3.1 Empirical model
To explore the impact of environmental protection investment on firm performance under the situation of financing constraints, this paper constructs the following regression model and divides the sample into two groups according to the level of financial constraints that a firm face to do the GMM regression:
where CFPit is one of the two measures of corporate financial performance i.e. market performance TQit and accounting financial performance ROAit and; EINVit is one of the two measures of a firm’s environmental investment: first, the ratio of the total amount of environmental investment multiplied by 100 divided by total assets (EINV1it), and second, the natural logarithm of the total environmental protection investment+1 (EINV2it); CNTitrepresents five firm-level financial controls i.e. SIZEit natural logarithm of total assets of the firm, LEVit is total debt over total assets, GRit is percentage growth in total assets, CASHit is cash and cash equivalents over total assets, and TANit is total fixed assets over total assets; and GOVit represents firm-level governance variables i.e. CTLit percentage shares held by largest shareholder, DUALit proxy for chairman-CEO duality, INDit percentage of independent directors on board, and STit proxy for the state-owned firm.
Table 1 provides definitions of all variables included in our empirical model.
3.2 Measurement of financial constraints
We construct a dummy variable of corporate financing constraints following Kaplan and Zingales (1997) and He and Ye (2017) to classify the sample firms as financially constrained and non-constrained firms. The calculation methodology is as follows. As a first step, we calculate five different ratios for each year: i. cash flow from operating activities over total assets (CFOA), ii. cash dividends over total assets (CD), iii. cash and cash equivalents over total assets (Cash & Kilcullen), iv. total debt over total assets (Levy & Murnane), v. Tobin’s Q as market value to total equity plus book value of total debt over book value of total assets (TQ). In the second step, we assign 1 or zero values to dummy variables kz1 to kz5 for each year. If CFOA is less than the median, then kz1 will assume value 1 otherwise zero. If CD is less than the median, then kz2 will assume vale 1 otherwise zero. If Cash is less than the median, then kz1 will assume vale 1 otherwise zero. If Lev is greater than the median, then kz1 will assume vale 1 otherwise zero. If TQ is greater than the median, then kz1 will assume vale 1 otherwise zero. In the third step, we add up the values of kz1 to kz5 to make KZ = kz1 + kz2 + kz3 + kz4 + kz5. In the fourth step, we take KZ as the dependent variable and use the ordered logit model to regress the five ratios calculated in the first step to get the regression coefficient of each variable. In the fifth step, we use the value of each ratio and the regression coefficient of each ratio we sum up the KZ index to measure the degree of financing constraints of listed companies. The larger the KZ index, the stronger the financing constraints of listed companies. After getting the KZ index, we calculate a dummy variable that equals 1 if KZ value is less than the mean value otherwise it equals 0 and uses it as a grouping purpose. This whole procedure divides our sample of 1988 firms into 817 financially non-constrained firms and 1171 financially constrained firms.
To mitigate the risk of obtaining biased results, we employ panel-based two-step system GMM technique considered as the most efficient method to address the issues of unobserved heterogeneity, endogeneity, and heteroskedasticity in the panel data (Hsiao, 1985; Abuzayed, 2012; Baños-Caballero et al., 2014; Tahir and Anuar, 2016). Moreover, two-step system GMM based models require only two diagnostics: Hansen test for over-identification restrictions to test the validity of instruments and AR2 to test the assumption that the differenced error terms do not have a second-order serial correlation. The reported results under each regression model and the p-values of Hansen test and AR2 are insignificant in all cases suggesting that the instruments are valid and the results do not suffer due to auto-correlation.
3.4 Data description
The data of this study is composed of 1988 Chinese firms listed on Shanghai and Shenzhen stock exchanges during 2009-2016. We extracted environmental investment data manually from the CSR report of each firm. Moreover, data of all the other variables of sample firms was obtained from the CSMAR (China Stock Market & Accounting Research) database.
Table 2 reports the descriptive statistics of the variables. With a mean of 1.628 and a standard deviation (SD) of 1.396, the sampled firms have a lot of variation in their market performance and 50% of the sampled firms have market performance below the mean value. Similarly, the operating performance of the sampled firms shows that the performance of different firms is quite different (with a mean of 0.026 and SD of 0.053). A wide range of book financial performance and market performance plausibly indicates the potential impact of environmental protection investment under financing constraint situation that calls for an investigation. The mean value of the two environmental investment variables is 2.027 and 1.670 respectively, and the range of each variable is relatively large, which shows that different firms have different environmental investment preferences. The values of control variables are generally consistent with the previous literature (Jiang & Akbar, 2018; Wang et al., 2020).
Table 3 reports the correlation among the variables. Quite intuitively, there is a positive correlation between TQ and ROA, and between the two environmental protection investment variables. To avoid multicollinearity in regression analysis, we use two environmental protection investment variables separately in two different models. The correlation coefficients among all other variables are relatively small and below 0.5, which means the regression model will not produce serious multicollinearity problems. Further, we calculate the variation inflation factor (VIF) and a VIF of less than 10 for all of our regression models suggests that our models are robust for multicollinearity (Ott & longnecker, 2015).