Corruption and Product Market Competition: An Empirical Investigation. Michael Alexeev* and. Yunah Song** October 2012

Transcription

1 Corruption and Product Market Competition: An Empirical Investigation by Michael Alexeev* and Yunah Song** October 2012 * - Department of Economics, Indiana University, Bloomington, IN 47405; malexeev@indiana.edu ** - Korea Insurance Research Institute, 35-4 Yoido-dong Youngdeungpo-gu Seoul, , Republic of Korea, yunah.song0@gmail.com We would like to thank the anonymous referees and the participants of the conference Enforcement, Evasion and Informality: Theory and Practice (Ann Arbor, June 4-6, 2010) for insightful and helpful comments. All errors and omissions are our own. 1

2 Corruption and Product Market Competition: An Empirical Investigation Abstract We analyze the relationship between product market competition and corruption. The existing literature typically views corruption as extortion of pre-existing rents. This perspective suggests that competition usually reduces corruption, although generally the sign of this relationship is ambiguous. Shleifer and Vishny (1993), however, show that cost-reducing corruption is promoted by product market competition. That is, the effect of competition on corruption depends of the nature of corruption. Unlike the existing empirical studies that employ cross-country data and general measures of corruption, we test the competition-corruption relationship using firm-level information. Our approach overcomes significant estimation difficulties that result from relying on cross-country data; for instance, we include country fixed effects, and we deal with potential endogeneities by instrumenting competition with US capital-labor ratios for the appropriate industries. Contrary to the existing empirical work, we show that stronger product market competition is associated mostly with greater corruption of the costreducing variety. JEL Classification: D73, D21, O17 Keywords: Competition, corruption 2

3 Corruption and Product Market Competition: An Empirical Investigation 1. Introduction Control of corruption has been an important public policy issue both in developed and developing countries. Encouraging competition in product markets represents one potential approach to dealing with corruption among the officials regulating these markets, and this approach has attracted considerable attention in the theoretical literature. 1 This literature has demonstrated that the relationship between corruption and competition is complicated and depends on various factors such as the nature of corruption, technologies employed by the firms, preferences of corrupt officials, probability of punishment, and the information that the officials possess about firms. Most of the models treat corruption as appropriation by government officials of rents that accrue to incumbent firms in the industry. Shleifer and Vishny (1993) and Sequeira and Djankov (2010), however, point out that the effects of product market competition on corruption depend strongly on whether corruption is coercive (extortion) or collusive (cost-reducing). We argue that while the link between product market competition and coercive corruption is theoretically ambiguous, collusive corruption is promoted by competition and, therefore, empirical work should be cognizant of this distinction. The main goal of this paper is to examine empirically the competition-corruption relationship in a framework where the nature of corruption is more specific than in the previous empirical studies such as Ades and Di Tella (1999) and Emerson (2006). These papers show that countries characterized by a greater degree of product market competition tend to have less corruption. The reliance of these papers on cross-country 1 See Bliss and Di Tella (1997), Ades and Di Tella (1999), Straub (2005) and Emerson (2006). 3

4 data, however, has obvious drawbacks, including a small number of observations and the possibility of omitted variable bias. In addition, the degree of market competition in these papers is usually measured in rather indirect ways. For example, Ades and Di Tella use such measures as the share of imports in GDP and the distance to world s major exporters while Emerson uses indicators of an economy s competitiveness as reflected in the World Economic Forum s Global Competitiveness rankings and the Heritage Foundation s Index of Economic Freedom. Another drawback of using cross-country data is the difficulty of dealing with the potential for reverse causality between corruption and competition. As was noted in the aforementioned papers and elsewhere, corrupt officials may exercise their power to limit competition in order to generate rents for the incumbent firms rents that then can be extorted through bribes. 2 And perhaps most importantly, it is unclear what type of corruption is reflected in country-wide measures of corruption. As note earlier, corruption can be collusive (cost-reducing) or coercive (rent extraction) and the relative amounts of each type of corruption may vary significantly from country to country or sector to sector. Meanwhile, the consequences of product market competition for different types of corruption may be quite different. We attempt to complement and improve on the existing empirical work by relying on firm-level survey data that allow for better measures of competition, more specific 2 Both papers attempt to deal with potential reverse causality by using 2SLS estimation. Ades and Di Tella instrument the intensity of competition (proxied by share of imports in GDP) with the logarithm of population and logarithm of land area. Emerson instruments corruption, which is a right-hand side variable in his empirical model, with a civil liberties index and variables reflecting educational level in a country. Neither author presents formal tests of the validity of these instruments. While there is little doubt that the instruments used in each paper are correlated with the variables being instrumented, it is unclear why these instruments would be uncorrelated with the residuals. In both cases, instruments could affect the dependent variables through channels other than the variable being instrumented, -- an issue that regularly arises with the instrumental variable approach. 4

5 measure of corruption, and better controls and instruments than are available with crosscountry data. Our results are quite different from both Ades and DiTella and Emerson; we suggest that the differences could be due not only to the empirical approach, but also to the type of corruption captured in the studies. Our firm-level data appear to refer mostly to cost-reducing corruption, while country-wide indices may predominantly reflect coercive corruption. We use the data from the World Bank s Productivity and the Investment Climate Private Enterprise Survey (henceforth, PICS); this survey contains responses from several thousand firms across a number of countries. PICS covers the years , and some of the countries are included in more than one round of the survey, although this is not a panel dataset. Our basic approach is to regress the survey-based corruption measure on various measures of competition and some controls. Corruption is measured as the share of sales that firms similar to respondent s pay in the form of informal payments to get things done. We measure competition in several ways: the number of competitors the firm faces, hypothetical customer reaction to price increases, firm market share, markup over operating costs, and industry-level Herfindahl-Hirschman indices calculated from the same survey. Our controls include firm characteristics that are likely to be exogenous to corruption, as well as country and year fixed effects. We demonstrate that for the most reliable measures of competition, firms in more competitive environments tend to pay a greater percentage of their sales in bribes. While this relationship does not always hold strongly, we do not find any evidence that competition and corruption are inversely related, particularly if we control for potential endogeneity between competition and corruption by instrumenting competition with US 5

6 capital-labor ratios for relevant industries as a proxy for the firm s fixed costs -- one of the deep competition parameters suggested by Bliss and Di Tella (1997). 3 The advantage of this instrument is its clear exogeneity with respect to corruption in the surveyed countries. One disadvantage of this instrument is that it does not necessarily reflect the technological constraints in those narrow sectors where the survey respondents operate. 4 All statistically significant coefficients indicate a positive relationship between the strength of competition and the extent of corruption, though in several regressions the relevant coefficients are statistically insignificant. All coefficients in the IV regressions have signs consistent with a positive competition/corruption relationship, while the coefficients associated with the most reliable measures of competition are typically statistically significant. The rest of the paper is organized as follows. The next section briefly reviews the existing literature, focusing on the empirical implications of the theory. In addition, we suggest another simple model that implies a positive relationship between product market competition and cost-reducing corruption under some reasonable assumptions. We describe the data in Section 3. Our main results are presented and discussed in Section 4. Some robustness checks are performed in Section 5. Section 6 concludes. 3 Deep competition parameters are technologically based and are not influenced by institutionally created opportunities for corruption (Bliss and DiTella 1997, p. 1002). 4 The manufacturing sectors identified in the survey constitute 15 rather broad industries such as Textiles or Metals and Machinery. Therefore, both corruption and the competitive environment faced by firms within these industries may vary greatly, depending on what specific part of the sector the firm operates in and in what part of the country it is located. 6

7 2. The existing theory and evidence Most of the existing models of the relationship between product market competition and corruption produce ambiguous implications with respect to its sign. In the first paper to focus on this relationship, Bliss and Di Tella (1997) assume that each official deals with only one firm; officials do not know the precise amount of rent enjoyed by the firm they oversee, but they know the distribution of these rents. An official s problem then is to demand the bribe that maximizes the expected value of bribe revenue, while the firm agrees to pay the bribe as long as it is smaller than the firm s rent. Otherwise, the firm exits the market. The degree of competition in this model is based on three deep competition parameters: (1) the degree of substitutability of the firms products; (2) the degree of similarity of the firms production functions; and (3) the amount of fixed costs in the industry. In other words, in this paper, the extent of competition is determined by technological factors that are assumed to be exogenous with respect to the degree of corruption. When the degree of competition is determined by either the first or the second of the deep parameters, the relationship between competition and corruption (measured by the size of the bribe demanded) is ambiguous. If the degree of competition is determined by the third parameter (i.e., fixed costs), greater competition always increases corruption. 5 An increase in fixed costs has two effects. Higher fixed costs reduce the number of incumbents, generating greater operating profits for the remaining firms. But a fixed cost increase also reduces each incumbent s total profits available for extortion. In Bliss and DiTella s model, these two opposed effects result in lower overall profit and, therefore, lower bribes. 5 To prove this result, Bliss and Di Tella assume that the distribution of the firms overhead costs is uniform. 7

8 Ades and Di Tella (1999) also assume that each official deals with only one firm, but unlike Bliss and Di Tella, they assume that the official knows precisely the firm s amount of profit (which is random and is not observed by the state). The official may collude with the firm to hide the true amount of profit in exchange for a bribe. If the bribe is detected by the state, however, the official loses his wage. The state s problem is to set the officials wages in such a way as to reveal (and collect as a tax) the greatest amount of profit net of the officials wages. The degree of competition in this model is measured by the exogenous number of firms in the market and the extent of corruption is defined as the frequency of bribes of exogenous size. The assumed exogeneity of the number of firms implies that the direction of causality on which the model focuses is from competition to corruption. In this framework, corruption decreases in the number of firms unless increased competition leads the state to decrease substantially the officials wages. If the wages do not depend on the number of firms in the economy, an increase in competition would always reduce the frequency of bribes. 6 The two models outlined above assume that the bribes represent pure extortion and are obtained from the firms pre-existing rents. However, bribes are often paid in return for some service, even if this service consists in letting the firm bypass some regulation or informal red tape that exists to facilitate bribe-taking. The important point here is that corrupt officials are not always free to extract firms rents, but can charge bribes only up to the value of the service they are providing to the firm. As Shleifer and 6 Straub (2005) contains a model that is somewhat similar to Ades and Di Tella s in that the number of firms is exogenous and the officials assigned to each firm have perfect information about the firms rents. The firm s rents in this model depend on whether the firm uses a good technology that does not produce externalities or bad technology that is cheaper for the firm but generates a non-pecuniary negative externality for the consumers. The main result of Straub s model relevant for our purposes is that a greater number of firms may either increase or decrease corruption measured as the aggregate amount of bribes. 8

9 Vishny (1993) point out, extortion is more difficult to hide and easier to fight than costreducing corruption, such as when an importer pays a bribe instead of the official customs duty, or a firm pays a bribe to avoid complying with costly regulations. Shleifer and Vishny note that [c]ompetition between buyers of government services assures the spread of cost-reducing corruption (p. 604) while such competition does not promote the spread of corruption of the extortion kind. Sequeira and Djankov (2010) expand on Shleifer and Vishny s argument by classifying corruption into collusive and coercive. Collusive corruption emerges when public officials and private agents collude to share rents generated by the illicit transaction Coercive corruption takes place when a public bureaucrat coerces a private agent to pay a fee just to gain access to the public service. (pp ). Collusive corruption is always cost-reducing in the Shleifer and Vishny sense and leads to an increased demand for public services, while coercive corruption is cost-increasing, leading to lower demand. These arguments also suggest that cost-reducing corruption is likely to be more widespread than extortion. Cost-reducing bribes may be charged for reducing fixed costs or for reducing variable costs. This latter type of corruption would be more natural when bribes are charged in the course of day-to-day business rather than in a lump-sum fashion, e.g., when the firm needs to import supplies from time to time or wants to evade excise taxes on its output. It is straightforward to show that in a basic model of Cournot competition with linear demand curves, when firms are paying bribes to reduce their costs by a given amount, a decrease in the firms fixed costs (or, equivalently, an increase in the number of competitors in the market) results in greater corruption measured either as the percentage of sales (the so-called bribe tax) or as the total amount of bribes. 9

10 Specifically, consider a market with Cournot competitors producing a homogeneous good at a constant marginal cost and with a common fixed cost,. 7 The firms face an inverse demand curve, where and is the output of firm. The fixed and marginal costs determine the number of firms in this market via a zero-profit condition. 8 Let s assume that all firms have the same opportunities for cost reduction via a bribe and that the corrupt official obtains 100% of the cost saving. Denote each firm s fixed cost reduction by and marginal cost reduction by. Then the total bribe and bribe tax in each case would be: Fixed costs Marginal cost reduction reduction Total bribe: Bribe tax: Both the total bribe and the bribe tax increase in (and, therefore, decrease in ) in either case. We conclude that when the number of competitors is determined by the size of technological fixed costs (Bliss and Di Tella, 1997) or when corruption is cost-reducing, the bribe tax is likely to be positively related to the degree of competition. Most other recent papers that address the issue of product market competition and corruption treat the number of competitors and the degree of corruption as jointly 7 We use common fixed costs for simplicity. The outcome does not change if the fixed costs are heterogeneous as long as they change in a uniform fashion for comparative statics purposes. 8 If all firms marginal costs are the same, _ 2_, where _._ denotes the greatest integer function. We treat as a continuous variable. 10

11 determined, but emphasize the effect of corruption on competition rather than the other way around. The standard approach is to assume that in one way or another (e.g., via an entry fee bribe, or regulation, or by issuing licenses) the officials can restrict entry into the relevant industry and create rents that can be extorted. The officials then determine the extent of their restrictive action in order to maximize their bribe revenue. Recent examples of such papers include Campos et al. (2010), Emerson (2006), Dutta and Mishra (2004), and Aidt and Dutta (2001). Typically, an entry fee imposed by corrupt officials on firms that want to enter the market reduces competition, relative to a completely free-entry case. However, given the existing regulations (or taxes), corruption may result in greater entry than would be the case if regulators were honest, because paying bribes may be less costly for at least some firms than following regulations. Dutta and Mishra (2004) provide one example where, in the presence of regulations, wealth inequality can lead to both greater corruption and increased product market competition. 9 To summarize, most models of the effect of product market competition on corruption have ambiguous implications with respect to the sign of this relationship, making the empirical investigation of this issue particularly relevant. Moreover, it is likely that cost-reducing corruption increases in the number of competing firms (or, equivalently, decreases in fixed costs). This implies that the outcome of an empirical test of the relationship between competition and corruption depends on whether corruption measures predominantly reflect cost-reducing corrupt practices or pure extortion. Also, such tests need to take into account the potential endogeneity between corruption and product market competition. 9 Regulations may, of course, be a result of corruption, but presumably some taxes and regulations would also exist without corruption. We expand on the issue of the effect of corruption on competition later in the paper. 11

12 3. The Data and Estimation Approach The existing empirical work that suggests that product market competition reduces corruption is based primarily on cross-country data. This limits the number of observations and creates a potential for omitted variable bias. 10 Also, it is unclear whether the corruption measures used in these studies reflect rent extortion or cost-reducing corruption. We use firm-level data and, contrary to the results based on cross-country studies, show that cost-reducing corruption tends to be promoted by product market competition. In our tests we rely on the firm-level Productivity and the Investment Climate Private Enterprise Survey (PICS). This World Bank-sponsored survey was administered to several thousand firms, mostly in developing and transitional countries, in We restrict our attention to manufacturing firms. After dropping observations that do not contain information on the variables relevant to our analysis, we end up with a range of about 4,700 to 15,000 observations from developing countries in our baseline regressions. 11 The large difference in the number of observations in different regression specifications is mainly due to some respondents not providing answers to questions 10 Dutta and Mishra (2004) use a subset of the firm-level data from a precursor to the survey that we use in the next section, but their empirical work is limited to motivating their theoretical model. Their empirical exercise consists of regressing the ratio of corrupt to non-corrupt firms in a survey of 23 economies in transition on the number of firms in each country sample and on the Gini index for the country. (They assume that the number of firms in the survey reflects the number of firms in the country.) Also, they do not claim that causality runs from the degree of competition to corruption. Campos et al. (2010) use a survey of Brazilian manufacturing firms to show that corruption is positively associated with incumbent firm performance. They argue that this presents indirect evidence of corruption acting as a barrier to entry. Here again, causality runs from corruption to competition. Also, competition in their empirical work is reflected in a highly indirect way by assuming that incumbent firm performance is inversely related to the strength of competition. 11 In addition to developing countries, PICS includes several high-income OECD countries. While we use only non-oecd countries in our baseline regressions, we include all available countries when we perform some of the robustness checks discussed in the next section. 12

13 associated with all of the measures of competition. Because these questions do not have political connotations, we do not believe that variable response rates imply any significant biases. We discuss this issue in greater detail in Section 5 of the paper. Our dependent variable is a measure of corruption that we refer to as the bribe tax. It equals the fraction of annual sales paid in bribes reported by firm in country and in year. 12 Note that the bribe tax reflects informal payments to public officials to get things done with regard to customs, taxes, licenses, regulations, services etc. (see Table 1). That is, at least on its face, this question can be interpreted as being about costreducing corruption rather than extortion of rents. In fact, we essentially test whether the answers to this question reflect cost-reducing corruption. As we argued above, if this bribe tax reflects cost-reducing corruption, it should be positively influenced by the strength of competition. We use the following regression specification: _, (1) where _ is defined in the paragraph above, competition is measured by several different indicators listed below, is a vector of firm characteristics, and stands for country and year fixed effects. As with any survey-based measure of corruption, the interpretation of the bribe tax is open to debate. In particular, it is unclear whether the respondent refers to his own firm or to other firms like yours. In a sense, however, it does not matter for our purposes, because if the answer is really about firms like the respondent s, we can still use the respondent s firm s characteristics as controls. It is also possible that some respondents conflate the strength of the competition that they face with corruption by 12 As is common practice in surveys, the questions about the extent of corruption concern unofficial payments typically made by firms like yours rather than by the respondent himself. 13

14 other firms, perhaps thinking that the other firms can be so competitive only if they bribe their way out of regulations. We believe, however, that this approach to answering the corruption question is unlikely to be frequent and that most respondents are sufficiently aware of the situation in their industry to distinguish between competition and corruption. The survey includes or allows for calculating several measures of the intensity of competition faced by respondents: (1) number of competitors; (2) markup over firm s costs; (3) the extent of customer reaction to a hypothetical price increase; (4) national market share; (5) local market share; (6) and industry concentration (Herfindahl- Hirschman Index or HHI). Each of these measures has advantages and disadvantages that we discuss in detail in Appendix 1 where we also explain precisely how we calculated these measures. All of our variables are also described in Table 1. We view the first three competition measures as more reliable than the last three because the first three measures do not depend on the respondent s knowledge of other firms output and because these measures are not influenced by the survey s overly broad definitions of industries. As we mentioned earlier, competition and corruption are likely to be endogenous. In particular, in the case of coercive corruption, corrupt officials may attempt to limit competition among the firms they oversee in order to increase the available rents. Also, as Emerson (2006) suggests, corruption is easier to hide when the official deals with fewer firms. At the same time, cost-reducing corruption can facilitate competition relative to the case of no corruption, because corruption makes it easier to evade entry-impeding rules. However, assuming that a more onerous regulatory environment makes it easier to induce firms to pay bribes, more corrupt governments would generally develop more 14

15 onerous regulations regulations that would result in less competition even if they can be evaded by paying bribes. That is, competition would have been stronger and corruption would have been smaller if no such onerous regulations existed to begin with. Under this argument, corruption could generate greater competition relative to the efficient case but only if bribes are used mainly to avoid efficiency-enhancing or innocent regulations rather than excessive regulations. If corruption acts to reduce competition, OLS and Tobit estimates would be biased towards a negative association between competition and corruption. Since, as we demonstrate below, all of our statistically significant OLS and Tobit estimates point towards a positive relationship between competition and corruption, the potential endogeneity is likely to strengthen our results rather than weaken them. In order to adjust for possible endogeneity, we instrument competition by the capital-labor ratios in the corresponding US manufacturing industry. 13 We view the capital-labor ratio as a proxy for a firm s fixed costs one of Bliss and DiTella s deep competition parameters. Obviously, it is not a perfect proxy for fixed costs, but the technologically determined capital-labor ratio is probably a fairly good indicator of socalled innocent entry barriers. Such barriers arise due to technological industry characteristics, even in the absence of other factors influencing competition such as corruption or strategic behavior by profit-maximizing firms. Other things equal, the higher the capital-labor ratio, the more difficult it is to enter the industry. 13 We use the capital-labor ratios for the appropriate 3-digit US industries lagged three years relative to the survey observations. This lag has proven to provide for the best fit in our regressions. Also, because the US industry classifications changed somewhat in 1997, we could not use much longer lags if we were to assure maximum consistency in industry definitions over time. 15

16 US capital-labor ratios are clearly exogenous to competition measures in other countries. One could argue, however, that the capital-labor ratio could affect corruption via channels other than competition, because high capital requirements may necessitate more permits and make it more difficult for the firm to move to a different jurisdiction, implying that high capital labor ratio would result in more corruption. This is indeed a possibility, but we believe that this connection between the capital-labor ratio and corruption, if present, is much weaker than the competition channel. This view is supported by the fact that when we replace a measure of competition with the US capitallabor ratio in regression (1), the coefficient associated with the capital-labor ratio is negative and significant at the 5% level. Our instruments are strongly negatively related to most of the competition measures reported by respondent firms (see first stage regressions in Table 6) although they are only weakly correlated with HHI and virtually uncorrelated with the customer reaction measure. The weakness of the instrument for the customer reaction is probably due to the ordinal nature of this measure of competition. The relative lack of strength of the capital-labor ratio as an instrument for HHI is more difficult to explain, except that the survey sample may not be representative of the true distribution of firms according to the amount of sales. We also examined the US HHI index for 3-digit industries, but it performed poorly as an instrument. In fact, the US HHI was negatively and statistically significantly correlated with the corresponding HHI variable derived from the survey. We also control for firm characteristics that can be expected to influence the degree of corruption engaged in by the firm and that (plausibly) are not in turn influenced by corruption. In our baseline regressions these firm-specific control variables include 16

17 firm location characteristics (population of the city in which the firm is located and a dummy variable for the capital city) and the age of the firm. In addition, all of our regressions include country and year fixed effects. The inclusion of country fixed effects is particularly important, because these variables control for the country-wide degree of corruption and competition, thus isolating the environment within particular industries. The summary statistics and pairwise correlations for our variables are presented in Tables 2 and 3, respectively. 4. The Results OLS and Tobit regressions We begin by OLS estimation of equation (1) using our six measures of competition. The errors in the regressions could be correlated (clustered) within a given industry and a given country. For this reason, we use the variance estimator described by Cameron et al. (2006) that provides for robust inference when there is two-way nonnested clustering of errors. 14 This technique is appropriate for OLS as well as for nonlinear estimators, including GMM. Under this approach, the standard errors are calculated as, where is the conventional standard error with one-way clustering by, denotes industry, stands for country, and superscript refers to the observations that belong to both industry and country. In our regressions usually, but not always,. All of our standard errors are robust to the presence of arbitrary heteroskedasticity. The OLS results shown in Table 4 present a somewhat mixed picture. The coefficients associated with the number of competitors and the customer reaction 14 In this context, non-nested means that errors can be clustered within a given industry that cuts across countries and within a country that contains several industries. 17

18 measures are significant at the 5% and 10% levels, respectively. Both of these coefficient estimates are consistent with a positive relationship between corruption and competition. No other coefficients are statistically significant, although the signs of all but the coefficient for national market share are also consistent with a positive corruptioncompetition relationship. As we argue in Appendix 1, however, the market share variables in PICS are the least reliable measures of competition, because it is unclear which market (local, national or even international) is most relevant for each firm and because national market share survey responses from firms in the same country and industry often add up to much more than 100%. OLS may not be the appropriate estimation approach for equation (1) because the bribe tax is censored from below and many respondents report the bribe tax at zero. If the firm s willingness to pay a bribe is a monotonic and increasing function of the strength of competition, it is possible that the willingness to pay a bribe would be reduced to zero before competition disappears, i.e., before the firm becomes a monopoly. Meanwhile, we cannot observe any negative willingness to pay a bribe. In this situation, we would observe a zero bribe tax even as competition decreases beyond the point where the willingness to pay a bribe is zero. Under these circumstances, the appropriate estimation technique is Tobit rather than OLS. 15 Tobit estimates presented in Table 5 are quite similar to OLS estimates. Again, the coefficients of the number of competitors and of customer reaction measures are 15 More formally, suppose that for a firm with characteristics (or firm for short), the costs and benefits of paying a bribe in order to reduce operating cost per unit of output by amount depend on the strength of competition, 0,. Denote these costs and benefits by and, respectively, and let firm be willing in principle to pay a bribe,, that is equal to the net benefit it gets from the bribe, i.e.,. A positive relationship between competition and corruption would imply that 0. Then it would be reasonable to assume that at least for some firms, there exists 0 such that =0. Because negative bribes are not feasible, we would observe firm paying zero bribe even when. In other words, we observe such that when and 0 when 0. 18

19 statistically significant at the 5% and 10% levels, respectively, with the customer reaction coefficient s p-value equal to 0.057, i.e., very close to The remaining coefficients are not statistically significant and have the same signs as the respective OLS coefficients. The sign and statistical significance of the customer reaction coefficient in OLS and Tobit regressions are particularly important, because our instrument for this measure of competition is weak. Instrumental variables estimation The potential simultaneity bias between corruption and the intensity of product market competition represents a problem in estimating the relationship between them. As we argued earlier, we expect simultaneity to bias downwards the estimates of the coefficients for the number of firm competitors and for customer reaction in OLS and Tobit, while the coefficients of market shares, markup, and HHI will be biased upwards. The results presented in Tables 6 and 7 support these expectations. In all specifications, the IV point estimates of the coefficients change in the predicted fashion relative to the non-iv estimates, and the signs of all IV estimates are consistent with a positive relationship between competition and corruption. In the two-stage GMM regressions, the coefficients for the number of competitors and for markup are statistically significant at the 5% and 12% levels, respectively, and the standard error of the markup coefficient estimate is rather conservative. 16 In the IV Tobit regressions the same coefficients are significant at the 12% and 5% levels, respectively; other coefficients are not statistically significant. Note, however, that our instruments do not work for the customer reaction 16 For the standard error for Markup we follow Cameron et al. s (2006) suggestion and use the maximum of the and, because 0. This is a conservative approach, because had been between and, which is typically the case in the other regressions, the statistical significance would have improved. 19

20 measure (presumably because of its ordinal nature) and are rather weak in the case of HHI. Because of weak instruments, we discount the results of these regressions. The instruments are fairly strong for the market share measures, but as we argued earlier, market shares are the least reliable measures of competition. We conclude then that the coefficients for the measures of competition that are both reliable and for which our instrument is strong support the hypothesis that more competition leads to more costreducing corruption. The numerical importance of the competition measures is substantial, albeit not overwhelming. According to our point estimates for statistically significant IV Tobit coefficients, a one standard deviation increase in the markup measure would decrease the bribe tax by about 0.8 percentage points, or by about 60% if we start at the mean value of the bribe tax in the markup regression (i.e., at 1.31%). In terms of standard deviations of the bribe tax, this change would mean a decrease of about one-sixth. The bribe tax elasticity with respect to the number of competitors is close to Robustness checks The above results refer to the sub-sample of PICS countries that excludes high income OECD countries according to World Bank classification. We exclude these countries in our baseline estimates because the problem of corruption in these countries is less obvious than in the lower income countries, and it is quite possible that the relationship between product market competition and corruption in the OECD countries is different from that in lower income countries. In our full sample, these OECD countries (Germany, Greece, Ireland, Portugal, South Korea, and Spain) contribute from 20

21 none to almost one thousand observations depending on the measure of competition used in the regressions. While country fixed effects can pick up some of the differences between developed and developing countries, they can do so only to the extent the true relationships among our variables differ only in the constant terms. If the true slope coefficients differ significantly between OECD and non-oecd countries, combining all countries in the same regression would result in noisier estimates. The results for the full sample of countries are shown in Tables 8-9. They are broadly similar to those for non- OECD subsample. Note first that OECD countries contribute no additional observations for the number of competitors and for market shares. 17 Therefore, the only results that can change are those where markup, customer reaction, or HHI are used as measures of competition. The only more-or-less significant difference is that neither the markup nor HHI coefficients in the two-stage GMM regressions are now significant even at the 15% level, although the actual p-value of the HHI coefficient increases only marginally from to (Table 9). However, the more appropriate IV Tobit specification produces results that are very similar to the estimates for the non-oecd subsample. The PICS dataset contains many missing values both in the bribe tax variable and in competition measures. While we do not expect firms to be biased in reporting the strength of competition facing them, the bias may be present in the bribe tax variable. Jensen et al. (2007) demonstrate that firms in countries with less political freedom tend to exhibit a higher non-response rate with respect to corruption measures. Also, Jensen et al. argue that firms in these countries tend to report a degree of corruption that is considerably lower than the level of corruption suggested by the corruption control measure from the World Governance Indicators in Kaufmann et al. (2007). We do not 17 Presumably these questions were not asked in those countries. 21

22 expect these problems to affect our results, because observations with missing values of the bribe tax are excluded from our analysis and, more importantly, because we use country fixed effects that would absorb the cross-country variation in overall levels of corruption and political freedom. That is, our results are based on the inter-industry variation of corruption and competition and thus should not be affected by countrywide rates of non-responses or even false responses, except maybe by introducing extra noise into the data and reducing the statistical significance of our estimates. Moreover, the Jensen et al. (2007) results with respect to the truthfulness of survey responses may be irrelevant for our purposes, because they use the answers to a question about corruption being an obstacle (rather than the amount of the bribe tax) to test the truthfulness of firm responses. The degree to which corruption presents an obstacle for a business may not necessarily be correlated with the size of corruption. In fact, incumbent firms may be perfectly happy to function in a corrupt environment if corruption protects these firms from competition and permits them to lower their costs by evading regulations. For the data used in our regressions, there is no statistically significant relationship between the bribe tax and either Voice and Accountability or Rule of Law measures from the World Governance Indicators when we control for the country s corruption measure. Meanwhile, bribe tax is strongly negatively related to corruption control measure. 18 In any case, we cannot directly test the effect of the truthfulness of responses with respect to the bribe tax, because it is unclear how to measure truthfulness when controlling for country fixed effects as we do in all our competition-corruption 18 For example, in the Tobit regression of the bribe tax on the rule of law and corruption control WGI measures for 2002, their respective coefficients are (1.325) and (1.202). The robust standard errors clustered by country are shown in parentheses. The rule of law coefficient has p-value of 0.9 while the coefficient for corruption control is much larger and significant at 0.1% level. 22

23 regressions. But we can evaluate whether the rate of non-responses to the bribe tax question is correlated with our measures of competition and, therefore, might possibly affect our results. To this end, we create a dummy variable that takes on a value of 1 if the bribe tax variable is missing and a value of zero otherwise. We then use this dummy variable as a dependent variable in linear probability regressions, with our standard set of controls and error clustering. In the OLS regressions, only the HHI coefficient is significant at the 15% level, although even this coefficient is very small numerically. In the IV linear probability specification no relevant coefficients are even close to statistical significance. This suggests that missing values of the bribe tax are unlikely to be relevant to our estimation. 19 The regression results for testing the effects of missing values are available upon request. Finally, we tried two other specifications, both of which yielded results broadly similar to those of our baseline regressions. First, we ran the regressions without any control variables except for country and year fixed effects. In the Tobit regressions without controls, all competition coefficients have signs consistent with a positive relationship between competition and corruption, although only the number of competitors and customer reaction coefficients are statistically significant, both at the 5% level. The signs of the coefficients in the IV Tobit regressions without control variables also all point to a positive effect of competition on corruption, but only the national share coefficient is significant at the 10% level. The results of these regressions are shown in 19 Another approach we tried was to set missing values of the bribe tax to its mean levels for non-missing observations for the same industry and country. The Tobit and IV Tobit results with these modified data were similar to the Tobit results with the original dataset. All the relevant coefficients signs were the same although statistical significance in the IV regressions was lower than in the regressions based on the original data. This is not surprising, given the way we generated the additional data. We conclude that there is no meaningful evidence that missing values in the bribe tax variable affect the results of our benchmark regressions. 23

24 Tables A1 and A2 of Appendix 2. Second, we used linear probability specifications where the dependent variable was set to 1 for positive values of the bribe tax and to 0 otherwise. The only statistically significant competition coefficients in OLS regressions were those for markup (p-value of 0.054) and customer reaction (p-value of 0.034). In the IV specifications, none of the relevant coefficients were statistically significant. The lack of statistical significance is not surprising, given that the dependent dummy variable aggregates all positive bribe tax values into unity. This prevents different degrees of competition from mapping into different values of the bribe tax as long as the latter is positive. These results are presented in Tables A3 and A4 of Appendix Conclusion We study the relationship between product market competition and corruption. Most of the existing theoretical literature arrives at ambiguous results with respect to this relationship. We argue, however, that the link between product market competition and corruption to a large extent depends on the nature of corruption. While surplus-shifting corruption (extortion of rents) might indeed be negatively linked to the strength of competition, cost-reducing corruption is likely to be promoted by competition. The empirical work on this issue has generally dealt with broad measures of corruption and has uniformly claimed to show that stronger product market competition is associated with lower corruption. In contrast, our estimates suggest that more competition is associated with greater corruption, implying that the firms in the World Bank s enterprise survey face cost-reducing corruption. This result strengthens when we adjust for the potential simultaneity between corruption and competition. 24

25 An important advantage of our approach is the use of firm-level instead of crosscountry data. Firm-level data let us utilize information specific to the competitive and institutional environment of particular firms, rather than rely on countrywide measures that reflect the degree of competition and corruption in highly aggregated and sometimes indirect ways. In addition, our data allow for the use of apparently valid instruments to deal with the potential simultaneity between competition and corruption. Finally, the large number of observations and the use of firm-level controls and country and year fixed effects add to the reliability of our results. Our findings do not necessarily contradict the existing literature, but rather call for a more nuanced view of corruption, emphasizing the need to distinguish between surplus-shifting and cost-reducing corruption. We certainly do not view our results as an argument against promoting product market competition among firms. The effect of competition even on cost-reducing corruption depends significantly on the factors that restrict competition. For example, if competition is restricted by excessive regulations, the removal of these regulations is likely to both reduce corruption and facilitate competition. Our findings do imply that other things being equal competition by itself does not tend to reduce corruption and may even promote it. Moreover, whatever effect competition has on corruption, it is presumably dwarfed by its well-known welfare-improving properties. Our results suggest, however, that corruption reduction is not necessarily one of them. 25

26 References Ades, A., and R. Di Tella. (1999). Rents, Competition, and Corruption, American Economic Review, 89(4): Aidt, Toke & Jayasri Dutta. (2002). "Policy compromises: corruption and regulation in a dynamic democracy," Royal Economic Society Annual Conference , Royal Economic Society. Bliss, C. and R. Di Tella. (1997). Does Competition Kill Corruption? Journal of Political Economy, 105(5): Cameron, Colin A., Jonah B. Gelbach, and Douglas L. Miller. (2006). Robust Inference with Multi-Way Clustering, NBER Technical Working Paper 327. Campos, Nauro F., Saul Estrin, and Eugenio Proto. (2010). Corruption as a Barrier to Entry, unpublished working paper. Dutta, I., and A. Mishra. (2004) Corruption and Competition in the Presence of Inequality and Market Imperfections, Center for Development Economics Working Paper No. 123, Department of Economics, Delhi School of Economics. Emerson, P. (2006). Corruption, Competition, and Democracy, Journal of Development Economics, 81: Jensen Nathan M., Quan Li, and Aminur Rahman. (2007). Heard Melodies Are Sweet, but Those Unheard Are Sweeter: Understanding Corruption Using Cross-National Firm-Level Surveys, World Bank WPS 4413, November. Sequeira, Sandra, and Simeon Djankov (2010). An Empirical Study of Corruption in Ports, mimeo, accessed on September 8,

27 Shleifer, Andrei, Vishny, Robert, Corruption, Quarterly Journal of Economics, 108(3): Straub, S. (2005). Corruption and Product Market Competition, mimeo, accessed on August 22, World Bank. (2002). Productivity and the Investment Climate Private Enterprise Survey. 27