Corruption and business procedures: an empirical investigation S. Roy*, Department of Economics, High Point University, High Point, NC - 27262, USA. Email: sroy@highpoint.edu Abstract We implement OLS, IV, and quasi-maximum-likelihood estimation techniques to estimate the impact of business procedures on the "control of corruption" index from the World Governance Indicators. We find that increases in business procedures lead to greater corruption. The result remains robust to various estimation strategies as well as inclusion of control variables. Keywords: Corruption; Business procedures; Fractional response; Instrumental variables. JEL Classification: C25, C26, D72, D73. ------------------------------------------------------------------------------------------------------------ *Corresponding author. Email: sroy@highpoint.edu 1
1. Introduction Do business regulations engender corruption? Major strands of public choice theory view regulations to be socially inefficient and create rent seeking opportunities. According to the theory of regulatory capture (Stigler (1971)), regulation of entry prevents new firms from entering a market. Thus they are acquired by incumbent businesses to create rents for themselves. On the other hand, the "tollbooth" version of the public choice theory (Djankov et al., 2002) proposes that regulations are devised by politicians and bureaucrats both to create rents, as well as to extract them through campaign contributions, votes, and bribes. While the former theory points to benefits of regulations derived by the industry, the latter viewpoint emphasizes the benefits to the politicians at the expense of businesses. Both the hypotheses suggest that business regulations give rise to rent seeking activities in the form of corruption. Svensson (2005) and Kaufman et al. (2007) found positive evidences to this effect. However, Treisman (2007) found that business regulations, measured by the number of days required to start a business, was not statistically significant in explaining the "control of corruption" measure from the World Bank's World Governance Indicators (WGI) in the two-stage least squared (2SLS) regressions. Thus, Djankov (2009) notes "that the link between regulation and corruption is still unexplored fully and that more rigorous empirical tests are needed to establish causation". In this paper, we investigate the effect of business regulations on corruption for over 100 countries using the above-mentioned WGI corruption measure as the dependent variable, and an alternate measure of regulatory burden- the number of procedures necessary to start a business in different countries as the proxy for business regulations. Our results suggest that across countries, increase in business procedures leads to greater corruption. The result remains robust to different estimation techniques and inclusion of control variables suggested to be important determinants of corruption in other studies. The paper contributes to the empirical literature on the effect of business regulations on corruption. While the existing studies in this area (e.g. Djankov (2002), Svensson (2005), Lichetti and Madani (2010)) have pointed to high correlations between corruption and regulations, these studies did not establish a causal link between the two variables. Treisman (2007) found the number of business procedures to be significant in explaining the Corruption Perceptions Index (CPI) constructed by the Transparency International (TI), while, in this paper, we use corruption data from the WGI. The paper also contributes to the literature on the cross- 2
country determinants of corruption by treating the corruption index as a fractional response dependent variable and implementing robust methods for estimation and inference applicable for bounded dependent variables. Studies on the determinants of corruption using corruption indexes have so far treated these measures as unbounded continuous variables. In addition, we allow feedback effects from corruption to business procedures by using suitable instruments for the latter. The remainder of the paper is organized as follows. Section 2 describes the data sources and summary statistics of the major variables of interest; Section 3 lays out the estimation models used in the study and discusses the major findings of the paper; and Section 4 concludes. 2. Data sources and summary statistics Table 1 reports the summary statistics for the major variables of interest. The measure for corruption used in the study is the Control of corruption variable from the World Governance Indicators (available online at www.govindicators.org) This measure captures perceptions of the extent to which public power is exercised for private gain, including both petty and grand forms of corruption, as well as "capture" of the state by elites and private interests. While the CPI measures perceptions regarding the extent of corruption in the public sector, the WGI is more comprehensive and measures effects of both public as well as private corruption. The original WGI corruption index (ranging from -2.59 to 1.68 in the sample) is transformed to a 0-1 relative scale with higher values of the index denoting greater corruption. Data on the number of business procedures are from the Doing Business database (available online at www.doingbusiness.org). This variable measures all mandatory procedures for an entrepreneur to start up and formally operate an industrial or commercial business. To reduce measurement error, we use the average of both corruption and business procedures for 2004-2010 during which period data for both variables are simultaneously available. Following previous studies, real GDP per capita (PCI) obtained from the World Bank's World Development Indicators is used to proxy for economic development. In order to mitigate potential reverse causality from corruption to development, we use lagged values of PCI (average of 1991-2000). We also introduce PCI-squared as an additional explanatory variable. This is motivated by the stylized fact that low-income countries exhibit greater corruption levels than high-income countries. Thus, the impact of per capita 3
income on corruption might be conditional on the level of economic development hereby captured by the PCI-squared term. 1 To address the feedback effect from corruption to regulations, we instrument for business procedures using the British legal origin dummy and the absolute value of latitude obtained from La Porta et al. (1999). Legal origins are viewed as proxy for the state's power to intervene in economic life, and countries with British legal origin are usually found to be less interventionist (Djankov et al., 2002) compared to others. The same study found that countries located further from the equator (measured by absolute value of latitude) are generally found to have better government performance. For robustness check, we introduce additional control variables representing the cultural, political, and economic determinants of corruption (see Treisman (2007) for a comprehensive survey on the determinants of corruption perceptions in cross-country regressions). These include: fractionalization index (for 1985); percent of Muslim population (for 1980); an inverse measure of democracy indicating the power of the President (for 2000); countries with plurality voting rule for Parliamentary electoral systems (for 2000); number of armed conflicts (average between 1995-2000); (logarithm of) import-gdp ratio (for 2000); and (logarithm of) average inflation between 1998-2000. Data for all the controls are based on Treisman (2007) and obtained from his homepage (http://www.sscnet.ucla.edu/polisci/faculty/treisman/pages/publishedpapers.html). After incorporating all the controls, our regressions are based on cross-country data covering a minimum of 123 developed and developing countries. 3. Estimation strategy, models and results 3.1 Linear model The conditional mean of our linear model can be expressed as: E( Corruption / x) PCI 2 1 2 log( Bu sin ess Procedure ) 3 log( PCI ) 4(log( )) (1) 1 Without the square of the per capita income term in the regression, the model is not properly specified as indicated by a significant RESET test that suggests neglected nonlinearities. For brevity, the result is not reported and can be obtained from the author on request. A square of business procedures term subsequently introduced, however, is found to be not significant in all the specifications. 4
3.2 Linear model results Col. (1) in Table 2 reports OLS results for the baseline model in equation (1). Logarithm of business procedures is found to be positive and highly significant, (logarithm of) PCI is also positive and significant, while the (logarithm of) PCI-squared term is negative and significant (all at 1% level of significance). Based on the results, 10% increase in business procedures is associated with approx. 1% increase in corruption levels across countries. The coefficients associated with the PCI and the PCI-squared terms allow us to estimate a threshold level of per capita income below which an increase in income is associated with greater corruption while further increases in income beyond this level reduces corruption. 2 For the baseline specification, this threshold per capita income is estimated to be around $200. This suggests that increases in per capita income is associated with reduced corruption for countries with real GDP per capita above $200, while income increases might be associated with greater corruption for countries with per capita incomes below this threshold, i.e. the low-income countries. Based on the adjusted R-squared, our basic linear model is able to explain about 78% of the variability of the corruption index. In order to address the potential reverse causality from business regulations to corruption which would render the estimates biased and inconsistent, we instrument for business procedures in our baseline model using the legal origin dummy and the distance of a country from the equator. The results are reported in Col. (2) of Table 2. Business procedures remain highly significant and strongly associated with higher corruption levels. Based on the IV results, 10% increase in business procedures is associated with approx. 1.4% increase in corruption levels. Thus instrumenting for business procedures substantially increases the impact of regulations on corruption. The instrument set (British legal origin and latitude) is found to satisfy tests for weak instruments based on the first-stage F-statistic as well as the p-values associated with the conditional likelihood ratio (CLR) test which is robust to weak instruments. The (insignificant) p-values associated with the J-statistic indicate that the instruments satisfy the over-identification tests and thereby impact corruption only through business procedures. 2 From equation (1), the estimated value (logarithm of) of real per capita GDP can be calculated as 3 2 ). ( 4 5
Cols. (3) and (4) in Table 2 report the results for the OLS and the IV model for equation (1) in the presence of control variables. The effect of business procedures, real GDP per capita, and its squared term on corruption remain substantively unaffected. The final specification in col. (4) suggests 10% increase in business procedures to increase corruption by about 1.6%. Apart from plurality, import-gdp ratio, and average inflation, the other controls are found to be not significant. For example, in the IV regression, only country-level inflation is found to be significant (at 10% level) and positively related to corruption. Across all specifications, the (robust) model specification test for the presence of omitted variables fails to reject the null hypothesis that the model is correctly specified 3. Thus the model is able to capture substantial variability of the dependent variable (reflected in the adjusted R- squared ranging from 0.77 to 0.79). Moreover, the results also indicate a broadly positive effect of per capita income in reducing corruption, except for a small number of low-income countries (below $148 real per capita income in the final specification). 3.3 Fractional logit model Given that the corruption index is bounded, the linear model might not fit the data, viz. if we have large number of observations on corruption close to 0 or 1, or if the predictions from the linear model fall outside the (0,1) bounded range. 4 In such scenario, Wooldridge (2010) shows that an alternative estimation procedure will be to specify the conditional mean of the dependent variable as a logistic function. Thus equation (1) can be expressed as: z z E( Corruption / z) E( y / z) e /[1 e ] G( z ) (2) where, y denotes Corruption, e denotes exponential, and z denotes the right-hand side of equation (1). The fractional logit model entails maximizing the following log-likelihood function (with respect to ): L( ) y log[ G( z )] (1 y)log[1 G( z )] (3) Since equation (3) is a member of the linear exponential family, the quasi-maximum likelihood estimator (QMLE) of obtained from maximizing (3) is consistent, provided the 3 This test is implemented using the -ovtest- command in Stata. Non-rejection of the null hypothesis, as in our case, indicates the null hypothesis that the model has no omitted variables is not rejected. 4 From Table 1, corruption data in our sample is bounded between [0.09, 0.76]. 6
conditional mean function in equation (2) is correctly specified. We also allow the effect of business procedures on corruption to be endogenous by implementing a control function method discussed in Wooldridge (2005) (here referred to as the IV-QMLE regression). In the first stage, the explanatory variable suspected to be endogenous, i.e. business procedures, is regressed on the included explanatory variables (GDP per capita, its squared term, controls) and the excluded explanatory variables (British legal origin and latitude). In the second stage, the predicted value of the residuals from the first-stage regression is incorporated as an additional explanatory variable in equation (2). 3.4 Fractional logit model results Col. (5) in Table 3 reports the results for the fractional logit model for the specification in equation (2). We observe that the effect of business procedures is highly significant and substantially larger compared to the linear model (col. (1)) - 10% increase in business procedures leads to an approx. 4.5% increase in the corruption levels. As in the linear model, instrumenting for business procedures makes its impact on corruption larger in comparison to when it is treated as strictly exogenous. For example, according to the IV-QMLE results in col. (6), 10% increase in the business procedures increases corruption levels by around 7%. The predicted value of the first-stage residual included in equation (2) is found to be significant at the 5% level, thereby providing evidence that the business procedures term is endogenous (and hence needs to be instrumented). As in the linear regressions, PCI is positive and significant, while the PCI-squared term is negative and significant. In the presence of the control variables, the impact of business procedures remain unchanged as reported in cols. (7) and (8) of table 3. Among the controls, only the plurality and the inflation terms are found to be significant and associated with greater corruption in both the QMLE and the IV-QMLE regressions. Across all specifications, the fractional logit model explains between 78% - 81% of the variability in the corruption index. Thus accounting for the bounded nature of the dependent variable slightly improves the prediction power compared to the linear model. 4. Conclusion Do business regulations corrupt? Our results, based on a variety of linear and nonlinear estimation strategies for a large cross-section of countries suggest an answer in the affirmative. 7
Business procedures across countries are found to be consistently associated with greater corruption. Given that the nonlinear estimation strategy proposed in this paper yields larger estimates for business regulations in comparison to the linear estimates, it is likely that we will obtain more accurate estimates of the effects of various explanatory variables on corruption by treating the latter as a bounded dependent variable instead of a continuous variable. 8
Table 1. Summary statistics # of Mean Standard Minimum Maximum observations deviation Corruption 175 0.51 0.16 0.09 0.76 log(business 175 2.13 0.40 0.54 3.03 procedures) log(pci) 172 7.49 1.57 4.77 10.52 log(pci)-squared 172 58.66 24.25 22.71 110.72 9
Explanatory variables log(business procedures) OLS (1) 0.098*** (0.017) Table 2. Linear model Dependent variable- Corruption IV (2) 0.137*** (0.039) OLS (3) 0.08*** (0.02) IV (4) 0.156*** (0.05) log(pci) 0.181*** (0.036) 0.175*** (0.037) 0.18*** (0.05) 0.13** (0.05) log(pci)-squared -0.017*** (0.002) -0.016*** (0.002) -0.016*** (0.003) -0.013*** (0.004) Muslim 0.01 (0.008) 0.0003 (0.0002) Fractionalization -0.01 (0.02) -0.007 (0.03) Democracy measure -0.001 (0.014) -0.006 (0.008) Plurality 0.03* (0.02) 0.02 (0.015) Armed conflict -0.005 (0.005) -0.008 (0.008) log(import/gdp) -0.21* (0.11) -0.013 (0.018) log(averageinflation) 0.07* (0.04) 0.017* (0.01) Observations 172 170 123 123 Robust RESET 1.31 [0.25] 0.73 [0.39] 0.65 [0.45] 1.07 [0.30] Adjusted R-squared 0.78 0.77 0.79 0.77 Threshold per capita $204 $237 $279 $148 real income First-stageF-statistic 21.90 12.54 CLR-confidence interval 0.065, 0.219 0.062, 0.297 CLR-robust p-value 0.0004 0.002 J-statistic 0.55 [0.46] 1.16 [0.28] Note: ***, **, * denote significance at 1% or better, 5% or better, 10% or better respectively; clusterrobust standard errors in ( ) parentheses; p-values for test statistics within ([ ]) brackets 10
Explanatory variables log(business procedures) QMLE (5) 0.452*** (0.08) Table 3. Fractional logit model Dependent variable- Corruption IV-QMLE (6) 0.705*** (0.163) QMLE (7) 0.375*** (0.095) IV-QMLE (8) 0.701*** (0.208) log(pci) 0.877*** (0.164) 0.739*** (0.194) 0.899*** (0.205) 0.692*** (0.226) log(pci)-squared -0.078*** (0.011) -0.067*** (0.014) -0.079*** (0.014) -0.063*** (0.016) Muslim 0.001 (0.0009) Fractionalization -0.066 (0.104) Democracy measure -0.001 (0.035) Plurality 0.12* (0.065) Armed conflict -0.034 (0.032) log(import/gdp) -0.131** (0.066) log(averageinflation) 0.069* (0.037) Predicted first-stage -0.399** residuals (0.185) 0.001 (0.0009) -0.046 (0.104) -0.027 (0.034) 0.108* (0.065) -0.037 (0.03) -0.06 (0.082) 0.071* (0.038) -0.398* (0.234) Observations 172 123 123 123 Robust RESET 4.32 [0.23] 4.30 [0.23] 4.81 [0.19] 3.79 [0.28] Adjusted R-squared 0.78 0.79 0.81 0.81 Threshold per capita real income $275 $247 $295 $242 Note: ***, **, * denote significance at 1% or better, 5% or better, 10% or better respectively; cluster-robust standard errors in ( ) parentheses; p-values for test statistics within ([ ]) brackets 11
References Djankov, S., R. La Porta, F. Lopez-Di-Silanes, and A. Shleifer, (2002) "The regulation of entry" Quarterly Journal of Economics 117, 1-37. Djankov, S. (2009) "The regulation of entry: a survey" World Bank Research Observer 24, 183-203. Kaufman, D., A. Kraay, M. Mastruzzi (2007) "Governance matters VI: governance indicators for 1996-2006" World Bank Policy Research Working Paper 4280. Lichetti, M., D. Madani (2010) Draft note on governance and corruption World Bank, Investment Climate Department, Washington D.C. La Porta, R., F. Lopez-Di-Silanes, A. Shleifer, and R. Vishny (1999) "The quality of government" Journal of Law, Economics, and Organization 15, 222-279. Stigler, G. (1971) The theory of economic regulation Bell Journal of Economics and Management Science 2, 3-21. Svensson, J. (2005) "Eight questions about corruption" Journal of Economic Perspectives 19, 19-42. Treisman, D. (2007) "What have we learned about the causes of corruption from ten years of cross-national empirical research?" Annual Review of Political Science 10, 211-244. Wooldridge, J.M. (2005) "Unobserved heterogeneity and estimation of average partial effects", in Identification and inference for econometric models: Essays in Honor of Thomas Rothenberg, by D.W.K. Andrews, and J. H. Stock, Eds., Cambridge University Press, Cambridge, pp. 27-55. Wooldridge, J. M. (2010) Econometric analysis of cross section and panel data. Second ed., MIT Press, Cambridge, Mass. 12