77579 How Important Are Financing Constraints? The Role of Finance in the Business Environment ¨c Meghana Ayyagari, Asli Demirgu ¸ -Kunt, and Vojislav Maksimovic What role does the business environment play in promoting or restraining �rm growth? Recent literature points to a number of factors as obstacles to growth. Inef�cient functioning of �nancial markets, inadequate security and enforcement of property rights, poor provision of infrastructure, inef�cient regulation and taxation, and broader governance features such as corruption and macroeconomic stability are all discussed without any comparative evidence on their ordering. Using �rm-level survey data on the relative importance of different features of the business environ- ment, the article �nds that although �rms report many obstacles to growth, not all the obstacles are equally constraining. Some affect �rm growth only indirectly through their influence on other obstacles, or not at all. Analyses using directed acyclic graph methodology and regressions �nd that only obstacles related to �nance, crime, and policy instability directly affect �rm growth. The �nance result is shown to be the most robust. The results have important implications for the priority of reforms. Maintaining policy stability, keeping crime under control, and undertaking �nancial sector reforms to relax �nancing constraints are likely to be the most effective routes to promote �rm growth. JEL codes: D21, G30, O12 Firm growth is at the center of the development process, making it a much researched area in �nance and economics. The �eld has seen resurgence in interest from policymakers and researchers, with a new focus on the broader business environment in which �rms operate. Through surveys, researchers have documented that �rms report many features of their business environment as obstacles to their growth. Firms report being affected by inadequate security Meghana Ayyagari is an assistant professor in the School of Business at George Washington University; her email address is ayyagari@gwu.edu. Asli Demirgu ¸ -Kunt (corresponding author) is a ¨c senior research manager, Finance and Private Sector Development, in the Development Economics Research Group at the World Bank; her email address is ademirguckunt@worldbank.org. Vojislav Maksimovic is Dean’s Chair Professor of Finance in the Robert H. Smith School of Business at the University of Maryland; his email address is vmaksimovic@rhsmith.umd.edu. The authors would like to thank Gerard Caprio, Rajesh Chakrabarti, Stijn Claessens, Patrick Honohan, Leora Klapper, Aart Kraay, Norman Loayza, David Mckenzie, Dani Rodrik, L. Alan Winters, and seminar participants at the World Bank’s Economist Forum, George Washington University, and the Indian School of Business for their suggestions and comments. THE WORLD BANK ECONOMIC REVIEW, VOL. 22, NO. 3, pp. 483 –516 doi:10.1093/wber/lhn018 Advance Access Publication November 20, 2008 # The Author 2008. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development / THE WORLD BANK. All rights reserved. For permissions, please e-mail: journals.permissions@oxfordjournals.org 483 484 THE WORLD BANK ECONOMIC REVIEW and enforcement of property rights, inef�cient functioning of �nancial markets, poor provision of infrastructure services, inef�cient regulations and taxation, and broader governance features such as corruption and macroeconomic instability. Many of these perceived obstacles are correlated with low �rm performance. These �ndings can inform government policies that shape the opportunities and incentives facing �rms, by influencing their business environment. But even if �rm performance is likely to bene�t from improvements in all dimensions of the business environment, addressing all of them at once would be challenging for any government. Thus, understanding how these different obstacles interact and which ones influence �rm growth directly is important in prioritizing reform efforts. Further, since the relative influence of obstacles may also vary with the level of development of the country and with �rm characteristics such as size, it is important to assess whether the same obstacles affect all subpopu- lations of �rms. This article identi�es the features of the business environment that directly affect �rm growth, using evidence from the World Business Environment Survey (WBES), conducted by the World Bank in 1999 and 2000 in 80 devel- oped and developing economies around the world. These data are used to assess whether each feature of the business environment that �rms report as an obstacle affects their growth, the relative economic importance of the obstacles found to constrain �rm growth, whether an obstacle has a direct effect on �rm growth or acts indirectly by reinforcing other obstacles that have a direct effect, and whether these relationships vary with the level of economic development and �rm characteristics. An obstacle is de�ned as binding if it has a signi�cant impact on �rm growth. Of the 10 business environment obstacles that �rms report, only 3 emerge from the regressions as binding constraints with a direct association with �rm growth: �nance, crime, and policy instability. To reduce the dimensionality of the different business environment factors in a systematic structured approach, directed acyclic graph (DAG) methodology is implemented by an algorithm used in arti�cial intelligence and computer science (Sprites, Glymour, and Scheines 2001). The DAG algorithm also con�rms �nance, crime, and policy instability as the binding constraints, with other obstacles having an indirect association, if at all, with �rm growth through the binding constraints. Further tests �nd �nance to be the most robust, in that the �nance obstacle is binding regardless of which countries and �rms are included in the sample. Regression analysis also shows that �nance has the largest direct effect on �rm growth. These results are not due to influential observations, reverse causality, or perception biases likely to be found in survey responses. Policy instability and crime, the other two binding constraints in the full sample, are driven by the inclusion of transition and African economies where, arguably, they might be the most problematic. Instrumental variable regressions also show �nance to be the most robust result. ¨c Ayyagari, Demirgu ¸ -Kunt, and Maksimovic 485 The relative importance of different factors is found to vary according to �rm characteristics. Larger �rms are signi�cantly less affected by the �nance obstacle, but being larger does not relax the obstacles related to crime or policy instability to the same extent. Although �rms identify many speci�c �nancing obstacles such as collateral requirements and lack of access to long-term capital, only the cost of borrow- ing is directly associated with �rm growth. But the cost of borrowing is itself affected by imperfections in �nancial markets. Firms that face high interest rates also perceive that the banks to which they have access are corrupt, under- funded, and require excessive paperwork. Dif�culties with posting collateral and limited access to long-term �nancing are also correlated with high interest rates. These obstacles are also likely to be aggravated by underdeveloped institutions.1 Several studies point to the importance of �nancing obstacles. Using �rm- level data, Demirgu ¸ -Kunt and Maksimovic (1998) and others provide evi- ¨c dence on how the �nancial system and legal enforcement relax �rms’ external �nancing constraints and facilitate their growth. Rajan and Zingales (1998) show that industries that depend on external �nance grow faster in countries with better developed �nancial systems.2 Although these studies investigate different obstacles to �rm growth and their impact, they generally focus on a small subset of broadly characterized obstacles. The current study is most closely related to Beck, Demirgu ¸ -Kunt, and ¨c Maksimovic (2005) but differs signi�cantly from that study in the question being asked, the execution, and the �ndings. Beck, Demirgu ¨c¸ -Kunt, and Maksimovic examine whether three obstacles (�nance, corruption, and legal obstacles) selected on a priori grounds individually influence �rm growth rates; they do not compare the obstacles to identify the most binding constraint. This is crucial since, as the current study shows, most obstacles when entered individually are signi�cant in growth regressions. The current study also differs in methodology, since Beck, Demirgu ¸ -Kunt, and Maksimovic do not incorporate country-�xed ¨c effects (or the DAG methodology) and have limited discussion of causality. The current study looks at the full set of business environment obstacles— �nance, corruption, infrastructure, taxes and regulations, judicial ef�ciency, crime, anticompetitive practices, policy instability and uncertainty, inflation, 1. Fleisig (1996) highlights the problem with posting collateral in developing and transition economies with the example of �nancing available to Uruguayan farmers raising cattle. While cattle are viewed as one of the best forms of loan collateral in the United States, a pledge on cattle is worthless in Uruguay. Uruguayan law requires speci�c description of the pledged property, in this case, identi�cation of the pledged cows. The need to identify collateral so speci�cally undermines the secured transaction, since the bank is not allowed to repossess a different group of cows in the event of nonpayment. 2. Here is a parallel literature on �nancial development and growth at the country level. Speci�cally, cross-country studies (King and Levine 1993; Beck, Levine, and Loayza 2000; Levine, Loayza, and Beck 2000) show that �nancial development fosters economic growth. See Levine (2005) for a review of the �nance and growth literature. 486 THE WORLD BANK ECONOMIC REVIEW and exchange rate—and �nds �nance, crime, and policy instability to be the most binding and �nancial to be the most robust. Thus this study has impli- cations for the priority of reform efforts, while the study by Beck, Demirgu ¨c¸ -Kunt, and Maksimovic does not. Several quali�cations need to be emphasized. First, as is common in the lit- erature, the current study takes as given the existing population of �rms in each country and studies the constraints they face. But, as described by Hausman, Rodrik, and Velasco (2008), it must be noted that in a more general setting the population of �rms is itself endogenous. For example, Beck, Demirgu ¸ -Kunt, and Maksimovic (2006) show that �rm size distri- ¨c bution adapts to the business environment, and Demirgu ¸ -Kunt, Love, and ¨c Maksimovic (2006) show that certain organizational forms are better adapted to speci�c business environments. Nevertheless, the analysis in this article can be seen as a way of identifying and targeting the most binding constraints for existing �rms, conditional on having entered, but not necessarily as a way of identifying the constraints to entry. Second, this article examines cross-country �rm-level regressions and therefore does not detail the experience of any single country in depth. But controlling for country-�xed effects provides useful—although not de�nitive—information from the cross-country set-up on the binding constraints to �rm growth. Finally, in the absence of panel data and �rm-�xed effects, potential reverse causality concerns are endemic to the growth literature. These issues are addressed in detail using instrumental variables. The article is organized as follows. Section I describes the methodology. Section II discusses the data and summary statistics. Section III presents the main results. Section IV presents conclusions and policy implications. I . M E T H O D O LO GY : I D E N T I F I CAT I O N OF BINDING CONSTRAINTS Numerous studies argue that differences in the business environment can explain much of the variation across countries in �rms’ �nancial policies and performance. While much of the early work relied on country-level indicators and �rms’ �nancial reports, more recent work has relied on surveys of �rms, which provide data on a wide range of potential obstacles to growth.3 Surveys have identi�ed a large number of potential obstacles to growth, making it dif�cult to identify the obstacles that are truly constraining. Enterprise managers may identify several operational issues, not all of them constraining. Therefore, it is necessary to identify the extent to which reported obstacles affect the growth rates of �rms. An obstacle is to be considered a “constraint� or a “binding constraint� only if it has a signi�cant impact on �rm growth. Signi�cant impact requires that the coef�cient of the obstacle in 3. See Dollar, Hallward-Driemeier, and Mengistae (2005), Gelb and others (2007), Carlin, Schaffer, and Seabright (2005), and Svejnar and Commander (2007). ¨c Ayyagari, Demirgu ¸ -Kunt, and Maksimovic 487 the �rm growth regression be signi�cant and that the enterprise managers identi�ed the factor as an obstacle.4 To the extent that the characteristics of a �rm’s business environment are correlated, it is likely that many perceived business environment characteristics will be correlated with realized �rm growth. It is important to sort these into obstacles that directly affect growth and those that may be correlated with �rm growth but affect it only indirectly. Since there is no theoretical basis for classifying the obstacles, empirical measures are required. The DAG methodology is used to reduce dimensionality in a structured way. The DAG algorithm begins with a set of potentially related variables and uses the conditional correlations between them to rule out possible relations among them. The �nal output of the algorithm is a pattern of graphs listing potential relations among the variables that have not been ruled out, which shows variables that have direct effects on the dependent variable or other variables, variables that have only indirect effects on the dependent variable through other variables, and variables that lack a consistent statistical relation with the other variables. If DAG identi�es an obstacle as having a direct effect on �rm growth, that obstacle would also have a signi�- cant coef�cient in all ordinary least squares regressions regardless of which subset of other obstacles is entered as control variables in the regression equation. Ayyagari, Demirgu ¸ -Kunt, and Maksimovic (2005) further illustrate ¨c the use of this methodology.5 Regression analysis is also used for further robustness tests, such as testing for possible endogeneity bias using instrumental variable methods and controlling for additional variables at the �rm and country level, growth opportunities, influential observations, and potential perception biases in survey responses. While the obstacles a �rm faces depend on the institutions in each country, the obstacles are not likely to be the same for each �rm in each country. Thus, the unit of analysis is the �rm. As described in what follows, the regressions have country-level �xed effects. 4. In the survey, managers indicate that an obstacle is a problem by assigning it a value of 1 to 4. The signi�cance of the coef�cient in the growth regression is usually suf�cient to determine whether an obstacle is binding since the mean value of all obstacles exceeds 1. But in determining the relative impact, it is important to take into account the level of the obstacles. 5. DAG analysis is related to the use of different analytical methods to identify the most reliable predictors of economic growth such as the extreme bounds analysis (EBA) used in Kormendi and Meguire (1985), Barro (1991), and Levine and Renelt (1992), and the technique in Sala-i-Martin (1997). DAG analysis has several advantages over these methods. While these methods start from an equation speci�ed by the researcher that embodies a causal ordering that is then tested, DAG can endogenously discover the causal ordering. Moreover, whereas EBA treats one relation at a time, the graphs produced by DAG show robust relations among all the variables being analyzed, taking into account the implications of robust relations elsewhere in the system on the ordering in a speci�c relation. 488 THE WORLD BANK ECONOMIC REVIEW I I . D ATA AND S U M M A RY S TAT I S T I C S As the main purpose of the WBES is to identify obstacles to �rm performance and growth around the world, it contains many questions on the nature and severity of different obstacles. Speci�cally, �rms are asked to rate the extent to which �nance, corruption, infrastructure, taxes and regulations, judicial ef�- ciency, crime,6 anticompetitive practices, policy instability and uncertainty, and macro issues such as inflation and exchange rate constitute obstacles to their growth. In addition to the detail on obstacles to growth, one of the great values of this survey is its wide coverage of smaller �rms. The survey is size-strati�ed, with 40 percent on observations on small �rms (de�ned as employing 5 –50 employees), 40 percent on medium-size �rms (51–500 employees), and the remainder from large �rms (more than 500 employees). The �rm-level obstacles are reported in table 1. The WBES asked enterprise managers to rate each factor as an obstacle to the operation and growth of their business on a scale of 1–4, with 1 denoting no obstacle; 2, a minor obstacle; 3, a moderate obstacle; and 4, a major obstacle. Firms in high-income countries tend to face lower obstacles in all areas ( panel A of table 1). In the sample of developing economies, regional analysis indicates that African �rms report corruption and infrastructure as the highest obstacles, Latin American �rms report crime and judicial ef�ciency as the highest obstacles, and Asian countries report �nancing as the lowest obstacle ( panel B). Smaller �rms face higher obstacles than larger �rms in all areas except in those related to judicial ef�ciency and infrastructure, where the ranking is reversed ( panel C). Firm sales growth over the past three years is used as a measure of �rm per- formance. Sales growth is used rather than productivity because productivity measures are noisier and available for a much smaller sample of �rms. Information on other performance measures such as pro�ts was not available. Appendix table A-1 reports �rm growth and the obstacles �rms report, aver- aged over all sampled �rms in each country. Average �rm growth across countries shows a wide dispersion, from negative rates of 20 percent for Armenia and Azerbaijan to 64 percent for Malawi and Uzbekistan. Firms report taxes and regulations to be their greatest obstacles. Inflation, policy instability, and �nancing obstacles are also reported to be highly constraining. In contrast, factors associated with judicial ef�ciency and infrastructure are ranked as the lowest obstacles faced by entrepreneurs. The correlations among the obstacles reported by �rms are signi�cant but fairly low, with few above 0.5 (correlation matrix not shown). As expected, the two macro obstacles, inflation and exchange rate, are highly correlated, at 6. The survey includes two obstacles on crime, one capturing street crime and the other organized crime. Since the correlation between the two obstacles is higher than 70 percent, only street crime, which is more strongly correlated with �rm growth, is used in the analysis. T A B L E 1 . Economic Indicators and General Obstacles General obstacles Taxes GDP per Firm Policy Exchange Judicial Street and Anticompetitive Classi�cation capita growth Financing instability Inflation rate ef�ciency crime Corruption regulation behavior Infrastructure A: Averaged across country income groups a High (N ¼ 11) 21,376.34 0.19 2.19 2.2 2.04 1.93 1.81 1.71 1.59 2.67 2 1.72 Upper middle 4,131.817 0.19 2.75 2.62 2.54 2.27 2.13 2.38 2.29 2.93 2.18 1.99 (N ¼ 18) Lower middle 1,984.852 0.11 3 3.14 3.1 2.94 2.31 2.72 2.73 3.24 2.59 2.31 (N ¼ 26) Low income 435.3 0.14 2.85 2.84 3.02 2.61 2.15 2.78 2.98 2.73 2.53 2.7 (N ¼ 25) B: Averaged across geographic regions Europe and North 22,863.72 0.19 2.2 2.22 2.06 1.89 1.79 1.78 1.63 2.77 1.98 1.76 America (N ¼ 9) Latin America 3,022.2 0.09 2.83 3.02 2.84 2.8 2.39 2.95 2.74 3.01 2.43 2.4 (N ¼ 20) Asia (N ¼ 10) 2,772.52 0.05 2.59 2.82 2.74 2.66 1.99 2.62 2.71 2.51 2.44 2.43 Transition (N ¼ 23) 2,417.02 0.19 3.05 2.99 3.06 2.7 2.17 2.39 2.5 3.28 2.44 2.09 Ayyagari, Demirgu ¨c Africa (N ¼ 18) 1,115.81 0.23 2.77 2.43 2.75 2.21 2.64 2.80 2.32 2.75 C: Averaged across �rm size groups Small 3,759.33 0.13 2.89 2.84 2.90 2.59 2.13 2.64 2.62 2.94 2.43 2.24 Medium 4,377.98 0.16 2.86 2.87 2.84 2.60 2.18 2.46 2.53 3.00 2.41 2.26 Large 4,365.68 0.17 2.54 2.75 2.65 2.55 2.19 2.49 2.43 2.70 2.23 2.36 Note: The variables are described as follows: GDP per capita is real GDP per capita in U.S. dollars averaged over 1995– 99. Firm growth is the percen- tage change in �rm sales over the past three years (1996 – 99). Financing, policy instability, inflation, exchange rate, judicial ef�ciency, street crime, corrup- tion, taxes and regulation, anticompetitive behavior, and infrastructure are general obstacles as indicated in the �rm questionnaire. They take values of 1 – ¸ -Kunt, and Maksimovic 4, where 1 indicates no obstacle and 4 indicates a major obstacle. In panels A, B, and C, �rm variables are averaged over all �rms in the speci�ed group. a Income groups are de�ned according to World Bank (2005). Source: Authors’ analysis based on WBES data described in text. 489 490 THE WORLD BANK ECONOMIC REVIEW 0.58. The correlations of corruption with crime and judicial ef�ciency are also relatively high, at 0.55 each, indicating that in environments where corruption and crime are widespread, judicial ef�ciency is adversely affected. The corre- lation between the �nancing obstacle and all other obstacles is among the lowest, indicating that the �nancing obstacle may capture different effects than those captured by other reported obstacles. All obstacles are negatively and sig- ni�cantly correlated with �rm growth. These relations are explored further in the next section. III. FIRM GROWTH AND R E PO R T E D OB S TAC L E S This section explores the link between the obstacles that �rms report and �rm growth rates using country-�xed effect regressions and DAG analysis. It �nds that �nance, crime, and policy instability are most signi�cantly associated with �rm growth, suggesting that these are the binding constraints. The results are robust to a number of checks, including variation across different �rm sizes and country income levels, endogeneity concerns, removal of outliers, and per- ception biases. Of the individual �nancing obstacles, high interest rates are found to be most signi�cantly associated with �rm growth. Obtaining the Binding Constraints Firm growth rates are regressed on the different obstacles �rms report. All regressions are estimated with �rm-level data using country-level �xed effects.7 The standard errors are adjusted for clustering at the country level. Speci�cally, the regression equations take the form: Firm growth ¼ a þ b1  obstacle þ b2  firm size þ country-fixed effects þ 1: ð1Þ The hypothesis that a reported obstacle is a binding constraint (has a signi�- cant impact on �rm growth) is tested by determining whether b1 is signi�cantly different from 0. Signi�cant impact also requires that the obstacle has a value higher than 1, which is true for all obstacles. When individual obstacles are analyzed separately, all but corruption, exchange rate, anticompetitive behavior, and infrastructure are signi�cantly related to �rm growth (table 2). The regressions explain up to 7.4 percent of the variation in �rm growth across countries. The coef�cients of the signi�cant obstacles range from 0.021 for the judicial ef�ciency obstacle to 0.032 for the 7. In unreported regressions, the robustness of the results was also checked by including additional control variables in the regression. Speci�cally, adding variables at the �rm level to capture a �rm’s industry, number of competitors, organizational structure, and whether it is government or foreign owned, an exporter, or a subsidy receiver reduces country coverage from 80 to 56 but does not signi�cantly affect the results for individual obstacles. Of the three binding constraints identi�ed earlier, only the policy instability obstacle loses signi�cance. Results are similar with country random effects controlling for GDP per capita and inflation at the country level. T A B L E 2 . Impact of Obstacles on Firm Growth Variable 1 2 3 4 5 6 7 8 9 10 11 12 Constant 0.205*** 0.165*** 0.193*** 0.170*** 0.180*** 0.140*** 0.152*** 0.117*** 0.111*** 0.126*** 0.332*** 0.297*** (0.028) (0.036) (0.034) (0.029) (0.040) (0.026) (0.032) (0.029) (0.028) (0.033) (0.059) (0.047) Size 0.003 0.005** 0.004 0.004 0.005* 0.005* 0.005* 0.003 0.005* 0.005 0.004 0.004 (0.002) (0.003) (0.002) (0.002) (0.003) (0.003) (0.003) (0.002) (0.003) (0.003) (0.003) (0.002) Financing 2 0.032*** 2 0.034*** 2 0.028*** (0.008) (0.009) (0.008) Policy instability 2 0.024*** 2 0.022* 2 0.014 (0.010) (0.013) (0.009) Street crime 2 0.030*** 2 0.033** 2 0.025* (0.013) (0.015) (0.014) Inflation 2 0.020** 2 0.002 (0.009) (0.011) Taxes and 2 0.027** 0.001 regulation (0.012) (0.013) Judicial 2 0.021** 2 0.003 ef�ciency (0.010) (0.009) Ayyagari, Demirgu Corruption 2 0.017 0.011 ¨c (0.011) (0.012) Exchange rates 2 0.000 (0.009) Anticompetitive 2 0.004 behavior (0.007) Infrastructure 2 0.009 (0.008) (Continued ) 491 ¸ -Kunt, and Maksimovic 492 TABLE 2. Continued Variable 1 2 3 4 5 6 7 8 9 10 11 12 Number of 6,235 6,133 5,964 6,175 6,343 5,142 5,620 6,068 5,091 6,205 4,551 5,778 �rms Number of 79 79 79 79 79 61 78 79 60 79 59 78 countries Adjusted R 2 0.07 0.073 0.07 0.068 0.069 0.07 0.072 0.069 0.069 0.068 0.074 0.072 THE WORLD BANK ECONOMIC REVIEW *Signi�cant at the 10 percent level; **signi�cant at the 5 percent level; ***signi�cant at the 1 percent level. Note: Numbers in parentheses are standard errors clustered at the country level. The regression equation estimated is �rm growth ¼ a þ b1  size þ b2  �nancing þ b3  policy instability þ b4  inflation þ b5  exchange rates þ b6  judicial ef�ciency þ b7  street crime þ b8  corruption þ b9  taxes and regulation þ b10  anticompetitive behavior þ b11  infrastructure þ b12  country-�xed effects þ 1. The variables are described as follows: �rm growth is the percentage increase in �rm sales over the past three years. Firm size is the log of �rm sales. Financing, policy instability, inflation, exchange rate, judicial ef�ciency, street crime, corruption, taxes and regulation, anticompetitive behavior, and infrastructure are general obstacles as indi- cated in the �rm questionnaire. They take values of 1– 4, where 1 indicates no obstacle and 4 indicates a major obstacle. In speci�cations 1 – 10, each of the obstacle variables is included individually. Speci�cation 11 includes all the obstacles that were signi�cant in speci�cations 1 – 10; speci�cation 12 includes only �nancing, policy instability and street crime obstacles. All regressions in speci�cations 1 – 12 are estimated using country-�xed effects with clustered standard errors. Source: Authors’ analysis based on WBES data described in text. ¨c Ayyagari, Demirgu ¸ -Kunt, and Maksimovic 493 �nance obstacle. Thus, for instance �rms that say �nancing is a minor obstacle grow 3.2 percent slower than those that say �nance is not an obstacle. Alternatively, a one-standard deviation increase in the �nancing obstacle decreases the �rm growth rate by 3.6 percent. Column 11 of table 2 includes all the signi�cant obstacles in the regression equation. In this speci�cation, only the �nance, policy instability, and crime obstacles have a signi�cant constraining effect on growth. Dropping the remaining obstacles from the regression (which are jointly insigni�cant as well), as in speci�cation 12, shows only �nance and crime as having a con- straining effect on growth. The economic impact of the �nance obstacle is higher than that of crime, but the difference is not statistically signi�cant. It is also possible to do such impact evaluation at the regional, country, or �rm level, instead of at the sample mean. Looking at the mean obstacles for individual countries reported in the appendix table A-1, it is clear that the binding obstacles are not equally important in every country. For example, in Singapore, where the mean value of the binding obstacles is all close to one, the economic impact of the obstacles is much smaller than in Nigeria, where the mean value of all three obstacles is more than 3, indicating severe con- straints. Thus, it is possible to use these cross-country results to do growth diagnostics at the country level as discussed in Hausmann, Rodrik, and Velasco (2008). Looking more closely at the �rm level, there may be some �rms in Nigeria for which the constraints are not binding (depending on the value of the obstacles they report) and some in Singapore for which they are. In fact, average values of obstacles by �rm size, as shown in table 1, suggest that the three obstacles will always be more binding for smaller �rms than for larger �rms. Overall, these results suggest that the three obstacles—�nance, crime, and policy instability—are the only true constraints, in that they are the only obstacles that affect �rm growth directly at the margin. The other obstacles may also affect �rm growth through their impact on each other and on the three binding constraints, but they have no direct effect on �rm growth. Have the Key Constraints Been Identi�ed? Robustness Checks The DAG methodology is used to check the robustness of the regression �nd- ings since DAG is useful in simplifying the set of independent variables in a systematic way, as described in Ayyagari, Demirgu ¨c¸ -Kunt, and Maksimovic (2005). The DAG analysis is implemented using the software program TETRAD III (Scheines and others 1994). In keeping with common practice, the business environment obstacles are assumed to cause �rm growth, not the other way around, and the model is assumed to contain all common causes of the vari- ables in the model. To be consistent with the �xed effects speci�cation in table 2, demeaned values of the business environment obstacles are used, 494 THE WORLD BANK ECONOMIC REVIEW F I G U R E 1. DAG Analysis of the General Obstacles to Firm Growth Source: Authors’ analysis based on WBES data described in text. where the country average of each obstacle is subtracted from the correspond- ing obstacle. Figure 1 illustrates the application of this algorithm to the full sample. The input to the algorithm is the correlation matrix between �rm growth and the 10 demeaned business environment obstacles from the sample of 4,197 �rms.8 Figure 1 shows that the only business environment obstacles that have a direct effect on �rm growth are �nancing, crime, and policy instability. Financing in turn is directly affected by the taxes and regulation obstacle, which include factors such as taxes and tax administration, and regulations in the areas of business licensing, labor, foreign exchange, environment, �re, and safety. Crime is directly affected by the corruption obstacle, and policy instabil- ity is affected by corruption, infrastructure, and anticompetitive behavior.9 The 8. In addition, the signi�cance level was selected for the tests of conditional independence performed by TETRAD. Because the algorithm performs a complex sequence of statistical tests, each at the given signi�cance level, the signi�cance level is not an indication of error probabilities of the entire procedure. Spirtes, Glymour, and Sheines (2001, p. 116), after exploring several versions of the algorithm on simulated data, conclude that “in order for the method to converge to correct decisions with probability 1, the signi�cance level used in making decisions should decrease as the sample size increases, and the use of higher signi�cance levels may improve performance at small sample sizes.� For the results in this article obtained from samples ranging from 2,659–4,197 observations, a signi�cance level of 0.10 was used. At the 5 percent signi�cance level, �nance, crime, and policy instability have a direct effect on �rm growth, whereas at the 1 percent level only �nance and crime have a direct effect on growth. 9. The DAG analysis and the set of causal structures determined by the algorithm are useful for an objective selection of variables, with the heuristic interpretation that if DAG analysis shows that obstacle X causes obstacle Y, then �rms’ reports of X as an obstacle are also likely to affect the probability that they report Y as an obstacle. For details refer to formal de�nitions. ¨c Ayyagari, Demirgu ¸ -Kunt, and Maksimovic 495 dashed double-headed arrows between policy instability and crime, inflation, taxes and regulation, and judicial ef�ciency indicate that the direction of orien- tation between policy instability and these variables changes between patterns. The output also shows that the relations between the obstacles themselves are quite complex and that there are multiple relations in the DAG among the business environment obstacles.10 Since the main focus of this article is to identify the business environment obstacles with a direct effect on growth, the interactions among the different variables are left for future work. Hence, rather than focusing on the farthest variables in the �gure, which are indirectly related to �rm growth and are thus likely to have a very diluted impact on �rm growth, we focus on the variables with direct effects, which are likely to have the biggest impact on growth. Most important, the DAG analysis also identi�es �nancing, crime, and policy stability as the only variables having direct effects on �rm growth, as suggested by speci�cation 11 of table 2. As discussed in section II, the analysis identi�es direct effects after conditioning on all subsets of the other variables. This suggests that in regression analysis, �nancing, crime, and policy instability will always have signi�cant coef�cients irrespective of the subsets of other obstacles included in the regression. Thus, these are binding constraints, and policies that relax these constraints can be expected to directly increase �rm growth. Binding Constraints and Firm Size and Level of Development This section explores whether these relationships are different for �rms of different sizes and at different levels of development. The �rst three columns of table 3 include speci�cations that interact the three obstacles with �rm size, given by the logarithm of sales. The interaction term with the �nancing obstacle is positive and signi�cant at the 1 percent level, suggesting that larger �rms are less �nancially constrained, con�rming the �ndings of Beck, Demirgu ¸ -Kunt, and Maksimovic (2005). The interaction terms with policy ¨c instability and crime are also positive but not signi�cant. When all the inter- actions are entered together in speci�cation 4, only the interaction term with the �nancing obstacle is signi�cant. Thus, although there is also some indi- cation that large �rms are also affected less by crime and policy instability, this evidence is much weaker. The three obstacles are also interacted with dummy variables for country income—upper middle income, lower middle income, and low income. The excluded category is high income. The results indicate that all three obstacles tend to be more constraining for middle-income countries. This �nding suggests that middle-income countries, having overcome country-speci�c 10. In addition to the directed arrows and bidirectional arrows, �gure 1 also shows that in some cases common latent causes drive associations between some variables (such as �nancing and corruption) and that in other cases the direction of orientation is inconsistent: some statistical tests indicate that an edge should be oriented as x1 ! x2, and other statistical tests indicate that it should be oriented as x1 x2. 496 T A B L E 3 . Firm Growth Interaction Effects Interaction with �rm size Interaction with country income dummy variables Variable 1 2 3 4 1 2 3 4 Constant 0.278*** 0.218*** 0.225*** 0.421*** 0.207*** 0.177*** 0.184*** 0.299*** (0.050) (0.061) (0.058) (0.089) (0.029) (0.039) (0.030) (0.046) Firm size 2 0.004 (0.004) 2 0.000 2 0.000 2 0.009 (0.006) 0.004 (0.002) 0.005* (0.003) 0.004 (0.002) 0.004 (0.002) (0.004) (0.004) Financing 2 0.058*** 2 0.053*** 2 0.002 2 0.004 (0.016) (0.015) (0.013) (0.015) Financing  Size 0.003*** 0.003** THE WORLD BANK ECONOMIC REVIEW (0.001) (0.001) Financing  Upper 2 0.041* 2 0.034 middle (0.023) (0.022) Financing  Lower 2 0.041** 2 0.027 middle (0.019) (0.019) Financing  Low 2 0.016 2 0.019 income (0.019) (0.022) Policy instability 2 0.042** 2 0.024 (0.019) 0.008 (0.012) 0.014 (0.012) (0.020) Policy 0.002 (0.001) 0.001 (0.001) instability  Size Policy instability  2 0.056*** 2 0.045** Upper middle (0.021) (0.018) Policy instability  2 0.055** 2 0.043* Lower middle (0.024) (0.025) Policy instability  0.005 (0.019) 2 0.008 Low income (0.017) Street crime 2 0.042* 2 0.034 (0.025) 2 0.010 2 0.014 (0.024) (0.014) (0.014) Street crime  Size 0.001 (0.001) 0.001 (0.002) Street crime  2 0.021 2 0.010 Upper middle (0.026) (0.025) Street crime  2 0.052** 2 0.039 Lower middle (0.025) (0.027) Street crime  0.039* 0.044** Low income (0.021) (0.020) Number of �rms 6,235 6,133 5,964 5,778 6,235 6,133 5,964 5,778 Number of countries 79 79 79 78 79 79 79 78 Adjusted R 2 0.071 0.074 0.071 0.074 0.070 0.075 0.073 0.075 F-test of interactions 0.0503 0.1184 0.0088 0.0039 0.0022 *Signi�cant at the 10 percent level; **signi�cant at the 5 percent level; ***signi�cant at the 1 percent level. Note: Numbers in parentheses are standard errors clustered at the country level. The regression equation estimated is �rm growth ¼ a þ b1  size þ b2  �nancing þ b3  policy instability þ b4  street crime þ b5  �nancing  income dummy variables þ b6  �nancing  size þ b7  policy instability  income dummy variables þ b8  policy instability  size þ b9  street crime  income dummy variables þ b10  street crime  size. The variables are described as follows: �rm growth is the percentage increase in �rm sales over the past three years. Firm size is the log of sales. Financing, Ayyagari, Demirgu policy instability, and street crime are general obstacles as indicated in the �rm questionnaire. They take values 1 – 4, where 1 indicates no obstacle and 4 ¨c indicates a major obstacle. Income dummy variables are country dummy variables based on the income level of the country. High-income dummy vari- able takes the value of 1 for countries belonging to the high-income group and 0 otherwise, upper middle-income dummy variable takes the value of 1 for countries belonging to the upper middle-income group and 0 otherwise, lower middle-income dummy variable takes the value of 1 for countries belonging to the lower middle-income group and 0 other wise, low-income dummy variable takes the value of 1 for low-income group countries and 0 otherwise. In speci�cations 1 – 3 in each panel, the obstacle variables and its interactions are included individually. Speci�cation 4 in both panels includes the full model. All regressions are estimated using country-�xed effects with clustered standard errors. Each speci�cation also reports the p-value of the joint signi�cance test of the interaction terms. Source: Authors’ analysis based on WBES data described in text. 497 ¸ -Kunt, and Maksimovic 498 THE WORLD BANK ECONOMIC REVIEW institutional obstacles, are now more constrained by a common set of obstacles pertaining to �nance, crime, and policy instability. This is consistent with Gelb and others (2007), who �nd that �rms’ levels of complaints about different obstacles vary with the income level of the countries. The F-tests for the hypotheses that all the entered interactions are jointly equal to 0 are rejected at the 1 percent level of signi�cance for the crime and policy instability obstacles but not for the �nancing obstacle. This suggests that �rms in countries in all income groups are similarly affected by the �nancing obstacle. Checking for Reverse Causality While �nancing, crime, and policy instability have been identi�ed as �rst-order constraints, signi�cantly affecting �rm growth, it is possible that the relations observed may also be due to reverse causality, with inef�cient, slow growing �rms blaming the environment for their performance. But while reverse causal- ity is potentially a concern, it does not explain why poorly performing �rms would systematically complain most about �nancing, crime, and policy instability and not about the other obstacles. While there might be a causal relation between poor performance and availability of �nancing, examined in what follows using instrumental variables, it is harder to posit a causal relation between poor performance and crime and policy instability. The approach recommended by Carlin, Schaffer, and Seabright (2005) is used to check for reverse causality for the street crime and policy instability obstacles. They compare the coef�cients of the �xed effects “within-estimator� and “between-estimator� and test for sign changes, arguing that since reverse causality is more likely to be signi�cant at the �rm level, it will cause the within-estimator and the between-estimator to change signs.11 When the �xed effects model is run using the within-estimator, the obstacle coef�cients are negative when entered individually. None of the coef�cients are perversely positive, which might have suggested reverse causality. The between-estimator also shows the obstacle coef�cients to be negative. Furthermore, as seen in table 1, some factors such as taxes and regulation are rated as very high obstacles compared with others but do not appear as binding constraints, whereas street crime is not rated very highly (except in Latin America) yet still emerges as a binding constraint. This suggests that �rms may complain about many factors when surveyed but controls are needed for country differences and �rm heterogeneity to identify the obstacles with the largest association with �rm growth. To assess the robustness of the results, instrumental variable regressions (limited information maximum likelihood estimators) are used to extract the exogenous component of the three obstacles. Two sets of instruments are used 11. Carlin, Schaffer, and Seabright (2005) argue that only in the case of the �nancing constraint, reverse causality makes the within-coef�cient more negative than the true value, thus making this method inapplicable. ¨c Ayyagari, Demirgu ¸ -Kunt, and Maksimovic 499 for �nancing, crime, and policy instability. The �rst is the average value of the obstacles for the industry groups in each country. While it is likely that individ- ual �rms may blame the obstacles for their poor performance, it is less likely that all �rms in a given country-industry group will engage in such blame shift- ing. Instrumenting the obstacles with the average obstacle for each industry group in the country isolates the exogenous part of the possibly endogenous obstacle the �rm reports, which can be used to predict growth. When the obstacles are considered at the country-industry level of aggregation, causality is likely to run from the average obstacles to individual �rms, not vice versa. In addition, country-industry averages also help with potential measurement errors that are largely idiosyncratic to the �rm and hence uncorrelated with the average values of the obstacles.12 The second set of instruments is �rm responses to the survey question: Does your �rm use international accounting standards? A �rm’s adoption of international accounting standards is likely to influence its business environment constraints, in particular the �nancing con- straint, but is not necessarily independently linked to �rm growth rates. The analysis is also conducted at the country level, averaging the obstacle variables and �rm growth rates across countries and controlling for log GDP per capita rather than for any of the �rm-level variables. The instruments for �nancing and policy instability obstacles are a “Common law� dummy vari- able, which takes a value of 1 if the country follows common law tradition, and three religion variables, Protestant, Muslim, and Catholic, which represent the percentages of the population that are Protestant, Muslim, or Catholic in each country. The instrument for street crime is the common law dummy vari- able and the “latitude� of a country’s capital city. An extensive literature has identi�ed these institutional variables as good instruments for institutional development, and hence they are not used as explanatory variables in the short- term growth regressions in the second stage. When country-industry averages of the obstacles are used as instruments, only the �nancing obstacle is negative and signi�cant (table 4, columns 1 –3). The �rst stage F-statistic is large, indicating that the country-industry average of the �nancing obstacle is a good instrument.13 While the country-industry averages pass the instruments test for policy instability and street crime, these obstacles are now insigni�cant in the regression. In addition, when all three obstacles are implemented together, �nancing is again the only signi�cant con- straint (column 4). This reinforces the �nding that �nancing is the most robust of the three binding constraints. When �rms’ adoption of international accounting standards is used as an instrument, all three obstacles have a signi�cant negative impact on �rm 12. Use of group averages as instruments is a common technique, as used in Fisman and Svensson (2007) and described in Krueger and Angrist (2001). 13. This is further con�rmed by the weak identi�cation test statistic (Kleibergen-Paap Wald statistic), which is much larger than the critical value of 16.38 500 T A B L E 4 . Robustness Test—Instrumental Variables, Firm-level Regressions Country-industry average of the obstacle variable Does the �rm follow international accounting standards? Instrument 1 2 3 4 5 6 7 Size 0.002 (0.002) 0.006*** (0.002) 0.004* (0.002) 0.003 (0.002) 2 0.004 (0.004) 0.011* (0.006) 2 0.005 (0.005) Financing 2 0.066*** (0.025) 2 0.067** (0.028) 2 0.285*** (0.101) Policy instability 2 0.045 (0.029) 2 0.041 (0.031) 2 0.897* (0.499) Street crime 2 0.011 (0.029) 0.014 (0.032) 2 0.529** (0.232) Number of �rms 6,235 6,133 5,964 5,778 5,846 5,747 5,592 First-stage test of excluded instruments F-statistic (�nancing) 382.32 (0.000) 112.13 (0.000) 36.48 (0.000) F-statistic (policy 334.57 (0.000) 106.44 (0.000) 4.66 (0.031) instability) THE WORLD BANK ECONOMIC REVIEW F-statistic (crime) 351.30 (0.000) 110.22 (0.000) 11.11 (0.001) Underidenti�cation 549.12 (0.000) 405.91 (0.000) 453.67 (0.000) 366.18 (0.000) 35.90 (0.000) 4.54 (0.033) 11.20 (0.001) test—Kleibergen-Paap rk Wald statistic Weak instrument robust 7.06 (0.008) 2.41 (0.121) 0.14 (0.704) 3.55 (0.014) 9.43 (0.002) 9.82 (0.002) 9.18 (0.002) inference—Anderson Rubin Wald test *Signi�cant at the 10 percent level; **signi�cant at the 5 percent level; ***signi�cant at the 1 percent level. Note: Two-stage instrumental variable regressions are used. Numbers in parentheses are standard errors clustered at the country level. The �rst-stage regression equation estimated is �nancing (or policy instability or street crime) ¼ a þ g1  country-�xed effects þ g2  �rm size þ g3  instrument. The second-stage regression equation estimated is �rm growth ¼ a þ b1  country-�xed effects þ b2  �rm size þ b3  �nancing ( predicted value from �rst stage) þ b4  policy instability ( predicted value from �rst stage) þ b5  street crime ( predicted value from �rst stage). In speci�cations 1 – 4, the instru- ment used is the average value of the obstacle across each industry in each country. In speci�cations 5 – 7, the instrument used is �rm response to the vari- able, “Does the �rm adopt international accounting standards?� The variables are described as follows: �rm growth is the percentage increase in �rm sales over the past three years. Firm size is the log of sales. Financing, policy instability, and street crime are general obstacles as indicated in the �rm ques- tionnaire. They take values 1 – 4, where 1 indicates no obstacle and 4 indicates a major obstacle. Source: Authors’ analysis based on WBES data described in text. ¨c Ayyagari, Demirgu ¸ -Kunt, and Maksimovic 501 growth. While the �rst-stage F-statistic is signi�cant in each case, it is greater than 10 only for the �nancing and crime obstacles (Stock and Watson 2003 rule of thumb for good instruments). But the Anderson Rubin Wald test, which is the preferred test for robust inference in the weak instrument case, is rejected in all three cases, suggesting that all three obstacles are individually important in affecting �rm growth. Over-identi�cation tests are not reported since the equation is just identi�ed in each case. Cross-country regressions are also run using historical institutional vari- ables as instruments (table 5). All three obstacle variables are negative and signi�cantly associated with �rm growth. While the �rst-stage F-tests are sig- ni�cant at least at the 5 percent level in each case, the F-statistic is less than 10, suggesting that the instruments may be weak. Hence, tests for robust inference under weak identi�cation are considered. The Anderson Rubin Wald test of the null hypothesis that the obstacle coef�cient is 0 is rejected in all cases. Con�dence intervals for these coef�cients are also computed. Following Moreira and Poi (2001) and Mikusheva and Poi (2006), critical values of the likelihood ratio tests are obtained, which yield correct rejection probabilities even when the instruments are weak. The con�dence region and the p-value for the coef�cient on the obstacle variable based on the con- ditional likelihood show that the estimated coef�cients belong to the con�- dence region. The underidenti�cation test (Kleibergen-Paap rk Wald statistic) is rejected in each case, indicating that the equation is identi�ed and that instruments pass the test of instrument relevance. The Hansen J-statistic of overidenti�cation is never rejected, suggesting that the instruments are valid. After controlling for a number of other country-level variables, including growth rates, inflation, property rights protection, level of �nancial develop- ment, and level of institutional development, the (unreported) results are unchanged. Overall, with different sets of instruments at the �rm and country level, the results suggest that there are exogenous components of the �nancing, crime, and policy instability obstacles that predict �rm growth and that the results are not due to reverse causality. The instrumental variable estimations also show that �nance is the most robust of the binding obstacles. It must be noted, however, that it is dif�cult to �nd perfect instruments at the level of the �rm in cross-country regressions and hence that some caveats regarding the instru- ments are in order. The country-industry averages of the instruments could potentially be correlated with the error term, so there could be systematic differences in growth rates and �rm complaints across country-industry groups that raise reverse causality concerns. On the use of international accounting standards as an instrument, it should be noted that �rm-�xed effects could not be used in the absence of panel data, so there is always the risk that a �rm’s adoption of accounting standards might be correlated with unobservables that affect �rm growth. Finally, while the instruments in the country-averages regressions can be considered exogenous since historical institutional variables 502 T A B L E 5 . Robustness Test—Instrumental Variables, Firm-Level Regressions, Cross-country Regressions 1 2 3 Common law dummy variable, three Common law dummy variable, three Common law dummy Instrument religion dummy variables religion dummy variables variable, latitude Constant 2.385** (1.013) 1.122*** (0.344) 1.206*** (0.465) GDP per capita 2 0.091** (0.043) 2 0.031* (0.016) 2 0.052* (0.029) Financing 2 0.556** (0.255) Policy instability 2 0.270*** (0.093) Street crime 2 0.264*** (0.102) Number of countries 79 79 80 F-statistic 2.71 (0.037) 6.44 (0.000) 6.95 (0.002) Underidenti�cation test—Kleibergen-Paap 11.74 (0.019) 27.86 (0.000) 14.63 (0.001) rk Wald statistic Weak instruments robust inference— 3.30 (0.015) 3.30 (0.015) 6.69 (0.002) THE WORLD BANK ECONOMIC REVIEW Anderson Rubin Wald test Moreira and Poi Conditional Likelihood (2 2.264, 2 0.213) (2 0.569, 2 0.115) (2 0.726, 2 0.085) Ratio test (0.986) (0.921) (0.983) Overidenti�cation test of all instruments— 0.966 (0.809) 1.227 (0.747) 0.562 (0.453) Hansen J-statistic *Signi�cant at the 10 percent level; **signi�cant at the 5 percent level; ***signi�cant at the 1 percent level. Note: Two-stage instrumental variable regressions are used. Numbers in parentheses are robust standard errors. The �rst-stage regression equation esti- mated is �nancing (or policy instability or street crime) averaged across countries ¼ a þ g1  common law dummy variable þ g2  latitude þ g3  Protestant þ g4  Catholic þ g5  Muslim þ g6  GDP per capita þ 1. The second-stage regression equation estimated is �rm growth ¼ a þ b1  GDP per capita þ b2  �nancing ( predicted value from �rst stage) þ b3  policy instability ( predicted value from �rst stage) þ b4  street crime ( predicted value from �rst stage. The variables are described as follows: �rm growth is the percentage increase in �rm sales over the past three years. GDP per capita is the log of real GDP per capita in U.S. dollars. Financing, policy instability, and street crime are general obstacles as indicated in the �rm questionnaire. They take values 1 – 4, where 1 indicates no obstacle and 4 indicates a major obstacle. English Common law is a dummy variable that takes the value of 1 for common law countries. Latitude is the absolute value of the latitude of the country scaled between 0 and 1. Protestant, Catholic, and Muslim vari- ables are the percentage of Protestant, Catholic, and Muslim religions in each country from La Porta and others (1997). “Does the �rm adopt inter- national accounting standards?� is a dummy variable that takes the value of 1 if the �rm adopts international accounting standards and 0 otherwise. Source: Authors’ analysis based on WBES data described in text. ¨c Ayyagari, Demirgu ¸ -Kunt, and Maksimovic 503 are being used, there is the possibility of omitted-variable bias in the absence of country-�xed effects. Other Robustness Checks This section describes several robustness checks of the main �ndings. First is an investigation of whether the results are driven by a few countries or �rms. Chandra and others (2001) suggest that �rms in African countries may exhibit different responses than the other �rms in the sample. A report by the United States General Accounting Of�ce (2004) analyzes several �rm-level surveys on Africa, including the WBES, and concludes that perceptions of corruption levels vary greatly for African countries, presenting a challenge for broad-based U.S. anticorruption programs. Ayyagari, Demirgu ¸ -Kunt, and Maksimovic ¨c (2008) argue that transition economies are fundamentally different from others in their perceptions of protection of property rights. The �rst four columns of table 6 present the results for preferred speci�ca- tion on different samples after eliminating transition and African economies. While �nancing and crime remain binding constraints, policy instability loses signi�cance when these countries are dropped from the sample. These results suggest that the type of policy instability present in transition and African economies is particularly damaging to �rm expansion. High inflation rates may be responsible for the very high �rm growth rates observed in some countries, particularly in Bosnia and Herzegovina, Estonia, and Uzbekistan. Constructing real �rm growth rates and replicating all the ana- lyses in this article do not change the main results, however. To check whether the results are driven by speci�c outlier �rms, �rms with very high growth rates (higher than 100 percent) are eliminated. Firms report- ing very high growth rates are typically from transition and African economies, where political connections could be behind the high growth rates and �rms thus may not be affected by business environment obstacles. The experience of these �rms may therefore differ from that of the typical �rm. In the reduced sample, �nancing remains the most binding constraint to �rm growth, con�rm- ing that the results are not driven by the fastest growing �rms in the sample. The impact of crime on �rm growth is less robust to eliminating high growth rate �rms, however. It is also possible that young �rms are affected differently by business environment obstacles. Excluding all �rms younger than �ve years old from the sample leaves the �nancing result unchanged, while crime and policy instability are not signi�cant in the regressions (results not reported). This suggests that ensuring policy stability and controlling crime are particularly important to the growth of younger �rms. Financing is still the main binding constraint to growth when robust regression analysis or quintile regressions are used to control for the presence of possible influential outliers. Several other robustness checks of the main �ndings were also conducted (results are available on request). First, the variation at the �rm level and the T A B L E 6 . Robustness Test—Varying Samples 504 High-growth �rms included, countries excluded High-growth �rms excluded, countries excluded Uzbekistan, Uzbekistan, African and Bosnia and African and Bosnia and Transition African transition Herzegovina, Transition African transition Herzegovina, economies economies economies Estonia None economies economies economies Estonia Variable 1 2 3 4 5 6 7 8 9 Constant 0.227*** 0.307*** 0.233*** 0.226*** 0.172*** 0.225*** 0.175*** 0.236*** 0.165*** (0.045) (0.045) (0.045) (0.041) (0.028) (0.039) (0.029) (0.042) (0.028) Firm size 2 0.000 0.005 0.000 0.004 (0.003) 0.003 0.001 0.003 0.000 0.003 (0.002) (0.002) (0.003) (0.002) (0.002) (0.002) (0.002) (0.002) Financing 2 0.012* 2 0.033*** 2 0.020*** 2 0.019*** 2 0.018*** 2 0.017*** 2 0.020*** 2 0.022*** 2 0.016*** (0.006) (0.008) (0.007) (0.007) (0.005) (0.006) (0.005) (0.006) (0.005) Policy 2 0.007 2 0.015* 2 0.010 2 0.008 2 0.015*** 2 0.011 2 0.015** 2 0.010 2 0.014*** THE WORLD BANK ECONOMIC REVIEW instability (0.008) (0.009) (0.008) (0.008) (0.005) (0.007) (0.006) (0.008) (0.005) Street crime 2 0.016** 2 0.027*** 2 0.020*** 2 0.021*** 2 0.007 2 0.018*** 2 0.008 2 0.020*** 2 0.009* (0.007) (0.008) (0.007) (0.007) (0.005) (0.006) (0.005) (0.007) (0.005) Number of 3,224 5,236 2,682 5,534 5,631 3,202 5,107 2,678 5,421 �rms Number of 54 62 38 75 78 54 62 38 75 countries Adjusted R 2 0.073 0.072 0.056 0.053 0.086 0.074 0.082 0.068 0.084 *Signi�cant at the 10 percent level; **signi�cant at the 5 percent level; ***signi�cant at the 1 percent level. Note: Numbers in parentheses are standard errors clustered at the country level. The regression equation estimated is �rm growth ¼ a þ b1  GDP per capita þ b2  size þ b3  �nancing þ b4  policy instability þ b5  street crime. The variables are described as follows: �rm growth is the percentage increase in �rm sales over the past three years. GDP per capita is the log of real GDP per capita in U.S. dollars. Firm size is the log of �rm sales. Financing, policy instability, and street crime are general obstacles as indicated in the �rm questionnaire. They take values 1 – 4, where 1 indicates no obstacle and 4 indicates a major obstacle. Speci�cations 1 – 4 exclude certain countries from the full sample of �rms, while speci�cations 5 – 9 exclude the countries from a reduced sample that does not include �rms reporting very high (or very low) growth rates ( . + 100 percent). All regressions are esti- mated using country-�xed effects with clustered standard errors. Source: Authors’ analysis based on WBES data described in text. ¨c Ayyagari, Demirgu ¸ -Kunt, and Maksimovic 505 variation at the country level were separated—that is, both the individual �rm-level effect of the obstacle (the demeaned value of the obstacle, or obstacle minus the country average of the obstacle) and the cross-country effect (the country average of the obstacle) are included. Once again, in the full speci�cation with the �rm-level and country-level effects of all the 10 business environment obstacles included, the only individual �rm-level obstacles that are binding constraints to growth are �nancing, policy instabil- ity, and crime. Next, various tests were performed to detect outliers and influential points. DFBETA statistics were computed for each obstacle variable. The DFBETAs for regressor i measure the distance that this regression coef�cient shifts when the jth observation is included or excluded from the regression, scaled by the estimated standard errors of the coef�cient. None of the obstacles in the regressions have jDFBETAj . 1 or the even the stricter cutoff of jDFBETAj . p 2 (N), as suggested by Besley, Kuh, and Welsch (1980). This implies that the results are not driven by influential observations. Financing and crime have a signi�cant negative effect on �rm growth, while policy instability is insigni�cant.14 The observed association between obstacles and �rm growth might occur because �rms that face higher obstacles are also those that face limited growth opportunities. After controlling for growth opportunities using average indus- try growth or �rm-level dependence on external �nance, the results remain unchanged using either measure of growth opportunities. Financing, policy instability, and street crime are signi�cant when entered individually, and only �nancing and street crime are signi�cant when entered together. Also investigated is whether �rm ownership drives the results. The sample includes 203 �rms with government ownership. Excluding these �rms leaves the �nancing and crime results unchanged. The sample also includes 1,340 �rms with more than 50 percent foreign ownership. When these foreign �rms are excluded from the analysis, only the �nancing obstacle remains signi�cant. This suggests that foreign-owned �rms are particularly sensitive to policy instability and crime. Including dummy variables to control for government and foreign ownership also leads to similar results, in that only �nancing and crime are signi�cant. Finally, the results are checked for robustness subject to controlling for per- ception biases. Following Kaufmann and Wei (1999), two kvetch variables were constructed, Kvetch1 and Kvetch2, which are deviations of each �rm’s response from the mean country response to two general survey questions. 14. The DFITS statistic of Welsch and Kuh (1977), which identi�es the influence of each observation on the �tted model, was also computed (unreported results). Besley, Kuh, and Welsch p (1980) suggest that a cutoff of jDFITSjj . 2 (k/N) indicates influential observations, where k is the number of estimated coef�cients and N is the number of observations. There are 145 observations in the current sample with jDFITSj greater than the cutoff value. When these influential observations are dropped, the �nancing, policy instability, and crime obstacles are all negative and signi�cant. 506 THE WORLD BANK ECONOMIC REVIEW Kvetch1 uses the responses to the question: How helpful do you �nd the central government today towards businesses like yours? Kvetch2 is constructed using the responses to the question: How predictable are changes in economic and �nancial policies? Since higher values correspond to unfavorable responses, positive deviations from the country mean indicate pessimism, and negative deviations indicate optimism. Controlling for differences in perceptions using the kvetch variables leaves only �nancing and crime results unchanged. Policy instability remains insigni�cant. Individual Financing Obstacles The results indicate that �nancing is one of the most important obstacles that directly constrain �rm growth. To get a better understanding of what type of �nancing obstacles are constraining �rm growth, entrepreneurs were asked to rate the extent to which the following �nancing factors represent an obstacle to their growth: collateral requirements, paperwork and bureaucracy, high interest rates, need for special connections, banks lacking money to lend, access to foreign banks, access to nonbank equity, access to export �nance, access to �nancing for leasing equipment, inadequate credit and �nancial information on customers, and access to long-term loans. The ratings are again on a scale of 1 to 4, increasing with the severity of obstacles. Table 7 reports regressions that parallel those in table 2, but focusing on speci�c �nancing obstacles. A residual is also included for the component of the general �nancing obstacle not explained by the individual obstacles. The results indicate that not all �nancing obstacles reported by �rms are constrain- ing. Only the coef�cients of collateral, paperwork, high interest rates, special connections, banks’ lack of money to lend, lease �nance, and the residual are signi�cant when entered individually. High interest rates have the highest econ- omic impact—a one-standard deviation increase in the obstacle results in a 3.3 percent decrease in �rm growth. Unlike the obstacles examined previously, speci�c �nancing obstacles are highly correlated with each other. Speci�cation 13 includes all obstacles that are signi�cant when entered individually. Only the high interest rates coef�- cient is signi�cant and only at the 10 percent level. If the residual is also included, as in speci�cation 14, only the residual remains signi�cant. The residual is likely to summarize how different �rms are affected differently by the structure and ownership of the �nancial system, the level of competition, and other factors that are not fully captured by the speci�c �nancial obstacles, thus proxying for general access to credit.15 Looking at the correlations among obstacles using DAG analysis shows that high interest rates are the only �nancial obstacle directly constraining �rm growth. (It may be noted that while the direction of causation is restricted to go 15. The residual remains signi�cant if all the general obstacles are included in addition to the residual and the signi�cant individual �nancing obstacles. T A B L E 7 . Impact of Individual Financing Obstacles on Firm Growth Variable 1 2 3 4 5 6 7 8 9 10 11 12 13 14 Constant 0.180*** 0.172*** 0.211*** 0.132*** 0.166*** 0.158*** 0.129*** 0.129*** 0.106*** 0.122*** 0.121*** 0.094*** 0.264*** 0.212*** (0.048) (0.031) (0.029) (0.033) (0.030) (0.034) (0.028) (0.040) (0.039) (0.034) (0.036) (0.039) (0.032) (0.040) Firm size 0.003 0.004 0.004 0.005 0.004 0.004 0.004 0.005 0.005* 0.004 0.005* 0.006* 0.002 0.005 (0.003) (0.003) (0.002) (0.003) (0.003) (0.003) (0.003) (0.003) (0.003) (0.003) (0.003) (0.003) (0.003) (0.003) Collateral 2 0.023*** 2 0.006 2 0.008 (0.011) (0.007) (0.010) Paperwork 2 0.025*** 2 0.010 2 0.015 (0.011) (0.009) (0.010) High interest rates 2 0.032*** 2 0.020* 2 0.011 (0.012) (0.010) (0.011) Special connections 2 0.015** 2 0.001 2 0.002 (0.014) (0.007) (0.010) Lack money to lend 2 0.024*** 2 0.011 2 0.007 (0.012) (0.008) (0.009) Lease �nance 2 0.015 (0.009) Ayyagari, Demirgu Access to foreign 2 0.002 ¨c banks (0.007) Access to nonbank 2 0.005 equity (0.008) Export �nance 0.004 (0.009) Credit 0.003 (0.007) Long-term loans 2 0.008 (0.008) ¸ -Kunt, and Maksimovic Financing residual 2 0.022** 2 0.023** (0.011) (0.011) (Continued ) 507 508 TABLE 7. Continued Variable 1 2 3 4 5 6 7 8 9 10 11 12 13 14 Number of �rms 6,024 6,133 6,298 6,002 5,808 5,076 5,093 5,037 4,440 5,332 5,030 2,988 5,317 2,988 Number of countries 79 79 79 79 79 78 78 78 78 78 60 58 79 58 Adjusted R 2 0.070 0.069 0.070 0.064 0.074 0.070 0.065 0.070 0.071 0.072 0.068 0.006 0.071 0.065 *Signi�cant at the 10 percent level; **signi�cant at the 5 percent level; ***signi�cant at the 1 percent level. Note: Numbers in parentheses are standard errors clustered at the country level. The regression equation estimated is �rm growth ¼ a þ b1  size þ THE WORLD BANK ECONOMIC REVIEW b2  collateral þ b3  paperwork þ b4  high interest rates þ b5  special connections þ b6  lack money to lend þ b7  access to foreign banks þ b8  access to nonbank equity þ b9  export �nance þ b10  lease �nance þ b11  credit þ b12  long-term loans þ b13 (residual). The variables are described as follows: �rm growth is the percentage increase in �rm sales over the past three years. Firm size is the log of sales. Collateral, paperwork, high interest rates, special connections, lack money to lend, access to foreign banks, access to nonbank equity, export �nance, lease �nance, credit, and long-term loans are individual �nancing obstacles as indicated in the �rm questionnaire. They take values of 1 – 4, where 1 indicates no obstacle and 4 indicates a major obstacle. In speci�cations 1– 11, each of the obstacle variables is included individually. Residual is the residual from a regression of the general �nancing obstacle on all the individual �nancing obstacles. Speci�cation 13 includes collateral, paperwork, high interest rates, special connections, lack of money to lend, and lease �nance. Speci�cations 12– 14 include the �nancing residual. All regressions are estimated using country-�xed effects with clustered stan- dard errors. Source: Authors’ analysis based on WBES data described in text. ¨c Ayyagari, Demirgu ¸ -Kunt, and Maksimovic 509 from the �nancing obstacles to growth, no ordering is imposed among the indi- vidual �nancing obstacles.) That �nding is not surprising since the high interest rate obstacle captures the cost of �nancing and is itself an endogenous variable that depends on the ability of the �nancial system to satisfy the demand for capital. It can be expected to constrain all �rms in all countries. Collectively, speci�c �nancing obstacles still do not capture everything measured by the general �nancing obstacle, as illustrated by the effect of the residual. This also suggests that the general access to credit is an important constraint for �rms. The DAG analysis also suggests that perceptions of high collateral require- ments and paperwork influence the perceptions of high interest rates. High interest rates also influence perceptions of lack of access to lease �nance, banks lacking money to lend, and the need for special connections in banking. Regressions of the high interest rate obstacle on individual �nancing obstacles found speci�c �nancing obstacles all to be individually correlated with high interest rates. When all �nancing obstacles are considered together, only collat- eral, paperwork, special connections, lack of money to lend, and access to long-term loans are correlated with high interest rates, as in the DAG analysis. I V. C O N C L U S I O N AND PO L I C Y I M P L I CAT I O N S Although �rms report many obstacles to their growth, not all of them are equally constraining. Some may affect �rm growth only indirectly, through their influence on other factors, or not at all. Analyses using regressions and DAG methodology found only �nance, crime, and policy instability to be binding constraints, with a direct association with the growth rate of �rms. Thus, while the other obstacles studied in this article are also associated with �rm growth through their impact on each other and on the direct obstacles, maintaining policy stability, keeping crime under control, and undertaking �nancial sector reforms to relax �nancing constraints are likely to be the most effective means of promoting �rm growth. The �nancing obstacle’s impact on growth is robust to varying samples of countries, while the policy instability and crime results are less robust to the exclusion of transition and African economies, where they might be the most problematic for business growth. The results were subject to a battery of robustness tests, including changing the sample and controlling for reverse causality, growth opportunities, and poten- tial perception biases in survey responses. The �nancing obstacle was the most robust to all these tests. This was further con�rmed through instrumental vari- able regressions. This suggests that �nancial sector reform should be a priority for governments contemplating reform of their business environments.16 Further investigation of the �nancing obstacles revealed the importance of high interest rates in constraining �rm growth. This result highlights the 16. An implicit assumption with the use of any survey data is that �rm managers are knowledgeable about the different obstacles and understand the true workings of the �nancial and legal systems. 510 THE WORLD BANK ECONOMIC REVIEW importance of macroeconomic policies in influencing growth at the �rm level, as indicated by the correlation between high interest rates and banks’ lack of money to lend. High interest rates are also correlated with high collateral and paperwork requirements, the need for special connections with banks, and the unavailability of long-term loans. These results suggest that bureaucracy and corruption in banking, greater collateral requirements, and lack of long-term loans are common in high-interest-rate environments. In addition to the cost of �nancing, general access to credit is an important constraint to �rm growth. Country- and �rm-level determinants of �nancing obstacles would bene�t from further investigation. FUNDING This research was supported by the National Science Foundation (NSF Grant # SES-0550454/0550573). APPENDIX T A B L E A - 1 . General Obstacles General obstacles Taxes Firm Number Policy Exchange Judicial Street and Anticompetitive Country growth of �rms Financing instability Inflation rate ef�ciency crime Corruption regulation behavior Infrastructure Albania 0.22 103 3.04 3.48 2.75 2.61 2.69 3.42 3.34 3.15 2.72 3 Argentina 0.08 82 3.01 3.07 1.77 1.73 2.27 2.39 2.58 3.34 2.41 1.93 Armenia 2 0.2 96 2.45 2.87 2.73 2.69 1.5 1.85 1.96 3.39 1.9 1.77 Azerbaijan 2 0.2 70 3.11 2.55 2.9 2.61 2.59 2.39 3 3.17 2.96 2.43 Bangladesh 0.13 34 2.6 3.08 2.86 3.09 2.38 3.07 3.61 3.03 2.4 Belarus 0.1 97 3.33 2.95 3.63 3.16 1.55 2.17 1.88 3.34 1.99 1.7 Belize 0.12 26 2.81 2.38 2.04 1.73 1.56 2.12 1.96 2.77 1.96 2.19 Bolivia 0.04 80 3.03 3.1 2.58 2.46 2.78 2.76 3.56 3.15 2.71 2.63 Bosnia and 0.63 76 3.09 3.19 1.33 1.25 2.54 1.86 2.56 3.16 2.58 2.65 Herzegovina Botswana 0.32 72 2.24 1.55 1.93 1.33 1.88 1.65 1.89 2.16 Brazil 0.03 148 2.67 3.53 2.8 2.94 2.56 2.83 2.53 3.66 2.49 2.18 Ayyagari, Demirgu ¨c Bulgaria 0.15 101 3.16 3.03 2.76 2.37 2.26 2.64 2.64 3.1 2.34 2.23 Cambodia 0.07 298 2.04 2.9 2.61 2.32 2 3.29 2.23 2.21 2.33 Cameroon 0.12 44 3.14 2.03 2.03 2.28 2.94 3.36 2.7 3.44 Canada 0.17 74 2.1 2.18 2.15 2.16 1.47 1.32 1.4 2.59 1.62 1.41 Chile 0.09 81 2.36 2.58 2.16 2.59 1.97 2.4 1.86 2.36 1.91 1.86 China 0.05 70 3.36 2.1 2.23 1.83 1.5 1.83 1.94 2.03 2.13 1.89 Colombia 0.06 83 2.67 3.49 3.01 3.34 2.4 3.37 2.87 3.17 2.33 2.46 Costa Rica 0.25 81 2.62 2.67 2.93 2.75 2.2 2.89 2.52 2.8 2.44 2.63 ¸ -Kunt, and Maksimovic Cote d’Ivoire 0.05 47 2.78 2.85 2.37 1.97 3.29 3.24 2.49 2.29 Croatia 0.1 97 3.26 3.11 2.47 2.86 2.74 2.09 2.59 3.34 2.04 1.94 511 (Continued ) TABLE A-1. Continued 512 General obstacles Taxes Firm Number Policy Exchange Judicial Street and Anticompetitive Country growth of �rms Financing instability Inflation rate ef�ciency crime Corruption regulation behavior Infrastructure Czech 0.1 80 3.18 2.95 3 2.46 2.18 2.09 2.1 3.44 2.16 2.5 Republic Dominican 0.21 95 2.63 3.02 2.85 2.88 2.43 3.22 3 3.96 2.75 2.63 Republic Ecuador 2 0.06 74 3.27 3.6 3.76 3.78 3.04 3.49 3.53 3.07 2.55 2.67 Egypt, Arab 0.16 44 2.91 3.14 2.68 2.9 2.24 3.14 3.43 3.23 Republic El Salvador 2 0.02 73 2.93 2.97 3.16 2.55 2.65 3.67 3.06 2.93 2.36 2.52 Estonia 0.63 109 2.47 2.62 2.41 1.89 1.72 2.09 1.88 2.67 1.85 1.64 THE WORLD BANK ECONOMIC REVIEW Ethiopia 0.26 70 3.02 2.38 2.26 2.47 1.51 2.46 2.33 3.04 France 0.2 62 2.61 2.2 2.03 1.82 1.79 1.77 1.62 3.13 2.02 1.81 Georgia 0.14 78 3.29 2.84 3.29 2.94 1.86 2.32 3.04 3.22 2.18 2.14 Germany 0.11 60 2.59 1.63 1.87 1.64 2.12 1.56 1.88 3.17 2.3 1.71 Ghana 0.19 58 3.1 2.37 3.43 2.58 2.37 2.78 2.83 2.74 Guatemala 0.18 84 2.99 3.16 3.32 3.6 2.5 3.22 2.7 2.75 2.28 2.52 Haiti 0 62 3.28 3.18 2.92 2.9 2.35 3.81 3.08 2.73 3.1 3.89 Honduras 0.1 65 2.97 2.53 3.41 3.3 2.41 3.23 2.9 2.83 2.79 2.56 Hungary 0.28 98 2.6 2.61 2.59 1.6 1.32 1.76 1.95 3.01 2.14 1.53 India 0.15 152 2.59 2.81 2.77 2.42 2.02 1.98 2.8 2.43 2.8 Indonesia 2 0.05 70 2.83 3.14 3.21 3.4 2.26 2.69 2.69 2.59 2.96 2.37 Italy 0.16 64 1.97 2.97 2.23 1.83 2.22 2.22 1.76 3.25 2.19 2.24 Kazakhstan 0.1 89 3.29 2.88 3.62 3.48 2.08 2.6 2.7 3.37 2.55 2.1 Kenya 0.03 70 2.76 3.03 2.8 1.75 3.27 3.56 2.53 3.64 Kyrgyz 0 68 3.47 3.23 3.78 3.48 2.13 3.26 3.19 3.59 3 1.98 Republic Lithuania 0.08 68 3.03 2.27 2.3 1.91 2.25 2.52 2.44 3.26 2.31 1.82 Madagascar 0.16 67 3.08 2.67 3.32 2.3 2.79 3.44 2.75 3.03 Malawi 0.64 30 2.81 2.2 3.56 2.54 3.08 2.65 2.37 3.76 Malaysia 0.01 37 2.57 2.14 2.44 1.94 1.63 1.78 2 2.03 1.91 1.92 Mexico 0.24 71 3.24 3.27 3.48 3.13 2.77 3.37 3.31 3.21 2.75 2.23 Moldova 2 0.15 84 3.42 3.6 3.86 3.51 2.51 3.11 2.93 3.58 2.93 2.64 Namibia 0.3 52 2 1.66 2.08 2.08 1.96 1.71 1.98 1.63 Nicaragua 0.21 76 3.05 2.91 3.39 3.07 2.33 2.8 2.88 2.96 2.42 2.71 Nigeria 0.26 63 3.11 3.43 3.21 2.92 3.3 3.37 3.1 3.68 Pakistan 0.05 61 3.28 3.64 3.21 2.87 2.56 3.03 3.54 3.2 2.67 3.08 Panama 0.09 81 2.06 2.72 2.04 1.42 2.4 2.98 2.8 2.38 2.44 2.19 Peru 2 0.02 83 3.09 3.21 2.85 2.99 2.55 2.81 2.83 3.35 2.68 2.27 Philippines 0.07 84 2.69 2.85 3.36 3.43 2.24 2.8 3.13 3.08 2.9 2.88 Poland 0.33 175 2.47 2.75 2.58 2.27 2.3 2.37 2.27 3.08 2.23 1.67 Portugal 0.12 52 1.8 2.08 2.1 1.74 1.88 1.64 1.73 2.15 2.18 1.75 Romania 0.07 96 3.26 3.44 3.75 3.19 2.59 2.45 2.88 3.57 2.52 2.44 Russian 0.29 384 3.2 3.49 3.53 3.15 2.17 2.65 2.62 3.58 2.67 2.12 Federation Senegal 0.15 38 3 2.21 2.56 2 2.61 3.04 2.97 2.88 Singapore 0.12 74 1.97 1.5 1.61 1.88 1.32 1.22 1.28 1.55 1.58 1.42 Slovak 0.14 91 3.34 1.53 3.13 2.43 2.13 2.49 2.47 3.25 2.26 1.98 Republic Slovenia 0.29 101 2.3 2.6 2.23 2.21 2.29 1.68 1.64 2.91 2.43 1.74 South Africa 0.26 87 2.34 1.97 2.45 2.39 3.58 2.58 2.64 1.83 Ayyagari, Demirgu ¨c Spain 0.25 66 2.21 2.17 2.27 1.93 1.97 1.92 2.08 2.65 2.25 1.94 Sweden 0.23 73 1.83 2.46 1.66 1.78 1.46 1.54 1.18 2.67 1.97 1.52 Tanzania 0.25 40 2.85 2.48 2.65 2.07 1.96 2.88 2.7 3.21 Thailand 2 0.02 337 3.1 3.49 3.4 3.62 2.13 3.48 3.47 3.54 3.6 2.76 Trinidad and 0.18 80 3.03 1.81 2.49 2.41 1.45 2.18 1.68 2.78 1.79 2.1 Tobago Tunisia 0.14 41 1.79 1.94 1.7 1.94 1.55 2.11 2.12 2.1 Turkey 0.1 115 3.12 3.55 3.61 2.83 2.3 2.09 2.89 3.16 2.79 2.22 Uganda 0.18 67 3.17 2.47 2.68 1.78 2.27 2.93 2.48 2.81 ¸ -Kunt, and Maksimovic Ukraine 0.03 170 3.45 3.22 3.43 3.05 2.16 2.49 2.51 3.7 2.86 2.22 (Continued ) 513 514 TABLE A-1. Continued General obstacles Taxes Firm Number Policy Exchange Judicial Street and Anticompetitive Country growth of �rms Financing instability Inflation rate ef�ciency crime Corruption regulation behavior Infrastructure United 0.27 62 2.33 2.19 2.16 2.28 1.5 1.95 1.24 2.87 1.72 1.69 Kingdom United States 0.16 66 2.38 2.05 2.12 1.71 1.84 2.14 1.88 2.39 1.7 1.83 THE WORLD BANK ECONOMIC REVIEW Uruguay 0 72 2.73 2.61 2.03 2.39 1.91 2.07 2 3.21 1.71 1.9 Uzbekistan 0.64 94 2.77 2.03 3.04 2.6 1.68 1.77 2.22 2.66 2.28 1.95 Venezuela 2 0.02 78 2.62 3.64 3.48 3.12 2.65 3.18 3 3.1 2.63 2.31 Zambia 0.18 46 2.95 2.57 3.45 1.88 3.18 2.78 2.39 3.07 Zimbabwe 0.47 91 3.05 2.73 3.83 2.93 2.57 2.87 2.87 2.53 Average 0.15 86.73 2.8 2.72 2.76 2.49 2.15 2.51 2.56 2.9 2.37 2.34 Note: The variables are described as follows: �rm growth is the percentage change in �rm sales over the past three years (1996 – 99). Financing, policy instability, inflation, exchange rate, judicial ef�ciency, street crime, corruption, taxes and regulation, anticompetitive behavior, and infrastructure are general obstacles as indicated in the �rm questionnaire. They take values 1 –4, where 1 indicates no obstacle and 4 indicates a major obstacle. Firm obstacles are averaged over all �rms in each country. The number of �rms reported is the number of �rms with nonmissing �rm growth rates. Source: Authors’ analysis based on WBES data described in text. ¨c Ayyagari, Demirgu ¸ -Kunt, and Maksimovic 515 REFERENCES Ayyagari, M., A. Demirgu ¸ -Kunt, and V. Maksimovic. 2005. “What Determines Protection of Property ¨c Rights? Analysis of Direct and Indirect Effects Using DAG Methodology.� George Washington University, World Bank Development Economics Group, Washington, D.C.; College Park, Maryland: University of Maryland. ———. 2008. “How Well Do Institutional Theories Explain Firms’ Perceptions of Property Rights?� Review of Financial Studies 21(4):1833–71. Barro, R.J. 1991. “Economic Growth in a Cross-Section of Countries.� Quarterly Journal of Economics 106(2):407–43. ¨c Beck, T., A. Demirgu ¸ -Kunt, and V. Maksimovic. 2005. “Financial and Legal Constraints to Firm Growth: Does Firm Size Matter?� Journal of Finance 60(1):137–77. ———. 2006. “The Influence of Financial and Legal Institutions on Firm Size.� Journal of Banking and Finance 30(11):2995 –3015. Beck, T., R. Levine, and N. Loayza. 2000. “Finance and the Sources of Growth.� Journal of Financial Economics 58(1):261– 300. Belsley, D.A., E. Kuh, and R.E. Welsch. 1980. Regression Diagnostics. New York, N.Y.: Wiley. Carlin, W., M. Schaffer, and P. Seabright. 2005. “Where Are the Real Bottlenecks? A Lagrangian Approach to Identifying Constraints on Growth from Subjective Survey Data.� CEPR Discussion Paper 5719. Washington, D.C.: Center for Economic Policy Research. Chandra, V., L. Moorty, J. Nganou, B. Rajaratnam, and K. Schaefer. 2001. “Constraints to Growth and Employment in South Africa.� Discussion Paper No.15. World Bank, Southern Africa Department, Washington, D.C. ¨c Demirgu ¸ -Kunt, A., and V. Maksimovic. 1998. “Law, Finance, and Firm Growth.� Journal of Finance 53(6):2107 –37. ¨c Demirgu ¸ -Kunt, A., I. Love, and V. Maksimovic. 2006. “Business Environment and the Incorporation Decision.� Journal of Banking and Finance 30(11):2967 –93. Dollar, D., M. Hallward-Driemeier, and T. Mengistae. 2005. “Investment Climate and Firm Performance in Developing Countries.� Economic Development and Cultural Change 54(1):1–31. Gelb, A., V. Ramachandran, M.K. Shah, and G. Turner. 2007. “What Matters to African Firms? The Relevance of Perceptions Data.� Policy Research Working Paper 4446. World Bank, Washington, D.C. Fisman, R., and J. Svensson. 2007. “Are Corruption and Taxation Really Harmful to Growth? Firm Level Evidence.� Journal of Development Economics 83(1):63–75. Fleisig, H. 1996. “Secured Transactions: The Power of Collateral.� Finance and Development 33(2):44–46. Hausmann, R., D. Rodrik, and A. Velasco. 2008. “Growth Diagnostics� In J. Stiglitz, and N. Serra eds., The Washington Consensus Reconsidered: Towards a New Global Governance. New York: Oxford University Press. Kaufmann, D., and S. Wei. 1999. “Does Grease Money Speed up the Wheels of Commerce?� Policy Working Paper 2254. World Bank, Washington, D.C. King, R.G., and R. Levine. 1993. “Finance and Growth: Schumpeter Might Be Right.� Quarterly Journal of Economics 108(3):717–38. Kormendi, R.C., and P.G. Meguire. 1985. “Macroeconomic Determinants of Growth: Cross-Country Evidence.� Journal of Monetary Economics 16(2):141–63. Krueger, A.B., and J. Angrist. 2001. “Instrumental Variables and the Search for Identi�cation: From Supply and Demand to Natural Experiments? Journal of Economic Perspectives 15(4):69– 85. La Porta, R., F. Lopez-de-Silanes, A. Shleifer, and R.W. Vishny. 1997. “Legal Determinants of External Finance.� Journal of Finance 52(3):1131 – 50. 516 THE WORLD BANK ECONOMIC REVIEW Levine, R. 2005. “Finance and Growth: Theory and Evidence.� In P. Aghion, and S. Durlauf eds., Handbook of Economic Growth. Amsterdam: Elsevier Science. Levine, R., N. Loayza, and T. Beck. 2000. “Financial Intermediation and Growth: Causality and Causes.� Journal of Monetary Economics 46(1):31–77. Levine, R., and D. Renelt. 1992. “A Sensitivity Analysis of Cross-Country Growth Regressions.� American Economic Review 82(4):942–63. Mikusheva, A., and B.P. Poi. 2006. “Tests and Con�dence Sets with Correct Size When Instruments Are Potentially Weak.� The Stata Journal 6(3):335–47. Moreira, M.J., and B.P. Poi. 2001. “Implementing Tests with Correct Size in the Simultaneous Equation Model.� The Stata Journal 1(1):1–15. Rajan, R., and L. Zingales. 1998. “Financial Dependence and Growth.� American Economic Review 88(3):559–87. Sala-i-Martin, X. 1997. “I Just Ran Two Million Regressions.� American Economic Review 87(2):178–83. Scheines, R., P. Spirtes, C. Glymour, and C. Meek. 1994. TETRAD II: Users Manual. Hillsdale, N.J.: Lawrence Erlbaum Associates. Spirtes, P., C. Glymour, and R. Scheines. 2001. Causation, Prediction, and Search. 2nd edition. Cambridge, Mass.: MIT Press. Stock, J., and M. Watson. 2003. Introduction to Econometrics. Reading, Mass.: Addison– Wesley. Svejnar, J., and S. Commander. 2007. “Do Institutions, Ownership, Exporting, and Competition Explain Firm Performance? Evidence from 26 Transition Countries.� IZA Discussion Paper 2637. Bonn, Germany: Institute for the Study of Labor. United States General Accounting Of�ce. 2004. “U.S. Anticorruption Programs in Sub-Saharan Africa Will Require Time and Commitment.� GAO-04-506. Report to the Subcommittee on African Affairs, Committee on Foreign Relations, U.S. Senate. Washington, D.C.: United States General Accounting Of�ce. Welsch, R.E., and E. Kuh. 1977. Linear Regression Diagnostics. Working Paper 173. Cambridge, Mass.: National Bureau of Economic Research. World, Bank. 2005. World Development Indicators 2005. Washington, DC: World Bank.