*4.2. Model Specification and Estimation Techniques*

The fixed-effects model and system GMM proposed by Blundell and Bond (1998) were used to test our hypothesis and cater to the un-observed endogeneity problem (Nakano and Nguyen 2013; Nguyen et al. 2014).

In the first equation, we want to explore the effect of agency cost and corporate governance on firm performance.

$$FP\_{it} = \alpha\_o + \alpha\_1 AC\_{it} + \alpha\_2 CGQ\_{it} + \sum \alpha\_n CTR + \mu\_o \tag{1}$$

where FP represents firm performance and has more than one measure, AC is the measure of agency cost, while CGQ is the corporate governance quality index. CTR represents the control variables used in the equations.

We added the interaction term in Equation (1) to capture the impact of corporate governance quality and ownership structure on the relationship between agency cost and firm performance.

$$FP\_{it} = a\_0 + a\_1 A C\_{it} + a\_2 C G Q\_{it} + a\_2 \left( A C \times M oderators\_{it} \right) + \sum a\_n CTR + \mu\_o \tag{2}$$

where the moderators are corporate governance quality, ownership concentration, and ownership type, respectively.

The financial performance of the firm is time-dependent, i.e., the current performance of the firm is affected by the past performance and testing the effect of two-year lagged performance on current performance does not give us a significant impact. This leads us to conclude that the AR (1) dynamic panel model is sufficient. The literature on corporate governance suggests that corporate governance, as well as ownership structure, are endogenously determined (Nguyen et al. 2015). Therefore, this study first uses the fixed effect model to control the governance variables when estimating the agency–performance relationship. However, the use of a fixed-effect estimator does not eliminate the endogeneity lag performance measures. Following Nguyen et al. (2015), this study uses the system GMM recommended by Blundell and Bond (1998). The major advantage of constructing the system GMM estimator is that it enhances the power of estimation (Hoechle et al. 2012).
