CHAPTER 8 CONCLUSIONS, IMPLICATIONS AND LIMITATIONS . 239
5.2.3.1 The sensitivity of the results to the reduction of instruments
The potential danger of the System GMM estimation implementation is instrument proliferation. Numerous numbers of instrumental variables in the System GMM estimator may bias the estimated coefficients towards those from non-IV estimators, such as the OLS, and have potential to severely weaken the power of the Hansen-J test in detecting the invalidity of the instruments employed (Roodman, 2009b). Therefore, it is essential to check whether the results are sensitive to the reduction of instrumental variables.
It is obvious that 28 instruments used in the System GMM model (Table 5.11) are small relative to the total of 120 clusters. This suggests that instrument proliferation is unlikely to be the problem. More carefully, following good standard practices in using the System GMM approach suggested by Roodman (2009a, 2009b), this subsection checks the sensitivity of the results reported in Table 5.11 with the reduction in the number of instrumental variables.
Specifically, the instrument count is reduced from 28 instruments (Table 5.11) to 21 instruments (Table 5.12)66. As shown in columns 1 and 2 of Table 5.12, the results generally remain unchanged, suggesting that the findings are robust to the instrument reduction.
The results reported in Table 5.12 show that the percentage of female directors is positively and statistically significantly related to Tobin’s Q at the 5% level (p =
66 Besides using only one lag of each instrumenting variable rather than all available lags for instrumental variables, the author also applies a collapsing instruments approach to reduce the instruments’ count. See Roodman (2009b) for more details about the techniques for reducing the instrument count in the System GMM estimation.
150
0.037), thus supporting hypothesis HVN1. The coefficient on female (𝛽 = 0.021) means that a ten percentage point increase in the ratio of female directors will, on average, increase the predicted Tobin’s Q by approximately 21%, holding all other factors fixed. As mentioned in Subsection 5.2.1.1, this is an economically strong effect given that the board size of Vietnamese listed companies is small.
This System GMM model result is consistent with those obtained by using both static and dynamic OLS models, thereby suggesting that the findings are robust to alternative econometric approaches.
This result is also consistent with the findings of several prior studies that confirm the positive relationship between gender diversity and firm performance (e.g.
Campbell & Mớnguez-Vera, 2008; Dezsử & Ross, 2012). This finding implies that board gender diversity seems to affect firm value, a point which is in general agreement with Adams et al. (2011, p. 31), who suggest that “shareholders may value female directors because they are better monitors and because they may alleviate value-decreasing stakeholder conflicts”.
As reported in Table 5.12, the coefficient on one-year lagged Tobin’s Q is significantly positive at the 1% level (𝛽 = 0.633, p = 0.00), thus suggesting that past performance can help to control for the unobserved historical factors in the relationship between corporate governance and firm performance. This empirical evidence strongly supports the arguments of Wintoki et al. (2012), among others, that the link between corporate governance and firm performance should be examined in a dynamic framework.
Regarding the variable nonexe, the results obtained from the static OLS, FE, and dynamic OLS models show that the presence of non-executive directors has no
151
significant impact on firm performance, thus supporting the hypothesis HVN2. However, when moving to the System GMM model, the results reported in Table 5.12 show that the relationship is significantly negative at the 5% level (𝛽 = – 0.019, p = 0.017). This conclusion is in line with Nowland (2008), who challenged agency theory’s viewpoint regarding the vital role of non-executive directors in monitoring managerial behaviours and in improving firm performance.
Regarding the other corporate governance variables, it is observed that there is statistical evidence of a significantly positive link between concentrated ownership and firm performance (𝛽 = 0.014, p = 0.027), thus supporting the hypothesis HVN5. This result is consistent in all four models applied in this study and similar to that obtained by Victoria (2006) and Nguyen et al. (2014), among others. The positive relationship between concentrated ownership and performance is in agreement with the agency theory perspective that ownership concentration helps to reduce agency problems arising from the separation of ownership and control (Shleifer & Vishny, 1986). This, in turn, is expected to improve firm performance.
However, the significantly positive relationship between board size and firm performance, revealed by the static OLS, FE, and dynamic OLS models (Table 5.7 and Table 5.8), disappears when dynamic endogeneity and simultaneity are controlled by using the System GMM approach (Table 5.11 and Table 5.12). This result, supporting hypothesis HVN4, accords with the findings of Pham et al.
(2011); Schultz et al. (2010); and Wintoki et al. (2012), who argued that such significant links, estimated by the pooled OLS and FE models, may be the result
152
of spurious correlations. Similarly, the relationship between CEO duality and firm performance changes from significantly positive to insignificantly negative when the author moves from the static OLS and FE models to the System GMM model.
Although this result does not support hypothesis HVN3, it does support the argument of Schultz et al. (2010); and Wintoki et al. (2012), among others, that taking the dynamic nature of the relationship between corporate governance structures and firm performance into consideration is essential to ensure the reliability of causal inferences.
With regards to the capital structure variable leverage, it is evident from Table 5.11 and Table 5.12 that the positive relationship between financial leverage and firm performance revealed by the FE and the dynamic OLS models disappears when the potential sources of endogeneity are taken into consideration. Several robustness checking models reported in Table 5.14 and Table 5.15 also confirm that the estimated coefficient on leverage is not statistically different from zero at any conventional levels of significance, suggesting that financial leverage has no impact on firm performance, thus not supporting hypothesis HVN6. Although this finding is consistent with that of Nguyen et al. (2014); Schultz et al. (2010) and others, the relationship between financial leverage and firm performance is not really clear in practice. The discussion below provides some possible explanations for this finding.
A recent study undertaken by Jiraporn, Kim, Kim, and Kitsabunnarat (2012) suggested that debt financing and corporate governance mechanisms may substitute for each other to alleviate agency cost, whereby firm performance is improved. If that is the case, it is plausible to argue that the potential performance
153
effect of financial leverage in Vietnamese companies is likely to be replaced by the stronger effects of other corporate governance mechanisms, including ownership concentration (measured by block) and board gender diversity (measured by female). In consequence, the estimated coefficient on leverage should not be statistically different from zero.
In a similar vein, González (2013) argued that the relationship between financial leverage and firm performance is likely to be contingent upon two contradictory antecedents: (i) the cost of financial distress; and (ii) the benefits of the disciplinary role of debt financing. A firm with higher financial leverage may suffer from higher costs of financial distress but may also benefit from the disciplinary role of debt financing, by which managers are forced to take value- maximising decisions (González, 2013). Therefore, the net effect of financial leverage on firm performance can be neutralised if neither of these two antecedents is predominant.
It is worth noting that Hansen-J test of over-identification and difference-in- Hansen tests of exogeneity of instrument subsets have been implemented to confirm the validity of the robustness model. Accordingly, the Hansen-J test, as reported in the last row of Table 5.12, yields a p-value of 0.22, suggesting that the instruments employed in the robustness model are valid. The results of difference- in-Hansen tests reported in Table 5.13 also suggest that the subsets of instruments in the robustness model are econometrically exogenous. In addition, as reported in Table 5.12, the F-test statistic (12.721) for the overall significance of the robustness regression also supports the model specification.
154
Table 5.11: The relationship between corporate governance structures and performance of Vietnamese listed companies: A System GMM estimation
Dependent variable: Tobin's Q ratio [lnq]
Explanatory variables [notation] b/[p] (t)
(1) (2)
Intercept -5.391* (-1.755)
[0.082]
One-year lagged Tobin's Q [laglnq] 0.611*** (3.123)
[0.002]
Percentage of female directors (%) [female] 0.018* (1.858)
[0.066]
Percentage of non-executive directors (%) [nonexe] -0.017** (-2.131)
[0.035]
Duality [dual] -0.011 (-0.054)
[0.957]
Board size [lnbsize] -1.178 (-1.248)
[0.214]
Ownership concentration (%) [block] 0.010* (1.881)
[0.062]
Firm age [lnfage] 0.369** (2.291)
[0.024]
Firm size [fsize] 0.243* (1.850)
[0.067]
Leverage (%) [lev] -0.003 (-0.352)
[0.725]
Industry dummy variables [industry] no
Firm fixed-effects yes
Year dummy variables [year] yes
Number of observations 352
F statistic 12.806***
Number of instruments 28
Number of clusters 120
Arellano-Bond test for AR(1) in first differences (p-value) 0.146 Arellano-Bond test for AR(2) in first differences (p-value) not defined Hansen-J test of over-identification (p-value) 0.299
Note: This table reports the result of the System GMM regression of firm performance (lnq) on board structure variables and other control variables. The variables are as defined in Table 4.6.
Asterisks indicate significance at 10% (*), 5% (**), and 1% (***). The t-statistics are reported in parentheses and are based on Windmeijer-corrected standard errors. The p-values are presented in brackets. Lags 2 and 3 of the levels of firm performance variable (lnq), board structure variables (female, nonexe, dual, and lnbsize) and other control variables (block, fsize, and lev) are employed as GMM-type instruments for the first-differenced equation. Lag 1 of the first differences of firm performance, board structure variables, and other control variables are used as GMM-type instruments for the levels equation. Year dummies and lnfage are treated as exogenous variables.
Year dummy variables are included in the regression but not reported.
155
Table 5.12: Robustness check of the sensitivity of the results to the instrumental variables’ reduction
Dependent variable: Tobin's Q ratio [lnq]
Explanatory variables [notation] b/[p] (t)
(1) (2)
Intercept -4.795* (-1.680)
[0.096]
One-year lagged Tobin's Q [laglnq] 0.633*** (3.791)
[0.000]
Percentage of female directors (%) [female] 0.021** (2.109)
[0.037]
Percentage of non-executive directors (%) [nonexe] -0.019** (-2.429)
[0.017]
Duality [dual] -0.017 (-0.084)
[0.933]
Board size [lnbsize] -1.429 (-1.373)
[0.172]
Ownership concentration (%) [block] 0.014** (2.237)
[0.027]
Firm age [lnfage] 0.430** (2.578)
[0.011]
Firm size [fsize] 0.227* (1.744)
[0.084]
Leverage (%) [lev] -0.000 (-0.013)
[0.990]
Industry dummies [industry] no
Firm fixed-effects yes
Year dummies [year] yes
Number of observations 352
F statistic 12.721***
Number of instruments 21
Number of clusters 120
Arellano-Bond test for AR(1) in first differences (p-value) 0.085 Arellano-Bond test for AR(2) in first differences (p-value) not defined Hansen-J test of over-identification (p-value) 0.220
Note: This table presents robustness check of the sensitivity of the results obtained from the System GMM to the instruments’ reduction. The variables are as defined in Table 4.6. Asterisks indicate significance at 10% (*), 5% (**), and 1% (***). The t-statistics are reported in parentheses and are based on Windmeijer-corrected standard errors. The p-values are presented in brackets. Lags 2 and 3 of the levels of firm performance variable (lnq), lag 2 of the levels of board structure variables (female, nonexe, dual, and lnbsize) and other control variables (block, fsize, and lev) are employed as GMM-type instruments for the first-differenced equation. Lag 1 of the first differences of firm performance, board structure variables, and other control variables are used as GMM-type instruments for the levels equation. Year dummies and lnfage are treated as exogenous variables. Year dummy variables are included in the regression but not reported.
156
Table 5.13: Difference-in-Hansen tests of exogeneity of instrument subsets used in the robustness model
Tested instrument subsets
Test statistics
Degrees of freedom
p- value
Panel A: System GMM-type instruments
Instruments for levels equation as a group 10.82 8 0.212
lnqit-2 and lnqit-3 (for transformed equation) 1.36 2 0.506
Δlnqit-1 (for levels equation) 1.55 1 0.213
Instruments for board structure variables 11.60 8 0.170
Instruments for control variables 10.56 6 0.103
Panel B: Standard instruments
2009 and 2010 year dummies, and lnfage 6.25 3 0.100
Note: This table presents difference-in-Hansen tests of exogeneity of instrument subsets employed in the robustness model to check the sensitivity of the results to the instrumental variables’ reduction. The test is under the null hypothesis of joint validity of a specific instrument subset. The variables are as defined in Table 4.6 The test statistics are asymptotically Chi- squared distribution with degrees of freedom equal to the number of questionable instrumental variables (Roodman 2009).
GMM instrument subset used for levels equation includes one-year lagged differences of firm performance, board structure, ownership structure, capital structure, and other control variables (Δlnqit-1 ; Δfemaleit-1 ; Δnonexeit-1 ; Δdualit-1 ; Δlnbsizeit-1 ; Δblockit-1 ; Δfsizeit-1 ; and Δlevit-1).
GMM instrument subset used for board structure variables includes lag 1 of the first difference and lag 2 in levels of board structure variables (femaleit-2 and Δfemaleit-1 ; nonexeit-2 and Δnonexeit-1 ; dualit-2 and Δdualit-1 ; lnbsizeit-2 and Δlnbsizeit-1).
GMM instrument subset used for the other corporate governance and control variables includes lag 1 of the first differences and lag 2 in levels of these variables (blockit-2 and Δblockit-1 ; fsizeit-2
and Δfsizeit-1 ; levit-2 and Δlevit-1). The subset of standard instruments for levels equation includes 2009 and 2010 year dummies, and lnfage. 2008 and 2011 year dummies are dropped due to collinearity.