating firm value and growth. The author takes advantage of Indonesia’s centralized and stable political structure during President Suharto’s reign to construct an index measuring the extent to which Indonesian firms depend on their political connection for their profitability. To understand the value of these connections, Fisman (2001) exploits episodes of rumors about the president’s health to compare returns of firms with different levels of connectedness to the Indonesian political elite. As the returns of the well-connected firms suffer more from more severe rumor episodes compared to less-connected firms, the study concludes that specialization in rent-seeking activities and corruption does result in greater firm value and growth. Studies utilizing subjective measures of corruption also display corruption’s detrimental effects on firm operations. Using firm-level data on 80 electricity distribution firms from 13 South American countries, Dal Bo & Rossi (2007) find that more corruption in the country of operation is strongly associated with larger firm inefficiencies, since firms in more corrupt countries utilize more inputs to produce a given level of output, which highlights corruption’s distortive effects through misallocation of entrepreneurial skills and effort. The authors make use of Transparency International and International
Country Risk Guide corruption perception indices and show that their results
are robust to the choice of the perception index used in the analysis.
The most common policy response to widespread corruption is the establishment of independent Anti-Corruption Agencies (ACAs) that have a threefold function of ‘investigation, prevention, and education’ about corruption (Heilbrunn, 2004, p. 10). The core role of ACAs is to act upon reports of corrupt behavior of public officials filed by legal (firms) or natural (individuals) persons. Based on these reports, the ACAs investigate cases and prosecute those responsible. The implicit assumption of this policy is that having an independent body responsible for better monitoring of public officials and enforcement of rules and regulations will ultimately lead to a decline in corruption and increase the efficiency, credibility, and accountability of public institutions, thus improving the business climate and confidence of entrepreneurs and stimulating economic activity (Svensson, 2005). Additionally, it is argued that revenue collected by the government through putting more and more entities into the formal sector as a result of the ACA’s anti-corruption initiatives can exceed the cost of establishing and running them (Wei, 1999).
The success stories of Singapore’s and Hong Kong’s ACAs serve as the main motivation for the replication of the policy in other countries – both Hong Kong and Singapore saw a sharp decline in corruption that came together with the establishment of their respective ACAs (Quah, 2010; Wei, 1999). The main characteristic of ACAs responsible for their success is their indisputable power in implementing their initiatives. As Svensson (2005) documents, the Hong Kong’s Independent Commission Against Corruption (ICAC) made ‘legal precedents such as guilty until proven innocent’ (p. 35) that increased the efficiency with which corruption cases were processed and sent a signal to public officials that any form of misconduct would not be tolerated. Nevertheless, there is a consensus in the literature that the operational context of an ACA ultimately determines its effectiveness (see among others, Klitgaard, 1988; Olken, 2007; Quah, 2010; Svensson, 2005; Wei, 1999). The governments of Hong Kong and Singapore were committed and proactive in creating an environment that stimulated the battle against corruption. First of all, public sector employees’ wages rose significantly relative to wages of their private sector counterparts. Singapore required their civil servants to frequently rotate positions, which prevented them from building relationships with their clients. Officials who reported clients that attempted to bribe them were given financial bonuses. The governments of both countries ensured the agencies were staffed and funded adequately. The governments of Hong Kong and Singapore worked on cutting down cumbersome bureaucratic procedures and fees that previously acted as a breeding ground for corrupt behavior, such as licenses and import duties. Finally, and most importantly, the dedicated political leadership of both countries minimized its interference in corruption investigations lead by the agencies and granted impartial implementation of the legal framework (Quah, 2010; Svensson, 2005).
The main contributors to the failure of ACAs in countries that attempted implementing this policy are (1) the absence of appropriate operating context and (2) general institutional weakness of the newly setup ACAs that reflects in poor execution of designated tasks. In other words, in countries where the political will was not on the side of the ACAs, the policy did not breed desired outcomes. The cases of Botswana and Benin are typical examples in this regard1. Botswana’s Directorate on Corruption and Economic Crimes (DCEC) lacks the support of its government. DCEC is supposed to fully investigate corruption cases and pass on their findings to the prosecution. The number of cases that are followed through by the prosecution is extremely low, illustrating the government’s attitude towards fighting corruption. In addition, for the DCEC’s annual report to be published, it needs the president’s approval. This pressures the Directorate to investigate and/or report on cases that are in line with the president’s political preferences, making DCEC a tool
to discredit the opposition rather than tackle institutionalized corruption (Heilbrunn, 2004). In addition to facing issues similar in nature to those of its Botswanan counterpart, Benin’s Office for the Improvement of Morality in Public Life (CMVP in French) lacks adequate financial support form its government, which points towards Benin’s lack of commitment to its anti-corruption strategy (Heilbrunn, 2004).
Although the effectiveness of ACAs has been extensively discussed with reference to measures such as numbers of cases prosecuted, numbers of employees and the size of ACAs budget in relation to the country’s GDP, there has not been an empirical evaluation of this policy measure. This paper provides first empirical evidence on the efficacy of the policy by estimating a causal effect of establishing an ACA on corruption levels in the context of a transition economy of Serbia. Utilizing the micro data on private sector’s corruption perceptions collected as part of the World Bank’s Enterprise Surveys, I employ a Difference-in-Differences methodology to assess the policy’s effect in a quasi-experimental setting. I exploit the drawbacks in the implementation of this state-level policy and conduct an inter-regional comparison in which I compare firms in the capital region of Belgrade, where the effects were most present, with firms based in 4 other regions within Serbia. To account for potential spillovers, I conduct a second, inter-country, comparison where Serbian firms act as treatment group and FYR Macedonian firms, a comparable neighboring country whose ACA was established long before its Serbian counterpart, act as control group. Due to the data’s observational nature, I make use of the Propensity Score Matching (PSM) technique to address selection bias. I do not find evidence of reduced corruption as a result of the intervention in Serbia when considering both the inter-regional and the inter-country comparison. Based on a closer evaluation of the Serbian ACA and its operating environment, I hypothesize that the lack of support from other relevant state institutions explains the absence of an effect. Nonetheless, the reported results are not conclusive due to the susceptibility of my analysis to violations of identifying assumptions.
The paper is structured as follows. Section 2 provides context to my research by examining the Serbian anti-corruption policy. Section 3 describes the data. Section 4 outlines the methodology. Section 5.1 presents the results of the regional analysis and section 5.2 presents the results of the inter-country comparison. Section 6 concludes.
2 Background and Context: Anti-Corruption Agency of Serbia
(ACAS)
Corruption is a systematic problem of the Serbian society, present at all levels of the bureaucratic hierarchy and taking a variety of forms. According to the Transparency International's (2018) Corruption Perception Index, Serbia is consistently ranked as one of the most corrupt countries in Europe. The United Nations Office on Drugs and Crime (UNODC) reports that Serbian citizens rank corruption as the third most pressing issue2 the country faces, while Serbian entrepreneurs rank corruption as the fifth biggest obstacle3 to their operations (United Nations Office on Drugs and Crime, 2011, 2013). The independent authorities within Serbia also stress the prevalence of corruption in their respective fields. For example, in 2009, the independent Public Procurement Office noted serious irregularities in procurement projects whose total worth exceeded 100 million euros. Additionally, the Office reported that in one fourth of all procurement projects there were no public tenders, which are required by law, pointing at the preferential treatment of private firms that have ties with the political elite (Dedeić et al., 2018a).
The Serbian government designed a policy to address the issue. In the light of the preparations for the launch of the first comprehensive anti-corruption intervention, the president of the National Assembly said that ‘the omnipresent corruption must be acted upon, as it threatens development of democracy, human rights, the consolidation of the state and economic development’ (Radio-televizija Vojvodine, 2009). The Serbian Minister of Justice at the time emphasized that both the Serbian citizens and the international community demand and deserve government’s swift and effective response to widespread corruption (Dedeić et al., 2018a).
The establishment of the Anti-Corruption Agency of Serbia (ACAS) was the long-awaited response. The body was instituted and began its operations on January 1st, 2010 (Anti-Corruption Agency of Serbia, 2010). The ACAS’ establishment represented a formalization of the government’s fight against corruption as the agency succeeded the government’s formerly scattered and incomplete anti-corruption initiatives and became the central anti-corruption authority. The policy was modeled after Hong Kong’s Independent Commission Against Corruption (ICAC) following the investigation-prevention-education principle. This is reflected in the main functions of the
2 The first two being unemployment and low standard of living (United Nations Office on
Drugs and Crime, 2011, p. 7).
3 The first four being: currency fluctuations, complicated tax laws, frequent changes in laws
and regulation, and political instability (United Nations Office on Drugs and Crime, 2013, p. 9).
ACAS, of which the most noteworthy are (Anti-Corruption Agency of Serbia, n.d.; Heilbrunn, 2004):
Oversight of the utilization of public resources for which reason the ACAS investigates potential conflict of interest between public officials and the private sector, keeps records and tracks changes in the asset holdings of public officials, and oversees financing of political parties.
Taking investigative action upon the reports of corrupt behavior. Both corporate subjects and individuals can file reports documenting their experiences of corruption.
Establishing collaboration and strengthening ties with other relevant
independent institutions to eliminate incidences of corrupt behavior.
For example, the collaboration with the judiciary branch is essential as the justice system is responsible for prosecuting corrupt officials or entities based on lawsuits submitted by the ACAS.
Cooperating with the media and the civil sector to educate and spread
awareness about corruption. This is done by issuing updates regarding
ACAS’ effectiveness in reducing corruption through the implementation of the aforementioned functions, and organizing events to raise awareness, promote the importance of fighting corruption, and emphasize the role ACAS plays in the process.
The ACAS is composed of 2 bodies - the Board and the Director. The Board consists of 9 experts who together appoint the Director. The Board is appointed by the National Assembly, the institution the ACAS is accountable to. The National Assembly votes on the candidates proposed by: Administrative Committee of the National Assembly, President of the Republic, Government, Supreme Cassation Court, State Audit Institution, Protector of Citizens and Commissioner for Information of Public Importance, Socioeconomic Council, and associations of journalists in the Republic of Serbia (Anti-Corruption Agency of Serbia, 2010, p. 20). Once the two bodies of the ACAS are appointed, they independently organize departments that specialize in investigation of various forms of corruption thus together tackling the issue in a holistic manner.
The ACAS made notable progress in executing designated responsibilities as documented by the Agency’s annual reports. The Agency exercised its authority granted by the law to compile several important registers that are used in fulfilling the Agency’s functions. These registers are: the register of public officials’ asset holdings, the register documenting gifts received by public officials, the register noting all public functions held by individual civil servants, and the register of officials who also hold positions in private firms that cooperate with the government through public-private partnerships. The reports also document a manyfold increase in the number of cases processed in
various lines of operation, particularly focusing on the capital region. For example, the number of resolved cases of conflict of interests increased from 17 in 2010 to 958 in 2013, with a ratio of the number of cases resolved to the total number of cases received around 90 percent each year. The Agency pressed charges against a few ministers and other higher-up officials, sending the message of equality before the law. According to previous studies, these results do make the ACAS comparable to some of the most successful ACAs in the world (ACAs of Singapore and Hong Kong) (Heilbrunn, 2004; Quah, 2010). Additionally, the ACAS was committed to establishing better relations with relevant institutions, civil society organizations and the general public, which the reports document in the form of numbers of meetings, seminars, workshops, and events organized, with a brief discussion of their content and related accomplishments (Anti-Corruption Agency of Serbia, 2010, 2011, 2012, 2013).
Numerous challenges overshadowed the Agency’s achievements. The government, in cooperation with the National Assembly, consistently obstructed the Agency. A notorious example from 2010 helps illustrate this. The Ministry of Justice interfered by presenting their own interpretation of Article 82 of the Law on the Anti-Corruption Agency, which stipulates that a civil servant cannot perform 2 public functions simultaneously, such as holding positions as mayor and parliament representative. The Ministry’s interpretation implied that this article does not apply retroactively, meaning that the officials with double positions who started their terms before January 1st 2010, the date of the enactment of the law, would await the end of their terms after which they would be allowed to apply and obtain only one function. Without consulting the Agency, the amendment “clarifying” this imprecision found its way to the National Assembly representatives who voted in favor of the amendment. After learning about the incident, the Agency intervened by publicly asking the president not to sign the law into action, which the president disregarded. Consequently, the Agency appealed to the Constitutional Court asking for a verification of the constitutionality of the amendment that allowed more than 5000 public officials to hold double positions. With some hesitation, the Constitutional Court declared the amendment void a year later, which allowed the Law’s retroactive application. Meanwhile, intending to strip the Agency of its autonomy and independence, the National Assembly instituted a regulation that would allow the representatives to discharge particular board members and/or the Director of the Agency in case their performance is deemed unsatisfactory. The regulation stated that the Assembly representatives would judge the performance of the Agency based on the annual reports. In February of 2011, the National Assembly withheld this regulation after being pressured by the media and the general public (Anti-Corruption Agency of Serbia, 2010, 2011; Dedeić et al., 2018c). It is important to note that the government and the National Assembly often ignored the Agency’s recommendations regarding the refinement of the
Law that would enable the Agency to fulfill its tasks effectively and comprehensively. The authorities decided to be proactive in this specific case as the “clarifications” of the Law offered an opportunity to subdue the Agency.
The government used other channels to hinder the Agency’s anti-corruption efforts. Namely, the ACAS representatives repeatedly voiced their concerns regarding the ability of the Agency to employ competent experts and function optimally, hinting at the deficient financial resources. While promising the general public the issue would be resolved, the government used this opportunity to subject the Agency to their interest. In September 2010, the Agency’s director revealed that they received the official budgeting guidelines for 2011 exactly 7 days before the deadline for the submission of the budget proposal. Additionally, the official letter containing the information was addressed to the Committee for Resolving Conflicts of Interest, an institution that ceased to exist with the establishment of the Agency (Dedeić et al., 2018c). This example shows that the government’s ultimate goal was not just to obstruct the Agency, but also to discredit the institution as a whole.
The ACAS did not find an ally in other independent institutions. In January 2012, the Director of the ACAS came out with information that out of 17 lawsuits filed against some of the most well-known members of the Serbian political elite, only 1 was settled – the court issued a formal warning. The Director pointed out that the stalling of these court cases and inadequate punishments would only deepen the distrust in the public administration and send a message to the elite that corrupt behavior would still be tolerated (Dedeić et al., 2018c). The judiciary followed the same practice in the subsequent period. Between 2013 and 20164, in 45 percent of settled lawsuits the Agency filed against public officials, the consequences were either a formal warning issued by the court (20 percent of the time) or a fine that is in its value below the minimum that the law stipulates (25 percent of the time). Over the same period, approximately a third of concluded judicial proceedings were suspended due to obsolescence (Anti-Corruption Agency of Serbia, 2017). These figures show that the collaboration between the Agency and the judiciary in practice was not functional. The justice system could not process cases against politicians adequately due to the interference of officials at higher levels of the bureaucratic hierarchy.
The complete demise of the ACAS followed the May 2012 general and presidential elections after which the Serbian Progressive Party held the absolute majority in the Serbian parliament5. The newly elected president exercised his power to propose a new Board member that was affiliated with the party in power. Slowly but surely, in the period between 2012 and 2018,
4 There is no data available for the period before 2013.
5 A coalition government lead by the Democratic Party marked the period between the
the structure of the Board and the Director changed and currently consists of individuals deemed suitable for those positions by the ruling party (Dedeić et al., 2018b). As a result, the Agency lost its purpose, as it came under complete control of the executive branch, thus limiting its ability to fundamentally reduce corruption.
In summary, the efficacy of the ACAS has been compromised by the absence of genuine political will to end the institutionalized and widespread corruption. The challenges the ACAS faced significantly limited its capability to respond to incidences of corruption as the Agency had to, in addition to that, work on securing its autonomy and independence.
3 Data
3.1 World Bank’s Enterprise Surveys
The analysis makes use of the World Bank’s Enterprise Surveys, a repeated cross-sectional survey containing firm-level data from 139 countries. More specifically, I utilize the data on Serbian and FYR Macedonian firms collected in the 2009 and 2013 waves. These are the only years for which the data is available as part of this survey. The survey provides a comprehensive snapshot of the state of affairs in the private sector of various countries by offering a wide range of firm-level indicators, among which are those capturing corruption experiences.
To analyze potential changes in corruption levels as a result of the intervention in question, the analysis makes use of subjective corruption measures6. Namely, these variables indicate the firms’ perceptions of corruption, capturing both the frequency of having to engage with public administration in a corrupt manner, and their assessment of the negative influence of corruption on daily operations. More concretely, the firms answered the following questions and indicated their response using a 5-degree scale:
“To what extent is corruption an obstacle to your daily operations?”
“How common is it for a firm in your line of business to make informal payments/gifts to ‘get things done’ regarding licenses, regulations, and general governmental services?”
“How common is it for a firm in your line of business to make informal payments/gifts to deal with customs/imports regulation?” “How common is it for a firm in your line of business to make
informal payments/gifts to deal with the judiciary system/courts?”
“How common is it for a firm in your line of business to make informal payments/gifts to deal with tax administration?”
Additionally, I generate a composite index combining the four aforementioned corruption frequency measures to provide a summative indicator against which the effectiveness of the intervention can be measured. This can be done as these 4 variables are measured on the same scale and follow the same direction (i.e. the variables record the frequency of engaging in corrupt activities in the same manner – from not frequent to very frequent). Following the theoretical considerations on creating such an indicator outlined in the OECD's (2008) Handbook on Constructing Composite Indices, I generate a standardized index that gives equal weights to the 4 dimensions that comprise it. Noting this mathematically yields the following expression:
𝐼𝑖 = 1
𝐽∑
𝑥𝑖𝑗 − 𝑥̅𝑗
𝜎𝑗 (4)
where 𝐼𝑖 is the value of the index for firm 𝑖, 𝐽 is the total number of variables used to construct the index, 𝑥𝑖𝑗 is the firm 𝑖’s value on variable 𝑗, 𝑥̅𝑗 is the baseline mean in the control group on variable 𝑗 , and 𝜎𝑗 is the baseline variance in the control group on variable 𝑗. By construction, the mean of the index in the control group is centered at zero.
3.2 Comparisons and Descriptive Statistics
The choice of the adequate counterfactual is crucial in estimating the true causal effect of an intervention in question, particularly in quasi-experimental studies. In order to provide a thorough examination of the policy, exploiting the Word Bank’s Enterprise Surveys introduced above, this paper considers two comparisons: inter-regional and inter-country.
For the first comparison, I exploit the information on the implementation of the Law on the Anti-Corruption Agency. Although a state-level policy, the effects of the establishment of the ACAS were mostly felt in the Serbia’s capital region of Belgrade. The annual reports state that most cases the Agency investigated concurrently with investing significant efforts to eliminate the obstacles discussed in Section 2 were related to public officials based in the capital (Anti-Corruption Agency of Serbia, 2010, 2011, 2012, 2013). While the figures indicating the fraction of Belgrade-based public officials the Agency interacted with are not available, auxiliary documents of the annual reports list names and functions of officials the Agency issued warnings to or prosecuted. These point to the fact that the vast majority of those officials perform functions in the parliament, ministries, courts, or public enterprises that are all based in the capital. The lack of the Agency’s presence in other regions of Serbia is also demonstrated by the fact that since its
establishment, the Agency, although entitled to by the law, has not opened any branches outside its headquarters in Belgrade. Operating regional branches would help the Agency’s experts have better understanding of the state of affairs and corruption locally, and allow for a swifter and more informed response to corruption. It is due to these setbacks in the policy’s countrywide implementations that I perform a comparison between firms in the Serbia’s capital, Belgrade, making them the treatment group, and firms in 4 other regions specified in the dataset, making them the control group. I work with binary treatment variable since the dataset does not contain indicators such as the geographical coordinates of the firms, which would allow me to calculate their distance to Belgrade and work with continuous treatment, thus providing greater variation in the analysis and potentially more intuitive interpretation of the results.
To account for the spillovers arising from the fact that the Agency is ultimately a state-level anti-corruption entity, I conduct a second comparison. This required choosing another country as the comparison group. Naturally, the focus was placed on finding a counterfactual in the region of Western Balkans, since, in addition to having joint history as federative republics of former Yugoslavia, these countries still experience similar socioeconomic conditions7, language, and culture, which further improves the comparability of the treatment and comparison groups.
The choice was dependent on two conditions. First, the potential counterfactual’s firm-level data needed to be available and compatible with the firm-level data from Serbia. This means that the data collection instruments and the timing of the surveys needed to be the same across countries to allow for a credible comparison. Taking into account that the data on Serbia comes from World Bank’s Enterprise Surveys that employ uniform data collection methods across all countries the survey considers, finding a counterfactual satisfying this condition was not problematic. Second, the potential counterfactual needed to either have a similar anti-corruption institutional framework instituted before the year for which baseline data on Serbia exist or miss such institutional framework altogether in order to satisfy the comparison group requirements of the Difference-in-Differences method I discuss in the following section. The countries of the Western Balkans that satisfied both of these conditions were Croatia and the Former Yugoslav Republic of Macedonia, with the choice falling on the FYR Macedonia due to a greater degree of comparability. Namely, the difference in the baseline corruption levels as reported by the Serbian and FYR Macedonian firms was considerably smaller than in the Serbia-Croatia comparison. In addition, the Serbian and FYR Macedonian legal frameworks regarding the fight against corruption are perfectly comparable, as FYR Macedonia’s Commission for
7 The countries of the Western Balkans region comprise a single market under the Central
Preventing Corruption perfectly matches its Serbian counterpart with respect to its designated tasks and modus operandi (The Government of FYR Macedonia, 2002).
Tables 1 and 2 provide descriptive statistics for baseline characteristics and outcomes in the inter-regional comparison respectively. Columns (3) and (6) in both tables provide the differences in ‘raw’ (before Propensity Score Matching) and differences in weighted (after Propensity Score Matching) means at baseline respectively. Tables 3 and 4 present the same information for the inter-country comparison. The tables show a decrease in the number of observations after Propensity Score Matching is performed - this is due to sample trimming on the basis of common support. The number of observations after Propensity Score Matching can be interpreted as the sum of weights. I discuss the PSM method in detail in the following section.
As expected, the heterogeneity amongst firms is more pronounced in the inter-country comparison, particularly regarding the outcome variables. In both comparisons, treatment territories report higher levels of corruption compared to the control territories. On average, firms in treatment territories report that corruption is a moderate obstacle to operations, while frequency of informal gifts/payments to public officials is characterized as ‘sometimes’ (just below ‘frequent’).
Table 1 – Inter-regional comparison descriptive statistics: Differences in non-weighted and weighted means on selected background characteristics at baseline
(1) (2) (3) (4) (5) (6)
Treatment Control Difference (1)-(2) Common Support Trimmed Treatment Weighted Control Difference (4)-(5) Inflation-adjusted total annual salesa 18.999 [2.092] 18.054 [1.845] 0.945*** (0.209) 18.255 [1.842] 18.196 [1.512] 0.059 (0.251) Firm in manufacturing sector (=1) 0.279 [0.450] 0.415 [0.494] 0.135*** (0.049) 0.276 [0.450] 0.321 [0.469] -0.045 (0.067) Firm officially registered
when established (=1) 0.967 [0.179] 0.930 [0.256] 0.037 (0.024) 0.947 [0.225] 0.921 [0.270] 0.026 (0.036) Current no. of employeesa 3.854
[1.699] 3.445 [1.406] 0.410** (0.159) 3.328 [1.648] 3.290 [1.285] 0.038 (0.220) Starting no. of employeesa 2.139 [1.989] 2.111 [1.551] 0.028 (0.197) 1.937 [2.072] 1.929 [1.392] 0.007 (0.269) Overdraft facility (=1) 0.601 [0.491] 0.621 [0.486] -0.020 (0.052) 0.592 [0.495] 0.608 [0.490] -0.016 (0.073) Firm purchased fixed
asset in last yr (=1) 0.760 [0.429] 0.631 [0.484] 0.129*** (0.048) 0.711 [0.457] 0.666 [0.473] 0.045 (0.068) Manager’s yrs of experience 18.276 [10.936] 16.374 [9.505] 1.902* (1.059) 17.303 [9.936] 17.705 [9.467] -0.403 (1.448) Security expenditures in last yr (=1) 0.526 [0.501] 0.385 [0.488] 0.141*** (0.051) 0.421 [0.497] 0.500 [0.502] -0.079 (0.074) International certificates (=1) 0.300 [0.460] 0.286 [0.453] 0.014 (0.049) 0.197 [0.401] 0.225 [0.419] -0.027 (0.061) Females among owners
(=1) 0.420 [0.495] 0.364 [0.482] 0.056 (0.051) 0.382 [0.489] 0.342 [0.476] 0.040 (0.071) Obstacle to operations: Access to finance 1.669 [1.325] 1.732 [1.288] -0.063 (0.136) 1.697 [1.265] 1.822 [1.240] -0.125 (0.187) Crime 1.242 [1.298] 0.921 [1.155] 0.321** (0.127) 1.145 [1.262] 1.192 [1.197] -0.047 (0.182) Tax rates 1.856 [1.320] 1.598 [1.264] 0.258* (0.134) 1.816 [1.293] 1.817 [1.214] -0.001 (0.187) Education of labor force 1.435
[1.318] 0.983 [1.185] 0.452*** (0.129) 1.289 [1.422] 1.052 [1.247] 0.238 (0.203) Observations 154 234 76 145
Notes: The table provides the differences in non-weighted and weighted means between the capital (treatment) and other
(control) regions within Serbia on selected background characteristics at baseline. Standard deviations are shown in parentheses in columns (1), (2), (4), and (5). Standard errors are shown in parentheses in columns (3) and (6). The statistical significance of the differences in means at 10%, 5% and 1% significance levels is denoted by *, ** and *** respectively.
Table 2 – Inter-regional comparison descriptive statistics: Differences in non-weighted and weighted means on outcome variables at baseline
(1) (2) (3) (4) (5) (6)
Treatment Control Difference (1)-(2) Common Support Trimmed Treatment Weighted Control Difference (4)-(5) Corruption – obstacle to operations 1.920 [1.417] 1.566 [1.426] 0.355** (0.150) 1.880 [1.461] 1.621 [1.431] 0.259 (0.216) Bribes to “get things
done” 2.561 [1.374] 2.300 [1.336] 0.261* (0.154) 2.548 [1.456] 2.461 [1.396] 0.087 (0.230) Bribe to deal with
import-related issues 1.800 [1.306] 1.662 [1.038] 0.138 (0.134) 1.915 [1.418] 1.777 [1.159] 0.138 (0.225) Bribe to deal with the
judiciary system 2.043 [1.341] 1.984 [1.197] 0.059 (0.148) 2.153 [1.495] 2.103 [1.223] 0.050 (0.237) Bribe to deal with tax
administration 1.936 [1.287] 1.968 [1.265] -0.032 (0.153) 2.054 [1.407] 2.005 [1.257] 0.048 (0.227) Corruption index 0.059 [0.906] 0.000 [0.820] 0.059 (0.105) 0.121 [1.004] 0.118 [0.830] 0.003 (0.163) Observations 150 226 75 141
Notes: The table provides the differences in non-weighted and weighted means between the capital (treatment) and other
(control) regions within Serbia on outcome variables at baseline. Standard deviations are shown in parentheses in columns (1), (2), (4), and (5). Standard errors are shown in parentheses in columns (3) and (6). The statistical significance of the differences in means at 10%, 5% and 1% significance levels is denoted by *, ** and *** respectively.
Table 3 – Inter-country comparison descriptive statistics: Differences in non-weighted and weighted means on selected background characteristics at baseline
(1) (2) (3) (4) (5) (6)
Treatment Control Difference (1)-(2) Common Support Trimmed Treatment Weighted Control Difference (4)-(5) Inflation-adjusted total
annual salesa,b
14.228 [2.002] 13.734 [1.916] 0.494*** (0.152) 13.632 [1.715] 13.749 [1.783] -0.117 (0.326) Firm in manufacturing sector (=1) 0.361 [0.481] 0.306 [0.461] 0.055 (0.034) 0.214 [0.413] 0.316 [0.468] -0.102 (0.080) Firm officially registered
when established (=1) 0.945 [0.229] 0.972 [0.164] -0.028* (0.015) 0.957 [0.204] 0.987 [0.115] -0.030 (0.028) Current no. of employeesb 3.607
[1.540] 3.379 [1.462] 0.228** (0.110) 3.062 [1.538] 3.330 [1.282] -0.268 (0.257) Starting no. of employeesb 2.122
[1.726] 1.942 [1.566] 0.179 (0.130) 1.972 [1.683] 2.002 [1.540] -0.030 (0.290) National sales (% of total
sales) 89.781 [19.763] 77.454 [35.894] 12.327*** (2.095) 86.100 [24.568] 79.371 [31.930] 6.729 (5.216) Main product (% of total
sales) 71.353 [25.584] 83.050 [22.126] -11.698*** (1.764) 80.043 [19.129] 79.918 [23.358] 0.124 (3.775) Overdraft facility (=1) 0.613 [0.488] 0.279 [0.449] 0.334*** (0.035) 0.429 [0.498] 0.558 [0.500] -0.130 (0.088) Outstanding loan (=1) 0.698 [0.460] 0.599 [0.491] 0.099*** (0.035) 0.571 [0.498] 0.631 [0.486] -0.060 (0.087) Firm purchased fixed
asset in last yr (=1) 0.682 [0.466] 0.712 [0.453] -0.030 (0.034) 0.614 [0.490] 0.713 [0.456] -0.098 (0.085) Manager’s yrs of experience 17.137 [10.132] 16.361 [9.068] 0.776 (0.707) 18.400 [11.323] 15.758 [7.247] 2.642 (1.639) Security expenditures in last yr (=1) 0.441 [0.497] 0.779 [0.416] -0.338*** (0.033) 0.529 [0.503] 0.700 [0.461] -0.171** (0.084) International certificates (=1) 0.291 [0.455] 0.293 [0.456] -0.002 (0.035) 0.157 [0.367] 0.208 [0.409] -0.051 (0.070) Females among owners
(=1) 0.386 [0.487] 0.400 [0.491] -0.014 (0.036) 0.371 [0.487] 0.426 [0.498] -0.054 (0.088)
Firm’s legal statusc:
Shareholding – traded shares (=1) 0.218 [0.413] 0.084 [0.279] 0.133*** (0.028) 0.100 [0.302] 0.057 [0.233] 0.043 (0.050) Shareholding – non-traded shares (=1) 0.422 [0.495] 0.828 [0.378] -0.405*** (0.035) 0.700 [0.462] 0.750 [0.436] -0.050 (0.083) Partnership (=1) 0.044 [0.205] 0.017 [0.129] 0.027** (0.014) 0.029 [0.168] 0.036 [0.188] -0.007 (0.032) Obstacle to operations: Access to finance 1.707 [1.301] 1.583 [1.308] 0.124 (0.096) 1.814 [1.333] 1.399 [1.178] 0.416* (0.213) Crime 1.050 [1.223] 1.077 [1.300] -0.027 (0.092) 1.143 [1.171] 1.265 [1.241] -0.122 (0.215) Tax rates 1.700 [1.291] 1.542 [1.243] 0.158* (0.093) 1.571 [1.199] 1.629 [1.288] -0.058 (0.220) Education of labor force 1.163
[1.258] 0.904 [1.186] 0.258*** (0.089) 1.043 [1.197] 1.044 [1.184] -0.001 [0.210] Observations 388 366 70 73
Notes: The table provides the differences in non-weighted and weighted means between Serbia (treatment) and FYR
Macedonia (control) on selected background characteristics at baseline. Standard deviations are shown in parentheses in columns (1), (2), (4), and (5). Standard errors are shown in parentheses in columns (3) and (6). The statistical significance of the differences in means at 10%, 5% and 1% significance levels is denoted by *, ** and *** respectively. a Sales were
Table 4 – Inter-country comparison descriptive statistics: Differences in non-weighted and weighted means on outcome variables at baseline
(1) (2) (3) (4) (5) (6)
Treatment Control Difference (1)-(2) Common Support Trimmed Treatment Weighted Control Difference (4)-(5) Corruption – obstacle to operations 1.707 [1.431] 1.371 [1.373] 0.337*** (0.104) 1.681 [1.510] 1.788 [1.411] -0.107 (0.258) Bribes to “get things
done” 2.397 [1.354] 1.912 [1.124] 0.485*** (0.096) 2.475 [1.337] 1.889 [1.033] 0.587*** (0.215) Bribe to deal with
import-related issues 1.712 [1.144] 1.421 [0.878] 0.292*** (0.081) 1.741 [1.163] 1.484 [0.923] 0.257 (0.199) Bribe to deal with the
judiciary system 2.007 [1.252] 1.688 [1.117] 0.319*** (0.095) 2.140 [1.420] 1.762 [1.307] 0.379 (0.267) Bribe to deal with tax
administration 1.956 [1.271] 1.306 [0.721] 0.650*** (0.081) 1.927 [1.289] 1.499 [0.982] 0.429* (0.227) Corruption index 0.477 [1.107] 0.000 [0.824] 0.477*** (0.081) 0.533 [1.173] 0.056 [0.985] 0.477** (0.214) Observations 150 226 69 72
Notes: The table provides the differences in non-weighted and weighted means between Serbia (treatment) and FYR
Macedonia (control) on outcome variables at baseline. Standard deviations are shown in parentheses in columns (1), (2), (4), and (5). Standard errors are shown in parentheses in columns (3) and (6). The statistical significance of the differences in means at 10%, 5% and 1% significance levels is denoted by *, ** and *** respectively.
3.3 Missing Data
The datasets used for the analysis of both comparisons suffer from severe data missingness. Tables A1 through A4 available in the appendix report the fraction of missing values on each of the variables used in this analysis across groups, time and comparisons. This paper employs Complete Case (CC) analysis throughout, meaning that observations that miss values on any of the variables used in the analysis are disregarded. This approach replies on the assumption that there are no systematic differences between the complete and the incomplete records, or in other words, that the missingness is random and that the CC sample is still representative at the population level, thus making the inference of the results valid (Bartlett, 2012). This assumption is typically referred to as the Missing Completely At Random (MCAR) assumption.
Since I am not able to fully investigate whether this assumption holds, I attempted to partly correct for the potential bias that arises from the violation of MCAR following Wooldridge (2002). In the case of right-hand-side variables, I generated binary variables indicating whether the observation is not missing a value on each of the independent variables respectively. I then interacted these with the independent variables they refer to and included the binary variables and the aforementioned interactions in the analysis whenever covariates were considered. This approach would enable me to isolate the true potential of the covariates to correct for the differences between the groups, from the non-response bias induced by the decrease in the number of observations in the regressions that consider covariates due to missing values. As a result, I would ensure that the analysis effectively considered the same sample regardless of whether or not the covariates were included. However, taking into account that the vast majority of the control variables are dummy variables and hence do not take on many different values, I was not able to implement this non-response bias correction method due to collinearity. The dummy variables and the interactions generated were highly correlated with the covariates themselves, thus preventing the possibility to address the potential violation of the MCAR assumption with right-hand-side variables.
Tables A1 through A4 in the appendix indicate that the issue of missing data is most notable in the outcome variables. I recognize that corruption-related questions can be viewed as sensitive, which prompted non-response. To understand if non-response on outcome variables is independent of firm characteristics and thus could be seen “as if” random conditional on those characteristics, I regress a dummy variable indicating missingness on the corruption index I generated earlier on firm characteristics. I perform this check in both comparisons. I find that non-response is independent of all firm characteristics except for access to finance, in which case firms that reported access to finance being a greater obstacle to their operations were on average more likely to not respond. This result holds for both comparisons. Since
non-response is largely independent of firm characteristics I consider, I assume that missing values on corruption indicators are MCAR.
It is understandable that firm representatives interviewed as part of the survey were not confortable or allowed to disclose sensitive or confidential information about their firms, even when full anonymity was ensured. This is reflected in the substantial amount of missing values, particularly in variables expressed in monetary terms. In addition to the potential violations of the MCAR assumption discussed above, missing data impacted the analysis presented in this paper in two ways.
First, the analysis does not assess the effect of the intervention on objective measures of corruption, such as the estimated percentage of total annual sales paid in informal payments/gifts to civil servants. Exploiting such a variable would enable us to quantify the magnitude of corruption. While focusing on subjective corruption indicators ignores the corruption’s quantitative dimension, they still represent valuable tools in assessing the effectiveness of the policy.
Second, the missing values limit the array of variables available as covariates. As will be discussed in detail in the methodology section, the success with which the Propensity Score Matching technique neutralizes the effect of observable confounders is highly dependent on the quality of variables that enter the propensity score function. Exploiting variables impaired with data missingness would result in a substantial decrease in the number of observations, which leads to issues discussed earlier in this subsection. On the other hand, disregarding some of these variables limits the ability of the PSM technique to address selection bias, as potentially crucial confounders are left out. For example, a variable such as the net book value of fixed assets would have been a valuable dimension to be considered in PSM as it acts as an indicator of firms’ size and value, thus allowing matching like-with-like in that respect. In both comparisons, the fraction of missing values on that variable is around 68 percent. Alternatives had to be considered, such as the inflation-adjusted sales figures, to allow the PSM model to include this important dimension firms differ on and improve comparability across groups. Since data limitations dictated the choice of variables that entered the PSM model, it is likely that perhaps important confounders were not used.
4 Methodological Considerations
This section outlines the methodological approaches employed by this paper. As the establishment of the ACAS created a quasi-experimental setting, I utilize the Difference-in-Differences (DD) estimator to gauge the causal effect of the intervention. Propensity Score Matching (PSM) is used to improve the comparability of the control and treatment groups. The application of both
methods is described and motivated, with special reference to their identifying assumptions whose violations are discussed in detail in the results section.
The primary econometric technique applied is the Difference-in-Differences (DD). This method merges the two counterfeit counterfactuals, (a) the before-and-after intervention comparison and (b) the enrolled-versus-not-enrolled comparison, to estimate an intervention’s causal effect. Focusing only on the before-and-after comparison that controls for the time-invariant factors influencing the outcome of interest cannot provide an intervention’s causal effect as many other time-varying factors can affect the outcome concurrently with the intervention. Similarly, focusing only on the enrolled-versus-not-enrolled comparison is also problematic, as there could be (unobservable) factors influencing treatment status. The DD estimator addresses these issues by subtracting the effects of time-varying factors influencing the outcome from the before-and-after difference in the treated group. The DD estimator captures the effect of time-varying factors by performing a before-and-after comparison in the untreated group exposed to similar environmental conditions (Gertler, Martinez, Premand, Rawlings, & Vermeersch, 2011, p. 96).
The DD estimator can be written in form of an OLS regression. In the context of the issue this paper studies, we obtain the following population model:
𝐶𝑖𝑡 = 𝛼 + 𝛽𝑌𝑒𝑎𝑟𝑡+ 𝛾𝑇𝑟𝑒𝑎𝑡𝑖 + 𝛿(𝑌𝑒𝑎𝑟𝑡× 𝑇𝑟𝑒𝑎𝑡𝑖) + 𝜃𝑋𝑖𝑡+ 𝜖𝑖𝑡 (1) where 𝐶𝑖𝑡 captures the firm 𝑖 ’s corruption perception at time 𝑡 , 𝑌𝑒𝑎𝑟𝑡 is a dummy variable taking the value of 1 in the post-treatment period, 𝑇𝑟𝑒𝑎𝑡𝑖 is a dummy variable taking the value of 1 for firms located in the treated territory (the capital region of Belgrade in the inter-region comparison and the Republic of Serbia in the inter-country comparison), 𝑋𝑖𝑡 is a vector of the firms’ background characteristics, and 𝜖𝑖𝑡 is a disturbance term. The coefficient on the interaction, 𝛿 , gives the causal effect8 of the ACAS’ establishment on corruption as perceived by the private sector firms.
In order to obtain an unbiased estimate of 𝛿, the common trend assumption needs to hold. This means that the control group must be a true counterfactual – it needs to be an accurate representation of what the trend of the outcome would have looked like in the treatment group had there been no intervention. It is necessary to assume this to be able to say that the difference in trends that we observe between the groups after the intervention is truly due to that intervention. While this cannot be proven, showing that the trends of the outcome of interest in treatment and control groups move in tandem in the periods before the intervention can provide sufficient evidence to believe that
8 DD estimator by construction gives the Average Treatment Effect on the Treated (ATET).
the common trend assumption holds. The common reason behind the absence of the common trend is selection bias. This happens when the treated and the untreated entities differ in dimensions other than treatment status. The imbalance between groups on these other dimensions is worrisome if they can influence the outcome. In that case, the DD estimator yields invalid or biased estimates, as it erroneously attributes these differences to the intervention (Gertler et al., 2011; Imbens & Wooldridge, 2009). Conditioning the analysis on the variables on which imbalances exist could be beneficial, as parallel trend could be conditional on certain covariates, but does not tackle the issue comprehensively.
A more potent approach is to make use of matching techniques to create a counterfactual that is as similar as possible to the treatment group based on observable pre-intervention characteristics. This paper makes use of Propensity Score Matching (PSM) technique that matches observations on the basis of the probability of being treated given a vector of background characteristics, also called a propensity score. The propensity score is a single value ranging from 0 to 1 that encapsulates the observed characteristics of the entities as they affect the treatment status. The identifying assumption of the method stipulates that the characteristics on the basis of which the observations are matched cannot simultaneously be outcomes – the intervention should not influence these characteristics (Gertler et al., 2011). Empirically, the propensity scores can be obtained using a probit regression. In the context of this paper, the regression takes the following form:
Pr(𝑇𝑟𝑒𝑎𝑡𝑖 = 1) = 𝑓(𝑋𝑖) (2)
where 𝑋𝑖 is a vector of background characteristics of the firms in treated and untreated territories. Tables 1 and 3 in the data section give a list of the variables considered. After the propensity scores are obtained, it is important to restrict the analysis to observations in treatment and control groups that are as similar to each other as possible based on baseline characteristics. This is done by plotting the baseline distribution of propensity scores in treatment and control groups and restricting the analysis to the overlapping area of the two distributions. Once the sample is reduced so that it consists of firms that are similar on the confounders considered, inverse probability weighting algorithm is used to finally match the observations and estimate the population model given by the equation (1).
The main advantage of this technique, as illustrated by Rosenbaum & Rubin (1983), is that it substantially reduces selection bias. That is, the outcomes of interest become independent of enrollment status conditional on the propensity scores. PSM mimics Randomized Controlled Trials (RCTs) by making assignment into treatment as good as random conditional on the propensity scores. However, one must be cautious about interpreting the results of the analysis, as this technique does not eliminate selection bias in its
entirety. PSM can only correct for the selection bias that arises from the non-random distribution of observable confounders across the treatment and control groups, as only those were considered in the process of generating propensity scores. The non-random distribution of unobservable confounders may still bias the results. In a panel data setting, this issue could be partly corrected for by performing a fixed effects analysis in which case the time-invariant unobservable confounders would be corrected for. Since the analysis is based on cross-sectional data as indicated in the previous section, the selection bias arising from unobservable confounders persists. I discuss this issue in the results section.
Even in situations when parallel trends can be assumed, or at least their absence corrected for through PSM, the DD estimator can still yield invalid or biased results. The DD assigns the intervention any difference in trends between the treatment and control groups that takes place from the moment the intervention is implemented (Gertler et al., 2011, p. 104). Therefore, the DD will falsely estimate the intervention’s effect if, concurrently with the intervention, other factors influence the trend of the outcome of interest. One must be cautious of the existence of such factors and account for them in the analysis. Based on the extensive evaluation of the context of this policy I conducted as part of this research, I identified no such factors.
To check the robustness of my estimates, I perform a ‘sub-block’ analysis proposed by Imbens (2015). After the matching is performed, the regression (1) is run in four sub-blocks containing similar numbers of observations of the area of common support with and without covariates. The covariates considered are those used in the propensity score matching exercise. Based on the estimates in each sub-block, the estimate for the entire area of common support is calculated. This is a sum of the weighted estimates in each sub-block, where the weight is the ratio of the sum of the number of observations in the treatment and control groups in that specific sub-block and the total number of observations in the four sub-blocks. The same procedure is performed on all outcomes the paper considers. Noting this mathematically:
𝜏̂ = ∑𝑁𝑗 𝐶 + 𝑁
𝑗𝑇
𝑁 𝜏̂𝑗 (3)
where 𝜏̂𝑗is the estimate of the intervention’s effect in the 𝑗th sub-block of the area of common support, 𝑁𝑗𝐶 and 𝑁𝑗𝑇 is the number of observations in control and treatment groups in the 𝑗th sub-block respectively, 𝑁 is the total number of observations in the entire area of common support and 𝜏̂ represents the calculation of the estimate of the intervention’s effect following the aforementioned procedure.
This form of robustness check showcases the sensitivity of the estimate of 𝛿 in equation (1) to the presence of additional covariates in the analysis. As
argued by Imbens (2015), observing whether the estimate fluctuates substantially when stress-tested in the aforesaid fashion helps assess the plausibility of unconfoundedness, an untestable yet crucial assumption, particularly in impact evaluation studies utilizing observational data.
5 Results
5.1 Inter-regional Analysis
I first present the results of the comparison in which firms based in the capital region are used as the treatment group and firms based in other regions of Serbia comprise the control group. As highlighted earlier, there have been noticeable regional disparities in the implementation of the policy that are exploited to conduct this comparison.
Table 5 below presents the results of the DD estimator capturing the impact of the policy on 5 variables measuring corruption perception and the composite index. Regressions in columns (1) through (6) do not consider covariates, while these are included in regressions in columns (7) through (12). The covariates used are those presented in Table 1. The decline in the number of observations we see is due to the Complete Case (CC) analysis – the observations with missing values are not considered in the analysis. As explained earlier, the DD method provides an estimate of the causal effect of the policy by combining two counterfeit counterfactuals thus accounting for the before-and-after and enrolled-versus-non-enrolled comparisons. The estimated causal effect is given by the estimated coefficient on the interaction embodied by the Region×Year explanatory variable in Table 5. The results seem to point towards no impact of the policy.
There are caveats hampering the causal inference of the estimates in regressions (1) through (6). Taking into account that there is only one cross-section prior to the intervention, it is impossible to plot the trend of the outcome variables across groups. This means the parallel trends assumption cannot be visually inspected. Thus, it is not clear whether the chosen counterfactual (non-capital regions of Serbia) accurately pictures the trend of the outcomes of interest in the absence of the intervention, consequently leading to potentially invalid estimates. As illustrated by Table 1, there are substantial differences between groups at baseline, which raises further concerns that the estimates are impaired by selection bias. Expanding the analysis by considering additional covariates, showcased in regressions (7) through (12), does not seem to be fruitful. Although covariates help control for the differences between the groups, since parallel trend assumption cannot be evaluated, it cannot be concluded that the comparability of the groups indeed improved by making the trends parallel conditional on these covariates. Therefore, the DD estimator may still be biased or invalid.
Table 5 – Testing the effect of the policy on perceived corruption: Non-weighted regressions using the entire sample – Inter-regional comparison – (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) Corruption – obstacle to operations Bribes to “get things done” Bribe to deal with import-related issues Bribe to deal with the judiciary Bribe to deal with tax administration Corruption Index Corruption – obstacle to operations Bribes to “get things done” Bribe to deal with import-related issues Bribe to deal with the judiciary Bribe to deal with tax administration Corruption Index Region 0.354** (0.150) 0.261* (0.155) 0.138 (0.142) 0.059 (0.152) -0.032 (0.154) 0.059 (0.108) 0.227 (0.185) -0.056 (0.196) -0.071 (0.188) -0.075 (0.198) -0.097 (0.184) -0.123 (0.135) Year -0.482*** (0.125) 0.026 (0.122) 0.197* (0.106) -0.108 (0.115) -0.128 (0.116) -0.005 (0.082) -0.412** (0.166) -0.063 (0.178) 0.224 (0.159) -0.032 (0.162) -0.039 (0.155) 0.017 (0.117) Region×Year -0.181 (0.211) -0.288 (0.215) -0.247 (0.203) -0.192 (0.216) -0.068 (0.206) -0.132 (0.159) -0.092 (0.238) -0.084 (0.252) -0.223 (0.239) -0.161 (0.252) -0.121 (0.234) -0.108 (0.180) Additional
Covariates No No No No No No Yes Yes Yes Yes Yes Yes
R2 0.0519 0.0063 0.0055 0.0065 0.0048 0.0021 0.1981 0.1044 0.1411 0.1373 0.1374 0.1783
Observations 723 629 591 582 572 530 534 467 440 434 428 400
Notes: The table shows the effect of the policy on a number of variables capturing private sector’s perceptions of corruption. All regressions follow the Diff-in-Diff specification. The estimated
coefficient on the interaction term provides the effect of the policy. Regressions (1) to (6) do not consider covariates. Covariates in regressions (7) to (12) are background characteristics the firms in the treatment and the control groups exhibit differences on, as shown in Table 1. Robust standard errors are shown in parentheses. The statistical significance of the estimated coefficients at 10%, 5% and 1% significance levels is denoted by *, ** and *** respectively.
In my attempt to improve the causal inference of the results, I perform Propensity Score Matching. The implementation of the PSM slightly differs in this case compared to how it is commonly applied in a panel setting. In a panel, the propensity scores obtained by running the probit regression on baseline data are given to a specific entity. The PS that entity receives is constant over time. The trimming on common support is done, as explained in the methodology section, based on PS distributions across groups at baseline. In the cross-sectional setting, since the same firms do not appear in both time periods, the probit regression is run over all observations (both pre- and post-intervention), so that every observation receives their unique propensity score (i.e. it is not constant over time). Nevertheless, the trimming on common support is done in the exact same manner. I perform a Likelihood-Ratio Chow test to see whether the probit model generates the probability of treatment differently across time by comparing the differences in coefficients on the vector of background characteristics the model considers across time. The test reported non-significant differences.
PSM mimics random treatment assignment and corrects for selection bias by controlling for the non-random distribution of observable confounders across groups at baseline. As outlined earlier, the concern regarding the implementation of the method in this analysis is that, due to missing values, perhaps critical confounders were left out. Nevertheless, this issue was not pronounced in the inter-regional comparison, as I was able to create a control group similar to the treatment group in baseline characteristics9. Column (6) in Tables 1 and 2 presented earlier shows that PSM successfully neutralized the differences in background characteristics as well as outcomes at baseline between quasi-experimental groups.
Table 6 below reports the results in the exact manner as Table 5. However, the regressions in this table are performed on the matched, restricted sample.
9 Please see Figure 1 of the appendix for a plot of the baseline distribution of propensity
