*4.2. Main Results*

We report the regression results of Equation (1) in Table 3. Columns 1–2 show the regression results for the innovation outputs—*LnPatents* and *LnInventions*—respectively. The coefficient of the interaction term, *Treated* × *Post*, is positive and statistically significant in column 1, suggesting the positive impact of MSER on firms' patent counts. In other words, the treatment firms that were required to disclose social and environmental governance information experienced a striking increase in patents relative to the control firms. The economic effect was also sizeable, that is, the number of patents for the treatment

firms increased by 19.7% three years after the regulations. Column 2 shows a positive and significant coefficient and suggests that the number of invention patents increased by 16.6% for the treatment firms compared to the control firms. To mitigate concerns regarding potentially omitted variables, we also performed an alternative specification that included the firm and year fixed effects in our DID estimation. The inclusion of firm fixed effects can rule out the interpretation from time-invariant firm characteristics; year fixed effects can absorb common shocks into all firms in a given year. Columns 3 and 4 in Table 3 indicate that our inferences remained unchanged with this alternative specification.


**Table 2.** Descriptive Statistics and Correlation Matrix (after Matching).

N = 2958. The bold correlation coefficient represents *p* < 0.05.

We further assessed the dynamics of the treatment effect. To do so, we used 2008 as the benchmark year, *Year* (0), and replaced *Post* with five year dummies: *Year* (−2) and *Year* (−1) for the two years prior to the regulations; *Year* (1), *Year* (2), and *Year* (3) for the first, second, and third years after the regulations, respectively. We then made these year dummies interact with *Treated* to capture the dynamic impact of the regulations on firm innovation. As shown in Table 4, the coefficients of all pre-regulation dummies (i.e., *Treated* × *Year* (−2) and *Treated* × *Year* (−1)) were small and insignificant. This evidence reassured us that this matched sample had no pre-existing trend. Thus, the parallel trends were stratified. Moreover, we found that the effect became significant only two years after the regulations, thus suggesting that the MSER took 24 months to translate to higher innovative output, which is consistent with the innovation lag found in previous studies [11,37]. Overall, our main results revealed that MSER in China had a significant effect on firm innovation output. Specifically, the innovation quantity (i.e., total patents) and innovation quality (i.e., the number of invention patents) experienced a striking increase for the treatment firms.

**Table 3.** The Impact of MSER on Firm Innovation.



**Table 3.** *Cont*.

Note. This table reports the estimated coefficients and heteroskedasticity-adjusted robust standard errors clustered at the firm level (in parentheses). Significance at \* *p* < 0.1; \*\* *p* < 0.05; \*\*\* *p* < 0.01.



Note. The results of the year dummies are omitted for conciseness. This table reports the estimated coefficients and heteroskedasticity-adjusted robust standard errors clustered at the firm level (in parentheses). Significance at \* *p* < 0.1; \*\* *p* < 0.05; \*\*\* *p* < 0.01.

#### *4.3. Robustness Checks*

**Nonpatenting Firms.** Our regression sample contained firms that did not file patents throughout the sample period, which accounted for 18.7% of all observations. Therefore, our results may have been affected by these nonpatenting firms. Although the inclusion of firm fixed effects in previous analyses has helped to mitigate this concern, we re-estimated our regression after excluding the nonpatenting firms. The results were unchanged (see columns 1–2 in Table 5).

**Alternative Observation Windows.** In our main analyses, we adopted a three-year observation window to observe the impact of MSER on innovation. In this part, we extend our observation window to four or five years to verify whether our results were sensitive to the selection of the observation window. As shown in Table 5, our results were robust in the four- and five-year observation windows. Moreover, the coefficient of the interaction term, *Treated* × *Post*, increased with the rise in the observation window. In particular, column 5 indicates that the number of patents for the treatment firms increased by 21.1% five years

after the regulations, which was larger than the influence after four years (20.6%, column 3) and three years (19.7%, column 1). Similar to the total patents, the impact on invention patents increased with the extension of the observation window (19.6%, 17.6%, and 16.6% after five, four, and three years of the regulations, respectively). The evidence shows that the influence of the MSER still existed after five years.


**Table 5.** Robustness Checks I—Changing Observation Window

Note. This table reports the estimated coefficients and heteroskedasticity-adjusted robust standard errors clustered at the firm level (in parentheses). Significance at \* *p* < 0.1; \*\* *p* < 0.05; \*\*\* *p* < 0.01.

**Alternative Matching Rules.** In our main regression, we used the PSM approach with a combination of one-to-two nearest-neighbor matching and replacement. Here, we selected different matching rules to check the robustness of our findings. Table 6 presents the corresponding results. First, we adopted different nearest-neighbor algorithms, i.e., oneto-one nearest-neighbor matching and nearest-three-neighbors matching, to generate the matched sample. Columns 1–4 indicate that the impact of MSER on the total number of patents and invention patents was unchanged. Second, we set a caliper of 0.02 to the one-to-two nearest neighbor because the *propensity score* between the matched pair may have been too far even if we used the "nearest neighbor". Common practice is to set the caliper to a 0.25\* standard error of the propensity score. However, the difference in the propensity scores of all matched pairs in our sample was smaller than a 0.25\* standard error of the propensity score (0.07) using default one-to-two nearest-neighbor matching. Therefore, we selected 0.02, a smaller and more restrictive rule, as the size of the caliper. As shown in columns 5 and 6, our results were robust to this setting.

**Table 6.** Robustness Checks II—Alternative Matching Rules.


Note. This table reports the estimated coefficients and heteroskedasticity-adjusted robust standard errors clustered at the firm level (in parentheses). Significance at \* *p* < 0.1; \*\* *p* < 0.05; \*\*\* *p* < 0.01.

**Placebo Test.** Here, we conducted a placebo test by replicating our analysis using a fabricated shock [45]. We set 2007 as the time when the false "MSER" occurred. To avoid "contamination" due to real treatment effects, we only used observations from 2006 to 2008 and supplemented the data on these observations from 2005 to construct balanced beforeand-after observation windows. Thus, in this placebo test, our pre-regulation period ranged from 2005 to 2006, and the post-regulation period ranged from 2007 to 2008. This test can help us to rule out alternative explanations caused by unobservable factors. As shown in Table 7, the estimates for *Treated* × *Post* were all attenuated to zero and both coefficients were insignificant, thus suggesting that our main findings were not attributable to any unidentified factors.

**LnPatens LnInventions (1) (2)** *Treated* × *Post* 0.040 0.040 (0.070) (0.056) Treated 0.265 \*\*\* 0.188 \*\* (0.092) (0.074) Post 0.145 \*\* 0.102 \* (0.067) (0.057) Controls Yes Yes Industry fixed effects Yes Yes Observations 1861 1861 Adjusted *R*<sup>2</sup> 0.385 0.310

**Table 7.** Robustness Checks III—Placebo Test (2005–2008).

Note. This table reports the estimated coefficients and heteroskedasticity-adjusted robust standard errors clustered at the firm level (in parentheses). Significance at \* *p* < 0.1; \*\* *p* < 0.05; \*\*\* *p* < 0.01.

#### **5. Mechanisms**

The above estimates indicate that MSER were associated with an increase in firm innovation output, i.e., the number of patents and invention patents. In this section, we explore the underlying mechanisms behind this association. Three potential channels can explain why MSER influenced firm innovation activities: the CSR-improving effect, information-disclosing effect, and market-reaction effect. Below, we discuss each channel in detail and conduct estimates to verify them.
