*3.3. Corporate Governance*

After estimating the extent of discretionary accruals, the relation between earnings management and the corporate governance practices was investigated. In the regression, the corporate governance practices represented the following independent variables:

Board independence: referred to the percentage of shareholder-elected directors that were evaluated as independent with respect to the company's executive management, material business contacts and main shareholders.

Employee representatives: referred to the presence of employee representatives or not. The variable was calculated as a dummy variable assigned the value 1 if the board has employee representatives, 0 otherwise.

Share ownership by directors: referred to the percentage of directors who directly or indirectly holds shares in the company. The variable was calculated by scaling the total number of directors who holds shares by total board size.

Directors as majority shareholders: referred to the percentage of directors who directly or indirectly is listed amongst the company's 20 largest shareholders. The variable was calculated by scaling the total number of directors who are majority shareholders by total board size.

Board activity: referred to the total number of meetings held during a year, scaled by quarter. The variable was calculated using the natural logarithm of total board meetings2.

Audit committee: referred to the presence of an audit committee or not. The variable was calculated as a dummy variable assigned the value 1 if the firm has an audit committee, 0 otherwise.

Earnings management decisions can also be influenced by factors other than the explanatory variables included in this analysis. To control for this and for any spurious relations between board characteristics and earnings management, the control variables firm size, return on assets and return on equity were included.

Firm size: the natural logarithm of total assets was used as a proxy for firm size.

Return on assets: net income divided by total assets was used as a measure for firm performance. Return on equity: total equity divided by total assets was used as a measure for firm profitability. To test the hypotheses', the following equation was used:

$$\begin{aligned} \text{absDA}\_{\text{it}} &= \beta\_0 + \beta\_1 (\text{BISE}\_{\text{it}}) + \beta\_2 (\text{DER}\_{\text{it}}) + \beta\_3 (\text{SOD}\_{\text{it}}) + \beta\_4 (\text{MJS}\_{\text{it}}) + \beta\_5 (\text{BA}\_{\text{it}}) \\ &+ \beta\_6 (\text{AC}\_{\text{it}}) + \beta\_7 (\text{FS}\_{\text{it}}) + \beta\_8 (\text{ROA}\_{\text{it}}) + \beta\_9 (\text{ROE}\_{\text{it}}) + \varepsilon\_{\text{it}} \end{aligned} \tag{3}$$

absDAit = absolute value of discretionary accruals for company i in quarter t

BISEit = board independence for company i in quarter t

DERit = dummy variable that equal 1 if the company has employee representatives on the board, 0 otherwise

<sup>2</sup> The natural logarithm is used to correct for heteroscedasticity (Benoit 2011).

SODit = share ownership by directors for company i in quarter t MJSit = directors as majority shareholders for company i in quarter t BAit = board activity for company i in quarter t ACit = dummy variable that equal 1 if the company has an audit committee, 0 otherwise FSit = firm size for company i in quarter t ROAit = return on assets for company i in quarter t ROEit = return on equity for company i in quarter t

Our study used panel data, featured by exploring the cross-section and time-series data simultaneously. A Hausman test (Table A3), showed that fixed effects estimator was a better fit for the model than the random effects estimator3. Moreover, Equation (3) using OLS was estimated. Additional analysis of the residuals from this estimation displayed significant heteroscedasticity. Consequently, the regression using robust standard errors was estimated. In regression estimates, multicollinearity due to a significant linear relationship between the explanatory variables can affect the estimation of the coefficients of the variables, leading to imprecise results. To test the severity of multicollinearity in the data, a correlation matrix and the Variance Inflation Factor (VIF) method was used. According to Brooks (2019), severe multicollinearity is indicated if the correlation between 2 variables exceeds 0.80 and the VIF index exceed 5. The VIF for each explanatory variable was under 5, with a total mean of 1.6. Supported by the correlation matrix, multicollinearity was not a problem to the model. The correlation matrix and VIF index for the variables are reported in the Appendix A (Tables A4 and A5).
