Regression

After considering the correlation coefficients, a hierarchical regression was performed with all variables, including the dummy variables, in order to support or reject the developed hypotheses. In order to check the hypothesized relationships, a total number of 4 regressions were run. The industry dummy Energy was taken as the reference group because this dummy contains the largest group of firms, i.e. 16%. For country, the dummy USA was taken as the reference group since this dummy contains 31.5% of all firm observations. Lastly, the year dummy 2019 was taken as the reference group since this is the most recent year of analysis.

5.3.1 Assumptions of a multiple linear regression

Whilst running a multiple linear regression test, four main assumptions need to be met in order for the analysis to be valid and reliable (Tabachnick & Fidell, 2007). First, according to the linearity assumption, there must be a straight-line linear relationship between the outcome variable and the independent variable. The scatterplot from the performed regression portrays that the first assumption of linearity is met. Second, the assumption of normality indicates that the residuals of the dependent variable FFP should be more or less

normally distributed, which can be assessed by means of a histogram. The first run of the regression models portrayed a histogram that deviated slightly from a normal distribution.

The casewise diagnostics indicated that there were 4 cases with a standardized residual larger than 4. We decided to remove these cases in order for the regression to better fit the assumption of normality. The case numbers that were removed are #743, #744, #745, and

#890. Consequently, the final regression model has N = 886. After the removal of the four cases the histogram of the FFP showed a distribution that does not deviate from normality.

Hence, it can be concluded that the second assumption of normality is met. Third, the assumption of homoscedasticity signifies that the residuals of the independent variable are the same across all values of that independent variable. Whether the residuals of the independent variable are homoscedastic or heteroscedastic can be assessed through a scatterplot. The scatterplot of the independent variable FFP does not show a clear pattern in the residuals such as a cone-shaped form. Therefore, it can be assumed that the model is not heteroscedastic, hence meeting the fourth assumption of a multiple linear regression.

The fourth assumption of a multiple linear regression is with regards to the absence of multicollinearity. Multicollinearity occurs when independent variables in a regression model are heavily correlated. This is a problem because, as the name implies, independent variables should be independent. Multicollinearity can be measured by the Tolerance value and the VIF value, which should lie above 0,10 and under 10, respectively (Hair, Black, Babin &

Anderson, 2010). The first 4 regression models have been run including the control variable country. For the first two models there was no problem with multicollinearity. However, the third model had serious multicollinearity issues with regards to the country dummy China.

The VIF value for this dummy was more than 42, and the Tolerance value was 0.023. These extreme values can be explained by the high correlation between the variable country and corruption, because certain corruption scores are linked to a specific country. To solve the

issue of multicollinearity, it has been tested whether taking another country as the reference group would lower the VIF and Tolerance scores. In the second test, China was put as the reference group since China is the second largest country dummy containing 12% of all firm observations. However, this created even more multicollinearity issues with regards to several country dummies. For example, the VIF and Tolerance scores for the USA were 87 and 0.011, respectively. Therefore, to adhere to the assumption of no multicollinearity, the decision has been made to remove the control variable country from the regression. After running the 4 regression models without the country dummies, all the values for Tolerance and VIF were within the acceptable range.

Moreover, according to the urgent advise of Field (2013), the independent variables GII, internationalization, and corruption were centred before running the regression models.

The author mentions that when interactions between independent variables are included in the model, there is a high probability of multicollinearity issues. Therefore, the independent variables have been converted to deviation scores in order for each variable to have a mean of zero.

5.3.2 Hierarchical regression

After checking the four assumptions belonging to a multiple linear regression, a hierarchical regression was executed with four models. The first model, model 1, included the control variables firm size, firm age, industry, and year. The second model, model 2, included the independent variable GII. The third model, model 3, added the direct effects of the variables internationalization and corruption on FFP. Finally, model 4 added the interaction effect between GII and internationalization and GII and corruption, thus measuring the potential moderating effect of the variables internationalization and corruption on the relationship between green innovation and financial performance. The outcomes of the

Table 2: Hierarchical regression with FFP as the outcome variable

Model 1 Model 2 Model 3 Model 4

Firm size -.246 -.159 -.201 -.197

(.150) (.150) (.154) (.154)

Firm age -.193 -.159 -.149 -.163

(.150) (.148) (.152) (.153)

Main effect:

Green Innovation Intensity -4.307*** -4.294*** -4.195***

(.931) (.934) (.953)

Direct effect:

Internationalization -.926* -.892

(.542) (.544)

Corruption .018* .017*

(.010) (.010)

Moderating effect:

GII x internationalization .540

(3.193)

GII x corruption -.892

(.063)

Year fixed effects Yes Yes Yes Yes

Industry fixed effects Yes Yes Yes Yes

R² .311 .328 .332 .332

Adjusted R² .292 .308 .311 .310

F 16.21 21.405 2.481 .377

N 886 886 886 886

Unstandardized B, standard errors in parentheses

*** Coefficient is significant at the 0.10 level (2-tailed)

* Coefficient is significant at the 0.01 level (2-tailed)

The first model, model 1, included the four control variables, i.e. firm size, firm age, industry, and year. The control variables accounted for 31.1% of the variance in the dependent variable FFP. This model was significant with a p-value smaller than 0.001, indicating that the aforementioned variables have significant predictive utility. In regression model 2 it became apparent that the addition of the variable GII significantly increased the predictive power of the model (R²=0.328, Adjusted R²=0.308, p<0.001). In order to accept or reject hypothesis 1, the unstandardized B coefficients of model 2 were examined. While hypothesis 1 predicted

that green innovation is positively related to firm financial performance, model 2 portrayed that the variable GII has a B of -4.307, signifying that there is a significant negative relationship between green innovation and firm financial performance (B= -4.307, SE=0.931, p<0.001, 95% CI [-6.134, -2.480]). The coefficient suggested that, ceteris paribus, an increase of 1 unit in GII will results in a 4,307 decrease of FFP. Therefore, it can be concluded that hypothesis 1 needs to be rejected since there is a significant negative relationship between green innovation and financial performance as opposed to the predicted positive relationship.

Hypothesis 2 looked at the moderating effect of internationalization on financial performance, stating that a high scale of internationalization will have a positive moderating effect on the relationship between green innovation and firm financial performance.

However, the outcome of model 4 demonstrated that there is no significant moderating effect of the variable internationalization on the relationship between green innovation and financial performance (B=0.540, SE=3.193, p=0.866, 95% CI [-5.738, 6.807]). The value for R² (0.332) remained the same throughout model 3 and 4 and the adjusted R² (0.310) even went down by 0.01 as opposed to model 3. The F-change was not significant (p=0.686), indicating that the addition of the interaction effects between GII and internationalization, and GII and corruption did not improve the predictive utility of the model. Concluding, the insignificant moderating relationship of internationalization results in the rejection of hypothesis 2.

Furthermore, hypothesis 3 predicted a negative moderating effect of the variable corruption on the relationship between green innovation and financial performance.

Regression model 4, in which the interaction effect of green innovation and corruption were added, showed that there was no significant moderating relationship (B=-0.054, SE=0.063, p=0.389, 95% CI [-0.178, 0.069]). Based on the outcome of the regression, hypothesis 3 needs to be rejected.