ENDOGENEITY OF BRAZILIAN CORPORATE GOVERNANCE QUALITY DETERMINANTS

Transcription

1 ENDOGENEITY OF BRAZILIAN CORPORATE GOVERNANCE QUALITY DETERMINANTS Alexandre M. da Silveira Assistant Professor of Finance and Accounting at the School of Economics, Management and Accounting of the University of São Paulo (FEA/USP), Ricardo P. C. Leal* Professor of Finance at the Coppead Graduate School of Business, Federal University of Rio de Janeiro (Coppead/UFRJ), André L. Carvalhal-da-Silva Assistant Professor of Finance at the Coppead Graduate School of Business, Federal University of Rio de Janeiro (Coppead/UFRJ), Lucas A. B. de C. Barros Assistant Professor of Finance at the Center of Applied Social Sciences of the Mackenzie Presbyterian University and at the School of Economics, Management and Accounting of the University of São Paulo (FEA/USP), November, 2008 * Corresponding author. Ricardo Leal thanks grants from the Rio de Janeiro Research Foundation (FAPERJ) and of the National Scientific Research Council (CNPq).

2 ENDOGENEITY OF BRAZILIAN CORPORATE GOVERNANCE QUALITY DETERMINANTS STRUCTURED ABSTRACT Category Purpose Design and Methodology Findings Research Limitation and Implications Practical Implications What is original? Research Paper This paper investigates the determinants and the evolution of voluntarily adopted firm-level corporate governance practices in Brazil from 1998 to 2004 using broad corporate governance scores. We employ a robust panel-data procedure that accounts for the main sources of endogeneity to a very representative panel of Brazilian firms over a 6-year period. We address the endogeneity that arises from the simultaneous determination of the quality of corporate governance practices, the dependent variable, and possibly several firm attributes that are commonly employed as the determinants of such practices and are supposedly independent. Specifically, theoretical arguments and empirical evidence strongly suggest that the quality of corporate governance practices may influence some of the variables commonly used as its determinants just as much as they may be influenced by them. Firm-level corporate governance practices are steadily improving but there is much room for improvement. Heterogeneity has increased. Voluntarily adhering to new stricter listing requirements is associated positively with improvements in firm-level corporate governance practices. Reducing or not using non-voting shares improves corporate governance practices. We found no clear evidence of the influence of other potential determinants of the quality of corporate governance, such as growth prospects, firm size, firm value, and ownership structure. Thus, we doubt previous findings that suggest a causal relationship from value and ownership to corporate governance practices because value and ownership seem to be determined endogenously. Policies directed to reduce the use of non-voting shares should be implemented. Creating strict listing requirements that may be adopted voluntarily by firms could be a feasible solution to improve the quality of corporate governance practices in emerging market countries. Firms in an emerging market that find that issuance in the US became too expensive or demanding may offer a substitute listing environment with credible requirements to foreign investors. Premium listings may partially compensate emerging market exchanges for their loss of trading to major markets. We examine the evolution of the voluntary adoption of corporate governance practices in Brazil from 1998 through 2004 while most studies use cross-section samples over one or a few years. Further, this is one of a few papers to analyze the impact of ownership structure on the quality of corporate governance practices by segregating control and cash flow rights. Key-words: Corporate Governance, Corporate Governance Index, Ownership Structure, Reverse Causality. 2

3 Introduction What motivates firms to improve voluntarily their corporate governance practices? We employ a corporate governance (CG) practices score based on practices recommended by the Brazilian codes of best practices as a proxy for the actual quality of CG of the firm. We investigate an extensive sample of listed firms in Brazil, where there is no comply or explain requirement. This line of research presents a new perspective because it seeks the determinants of CG practices while a large number of academic papers deal with the impact of CG practices on the value of the firm. We also use a longer time period (1998 to 2004) while most studies use cross-section samples over one or a few years and a more robust and little used econometric procedure designed to eliminate or substantially reduce the problems generated by the endogeneity among the variables that plague the typical panel analysis research. We believe that our research helps to understand what leads firms to voluntarily improve their CG practices. Theoretical arguments and empirical evidence strongly suggest that the quality of CG practices may influence some of the variables commonly used as its determinants just as much as they may be influenced by them. Studies that employ large multi-country samples that include only the largest firms in each country provide broad insights but may lack depth. For example, Klapper and Love (2004) include data of only 24 Brazilian firms from 2000 and Lins (2003) includes data of 59 Brazilian firms from Single-country studies may be narrow in scope but may offer the potential of deep and detailed insights by way of larger and more representative samples and possibly better measured variables. Ownership structures and CG practices of firms may be seen as an equilibrium response to the legal environment in which they operate (see La Porta, Shleifer, Lopez-De-Silanes, and Vishny (1998); La Porta, Shleifer, Lopez-De-Silanes, and Vishny, 2000; Claessens, Djankov, Fan, and Lang, 2002; and Beck, Demirguç-Kunt, and 3

4 Levine 2001). Himmelberg, Hubbard, and Palia (1999) and Himmelberg, Hubbard, and Love (2004) argue that the level of protection offered to investors has two components: an external one, related to the legal environment in which the firm operates, and an internal one, related to the type of activity carried out by the firm and to other observable characteristics of the firm. Himmelberg, Hubbard, and Love (2004, p. 2) also say that investor protection refers collectively to those features of the legal, institutional, and regulatory environment and characteristics of firms or projects that facilitate financial contracting between inside owners (managers) and outside investors. Thus, it is possible that firms within a country offer different levels of protection to their investors because of their operational specificities and their motivation to voluntarily adopt better CG practices. Motivation varies because the potential gains from better CG may not be the same for all firms. This article belongs to a body of literature in corporate governance that investigates why firms within the same contractual environment choose voluntarily different levels of quality of their corporate governance practices. Potential Determinants of Firm-Level Quality of Corporate Governance Practices We review selected papers that we see as the closest to our work and that suggest the potential determinants of CG practices we employ. Klapper and Love (2004) indicate three main potential determinants of firm-level CG quality: the utility of CG, the nature of the operations of the firm and its size. They hypothesize that firms with greater needs to finance their growth opportunities would derive more utility from the adoption of good CG practices. The expropriation of investors may depend on the nature of the operations of the firm, according to Himmelberg, Hubbard, and Palia (1999, p. 358). Firms that invest mostly in tangible assets face more difficulty to divert or steal those assets because they are easily monitored and hard to be channeled to other uses. On the other hand, firms investing 4

5 predominantly in intangible assets have more incentives to adopt good CG practices because they may need to signal that they do not intend to divert or use them inadequately. The size of the firm is the third potential determinant of firm-level CG. Large firms may adopt good CG practices in order to mitigate the agency costs of free cash flow. Firms with large free cash flows may not use them in the best interest of all shareholders. Small firms may grow faster than large firms and will need more external financing. Both large and small firms could have incentives to adopt good CG practices voluntarily (Klapper and Love, 2004). Durnev and Kim (2005) analyzed potential determinants of the quality of firm-level CG practices. They investigated how certain attributes of firms influence their choice of CG practices and how they interact with the legal environment in which they operate. Their theoretical model yields three predictions confirmed by their empirical evidence: (1) the three main attributes that make firms adopt good CG practices are their growth opportunities, their need for external funding, both debt and equity, and their concentration of ownership; (2) the market value of firms increases with good CG practices; and (3) the adoption of good CG practices is most relevant in countries where investor legal protection is poor. Anand, Milne, and Purda (2006) empirically examined the adoption of recommended CG guidelines in Canada and found an increasing voluntary adoption and convergence of good CG practices over time. Their results suggest that the presence of a majority shareholder or an executive blockholder is negatively associated with good CG practices. Significant investment opportunities and research and development expenditures are positively associated to a proxy for board quality. The authors argue that the need to obtain external financing is a major motivation to adopt good CG practices. Cash flow rights are defined as the amount of the total equity capital held by a shareholder. Control or voting rights are as the amount of the voting equity capital held by a shareholder. The ratio between the proportions of control rights to cash flow rights is called 5

6 the wedge. The wedge reflects the leverage achieved by the controlling shareholders by way of indirect control structures, cross-holdings, and non-voting shares. The average wedge for the controlling shareholder in Brazil was approximately 2. This asymmetry may be a source of conflicts of interest between controlling and minority shareholders. Lins (2003) reports an average wedge of 2.68 for 719 firms in 18 countries. The ownership structure deserves more discussion (see Claessens, Djankov, Fan, and Lang, 2002). A large concentration of control rights by controlling shareholders or managers entitles them to have little need to secure the votes of minority shareholders in order to manage the firm. Firms with such dominant shareholders may be less likely to adopt good CG practices voluntarily. Alternatively, firms with dominant shareholders could adopt good CG practices voluntarily to reduce the perception of its minority shareholders that they are more exposed to expropriation. Thus, the influence of the concentration of control rights over the presence of good CG practices is ambiguous. A large concentration of cash flow rights by dominant shareholders may align their interests with those of minority shareholders. The probability of expropriation of minority shareholders could be reduced by such alignment of interests. The presence of good CG practices could be positively associated to a large concentration of cash flow rights by the dominant shareholders. Alternatively, if a large concentration of cash flow rights by the dominant shareholders is seen as a substitute for good CG practices then there would be no need to adopt more good CG practices voluntarily. Thus, the influence of the concentration of cash flow rights over the presence of good CG practices may be ambiguous as well. We employ a dummy (N2NM) to indicate if the firm is listed in the most demanding premium listings of the São Paulo Stock Exchange (Bovespa). Bovespa has four listing levels. The listing levels differ in terms of the requirements for liquidity, disclosure and accounting practices, conflict resolution practices, and CG practices. Companies may migrate 6

7 between listing levels voluntarily. The traditional listing level has no requirements. Level 1 is the first premium level and requires a minimum liquidity and certain corporate disclosure and accounting practices. Level 2 encompasses all Level 1 requirements and requires many additional CG practices, dispute resolution through arbitration, and the use of international accounting standards. Novo Mercado (NM) is the most demanding premium listing and includes all Level 1 and 2 requirements plus a ban on non-voting shares. Levels 2 and 3 ADR programs require that firms comply with Sarbanes-Oxley Law and file a 20-F form annually. We employ the ADR23 dummy for the presence of level 2 or level 3 ADR programs. We also use control dummies for the type of controlling shareholders (family, foreign, state-owned, and control blocks). The operational definitions of all variables are presented in the Appendix. Methodology Sample and corporate governance practices score CG practices can be measured by using the wedge as in Claessens, Djankov, Fan, and Lang (2002). However, the wedge may capture only part of CG practices. Recent literature has employed CG scores that consider many different aspects both objectively and qualitatively. Klapper and Love (2004) use scores from a questionnaire that involves qualitative evaluations. Durnev and Kim (2005) use a transparency and disclosure index computed by Standard and Poor s (using 98 0 or 1 questions) that is largely objective. Gompers, Ishii and Metrick (2003) use a CG index built from provisions followed by the Investors Responsibility Research Center (IRRC). Black, Jang and Kim (2006) use only the objective portion of a Korean Stock Exchange survey. We measure the quality of CG practices using an index (CGI) designed by Leal and Carvalhal-da-Silva (2007). They include only features that can be objectively assessed without the need to interview or survey interested parties. Their CGI was validated by a 7

8 group of market practitioners, including lawyers, regulators, and CG professionals, as well as the many academics participating in the Inter-American Development Bank project that sponsored its creation. The CGI is computed from the responses to twenty-four binary and objective questions that probe for the presence of practices that can be verified by means of publicly available information and filings. Each positive answer adds one point to the score. The final score for each firm ranges from 0 (worst) to 24 (best). The CGI was computed for every other year during the sample period because of the time consuming work to compute the index manually for every year and because of its relative stability from one year to the next. We use an equally weighted version of the CGI because it is easier to reproduce. Indexes constructed in other studies have also followed this method (see La Porta, Shleifer, Lopez-De-Silanes, and Vishny (1998), Gompers, Ishii and Metrick (2003), Black, Jang and Kim (2006), among others). Experimentation with different weighing schemes did not substantially change the ranking of firms. For lack of space, the reader can see the complete CGI questionnaire and supporting information about it, including the theoretical evidence supporting each question, in Leal and Carvalhal-da-Silva (2007), available in the Internet at The sample is comprised of financial and non-financial firms listed at Bovespa and does not include firms with: (1) incomplete or unavailable information; (2) negative book value of assets; (3) negative book value of common equity; and (4) that do not present a minimal level of trading liquidity. The final sample consists of approximately 200 firms in each year and 823 firm-year observations that represent about 90% of the stock market capitalization in Brazil in the years of 1998, 2000, 2002, and This is a much larger and representative sample than those employed in previous multi-country studies and it encompasses a period of 6 years. All data, except the CGI, were collected using the DIVEXT System of the Brazilian 8

9 Securities Commission (CVM, Comissão de Valores Mobiliários) and the Economatica data base. Econometric method and model From the determinants identified in our brief literature review, we initially estimated the model in Equation 1 using the panel data regression procedure developed by Blundell and Bond (1998) known as the System GMM estimator (GMM-Sys, for short). In Equation 1, i represents the firm and t the year (with t 1998, 2000, 2002, 2004 ). it is the random error term from the i-th firm in the t-th year. The term ui captures unobserved firm characteristics that are time-invariant or that did not vary within the sample period. OWN represents the set of ownership variables that are not used simultaneously: 1VDIR, 3VDIR, 1TDIR, 3TDIR, WEDGE1, WEDGE3, and VOTE. They are the proportions of control and cash flow rights and the wedge of the major and of the 3 major shareholders. VOTE is an additional variable with the proportion of voting shares in the total equity capital of the firm, not of the largest shareholders, because the use of non-voting shares is widespread and because non-voting shares are a major reason for a wedge in Brazil. VALUE is represented either by the Tobin s Q or the price-to-book ratio (PBV). Firm age (AGE) and total leverage (LEVER) are additional control variables. We did not use industry dummies in the analysis because they show no time variation and the time invariant firm characteristics are explicitly modeled byu i. CGI GROWTH TANG SIZE ADR23 N2NM OWN it 1 it 2 it 3 it 4 it 5 it 6 it 3 3 VALUE PROFIT AGE LEVER TYPE YEAR u 7 it 8 it 9 it 10 it l lit m mt i it l 1 m 1 Equation 1 Based on the previous discussion, we expect that 1, 4, 5, 7, 8 0 and that 2 0. Because the signs of the relationship between the size of the firm and between the proportion 9

10 of control and cash flow rights and the CGI, representing the quality of firm-level CG practices, are ambiguous, we do not have an expected sign for the coefficients 3 and 6. The remaining coefficients relate to control variables. For lack of space, we omit the descriptive statistics and other information about the independent variables. This information is available from the authors. The estimation of Equation 1 by Ordinary Least Squares (OLS) or even by more sophisticated traditional panel data regression methods, such as Random Effects or Fixed Effects, is likely to be inappropriate because these methods fail to address some very important potential sources of endogeneity. By endogeneity we mean here any phenomena that creates some degree of correlation between the error term ( it or it ui, depending on the method employed) and one or more regressors, thus violating a key assumption of these methods and rendering inconsistent the coefficient estimator. Such correlation can stem from measurement errors, from relevant omitted variables that are correlated with the regressors or from feedback effects (discussed in more detail below). However, of paramount importance for our model in Equation 1, we want to handle the endogeneity that arises from the simultaneous determination of the CGI (the dependent variable) and its potential determinants, as we discuss next. Theoretical arguments and empirical evidence strongly suggest that the dependent variable that represents the quality of CG practices can influence some of the (supposedly independent) variables we use as regressors, just as much as these regressors may influence the quality of CG practices. For example, it is easy to hypothesize that the CG practices of a firm may substantially impact its level of financial leverage, profitability, market value, and ownership structure. Firms that already practice good CG would find it easier and less costly to issue ADRs or to list in one of the premium Bovespa lists. 10

11 In our empirical analysis, we treat the two-way causality as a simultaneous determination problem that creates endogeneity in the form of contemporaneous correlation between our variables and the error term. We assume that E x 0, where. it it E is the expectation operator and x it can be any of the regressors in Equation 1 with the exception of the year and type of controlling shareholder dummies and the age of the firm, which are assumed to be strictly exogenous. The GMM-Sys is one of the most suitable procedures to deal with endogenous covariates in panel data in many empirical corporate finance exercises like ours, in which there are no reliable exogenous external instruments available. The GMM-Sys estimator allows the efficient use of appropriate lags of the potentially endogenous regressors as their own instrumental variables. By appropriate we mean any lagged value of the endogenous variable that is known or assumed to be uncorrelated with the error term of the model. Specifically, our identification strategy is as follows: we assume that the shocks that affect the CGI are uncorrelated with past values of our endogenous regressors, i.e. E x 0 it s it, with 1 s. Thus, we allow for current shocks to affect current and future values of x. In this setting, we are able to address the simultaneous determination problem (allowing for 0 E x ) and also feedback effects running from the response variable to it it the regressors. Feedback effects arise when E x 0 it s it, with 1 s. That would happen if, for example, a new regulation (e.g. change in corporate law) led the firm to contemporaneously adopt better CG practices. Clearly, this new regulation (a shock to CGI captured in part by it ) might subsequently influence the evolution of other corporate variables and such feedback mechanism is ruled out by traditional fixed effects procedures. By employing the GMM-Sys procedure we are able to isolate unobserved heterogeneity ( u i 11

12 in Equation 1) while relying on sequential exogeneity assumptions, which are much less restrictive (i.e. more realistic) than the strict exogeneity assumption. Controlling for unobserved firm heterogeneity is important because it makes it less likely that our results are driven by omitted variables (e.g. any fixed firm characteristic that correlates with both its CGI score and the covariates). In our regressions, we eliminate u i by applying to all endogenous variables, including the CGI, the first-difference transformation (e.g. CGI it CGIit CGIit 1). Thus, the identifying assumptions outlined above translate to E x 0, with s 2 it s it. These orthogonality conditions (i.e. non-correlation between selected lags of the covariates and the first-differenced error) are then used for estimating the regression coefficients. The GMM-Sys also allows us to explore the complementary orthogonality condition defined by E x u it 1 it i 0. In the presence of simultaneous determination and feedback effects, the above orthogonality conditions can only be valid E it it 1 0, i.e. it if is not autocorrelated, which implies that it must display negative first-order autocorrelation and no second-order autocorrelation. Therefore, for assessing the plausibility of our identification strategy, we first run autocorrelation tests. As an additional diagnostic tool, we check the overall plausibility of the set of instruments used (i.e. the appropriate lags of our endogenous regressors) running the well-known Sargan/Hansen test of overidentifying restrictions. Finally, our identification strategy might also mitigate any spurious correlation originated by measurement errors in the regressors provided that these errors are not perfectly autocorrelated. In sum, GMM-Sys permits us to rely on much weaker assumptions than those necessary to produce the correct coefficient estimates in most regression methods such as the wellknown random or fixed effects panel methods. Thus, the procedure we use here may be regarded as more robust than those applied in most of the related previous research. It can 12

13 also be argued that dealing with the problem of reverse causality employing single-equation GMM-Sys is generally better than resorting to simultaneous equations methods because GMM-Sys neither relies on the correct specification of a potentially complex system of equations nor on instruments derived from questionable exclusion restrictions. See Blundell, Bond, and Windmeijer (2000) for a detailed discussion of this method and a comparison among several different regression strategies in empirical exercises similar to ours. Analysis of Results Evolution of corporate governance practices in Brazil. The descriptive statistics of the CGI from 1998 to 2004 scaled to a 0-10 range are presented in Table 1 and four main conclusions can be drawn: 1. CGI scores are increasing at a slow pace. The average CGI score increases from 4.16 in 1998 to 5.00 in Conventional mean comparison tests not reported here show that these differences are statistically significant. The changes from 2000 to 2002 and from 2002 to 2004 are statistically significant and suggest an improvement in CG practices. 2. The average CGI score of 5.0 out of 10.0 can be considered low because the CGI questionnaire comprised several questions that check for the presence of CG practices that are easy to adopt. We conclude that there is still much room for improvement. For example, see questions 2, 3, 9, 13, 14, and 24 in Leal and Carvalhal-da-Silva (2007). 3. We observed an increasing dispersion of the CGI scores over time. The standard deviation of the CGI increases from 2.07 in 1998 to 2.88 in The increasing inter-quartile range conveys the same information. The creation of the premium listing levels in 2000 where all recent listings took place, plus the migration from the traditional listing to the premium listing is the cause for the larger CGI score 13

14 dispersion in recent years. The Brazilian market is becoming two-tiered and is in a transitional phase. ************** TAKE IN TABLE 1 ***************** Determinants of firm-level CG quality. We ran diagnostic tests on an initial set of GMM-Sys regressions. The Sargan/Hansen tests could not reject the null of validity of the set of instruments used at the 5% level. Nevertheless, there was a pronounced error autocorrelation that cast doubt on the validity of our instruments and suggest that Equation 1 fails to capture all relevant systematic information about the behavior of CGI. Clearly, the diagnostics of our initial runs indicated that a dynamic version of Equation 1 is more appropriate. We then proceeded to estimate the same initial specifications with one difference: the inclusion in the set of regressors of the first lag of CGI ( CGI it 1 ). This dynamic term proved to be highly significant across specifications and it did completely capture the observed pattern of autocorrelation in the error term 1. The estimates from the initial static model, without the lagged dependent variable, are available upon request. The results are shown in Table 2. Our choice of instrumental variables is based on sequential exogeneity assumptions that allow for the simultaneous determination of these variables and the CGI (i.e. reverse causality) and for feedback effects (i.e. past shocks influencing current values of the variables). Specifically, we use the lags x it s, with s 2, and xit 1. The instrumented variables ( x ) include the CGI (in our dynamic models) and all the regressors shown in Equation 1, with the exception of the year and type of controlling shareholder dummies and 1 We ran the Arellano-Bond tests m1 and m2 (respectively, for first-order and second-order autocorrelation of the first-differenced errors) after GMM-Sys regressions and the heteroscedasticity robust autocorrelation test for short panels suggested by Wooldridge (2002, p ) after OLS regressions. In all static models, the null of no autocorrelation was rejected at the 5% level with both tests. When we estimate a dynamic version of Equation 1 we are left with too few time periods to run the Arellano-Bond tests m2. Therefore, we rely on the Wooldridge test to infer that there is no remaining error autocorrelation after the inclusion of the lagged dependent variable (in all specifications we cannot reject the null of no autocorrelation at the conventional significance levels). 14

15 the age of the firm, which are assumed to be strictly exogenous. As a robustness check, we ran regressions with different sets of instruments that are compatible with our identification strategy and pass the diagnostic tests. We do this by using deeper lags of some covariates (e.g. s 3), which allows us to check the sensitivity of our inferences to changing the instrument set (i.e. reducing the number of instruments). These variations yield qualitatively similar results. ************** TAKE IN TABLE 2 ***************** The first notable result of our regressions is the magnitude and significance of the variable ADR23. As hypothesized, firms that choose to issue ADRs levels 2 or 3 tend to have a substantially higher CGI score. This result is quite robust across specifications. Similar results, though less strong, are observed for our dummy for listing in Level 2 or Novo Mercado premium listings of Bovespa. When we estimate the models excluding ADR23, the estimated coefficients for N2NM increase considerably. At first glance, these results may look trivial because issuing ADRs or adhering to premium listings at Bovespa requires a commitment to many good CG practices. But one should notice that we explicitly accounted for a possible reverse causality between these variables and CGI. Thus, we can interpret our results as indicative that the CG practices of firms that adhere to more stringent CG rules improve after they do it. This is not obvious because it could be the case that higher CGI score firms choose to issue ADRs or to list in a Bovespa premium listing to signal to the market that they already have good CG practices. Firms need to provide costly signals to investors to be credible. Issuing ADRs and adhering to premium listings are types of costly signaling. The second strong result relates to the year dummies that are significant, both statistically and economically, in all cases. The dummy for the first year was not included in the dynamic version of our model depicted in Table 2 because of the lagged value of the 15

16 dependent variable. The interpretation of these coefficients is straightforward. They clearly reflect the steady improvement of the overall CG practices throughout the sample period. They show that there was a clear trend to adopt good CG practices between 1998 and 2004 and that this trend is actually more important to explain the cross-sectional variation in our sample than most of our potential CG determinants. The use of non-voting shares is still widespread in Brazil. The law used to allow 2 nonvoting shares for each voting share and the current proportion is one to one. The proportion of voting shares in the equity capital (VOTE) is positively and significantly associated with the CGI. This is an important result because the policy trend to replace non-voting shares with voting shares is supported by the data. Table 2 shows no clear evidence of the impact of the ownership structure variables on the CGI. The same can be said about the other potential determinants in our regressions. These results are somewhat different than the ones obtained using traditional methods that do not account for endogeneity. For instance, OLS estimates that are not reported here suggest that the size of the firm and its Tobin s Q are important explanatory variables, net of the effects of all other regressors. It is likely that these findings from traditional methods reflect spurious correlations. Therefore, our GMM-Sys results cast doubt on the conclusions offered by some of the previous empirical research in the field. Leal and Carvalhal-da-Silva (2007) use OLS regressions for each year (1998, 2000, and 2002) and find results that are similar to ours. They do not include ownership and market value variables and find that firm size is significant in each year. Their Novo Mercado dummy is used only in 2002, because this market did not exist prior to that, and is significant. Their ADR dummy is significant only in 2002 but presents a positive coefficient in each year. Thus, despite considering their different modeling and their potential endogeneity problems, their results are qualitatively similar to ours. 16

17 Besides using alternative instrument sets, we test the robustness of our results in three ways. First, we rerun our regressions after excluding subsets of the covariates whose coefficients shown in Table 2 are not statistically significant. Second, we rerun the models shown in Table 2 after excluding possible outliers (observations with the largest associated standardized residuals). Finally, although our models do not suffer from severe multicollinearity problems (as attested by the low mean Variance Inflation Factor and the correlation matrix) we rerun regressions after excluding covariates with correlation coefficient (with any other covariate) above 0.6 in absolute value. These exercises (available from the authors) yield qualitatively similar results. Concluding Remarks Firm-level CG practices are improving at a slow pace in Brazil. The voluntary adoption of good CG practices is increasing the dispersion of CGI scores rather than inciting convergence, leading to more CG practices heterogeneity throughout the period. Decreasing the number of or eliminating non-voting shares appeared as a strong determinant of good CG practices. Adherence to rigorous listing requirements by way of Levels 2 and 3 ADRs or the two most demanding of the premium listings at Bovespa are positively related to the CGI scores and it appears that firms that upgrade their listing environment subsequently improve their CG quality. There is no clear evidence that the ownership structure influences CG quality but for the trivial fact that ownership structure itself can be regarded as a CG mechanism. This same conclusion applies to other potential determinants of the quality of CG practices such as growth opportunities, size, and value. We cast doubt on most potential determinants of CG practices, including value and ownership structure measures, claiming that they are likely to be endogenous. We employ a robust GMM-Sys panel data procedure 17

18 that explicitly addresses the main sources of endogeneity and our results suggest that ignoring this problem will lead to incorrect coefficient estimates and conclusions. Some cautionary notes are warranted in spite of our efforts. Our results should be interpreted with caution because our variables, including the CGI, may not be suitable representations of their associated theoretical constructs. Moreover, not even the most sophisticated econometric method can completely assure that researchers have controlled all sources of spurious correlation. If the results from Brazil are valid for other emerging countries, policies directed to reduce the use of shares with inferior voting should be implemented. Stock exchanges may consider creating premium listing requirements that may be voluntarily adopted by firms. Many times, changing legislation is not an easy and quick task and this voluntary private contracting arrangement with the stock exchange could be a feasible solution to improve the quality of CG. Actually, many exchanges in emerging markets have been implementing this inspired by the successful experience of Bovespa. Firms that find that issuance in the US became too expensive or demanding may offer a substitute listing environment with credible requirements to investors, especially foreigners. Premium listings may partially compensate emerging market exchanges for their loss of trading to major markets. References Anand, A., Milne, F., and Purda, L. (2006), Voluntary Adoption of Corporate Governance Mechanisms, Queen s University, Department of Economics, Working Paper Beck, T., Demirguç-Kunt, A., and Levine, R. (2001), Law, Politics and Finance, The World Bank Policy Research Working Paper n

19 Black, B., Jang, H. and Kim, W. (2006), Does Corporate Governance Affect Firms Market Values? Evidence from Korea, Journal of Law, Economics and Organization, Vol. 22, Blundell, R. and Bond, S. (1998), Initial Conditions and Moment Restrictions in Dynamic Panel Data Models, Journal of Econometrics, Vol. 87, No. 1, pp Blundell, R., Bond, S., and Windmeijer, F. (2000), Estimation in Dynamic Panel Data Models: Improving on the Performance of the Standard GMM Estimator, in Baltagi, B. H. (Ed.), Nonstationary Panels, Panel Cointegration, and Dynamic Panels, Advances in Econometrics, Vol. 15, Elsevier Science, JAI, Amsterdam, pp Claessens, S., Djankov, S., Fan, J., and Lang, L. (2002), Disentangling the Incentive and Entrenchment Effect of Large Shareholdings, Journal of Finance, Vol. 57, No. 6, pp Durnev, A. and Kim, H. (2005), To Steal or Not to Steal: Firm Attributes, Legal Environment, and Valuation, Journal of Finance, Vol. 60, No. 3, pp Gompers, P., Ishii, J. and Metrick, A. (2003), Corporate Governance and Equity Prices, Quarterly Journal of Economics, Vol. 118, No. 1, pp Himmelberg, C., Hubbard, G., and Love, I. (2004), Investor Protection, Ownership and the Cost of Capital, The World Bank Policy Research Working Paper Himmelberg, C., Hubbard, G., and Palia, D. (1999), Understanding the Determinants of Managerial Ownership and the Link Between Ownership and Performance, Journal of Financial Economics, Vol. 53, No. 3, pp

20 Klapper, L. and Love, I. (2004), Corporate Governance, Investor Protection, and Performance in Emerging Markets, Journal of Corporate Finance, Vol. 10, No. 5, pp La Porta, R., Shleifer, A., Lopez-De-Silanes, F. and Vishny, R. (1998), Law and Finance, Journal of Political Economy, Vol. 106, No. 6, pp La Porta, R., Shleifer, A., Lopez-De-Silanes, F. and Vishny, R. (2000), Investor Protection and Corporate Governance, Journal of Financial Economics, Vol. 58, No. 1, pp Leal, R. P. C. and Carvalhal-Da-Silva, A. L. (2007), Corporate Governance and Value in Brazil (and in Chile), in Chong, A., and Lopez-de-Silanes, F. (Eds.), Investor Protection and Corporate Governance Firm Level Evidence Across Latin America, Palo Alto: Stanford University Press, pp , Working paper version available at Lins, K. (2003), Equity ownership and firm value in emerging markets, Journal of Financial and Quantitative Analysis,Vol. 38, No. 1, pp Wooldridge, J. Econometric analysis of cross section and panel data. Cambridge, Massachusetts: MIT Press,

21 Table 1 Summary statistics for the Corporate Governance Index (CGI) (scaled from 0-10) Mean Standard-Dev Corporate Governance Index (CGI) Scaled from 0-10 range Minimum o Quartile Median o Quartile Maximum N (sample)

22 Table 2 Determinants of firm-level corporate governance Dynamic GMM-Sys regressions The Corporate Governance Practices Index (CGI) is the dependent variable. Operational definitions of all explanatory and control variables are presented in the Appendix. Figures in parentheses are t statistics. ***, **, and * denote statistical significance at 1%, 5%, and 10%, respectively. The coefficients were estimated with one-step GMM-Sys assuming that all regressors are endogenous, with the exception of the year and type of controlling shareholder dummies and firm age, which are assumed to be strictly exogenous. We compute firmclustered standard-errors, which are robust to arbitrary forms of heterocedasticity and autocorrelation of the error term (additionally, by including year dummies we control for a possible cross-sectional dependence of the error term). At the bottom of the table we report: the number of observations, the number of instruments used by the GMM estimator, the F-test statistic (with degrees of freedom and p-value, respectively, in parentheses), and the Sargan/Hansen (test of overidentifying restrictions) statistic (with degrees of freedom and p-value, respectively, in parentheses). Corporate Governance Practices Index (CGI) CGI (t-1) GROWTH TANG SIZE ADR23 N2NM 1VDIR 3VDIR 1TDIR 3TDIR WEDGE1 WEDGE3 VOTE Q PBV PROFIT AGE LEVER FOR SBH FAM (1) (2) (3) (4) (5) (6) (7) (8) 0.445** (2.42) (-0.31) (-1.52) (0.05) 3.070** (2.42) (1.44) (0.31) 2.415* (1.96) (1.46) (-0.42) (0.10) (0.49) (0.11) (0.53) (-0.04) 0.430** (2.37) (0.04) (-1.58) (-0.59) 3.565*** (2.94) (1.64) (-0.29) 2.927** (2.35) (1.13) (0.10) (-0.12) (0.32) (0.04) (0.27) (-0.42) 0.450** (2.35) (-0.35) (-1.28) (0.03) 3.089*** (2.61) (1.40) (0.38) (0.08) (1.55) (1.47) (-0.44) (0.10) (0.54) (0.05) (0.49) (-0.05) 0.407** (2.21) (0.12) (-1.00) (-0.56) 3.626** (2.56) (1.61) (-0.16) (-0.28) (1.28) (0.95) (0.40) (-0.09) (0.51) (-0.08) (0.22) (-0.45) 0.362* (1.97) (-0.38) (-1.49) (0.12) 3.346** (2.27) 5.391* (1.77) (0.31) 2.381* (1.90) (0.79) (0.10) (0.48) (-0.15) (0.76) (1.11) (0.30) 0.388** (2.35) (0.37) (-1.46) (-0.23) 3.316*** (2.63) 4.522* (1.78) (-0.43) 2.922** (2.38) (-0.09) (0.67) (0.04) (0.24) (0.42) (0.77) (-0.13) 0.359* (1.93) (-0.36) (-1.13) (0.09) 3.388** (2.47) 5.426* (1.80) (0.28) (0.22) (1.48) (0.83) (0.10) (0.51) (-0.21) (0.84) (1.13) (0.31) 0.383** (2.22) (0.50) (-1.04) (-0.16) 3.126** (2.06) (1.64) (-0.51) (-0.02) (1.61) (-0.19) (0.82) (-0.06) (0.40) (0.42) (0.72) (-0.17) 22

23 YEAR2-0.86*** -0.85*** -0.86*** -0.83*** -1.02*** -0.99*** -1.02*** -0.98*** (-2.68) (-2.87) (-2.82) (-2.81) (-3.30) (-3.77) (-3.61) (-3.88) YEAR3-0.80*** -0.78*** -0.81*** -0.80*** -0.95*** -0.96*** -0.95*** -0.96*** (-2.87) (-3.21) (-2.92) (-3.37) (-4.16) (-5.51) (-4.21) (-5.48) Num. Obs Num. of instruments F Sargan/Hansen (17;222) (<0.001) (21;0.30) (17;222) (<0.001) (21;0.36) (18;222) (<0.001) (22;0.36) (18;222) (<0.001) (22;0.48) (17;222) (<0.001) (21;0.30) (17;222) (<0.001) (21;0.27) (18;222) (<0.001) (22;0.35) (18;222) (<0.001) (22;0.43) 23

24 Appendix Summary of research variables and their operational definitions Code Name of Variable Operational definition Dependent Variable CGI Corporate Governance Quality Corporate Governance Index proposed by Leal and Carvalhal-da-Silva (2007), based on binary questions, and scaled to a 0-24 range. Ownership Structure Variables (OWN) 1VDIR Control rights - largest shareholder Percentage of common stock (voting capital) owned directly by the largest shareholder *. 1TDIR Cash flow rights - largest shareholder Percentage of total shares (voting and non-voting capital) owned directly by the largest shareholder. 3VDIR Control rights three largest shareholders Percentage of common stock (voting capital) owned directly by the three largest shareholders. 3TDIR Cash flow rights three largest Percentage of total shares (voting and non-voting capital) WEDGE1 WEDGE3 VOTE ADR23 shareholders Wedge between control rights and cash flow rights largest shareholder Wedge between control rights and cash flow rights three largest shareholders N2NM Participation in Bovespa s premium listings owned directly by the three largest shareholders. Difference between the percentage of voting capital and total capital owned directly by the largest shareholder (voting capital minus total capital). Difference between the percentage of voting capital to total capital owned directly by the three largest shareholders (voting capital minus total capital). Percentage of voting shares to Ratio of voting capital to total capital. total shares Other Determinants of Corporate Governance Quality Presence of ADRs Programs Dummy variable equal to 1 if the firm issues ADR Level Levels 2 or 3** 2 or Level 3. Dummy variable equal to 1 if the firm is listed in the two most demanding lists: Level 2 or New Market at Bovespa. LEVER Leverage Ratio of total (non-equity) liabilities to total assets at year-end. See Durnev and Kim (2005). GROWTH Growth/investment opportunities Cumulative percentage variation of net revenues over the last three years. See Klapper and Love (2004) and Durnev and Kim (2005). Q Tobin s Q Estimated as the ratio of market value to book value of assets. Market value of assets is computed as the market value of equity plus book value of assets minus book value of equity at year-end. The numerator market value of equity was computed directly by the ECONOMATICA database as the most liquid share class (voting or non-voting) market price times the total number of shares (voting and non-voting). See Durnev and Kim (2005). PBV Price-to-Book-Value Market value of shares divided by their book value. See Durnev and Kim (2005). PROFIT Return on Assets Estimated as the ratio of operating income to total assets at year-end. See Durnev and Kim (2005). TANG Tangibility of assets (proxy for the nature of operations) Total fixed assets divided by net operational revenues. See Himmelberg, Hubbard, and Palia (1999). SIZE Firm size Natural logarithm of book value of total assets in thousands of Brazilian reais at year-end. See Klapper and Love (2004). Additional Control Variables 24

25 TYPE (FOR, SBH, FAM, SOE) Type of controlling shareholder YEAR Year dummies Dummy variables () Dummy variables regarding the identity of the controlling shareholder(s): FAM = family-owned (TYPE1), SOE = state-owned (TYPE2), SBH = shared block-holding (control in the hands of a group of national and/or international investors through a shareholder agreement - TYPE3), and FOR = foreign ownership (TYPE4). YEAR t defined as YEAR( t) 1 in the t-th year and YEAR( t) 0 otherwise, with t 1,...,4 (1998, 2000, 2002, and 2004). AGE Firm Age Number of years since the foundation of the firm. 25