國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
17
Chapter 3: Data and Methodology
Data and choice of analysis period
To examine the question of whether the capital markets view being included on the DJSI as a credible signal of sustainability thus adding value to companies, I measure the stock market reaction to the announcement that American firms were added to, or deleted from, the DJSI over the period 2002-2009. Obtained from DJSI website, Table1 provides summary data about the announcement and effective dates for DJSI index changes as well as information about the number of American firms that were added or deleted in each year. Otherwise, I use S&P500 firms to pair-match with these added and deleted firms. That is to say, S&P500 companies are used as control group. Firms are matched using data from CompuStat database with all numbers being denominated in US dollars.
Matching methods: propensity score matching
An alternative to the regression approach is to use matched pairs of firms to examine if there is a market performance differential between firms with the same scores.
Traditionally, researchers attempt to isolate the variable of interest by matching firms based on other characteristics that also drive the dependent variable. Following the work of Fama and French (1993) matching is often done on the basis of size and book to market ratio. Control firms are sorted into bins based on size and then further subdivided based on their book-to-market ratio. Each firm in the treatment group is then matched to the firm whose characteristics more closely match its own. The
difficulty with this approach is that it is sensitive to the order in which the matching is
‧
Table 1: Changes in USA Stocks in the Dow Jones Sustainability Index (DJSI) from 2002-2009
(Include the overlapped ones, existing both in Additions and Deletions list) Year Additions Deletions
Announcement
done and the number of criteria used. In order to minimize the likelihood of
mismatching firms confounding the results, we use propensity scoring as the method for matching. A propensity score is the conditional probability of assignment to a particular treatment given a vector of observed covariates.
Starting with a Probit regression to obtain the estimated propensity score Pr (Di = 1|Xi) for all separating 2 firm groups, which are denoted as Pi and Pj, respectively. The predicted probability is used as the single matching criterion between the treatment group and the control group. Propensity score matching is done by sampling from a large reservoir of potential firms (the controls) and finds those whose characteristics lead to them having the same propensity score as the firm in the treatment group. The advantage of this method is that it produces matched pairs that are similar on multiple dimensions. Our methodology is as follows: First, we calculate the propensity score
‧
for all firms. Then, we sort both the treatment group and the control group by the propensity score. We start with the first firm in the treatment group and match it with the first firm in the control group whose propensity score matches to four significant digits. If more than one control firm is a potential match, a random number generator selects the control firm. Both the treatment and the control firm are then placed in a file of matched firms, so no control is matched to more than one treatment firm. If no control firm matches to four significant digits, the treatment firm is not matched and the selection process moves to the next firm. Matching continues until all treatment firms have been matched or discarded. We expect the two sets of control variables to be identical after matching. I conduct all the statistics results on the platform of software STATA10.
Two criteria are often suggested in the literature to find the approximation. Here in this thesis we apply the Nearest-Neighbor Matching (Nearest hereafter), which matches each treatment sample with the control sample such that the difference of the two is minimized. That is,
C(Pi) = minj �Pi− Pj�
where C(P
i)
is a set of control units matched with the treated unit i, i.e., samples that has the nearest propensity score with added firm i.The second criterion is the caliper matching (Caliper hereafter), which requires the two groups to be not too distant. That is, it requires the propensity score of the
added-firm (deleted-firm) and S&P500-firm to fall within a prespecified caliper. That is,
C(Pi) = �Pi− Pj� < 𝜂
where η is a very small number and is specified as quarter of standard error of estimated propensity scores for all added-firms (deleted-firms) and S&P500-firms.
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
20
Note that under this criterion, the number of control samples contained in
C(P
i)
is arbitrary, from zero to all control samples.Confirmation of matching result
Once the matching S&P500 firms are selected, we could verify the matching by performing the following test: This test examines the null hypothesis of
, where are average of characteristic variables of treated firm i and control firm j, respectively. The two groups have the same characteristics if the null is not rejected.
Regression analysis to estimate the treatment result
Our independent variables are asset, financial leverage, ROE, industries and year, explained in detail in the following and also showing in Table 2.
Asset: Asset can be interpreted as scale, Fombrun and Shanley (1990) pointed out that the larger the scale, the more the attention that a company attracts from the public. Thus, the response from its philanthropic activities is noticeable.
Scale is often considered as a crucial characteristic variable and its effect on the probability of adopting CSR is positive.
Financial leverage: Return on Average Equity divided by Return on Average Assets. The firm's default risk as measured by financial leverage would affect the cumulative abnormal return (DHALIWAL, LEE, & FARGHER, 1991).
ROE: McGuire et al. (1988) proposed the available funds theory to argue that firms with abundant resources have more ability to engage in CSR activities.
Posner and Schmidt (1992), on the contrary, prove that firms with earn sufficient profits could have egocentric behaviors without fear of being challenged for not
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
21
Table 2: Abbreviation and definition of dependent and independent variables
Variable Definition
Independent Variables
Dadd A dummy variable which equal to one if firms were added into DJSI. Other S&P 500 components were equal to zero.
Ddelete A dummy variable which equal to one if firms were deleted from DJSI. Other S&P 500 components were
equal to zero.
Asset This item represents current assets plus net property, plant, and equipment plus other noncurrent assets.
FL Financial leverage, Return on Average Equity divided by Return on Average Assets.
ROE Return on equity, net income divided by total assets.
Industry A dummy variable which equal to zero if firms are among shunned and high contamination industries. Others are equal to one.
Year Indicated the year companies were added into or deleted from DJSI. Year 2002 is designated as year 1, and the following year 2003 as year 2, and the rest may be deduced by analogy.
Dependent Variables
CAR (-1,1) Cumulative abnormal return (CAR) in the intervals (-1,1), indicating stock performance, with 0 being the announcement date.
Note:
1. The definitions of shunned and high contamination industries are from KLD definition (ex alcohol, tobacco, gambling, armaments &
firearms and adult entertainment) and Cowen et al., respectively.
2. These independent and dependent variables are acquired from Compustat database.
‧
noticing the interests of the public. ROE’s effect toward inclusion in or exclusion from the index is uncertain, and we hope to find the relationship in between.
Industry: A dummy variable which equal to zero if firms are among shunned and high contamination industries. Others are equal to one. In our study, shunned companies are those that KLD classifies as associated with at least one of the following: tobacco, alcohol, gambling, firearms, military, or nuclear operations.Similar to Cowen et al. (1987), the chemical (SIC code 28xx, excluding pharmaceutical firms, code 283x), metals (33xx), paper (26xx), and petroleum (2911) industries are defined as more environmentally sensitive. The activities of such companies violate social norms, and some socially responsible indices avoid such stocks even if they yield higher returns than stocks in other industries.
We set this variable to explore the relationship between DJSI and CSR
controversial companies, and it can also serves as a control for various kinds of industries involved in DJSI.
Year: Indicated the year companies were added into or deleted from DJSI. Year 2002 is designated as year 1, and the following year 2003 as year 2, and the rest may be deduced by analogy. This control variable is resulted from the
insufficient samples we can collect within each year, therefore, we pool all the samples together and designate each a “year” variable.
Once obtaining the estimated propensity score Pr (Di = 1|Xi) for all firms, our
objective is to find the S&P500 firms whose propensity scores are sufficiently close to those of added firms or deleted firms by applying the nearest-neighbor and Caliper matching method.
Finally, I employ the regression analysis with Dadd and Ddelete dummy variable to examine the differences of market performances between three groups of firms. The regression model is
‧
Where Dadd are dummy variables which equal to one if firms were added into DJSI.
Other S&P 500 components were equal to zero.α, β and ε are coefficients to be estimated, and ε is error term. Estimated β captures the effect on firms’ market
performance of being philanthropic. A significant positive β suggests that added firms have better market performance than S&P500 firms, supporting the social impact hypothesis, whereas a significant negative β supports the shift of focus hypothesis.
The opposite interpretation would be applied if Ddelete is substituted in the above equation. Ddelete is the other dummy variable which equal to one if the firm was deleted from DJSI. Companies assigned with Dadd also represent a better sustainability than those assigned with Ddelete..
We use cumulative abnormal return as the proxy for market performance because index inclusion and index exclusion may affect market performance in the way of abnormal stock returns. This methodology computes the abnormal return on each added and deleted stock, equal to the return on the stock minus the return on major national index (the S&P 500 in the US), during every day in an announcement event window that includes three days (Day -1 until Day +1).
As already observed, one of the main limits of all the above mentioned analyses based on accounting data is the difficulty of controlling for endogeneity. In the
CSR-corporate performance relationship the problem is particularly severe as it is important to discern, for instance, in case of positive relationship, whether the move to CSR is an autonomous driver of improvement in corporate performance or, quite to the opposite, high cash flow and better performing firms are more likely to choose CSR, due to their higher cash flow availability. A second, almost insurmountable, limit is that accounting based analyses on the CSR-corporate performance nexus do not provide a risk adjusted measure of performance.
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
24
The two advantages of investigating the impact of CSR on corporate performance in financial markets are stated as followed. By calculating cumulative abnormal returns at the announcement date, we pick up the expected net effect of entry into/exit from CSR and hence separate the effect of change in CSR on corporate performance from the reverse causality effect. Furthermore, we may calculate it net of measurable risk factors.
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
25