24
CDP response status is AQ, AQL, or AQSA, then we code the firm as 1, and 0 otherwise.
19
To test H2, we classify each sample firm as in an industry where users i.e., SASB’s panel of experts judge CCR as either material coded = 1 or not material coded = 0.
V. RESULTS
Descriptive Statistics
Table 4 provides summary statistics for the variables in Equation 1. We winsorize all continuous variables at the one percent level on both tails of the distribution. Panel A of Table 4
shows that the mean median COE is 8.14 percent 8.05 percent.
20
The firms’ mean median BETA is about 1.15 1.07, which is consistent with the relatively low risk of SP 500 firms in
general. The firms’ mean median BM is 0.522 0.423, indicating that on average, the firms are healthy and have growth opportunities. For a few firms, foreign income represents a large
proportion of their total income. The mean FIPI is 30.2 percent, although the median is only 12.1 percent. The first three quartiles of the EXCH variable are 1. This reflects the composition
of our sample, whereby 2,417 firm-years 80.7 percent trade on NYSE coded = 1, and 19.3 percent trade on NASDAQ coded = 3 untabulated.
21
The mean STRNG and CNCRN is 1.021 and 0.454, respectively. Finally, about 65 percent of the firms participated in the CDP climate
survey and allowed their responses to be publicly available. Panel B of Table 4 shows summary statistics and univariate tests for the variables in
19
CDP uses the following response status legend: AQ: Answered the survey; AQL: Answered the survey late; AQSA: Answered the survey but the company is a subsidiary or has merged; NP: Answered the survey but the
response is not publicly available; IN: Information provided; DP: Declined to participate; NR: No response; X: the company did not fall into the CDP sample that year.
20
Damodaran 2015 estimates an average risk premium of 2.62 percent over the 2008 −2014 period for SP 500
firms using the dividend discounting DD model, and 5.50 percent using the free cash-flows-to-equity FCFE approach. With an average risk-free rate of 2.60 percent over this period, this translates into a COE of 5.22 percent
and 8.10 percent for the DD and FCFE approaches, respectively.
21
Two-thirds of the firms listed on the NYSE disclose CCR information in Form 10-K; in contrast, 37 percent of the firms listed on NASDAQ disclose this information in Form 10-K.
25
Equation 1, partitioned by whether the firms disclose or do not disclose CCR DISC_10K = 1 and DISC_10K = 0, respectively. In general, except for FIPI, the disclosers are significantly
different from the non-disclosers. Both the mean and median COE are significantly higher for the disclosers than for the non-disclosers p 0.05 and p 0.10, respectively. Although both the
mean BETA and BM are higher for the disclosers p 0.05, the median BETA is not significantly different between disclosers and non-disclosers. The significantly higher mean
BETA and BM for the disclosers suggests that, in general, these firms are riskier on these dimensions and therefore may have a higher COE than the non-disclosers. The mean and median
SIZE are significantly higher for the disclosers than for the non-disclosers p = 0.00. Contrary to expectation, the mean and median ROA are higher for the non-disclosers than for the disclosers
p = 0.00. Taken together, our univariate results reinforce the importance of correcting for self- selection. That is, as discussed earlier, using data from the disclosing firms to draw inferences
about the non-disclosing firms, without first correcting for these differences, will likely lead to biased coefficients and thus, erroneous conclusions.
Panel C of Table 4 shows summary statistics for the variables in Equation 1, partitioned by user-based SASB materiality judgments. The median COE is higher for firms in the material
CCR group than those of firms in the not-material CCR group, but the difference in means is not significant. The material CCR firms also have higher BM and are larger than the not-material
CCR firms, but have lower ROA. Further, material CCR firms also have higher STRNG and CNCRN scores.
| Insert Table 4 about here | Table 5 presents correlation coefficients for the variables in Equation 1. The tables
show Pearson and Spearman rank correlations below and above the diagonal, respectively. COE
26
is positively correlated with both DISC_10K mandatory disclosure and CDP voluntary disclosure p 0.10. Further, COE is correlated with all the other variables in our regression
model p 0.01 or better, except FIPI. The signs for all the correlations are as expected, except for the positive correlation between COE and SIZE Spearman rank = 0.243; p 0.01. This
result is consistent with Dhaliwal et al. 2011 and may be due to our sample firms drawn from the SP 500 index, which are among the largest in the world.
22
DISC_10K is significantly correlated with both STRNG and CNCRN, consistent with our descriptive statistics in Panels B
and C of Table 4. Interestingly, the correlation between DISC_10K and CDP, although highly significant, is small 0.057, p 0.01.
| Insert Table 5 about here |
Hypothesis 1 Tests
Table 6 presents the results of Equation 1 matching the firms that disclose CCR in Form 10K DISC_10K = 1 with those that do not DISC_10K = 0 on various firm-level
characteristics Panel A, and our tests of H1 to examine the COE effect of disclosing vs. not disclosing CCR after propensity score matching PSM Panel B. Panel A shows that, of the
total 2,996 firm-year observations, we are able to match 2,966 observations: 1,770 disclosers matched with 1,196 non-disclosers. We are unable to match 30 disclosers. Before matching, the
two groups of firms were significantly different on all but one of the firm characteristics included in Equation 1 see Table 4, Panel B. After matching, only four firm-level variables remain
significantly different between the two groups, BM and SIZE at p 0.05, and STRNG and CDP at p 0.01 Table 6, Panel A, Covariate Balance.
| Insert Table 6 about here |
22
See also Easton 2007 for a discussion of assessing the validity of COE measures using associations or correlations with other known risk factors, such as BETA and SIZE.
27
Panel B of Table 6 shows the t-tests of differences in the COE of matched disclosers versus non-disclosers. The difference in COE is positive and significant p 0.05; that is, the
COE of the disclosers is higher than the COE of non-disclosers before matching. However, after matching, the COE for the disclosers is lower than that of the non-disclosers, but the difference is
not statistically significant p 0.10. As discussed above, even after propensity score matching, our matched sample is
significantly different on four dimensions. Further, the standard errors from the PSM may not be unbiased. Therefore, to remove any residual misspecification that may remain after matching, we
estimate a doubly robust regression Imbens and Wooldridge 2007, clustering the standard errors on firm identifier Permno. Panel C of Table 6 shows that the difference in COE between
the DISC_10K = 1 and the DISC_10K = 0 firms is negative and significant p 0.05: the COE
of disclosers is approximately 21.3 bps lower than the COE of the non-disclosers, thus rejecting our null hypothesis H1 of no difference between the COE of disclosers and non-disclosers.
Hypothesis 2 Tests
Our tests of H2 examine the role of CCR materiality, as judged by users i.e., SASB’s panel of experts, on the association between disclosing CCR and COE. The PSM results in
Table 7, Panel A show that the matched sample for the material CCR firms has differences along the three risk dimensions, BETA, BM, and SIZE, as well as the two environmental performance
measures, STRNG, and CNCRN. We are able to match 1,095 firm-year observations out of the total 1,138: 809 disclosers to 286 non-disclosers. We are unable to match 43 disclosers.
Panel B of Table 7 shows the tests of differences in COE between matched disclosers and non-disclosers, partitioned by user-based SASB materiality. For the matched sample of the
material CCR firms SASB_MTRL=1, the COE of DISC_10K = 1 firms is lower than the COE of
28
DISC_10K = 0 firms, but the difference is not significant p 0.10. However, the doubly robust regression results in Panel C show that, for the material CCR group, the difference in COE
between the disclosers and non-disclosers is negative and significant p 0.05: the COE of disclosers is 49.1 bps lower than the COE of non-disclosers.
Next, we discuss the results for the not-material CCR firms SASB_MTRL=0. Panel A of Table 7 shows that the matched sample of 1,793 firms, 912 disclosers and 881 non-disclosers,
differs along three dimensions after matching: BETA, ROA, and STRNG. We are not able to find matches for 16 disclosers. Panel B of Table 7 shows that, for the not-material CCR group of
matched firms, the COE of disclosers is higher than the COE of non-disclosers, but the difference is not statistically significant p 0.10. Similarly, the doubly-robust regression
results Panel C, Table 7 show no statistical difference in the COE of disclosers versus non- disclosers for the not-material CCR group. Taken together, our results support H2.
| Insert Table 7 about here | In summary, our H1 results are consistent with lower COE for firms that disclose CCR,
compared to firms that do not disclose CCR. In addition, our H2 results indicate that disclosing CCR is associated with lower COE only for firms in industries where users judge CCR as
material. For firms where users judge CCR as not material, we find no association between disclosing CCR and COE. Overall, our results indicate that, on average, investors impose a risk
premium on firms that do not disclose CCR in their 10-K filings. However, after partitioning the sample on materiality of CCR from the users’ perspective, we find that this risk premium exists
only for firms where users judge CCR as material.
29
Sensitivity and Robustness Tests
Prior research finds that firms with better corporate governance have higher firm value and stock returns Gompers, Ishii, and Metrick 2003. To control for the effects of corporate
governance on COE, we construct a variable, CGOV, to proxy for firms’ climate-change governance measures. We obtain corporate governance data from Bloomberg on three separate
dimensions: Does the firm have: 1 a climate change policy; 2 a climate change committee; and 3 incentives tied to climate change management? We code each dimension as equal to one
if the firm answers “yes,” and zero otherwise. We add the scores on the three questions to construct the CGOV variable.
23
Tables 8 and 9 show our results for the full sample, and broken down by materiality, respectively. Our doubly robust regression results are consistent with our
main results. After matching on all firm-level variables, including corporate governance, we find a negative association between disclosing CCR and COE for the full sample Table 8, Panel C,
and for firms where users judge CCR as material Table 9, Panel C. We also test H1 and H2 after including industry fixed effects in our models untabulated.
We use the Fama-French five-industry classification for industries. Although we are able to find matches for more observations for the full sample, we are unable to match on six of the nine
firm-level variables. Our results remain unchanged. The COE coefficient for the full sample in the doubly robust regression shows that the COE for disclosers is 18.3 bps lower than the COE
for non-disclosers p 0.05. Our results after partitioning on user-based materiality judgments also remain unchanged. For firms in the CCR material group, the COE of disclosers is 73.7 bps
lower than the COE of non-disclosers, and the difference is significant p 0.01. In contrast, for
23
The breakdown of the scores untabulated for our 2,996 observations is: 0 42 percent, one 16 percent, two 24 percent and three 18 percent.
30
firms in the CCR not-material group, we find no significant difference in the COE of disclosers versus non-disclosers.
Although our hand-collected data on firms’ CCR disclosures includes 2015, the time period for our main analyses ends in 2014 because we do not have 2015 data on firms’
participation in the CDP climate survey. However, firms’ participation in the survey is sticky; that is, once a firm participates in the CDP survey, it is likely to continue to do so in subsequent
years. In our sample period, less than 10 percent of the firms change their reporting status from one year to the next. In addition, the correlation between CDP reporting status in 2013 and 2014
is 0.85 p 0.01. Consequently, we test H1 and H2 by extrapolating firms’ CDP reporting status for 2015 using CDP 2014 data untabulated. Our sample size increases to 3,395 firm-year
observations i.e., an increase of 399 observations relative to our main results, of which we are able to match 3,376 observations 2,045 disclosers to 1,331 non-disclosers. The results from the
doubly robust regressions are stronger relative to our main results. The COE of disclosers is 24.4 bps lower than that for non-disclosers p 0.01. For firms in the material CCR group, we are
able to match 927 disclosers with 316 non-disclosers. The doubly robust regression results show that the COE for the disclosers is 54.6 bps lower than the COE for non-disclosers, and this
difference is significant p 0.01. For firms in the not-material CCR group, the difference in COE between disclosers and non-disclosers is not significant.
In our next sensitivity analyses, we look at changes in whether CCR is disclosed in 10- K’s untabulated. The results of these analyses need to be interpreted with caution since only
about 5 percent of our sample firms i.e., about 155 observations change their disclosure practices. We find that firms which start disclosing CCR experience a decline in COE, but the
coefficient is not statistically significant. However, firms that stop disclosing CCR experience an
31
increase in COE for the full sample and for firms where users judge CCR as material p 0.05. Our results are not significant for non-material CCR firms.
Next, instead of matching on firm performance using ROA, we match on whether a firm suffered a loss during the year untabulated. We include an indicator variable equal to 1 if the
firm reported negative income before extraordinary items during the year, and 0 otherwise. Our results are inferentially similar to our main results. Finally, we calculate implied COE as the
average of the four COE measures, instead of the median of the four measures untabulated. Our results are inferentially similar to our main results.
VI. CONCLUSION
