Corruption and Firm Growth - Evidence From Around the World

Corruption and Firm Growth: Evidence from around the World Abstract We empirically investigate the relationship between corruption and growth using a firmlevel data set that is unique in scale, covering almost 88,000 firms across 141 economies in 2006-2020, with wide-ranging corruption experiences. The scale and detail of our data allow us to explore the corruption-growth relationship at a very local level, within industries in a relatively narrow geography. We report three empirical regularities. First, firms that make zero informal payments tend to grow slower than bribers. Second, this result is driven by non-bribers in high-corruption countries. Third, among bribers growth is decreasing in the amount of informal payments — in both high- and low-corruption countries. We suggest that this set of results may be reconciled with a simple model in which endogenously determined higher bribe rates lead to lower growth, while non-bribers are often excluded entirely from growth opportunities in high-corruption settings. 1 Introduction A large literature has documented the adverse impact of poor governance on economic growth (e.g., Mauro, 1995; Robinson et al., 2005). A business environment characterized by corruption or weak rule of law makes investment riskier while depriving governments of revenue (e.g., Wei, 2000). It leads to an increased reliance on political and personal connections, which in turn distorts market signals and results in suboptimal allocation of capital and labor (Faccio, 2006; Murphy et al., 1991; Khwaja and Mian, 2005). At the level of an individual firm, however, the relationship between corruption and growth is more ambiguous conceptually and less studied empirically. Faced with onerous regulations and inefficient bureaucracies, a firm may find that its best option is to make informal payments in order to “grease the wheels.” The opportunity to jump the queue may be particularly valuable for firms with greater opportunity cost of wasted resources (Lui, 1985). In this scenario, firms that report making informal payments will conceivably grow relatively quickly. The relationship also depends, however, on a firm’s ability to negotiate with corrupt officials in a market that often lacks clear pricing benchmarks. Indeed, Svensson (2003), Bai et al. (2017), Amodio et al. (2021) and others document a broad range of corruption experiences among firms in narrowly defined industries or geographies. While arguments about the costs and benefits of corruption are well established, empirical evidence on the value (positive or negative) of corruption to a typical firm, holding the business environment constant, is relatively scarce. We aim to begin filling this gap by establishing several stylized facts about corruption and firm growth in a large cross-country firm-level data set covering almost 88,000 firms across 141 economies during 2006-2020. Our data are derived from the Enterprise Surveys, a set of repeated cross-sectional surveys conducted by the World Bank Group, which follow a common methodology. As part of face-to-face interviews, the respondents—senior managers or owners of firms—answer a wide range of questions about their firms’ activities and experiences, including those related to recent sales growth and employment, as well as corruption. The countries included in the surveys span a broad range of corruption environments, from Zimbabwe to Estonia (the latter has a corruption perceptions rating from Transparency International that is better than that of the U.S.). The range and scale of the data enable us to examine the correlation between firms’ corruption experiences and their growth, comparing businesses within a given industry and a subnational region at a point in time. We report several stylized facts. For the sample overall, firms that report making zero informal payments have relatively weak growth in sales and productivity when compared to other firms in their industry-region-year cell, as do firms with relatively high informal payments. (Thus, if one looks simply at a linear specification relating informal payments to growth—within an industry-region-year grouping—there is no significant association). Using a specification that 1 allows for considerable flexibility in functional form, we verify that there is indeed a discontinuity at zero in the relationship between informal payments and growth. We further argue that the zero bribe “penalty” does not reflect misreporting, as it is driven by survey respondents who also report (despite making zero informal payments) that corruption is a “moderate,” “major,” or “very severe” obstacle for doing business – responses that we would not expect from firms that wish to avoid complaining about corruption. We also examine how this relationship differs in high versus low corruption settings, splitting our sample countries at the median of the Worldwide Governance Indicator for Control of Corruption. We find that, whereas firms that make high rates of informal payments grow slowly in both subsamples, the zero bribe penalty is driven entirely by the high corruption subgroup. The magnitudes of the effects we describe above are substantial, implying a 3 percentage point lower annual sales growth rate for firms that do not make informal payments in countries where such practices are the norm. The negative relationship between informal payments and growth is such that non-bribers grow more slowly than firms making informal payments of up to 10% of sales. The relationships we document do not necessarily imply a causal link from corruption to firm growth and productivity – even with the inclusion of industry-region-year fixed effects, many of the caveats and critiques of cross-region or cross-country corruption analyses apply here as well. However, our data are able to reveal some nuance to the corruption-growth association: firms that pay high bribes grow more slowly, but so do firms that abstain from bribery altogether. While these regularities are not impossible to reconcile with reverse causation, we suggest that the simplest versions of such models would predict a subset of these results, but not the collective set of findings. We show that these patterns can be rationalized in a simple framework in which profit-maximizing firms have to deal with officials who issue permissions for the firm to pursue a certain profitable opportunity (such as construction permits or operating licenses). Firms which face (endogenously determined) steep bribe rates grow slowly, but so do firms which choose not to pay bribes and instead report bribe-demanding officials, and are unsuccessful in their petitions. We argue that in high-corruption settings officials may act with relative impunity, and these are thus the environments where we would observe non-bribing firms shut out of growth opportunities. We provide an illustrative model following our empirics to highlight the assumptions and implications of this framework. Our paper contributes to the literature on governance and economic development. Modern empirical research on this topic was launched by Mauro (1995), which documented the cross-country relationship between corruption, and investment and growth, during 1960-1985. In the intervening years, a sizable body of work has enriched our understanding of what leads to weak institutions, and the consequences for development (e.g., Hall and Jones, 1999, and Robinson et al., 2005; see also Besley and Mueller, 2018, and Acemoglu et al., 2019 on the importance of political accountability). Researchers have also turned their focus to microeconomic data to better understand whether and how corruption and other institutional failures 2 constrain firm growth overall and lead to allocative distortions across firms. We review this literature in the next section. We see our contribution as providing the largest-scale analysis of credible observational data on firm bribe payments and growth, which reveals several heretofore undocumented empirical regularities that may provide input into the development of better theoretical foundations for understanding corruption, and also inform future empirical research on the topic. The rest of the paper is structured as follows. The next section provides a more detailed review of prior research on corruption and firm performance. Section 3 outlines the data and methodology. Section 4 discusses our baseline empirical results, while in Section 5 we provide a set of heterogeneity and robustness results that build on and reinforce our main findings. We present in Section 6 a simple modeling framework that can rationalize our findings, and in Section 7 we discuss the extent to which alternative modeling approaches might fit with our data. Concluding remarks follow in Section 8. 2 Corruption and firm performance The link from institutional quality (as reflected in robust property rights enforcement and constraints on rent-seeking by those in power) to economic development plausibly runs in part through the impact on individual firms. Predation by officials and weak legal enforcement inhibits investment because of lower and more uncertain returns. Prior work suggests that bribery is thus several times more detrimental to firm growth than formal taxation, based on cross-country analysis of foreign investment (Wei, 2000) and firm-level evidence from Uganda (Fisman and Svensson, 2007). Furthermore, weakness in the rule of law tends to increase the reliance of individuals and firms on personal connections, thus blunting market signals and leading to labor and capital misallocation across firms (Bussolo et al., 2018). Governance can influence long-term economic outcomes in part by altering the structure of economic activity. Economies with stronger institutions tend to specialize in sectors that are more reliant on innovation and complex contracts, and utilize more production inputs to produce final goods (Nunn, 2007; Levchenko, 2007; Silve and Plekhanov, 2018). While extant evidence indicates that corruption lowers growth for firms overall, the relationship between the choices of individual firms—holding institutional context fixed—and performance is ambiguous. There are several well-known arguments for why bribe-paying firms may perform relatively well. Faced with onerous regulations and inefficient bureaucracies, firms may choose to make informal payments to “grease the wheels.” 1 Firms that are better connected, less principled, or otherwise face lesser frictions in making informal payments may 1 At the level of the economy overall, this argument may be turned on its head due to “endogenous red tape;” see, e.g., Guriev (2004). However, our analysis is within industry-region-year cells; this presumably allows us to control to some extent for the local regulatory environment. 3 also be afforded opportunities that are unavailable to other businesses, which may also allow bribe-payers to grow faster. Firm-level evidence on the corruption-performance relationship is limited. Fisman and Svensson (2007) find a strong negative association between informal payments and sales growth in Uganda in the 1990s; De Rosa et al. (2015) document a positive relationship in most economies in Central and Eastern Europe, apart from three high-corruption economies where the relationship is negative; Aterido et al. (2011) find inconclusive results on the link between the incidence of facilitation payments in a certain region/sector and employment growth. Foreshadowing the inevitable endogeneity concerns that will arise in our analysis, Bai et al. (2017) show that growth causes a reduction in bribe extraction, based on an instrumental variables strategy applied to Vietnamese firm-level data. Amodio et al. (2021) use a randomized experiment to study the relationship between firm characteristics and bribery in the Kyrgyz Republic. They show that rewarding tax inspectors for higher anonymous evaluations by firms decreases corruption but also shifts inspections toward firms with lower elasticity of demand, fewer competitors, larger market shares, higher revenues, and higher profits.2 As we noted at the outset, our contribution is not to solve the endogeneity problems inherent in studying the corruption-growth relationship, but rather to document in a vast and credible dataset a number of patterns that may inform our understanding of the theories and frameworks that we briefly delineate above, which have heretofore been studied only in more narrowly bracketed samples and settings. 3 Data Our empirical analysis draws on a cross-country firm-level dataset of unique scale. The Enterprise Surveys follow a common methodology and contain detailed data on sales, employment and the business environment for firms operating in 141 economies during 2006-2020. Most participating countries are low- and middle-income economies, although the sample also includes a number of advanced economies (such as Italy, Israel and Portugal) offering a wide range of corruption experiences, from, say, Zimbabwe to Estonia. The former is consistently ranked as among the most corrupt in global corruption perception surveys, while Estonia is ranked on par with Iceland (and eight ranks above the U.S.) in Transparency International’s 2020 Corruption Perceptions Index. These face-to-face representative surveys cover firms with at least five employees. Strat2 This evidence is consistent with an imperfect competition model developed in Amodio et al. (2021) where firms differ by demand elasticities. Like in our paper, higher bribe rates set by inspectors result in lower output. On the other hand, in Amodio et al.’s model, firms’ complaints about inspectors are an exogenous function of foregone profits (the complaints are anonymous and bring neither costs nor benefits to the firms); in our case, this is an outcome of strategic interaction between firm and bureaucrat given the country-specific level of enforcement. 4 ified random sampling is performed by broad sector (manufacturing, retail and other services, with further subsectors in selected economies), firm size (5-19 employees; 20-99 employees; and 100+ employees) and by subnational region. Surveys exclude enterprises fully owned by the state, and for other firms surveyors record the levels of state and foreign ownership. All participating firms operate in the formal sector. The dataset is a repeated cross-section, although a handful of firms enter the survey in more than one wave; in Section 5 we make use of the relatively small subset of respondents – 12 percent of our main sample – that were surveyed twice (or in some cases even three times). A typical (median) firm covered by the survey is a domestically-owned private-sector firm with 20 employees serving the domestic market, with US$ 570,000 in annual revenue. The median firm has been operating for 16 years and has two-year annual sales growth (in US dollars) of 5.2% (see Annex Table A2). We omit firms that report positive or negative change in sales in excess of 50 times over the two-year period as these likely reflect errors in the data, a total of 1.2% of sample firms. As part of the survey, respondents (who, recall, are all senior managers or owners of firms) are asked the following question: “It is said that establishments are sometimes required to make gifts or informal payments to public officials to ‘get things done’ with regard to customs, taxes, licenses, regulations, services etc. On average, what percentage of total annual sales, or estimated total annual value, do establishments like this one pay in informal payments or gifts to public officials for this purpose?” The wording seeks to elicit a truthful response given the topic’s sensitive nature. Additionally, the survey records various firm characteristics as well as data on sales and employment in the last fiscal year and the three years prior. On average, firms in the sample report spending 0.8% of revenues on informal payments. This figure ranges from more than 7% in Sierra Leone to 5% in Uganda to 3% in Albania, to less than 0.01% in most advanced economies. The informal payment rate—or the bribe rate— tends to be higher in economies where corruption is more pervasive based on the Worldwide Governance Indicator of the Control of Corruption. See Appendix Figure A1 for a country-level scatterplot depicting the relationship between Control of Corruption and the mean fraction of firms that report non-zero bribes (Panel A), as well as a scatterplot of the relationship between Control of Corruption and the average bribe rate for the subset of bribe-paying firms. As in Fisman and Svensson (2007), we exclude a small number of firms with unrealistic reports of informal payments, in excess of 50% of sales. While the majority of firms report making no informal payments, this does not necessarily imply corruption-free business environments for these firms. In fact, we see evidence that this is not the case in our own data. Surveyed firms were asked to evaluate corruption as a constraint to doing business on a five-point scale ranging from corruption being “no obstacle” (coded 0) to corruption being a “very severe obstacle” (coded 4). Of the firms that report no informal payments, 14% nonetheless described corruption as a severe obstacle to doing business, 5 suggesting that they may make a conscious choice to forego opportunities to make facilitation payments; only 37% of zero-bribe firms refer to corruption as “no obstacle” at all. Interestingly, the correlation between the country-level Control of Corruption variable and firms’ responses on this five-point scale is just as strong for firms that report zero informal payments as it is for firms that report positive bribes (Appendix Figure A1, Panels C and D). The nearly 88,000 firms that constitute our main estimation sample are drawn from a larger dataset of 167,286 firms. In addition to the small fraction of firms that are screened out as described above due to unrealistic data on informal payments or sales, a substantial number of firms do not provide responses to some of the key variables in our analysis. Specifically, 14% declined to answer the question about informal payments. This may reflect respondents’ genuine lack of knowledge of the issue or reluctance to answer the question if the firm is making informal payments. Our main concern is whether missing bribe data is correlated with sales growth, or with firms’ vulnerability to corruption. Importantly, we show in the first column of Appendix Table A1, missing bribe data is not predicted by firm growth; in the second column we show that it is also uncorrelated with country-level Control of Corruption index (since this varies at the country-level, column (2) has only sector-year fixed effects).3 The main dependent variable of interest is sales growth, calculated as the log difference between sales in the last fiscal year and the three years prior, divided by two. We convert sales figures in national currencies to US dollars. We consider growth of sales per worker, as a proxy for labor productivity, and employment growth as alternative dependent variables. Calculating growth rates from recollection of past sales may introduce measurement error. However, since all firms operate in the formal sector, they would be expected to have formal accounts that senior managers would have access to, and can be consulted during interviews. We generate a control to account for the fact that underperforming firms may have many complaints, real or imaginary, including extortion by corrupt officials – Kaufmann and Wei (1999) call such propensity to complain the “kvetch effect.” Our “kvetch index” is measured as the difference between a firm’s perception of transport, electricity and access to land as obstacles to its operations and the country average complaints about these aspects of the business environment. In each case, firms were asked to evaluate a potential constraint on the same five-point scale as for corruption. 3 Full details on observations lost from the original dataset are as follows: Of the full 167,286 observations, we exclude 43,137 observations with missing sales growth and a further 19,057 observations with missing data on informal payments. 1,382 further observations were discarded because of improbably high sales growth, and a further 193 because of informal payments above 50% of sales. We then exclude an additional 14,158 observations with missing information on age, ownership, kvetch index, lagged employment, sector or region. Finally, we drop 1,530 observations that were single observations in the region × sector × year cells that make up the fixed effects in our main model. Because so many of the observations are lost due to missing data on sales growth, we also look at the correlates of missing sales growth data in column (3) of Appendix Table A1; the most striking correlation by far is with age, which is unsurprising given that sales growth is defined retrospectively over two years, so growth is thus defined for firms that are at least two years old. In column (4), we measure age by a set of indicator variables; we find that indeed sales growth is missing primarily for very young firms (the omitted category is firms greater than 10 years old). 6 In Appendix Table A3 we show the patterns in bribe payments over time. Of the 141 countries in our sample, most only appeared in one or two waves; however, 37 countries were surveyed in each of the three waves, allowing us to look at how firms’ informal payments have changed in recent years. This kind of over-time comparison is generally difficult in perceptionsbased cross-country measures (see, e.g., Fisman and Golden, 2017). The most striking pattern is the dramatic decline in the fraction of firms that report paying bribes; the share of such firms fell by half across the three waves. Among the firms that do pay bribes, the average bribe rate has not changed substantially over time, remaining at 5-6% of sales. As we emphasize throughout, our findings are correlational, and we would not necessarily expect bribers to be identical on other dimensions to non-bribers. We can hope that our finegrained controls account for some of the differences between the two groups of firms. We may furthermore probe the extent to which the two groups are imbalanced on other covariates, which may give a sense of the extent to which omitted variable concerns are likely to loom large in the analysis below. In Table 1, we look at the characteristics of bribers versus non-bribers by regressing the non-zero bribe dummy on firm characteristics. We begin with a specification that includes only sector-location-year fixed effects and log(Sales), which serves as a proxy for size or visibility of the firm. The small and statistically insignificant coefficient on log(Sales) argue against any explanation for our results based on the targeting of high ability-to-pay firms with growth potential, as size is the most readily observable proxy for deep pockets; we return to a discussion of targeting after presenting our main results. In column (2) we include a number of other covariates that we use in our main regressions below, and which we might anticipate could be correlated with bribe payments. Exporters are more likely to pay bribes and foreign firms less likely to do so. The first of these very plausibly reflects interaction with the customs bureaucracy; the latter may reflect a host of factors that limit officials’ ability to extract payments from foreign firms, and their ability to resist such overtures. Also unsurprising – firms that tend to complain about other business impediments also are more apt to report paying bribes (though we note that this could be a bad control in this regression; comfortingly, its inclusion/exclusion has little effect on other reported coefficients here, nor does its inclusion affect the main results reported below). Neither age nor government ownership is correlated with paying bribes. In column (3) we add profits as a covariate, measured as revenues minus costs deflated by revenues. Note that the sample is far smaller in this specification, as most firms do not provide cost data. Profits are positively correlated with the likelihood of a bribe demand, which is consistent with some degree of targeting (though also with a correlation between profitability and growth prospects, a point we return to below). We note that none of the other coefficients change sign (or magnitude) relative to column (2), despite the shift in sample. 7 4 Baseline Results We regress the annualized change in the logarithm of the sales of firm i between time t − 3 and time t − 1 (∆log(Salesi )) on various parameterizations that capture the extent to which firms make informal payments to officials (Bribes): ∆log(Salesi ) = α + βBribesi + λZi + υrst(i) + i , (1) where Z is a vector of firm-level controls including the logarithm of lagged sales in US dollars, the logarithm of lagged employment, indicator variables for ownership type (state ownership, foreign ownership or private domestic ownership), the logarithm of firm age, and exporter status; υrst(i) is a set of 5,340 fixed effects for subnational region × sector × year, and i is the error term, with standard errors clustered at the region-sector-survey-wave level. The fixed effects absorb the many factors that may be correlated with both rent extraction and growth across industries or areas, and focuses our analysis on the relationship between bribery and growth within relatively narrow groupings of firms. Finally, Z also includes a control for a firm’s propensity to complain about various constraints on its operations, as captured by the kvetch index. Results based on specifications of this form appear in Table 2. Column (1) provides a minimalist specification which includes region-sector-year fixed effects, since we are focused on the within- rather than cross-economy relationship between informal payments and growth, and uses BribeShare, defined as informal payments as a fraction of sales, as our measure of bribery. Interestingly, while the coefficient on BribeShare is negative, it is not statistically significant, and is of modest magnitude.4 A one percentage point increase in the share of informal payments is associated with a reduction of 0.053 percentage points in the annual growth of sales; given the standard error of 0.052 percentage points, we can reject effect sizes larger than 0.15 percentage points at the 5% level, based on a two-tailed test. In column (2) we rerun this specification using instead an indicator variable for non-zero informal payments, Non-zero BribeShare, as our measure of firm-level bribery. Interestingly, the coefficient changes sign, with a positive relationship between making a non-trivial informal payment and firm growth. The coefficient is significant at the 5% level and large in magnitude: bribe-paying firms grow 1 percentage point per year faster than those with zero bribes. In column (3), we include the measure of bribe share as well as the indicator variable for non-zero bribes. Given the opposing forces observed in the prior columns, it is unsurprising that the coefficients on both terms increase in magnitude. The point estimate on BribeShare is 4 In Appendix Table A4, we present our baseline results in a weighted regression using the BEEPS sample weights; we favor the unweighted regressions, given the very wide range in sample weights – the most heavilyweighted firms receive a weight that is 50,000 times that of the least-weighted firms (and it is also unclear that weighted regressions are preferred in any event (Solon et al., 2015)). In practice, the results are very similar. 8 -0.14, and 0.018 on Non-zero BribeShare (p < 0.01 in both cases). In our preferred specification in column (4) we add the full set of firm-level controls described above; for brevity, we do not report the point estimates on these controls (these are presented in Appendix Table A5). In this preferred specification, both coefficients increase marginally in magnitude, and are estimated with comparable precision as in the previous column. The point estimates imply that firms paying even fairly high bribes grow more rapidly than bribe-abstainers, with the ‘crossing point’ coming only at BribeShare = 0.09 (0.021/0.232), which is the 78th percentile of the distribution of BribeShare conditional on taking a positive value. Finally, in column (5) we allow for greater flexibility in the relationship between the share of informal payments and firm growth via a set of indicator variables which capture various ranges of BribeShare. The coefficients reflect the link from informal payments to growth relative to the omitted category of BribeShare = 0. Broadly consistent with the prior results, we observe a positive relationship between relatively modest bribe payments and growth, with the relationship only changing sign at high rates of informal payments (BribeShare > 0.10). We believe these results to be the first documentation of such a non-monotonic relationship between firm bribe payments and firm performance, one which is enabled by the granularity, detail, and scale of our data. Below, we will argue that it is unlikely to be driven by misreporting of zeros, and furthermore can be reconciled with an intuitive framework of how firms respond to the demands of corrupt officials. Our intuition for this non-monotonicity is that firms which pay zero bribes are one of two types: (i) those which may operate without having to make informal payments; and (ii) those which choose not to pay bribes, and are potentially shut out of growth opportunities as a result. Intuitively, the latter effect should be more dominant in settings where corruption is pervasive. Motivated by this intuition, in Table 3, we split the sample according to country-level corruption rankings, based on the Worldwide Governance Indicator of Control of Corruption. This measure—available at the country-year level—aggregates data from a large number of available surveys of corruption. Each year, the resulting country-level measures are normalized to have a global mean of zero and a standard deviation of 1 (see Kaufmann et al., 2009 for details). Higher values correspond to lower corruption. The Control of Corruption measure is highly correlated with other cross-country measures of corruption such as Transparency International’s Corruption Perception Index. We classify countries as low-corruption and high-corruption using the median observation of Control of Corruption in our sample. As the Enterprise Surveys focus on lower-income, less well-governed economies, the cut-off of -0.43 is well below the global mean. Thus, countries like Bolivia, India or Vietnam belong (just barely) to the low-corruption subsample. As expected, informal payments are far higher in high-corruption countries relative to low-corruption ones. For example, in the former group 20% of firms report non-zero informal payments versus 9.6% 9 in the latter. We rerun our preferred specification that includes both BribeShare and Non-zero BribeShare, as well as firm-specific controls, for the two subsamples separately in Table 3. The coefficient on BribeShare is negative (p < 0.05) in both instances. However, the indicator variable denoting whether a firm makes informal payments at all is significant (p < 0.01) and large in magnitude only in the high-corruption subsample. That is, among bribe-abstainers, there is no growth penalty in lower-corruption settings, only in high-corruption environments.5 Before turning to the hetereogeneity and robustness analyses in the next section, we note that the effects we document above come primarily through labor productivity (sales per worker) rather than changes in employment. The results for labor productivity (Appendix Table A7) look very similar to those of Table 2, while the relationship between firm corruption experiences and employment growth is relatively weak (Appendix Table A8). 5 Heterogeneity and Robustness In this section, we begin by exploring the extent to which the primary relationships we document in the preceding section vary across firm, sector, or country characteristics. In addition to providing the reader with a fuller view of the data, this analysis aims to probe the robustness of our results, and to provide further insight into the classes of models that may rationalize the non-monotonic relationship we document above. In the latter part of the section, we consider further issues of robustness, in particular with respect to functional form, and possible reporting bias that may arise as a result of the sensitive nature of bribe responses. Throughout, we focus on our preferred specification that includes subnational region × sector × year fixed effects as well as firm-level controls, unless otherwise noted. 5.1 Firms reporting bribes as an obstacle As we observed earlier, a firm that makes no informal payments may nonetheless be vulnerable to corruption. Indeed, when bribery is common, such firms may be particularly impacted by corruption because they are unable to exploit business opportunities. We thus split the sample based on whether a firm self-reports that corruption is at least a “moderate” obstacle to growth, under the presumption that the zero-bribe penalty should be more severe amongst those complaining of corruption as an obstacle to growth. We further argue that the ‘high-obstacle’ sample is useful because it includes firms for which underreporting of informal 5 In Appendix Table A6, we further subdivide the sample into corruption quartiles; we again observe that the zero-bribe penalty is higher, and monotonically increasing, in country-level corruption. In the lowestcorruption quartile, there is no zero-bribe penalty. 10 payments is likely less of a concern, since these ‘complainers’ were willing to describe government corruption as a problem.6 We report the results of this sample split in Table 4. For the high-obstacle subsample in column (1), the point estimates on the bribe variables are greater in magnitude, relative to the full sample results, while the bribe coefficients for the low-obstacle subsample (column (2)) are smaller and, in the case of BribeShare, does not approach significance. (We obtain virtually identical results if we define complainers based on whether firms are above or below the country-level mean; see Appendix Table A9.) It may also be instructive to compare how growth compares between non-complainers and complainers, focused on firms that do not pay bribes. Intuitively, there may be firms that refuse to pay bribes and suffer from it (as bureaucrats may block their projects). Second, there are firms that do not pay bribes but do not suffer from it (e.g., they are not asked to pay or perhaps are asked to pay but complain and are fortunate enough to see the bureaucrat removed). We assume that firms of the first kind are more likely to complain about corruption as an obstacle than those of the second kind. It is also natural to conjecture that the first ones have slower growth (their projects are blocked) than the second ones. In Appendix Table A10, we show that the growth rates of non-bribing complainers are indeed lower than those of non-bribing non-complainers. 5.2 Competition Amodio et al. (2021) and other models of bribery and its consequences presume that firms are targeted based on their ability-to-pay, which is in turn a function of competition and thus demand elasticity. It is therefore natural to consider the extent to which the patterns we document vary by extent of product market competition. In a subset of surveys, firms were asked about the number of competitors they faced for their primary product. In most cases, they were given various ranges: 0; 1; 2-5; more than 5. For firms that provide a specific number, we collapse their responses into these categories and because so few firms report zero or one competitor, we pool these two groups. In the first column of Appendix Table A11, we revisit our main results, including the interactions of our main bribery measures with the firms’ selfreported product market competition. None of the interactions approach significance, providing at least suggestive evidence against demand-side explanations for our main results. (Note that competition is negatively correlated with sales growth, suggesting that firms’ responses are capturing some element of their competitive environments.) In columns (2) and (3) we explore a more fundamental assumption of demand-side models — whether competition is correlated with bribe demands, whether based on a dummy variable 6 We cannot rule out the opposite bias – that firms which report bribery as a problem are more likely to under-report bribes for fear of reprisal, though we believe that intuition is more straightforward in the other direction. 11 for non-zero bribes, or a continuous measure of bribe shares. In the first case, we find that firms in more competitive settings are more likely to report paying positive bribes, while the point estimates on the competition variables are insignificant in predicting bribe shares. 5.3 Vulnerability to government decision-making In this subsection, we look at heterogeneity in the bribe-growth relationship as a function of attributes that potentially reflect a vulnerability to decisions made by government officials. Specifically, we split the sample based on whether a firm (a) has secured or attempted to secure government contracts; (b) engages in import activity; (c) engages in export activity. We acknowledge that all three dimensions surely capture a broad set of firm characteristics, and thus any differences (or lack thereof) along any of these dimensions should be interpreted with caution. We also note that export and import status is missing for a very large number of firms. With those caveats in mind, we present the three sets of sample splits in Table A12. While the coefficients do vary from column to column, the most striking feature of the table is the fact that the main U-shaped relationship between bribery and growth exists across all subgroups.7 5.4 Sector- and Country-level heterogeneity We conclude with several tables that split the sample by sector and country characteristics. In Appendix Table A13, we disaggregate the sample into firms in Light and Heavy manufacturing, Wholesale and Retail, Utilities and Construction, and Others; we observe quite similar patterns across all sectors. In Appendix Table A14, we consider whether the bribe-growth relationship differs by other country-level attributes, in particular level of democracy and income. In columns (1) and (2), we split the sample based on whether a country’s Polity IV score is at least 6; we observe similar patterns for both democracies and non-democracies. In columns (3) and (4) we divide the sample based on whether it is classified as a lower-income or middle-income country by the World Bank (our sample includes very few high-income countries). Again, the main patterns are observed in both subsamples. 7 A related dimension of heterogeneity is the type of bribe paid. In particular, one may distinguish amongst, for example, bribes paid to sidestep labor or environmental rules, to obtain government contracts, or to prevent hold-up by customs officials. Unfortunately, our data are not well-suited to drawing these distinctions. While firms were asked about various types of bribes paid, there are many missing values for each type, and if one looks at the intersection of firms with non-missing data for all bribe types, the sample drops to about 1000 firms. 12 5.5 5.5.1 Robustness checks Functional Form In the preceding section, we explored the non-monotonicity of the bribe-growth relationship via our two main specifications, both of which place particular emphasis on the bribe versus no-bribe threshold. We begin our robustness checks by further examining the robustness of this discontinuity, and whether alternative functional forms fit the data better. A natural approach would be to provide a non-parametric plot or kernel-weighted local polynomial regression relating bribes to growth, to trace out the relationship in a less restrictive way. However, as seen in Appendix Figure A2, which shows the distribution of BribeShare responses in the range (0, 0.10], such an exercise is not feasible. Specifically, we observe large modes at round numbers, making it impossible to estimate a continuous relationship. This is unsurprising given that, as we describe in the Data section, most BribeShare data are based on respondents’ estimates of informal payments as a fraction of sales. When we telescope in to the range (0, 0.001], i.e., just above zero, we observe a considerably smoother distribution, albeit for a relatively modest number of observations (see Appendix Figure A3). To further probe the relationship between informal payments and growth in these barelypositive observations, we subdivide the category (0, 0.001] into roughly-equal quartiles – (0, .0001], (.0001, .0003], (.0003, .0006], (.0006, .001]8 – and repeat the analysis from Table 2, column (5) above, with this finer partition of the smallest category. These results appear in the first column of Appendix Table A15. We observe that for all very low levels of informal payments, the coefficients are positive and of approximately the same magnitude, consistent with the “zero bribe penalty” we described in our main results. Based on this specification, we further graph out the coefficients as well as confidence intervals in Figure 1, which provides a clear visual representation of the bribe-growth relationship. We also explore whether a higher-order polynomial provides a better fit for the data. In the second column of Appendix Table A15, we augment our main specification from Table 2 to include a quadratic BribeShare term. The coefficient on the linear term is somewhat more negative and still significant (p<0.01); the quadratic term is positive, implying some concavity in the relationship, though the point estimate does not approach significance. The point estimates imply a downward-sloping relationship for all bribe shares below 0.341/(2*0.332)=0.51. In column (3) we add a cubic term; the higher-order terms are again at most marginally significant, and their point estimates imply a downward-sloping linear relationship for all positive bribe rates. Overall, these patterns suggest that the parsimonious specification in our main table captures well the main features of the bribe-growth relationship in our data. 8 This leads to groupings of 386, 341, 370, and 318 observations. 13 Throughout, we have focused on bribes as a fraction of sales as our measure of bribe extraction. We consider in Appendix Table A16 a specification that uses the natural log of (one plus) total bribe payments instead. The patterns are similar to those described in our main results – among bribe-paying firms, those that make higher bribes also grow more slowly. 5.5.2 Respondent Demand Effects Social desirability bias and related issues inevitably arise in studying corruption via survey responses. As with most corruption surveys, the Enterprise Surveys are designed to minimize respondent bias, by asking firms about “establishments like this one” rather than themselves. The firm does not thus implicate itself directly (in a legal sense or otherwise) in illicit behaviour. While the particular wording used in the survey may alleviate some questions of legal culpability, the issue of social desirability bias surely remains, and is not fully obviated by the indirect wording nor the kvetch index control (which captures a general willingness to complain) that we utilize in our main analysis. In our main analysis, we use a kvetch index that most closely parallels that of earlier work, but in this subsection we consider an alternative ‘propensity to complain’ measure that is more specific to complaints about government officials, and thus may be a more direct proxy for a firm’s general willingness to be critical of the government. We use questions from the Business-Government Relations section of the questionnaire, focusing in particular on the questions about the extent to which the following are barrier to business: (a) tax rates; (b) tax administration; (c) business and licensing permits; (d) political instability. We aggregate these into a “Government Kvetch Index” (GKI) which has mean zero and standard deviation one, increasing in a tendency to complain about government. We first observe that the GKI is positively correlated with BribeShare (or the indicator variable for non-zero informal payments), but the correlation is far short of one. For example, the raw correlation between the GKI and informal payments is 0.09, and the correlation with the non-zero dummy variable is 0.13; the correlation continues to hold even when fine-grained controls are included. When we include the GKI as a covariate in our main regression, however, our results are essentially unaffected. We provide these additional results in Appendix Table A17. Overall, while we cannot fully rule out the possibility that respondent demand effects play a role in our main findings, we believe that the discussion and results we report here suggest that it is unlikely to be the case. 5.5.3 Analysis Based on Resurveyed Firms Enterprise Surveys re-interviewed a small subset of firms between two and eleven years after the initial survey. Of the 87,829 firms in our total sample, drawn from 141 countries, 14 10,691 (12 percent) were resurveyed, representing 86 countries. Of these, 2,474 were resurveyed a third time, between 2 and 9 years after the second survey. We may use this subsample to assess the extent to which it is problematic that, in our main analysis, we use contemporaneous bribe payments and growth over the previous two years. Implicit in the formulation we use in our baseline specification is that bribe extraction is stationary or at least slow-moving. We can provide a check on whether the timing of bribes versus growth might be driving our results by utilizing data from the resurveyed firms. Because of the much smaller sample size, it is not possible to include the full set of sector×countryregion×year fixed effects. We use instead sector, country-region, and year fixed effects (rather than their interactions). We present these results in Appendix Table A18, in which our measures of bribery come from lagged survey responses, and the two-year sales growth outcome is based on a followinground survey. We include the same covariates as in our main analysis, and also a set of dummy variables to capture the time elapsed between surveys (between two and eleven years, with 67 percent of firms resurveyed between four and six years after the initial survey). In this subsample, the coefficient on Non-zero BribeShare is almost identical to the estimates in Table 2, while the coefficient on BribeShare is negative though about half the magnitude of the coefficients estimated in our main table; owing to the much smaller sample size, these results are imprecisely estimated, however (p-values of 0.10 and 0.33 respectively). To assess the role played by the different samples in comparing this subsample to our main results, in the second column, we present results for this subsample, but using the specification from our main table (i.e., contemporaneous rather than lagged bribe variables on the right-hand side); these point estimates are virtually identical to those in Table 2 though again imprecisely estimated. Finally, to assess whether the different fixed effects are relevant for comparing these results to our main table, in the third column we use the full sample from Table 2 but the more limited set of fixed effects used in column (1) of Table A18; again the point estimates are very similar to those of our main table. While not dispositive, these patterns suggest that the particular time structure of our data is unlikely to account for the patterns we document in Table 2. 6 The model We present a simple game-theoretical model of corruption that shows how the empirical findings above can emerge in equilibrium. 15 6.1 Setup Consider three sets of risk-neutral agents: principal P , bureaucrat B, and a continuum of firms F normalized to 1. Firms are indexed by i ∈ [0, 1]; they differ in their cost structures; the cumulative distribution function is G(·). The principal is not a strategic player. If a firm reports corrupt behaviour by a bureaucrat, the principal fires the reported bureaucrat with probability 1 − k.9 The parameter k ∈ [0, 1] is a country-level measure of corruption. If fired, B incurs a non-pecuniary cost z. The parameter z should be understood as a bureaucrat-specific preference parameter. Each firm i has a growth opportunity which generates additional output y at a firm-specific cost ci (y). ci (y) is an increasing convex twice differentiable function: c0i (y) > 0, c00i (y) > 0. The growth opportunity may involve a fixed cost, so c(0) may be positive. In what follows, it will be convenient to use the profit function: πi (p) ≡ max {py − ci (y)} . y Here p is the price of output y. For simplicity, we normalize units of output so that the price in the absence of bribes is equal to 1. Thus, when the firm expects to pay a share b of its output in bribes, its after-bribe profit should be πi (1 − b). We denote the profit-maximizing level of output yi∗ (p) ≡ arg maxy {py − ci (y)}. By definition, yi∗ (p) is an increasing function. The envelope theorem implies that yi∗ (p) = πi0 (p). The bureaucrat extorts bribes by threatening to block the growth opportunity. He maximizes the expected amount of bribes minus the cost of being fired if caught. 6.2 Timing The game is as follows: 1. B makes a take-it-or-leave-it offer to F, that it pay a share b ≥ 0 of output as a bribe, or B will block its growth opportunity. B does not observe the cost structure, so the offer is the same for all firms.10 2. Each firm i chooses whether to accept or reject the offer. (a) If F accepts, it then chooses yi . F receives (1 − b)yi − ci (yi ). B receives byi . 9 10 It is easy to provide microfoundations of this behaviour with a strategic principal and multiple bureaucrats. In Appendix B we consider a more general model where B can offer contract menus with different bribe levels and different outputs and the firm self-select into contracts. In Section 7, we discuss why this model’s predictions are not consistent with our empirical findings. Our assumption that B makes the same offer to all firms is related to difficulties of enforcing sophisticated contracts outside the legal system. 16 (b) If F rejects, it complains to the principal P. The outcome depends on the country’s level of corruption k: i. With probability k, the complaint is neglected, and B denies the growth opportunity to F. Both get zero.11 ii. With probability 1 − k, B is fired, and there is no bribe b = 0. F chooses yi and receives yi − ci (yi ). B pays cost z. 6.3 Assumptions We make several assumptions that simplify the analysis below. Without loss of generality, (1) 12 we sort the firms by ξi ≡ ππi0 (1) . We denote the density function g(ξ) ≡ G0 (ξ). We assume that i   the density is finite and that the distribution G(·) of ξi has a finite support 0, ξ¯ : ¯ = 1. Assumption 1. There exists a finite ξ¯ > 0 such that G(ξ) The second assumption ensures that the equilibrium bribe share is small b  1 (as observed in the data): π 0 (1) Assumption 2. For all firms g(0) π00i (1)  1. i As becomes clear later, this assumption allows for a Taylor expansion at b  1 which simplifies the analysis. Finally, we make the following “comparability” assumption: Assumption 3. In the absence of bribes, firms have the same output growth yi∗ (1) = y ∗ . This assumption implies that the heterogeneity between firms relates to their sensitivity to bribes, while without bribes their output is the same.13 This implies that even if the bureaucrat observes the output before approaching a firm, he still does not have information regarding its profits and thus does not know whether the firm will accept or reject the offer to pay the bribe. 6.4 Analysis We first need to determine which firms accept to pay the bribe and which ones refuse. When firm i agrees to pay, it receives maxy {(1 − b)y − ci (y)} = πi (1 − b). If it refuses, the 11 The firm does not pay the fixed cost ci (0) as the latter is associated with the growth opportunity rather than the firm’s current business. 12 This parameter can be understood as the profit margin (profit/sales ratio) in the absence of bribes. Indeed, the sales in the absence of bribes are yi∗ (1) = πi0 (1). In terms of the cost function, ξi is a proxy for the relative importance of fixed versus variable costs. The higher the fixed costs, the lower ππi0 (1) . i (1) 13 The assumption of the same output in the absence of bribes does not rule out heterogeneity across firms in terms of cost structures. For example if firms have the same variable costs but different fixed costs, they have the same yi∗ (1) = πi0 (1) but different πi (1) and therefore different ξi . Firms may also differ in terms of πi00 (·) and other characteristics. 17 expected payoff is (1 − k)πi (1). Therefore, firm i agrees if and only if k≥ πi (1) − πi (1 − b) . πi (1) Assuming that the equilibrium bribe share is small (b  1), we obtain b πi (1) ≤ 0 = ξi . k πi (1) (2) Since firms are sorted in the order of increasing ξi , for a given b there exists a cutoff î(b/k) that for all i > î(b/k) firms agree to pay the bribe and for all i < î(b/k) firms refuse. Obviously, î(b/k) is increasing in bribe b, holding country-level corruption k constant. Condition (2) implies that firms with higher profit to sales ratios are more likely to accept paying a bribe. This is intuitive as firms that agree to pay a bribe lose a share b of their sales, while firms that refuse lose their profits with probability k. This result is consistent with the data (see Column 3 in Table 1). We can now solve for the bureaucrat’s optimal strategy. His payoff is Z ξ̄ b yi∗ (1 − b)dG(ξi ) − (1 − k)zG(b/k). (3) b/k where yi∗ (1 − b) is the optimal output of firm i, given the bribe, and yi∗ (1 − b) = πi0 (1 − b) = y ∗ − bπi00 (1) + o(b) is a decreasing function of b. When maximizing (3), the bureaucrat faces a standard trade-off: higher b increases the bribe collected from each firm that agrees to pay the bribe, but also in a greater share of firms refusing to pay the bribe.14 The trade-off leads to the optimal choice of b as a function of the distribution of firms’ types G(·), the bureaucrat’s cost of being fired z, and country-level corruption k. 6.5 Comparative statics We now discuss the comparative statics with regard to exogenous parameters k and z, and show that the results are consistent with the empirical findings from Section 4. 1. Extensive margin. Let us compare the output growth y of firms with b = 0 and b > 0. For a given b set by 14 This trade-off is similar to the one in a standard monopoly model where the monopolist raises its price even if a higher price reduces the volume of sales. 18 B, firms that refuse to pay a positive bribe have average output of h i y 0 ≡ (1 − k)E y ∗ |i < î(b/k) = (1 − k)y ∗ . Firms that agree to pay a positive bribe have average output of h i y + ≡ E yi∗ (1 − b)|i > î(b/k) < y ∗ . (4) We now compare y 0 and y + for countries with high and low corruption k. If a country is very clean k → 0 then b → 0;15 therefore both y 0 and y + converge to y ∗ . So for perfectly non-corrupt countries y 0 = y + . Now consider k = 1. In this case y 0 = 0. To find y + , we solve the bureaucrat’s maximization problem. Taking the first derivative of (3) with respect to b, we obtain the first order condition: Z ξ̄   00 [y ∗ − 2bπi00 (1)] dG(ξi ) = bg(b) y ∗ − bπI(b) (1) . b Using b  1, we find b = 1/g(0). As the bribe share is small, each firm that pays a positive bribe has a positive yi∗ (1 − b) = y ∗ − bπi00 (1). Therefore, the average output growth for these firms (4) is also positive so y + > y 0 . This analysis implies that if country-level corruption changes from high k = 1 to low k = 0, the output of firms refusing to pay bribes y 0 = (1 − k)y ∗ increases from zero to y ∗ . At the same time, the output of firms agreeing to pay bribes y + increases from a positive but inefficiently low level (y + < y ∗ ) to the efficient level y ∗ (the latter being the case in the perfectly clean country with k = 0 and b = 0). The intuition is straightforward. Firms that refuse to pay bribes b = 0 are either (i) the ones denied access to growth opportunities (with probability k) and have y = 0, or (ii) those that avoid paying bribes (with probability 1 − k) and produce y ∗ .16 As k increases, the average output growth of this group falls from y ∗ to zero. At the same time, firms with b > 0 produce a suboptimal but non-trivial amount yi∗ (1 − b), so for sufficiently high k we have y 0 < y + . If there is no corruption at all (k = 0 and b = 0), then there is no discontinuity at b = 0: both y 0 and y + converge to the efficient level y ∗ . If corruption is very high, there is a discontinuity: y 0 = (1 − k)y ∗ = 0, while firms that pay a small bribe produce y ∗ − bπi00 (1) which is strictly positive. 2. Intensive margin. 15 16 ¯ otherwise no firm agrees to pay a bribe. For a given k, B always chooses b ≤ ξk, Obviously, the former are more likely to complain about corruption than the latter. Our analysis therefore implies that among non-bribers, we should observe a negative correlation between complaining about corruption and growth. As noted earlier, this is consistent with what we observe in the data (see Appendix Table A10). 19 We shall now check that once a firm pays a positive bribe, an increase in b reduces the firm’s y. As b is an equilibrium outcome in a game between firm and bureaucrat, for exploring this issue, we need to change an exogenous parameter — such as the bureaucrat’s propensity to extort larger bribes. Let us consider the comparative statics with regard to the bureaucrat’s cost of being fired, z. Suppose that there is a bureaucrat who is less concerned about being fired, i.e., his z is lower. Then for the same country-level and firm-level characteristics (k and G(·)), the equilibrium bribe share b is higher. Indeed, the monotone comparative statics imply that the b that maximizes (3) decreases with z. Facing a higher b, some firms that previously paid the bribe will now refuse to pay. However, for those that continue to pay, output will decline. Indeed, yi∗ (1 − b) is a decreasing function of b. 7 Discussion of results and alternative explanations As we have emphasized from the outset, our empirical findings do not necessarily reflect causal links from corruption to growth. Causation may run the other way, with informal payments made by firms depending on their performance. Firms that grow faster may be targeted for extra checks and inspections by rent-seeking officials in expectation of receiving facilitation payments. In theoretical terms, this would require observability of firms’ cost structure (while we assume that firms’ costs are their private information). A model in which bureaucrats target more successful firms would imply that high growth causes higher bribe payments, or some third factor lies behind the positive bribe-growth correlation – in line with our result that non-bribers grow slower, but not consistent with the negative correlation between bribes and growth among bribers that we also document. As we show in the Appendix Table A16, this negative correlation is the case not only for bribe as a share of sales but also for the absolute amount of the bribe in dollars.17 In order to explain the negative correlation between bribes and growth for the firms with non-zero bribe shares, the targeting theory should therefore be supplemented by a countervailing mechanism. For example, one can assume that very high bribes may attract the attention of the bureaucrat’s principal and therefore result in punishment (somewhat similar to the effects of complaints we explicitly consider in our model). This would result in a disincentive for the bureaucrat to demand very high bribe shares from the most successful firms, thus generating a negative correlation between bribe shares and growth. However, this mechanism is unlikely to operate in the most corrupt countries — which is not what we observe in the data (see the Appendix Table A6). 17 Our model does predict a negative relationship between bribe rate b and output y but has no clear prediction regarding the total bribe amount by. The total bribe amount may increase or decrease with bribe rate and output depending on the second derivative of the cost function ci (·). 20 The targeting of firms by bureaucrat may also be driven by other characteristics such as size and profitability. As we discuss in Section 3 and show in Table 1, we find no evidence that larger firms are more likely to pay bribes. However, firms with higher profit margins are. This may be interpreted as evidence of targeting; on the other hand, as we show above, this finding is also consistent with our theory, as firms with higher profits to sales ratios are more likely to accept paying a bribe in our model (see equation (2)). Our theory thus allows for a reconciliation of the positive correlation between profit margins and paying bribes and the negative correlation between bribe share and firm growth. The negative correlation between bribe share and firm growth is also consistent with a “fixed fee” explanation. If operating a business involves a fixed amount of informal payments, firms that grow their sales will pay a smaller fraction of revenue in bribes (Bai et al., 2017 show that the experiences of firms in Vietnam is line with this theory). However, this “fixed fee” model is not consistent with our empirical results on the “zero bribe penalty” and the fact that the latter is only observed in corrupt countries. Another important assumption we make in our model is that the bureaucrat makes the same offer to all firms. In principle, even without observing the firms’ cost structures, bureaucrats could use second degree price discrimination. In Appendix B, we consider a model in which the bureaucrat offers a menu of contracts with different outputs and different bribe levels. We show that such a model predicts a positive correlation between total bribe amount and firm growth. This is not consistent with the data (see Table A16). While using menus should (at least weakly) increase the bureaucrat’s payoff, it requires enforceability of rather complex corruption contracts; our empirical results suggest that the assumption that such contracts are enforceable may not be realistic. 8 Conclusion This paper investigates the relationship between corruption and growth of individual firms using a comprehensive set of enterprise surveys conducted in 141 economies in 2006-20, based on the experiences of almost 88,000 firms. We document several empirical regularities. Firms that make zero informal payments have relatively weak growth in sales when compared to other firms in their industry-region-year cell, as do firms with relatively high rates of informal payments. Furthermore, whereas firms that make high informal payments grow slowly in both high- and low-corruption countries, the zero bribe penalty is driven entirely by firms in high-corruption economies. As we have emphasized throughout, these patterns need not reflect a causal relationship between corruption and growth; however, we also show that they can be rationalized in a simple framework in which profit-maximizing firms confront officials who administer permits 21 that firms require to operate; some agree to pay a bribe and others refuse, with the latter group suffering particularly negative consequences in high-corruption settings where corrupt officials shut bribe-abstainers out of growth opportunities with high probability. We hope that, by introducing these new facts to the literature as well as a framework for organizing them, we may inspire future work to better develop and test microeconomic models of corruption. Our theoretical framework not only helps to reconcile our empirical results but also explains why the zero bribe penalty should not be considered as evidence that corruption is socially beneficial. First, in low-corruption countries, there is no zero bribe penalty, and any increase in corruption involves a reduction in firm growth. Second, in high-corruption countries, the zero bribe penalty reflects the fact that corrupt bureaucrats punish honest firms and facilitate the growth of corrupt firms. Given that we always control for country-region fixed effects, we only study the growth differentials within countries and do not analyze the relationship between corruption and growth at the country level. Our evidence on firms that complain about corruption suggests that the even if bribers grow faster than non-bribers in high-corruption countries, the average growth of firms in such economies may be lower than in low-corruption countries. In addition, our analysis by definition considers growth of existing firms but does not include the impact of bribery on “missing entrants” that could have contributed to economic growth but are dissuaded by corruption. 22