Impact of Directors’ Network on Corporate Social Responsibility Disclosure: Evidence from China

The concept that a director possesses relatively many channels of communication or resource exchange is measured by degree centrality; The concept that a director lies on relatively more paths between pairs of other directors is measured by betweenness centrality; The concept that a director possesses relatively closer ties to other directors is measured by closeness centrality. The computation of degree, betweenness, and closeness centrality is explained in “Sample and Research Design” section.

In November 2021, China Mainland launched its third stock exchange in Beijing.

Please refer to http://​www.​csrc.​gov.​cn for detailed information on the Code of Corporate Governance for Listed Companies and the Standards for the Content and Form of Information Disclosure by Companies Publicly Offering Securities No. 2 Content and Form of Annual Reports (in Chinese).

Companies that are key pollutant discharge units announced by the environmental protection department shall disclose environmental information such as the discharge of major pollutants and the construction and operation of pollution prevention facilities. Voluntarily disclosure information on protecting the ecology, and preventing pollution is encouraged.

The need for social recognition is defined as an individual’s desire to be recognized in a social group or organization by his or her engagement in social activities (Lin et al., 2008).

Due to the lag of one year in the release of corporate social responsibility report, we made an adjustment of CSR data. The year in this paper means the year CSR report belongs to, rather than the year CSR report is published. For example, the CSR report in 2010 was published in 2011.

That means for a given quantile limit (such as 1%), the part that exceeds the upper and lower bounds is replaced by a quantile.

The 14 level-2 indicators of RKS’ CSR reports ranking system are grouped as follows: Macrocosm includes strategy and governance. Content includes economic performance, labor and human rights, environment, fair operation, consumers, and community engagement and development. Technical includes content of the balance, information comparability, report on innovation, credibility and transparency, normative and availability, and effectiveness of information delivery. Please refer to http://​www.​rksratings.​cn/​list-704-1.​html for detailed information (in Chinese).

To address the endogeneity issue, this baseline model is estamted using two-stage least squares. Details of two-stage least squares are presented in the section of “Endogeneity of Directors’ Network and CSR Disclosure Quality” combined with other regression models.

Kaplan and Zingales’s (1997) index measured at the end of fiscal year t, calculated as minus 1.002 × cash flow plus 0.283 × Tobin’s Q plus 3.139 × leverage minus 39.36 × dividends minus 1.315 × cash holdings.

In Norway and Spain, 40% of gender quota was allocated for female. In France, there was a rule that the proportion of female directors should not be lower than 40% by the year 2017 (Katmon et al., 2017). The Australian Institute of Company Directors once targeted 30% female board representations by the end of 2018, and the Japan Prime Minister once set a goal of increasing the percentage of female in executive positions in the country’s companies to more than 30% by year 2020 (Katmon et al., 2017).

Please refer to http://​www.​csrc.​gov.​cn.​html for detailed information (in Chinese).

In the following 2SLS tests in our research, the F-statistics of the first-stage regressions are all higher than the cutoff point of 10 suggested by prior research (Staiger and Stock 1997). The t-statistics for the instrumental variables in the first-stage regressions are all higher than the cutoff point of 3 suggested by prior research (Adkins and Hill 2008). Thus, we conclude that instrumental variables are valid. We only present the results of the second-stage regressions. The results of the first-stage regressions are not presented for brevity, but are available upon request.

Two-step GMM has been found to be more efficient than one-step GMM estimators (Windmeijer 2005). One-step GMM estimators use weight matrices that are independent of estimated parameters, whereas the efficient two-step GMM estimator weighs the moment conditions by a consistent estimate of their covariance matrix (Windmeijer 2005). This weight matrix is constructed using an initial consistent estimate of the parameters in the model (Windmeijer 2005). The extra variation due to the presence of these estimated parameters, in the efficient weight matrix, accounts for much of the difference between the finite sample and the estimated asymptotic variance for two-step GMM estimators based on moment conditions that are linear in the parameters (Windmeijer 2005). This difference can be estimated, resulting in finite sample corrected estimates of the variance. The proposed feasible correction to the estimate of the asymptotic variance is very simple to implement and is shown to approximate the finite sample variance of the two-step GMM estimator well in a Monte Carlo study of a panel data model, leading to more accurate inference (Windmeijer 2005).

We use the xtabond2 command with the option of noleveleq in Stata software to perform Difference GMM regressions. We use the laglimits option in Stata to restrict the lag ranges used in generating the instrument sets (Roodman 2009).

We check the reliability of the GMM estimates with the Hansen test of overidentification and Arellano and Bond (1991) test for serially uncorrelated error terms (Eugster 2020; Wintoki et al., 2012). The Hansen statistic of overidentification tests the null hypothesis of a correct model specification and valid overidentifying restrictions. A p-value of 10 percent or higher indicates that the lagged firm values are exogenous to the current values. The Arellano and Bond (1991) test for autocorrelation has the null hypothesis of no autocorrelation and is applied to the differenced residuals (Eugster 2020). Due to the construction of the dynamic GMM panel, the AR(1) test will be usually rejected. Nevertheless, the AR(2) test remains important to detect a serial correlation (Eugster 2020). A second-order serial correlation in the dynamic panel GMM indicates a specification error and a potential omitted variable bias (Arellano and Bond 1991; Eugster 2020).

We use the xtabond2 command in Stata software and perform two-step System GMM regressions. We use the laglimits option in Stata to restrict the lag ranges used in generating the instrument sets (Roodman 2009).

We use the lag option in the xtabond2 Stata command, which only uses instruments with the exact specified lag as instrumental variables. Omitting this option would lead to using the firm’s entire history as an instrumental variable for the current values (Eugster 2020).

When multiple lags are used as instruments, the Diff-in-Hansen tests of exogeneity verifies the assumption that any correlation between the endogenous variables and the unobserved effect is constant over time. Therefore, these tests verify whether the lagged differences are exogenous for the level equation (Eugster 2020). Furthermore, the additional lag specification test should help us in understanding the underlying model and should be more robust in terms of the exogeneity concerns of the instruments (Eugster 2020).

For simplicity, we did not report the results in the section “Other Robustness Tests.”

Please refer to http://​tv.​cctv.​com/​2013/​10/​30/​VIDE138313236905​5628.​shtml for detailed information on Document No. 18 (in Chinese).