*4.5. Robustness Tests*

To further validate the results and test their consistency, several robustness checks were performed. Our robustness test mainly relates to three concerns. Firstly, one concern exists about the appropriateness of a proxy for R&D investment intensity. Referring to the method of Kao and Chen (2020) [111], we firstly substitute the R&D investment intensity with the proportion of R&D expenses in total assets (See Table 7). Secondly, another concern involves how existing extreme observations may influence the accuracy of our estimations. Therefore, all continuous variables are once again winsorized by 5% instead of 1% (See Table 8). Thirdly, the last concern relates to improper designation for the omitted dependent variable. Considering the incompleteness of data, some firms' R&D investments maybe not be disclosed and are treated as omitted observations. For a large number of omitted observations, we cannot distinguish between zeros that represent a true zero level of R&D activity and zeros that were created by the statistical authorities because no figures were recorded in the database. During the previous empirical process, these omitted observations are made to be zero. However, this designation may lead to an underestimation of coefficients. Therefore, we exclude these samples and carry out the regressions again (See Table 9). Overall the results were highly robust to these changes in specification.


**Table 7.** Variable Replacement test.

**Notes:** Standard errors are in parentheses. \*\*\* *p* < 0.01, \*\* *p* < 0.05, \* *p* < 0.1

Additionally, we address endogeneity issues in our analysis by applying an instrumental variable approach. Specifically, we analyze whether our mediator variable is exogenous in model (3). A recent contribution proposes that firm innovation inversely increases the firm information environment overall, subsequently stimulating firms' access to financing [112]. There may exist a problem of mutual cause and effect between firm overborrowing and the R&D investment intensity. In addition, the Hausman test [113] for endogeneity also shows that there exists endogeneity in overborrowing. To alleviate the endogeneity problem, we carry out the 2SLS regression test [114]. We take the mean of industrial overborrowing as the instrument variable of a firm's overborrowing. It is correlated with the firm's overborrowing, whereas it is unlikely to be affected by the firm's R&D investment intensity, thus meeting the basic requirements of correlation and exogeneity of instrument variables. The under-identification test (Anderson LM statistic is 903.83, *p* value is 0.00) also reflects that the instrument variable is correlated with the endogenous variable. The weak ID test statistics (Cragg–Donald Wald F is 1075.54) are far beyond the 10% Stock–Yogo weak critical values of 16.38, further rejecting that the instrument is weak. Our results in Table 10 show that the mediator—overborrowing—does cause firm's R&D intensity to drop down significantly.


**Table 8.** Winsorize Test (5%).

**Notes:** Standard errors are in parentheses. \*\*\* *p* < 0.01, \*\* *p* < 0.05, \* *p* < 0.1.

**Table 9.** Eliminating the default sample of R&D investment intensity.


**Notes:** Standard errors are in parentheses. \*\*\* *p* < 0.01, \*\* *p* < 0.05, \* *p* < 0.1.


**Table 10.** 2 SLS Regression Results.

**Notes:** Standard errors in parentheses. \*\*\* *p* < 0.01, \* *p* < 0.1.
