University of Amsterdam
Amsterdam Business School
MSc Business Economics, Finance track
Master Thesis
Do Investors Walk the Walk?
The Effect of Institutional Ownership on Corporate Social
Performance
05-07-2016
Vincent Christiaan Altorffer
Statement of Originality
This document is written by Vincent Christiaan Altorffer, who declares to take full responsibility for the contents of this document. I declare that the text and the work presented in this document is original and that no sources other than those mentioned in the text and its references have been used in creating it. The Faculty of Economics and Business is responsible solely for the supervision of completion of the work, not for the contents.
Acknowledgements
I would like to offer my special thanks to dr. Torsten Jochem for his supervision, and his help in providing the necessary proprietary Russell and institutional holdings data. His willingness to give valuable feedback and make time so generously has been much appreciated. I would also like to thank my parents, friends, and fellow students for their continued support and for making this past year an unforgettable experience.
Abstract
This paper uses the Russell indices’ inclusion/exclusions in a regression discontinuity and simple instrumental variable design to determine the causal effect of institutional ownership on corporate social performance. Using a sample of 1600 firm-years from 2004 to 2013 I find that institutional investors do not have a statistically significant causal effect on corporate social performance, apart from the corporate governance performance dimension, which is negatively affected by more institutional ownership. On average, firms in the top 80 firms of the Russell 2000 index have 37% more institutional ownership which leads to a 0.211-point increase in the average number of corporate governance related concerns. This means that they either have more controversial governance structures, more controversial investments, are more involved with bribery and fraud, have more other governance related concerns, or experience an increase in any combination of these factors. I also find evidence that institutional investors do employ social activism to try to influence the corporate social performance outcome of their investees. Firms in the top of the Russell 2000 on average receive 2.771 more SRI-related proposals per firm-year than peers with less institutional ownership. My findings are robust with respect to different bandwidth sizes and placebo tests, and the effect becomes stronger the closer firms are to the cut-off between both indices, where the discontinuity in institutional ownership is the most pronounced.
Table of contents
1 Introduction ... 5
2 Literature review and hypothesis development... 7
2.1 Rationale for better social performance ... 7
2.2 The trade-off between screening and engagement ... 8
2.3 Empirical debate ... 9
2.4 Hypothesis development ... 11
3 Methodology and data ... 12
3.1 Methodology ... 12
3.1.1 Russell-indices ... 12
3.1.2 Regression discontinuity approach ... 13
3.1.3 Instrumental variable approach ... 14
3.2 Data ... 16 3.2.1 Dataset construction ... 16 3.2.2 Summary statistics ... 18 4 Empirical results ... 21 4.1 Results ... 21 4.1.1 Responsible investment ... 22 4.1.2 Window dressing... 28 4.1.3 Activism ... 32 4.2 Robustness checks ... 34
4.2.1 Placebo test for discontinuity ... 34
4.2.2 Bandwidth and index dependence ... 34
4.2.3 Activism full sample test ... 35
5 Discussion and conclusion ... 36
6 References ... 39
1 Introduction
Socially Responsible Investing (SRI) has long been regarded as an investment style employed by fringe activist groups and NGOs, rather than by mainstream institutional investors whose goal it is to maximize returns (Sparkes & Cowton, 2004). However, since the early 2000s, this attitude towards SRI has been changing, primarily due to a changing attitude towards social responsibility by large companies, who in turn responded to changing consumer awareness on social issues (Sparkes & Cowton, 2004; Mohr et al., 2001). A clear signal of this changing attitude was when in early 2005, a group of the largest institutional investors were invited by then UN Secretary-General Kofi Annan to co-develop the Principles of Responsible Investment (PRI) (PRI, 2016). Currently, 1453 signatories have underwritten the principles, who together own and/or manage over $59 trillion in assets (up from $4 trillion at launch in 2006), which represents almost 80% of the world’s total assets under management as of July 2015 (BCG, 2015; PRI, 2016). This certainly seems to suggest that institutional investors place an ever greater interest in incorporating corporate social responsibility (CSR) into their investment decisions. It is the performance outcome of these CSR-policies that this paper focusses on, which is known as corporate social performance1 (Clarkson, 1995).
Although underwriting of the PRI is on the rise, and one of the six principles calls for investors to be active owners who promote CSR within the firms they invest in, the empirical evidence on the association between institutional ownership and corporate social performance is mixed. Furthermore, although more recent studies do find a positive association between institutional ownership and corporate social performance, there is only limited suggestive evidence that this actually signifies a causal relationship (Dyck et al., 2015). These mixed empirical findings are not surprising from a theoretical viewpoint, since two conditions have to be met for a causal effect of institutional ownership on corporate social performance to exist. First of all, institutional investors must have incentives to pursue better social performance in their investees, which means that social performance may not hurt their main goal of pursuing financial returns. Second of all, the trade-off that institutional investors face between social screening (passive) and engagement (active) should be decided in favour of active engagement2. If not, institutional owners do not affect social performance,
but social performance would affect the amount of institutional ownership.
The current empirical literature also suffers from two important potential problems: simultaneous or reverse causality, and the possibility of spurious regression results. Simultaneous or reverse causality problems stem from the aforementioned trade-off that institutional investors face: if screening is being used, the association would be the result of a causal connection going from
1 Precisely how corporate social performance (CSP) is measured, is outlined in section 3 of this paper.
2 It is important to note, that even tough social screening is passive, it could also affect social performance under the
right conditions. If the financial incentives for firms and/or managers from increased institutional investment through better social performance are large enough (and this mechanism is known), it could be that the threat of exit/promise of entry from institutions is strong enough to cause better social performance in itself. Therefore, I also test the proposed engagement (activism) channel (see also section 2.4).
corporate social performance to institutional ownership; if institutional investors use engagement, the association would indeed signify a causal link going from institutional ownership to better social performance. Alternatively, the two effects could also be going on at the same time: institutional investors could use an optimal ‘mix’ of screening and engagement. Spurious regression results could arise from a time-series analysis due to the simultaneous rise of institutional ownership and corporate social performance over the last several decades (Aghion et al., 2013).
This current lack of understanding on the influence of institutional ownership gives cause for further research into this important topic within the broader corporate governance context. As sustainability and ethics become ever more important within society, knowing whether institutionally invested capital is allocated in a responsible manner also becomes ever more important for a variety of stakeholders including consumers, firms, and most importantly institutional owners themselves. Thus, the goal of this paper is to find out if institutional owners are a driving force behind increased corporate social performance in firms by answering the question: does increased institutional ownership cause an increase in corporate social performance?
To find this causal relationship, I employ two different approaches: a regression discontinuity approach, and a simple instrumental variable approach. The two methods both use the discontinuity in institutional ownership between firms in the bottom of the Russell 1000 index and the top of the Russell 2000 index. This discontinuity exists due to the fact that the Russell indices are market value weighted, where the first stock has the highest weight in the index, and the last stock has the lowest weight. Together with the fact that the Russell 2000 index is more popular with institutional investors, this means that the first firm in the Russell 2000 index on average has ten times more institutional ownership (as a percentage of total outstanding shares owned by institutions) than the last firm in the Russell 1000 (Chang et al., 2014; Lu, 2013). I can use this discontinuity to research whether corporate social performance is affected by institutional owners. The validity of using this discontinuity is also supported by the balance table presented in table 8 in the appendix, which shows that the firms on both sides of the cut-off are not statistically significantly different from each other, apart from the market capitalization and institutional ownership parameters3.
The dataset on which the two methods are used spans the time period of 2004 to 2013 and consists of the proprietary yearly Russell constitutions, institutional 13f holdings from the Thomson Reuters Institutional (13F) Holdings database, corporate social performance scores from the MSCI ESG KLD database, and proposal data from the ISS (RiskMetrics) database. The full sample consists of 29,818 firm-years, which is restricted to a bandwidth of 80 firms on either side of the cut-off based on Lu’s (2013) lowest estimate of the ideal bandwidth which balances effect strength and statistical
3 The statistical difference in market capitalization and institutional ownership between both sides of the cut-off is in
line with my expectations. The difference in market capitalization is the result of the ranking system used to compute the Russell indices (firms are ranked according to market capitalization), and the difference in institutional ownership is the result of the market value weighting in both indices.
power. This restriction results in a dataset of 1600 firm-years over a 10-year period.
My main finding is that institutional owners do not seem to have a positive influence on the social performance outcomes of the firms they invest in, but rather negatively affect the corporate governance dimension of social activism. In particular, firms with a higher percentage of institutional ownership experience a higher number of concerns, meaning that they either have more controversial governance structures, more controversial investments, are more involved with bribery and fraud, have more other governance related concerns, or experience an increase in any combination of these factors. However, institutions do appear to engage in (social) activism as evidenced by the statistically and economically significant increase in SRI-related proposals in firms that are delisted from the bottom of the Russell 1000 and listed in the top of the Russell 2000. These results present further empirical backing for the findings of David et al. (2007) who give an explanation for this empirical finding by theorizing that managers who are pressured by shareholders will divert resources from corporate social responsibility towards political activities and attention diverting behaviours. These diversions include symbolic actions that do not, in the end, have a positive impact on the performance outcome of implemented corporate social responsibility measures.
I will start by reviewing the recent relevant literature and developing my hypotheses in section 2. The methodology is described in section 3, which also contains a description of the data used and how the dataset was constructed. Section 4 contains the empirical results, including a number of robustness checks. Finally, the discussion and conclusion are presented in section 5.
2 Literature review and hypothesis development
This section provides the literary background to the subsequent empirical tests by providing an overview of the rationale for social improvement and the barriers faced by institutional investors, a discussion on the trade-off between screening and engagement, and a review of the recent empirical debate.
2.1 Rationale for better social performance and barriers for institutions
Carroll and Shabana (2010) present a broad review of the most important studies on the ‘business case’ for engaging in corporate social responsibility of the last decades. According to them, firms engaging in CSR-activities (and thus have better social performance) seem to be rewarded by the market, which in turn increases financial performance. This happens in two ways: directly through cost savings, and indirectly through enhancing the firm’s competitive advantage, reducing risk, and creating reputation benefits. Although the notion that firms are financially rewarded for engaging in CSR is not supported by all research, there exists a fairly broad consensus that better social performance will at least not negatively affect financial performance, and will in most contexts improve it (Carroll and Shabana, 2010). Further support comes from recent empirical literature, such
as Godfrey, Merrill and Hansen (2009), who find that CSR policies act as a kind of insurance policy which increases shareholder value; Servaes and Tamayo (2013), who find that firm value increases through increased engagement with CSR when customer awareness is high; Ioannou and Serafeim (2015), who present evidence that in recent years sell-side analysts value good social performance firms higher; and Flammer (2015), who uses a regression discontinuity approach and finds that the adoption of close-call CSR proposals lead to positive return announcements. This consensus, as well as the recent supporting evidence, provides the most important rationale for institutional owners to drive their investees towards better social performance.
However, three important practical barriers exist for institutional investors to de facto coherently act as a group to improve the social performance of their investees. One such barrier is differing investment horizons, which could lead to differing commitments to corporate social performance (Dyck et al., 2015). For instance, short-term holders could be uninterested in pursuing long-term gains from better corporate social performance due to the short-term investment requirements (Cox et al., 2004). Collective action problems form a second barrier against coherent action among investors, since each individual institution has an incentive to wait for other investors to take action instead (Dyck et al., 2015). Even if all institutional investors had similar investment horizons and could overcome collective action problems, differences in opinion would still form a third barrier (Dyck et al., 2015). Some investors might care more for certain dimensions of corporate social responsibility (such as the environment or human rights) than other dimensions (such as employee relations) and vice versa. These three barriers create ex-ante uncertainty as to whether institutional investors as a group will in practice be a driving force behind better social performance. 2.2 The trade-off between screening and engagement
Next to the practical barriers to coherent action-taking, investors also face a trade-off between using active engagement on the one hand, and a passive method to improve the social performance of their portfolios called “social screening” (Hill et al., 2007). This means that they simply direct more investments towards firms that already have good social performance, rather than actively engaging with the firms they invest in. Spencer (2001) describes a variety of these screens designed to determine whether or not a certain firm is included based on meeting or failing certain criteria (inclusionary and exclusionary). More recent literature on SRI includes research by Berry and Junkus (2013), who find that investors are inclined to reward those firms that act responsibly rather than punish those that do not, and Oh et al. (2013), who examine the way in which investors use rankings and indices to incorporate corporate social performance into their investment decisions.
This is not to say that institutional investors only passively allocate capital using social screens. For instance, next to examining the way investors allocate capital responsibly, Oh et al. (2013) also touch upon the possibility that investors actively encourage the firms they invest in to act more socially responsible. In fact, according to Dyck et al. (2015) institutional investors face a
trade-off when considering social screening versus active engagement (activism). On the one hand, social screening might be a cost-effective way to direct investments towards more socially responsible investments, since there are no direct costs of engaging with investees. On the other hand, there are indirect costs associated with both negative screening (blacklisting firms that do not score high enough on certain CSR dimensions) and positive screening (where investors actively seek firms with better CSP and invest more in them).
In the case of negative screening, excluding firms from the investment horizon could hurt diversification options. Hong and Kacperczyk (2009) research so called “sin-stocks”, being public companies in the alcohol-, tobacco-, and gaming industries. They find that these firms have 18 per cent lower institutional ownership ratios than other comparable firms, while at the same time outperforming the market in returns. This means, that negative screening does not only lessen diversification options, but also that the excluded firms represent perform significantly better than their counterparts (Hong and Kacqerczyk, 2009), meaning that negative screening could represent an indirect cost to institutional owners in the form of lower financial returns. Positive screening is more widely used, but comes with similar indirect costs to negative screening to the extent that it limits diversification options (Dyck et al., 2015). If positive screening is more widely adopted, it may also lead to overpricing in assets that are considered to be good social performers, which leads to indirect costs in the form of financial underperformance (Dyck et al., 2015).
Alternatively, institutional investors could opt out of limiting their diversification options while still incorporating CSR considerations into their investment portfolios by using active engagement (McCahery et al., 2016). This can be done in several different ways, including selling shares, holding critical speeches at annual meetings, meeting with the executive board, and submitting proposals at annual meetings (McCahery et al., 2016). Actively engaging with investees, however, comes with considerable direct costs, both for the investor and for the investee (Del Guercio & Tran, 2012; McCahery et al., 2016). Whether the emerging trade-off and its interaction with the aforementioned practical barriers results in active ownership among institutional investors has been subject of a debate in the empirical literature for some time.
2.3 Empirical debate
The empirical literature can broadly be separated into two camps. One camp finds that institutional ownership is a neutral or negative factor with respect to corporate social performance. For instance, Barnea and Rubin (2013) use a database of 3000 firms from the Russell indices that have been classified as either being socially responsible or socially irresponsible based on social performance ratings retrieved mainly from the KLD database. They find that the concentration of institutional ownership has no statistically significant effect on the social responsibility of the firm, thus making institutional ownership a neutral factor according to them. Likewise, Dam and Scholtens (2012) use cross-sectional data on the social performance of European firms from the Ethical-Investment
Research Service (EIRS) database for a measure of corporate social performance, and data on ownership percentage from the Amadeus database. They find that institutional investors have no effect on the level of social responsibility, and consequently social performance. David et al. (2007) take a somewhat different approach, and look at the connection between shareholder activism and social performance. They used data from the IRRC, KLD and Compustat databases on a sample of firms for which proposals were filed over a period of seven years, and found that the level of activism was negatively associated with consecutive corporate social performance. These studies give important insight into the ambiguity that surrounds the effect of institutional ownership (and consequently the necessity of further research), and the necessity of finding supporting evidence for the proposed channel through which institutional owners could influence social performance: activism.
In contrast, the other camp finds that institutional ownership is positively associated with corporate social performance. In an early study on the effect of ownership type on social performance, Johnson and Greening (1999) use data from the KLD database in a structural equation model approach, and find that long-term institutional owners have a positive influence on the social performance of firms. More recently, Neubaum and Zahra (2006) used handpicked data on the Fortune 500 population, as well as data from the KLD database for two different time periods (1993 – 1995 and 1998 – 2000). They find that in both time periods, long investment horizons and level of activism show a strong statistically significant relationship with social performance. Here, long-term investors are mainly pension funds and other long-term block holders. In a currently unpublished working paper, Dyck et al. (2015) find similar results. The authors use data on firm’s social responsibility practices from the Thomson Reuters ASSET4 ESG database, and institutional ownership data from the Factset Ownership (LionShares) database. They find that firms are increasingly committed to their environmental and social responsibilities, and that institutional ownership seems to be strongly associated with this. They even find some evidence to suggest that there is a causal link going from IO to CSP by looking at the BP Deepwater Horizon oil spill, by using a difference-in-differences estimation to show that the relationship between IO and a firm’s environmental commitments strengthened after the event. There is some evidence that this positive relationship between IO and CSP may also hold for firms outside of the U.S. and Western Europe, since Oh et al. (2011) find similar results for Korean firms. On this side of the debate there is evidence surrounding shareholder activism, too. Del Guercio and Tran (2012) used data from the IRRC database on CSR-related shareholder proposals over a period of nineteen years, and found that both instances and success rates of these proposals increased over time. These studies, which point towards both a positive association between IO and CSP and the proposed activism-channel, are important for my research in that they give support for the possibility of a causal link that exists between IO and CSP through shareholder activism.
What all these studies have in common, however, is that they only provide evidence on association, not causation between institutional ownership and corporate social performance. There are two main problems in determining causality. First of all, institutional ownership has been on the rise over the last 30 years (Aghion et al., 2013). Thus, in a time series analysis, a positive association between institutional ownership and corporate social performance might simply be a spurious one, as a result of the simultaneous rise in IO and increased CSR in firms. Second of all, there is the problem of simultaneous causality. Given the current literature, it is very likely that institutional ownership is (also) determined by the corporate social performance of the firm, so that institutional investors invest more in firms with good CSP in order to make their investments more socially responsible (Sievänen et al., 2013), which can cause biased and inconsistent results when doing a cross-sectional analysis. Given the aforementioned ambiguity in the empirical results of the current literature and the possible problems surrounding the methodology of the mentioned papers, there is cause for a new way of approaching the link between IO and CSP. Therefore, this thesis will add to the current literature by using a regression discontinuity approach as outlined in section 3, in order to identify a possible causal link going from institutional ownership to corporate social performance.
2.4 Hypothesis development
The comprehensive literature review by Carroll and Shabana (2010) provides a strong ‘business case’ for firms, and consequently institutions, to pursue better corporate social performance. Importantly, there also appears to be no statistically significant difference in financial performance between CSR-conscious funds and other institutional investors (Schröder, 2007). Instead, there is evidence to suggest that CSR-committed institutions perform better than their peers (Schröder, 2007). Together with empirical evidence that supports a link between institutional ownership and corporate social performance from Neubaum and Zahra (2006) and Johnson and Greening (1999), I suspect that institutional owners actively encourage CSR measures. This first hypothesis is called the “responsible investment hypothesis”:
H1: More institutional ownership in a firm causes better corporate social performance.
The second objective of this paper is to see, once the nature of the causal relationship has been determined, whether institutional owners as a group tend to focus on maximizing strengths or minimizing concerns when they engage with firms on social performance issues. This is interesting, since minimizing concerns might be compliance-driven and simply a way to mitigate the effects of potential law suits or societal outrage, whereas increasing a firm’s strengths might signal a true commitment towards increasing the social impacts companies can have. Due to the primary financial goal of most institutions, I suspect that institutional owners as a group tend to focus on the cheaper, short-term gains in social performance. This hypothesis is the “window dressing hypothesis”:
H2: Institutional owners have a positive effect on social concerns, but are neutral with respect to social strengths.
My third objective is to find out whether or not institutional owners actually use active engagement to influence the social performance of the firms they invest in. This will answer whether institutional owners use active engagement instead (or in addition to) social screening. It also tests whether activism is necessary to get positive performance outcomes. This is important, since passive screening could cause better social performance through a backdoor. In the case where the (financial) incentives for firms and/or managers from more institutional investment through better social performance are large enough (and known to firms and/or managers), the threat of exit/promise of entry from institutional investors (that use passive social screens) might already be enough to cause better social performance in firms without the active engagement from institutions. Based on evidence from the survey done by McCahery et al. (2016), however, I suspect that the trade-off is decided in favour of active engagement. This leads to the third hypothesis, called the “activism hypothesis”:
H3: Institutional owners use activism to influence corporate social performance in firms. The next section will outline how these hypotheses will be tested.
3 Methodology and data
This section describes the two methodologies that I use to test the hypotheses as well as the construction and summary statistics of the used dataset.
3.1 Methodology
This paper uses two different methodologies: a regression discontinuity approach, and an instrumental variable regression approach. Both of these approaches use the (re)constitutions of the Russell-indices as an exogenous determinant of institutional ownership.
3.1.1 Russell-indices4
The Russell-indices are a product of FTSE Russell, a global indexing firm that provides services to most of the top financial institutions around the world (FTSE, 2016). The Russell U.S. Equity indices capture 99% of the total, and 100% of the investible market for U.S. equity. The biggest advantage for institutional investors to invest in the Russell-indices is that Russell’s methodology is transparent, whereas most other indices use so called ‘black box approaches’ to index building (Lu, 2013).
The basis for Russell’s methodology are the annual reconstitutions. Only U.S. common stocks listed on the major exchanges in the U.S. are eligible for inclusion. Stocks 1 – 1000 are included in the Russell 1000 index, and stocks 1001 – 2000 are included in the Russell 2000 index. From 2007
4 All information related to the Russell U.S. Equity Indices is based on the FTSE Russell Construction and
onwards, however, Russell has used a banding policy to lessen index turnover, meaning that firms close to the 1000 cut-off with a market capitalization very close to their market capitalization in the previous year remain in the index of the previous year, regardless of their actual position based on market capitalization alone (Chang et al., 2014). The last trading day in May is the annual ranking day, meaning that all (U.S. equity) securities are ranked according to their respective market capitalizations on that day. After this ranking process, the index reconstitution then occurs on the last Friday of June5. This yearly reconstitution schedule is supplemented by quarterly initial public
offerings (IPOs) additions where applicable. However, these quarterly IPOs are only additions to the existing indices, and therefore do not amount to reconstitutions. Therefore, only the observations directly following the reconstitutions cause shocks to the amount of institutional ownership, and it is only these observations that can be used to study the effect institutional ownership has on factors exogenous to the Russell reconstitutions (Chang et al., 2014).
3.1.2 Regression discontinuity approach
The regression discontinuity approach used in this paper is akin to the approach used by Lu (2013) in his unpublished working paper. He uses the differences in institutional ownership between the bottom firms in the Russell1000 and top firms in the Russell2000 indices to determine the effect of IO on bank loan pricing. I will use the ownership differential similarly, with the idea that a firm that is number 1000 in the Russell1000 index is almost indistinguishably similar to number 10016. The
Russell indices are value-weighted, however, so that the weight of firms just above the cut-off in the Russell1000 index is around ten times smaller than similar firms just below the cut-off in the Russell2000 index (Chang et al., 2014). Which firms are listed at the bottom of the Russell1000 index or the top of the Russell2000 index is exogenously determined by outside factors that influence size, as described in section 3.1.1, which makes this cut-off usable in a regression discontinuity design. Using this discontinuity design, I can determine the causal effect of an increase in IO on the CSP of that firm. This exogenous determination of the IO, which will increase with a de-listing from the Russell1000 to a listing on the Russell2000 (Chang et al., 2014), will be used to test my hypotheses by estimating the following (reduced form) model:
𝑌𝑖𝑡 = 𝛽1𝑅2000𝑖𝑡+ 𝛽2𝑅𝑎𝑛𝑘𝑖𝑡+ 𝛽3(𝑅2000𝑖𝑡∗ 𝑅𝑎𝑛𝑘𝑖𝑡) + 𝑓𝑖𝑥𝑒𝑑 𝑒𝑓𝑓𝑒𝑐𝑡𝑠 + 𝜀𝑖𝑡
Where:
Y is the dependent variable, which changes depending on the hypothesis being tested: CSP is measured by the scores in the KLD database,
5 Sometimes, the last Friday of June is determined as being “too proximal to exchange closures”, and will then occur
on the preceding Friday to ensure enough market liquidity.
6 As previously said, in practice the cut-off between the Russell1000 and the Russell2000 is not precisely at the 1000th
firm. However, for illustrative purposes it is convenient to refer to these firms as being precisely number 1000 and number 1001.
IO is measured by the total institutional ownership percentage listed in the Thomson Reuters Institutional (13F) Holdings database,
Activism is measured by the number of proposals backed by an institutional owner over a one-year period as listed in the ISS (RiskMetrics) database7,
R2000 is a dummy variable equal to 1 if the firm is in the Russell2000 index and 0 otherwise, Rank is the value of the distance to the cut-off (being the last firm in the Russell1000), and R2000*Rank is an interaction term between these two variables.
The regression also includes time, industry and/or firm fixed effects, and consequently excludes a constant, to avoid perfect multicollinearity. With firm fixed effects included, any variation in the dependent variable is purely the result of within-firm variation when firms change indices, which makes the inclusion of firm fixed effects a very tight specification (especially when relatively few firms change indices over time). Therefore, the in- or exclusion of any fixed effects differs over regressions to allow for variation in specification tightness. I will first test the strength of the IO-change at the cut-off, and thus the validity of R2000 as a forcing variable, by using IO as a dependent variable. I will then test the first hypothesis by using a measure of CSP as the dependent variable. The second hypothesis will be tested in a similar way, but I will do two separate regressions for strength- and concern-scores. To test the third hypothesis, I will use Activism (over year t) as a dependent variable.
3.1.3 Instrumental variable approach
There are three potential shortcomings of a regression discontinuity approach as described in section 3.1.2. The first is the trade-off between statistical power and effect strength: a broader bandwidth results in a larger sample, and therefore more statistical power; whereas within a narrower bandwidth the discontinuity of the effect is more pronounced. Closely related to this potential problem is the second problem: the exclusion of firm jumps. If a firm jumps from a position above the chosen bandwidth in the Russell 1000 to a position below the chosen bandwidth in the Russell 2000, this firm will not be included in the sample8. If enough firms make such a jump, the total effect of institutional
ownership on corporate social performance might not be captured. The third potential shortcoming is the persistence of firms within the bandwidth on one side of the cut-off. In this case, there is no real shock to institutional ownership due to the Russell reconstitutions. Due to this shortcoming, the initial effect of gaining a larger institutional ownership base is indistinguishable from the long term effect of having more institutional holdings. It might well be, however, that institutional owners bring newly held firms up to a certain standard and become neutral with regards to corporate social performance
7 The database includes both resolution type and sponsor type, thus making the aggregation of these proposals
relatively easy.
8 For example, consider a firm that is the 850th largest firm in year t, but is the 1130th largest firm in year t+1. This
firm will not be included in a discontinuity regression with a bandwidth of less than 130 on each side of the cut-off, whereas the jump in institutional ownership and corporate social performance might be very pronounced.
after the initial push. By using an instrumental variable regression that only looks at the effect of a shock in institutional ownership on corporate social performance for firms that are delisted from the Russell 1000 and make the jump to the Russell 2000, the problems with the regression discontinuity approach are mitigated.
It is not to say, however, that the instrumental variable approach does not have problems of its own. Most importantly, the sample of firms that make a jump from the bottom of the Russell 1000 to the top of the Russell 2000 over the period 2004 – 2013 is very small: only three firms make a jump within the bandwidth of the restricted sample9. Taking Lu’s (2013) broadest suggested bandwidth on
the Russell 1000 side of the cut-off (200 firms) only brings this total up to 25, which makes finding significant results unlikely if the causal effect of institutional ownership on corporate social performance is small. A second problem is the unobserved reason for larger jumps. A firm that jumps from the 800th place in the Russell 1000 to the 80th place in the Russell 2000 (the largest possible
jump within the restricted sample) might have become significantly smaller in terms of market capitalization due to a large special dividend pay-out for example; but the firm might also be in considerable financial distress. Especially the latter could provide a ‘back door’ for the instrument to influence the dependent variable directly, even though the chance of a significant number of jumps being the result of severe financial distress in the population of the largest U.S. public companies is very small (especially since the reason for making a jump is severe underperformance relative to peers, thus disregarding any systemic risk such as the 2008 – 2009 financial crisis). For these reasons, the two approaches should be viewed as complementary, making sure that together they paint a comprehensive picture of the causal effect of institutional ownership on corporate social performance. The instrumental variable regression has the following first stage:
𝐼𝑂̂𝑖𝑡 = 𝜋0+ 𝜋1𝐷𝑒𝑙𝑖𝑠𝑡𝑒𝑑𝑖𝑡
And second stage:
𝐶𝑆𝑃𝑖,𝑡+1 = 𝛾0+ 𝛾1𝐼𝑂̂𝑖𝑡+ 𝜆𝑡+ 𝜀𝑖𝑡
Where:
IO is measured by the total institutional ownership percentage listed in the Thomson Reuters Institutional (13F) Holdings database,
Delisted is a dummy variable equal to 1 if the firm was delisted from the Russell 1000 and included in the Russell 2000, and 0 otherwise,
CSP is measured by the scores in the KLD database. The regression also includes time fixed effects.
9 In other words, only three firms make the jump from inclusion in the bottom 80 firms in the Russell 1000 to the top
3.2 Data
In order to be able to perform these tests, data is needed on four main variables: CSP, composition of the Russell1000 and Russell2000 indices, a measure of institutional ownership, and a measure of activism. The first (CSP) is retrieved from the MSCI ESG KLD database via WRDS, which contains data on the CSP ratings of US listed firms (measured as a net value of strengths minus weaknesses). The second (compositions) is gained through the proprietary compositions of the Russell1000 and Russell2000 indices for the period 2004 – 201310. The third (IO) is retrieved from the Thomson One
Institutional (13F) Holdings database, which contains the 13f holdings of institutional investors per quarter. The measure for IO is the total percentage of shares outstanding owned by institutional investors, and the classifications of owners is gained from Brian Bushee’s personal website11 (the
classifications are made specifically to be merged with the Thomson database). The fourth and last variable (Activism), is retrieved from the ISS (RiskMetrics) database, which contains shareholder proposal data for all US listed firms, and is measured by the number of proposals over the period t to t+1. All data is yearly (as the Russell-indices are reconstituted yearly). The main restriction when time period is concerned is the Russell database, which only contains June observation data from 2004 – 2013. Therefore, I will examine the full Russell dataset period of 2004 – 2013 to create the largest sample possible, which contains 1600 observations when using a bandwidth of 80 firms on both sides of the cut-off point12 (This bandwidth is the lowest optimal bandwidth suggested by Lu (2013)).
3.2.1 Dataset construction A. Russell Indices
The Russell indices were gained in the form of an Excel-file, which is imported into Stata. The index constitutions are given per quarter, but the Russell indices are only fully reconstituted every year at the end of the second quarter (in the other quarters, the index is only updated to account for new IPOs). Thus, only the records for June of every year are kept. The firms in the indices are assigned an index-rank based on their respective weights in the index. This rank counts from the Russell1000 – Russell2000 cut-off so that the 1000th firm and the 1001st firm are both assigned number one, and then count upwards as the firms are located further from the cut-off between the indices.
B. Institutional Holdings
The Thomson One 13f institutional holdings database contains information on, among others, ticker symbols of the firm invested in, manager numbers of the institutional owners, filing dates, reporting dates, outstanding shares of the firm invested in (in millions and in thousands), and shares held at the
10 Gained through dr. Jochem of the Amsterdam Business School, University of Amsterdam.
11 http://acct.wharton.upenn.edu/faculty/bushee/IIclass.html
12 Lu (2013) uses a bandwidth of 100 firms on either side of the cut-off. However, the graphical results of the
discontinuity show that the effect of the index weighting is most pronounced for a bandwidth of 40-50 firms, which is why I have taken the low estimate of the amount of observations.
end of the reporting quarter by the institutional owner. The total dataset roughly consists of between 2 million and 3.5 million records per year.
Due to the way the Russell indices are reconstituted (reconstituted on a yearly basis in May, trading starts on the last day of June, also see section 4.2A), only the holdings at the end of the second quarter are needed (holdings as of June 30th in any given year). To ensure consistency, the reporting
date of the 13f holding is taken as leading, since these reporting dates occur at the end of each quarter. Filing dates and reporting dates do not always align in the database, causing duplicates to occur. To prevent double-counting, observations are only kept if the month and the year of filing align with the month and the year of reporting. To construct the amount of shares held by investor type, the Bushee classifications13 are merged into the institutional holdings dataset on manager number and year. The
total institutional holding percentages are then constructed by first dividing the number of shares held by the institution at the end of the quarter by the total number of shares outstanding of that company at the end of the quarter. The same is done by investor type. Then, the individual holdings are summed by company (ticker symbol) to arrive at the total institutional holdings by firm for a given quarter (defined as number of shares held by institutional investors normalized by the total number of shares outstanding). The holdings with missing ticker symbols are then dropped, since they cannot be attributed to any specific company.
Due to anomalies in the database, the total institutional holdings sometimes exceed 100 per cent. Here, the assumption is made that 100 per cent of the shares of that firm are owned by institutional investors for two reasons: there is no way to determine the actual holdings, and the institutional holdings in that firm will be substantial (close to 100 per cent). The holdings per investor type are then scaled by dividing the holdings of that investor type over the pre-adjusted total holdings, and then multiplying this scaling factor by one (so that all the holdings sum up to 100 per cent)14.
These procedures are carried out for every year, and the datasets are then appended to arrive at the institutional holdings for every firm in the Thomson One 13f holdings database at the end of the second quarter for the years 2004 – 2014. The institutional holdings are then merged into the Russell index database.
C. Corporate Social Performance
The measure for corporate social performance is gained from the MSCI ESG KLD database, and is built up of the number of strengths and the number of concerns in seven dimensions: the environment, community, human rights, employee relations, product, diversity, and corporate governance. The performance in each of these areas is defined as the number of strengths net of the number of concerns in each area. The measure of corporate social performance for each firm is then defined as
13 See footnote 11.
14 For example, say that firm A is reported to have 200% total institutional holdings, 70% of which are bank trust
the sum of the performance in each of the seven dimensions. To allow for some interaction time between investors and investees, the consecutive year’s corporate social performance measures are taken as a measure of performance (performance at the end of period t+1)15. The first differences in
corporate social performance over the period t to t+1 are also taken as a measure of change. This manner of measuring CSP makes the number of firm-year observations somewhat smaller than for institutional holdings, since the CSP scores used in this study are available until the year 2013, and can therefore only be matched to firm-year observations from 2012 or earlier.
D. Activism
Activism is retrieved from the ISS (RiskMetrics) database for the years 2003 – 2014, and contains 10,589 observed resolutions. Resolutions are split between SRI and non-SRI resolutions. All resolutions are given a value of 1, and collapsed by ticker and year to gain both the number of total resolutions per company per year and of the number of SRI-related resolutions per company per year. This will function as the measure of activism, where more resolutions means a larger degree of activism.
E. Construction of the master dataset
To create the master dataset, the Russell indices are taken as the master file. Then, the institutional holdings, CSP scores, and Activism measures are merged in, creating a master dataset consisting of 29,818 firm-year observations for the month June for the period 2004-2013 (the year 2014 Russell Index constitutions only runs until the first quarter, rendering it unusable). Industry designations, as needed for the industry fixed effects, were taken from the CRSP/Compustat Merged database link table and merged into the final dataset by year and company. Additional data needed for the balance table to test the statistical difference between both sides of the cut-off was gained from the Compustat and CRSP databases16. The resulting total sample was then restricted to the bandwidth of 80 firms
around the cut-off between the Russell 1000 and Russell 2000 indices. The restricted sample consists of 1600 firm-years. This is the final sample used in the regression analyses.
3.2.2 Summary statistics
Before discussing the summary statistics of the variables that are relevant to my three hypotheses, it is important to see whether the sample is statistically similar on either side of the cut-off in important aspects apart from the proposed channel (institutional ownership). If this is not the case, any variation in corporate social performance and/or activism could instead be the result of fundamental firm-
15 Meaning that if firm A has a certain score for 2013 this score is matched with year 2012, so that when the database
is merged with the Russell indexes and institutional holdings, the corporate social performance for 2013 is matched with the holdings for 2012, thus constituting a t+1 observation.
16 From Compustat I retrieved end of the year total assets (at); debt (dlc, dltt); equity (teq, seq); property, plants and
equipment (ppent); earnings (ebitda); and capital expenditures (capx). From CRSP I retrieved end of the second quarter shares outstanding (shrout) and share price (prc).
Table 1a – Summary statistics of responsible investment and activism hypothesis variables Panel 1 - Full Sample
Variables Mean Median SD P25 P75 N
IO 0.682 0.731 0.238 0.528 0.868 28356 Delisted 0.001 0.000 0.029 0.000 0.000 29818 Resolutions 2.238 1.000 2.363 1.000 3.000 3773 SRI Resolutions 0.763 0.000 1.154 0.000 1.000 3773 CSP -0.046 0.000 2.557 -2.000 1.000 22407 ∆CSP 0.111 0.000 1.687 -1.000 1.000 21960 Environment 0.071 0.000 0.778 0.000 0.000 22407 ∆Environment 0.046 0.000 0.549 0.000 0.000 21960 Community 0.071 0.000 0.495 0.000 0.000 22407 ∆Community 0.009 0.000 0.324 0.000 0.000 21960 Human Rights -0.020 0.000 0.232 0.000 0.000 22407 ∆Human Rights 0.014 0.000 0.188 0.000 0.000 21960 Employee Relations 0.321 0.000 0.860 0.000 0.000 22407 ∆Employee Relations 0.078 0.000 0.666 0.000 0.000 21960 Diversity -0.088 0.000 1.367 -1.000 1.000 22407 ∆Diversity -0.048 0.000 0.800 0.000 0.000 21960 Product -0.130 0.000 0.580 0.000 0.000 22407 ∆Product 0.014 0.000 0.408 0.000 0.000 21960 Governance -0.271 0.000 0.727 -1.000 0.000 22407 ∆Governance -0.002 0.000 0.707 0.000 0.000 21960
Panel 2 - Restricted Sample
Variables Mean Median SD P25 P75 N
IO 0.730 0.809 0.254 0.579 0.938 1511 Delisted 0.016 0.000 0.124 0.000 0.000 1600 Resolutions 1.390 1.000 1.170 1.000 1.000 118 SRI Resolutions 0.347 0.000 0.478 0.000 1.000 118 CSP -0.433 -1.000 1.986 -2.000 0.000 1243 ∆CSP 0.024 0.000 1.446 -1.000 0.000 1218 Environment -0.042 0.000 0.647 0.000 0.000 1243 ∆Environment 0.021 0.000 0.446 0.000 0.000 1218 Community 0.031 0.000 0.431 0.000 0.000 1243 ∆Community 0.001 0.000 0.323 0.000 0.000 1218 Human Rights -0.013 0.000 0.183 0.000 0.000 1243 ∆Human Rights 0.011 0.000 0.164 0.000 0.000 1218 Employee Relations 0.219 0.000 0.636 0.000 0.000 1243 ∆Employee Relations 0.037 0.000 0.522 0.000 0.000 1218 Diversity -0.131 0.000 1.163 -1.000 0.000 1243 ∆Diversity -0.085 0.000 0.749 0.000 0.000 1218 Product -0.114 0.000 0.514 0.000 0.000 1243 ∆Product 0.016 0.000 0.378 0.000 0.000 1218 Governance -0.382 0.000 0.685 -1.000 0.000 1243 ∆Governance 0.024 0.000 0.713 0.000 0.000 1218 v
This table shows the mean, median, standard deviation, and first and third quartiles of all variables relevant for testing the first and third hypotheses for both the full and restricted samples. The restricted sample is constructed by taking a bandwidth of 80 firms around the cut-off between the Russell 1000 and Russell 2000 indices. IO is the number of shares held by institutions normalized by total outstanding shares. Delisted is a dummy equal to 1 if the firm was delisted from the Russell 1000. Resolutions and SRI Resolutions are the total and SRI-related resolutions per firm per year. CSP and its seven components (Environment, Community, Human Rights, Employee Relations, Diversity, Product, Governance) are the strengths net of weaknesses in each category, where the CSP is the sum of all net scores per firm per year. ∆CSP and its components (∆Environment, ∆Community, ∆Human Rights, ∆Employee Relations, ∆Diversity, ∆Product, ∆Governance) are the changes in scores from period t to period t+1.
differences on either side of the cut-off. Table 8 in the appendix shows the balance table for the restricted sample split between the treatment group (Russell 2000 firms) and the control group (Russell 1000 firms). As expected, the two groups are not statistically significantly different when it comes to the market-to-book17, leverage, return on equity, tangibility, and relative capital expenditure
dimensions. Also as expected, the samples are statistically significantly different with respect to institutional ownership and, to a lesser extent, market capitalization (which is a logical result of Russell’s ranking procedure). Thus, the balance table provides statistical evidence to suggest that using a regression discontinuity design on the restricted sample is indeed warranted.
The summary statistics of variables relevant for testing hypothesis 1 and 3 for both samples are shown in table 1a, the relevant restricted sample statistics for testing hypothesis 2 are shown in table 1b18. The full sample mean institutional ownership, reported as IO in Table 1a, is 68%, with a
somewhat higher median of 73%. Both of these figures show the same pattern as in Lu’s (2013) sample, with mine being somewhat higher. These higher institutional holdings are relatively unsurprising since my sample is over a more recent period, and institutional holdings have been
17 I feel obliged to mention that the market-to-book ratio is very nearly statistically significant at the 5% level. This
can be explained by the buying pressure that firms in the Russell 2000 index experience from institutional investors as a result of being included in the index, making the market-to-book ratio partly endogenous with respect to the “treatment” (inclusion in the Russell 2000 index). However, since the t-statistic is still strictly lower than its critical value, there is no statistical reason to control for the market-to-book ratio in the discontinuity regressions.
18 Table 9 in the appendix shows the relevant full sample statistics for testing hypothesis 2.
Table 1b – Restricted sample summary statistics of window dressing hypothesis variables Restricted Sample
Variables Mean Median SD P25 P75 N
CSP Strengths 1.059 1.000 1.571 0.000 1.000 1243 CSP Concerns 1.809 2.000 1.482 1.000 2.000 1243 Environment Strengths 0.134 0.000 0.470 0.000 0.000 1243 Environment Concerns 0.176 0.000 0.489 0.000 0.000 1243 Community Strengths 0.091 0.000 0.379 0.000 0.000 1243 Community Concerns 0.060 0.000 0.248 0.000 0.000 1243
Human Rights Strengths 0.010 0.000 0.117 0.000 0.000 1243
Human Rights Concerns 0.023 0.000 0.151 0.000 0.000 1243
Employee Relations Strengths 0.242 0.000 0.619 0.000 0.000 1243
Employee Relations Concerns 0.341 0.000 0.591 0.000 1.000 1243
Diversity Strengths 0.421 0.000 0.808 0.000 1.000 1243 Diversity Concerns 0.552 0.000 0.663 0.000 1.000 1243 Product Strengths 0.066 0.000 0.248 0.000 0.000 1243 Product Concerns 0.180 0.000 0.466 0.000 0.000 1243 Governance Strengths 0.094 0.000 0.298 0.000 0.000 1243 Governance Concerns 0.476 0.000 0.590 0.000 1.000 1243 v
This table shows the restricted sample mean, median, standard deviation, and first and third quartiles of all variables relevant for testing the second hypotheses, being the strengths and concerns of CSP and its seven components (Environment, Community, Human Rights, Employee Relations, Diversity, Product, Governance) . The restricted sample is constructed by taking a bandwidth of 80 firms around the cut-off between the Russell 1000 and Russell 2000 indices.
consistently rising over the past decades (Aghion et al., 2013). The restricted sample means and medians are somewhat higher than for the full sample. This is most likely the result of the high percentages of institutional holdings amongst the top firms in the Russell 2000 index.
The average number of delistings in the full sample is very low: 0.1% of total observations; this jumps to 1.6% of observations in the full sample, resulting in 25 delistings over the full ten-year period. The mean number of resolutions shows a different pattern: here the number of resolutions is almost two times higher in the full sample than in the restricted sample, although the median number of resolutions stays the same at about one per firm per year. This is similar for the SRI resolutions. Both variables also have a lower standard deviation in the restricted sample than in the full sample. The other variables, CSP and its seven components, show low means and medians, especially when compared to the theoretical maximum range of -30 to 37 points when comparing all concerns and all strengths19. However, these low average scores are similar in terms of size to other papers,
such as David et al. (2007). It is important to note that given these low scores, any effect of institutional ownership will most likely be relatively small as well. As for the change variables (∆CSP and its components), it is clear that most change is positive, meaning that over time scores tend to improve both in the full sample and in the restricted sample, the only exception being Diversity. Table 1b shows the restricted sample summary statistics (and table 9 in the appendix the full sample statistics) for all strengths and concerns scores for both total CSP and its components, as well as the yearly changes in each. Here, we see an overall comparable picture to the CSP scores in Table 1a, with concerns showing an overall downward trend while changes in strengths tend to be positive (the exceptions being Human Rights and Diversity). The individual strengths and concerns scores are higher than the total CSP scores and its components, and also show a lower variability. This is unsurprising since total scores are strengths net of concerns, thus making the deviation of the net scores dependent on both the variance in strengths as well as the variance in concerns.
4 Empirical results
This section contains the results of the regressions per hypothesis, as well as a number of robustness checks. For every hypothesis, the results of the regressions discontinuity approach are discussed first, followed by the results of the instrumental variable approach.
4.1 Results
The results of the two approaches are presented per hypothesis, with the regression discontinuity approach being presented first, followed by the instrumental variable approach. The first hypothesis being tested is the responsible investment hypothesis.
19 For a full overview of the scoring methodology see the ESG ratings definitions, available at http://cdnete.lib.ncku.
4.1.1 Responsible investment A. Regression discontinuity
The results of the discontinuity approach used in this context are only meaningful if institutional ownership does indeed show the expected discontinuity. Therefore, the first dependent variable in this approach is institutional ownership. Table 2 shows the results of this first regression, with different combinations of fixed effects per regression. The coefficient of the R2000 variable shows that the discontinuity in institutional ownership between the Russell 1000 index and the Russell 2000 index is both statistically and economically significant, with regressions 1 to 4 showing a jump of between 34.9% and 37.1% in institutional ownership20. This means that when a firm is excluded from the
bottom of the Russell 1000 index and included in the top of the Russell 2000 index, on average institutional investors will buy an additional stake of between 34.9% and 37.1% in that firm21.
Column 5 shows a smaller jump of 13.3%. Due to the inclusion of firm fixed effects, this tight specification most closely measures the within-firm variation when firms jump from the Russell 1000 to the Russell 2000 index. The relatively large difference between the coefficients on the R2000 variable between specifications 1 to 4 on the one hand and specification 5 on the other suggests that there might be some other, unobserved factors driving institutional ownership. For instance, it could
20 This result is similar in size to Lu’s (2013).
21 For instance, suppose that a firm is number 995 in the Russell 1000 and has 20% institutional ownership. If the
following year it then becomes number 1005 (thus 5 in the Russell 2000) it will on average have between 54.9% and 57.1% institutional ownership in that year.
Table 2 – Institutional ownership regressions
(1) (2) (3) (4) (5) Variables IO IO IO IO IO R2000 0.369*** 0.371*** 0.349*** 0.352*** 0.133*** (0.047) (0.046) (0.048) (0.047) (0.050) Rank 0.004*** 0.004*** 0.003*** 0.003*** 0.001** (0.001) (0.001) (0.001) (0.001) (0.001) R2000*Rank -0.004*** -0.004*** -0.003*** -0.003*** -0.001* (0.001) (0.001) (0.001) (0.001) (0.001) Observations 1,511 1,511 1,311 1,311 1,511 R-squared 0.235 0.266 0.549 0.579 0.913
Year fixed effects NO YES NO YES NO
Industry fixed effects NO NO YES YES NO
Firm fixed effects NO NO NO NO YES
Robust standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1
This table shows the regression results of the regression discontinuity approach with institutional ownership as the dependent variable. Different columns have different fixed effects. IO is the firm-level fraction of total shares outstanding owned by institutional investors as declared in the 13f holdings at the end of June. R2000 is a dummy variable equal to 1 if the firm is listed in the Russell 2000 index and a 0 if it is listed in the Russell 1000 index. Rank designates the distance of a firm from the cut-off, starting at a value of 1 at both sides of the cut-off and counting outwards, where the cut-off is between the last firm in the Russell 1000 index and the first firm in the Russell 2000 index. Standard errors are clustered at the firm level.
be that not only the initial jump causes an increase in institutional ownership, but that staying in the top of the Russell 2000 index also increases institutional ownership over time. However, adding firm fixed effects to the specification in this case may also result in some statistical difficulties. The dataset contains 1600 firm-year observations and 800 unique firms, which means that on average each firm has only 2 observations. Due to these low numbers of observations per firm, adding firm fixed effects leaves far fewer degrees of freedom to find an effect than under the other specifications22.
The coefficients for the Rank and interaction variables are also in line with my expectations. The Rank variable means that firms in the Russell 1000 index get between 0.3% and 0.4% more institutional ownership, measured as the percentage of outstanding shares owned by institutional investors, as they move further away from the cut-off. Firms in the Russell 2000 index experience both the Rank and interaction variable effects, which cancel each other out, meaning that for the first 80 firms in the top of the Russell 2000 index institutional ownership is more or less stable (and higher than for the bottom of the Russell 1000 index). For these variables the estimates of the effect in the regression specification in column 5 is also lower for the same reasons as mentioned earlier. A visual representation of the institutional ownership discontinuity is given in graph 123.
22 In practice, adding firm fixed effects means adding a dummy for each individual firm. In this case this means
adding 800 dummies to the regressions specification, which means fewer degrees of freedom.
23 A breakdown of the discontinuities by institutional investor type is presented in the appendix.
Graph 1 – Discontinuity in institutional onwership
This graph shows the discontinuity of institutional ownership between the Russell 1000 and the Russell 2000 indices. Firms are aggregated into bins of ten firms each. Institutional ownerhip is represented on the vertical axis as a percentage of outstandingshares owned by institutional investors. The horizontal axis shows the rank of the firm-bins in both indices.
Since there is sufficient evidence to suggest that there is indeed a considerable discontinuity in institutional ownership, the question becomes whether this same discontinuity exists for corporate social performance. Table 3 shows the regression results for the corporate social performance regressions. To control for time effects and industry differences, the regressions include time and industry fixed effects24. The first panel shows the absolute social performance scores, and the second
panel shows the yearly change in social performance scores. Because jumps between the Russell 1000 and Russell 2000 indices are rare (only 25 in the ten-year sample within the range around the cut-off), persistence of firms being in one of the two indices is high. Given this persistence, the first panel can
24 Firm fixed effects are omitted for the statistical reasons mentioned in the interpretation of the institutional
ownership regressions.
Table 3 – Corporate social performance RD regressions Panel 1 - Social performance scores
(1) (2) (3) (4) (5) (6) (7) (8)
Variables CSP Environment Human
Rights Community Employee Relations Diver-sity Product Corporate Governance R2000 -0.497 -0.066 -0.020 -0.037 -0.215 0.024 0.108 -0.291*** (0.321) (0.089) (0.022) (0.062) (0.155) (0.175) (0.118) (0.101) Rank -0.005 -0.001 -0.000 0.001 -0.002 0.004 0.001 -0.006*** (0.005) (0.001) (0.000) (0.001) (0.002) (0.003) (0.002) (0.002) R2000*Rank 0.004 0.001 0.000 -0.001 0.001 -0.004 -0.001 0.008*** (0.006) (0.002) (0.000) (0.001) (0.002) (0.003) (0.002) (0.002) Observations 1,108 1,108 1,108 1,108 1,108 1,108 1,108 1,108 R-squared 0.478 0.515 0.460 0.367 0.331 0.497 0.371 0.453
Year fixed effects YES YES YES YES YES YES YES YES
Industry fixed effects YES YES YES YES YES YES YES YES
Panel 2 - Yearly social performance change
(1) (2) (3) (4) (5) (6) (7) (8) Variables ∆CSP ∆Environ-ment ∆Human Rights ∆Community ∆Employee Relations ∆Diver-sity ∆Product ∆Corporate Governance R2000 -0.176 -0.022 -0.028 -0.023 -0.189* 0.122 -0.049 0.012 (0.231) (0.080) (0.027) (0.039) (0.107) (0.104) (0.064) (0.091) Rank -0.004 -0.000 0.000 -0.001* -0.003 0.001 -0.001 -0.000 (0.004) (0.001) (0.001) (0.001) (0.002) (0.002) (0.001) (0.001) R2000*Rank 0.002 -0.000 -0.000 0.001 0.003 -0.003 0.000 0.001 (0.005) (0.001) (0.001) (0.001) (0.002) (0.002) (0.001) (0.002) Observations 1,086 1,086 1,086 1,086 1,086 1,086 1,086 1,086 R-squared 0.280 0.186 0.262 0.152 0.262 0.283 0.191 0.452
Year fixed effects YES YES YES YES YES YES YES YES
Industry fixed effects YES YES YES YES YES YES YES YES
Robust standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1
This table shows the regression results of the regression discontinuity approach with corporate social performance and the yearly change in corporate social performance as the dependent variables. In the first panel, regression 1 shows the results for the regression with total CSP scores as the dependent variable, and column 2 to 8 show the regression results for its component scores. The second panel uses the yearly changes in the scores as dependent variable. CSP is measured by the number of strengths net of the number of concerns in the KLD database. R2000 is a dummy variable equal to 1 if the firm is listed in the Russell 2000 index and a 0 if it is listed in the Russell 1000 index. Rank is the distance of a firm from the cut-off, starting at a value of 1 at both sides of the cut-off and counting outwards, where the cut-off is between the last firm in the Russell 1000 index and the first firm in the Russell 2000 index. All regressions include year and industry fixed effects. Firm fixed effects were omitted for statistical reasons. Standard errors are clustered at the firm level.
