國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
25
Chapter 4: Empirical Results
Basic statistics before matching
Our samples of S&P 500, added to and deleted from DJSI firms are 3301, 79 and 103 firm-years, respectively. Table3 reports basic statistics of the independent and
dependent variables. Based on samples before the matching, added-firms tend to have larger Asset and ROE, but lower financial leverage. For instance, the average ASSET and ROE of added and before-matching deleted firms are (70.44, 41.46) and (12.4%, 9.6%) billion US dollars, respectively; but they are 5.68 and 5.62 for the financial leverage, respectively. With respect to the market performance variables, the average cumulative abnormal return (CAR) of S&P500, added and deleted firms are -0.6%, -1.3%, and -1.0%, respectively. The results from basic statistics suggest that added-firms perform better than deleted-firms before-matching, especially on the accounting performance. However, when it comes to market performances S&P500 firms have the best results. According to Table3, the shunned and high-contaminated companies seem to have the tendency to enter and exit from the social index more frequently. The data shows as well that it’s more possible for companies to be added or deleted from the social index in the earlier years than recent years.
Table4 reports correlation matrix of accounting and market performance variables and dummy variable. For instances, dummy variable Dadd, , which equal to one if given sample is belong to added firms and equal to zero if it’s a S&P500 firm. The same principle can be applied to Ddelete, indicating those deleted firms if equal to one. From the first and second column we observe that, first, the correlation coefficient between Dadd and Ddelete is -0.0263. This number is close to zero because the large amounts of S&P500 firms are presented in samples, and they are all designated with the value
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
26
Table 3: Basic statistics of S&P500 Samples, samples of additions and deletions
S&P 500 firms Added firms Deleted firms
Mean Std. Dev. Min Max Mean Std. Dev. Min Max Mean Std. Dev. Min Max
Characteristic Variable
Asset 34.479 131.782 0.000 2187.630 70.441 179.485 0.670 1009.570 41.460 132.665 -0.020 887.520 Financial
Leverage 4.966 20.224 1.070 635.650 5.624 9.052 1.320 59.520 5.681 9.792 1.090 57.340 ROE 0.149 0.398 -6.296 9.305 0.124 0.296 -1.704 0.959 0.096 0.222 -0.797 0.920 Industry 0.920 0.272 0.000 1.000 0.855 0.354 0.000 1.000 0.871 0.337 0.000 1.000 Year 4.493 2.294 1.000 8.000 3.684 2.028 1.000 8.000 3.849 2.284 1.000 8.000 Performance Variable
CAR(1,-1) -0.006 0.034 -0.175 0.246 -0.010 0.037 -0.120 0.102 -0.013 0.036 -0.134 0.065 Note:
1. See Table 1 for the definitions of variables.
2. The unit of assets is billions of US dollars. The data of announcement date are applied.
3. There are totally 3301 S&P500 firm-year samples, where the samples of Added and Deleted firms are 79 and 103, respectively.
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
27
Table 4: Matrix of correlation coefficient
Dadd Ddelete Asset FL Industry ROE Year CAR
Dadd 1
Ddelete -0.0263 1
Asset 0.0369 0.0212 1
FL 0.0054 0.0063 0.096 1
Industry -0.0356 -0.0355 0.0268 -0.0088 1
ROE -0.0111 -0.0661 -0.0389 0.2122 -0.0284 1
Year -0.0416 -0.0433 0.0372 -0.0193 -0.0026 -0.0039 1
CAR(-1,1) -0.0177 -0.0257 0 0.0587 -0.0039 0.0066 0.1284 1.
Note:
1. See Table 1 for the definitions of variables.
2. The financial data of announcement date are used. There are totally 3134 observations.
‧
zero. Secondly, Asset and Dadd and Ddelete are 0.0369 and 0.0212, in line with the results we present in Table3. For financial leverage and industry, coefficient is relatively the same for both. With the respect of ROE, though both numbers are negative, the correlation coefficient between Dadd and ROE is much bigger than that of Ddelete and ROE. Thus, added-firms tend to have larger scale and higher ROE than deleted-firms. Thirdly, the correlation coefficients between Dadd, Ddelete and
cumulative abnormal return are both slightly negative, for examples, the correlation coefficients between Dadd and CAR(-1,1), between Ddelete and CAR(-1,1) are -0.0177 and -0.0257, respectively. Although one tends to conclude that added-firms slightly outperform deleted-firms using before matching data, we observe that there is also systematic divergence of independent variables between two groups of firms, at least on average. We cannot attribute performance difference solely to firms’ engaging to philanthropic activities and have to fix the differences in characteristics and get purer identifiable effect on market performance of firms entering DJSI, that’s the market performance difference between added- and deleted-firms at least not duo to difference in characteristics between them.
Basic statistics after matching
Table 5 presents the estimated results of propensity score function (PSF) by Probit model. We employ two models to examine the robustness. For
Model I with Dadd, contemporary characteristic variables are used to estimate PSF, and we observe that coefficients for Asset is significantly positive, suggesting that firms with large asset tend to be selected into DJSI. The coefficient for ROE is negative but insignificantly negative. As for Ddelete, the results are much the same, yet with smaller insignificant Asset and ROE. Companies are inclined to be deleted from DJSI with higher financial leverage, which may indicates DJSI review the index components by applying the risk criteria; on the contrary, a lower financial risk may
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
29
Table 5: Model I: Contemporary estimation of propensity score function, probit regression Pr (D
add=1)=F(β’x)
Probit Regression Numbers of obs = 3061
Log Likelihood = -349.01889 Psedo R2 = 0.0194
Dadd Coef. Std. Err. t P>|t| [95% Conf. Interval]
Asset 0.0005 0.0003 2.0800** 0.03800 0.00003 0.00100
FL 0.0004 0.0025 0.1700 0.86800 -0.00450 0.00533
Industry -0.3136 0.1497 -2.1000** 0.03600 -0.60691 -0.02024
ROE -0.1146 0.1368 -0.8400 0.40300 -0.38276 0.15365
Year -0.0569 0.0230 -2.4800** 0.01300 -0.10198 -0.01185
_cons -1.4650 0.1687 -8.6800 0.00000 -1.79570 -1.13433
Table 6: Model I: Contemporary estimation of propensity score function, probit regression Pr (D
delete=1)=F(β’x)
Probit Regression Numbers of obs = 3070
Log liklihood = -381.9537 Psedo R2 = 0.0173
Ddelete Coef. Std. Err. t P>|t| [95% Conf. Interval]
Asset 0.0004 0.0003 1.3300 0.1830 -0.0002 0.0009
FL 0.0010 0.0021 0.4600 0.6430 -0.0032 0.0052
Industry -0.3049 0.1450 -2.1000 ** 0.0360 -0.5892 -0.0206
ROE -0.1861 0.1248 -1.4900 0.1360 -0.4306 0.0584
Year -0.0547 0.0216 -2.5300 ** 0.0110 -0.0971 -0.0124
_cons -1.4187 0.1625 -8.7300 0.0000 -1.7372 -1.1003
The t-statistics are given here, and ***,**and * denotes the significance at the 1%, 5%, and 10% level, respectively.
‧
characterize the added firms. We also found that those companies related with shunned behaviors and high contaminations are easier to be presented in Dow Jones Sustainability Index than their counterparts, which could be possible due to a compensation mindset for pollutions which they caused. However, their CSR performances don’t seem to be stable or durable, because their coefficient between Ddelete and industry is also significantly negative. Finally, the correlations between year and both dummies are both significantly negative, in line with the results of Table1. This fact suggests that the phenomenon of higher turnover exists in DJSI components in early years, which could attribute to instability when the social index was just constructed in those starting years. Though there are some high correlation coefficients between each two independent variables, as we discussed earlier, the PSM is not the model of choosing determinants of CSR, but is to reduce the dimensions.
I use model I as our benchmark model to estimate propensity score function, that’s the timing of explanatory variables are all contemporaneous. One may concern that whether earlier characteristics affect the probability of being added to or deleted from the social index, for example, a firms with large earning this period could engage more in CSR activities in the next period because they have more available funds, suggested by McGuire et al. (1988) and Moore (2001). Thus, we also consider earlier factors, such as, Assett-1, ROE, and FLt-1 as explanatory variables to replace
contemporary ones to establish model II. Since the estimated results of PSF as roughly the same with results of model I (See Table 7 and 8), it confirms with the robustness we previously discussed.
Based on estimation results of PSF by model I, we obtain propensity score for each sample of firm. Thus, we can select the samples based on the nearest-neighbor
matching methods. Table9 compares the means of independent variables between two
‧
groups based on before- and after-matching samples. Unlike those in Table3 where the whole sample of S&P500, added firm and deleted firms are used, we only employ matched samples here. It is not surprising that, after nearest-neighbor matching, the average of each two pairs become approximately equal. In order to illustrate this, five characteristic variables between added and after-matching S&P500-firms become approximately equal because the differences in their means are insignificant according to t-statistics results. While in the situation between deleted and after matching
S&P500 enterprises, significant differences exist in Asset, but not in the remaining four variables. As mentioned earlier, the comparison of market performance is meaningful only when the characteristic variables are close to each other. It is worth noting that the observations of after-matching S&P500 companies still retain 76 and 83 in each case. They don’t lose much of the degree of freedom like applying other method such as Mahala Caliper in propensity score matching (Shen & Chang, 2009).
Market performance comparisons
Table10 presents the estimated results of regression analysis between dummy variable Dadd and cumulative abnormal return, and Table11 likewise shows the relationship between Ddelete and market performance. Employing the sample before matching, the estimated coefficients of both dummies are negative but only Ddelete is significantly negative. Thus, before matching, one may conclude that relations with DJSI could have negative impacts to market performances of CSR stocks no matter it’s in addition or deletion situation. In particular, passive socially responsible investments (SRI) funds could sell the deleted stocks while exits events regarding with violation of ethical criteria and cause the significantly inferior market performances. The results changed when the after matching samples are used. First, when nearest-neighbor matching is used, the estimated coefficient of Dadd and CARbecomes insignificantly positive, while the estimated coefficient of Ddelete maintains insignificantly negative.
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
32
Table 7: Model II: Estimation of propensity score function using early factors, probit regression Pr (D
add=1)=F(β’x)
Probit Regression Numbers of obs = 3081
Log Likelihood = -304.01279 Psedo R2 = 0.0204
Dadd Coef. Std. Err. t P>|t| [95% Conf. Interval]
Assett-1 0.0101 0.0040 2.5200** 0.0120 0.0023 0.0179
FLt-1 -0.0052 0.0160 -0.3300 0.7430 -0.0366 0.0261
Industry 0.0082 0.0455 0.1800 0.8570 -0.0810 0.0975
ROEt-1 0.0042 0.0014 2.8900*** 0.0040 0.0013 0.0070
Year -0.0660 0.0416 -1.5900 0.1130 -0.1476 0.0156
_cons 1.3013 0.8342 1.5600 0.0020 -0.3338 2.9363
Table 8: Model II: Estimation of propensity score function using early factors, probit regression Pr (D
delete=1)=F(β’x)
Probit Regression Numbers of obs = 3099
Log liklihood = -366.8517 Psedo R2 = 0.0166
Dadd Coef. Std. Err. t P>|t| [95% Conf. Interval]
Assett-1 0.9100 0.5400 1.6800* 0.0920 0.1500 1.9700
FLt-1 -0.0081 0.3571 -0.0200 0.9820 -0.7080 0.6918
Industry -0.0015 0.0431 -0.0300 0.9720 -0.0859 0.0829
ROEt-1 -0.0030 0.0359 -0.0800 0.9330 -0.0734 0.0674
Year -0.0462 0.0412 -1.1200 0.2620 -0.1270 0.0345
_cons 0.9064 0.8258 1.1000 0.2720 -0.7121 2.5248
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
33
Table 9: Descriptive statistics of independent variables: before- and after-matching samples
Before Matching
Added firms S&P 500 firms Differences (t-value) Deleted firm S&P 500 firms Differences (t-value)
Asset 70.441 34.479 35.962*(1.736) 41.460 34.479 6.981(0.500)
FL 5.624 4.966 0.658(0.597) 5.681 4.966 0.716(0.665)
Industry 0.124 0.149 -0.025(-0.724) 0.096 0.149
-0.053** (-2.219)ROE 0.855 0.920 -0.064(-1.573) 0.871 0.920 -0.049(-1.379)
Year 3.684 4.493 -0.809***(-3.425) 3.849 4.493
-0.644*** (-2.678)After matching (Nearest-neighbor matching)
Added firms S&P 500 firms Differences (t-value) Deleted firms S&P 500 firms Differences (t-value)
Asset 71.338 63.209 8.129(0.304) 57.308 24.582 32.726*(1.845)
FL 5.666 4.216 1.450(1.141) 5.752 4.683 1.069(0.909)
Industry 0.867 0.8667 0(0.000) 0.866 0.854 0.012(0.257)
ROE 0.124 0.095 0.030(0.719) 0.117 0.122 -0.005(-0.166)
Year 3.627 3.587 0.040(1.349) 3.621 3.597 0.024(0.498)
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
34
After matching (Caliper matching)
Added firms S&P 500 firms Differences (t-value) Deleted firms S&P 500 firms Differences (t-value)
Asset 51.928 48.357 3.571(0.138) 29.197 24.033 5.164(0.758)
FL 5.001 4.046 0.955(0.813) 4.764 4.836 0.072(0.072)
Industry 0.912 0.912 0(0.000) 0.905 0.919 -0.014(-0.376)
ROE 0.121 0.132 -0.011(-0.344) 0.119 0.124 -0.005(-0.150)
Year 3.721 3.589 0.132***(3.197) 3.784 3.757 0.027(0.575)
Notes:
1. The numbers presented are means of S&P 500, added and deleted-firms, respectively and their differences.
2. Before matching denotes the raw sample of the 3 groups of firms without adopting any matching methods. There are totally 3301 S&P500 firm-year samples, where the samples of Added and Deleted firms are 79 and 103, respectively.
3. After matching denotes the sample have been matched by adopting the nearest-neighbor and caliper matching methods. The numbers of observation for the added and deleted firms after matching are 76 and 83 under nearest-neighbor, and 66 and 69 under Caliper, respectively.
4. The t-statistics are presented in parentheses.
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
35
Second, significantly higher CAR (2.25) for the added firms under the matching of Caliper is in contrast with the insignificant lower result (-0.36) of the estimated coefficient of the deleted firms. Since most part of estimated coefficients of Dadd dummy are positive and Ddelete are negative, it implies that added-firms have superior market performance CAR(-1,1) than deleted-firms. Hence, adding to social index is beneficial for the value of firms in US, thereby supporting the social impact
hypothesis. My results suggest that companies’ value can be achieved through
recognized by reputational social index. Applying for DJSI membership appears to be a worthwhile investment when firms are seeking to demonstrate the sustainability of their business operations to shareholders. The increase in the firm’s market value may be large enough to offset the considerable effort and cost on the part of the companies that apply for inclusion on the DJSI.
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
36
Table 10: Performance regression of CAR (before- and after matching samples of added and S&P 500 firms) CAR = α + βD
add+ ε
CAR Coef. Std. Err. t P>|t| [95% Conf. Interval]
Before Matching
Dadd -0.004 0.004 -1.230 0.220 -0.011 0.003
_cons -0.006 0.001 -9.120 0.000 -0.007 -0.004
Nearest Matching
Dadd 0.005 0.006 0.870 0.384 -0.007 0.018
_cons -0.018 0.004 -4.160 0.000 -0.027 -0.010
Caliper Matching
Dadd 0.016 0.007 2.25** 0.026 0.002 0.029
_cons -0.021 0.005 -4.220 0.000 -0.031 -0.011
1. Before matching denotes the raw sample of the 2 groups of firms without adopting any matching methods. There are totally 3301 S&P500 firm-year samples, where the samples of added firms are 79.
2. After matching denotes the sample have been matched by adopting the nearest-neighbor and caliper matching methods. The numbers of observation for the added and S&P firms after matching are both 76 under nearest-neighbor, and 66 under Caliper, respectively.
3. The t-statistics are given here, and ***,**and * denotes the significance at the 1%, 5%, and 10% level, respectively.
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
37
Table 11: Performance regression of CAR (before- and after matching samples of deleted and S&P 500 firms) CAR = α + βD
delete+ ε
CAR Coef. Std. Err. t P>|t| [95% Conf. Interval]
Before Matching
Ddelete -0.007 0.004 -1.860* 0.063 -0.015 0.000
_cons -0.006 0.001 -9.130 0.000 -0.007 -0.004
Nearest Matching
Ddelete -0.029 0.024 -1.220 0.228 -0.076 0.019
_cons 26.858 15.719 1.710 0.093 -4.643 58.358
Caliper Matching
Ddelete -0.002 0.006 -0.360 0.721 -0.014 0.009
_cons -0.014 0.004 -3.390 0.001 -0.028 -0.006
Notes:
1. Before matching denotes the raw sample of the 2 groups of firms without adopting any matching methods. There are totally 3301 S&P500 firm-year samples, where the samples of deleted firms are 103.
2. After matching denotes the sample have been matched by adopting the nearest-neighbor and caliper matching methods. The numbers of observation for the deleted and S&P firms after matching are both 83 under nearest-neighbor, and 69 under Caliper, respectively.
3. The t-statistics are given here, and ***,**and * denotes the significance at the 1%, 5%, and 10% level, respectively.