Supporting Information

Transcription

1 A Supporting Information I Description of Covariates in Tables 1 & 2 Regarding the determinants of corruption in the literature, the most significant finding is that higher GDP per capita a proxy for economic development is associated with lower corruption levels, even when possible endogeneity is considered (see Treisman 2007). Economic development not only leads to the rationalization of economic and political systems, but it also contributes to the spread of education and literacy, all of which should help reduce corruption. However, GDP per capita is likely to be endogenous to both MNC activity and corruption and also highly correlated with MNC activity, posing a challenge of precisely estimating the coefficient of MNC activity in a relatively small sample. To deal with these challenges, I take advantage of a natural experiment created by China s reform and opening. Before the reform and opening, China had virtually no foreign investment. Thus, GDP per capita in the pre-reform era is unlikely to be affected by foreign investment. I utilize the data from two periods before China s reform and opening, and , as proxies for GDP per capita in and , respectively. 1 Utilizing GDP per capita before reform and opening helps mitigate the endogeneity and collinearity problem. 2 Furthermore, it is a very good proxy for GDP per capita from 1998 to The Pearson correlation between these two GDP per capita variables is The results are consistent when GDP per capita is lagged one period (see results in Table K). Additionally, I include total GDP and population to account for the effect of the size of the economy, as the scale of an economy is an important determinant of the total amount of rents in the market. 3 All data come from China Data Online. Corruption may rise with the size of government, as bigger government means that officials 1 This variable is logged to deal with the skewed distribution. 2 In the presence of collinearity, the estimated standard errors of the coefficients tend to be inflated. We can see in Table K that due to the collinearity problem, the standard errors of MNC activity s coefficients are much larger compared with those in Table 1. 3 Again, I use GDP before the reform and opening era as a proxy to deal with the possible endogeneity problem. Results are consistent when both GDP per capita and GDP variables are lagged one period. See results in Table K. 1

2 have more resources under their control, and thus more opportunities for extracting bribes. Empirical studies provide mixed results regarding the relationship between government size and corruption (see, e.g., Gerring and Thacker 2005; Montinola and Jackman 2002). Given government s significant role in steering economic growth in China, and also to be consistent with previous studies, I control for this variable in regressions. Government size is measured by the percentage of government expenditure to GDP and the share of employees in state-owned units including government agencies, Party organs, social organizations, and state-owned enterprises (SOEs). 4 Education is believed to help reduce corruption, since people s political participation and civic engagement are positively related with their education levels (see, e.g., Glaeser and Saks 2006). This variable is measured by the percentage of population aged six or above who have at least some college education. Scholars also suggest that the relatively high wages of public sector employees to private sector employees decrease incentives for corruption (Treisman 2000). When a public employee has a high-paying position, s/he has less incentive to jeopardize it by engaging in corruption. Thus, we should expect that the higher relative wages of public employees should be associated with less corruption. This variable is measured by the ratio of the average wage of the public sector to that of the private sector. Scholars have also found that gender impacts corruption (e.g., Swamy, Knack, Lee and Azfar 2001). Specifically, women tend to be more disciplined and less tolerant of corruption. Therefore, a government with more female employees should be less corrupt, though the results on gender have been disputed recently (e.g., Sung 2003). To be consistent with previous studies, I use the share of female employees in state-owned units to capture the influence of gender. Finally, I include a dummy variable for the second period of this study to account for the effect of the time trend, since corrupt funds tend to grow over time. The data used to measure education, government expenditure, size of public employees, and public employees relative wages come from China Statistical Yearbooks. Gender (the share of female public employees) is measured based on the data from China Labor Statistical Yearbooks. 4 Public employees in China are more broadly defined. For instance, managers and directors in SOEs and social organizations usually obtain the same status as government officials and sometimes are promoted into the government and Party systems. 2

3 Government expenditure is lagged one period to deal with possible endogeneity, and all other variables are averaged at the periods of and The descriptive statistics and the correlation matrix of explanatory variables are shown in Tables M and N. II Witnessed and Perceived Corruption Indices in Table 2 The China Module of 2008 Asian Barometer Survey (ABS) includes questions that ask inhabitants whether they have witnessed corruption, as well as their opinions about local corruption. I use the question, Have you or anyone you know personally witnessed an act of corruption or bribetaking by a politician or government official in the past year? 5, to construct an index of witnessed corruption. I generate a measure of perceived corruption using the question, How widespread do you think corruption and bribe-taking are in your local/municipal government? 6. For the index of witnessed corruption, I calculate the frequency of respondents who answered Never Witnessed in each province and then reverse this variable. 7 This measure is constructed based on respondents personal experience. Therefore, it is likely to capture the pervasiveness of corruption in local governments. For corruption perceptions, I take the means of respondents perceived corruption scores within each province and reverse the variable. 8 By doing so, I obtain an index of witnessed and perceived corruption for 26 provinces. 9 III Construction of the Instrumental Variable I take advantage of the spatial variation in China s levels of economic integration and use a weighted geographic distance as the instrumental variable for MNC activity. Following Jensen and Rosas (2007) and Larraín and Tavares (2004), I construct an instrumental variable for MNC 5 The choices are: 1. Witnessed; 2. Never witnessed; 8. Can t choose; 9. Decline to answer. 6 The choices are: 1. Almost everyone is corrupt; 2. Most officials are corrupt; 3. Not a lot of officials are corrupt; 4.Hardly anyone is involved; 5. Decline to answer. 7 The answer Witnessed could be biased because respondents might fear potential punishment. 8 By taking the means, we lose the information of respondents who declined to answer. Alternatively, I calculate the share of people who answered Almost everyone is corrupt and Most officials are corrupt. Empirical results are substantively the same. 9 The survey does not contain information for Gansu, Hainan, Xinjiang, and Tibet. I exclude Ningxia Province because it only has 8 observations due to the loss of information during the survey implementation (based on personal contact with ABS staff). 3

4 activity at the provincial level using the multiplicative inverse of bilateral geographic distance between each provincial capital and the five major economic centers Hong Kong, Seoul, Singapore, Taipei, and Tokyo around China, weighted by the five economies real GDP per capita. This instrumental variable is rooted in the gravity models of international trade and FDI flows (see Carr, Markusen and Maskus 2001; Loungani, Mody and Razin 2002; Markusen 1995). Countries tend to trade more with their neighbors, and FDI originating from wealthier countries is more likely to flow into geographically closer regions. That geographic distance affects trade patterns is one of the most robust empirical regularities in the economic literature. With regard to FDI, the knowledge-capital model suggests that efficiencyseeking (vertical) FDI tends to decrease with trade costs such as geographic distance, whereas market-seeking (horizontal) FDI increases with trade costs (see Carr, Markusen and Maskus 2001; Markusen 1995). In China before 1992, the Chinese government withheld access from marketoriented foreign firms, and thus all FDI was efficiency-seeking. Since 1992, when China started to open its market to foreign firms, efficiency-seeking FDI has still constituted a considerable portion of total FDI. For market-seeking FDI in China, foreign firms tend to be located in areas closer to their home countries, giving them advantages in importing parts and components from parent firms. For instance, Japanese and Koreans firms tend to concentrate in north and northeastern China, such as Beijing, Liaoning, Shandong, and Tianjian, while firms from Taiwan and Hong Kong operate mainly in southeastern coastal areas, such as Fujian, Guangdong, and Zhejiang. Therefore, geographic distance to the five economic centers is a good predictor of overall MNC activity in China. The inverse of bilateral geographic distance is weighted by these five economies real GDP per capita, which captures the fact that more developed countries tend to export more products and capital. On average, these five economies together account for approximately 59% of China s FDI inflows and 42% of trade from 1998 to One note is that most FDI from Taiwan for the period of this study actually took place via Hong Kong, since Taiwan had strict restrictions on Taiwanese firms direct investment in mainland China. 10 From the perspective of the standard 10 I thank one of the reviewers for pointing this out. 4

5 gravity model of foreign investment, the distance between Hong Kong and the mainland provinces should also serve to capture FDI from Taiwan. Therefore, when constructing the instrumental variable, I replace Taipei s distance to the mainland provinces with Hong Kong s and then weight them by Taiwan s real GDP per capita. 11 The instrument variable is constructed as follows: Z i,t = 5 j=1 1 dist i,j,t GDP per capita j,t (1) where i = 1, 2,..., N, j = 1,..., 5, and t = 1, 2. This instrumental variable measures the geographic closeness of China s provinces to the five economic centers. Note that, although geographic distance is constant, the assigned weights real GDP per capita of the five economies around China change over time. This captures the idea that if one economy, such as South Korea, has developed relatively quickly compared with others, it should be weighted more in the instrumental variable. Therefore, I expect that the closer a province is to the five economies, the more MNC activity it has. Geographic distance is calculated using the ArcGIS 9.3 program. Real GDP per capita data of the five economies between 1998 and 2007 are from Penn World Table. The first-stage regression results show that the instrumental variable is strong (see Table J). IV Sensitivity Analysis of the IV Exclusion Restriction A valid instrumental variable requires that it affects the dependent variable only through the endogenous variable. A potential violation of the exclusion restriction presents a challenge to the identification. In this case, we have reason to believe that exogenous geographic distance and the five economies real GDP per capita are unlikely to have a direct effect on China s provincial corruption, except through the channels of foreign investment and trade. Geographic closeness impacts corruption through the activities related to distance. It effects corruption primarily through the two major transnational economic activities: foreign investment and trade. How- 11 The empirical results are substantively the same when we use the distances between mainland provinces and Taiwan, or simply remove Taiwan from the construction of the instrumental variable. 5

6 ever, it is possible that geographic closeness could also influence corruption through labor migration/remittances, 12 tourists, or even exposure to foreign media. These kinds of effects should be at the margins, since cross-border labor movement and migration in China are still limited and the government highly restricts foreign media. In addition, Hong Kong, Japan, South Korea, Singapore, and Taiwan are all considered less corrupt than China. 13 Thus, we should expect a negative effect of geographic closeness on corruption through labor migration, tourists, and exposure to foreign media. If this is indeed the case, the 2SLS models tend to underestimate the positive coefficient of MNC activity on corruption. Nonetheless, we might still worry that geographic closeness impacts corruption in some unobservable ways, and hence the exclusion restriction could be violated. 14 As suggested by Conley, Hansen and Rossi (2012), I perform a sensitivity analysis to assess to what extent the empirical results are sensitive to the possible violation of the exclusion restriction. The model can be set up as follows: Y = Xβ + Zγ + ε (2) X = Zλ + υ (3) where Z is the (excluded) instrument (geographic closeness) for the endogenous variable of X (MNC activity); E(Xε) 0 and E[Zε] = 0. γ is a parameter measuring to what extent the exclusion restriction is satisfied. In a normal setup, the term Zγ does not appear in the structural equation (2). If the exclusion restriction holds, then γ 0. We can estimate the two equations using a normal 2SLS regression. If the exclusion restriction is violated, γ 0. Based on these two equations, we can conduct a sensitivity analysis using the prior knowledge about the magnitude of γ. 12 See Ahmed (2012) and Tyburski (2012) for how remittances may have an impact on corruption. 13 According to the TI, for instance, in 2007 the corruption perceptions indices (CPI) of Hong Kong, Japan, Korea, Singapore and Taiwan were 8.3, 7.5, 5.1, 9.3, 5.7 respectively. China s CPI was 3.5, far below the five economies CPIs. 14 One possibility is that geographic closeness may be associated with more overseas remittances that might undermine domestic governance (see Ahmed 2012). See also Tyburski (2012) for a counterargument. 6

7 Table A: Sensitivity Analysis of the IV Exclusion Restriction of Models 5-7 in Table 1 Dependent Variable Corrupt Funds Recovered Corrupt Funds Recovered Senior Cadres Disciplined Per Filed Case Per Capita Per 10,000 Public Employees (Model 5) (Model 6) (Model 7) Approach 1 Maximum γ for a Significant β Point Estimate β (MNC Activity) 95% Confidence Interval [0.00, 0.42] [0.00, 0.39] [0.00, 0.35] Approach 2 Prior Distribution of γ γ N(0, ) γ N(0, ) γ N(0, ) Point Estimate β (MNC Activity) % Confidence Interval [0.03, 0.88] [0.02, 0.56] [0.03, 0.41] In this case, we are more interested in the positive values of γ, because a negative γ will increase the positive slope of MNC activity. I focus on two of the approaches recommended by Conley, Hansen and Rossi (2012). The first is to specify a set of values for γ based on prior knowledge and obtain a union of confidence intervals for β. We want to see how large the γ could be so that we may still obtain a positive and significant coefficient of MNC activity at the 95% level. 15 The second approach is local-to-zero approximation, which adopts a large sample approximation and treats the uncertainty of γ as sample uncertainty to obtain an approximate distribution for β (Conley, Hansen and Rossi 2012, ). This approach allows the levels of confidence intervals of β to depend on the probabilities of observing specific values of γ. 16 Since we have reason to 15 We are more interested in positive values of γ. If γ is negative, a normal 2SLS regression in fact underestimates the coefficient of MNC activity. 16 In the case of a Gaussian prior for γ, βapprox N(β + Aµ γ, V 2SLS + AσγA 2 ), and A = (X Z(Z Z) 1 Z X) 1 (X Z), where β and V 2SLS are the point estimates and variance-covariance matrix obtained from a normal 2SLS regression; µ γ and σγ 2 are the mean and variance of the prior distribution of γ; X and Z are the 7

8 believe geographic closeness affects corruption primarily through the channels of inward FDI and trade, γ should be small even if the exclusion restriction is violated. Thus, I choose a normal distribution for γ with a mean of 0 and a variance of σ 2 to capture the fact that there is a high probability of observing γ = 0. I conduct the sensitivity analysis on Models 5-7 in Table 1 and present results in Table A. In Approach 1, we can see that γ can be as large as 0.27, 0.11, and in these three models, respectively, so that we still observe a positive and significant effect of MNC activity on corruption. The point estimates of MNC activity are 0.21, 0.19, and 0.17, respectively. After we subtract the direct effect of geographic closeness on corruption, all else being equal, one standard deviation increase in MNC activity will still raise corrupt funds recovered per filed case by 12,337 ($ 1,621), corrupt funds recovered per capita by 1.21 ($ 0.16), and senior cadres disciplined per 10,000 public employees by 1.19 units. The effects are substantively large. These results indicate that MNC activity affects corruption positively and significantly even if we allow for a considerable deviation from a perfect instrumental variable. Next, I draw γ from prior normal distributions in which σ= 0.21, 0.11, and for Models 5-7 in Table 1, respectively. Again, we can see that even if we take into account the uncertainty in γ, the 95% confidence interval of the coefficient of MNC activity is still above 0 in all three models. This finding confirms that the positive effect of MNC activity on corruption is not sensitive to the potential violation of the exclusion restriction. endogenous and instrumental variables, respectively. See, Conley et al. (2012, 264). 17 If these numbers were the true coefficients of geographic closeness in equation 2, one standard deviation increase in geographic closeness would raise recovered corrupt funds per filed case by 11,185 ($1,470), recovered corrupt funds per capita by 1.05 ($0.14), and senior cadres disciplined per 10,000 public employees by 1.03 units, when all other variables are held constant. These numbers would represent considerably large effects of geographic closeness on corruption. 18 If these three numbers were the true coefficients of geographic closeness, all else being equal, one standard deviation increase in geographic closeness would raise recovered corrupt funds per filed case by 10,910 ($1,434), recovered corrupt funds per capita by 1.22 ($ 0.16), and senior cadres disciplined per 10,000 public employees by 1.02 units. In fact, we allow γ to vary beyond σ. Approximately 95% of the γ values range between 2σ and 2σ. 8

9 V Foreign Presence, Market Concentration, and Firms ETCs My argument suggests that the entry and presence of MNCs contribute to rent creation by increasing market concentration, thereby giving rise to more rent-seeking activities. This argument implies that market concentration mediates the relationship between MNC activity and corruption. In the paper, I have shown that foreign presence of non-hmt firms is strongly associated with a high level of firms expenditures on ETCs our proxy for bribes and this relationship is indeed mediated by market concentration, measured by the weighted output ratio of the largest four firms. In Table B, I present the complete regression results that are omitted in the paper for space. To check the robustness of the findings, I use the weighted output ratio of the largest eight firms as an alternative measure of market concentration. We can see that the results are consistent. These results provide us with confidence that MNC activity contributes to rent creation through increasing market concentration, thereby giving rise to more rent-seeking activities. Yet, given the constraints of data availability, we are not able to explore the MNCs-concentration-corruption linkage at a more fine-grained industry level (preferably at the four-digit industry level). As an alternative exercise, I first regress market concentration levels (four- and eight-firm concentration ratios) on foreign preference or foreign presence of HMT and non-hmt firms and other covariates at the four-digit industry level. Then, I use the predicted values to construct market concentration levels at the two-digit level using industrial output as a weight and merge these two variables with the ETCs dataset. Finally, I regress firms ETCs on the two weighted market concentration variables based on the predicted values. We can see from Table C that foreign presence or foreign presence of HMT and non-hmt firms are all significant predictors of market concentration; the two weighted market concentration variables based on the predicted values are strongly and positively associated with firms expenditures on ETCs; furthermore, when the weighted market concentration variable is controlled for, foreign presence or foreign presence of HMT and non-hmt firms has no significant effect on firms ETCs. This exercise provides some preliminary results that both HMT and non-hmt firms are likely to contribute to rent creation through increasing market concentration, thereby leading to more rent-seeking activities. 9

10 Table B: Foreign Presence, Market Concentration, and Firms ETCs Model (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) DV CR4 CR8 ETCs ETCs ETCs CR4 CR8 ETCs ETCs ETCs Foreign Presence 0.08** 0.06** (0.03) (0.03) (0.14) (0.13) (0.14) Foreign Presence 0.19*** 0.16*** 0.50*** (Non-HMT Origin) (0.04) (0.04) (0.19) (0.18) (0.18) Foreign Presence (HMT Origin) (0.05) (0.05) (0.23) (0.22) (0.23) Concentration Ratio 1.34*** 1.28*** (Largest Four Firms) (0.18) (0.19) Concentration Ratio 1.27*** 1.20*** (Largest Eight Firms) (0.20) (0.20) Firm-Level Covariates SOEs -0.16*** -0.19*** -0.18*** -0.16*** -0.18*** -0.18*** (0.06) (0.06) (0.06) (0.06) (0.06) (0.06) Collective -0.29*** -0.30*** -0.30*** -0.30*** -0.30*** -0.30*** (0.07) (0.07) (0.07) (0.07) (0.07) (0.07) Private (0.24) (0.24) (0.24) (0.24) (0.24) (0.24) Foreign (Non-HMT) 0.23*** 0.21*** 0.22*** 0.22*** 0.21*** 0.21*** (0.06) (0.06) (0.06) (0.06) (0.06) (0.06) Foreign (HMT) (0.08) (0.08) (0.08) (0.08) (0.08) (0.08) Revenue (log) -0.43*** -0.44*** -0.44*** -0.43*** -0.44*** -0.44*** (0.01) (0.01) (0.01) (0.01) (0.01) (0.01) Employees (log) 0.29*** 0.29*** 0.29*** 0.29*** 0.29*** 0.29*** (0.02) (0.02) (0.02) (0.02) (0.02) (0.02) Out-of-Province Sales 0.48*** 0.48*** 0.48*** 0.48*** 0.48*** 0.48*** (0.04) (0.04) (0.04) (0.04) (0.04) (0.04) Sales to Government 0.89*** 0.89*** 0.89*** 0.88*** 0.88*** 0.88*** (0.16) (0.16) (0.16) (0.16) (0.16) (0.16) Sales to SOEs 0.48*** 0.46*** 0.47*** 0.48*** 0.46*** 0.47*** (0.05) (0.05) (0.05) (0.05) (0.05) (0.05) Years of Relationship 0.04*** 0.04*** 0.04*** 0.04*** 0.04*** 0.04*** (0.01) (0.01) (0.01) (0.01) (0.01) (0.01) Licenses 0.08*** 0.09*** 0.08*** 0.08*** 0.09*** 0.09*** (0.02) (0.02) (0.02) (0.02) (0.02) (0.02) Industry-Level Covariates Market Size -0.05*** -0.04*** *** 0.09*** -0.05*** -0.04*** *** 0.08** Continued... 10

11 Table B (cont.): Model (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) DV CR4 CR8 ETCs ETCs ETCs CR4 CR8 ETCs ETCs ETCs (0.00) (0.00) (0.03) (0.03) (0.03) (0.00) (0.00) (0.03) (0.03) (0.03) Growth Rate -0.04** -0.04** -0.21*** -0.16** -0.17** -0.04** -0.04** -0.20*** -0.16** -0.16** (0.02) (0.02) (0.08) (0.07) (0.07) (0.02) (0.02) (0.07) (0.07) (0.07) Asset Intensity (log) 0.13*** 0.11*** 0.20*** *** 0.11*** 0.17*** (0.01) (0.01) (0.05) (0.06) (0.06) (0.01) (0.01) (0.05) (0.06) (0.06) Scale Economies (log) 0.05*** 0.03*** -0.10*** -0.20*** -0.17*** 0.05*** 0.03*** -0.10*** -0.20*** -0.17*** (0.01) (0.01) (0.03) (0.03) (0.03) (0.01) (0.01) (0.03) (0.03) (0.03) Government Help 0.06** 0.05** -0.25** -0.28*** -0.27*** 0.06*** 0.06*** -0.24** -0.27*** -0.27*** (0.02) (0.02) (0.10) (0.10) (0.10) (0.02) (0.02) (0.10) (0.10) (0.10) Tax Burden *** 1.68*** 1.73*** *** 1.70*** 1.75*** (0.04) (0.03) (0.41) (0.40) (0.40) (0.04) (0.03) (0.41) (0.40) (0.40) Province-Level Covariates GDP per Capita (log) -0.09*** -0.06** 0.31*** 0.40*** 0.36*** -0.10*** -0.07*** 0.28** 0.38*** 0.34*** (0.03) (0.03) (0.12) (0.12) (0.12) (0.03) (0.03) (0.12) (0.12) (0.12) GDP *** -0.59*** -0.54*** *** -0.55*** -0.49*** (0.06) (0.05) (0.18) (0.18) (0.18) (0.06) (0.04) (0.18) (0.18) (0.18) Population (log) -0.08*** -0.04** 0.28*** 0.37*** 0.31*** -0.08*** -0.05** 0.26*** 0.36*** 0.30*** (0.03) (0.02) (0.09) (0.09) (0.09) (0.02) (0.02) (0.09) (0.09) (0.09) Constant 1.81*** 1.48*** -3.75** -5.98*** -5.39*** 1.96*** 1.62*** -3.08** -5.56*** -4.93*** (0.45) (0.36) (1.52) (1.52) (1.55) (0.43) (0.35) (1.51) (1.54) (1.57) Random Effects Var(Cons Province) 0.00*** 0.00*** 0.01*** 0.02*** 0.02*** 0.00*** 0.00*** 0.01*** 0.02*** 0.02*** (0.00) (0.00) (0.01) (0.01) (0.01) (0.00) (0.00) (0.01) (0.01) (0.01) Var(Cons Industry) 0.02*** 0.01*** 2.75*** 2.75*** 2.75*** 0.02*** 0.01*** 2.75*** 2.75*** 2.75*** (0.00) (0.00) (0.04) (0.04) (0.04) (0.00) (0.00) (0.04) (0.04) (0.04) Var(Resids) 0.10*** 0.07*** 0.07*** 0.10*** 0.07*** 0.07*** (0.02) (0.01) (0.01) (0.02) (0.01) (0.01) N ,359 11,359 11, ,359 11,359 11,359 No. of provinces Note: Multilevel models with varying intercepts at both the industry- and province-level; Standard errors in parentheses; ** significant at 5%; *** significant at 1%. 11

12 Table C: Foreign Presence, Market Concentration, and Firms ETCs (Alternative Specifications) Model (1) (2) (3) (4) (5) (6) (7) (8) DV CR4 ETCs CR8 ETCs CR4 ETCs CR8 ETCs Foreign Presence 0.04*** *** 0.12 (0.01) (0.13) (0.00) (0.13) Foreign Presence 0.01** ** (HMT Origin) (0.01) (0.22) (0.01) (0.22) Foreign Presence 0.07*** *** 0.33 (Non-HMT Origin) (0.01) (0.18) (0.01) (0.18) Concentration Ratio 2.25*** 2.14*** (Largest Four Firms) (0.36) (0.36) Concentration Ratio 3.22*** 3.06*** (Largest Eight Firms) (0.53) (0.53) Firm-Level Covariates SOEs -0.18*** -0.17*** -0.18*** -0.17*** (0.06) (0.06) (0.06) (0.06) Collective -0.30*** -0.30*** -0.30*** -0.30*** (0.07) (0.07) (0.07) (0.07) Private (0.24) (0.24) (0.24) (0.24) Foreign (Non-HMT) 0.22*** 0.22*** 0.21*** 0.21*** (0.06) (0.06) (0.06) (0.06) Foreign (HMT) (0.08) (0.08) (0.08) (0.08) Revenue (log) -0.43*** -0.43*** -0.43*** -0.43*** (0.01) (0.01) (0.01) (0.01) Employees (log) 0.29*** 0.29*** 0.29*** 0.29*** (0.02) (0.02) (0.02) (0.02) Out-of-Province Sales 0.48*** 0.48*** 0.48*** 0.48*** (0.04) (0.04) (0.04) (0.04) Sales to Government 0.88*** 0.88*** 0.87*** 0.87*** (0.16) (0.16) (0.16) (0.16) Sales to SOEs 0.46*** 0.47*** 0.46*** 0.47*** (0.05) (0.05) (0.05) (0.05) Years of Relationship 0.04*** 0.04*** 0.04*** 0.04*** (0.01) (0.01) (0.01) (0.01) Licenses 0.08*** 0.08*** 0.08*** 0.08*** Industry-Level Covariates (0.02) (0.02) (0.02) (0.02) Continued... 12

13 Table C (cont.): Model (1) (2) (3) (4) (5) (6) (7) (8) DV CR4 ETCs CR8 ETCs CR4 ETCs CR8 ETCs Market Size -0.10*** 0.10*** -0.07*** 0.09*** -0.10*** 0.09** -0.07*** 0.08** (0.00) (0.03) (0.00) (0.03) (0.00) (0.03) (0.00) (0.03) Growth Rate 0.07*** *** *** *** 0.06 (0.00) (0.05) (0.00) (0.05) (0.00) (0.05) (0.00) (0.05) Asset Intensity (log) 0.07*** -0.15*** 0.05*** -0.14*** 0.07*** -0.15*** 0.05*** -0.14*** (0.00) (0.03) (0.00) (0.03) (0.00) (0.03) (0.00) (0.03) Scale Economies (log) ** ** ** ** (0.00) (0.07) (0.00) (0.07) (0.00) (0.07) (0.00) (0.07) Government Help -0.24** -0.25** -0.24** -0.24** (0.10) (0.10) (0.10) (0.10) Tax Burden 1.99*** 1.97*** 2.00*** 1.99*** (0.41) (0.41) (0.41) (0.41) Province-Level Covariates GDP per Capita (log) *** ** *** ** (0.01) (0.11) (0.01) (0.11) (0.01) (0.11) (0.01) (0.11) GDP -0.08*** -0.35** -0.08*** *** *** (0.02) (0.18) (0.02) (0.18) (0.02) (0.18) (0.02) (0.18) Population (log) *** 0.02*** 0.27*** *** 0.02*** 0.26*** (0.01) (0.09) (0.01) (0.09) (0.01) (0.09) (0.01) (0.09) Constant 1.08*** -6.55*** 0.91*** -6.94*** 1.11*** -5.98*** 0.92*** -6.32*** (0.14) (1.53) (0.12) (1.56) (0.14) (1.54) (0.12) (1.57) Random Effects Var(Con Province) 0.00*** 0.01*** 0.00*** 0.01*** 0.00*** 0.01*** 0.00*** 0.01*** (0.00) (0.01) (0.00) (0.01) (0.00) (0.01) (0.00) (0.01) Var(Cons Industry) 0.00*** 0.08*** 0.00*** 0.08*** 0.00*** 0.07*** 0.00*** 0.07*** (0.00) (0.02) (0.00) (0.02) (0.00) (0.01) (0.00) (0.02) Var(Resids) 0.02*** 2.75*** 0.01*** 2.75*** 0.02*** 2.75*** 0.01*** 2.75*** (0.00) (0.04) (0.00) (0.04) (0.00) (0.04) (0.00) (0.04) N Industry-Level Groups 1, , , , Province-Level Groups Note: Multilevel models with varying intercepts at both the industry- and province-level; Standard errors in parentheses; ** significant at 5%; *** significant at 1%. 13

14 V.1 2-Digit Industry-Level Covariates in Tables 3, B, & C Four-Firm Concentration Ratio (CR4): output share of the largest four firms. Eight-Firm Concentration Ratio (CR8): output ratio by the largest eigth firms. Foreign Presence: output share of foreign firms. Foreign Presence (HMT Origin): output share of foreign firms originating in Hong Kong, Macao, and Taiwan. Foreign Presence (Non-HMT Origin): output share of foreign firms originating in countries other than Hong Kong, Macao, and Taiwan. Market Size: log of total value added. Growth Rate: growth rate of total output from 2001 to Asset Intensity: log of fixed assets per employee. Scale Economies: log of the average size of the largest firms that account for 50% of total output in the industry. Government Help: industry median of a linear combination of the percentages of the officials in government departments (taxation, public security, environment, and labor and social) who are perceived to contribute to the development of the company. Tax Burden: industry median of firms tax rates (total taxes divided by total sales). All the variables are constructed using data from the China Industrial Enterprises Database (2003), with the exceptions of Government Help and Tax Burden, which use the province-industry median of firm-level observations from the same firm survey of the ETCs data. All variables are lagged one year to deal with possible endogeneity. 14

15 V.2 Firm-Level Covariates in Tables 3, B, & C SOE: a dummy variable coded 1 if firms are registered as state-owned and the share of state capital is equal to or greater than 50%, and 0 if otherwise. Collective: a dummy variable coded 1 if firms are registered as collective-owned and the share of collective capital is equal to or greater than 50%, and 0 if otherwise. Private: a dummy variable coded 1 if firms are registered as privately owned and the share of private capital is equal to or greater than 50%, and 0 if otherwise. Foreign (HMT): a dummy variable coded 1 if firms are registered as foreign-owned (Hong Kong, Macao, and Taiwan) and the share of foreign capital is equal to or greater than 50%, and 0 if otherwise. Foreign (Non-HMT): a dummy variable coded 1 if firms are registered as foreign-owned (other than Hong Kong, Macao, and Taiwan) and the share of foreign capital is equal to or greater than 50%, and 0 if otherwise. Mixed: a dummy variable coded 1 if firms are not coded as SOEs, collective, private, or foreign, and 0 if otherwise. Revenue (log): log of total business income (core business income + other business income) in Employees (log): log of total employees in Out-of-Province Sales: a dummy variable coded 1 if firms sell to other provinces, and 0 if otherwise. Sales to government: proportion of sales to government in Sales to SOEs: proportion of sales to SOEs in Years of Relationship: total years of firms relationship with their major clients and suppliers. 15

16 Licenses: number of licenses and registrations (permanent and renewable annually) required for firms. CEO Pay (log): log of CEO annual income. This variable is not directly observed. The survey reports workers average wage as well as the ratio of CEO s annual income to the mid-level managers and the ratio of the mid-level managers annual income to the ordinary employees. CEO annual income is calculated using the product of workers average wage and the two ratios. Interaction with government: number of days a firm s general manager (GM) or vice GM spends on government assignments and communications. Government Appointed General Manager: a dummy variable coded 1 if the firm s general manager is appointed by the government, and 0 if otherwise. 16

17 VI Does Country of Origin Matter? In the paper, I have shown that provinces with more MNC activity are associated with a higher level of corruption. In the analyses, all FDI originating from different countries are treated the same. In China, a substantial part of FDI is from Hong Kong, Macao, and Taiwan (HMT). HMT investors are supposed to be more familiar with business practices in China and have more connections with local governments. Thus, we may wonder whether HMT investors behave differently from others. It could be the HTM FDI that actually drives the positive relationship between MNC activity and corruption. To examine whether HTM FDI really makes a difference, I disaggregate FDI into HTM and non-htm origins. Data on FDI by country of origin is collected from provincial statistical yearbooks. To be consistent with previous model specifications, I again average both variables for two five-year periods: and Since data on FIE trade by partner countries at the provincial level is not available, MNC activity is measured by inward FDI from either the HTM or non-htm region as a percentage of GDP. The dependent variables are the three corruption measures based on filed cases (corrupt funds recovered per filed case in Models 1 & 2; corrupt funds recovered per capita in Models 3 & 4; senior cadres disciplined per 10,000 public employees in Models 5 & 6). Given the high collinearity between these two FDI variables (r=0.91), they enter into the regression equation separately. As discussed before, OLS regression tends to underestimate the coefficient of MNC activity. If we observe a positive and significant coefficient in OLS regressions, we are confident that the actual coefficient will be larger when endogeneity is accounted for. Table D presents the results. We can see that both HTM and non-htm FDI are positively associated with the three measures of corruption. The substantive effects of these two variables are not significantly different. For instance, in Models 1 and 2 in Table D, one standard deviation increase in HTM FDI will raise corrupt funds recovered per filed case by 12,214 ($1,605), while the same change in non-htm FDI will do so by 11,853 ($1,556). These results suggest that FDI originating from both HMT and non-hmt regions are associated with corruption in China. In Table E, I report results when the weighted geographic distance is used as an instrument variable for both HTM and non-htm 17

18 Table D: Hong Kong, Macao, and Taiwan FDI vs Other FDI (OLS) Model (1) (2) (3) (4) (5) (6) Recovered Corrupt Funds Recovered Corrupt Funds Senior Cadres Disciplined per Filed Case per Capita per 10,000 Public Employees HMT FDI 0.20** 0.21** 0.22*** (0.10) (0.09) (0.05) Non-HMT FDI *** (0.11) (0.10) (0.06) GDP per Capita (log) *** -0.78*** (0.47) (0.51) (0.40) (0.45) (0.20) (0.24) GDP *** 5.79*** (3.93) (3.86) (2.91) (2.81) (1.44) (1.68) Population *** -1.13*** (0.68) (0.65) (0.55) (0.52) (0.29) (0.34) Government Expenditure (% of GDP, log) (0.39) (0.44) (0.36) (0.41) (0.20) (0.23) Public Employees *** 13.05*** (5.48) (5.93) (4.52) (4.70) (2.64) (3.22) Schooling (0.50) (0.49) (0.43) (0.43) (0.27) (0.32) Public Employees ** 0.36 Relative Wages (0.35) (0.38) (0.29) (0.32) (0.15) (0.20) Gender (5.14) (5.39) (4.17) (4.62) (2.94) (2.88) Time (0.55) (0.57) (0.49) (0.51) (0.23) (0.28) Bureaucratic Integration (0.11) (0.11) (0.08) (0.08) (0.04) (0.05) Four Municipalities *** 0.78*** (Dummy Variable) (0.32) (0.32) (0.31) (0.30) (0.12) (0.14) Trust in Courts (0.38) (0.39) (0.37) (0.37) (0.18) (0.18) Constant (3.53) (3.71) (2.85) (3.06) (1.93) (2.07) N R Note: Robust standard errors in parentheses; ** significant at 5%; *** significant at 1%. FDI. After endogeneity and selection bias are accounted for, we observe an even larger effect of both types of FDI (especially non-htm FDI) on corruption. One caveat of these results is that FDI data by country of origin at the provincial level is less reliable compared with the data on total FDI, since the reporting procedure has changed over time and some provinces do not distinguish between foreign capital in general and FDI. Therefore, the empirical results are more suggestive rather than conclusive. Given the limited availability of FDI 18

19 data by country of origin, we are not able to disaggregate FDI further into origins of OECD vs. non-oecd or developed vs. developing countries. Table E: Hong Kong, Macao, and Taiwan FDI vs Other FDI (2SLS) Model (1) (2) (3) (4) (5) (6) Recovered Corrupt Funds Recovered Corrupt Funds Senior Cadres Disciplined per Filed Case per Capita per 10,000 Public Employees HMT FDI 0.31*** 0.17** 0.27*** (0.09) (0.07) (0.05) Non-HMT FDI 0.41*** 0.23*** 0.33*** (0.11) (0.08) (0.09) GDP per Capita (log) *** -0.99*** (0.40) (0.46) (0.30) (0.33) (0.17) (0.24) GDP *** 5.98*** (3.39) (3.56) (2.25) (2.30) (1.19) (1.43) Population *** -1.29*** (0.65) (0.68) (0.46) (0.48) (0.26) (0.38) Government Expenditure 0.86*** 1.20*** (% of GDP, log) (0.30) (0.34) (0.27) (0.28) (0.17) (0.21) Public Employees 10.18** *** 13.59*** (5.04) (5.28) (4.32) (4.28) (2.19) (2.72) Schooling (0.46) (0.47) (0.38) (0.37) (0.22) (0.28) Public Employees *** 0.36 Relative Wages (0.28) (0.37) (0.24) (0.29) (0.13) (0.21) Gender (4.15) (4.37) (3.47) (3.39) (2.46) (2.50) Time (0.49) (0.50) (0.44) (0.44) (0.17) (0.23) Bureaucratic Integration (0.09) (0.11) (0.06) (0.07) (0.03) (0.03) Four Municipalities 0.58** *** 0.75*** (Dummy Variable) (0.29) (0.30) (0.27) (0.27) (0.10) (0.13) Trust in Courts (0.32) (0.35) (0.31) (0.31) (0.15) (0.16) Constant ** (3.06) (3.43) (2.23) (2.44) (1.65) (1.88) N R F Statistic (Excluded Instrument) P rob > F Note: Robust standard errors in parentheses; ** significant at 5%; *** significant at 1%. 19

20 VII Potential Outliers Given the relatively small sample size, we may worry that the results could be driven by some statistical outliers. To formally check the existence of outliers, I calculate the studentized residual for each observation in the sample for the second-stage regressions in Models 5-7 of Table 1. Observations with a studentized residual larger than 2 or smaller than -2 are potential outliers. I use the leverage measure and Cook s distance to assess the influence of these observations on the regression model. Table F lists observations with the absolute value of the studentized residual larger than 2, along with the values of the leverage measure and Cook s distance. It shows that all values of the leverage measure in the three models are actually below the threshold 3 k/n. 19 Regarding Cook s distance, scholars have used different thresholds of identifying influential observations. For instance, Heiberger and Holland (2004) and Weisberg (1985) suggest 1 as the threshold, while Fox (1991) proposes a more conservative criterion 4/(N k 1). Table F indicates that all Cook s distances in Model 5 do not exceed 4/(N k 1) and none of the Cook s distances in Models 6 & 7 is larger than 1. I treat observations with the absolute value of the studentized residual larger than 2 and the leverage greater than 3 k/n or Cook s distance exceeding 4/(N k 1) as influential outliers. 20 To check whether the results are driven by these influential outliers, I re-estimate Models 6 & 7 21 in Table 1 by dropping these influential observations and report results in Table G. We can see that MNC activity is still strongly and positively associated with corruption. The magnitude and significance level of its coefficients are not statistically different from those in Table N is the sample size and k is the number of explanatory variables. 20 Results are consistent, if we treat all observations with a studentized residual outside the ±2 range as outliers. 21 No influential outlier is identified in Model 5 of Table 1. 20

21 Table F: Studentized Residuals, Leverage, & Cook s Distances DV Province Time Studentized Cook s Leverage Period Residual Distance Recovered Corrupt Funds Jiangsu per Filed Case (Model 5) Shanxi Shanxi Recovered Corrupt Funds Shanxi per Capita (Model 6) Shanxi Senior Cadres Disciplined Hebei per 10,000 Public Employees Inner Mongolia (Model 7) Inner Mongolia Ningxia Note: observations with a studentized residual outside the ±2 range, along with the values of the leverage and Cook s distance. 21

22 Table G: Results without Potential Outliers (2SLS) Model (1) (2) DV Recovered Corrupt Funds Senior Cadres Disciplined per Capita per 10,000 Public Employees MNC Activity 0.28*** 0.22*** (Factor Score) (0.10) (0.08) GDP per Capita (log) *** (0.31) (0.16) GDP *** (2.10) (1.41) Population *** (0.46) (0.29) Government Expenditure (% of GDP, log) (0.27) (0.15) Public Employees 10.00*** 2.48 (3.88) (2.56) Schooling (0.35) (0.18) Relative Wages ** (0.22) (0.15) Gender (3.05) (1.84) Time *** (0.28) (0.15) Bureaucratic Integration (0.07) (0.04) Four Municipalities 0.53** 0.86*** (Dummy Variable) (0.22) (0.13) Trust in Courts (0.17) (0.15) Constant (2.40) (1.14) N R F Statistic (Excluded Instrument) P rob > F Note: Robust standard errors in parentheses; ** significant at 5%; *** significant at 1%. 22

23 VIII Pooled Time-Series Cross-Sectional Analysis As discussed in the paper, corruption cases take time to detect and the whole process of investigation and prosecution may last several years. Guo (2008) finds that the average latency period of corruption cases, referring to the time it takes to detect the corruption case since a public official commits a corrupt act for the first time, increases from about three years in the late 1990s to five years, or even more, in the early 2000s. Additionally, the actual number of corruption cases investigated each year may depend on leaders political willingness and considerations. Thus, the annual number reported by the procuratorate at the provincial level may not well reflect each year s actual level of corruption. The temporal variation in the dataset could be misleading. Here, I experiment with time-series cross-sectional data and check whether the results are consistent with those obtained from the two-period panel data. All right-hand time-variant variables are lagged one year to deal with possible endogeneity. Simple OLS regressions with both fixed province and year effects show that MNC activity is positively and strongly associated with the three different measures of corruption based on filed cases (see Table H). 23

24 Table H: Time-Series Cross-Sectional Analysis ( ) Model (1) (2) (3) (4) (5) (6) (7) (8) (9) DV Corrupt Funds Recovered Corrupt Funds Recovered Senior Cadres Disciplined per Filed Case per Capita per 10,000 Public Employees MNC Activity 0.27*** 0.41*** 0.32*** 0.25*** 0.35*** 0.28** 0.18*** ** (Factor Score) (0.08) (0.10) (0.11) (0.08) (0.11) (0.13) (0.07) (0.11) (0.12) GDP per Capita (log) ** (0.26) (0.64) (1.25) (0.29) (0.84) (1.39) (0.21) (0.35) (0.42) GDP *** (0.20) (0.27) (0.27) (0.20) (0.29) (0.28) (0.15) (0.20) (0.20) Population ** ** (0.41) (1.74) (1.44) (0.37) (1.81) (1.45) (0.30) (1.15) (1.38) Government Expenditure 0.78*** *** *** (% of GDP, log) (0.19) (0.43) (0.45) (0.18) (0.49) (0.53) (0.15) (0.38) (0.38) Public Employees *** (2.67) (2.97) (4.56) (2.43) (3.23) (4.71) (2.23) (2.67) (2.34) Schooling -0.26*** (0.07) (0.09) (0.13) (0.09) (0.10) (0.13) (0.05) (0.09) (0.11) Relative Wages (0.45) (0.65) (0.64) (0.41) (0.63) (0.66) (0.21) (0.25) (0.28) Time ** ** 0.08 (0.03) (0.07) (0.11) (0.03) (0.10) (0.13) (0.03) (0.05) (0.05) Bureaucratic Integration (0.04) (0.04) (0.04) (0.04) (0.04) (0.05) (0.04) (0.05) (0.04) Four Municipalities *** (Dummy Variable) (0.32) (0.21) (0.21) Constant *** ** (2.08) (6.38) (10.92) (2.30) (8.13) (12.16) (1.68) (3.47) (3.92) N R Number of provid Fixed Province Effects No Yes Yes No Yes Yes No Yes Yes Fixed Year Effects No No Yes No No Yes No No Yes Note: Robust standard errors in parentheses. ** significant at 5%; *** significant at 1%. 24

25 IX Other Supplementary Materials Table I: MNC Activity and Corruption (OLS) Model (1) (2) (3) (4) (5) (6) (7) MNC Activity 0.35*** 0.36*** 0.35*** 0.23** 0.23** 0.22** 0.18** (Factor Score) (0.10) (0.09) (0.09) (0.10) (0.09) (0.08) (0.07) GDP per Capita (log) ** (0.42) (0.40) (0.40) (0.47) (0.44) (0.38) (0.20) GDP ** (3.82) (3.49) (3.50) (3.67) (3.40) (2.67) (1.69) Population (0.76) (0.73) (0.62) (0.67) (0.59) (0.51) (0.37) Government Expenditure 0.93** 0.89** 1.08*** * (% of GDP, log) (0.40) (0.39) (0.37) (0.37) (0.38) (0.36) (0.20) Public Employees 11.77** 11.47** 13.55*** *** (5.67) (5.38) (4.54) (6.07) (5.25) (4.25) (3.15) Schooling (0.32) (0.32) (0.29) (0.49) (0.45) (0.41) (0.25) Relative Wages ** (0.32) (0.31) (0.31) (0.30) (0.34) (0.31) (0.19) Gender (3.72) (3.99) (3.70) (4.45) (4.59) (4.07) (2.64) Time (0.29) (0.29) (0.29) (0.43) (0.46) (0.42) (0.23) Bureaucratic Integration (0.09) (0.10) (0.08) (0.05) Four Municipalities 0.86*** 0.65** *** (Dummy Variable) (0.27) (0.31) (0.29) (0.16) Trust in Courts (0.24) (0.25) (0.23) (0.15) Constant (3.20) (3.04) (2.90) (3.48) (3.24) (2.77) (1.66) N R Note: Robust standard errors in parentheses; ** significant at 5%; *** significant at 1%. The dependent variable (DV) is recovered corrupt funds per filed case in Models 1-5, per capita recovered corrupt funds in Model 6, and senior cadres disciplined per 10,000 public employees in Model 7. 25

26 Table J: First Stage Regressions of Models in Table 1 (OLS) Model (1) (2) (3) (4) (5) (6) (7) Geographic Closeness 1.15*** 1.17*** 1.15*** 1.08*** 1.11*** 1.11*** 1.12*** (0.17) (0.17) (0.17) (0.15) (0.17) (0.17) (0.17) GDP per Capita (log) 1.01*** 1.06*** 0.95*** ** 0.75** 0.76** (0.28) (0.28) (0.30) (0.37) (0.36) (0.36) (0.36) GDP (2.99) (2.89) (3.06) (3.05) (2.93) (2.93) (3.07) Population (0.56) (0.57) (0.57) (0.61) (0.66) (0.66) (0.67) Government Expenditure -1.13*** -1.11*** -1.11*** -1.20*** -1.17*** -1.17*** -1.16*** (% of GDP, log) (0.22) (0.22) (0.24) (0.26) (0.28) (0.28) (0.29) Public Employees (4.63) (4.61) (4.51) (4.60) (4.73) (4.73) (4.90) Schooling (0.21) (0.21) (0.21) (0.37) (0.38) (0.38) (0.39) Relative Wages 0.81*** 0.84*** 0.82*** (0.30) (0.31) (0.30) (0.29) (0.30) (0.30) (0.30) Gender (3.21) (3.22) (3.41) (3.55) (3.60) (3.60) (3.62) Time (0.22) (0.22) (0.22) (0.28) (0.28) (0.28) (0.30) Bureaucratic Integration (0.07) (0.10) (0.10) (0.10) Four Municipalities (Dummy Variable) (0.20) (0.24) (0.24) (0.24) Trust in Courts (0.26) (0.26) (0.26) (0.28) Constant -5.65*** -5.91*** -5.39*** -4.52** -4.53** -4.53** -4.60** (1.76) (1.67) (1.73) (2.08) (1.92) (1.92) (1.90) N R F Statistic (Excluded Instrument) P rob > F Note: Standard errors in parentheses ** significant at 5%; *** significant at 1% 26