The Role of Proxy Advisory Firms: Evidence from a Regression-Discontinuity Design

Transcription

1 The Role of Proxy Advisory Firms: Evidence from a Regression-Discontinuity Design Nadya Malenko and Yao Shen April 2015 Abstract Proxy advisory rms have become important players in corporate governance. We use exogenous variation in Institutional Shareholder Services (ISS) recommendations generated by a cuto rule in ISS guidelines to answer two questions. First, what e ect do ISS recommendations have on voting outcomes? Second, what are the sources of ISS in uence? Using a regression discontinuity design, we nd that a negative ISS recommendation on a say-on-pay proposal leads to a 25 percentage point reduction in voting support. We next examine potential channels of ISS in uence and do not nd evidence that investors follow ISS recommendations for their information content. Thus, our evidence suggests that the more likely channel of ISS in uence is the certi cation role of ISS recommendations. Keywords: proxy advisors, ISS, proxy voting, shareholder activism, regression discontinuity, say-on-pay, executive compensation JEL Classi cation Numbers: G34, D72, J33 The authors are from the Carroll School of Management at Boston College, Finance Department. We are grateful to Bernard Black, Sudheer Chava, Mariassunta Giannetti, Dirk Jenter, Oguzhan Karakas, Darren Kisgen, Camelia Kuhnen, Andrey Malenko, Jordan Nickerson, Je rey Ponti, Jonathan Reuter, Miriam Schwartz-Ziv, Philip Strahan, Jerome Taillard, and seminar participants at Boston College, McGill University, and University of California, Davis for helpful comments and discussions. We also thank Jun Crystal Yang and Pouyan Foroughi for excellent research assistance. nadya.malenko@bc.edu, yao.shen@bc.edu. 1

2 1 Introduction Shareholder voting plays a key role in corporate governance. Each year, investors cast billions of votes that a ect the structure of boards, mergers and acquisitions, and rms corporate governance policies. Shareholder voting has become increasingly important due to a rise in shareholder activism, the increase in institutional ownership and regulations requiring institutions to disclose their votes, the introduction of mandatory say-on-pay, and the shift to majority voting for director elections. A signi cant development in recent years has been the growing in uence of proxy advisory rms, which counsel institutional investors on how to vote their shares. The largest proxy advisor, Institutional Shareholder Services (ISS), issues recommendations on about 39,000 companies in 115 countries. According to the SEC commissioner Michael Piwowar, proxy advisory rms may exercise outsized in uence on shareholder voting and the Dodd-Frank provisions, such as mandatory say-on-pay votes, make proxy advisory rms potentially even more in uential. 1 Multiple commentators have raised concerns about proxy advisors one-size- ts-all approach to governance matters, lack of transparency, and potential con icts of interests. 2 These concerns have prompted the SEC and regulatory bodies in other countries to take steps aimed at increasing the transparency and accountability of proxy advisors and investment advisors using their services. 3 Given the role of shareholder voting in corporate governance and the ongoing calls for regulation of the proxy advisory industry, it is important to answer two questions. First, do proxy advisors have a strong impact on voting outcomes? And second, if they do, what is the source of their in uence? In particular, do proxy advisors recommendations a ect votes because they convey new information to investors, or because they play a certi cation role, protecting institutions from potential criticism and legal liability for their voting practices? Answering these questions is challenging because of an identi cation problem. While a number of papers have shown a positive correlation between ISS recommendations and 1 See the transcript of the December 5, 2013 SEC roundtable. Say-on-pay is an advisory vote expressing shareholder approval or disapproval of the rm s executive compensation practices. The Dodd-Frank Act required almost all US public rms to provide shareholders with such a vote. 2 E.g., Outsized Power & In uence: The Role of Proxy Advisers by the SEC commissioner Daniel Gallagher (Gallagher, 2014) and the SEC Concept Release on the U.S. Proxy System (July, 2010). 3 The SEC sought public comment on the role of proxy advisors in July 2010, held a public roundtable in December 2013, and released guidance on investment advisors responsibilities in using proxy advisors in June 2014 (SLB No. 20). See also the Final Report and Feedback Statement on the Consultation Regarding the Role of the Proxy Advisory Industry by the European Securities and Markets Authority (ESMA) and Guidance for Proxy Advisory Firms by the Canadian Securities Administrators. 2

3 voting outcomes, 4 this relation is not necessarily causal: the same unobservable rm and management characteristics that lead ISS to give a negative recommendation can also lead shareholders to withdraw their support for the proposal. Accordingly, there is disagreement about whether ISS is truly as in uential as many commentators claim. 5 Even ISS itself, trying to alleviate regulators concerns about its oversized in uence, emphasizes: As many investors can be expected to share a general approach to assessing corporate governance practices, it is perhaps not surprising that correlations can be found between ISS recommendations and shareholder voting outcomes. However, this does not prove causality, nor are those correlations consistent. In our view, it is more logical to interpret broad correlation as indicating that ISS policies, analyses and recommendations are based on principles and approaches which are shared by many investors. (ISS Responses to ESMA Discussion Paper, Jun 21, 2012) In this paper, we address this empirical challenge in a regression discontinuity (RD) setting by using exogenous variation in ISS recommendations due to a cuto rule employed by ISS in its guidelines on say-on-pay proposals. Speci cally, ISS used to conduct an initial screen of companies focusing on their one- and three-year total shareholder returns (TSRs) relative to certain cuto s and only performed a deeper analysis of the company s executive compensation practices for companies below the cuto. For example, the ISS 2012 white paper Evaluating Pay for Performance Alignment notes: In the last few years, the approach has utilized a quantitative methodology to identify underperforming companies i.e., those with both 1- and 3-year total shareholder return (TSR) below the median of peers in their 4-digit Global Industry Classi cation Standard (GICS) group. Underperforming companies then received an in-depth qualitative review, focused primarily on factors such as the year-over-year change in the CEO s total pay, the 5-year trend in CEO pay versus company TSR, and the strength of performance-based pay elements. More speci cally, as we discuss in Section 2, the cuto for a given company is calculated as the median TSR of all rms that are both in the company s four-digit GICS industry 4 They include Alexander et al. (2010), Bethel and Gillan (2002), Cai, Garner, and Walking (2009), Ertimur, Ferri, and Oesch (2013), Iliev and Lowry (2015), Larcker, McCall, and Ormazabal (2014), and Morgan et al. (2011), among others. A more detailed review of the literature is provided below. 5 For example, BlackRock, in its comment to the SEC, states: We believe that the in uence of proxy advisory rms in general, and ISS in particular, have been overstated. ( 10/s pdf). See also Choi, Fisch, and Kahan (2010). 3

4 group and in Russell 3000, where the TSRs are computed on the last day of the calendar quarter closest to the company s scal-year-end. In the rest of the paper, we refer to this rule as the ISS cuto rule. The ISS rule implies that rms below the cuto undergo more scrutiny to get a positive recommendation from ISS than rms above the cuto, and hence the probability of a negative ISS recommendation should increase discontinuously just below the cuto. Indeed, we show that relative to rms just above the cuto, there is a 15% increase (from 10% to 25%) in the probability of a negative recommendation for rms just below the cuto. This jump is large given that the average probability of a negative ISS recommendation in our sample is 12.7%. At the same time, the somewhat arbitrary nature of the ISS cuto suggests that being just above or below the cuto is locally random, i.e., rms just above and just below the cuto are similar across all characteristics except, potentially, the ISS recommendation. We verify this conjecture by showing that rm characteristics are smooth around the cuto, that rms do not manipulate their TSRs to move above the cuto, and by conducting falsi cation tests on samples where the cuto rule is not used. Thus, any discontinuous decrease in voting support below the cuto can be attributed to the causal e ect of ISS recommendations. We therefore implement a fuzzy RD design by focusing on a narrow bandwidth around the cuto and instrumenting a negative ISS recommendation with an indicator variable that equals one for rms below the cuto (Imbens and Lemieux (2008), Roberts and Whited (2012)). Importantly, because the cuto is based on industry medians, the sample of rms in a bandwidth around the cuto is likely to be representative of the sample as a whole. Our results are as follows. First, we nd that ISS recommendations have a strong causal e ect on voting outcomes: a negative ISS recommendation leads to a 25 percentage point decrease in voting support for say-on-pay proposals. This e ect is economically large: dissent above 20% is generally viewed as an indication of substantial dissatisfaction (e.g., Del Guercio, Seery, and Woidtke, 2008; Ferri and Maber, 2013) and leads companies to change their compensation practices (Ferri and Maber, 2013; Ertimur, Ferri, and Oesch, 2013). 6 Our main speci cation is a local linear regression estimated on a 5% bandwidth, and our results are robust to multiple bandwidths, using exible polynomial functions, and controlling for 6 See also At Annual Meetings, 70 Is the New 50, Study Shows, the Wall Street Journal, March According to proxy advisors policies, support of less than 70% (ISS) or 75% (Glass Lewis) warrants an explicit board response, and investor respondents to the ISS policy survey share the same view. 4

5 various rm characteristics. We also nd some evidence that the e ect of ISS is stronger in rms with less concentrated institutional ownership, consistent with the hypothesis that larger shareholders have more incentives to invest in information acquisition and are therefore less likely to rely on proxy advisors recommendations (e.g., Iliev and Lowry, 2015). Given the signi cant e ect of ISS recommendations on voting outcomes, it is important to understand the sources of ISS in uence. There are two channels that can explain why investors follow ISS recommendations. One is the informational channel: investors follow ISS primarily because its recommendations and reports convey new valuable information to them. Indeed, the rationale for the existence of proxy advisors is that, given institutional investors diversi ed holdings, it may be more cost-e ective for them to outsource research to a third party than to do their own research for each proposal of each rm in their portfolios. This is especially so given that most investors believe that proxy statements are too long, obscure, and di cult to understand (see Section 5 for details). The second potential channel of ISS in uence is the certi cation role of its recommendations. Because institutions have a duciary duty to vote in their clients best interests, 7 they may have incentives to follow ISS recommendations regardless of their actual information content as it may protect them from criticism and litigation. For example, the 2003 SEC rule explicitly states that an institution could demonstrate that the vote was not a product of a con ict of interest if it voted client securities in accordance with a pre-determined policy, based upon the recommendations of an independent third party. The certi cation role of ISS is enhanced by its coordination e ect: by following ISS, institutions ensure that they vote similarly to many other shareholders, which may further decrease their chances of being criticized for their voting practices. 8 Distinguishing between the informational and certi cation channels is important to understand the e ect of ISS on voting outcomes, especially in light of the frequent calls to regulate the proxy advisory industry. To disentangle the two channels, we again rely on the ISS cuto rule, which states that rms below the cuto receive an in-depth qualitative review and only get a positive recommendation if the review shows that their compensation practices are appropriate. Consider two rms that both receive a positive ISS recommenda- 7 The 2003 SEC rule Proxy Voting by Investment Advisers (Release No. IA-2106) requires mutual fund managers to vote proxies in their clients best interests, and the Department of Labor s 1994 Interpretive Bulletin imposes a similar requirement on pension funds. 8 Relatedly, Matvos and Ostrovsky (2010) point out that management s ability to punish a given fund for its voting behavior is lower if it has to retaliate against a larger number of funds. This may also motivate funds to follow ISS recommendations even if these recommendations do not have any information content. 5

6 tion, but one falls just below and the other just above the cuto. If the in-depth review by ISS provides new valuable information to investors, then a positive recommendation for the rm below the cuto is a stronger positive signal about the rm s compensation practices than a positive recommendation for the rm above the cuto (which was given without an in-depth review). Hence, if the informational channel is important, voting support should be higher for the rm below the cuto. As we discuss in Section 5, this argument does not require investors to be aware of the ISS cuto rule because the research report distributed by ISS to its clients contains the summary of the analysis underlying each recommendation. Focusing on the subsample of rms with positive ISS recommendations, we show that the average di erence in voting support between rms just above and just below the cuto is about 2% with a standard error of 1.4%, which is statistically and economically insigni cant. For comparison, the standard deviation of voting support in the subsample with positive ISS recommendations is 7%. Moreover, the di erence has the opposite sign from that predicted by the informational channel: voting support above the cuto is higher. This suggests that either the ISS in-depth review does not convey new information to investors or, if it does, this new information does not a ect investors voting decisions. In other words, we do not nd evidence that investors follow ISS say-on-pay recommendations because of their information content. Hence, at least in our sample of say-on-pay proposals, the more likely channel of ISS in uence is the certi cation role of ISS recommendations. The key assumption of our RD design is that whether a rm falls just above or below the cuto is locally random. We perform several tests to verify this assumption and to show that our results are not driven by di erences in rm characteristics around the cuto. First, we redo our analysis on two samples for which the cuto rule does not apply: say-on-pay voting in 2012 (the year in which ISS stopped using this rule) and voting for director elections during our main sample period. In both cases, we do not nd any discontinuity in voting support around the cuto. Second, we show that the distribution of various elements of CEO compensation and other rm characteristics is smooth around the cuto. Third, we alleviate the concern that rms manipulate their TSRs to move above the cuto by performing the McCrary (2008) test and showing that the density of the forcing variable is smooth around the cuto. Such manipulation is indeed unlikely given that the cuto depends on the TSRs of all rms in the industry and the TSRs are determined by stock price movements. Finally, we consider several placebo cuto s and show no evidence of discontinuity around them. 6

7 Our paper contributes to the growing literature on shareholder activism 9 and the role of institutional investors in rms corporate governance. 10 the literature on shareholder voting. In particular, it contributes to Prior research shows that shareholder voting has a signi cant impact on rms policies and value, even when votes are nonbinding. 11 In an RD setting, Cuñat, Gine, and Guadalupe (2012, 2013) nd that relative to proposals that fail by a small margin, proposals that pass by a small margin yield an abnormal return between 1.3% and 2.4%, depending on the proposal type. 12 In the context of say-on-pay voting, Ferri and Maber (2013) show that about 80% of UK rms with substantial voting dissent respond by removing controversial compensation practices, and Ertimur, Ferri, and Oesch (2013) nd similar evidence for US rms. 13 Given the signi cance of shareholder voting, it is important to understand what factors a ect investors voting decisions. Our paper shows that a major determinant of shareholder votes are the recommendations of proxy advisory rms, but that the information content of these recommendations might not be the main reason why investors follow them. There is a growing literature on proxy advisory rms. A number of papers document a signi cant relation between ISS recommendations and shareholder support on various voting issues. They include Alexander et al. (2010), Bethel and Gillan (2002), Cai, Garner, and Walking (2009), Iliev and Lowry (2015), and Morgan et al. (2011), among others. In the context of say-on-pay, Ertimur, Ferri, and Oesch (2013) show that the sensitivity of votes to proxy advisors recommendations varies with the institutional ownership structure and the rationale behind the recommendation, and Larcker, McCall, and Ormazabal (2014) document a negative market reaction to compensation program changes made to avoid a negative recommendation. Our contribution to this literature is twofold. Our paper is the rst to identify the causal e ect of ISS recommendations on voting outcomes. Several prior 9 See Karpo (2001), Gillan and Starks (2007), and Brav, Jiang, and Kim (2009) for reviews. 10 E.g., Hartzell and Starks (2003), Almazan, Hartzell, and Starks (2005), Aghion, Van Reenen, and Zingales (2013), Mullins (2014), Crane, Michenaud, and Weston (2014), Appel, Gormley, and Keim (2015), and Boone and White (2015). 11 See Ferri (2012) for a survey of nonbinding voting and Levit and Malenko (2011) for a theoretical analysis. 12 While these papers also use RD in the context of shareholder voting, the approach in our paper is di erent since we compare rms that fall just below and just above the ISS cuto. Some other nance studies that use the RD design include Bakke and Whited (2012), Beshears (2013), Chava and Roberts (2008), Iliev (2010), Keys et al. (2010), Liberman (2013), Mullins (2014), Rauh (2006), and Reuter and Zitzewitz (2013). 13 Relatedly, Schwartz-Ziv and Wermers (2014) show that low say-on-pay support is followed by a decrease in excess compensation and better selection of peer rms when ownership is relatively concentrated. See also Ertimur, Ferri, and Muslu (2010) and Cai and Walkling (2011). 7

8 studies acknowledge that the positive association between recommendations and votes might not be causal and suggest that the causal e ect of ISS can range between 6% and 35% (Ertimur, Ferri, and Oesch (2013), Choi, Fisch, and Kahan (2010)). By using exogenous variation in ISS recommendations, we estimate the causal e ect of ISS to be about 25%. Second, our setting allows us to distinguish between di erent channels of ISS in uence and show that the more likely channel is the certi cation e ect. This nding is broadly in line with Ertimur, Ferri, and Oesch (2013) and Iliev and Lowry (2015), who nd evidence consistent with certain institutions conducting their own independent research, and with Aggarwal, Erel, and Starks (2014), who show that investors reliance on ISS has declined over time, partly due to information received through media. The remainder of the paper is organized as follows. Section 2 outlines our empirical identi cation strategy. Section 3 describes the data and variable construction. Section 4 analyzes the e ect of ISS recommendations on voting outcomes, and Section 5 studies the channels through which this e ect occurs. Section 6 presents tests of the validity of the RD approach and discusses the robustness of the results. Section 7 shows how the e ect of ISS on voting outcomes varies across rms. Finally, Section 8 concludes. 2 Methodology A positive correlation between shareholder support and ISS recommendations does not imply that ISS has a causal e ect on voting outcomes because of omitted variables that could a ect both ISS recommendations and shareholder votes. In this section, we discuss how we address this identi cation issue through a regression discontinuity design. The source of exogenous variation in ISS recommendations is the following cuto rule used by ISS in 2010 and When giving recommendations on say-on-pay proposals, ISS used to rst screen companies based on their TSRs and focused its e orts on underperforming rms. Speci cally, ISS identi ed a company as underperforming if both its one- and three-year TSRs were below the respective median TSRs of other rms in the company s four-digit GICS group, where the TSRs were calculated on the last day of the calendar quarter closest to the company s scal-year-end. Note also that ISS only included Russell 3000 rms in its de nition of the company s four-digit GICS group. 14 After identifying the 14 See, e.g., the 2011 U.S. Proxy Voting Guidelines Summary white paper and the Financial Highlights 8

9 underperforming companies using these cuto s, ISS conducted an in-depth qualitative review of their compensation practices before giving a say-on-pay recommendation. 15 This rule suggests that ISS is likely to give a positive say-on-pay recommendation without conducting deep analysis if the rm is not identi ed as underperforming, but will carefully scrutinize the rm s compensation practices (pay-performance alignment, the strength of performance-based elements, pay relative to peers, quality of disclosures, and other issues) before giving it a positive recommendation if the rm is underperforming. 16 As we show in Section 4.1, this leads to a discrete jump in the probability of a negative recommendation for underperforming rms. We therefore implement a fuzzy RD design by instrumenting a negative ISS recommendation with an indicator variable BelowCuto, which equals one if the rm s one- and three-year TSRs both fall below their respective industry medians, and zero otherwise (see Imbens and Lemieux (2008) and Roberts and Whited (2012) for reviews of the fuzzy RD methodology.) Formally, for rm i in year t, BelowCuto it is given by 8 < 1 if MaxT SR it < 0; BelowCuto it = : 0 otherwise; where MaxT SR is the forcing variable, measured in percentage points and de ned as MaxT SR it = max(t SR (1) it MedianT SR (1) it ; T SR(3) it MedianT SR (3) it ); T SR (n) it is the n-year TSR of rm i computed in year t, and MedianT SR (n) it is the median n-year TSR in year t computed across all Russell 3000 rms in the same four-digit GICS group as rm i. We describe the calculation of the TSRs in Section 3. Data Overview, which says The GICS Sector TSR displayed for US companies is the median TSR for companies in the same 4-digit GICS group and Russell 3000 index membership i.e., for a company included in the Russell 3000 index, sector peers will be drawn only from the Russell 3000 index. 15 See Evaluating Pay for Performance Alignment ISS Quantitative and Qualitative Approach, Feb 17, See also page 38 of the 2011 U.S. Proxy Voting Guidelines Summary and pages of the 2010 SRI U.S. Proxy Voting Guidelines for descriptions of the same rule. 16 The rule does not imply that ISS always gives a positive recommendation to a rm above the cuto. Indeed, in addition to the above methodology, ISS checks for three key problematic pay practices that, according to its guidelines, can lead to a negative recommendation on a stand-alone basis: single-trigger or modi ed single-trigger provisions in severance contracts, tax gross-ups, and repricing of stock options without shareholder approval (e.g., the ISS 2011 U.S. Proxy Voting Guidelines Summary and Pearl Meyer & Partners Client Alert: ISS Issues Policy Updates and FAQs for 2011 Proxy Season.) This policy is consistent with our ndings: to understand the ISS methodology better, we examine the corresponding (2010 or 2011) proxy statements of the 78 rms that received a negative ISS recommendation despite being above the cuto and nd that more than 90% of these rms had at least one of these three pay practices. 9

10 The identi cation assumption is that the relation between voting support and the variable MaxT SR would be smooth around the cuto in the absence of di erential ISS recommendations around the cuto. This assumption is plausible because the cuto used by ISS is somewhat arbitrary. First, it is based on the TSRs of a speci c group of rms (those that are both in Russell 3000 and in the rm s four-digit GICS group), 17 and second, the TSRs are calculated on a speci c date (the last day of the quarter closest to the rm s scal-year end). We formally test this identi cation assumption in Sections To identify the causal e ect of ISS recommendations on voting outcomes, we conduct the two-stage least squares (2SLS) procedure by estimating the following two-equation system. The rst (second) equation corresponds to the rst (second) stage: NegRec = BelowCuto + f 1 (MaxTSR) + BelowCuto f 2 (MaxTSR) + X + u, Votes = NegRec + g 1 (MaxTSR) + BelowCuto g 2 (MaxTSR) + X + "; where Votes is the percentage of votes in favor of a say-on-pay proposal, NegRec is an indicator variable equal to one if the ISS recommendation is negative and zero if positive, f 1 ; f 2 ; g 1 ; and g 2 are continuous functions of the forcing variable MaxT SR, and X is a vector of control variables. As is standard in the literature (e.g., Imbens and Lemieux, 2008), we estimate a linear probability model for the rst stage, and show the robustness to the probit formulation in Table 6. The variable BelowCuto serves as the excluded instrument for the endogenous variable NegRec in the second-stage regression. Our main coe cient of interest is 1, which captures the local average treatment e ect of a negative ISS recommendation on voting support. Standard errors are computed as the usual 2SLS standard errors. Our main speci cation, presented in Tables 2 and 3, is a local linear regression, i.e., f i and g i are linear functions estimated on a small bandwidth around the cuto (we show the robustness to including higher-order polynomials in Table 6). In particular, we estimate NegRec = BelowCuto + 2 MaxTSR + 3 BelowCuto MaxTSR + X + u Votes = NegRec + 2 MaxTSR + 3 BelowCuto MaxTSR + X + ": 17 According to the investor survey by Bew and Fields (2012), investors use di erent peer groups from those used by proxy advisors. The ISS peer group selection has been criticized by companies for including rms engaged in a di erent business than the company and not including the company s publicly recognized competitors (see, e.g., the proxy statement amendment of Allegheny Technologies on April 29, 2011). 10

11 Following the suggestion in Imbens and Lemieux (2008), we focus on the rectangular kernel and verify the robustness of the results to di erent bandwidths in Table 6. In model 1 of Tables 2 and 3, we set 3 and 3 to zero, thus restricting the slope of the linear regression to be the same on the two sides of the cuto. In models 2-5, we allow 3 and 3 to be di erent from zero, thus allowing for di erent slopes on the two sides of the cuto. Note that as long as the covariates are continuous around the cuto (the assumption that we verify for various rm characteristics in Section 6.2), fuzzy RD design does not require the inclusion of control variables other than the forcing variable to produce consistent estimates (Imbens and Lemieux, 2008). Nevertheless, we include several rm characteristics that may a ect voting support for say-on-pay proposals, such as characteristics of the CEO compensation package, the rm s ownership structure, and other rm characteristics, as well as year and industry xed e ects, and show the robustness of our results. We estimate the above equations on a narrow bandwidth around the cuto, i.e., we only include observations that satisfy h < MaxT SR < h, where h is the bandwidth. Following the practical considerations in Imbens and Lemieux (2008), we use the same bandwidth for the rst- and second-stage regressions. The tradeo in choosing the bandwidth is that a larger bandwidth increases precision by including more observations, but also introduces an additional bias. In our main analysis in Tables 2-5, we use a bandwidth of 5%. In Table 6, we repeat the analysis on many alternative bandwidths and show that our estimates are robust to the choice of the bandwidth. In addition, we apply the cross-validation procedure that is commonly used in the literature to determine the optimal bandwidth. The details of this procedure are outlined in Appendix C. The cross-validation procedure yields the optimal bandwidth between 4% and 5%, consistent with our baseline bandwidth of 5%. 3 Data and variable construction The data on ISS recommendations and voting outcomes come from the ISS Voting Analytics database, which covers Russell 3000 rms. For each rm and each proposal on the agenda, the database provides the percentage of votes for, votes against, and abstentions, whether the proposal passed or failed, and the ISS recommendation on the proposal. There is variation across rms in the way voting support is calculated. While the numerator is always the number of votes in favor of the proposal, the denominator, captured by the variable base, 11

12 is di erent across rms. Base can be the sum of the votes in favor and against the proposal (51.83% of the sample), the sum of the votes in favor and against the proposal plus the number of abstentions (47.77% of the sample), or the total number of shares outstanding (0.40% of the sample). We use the appropriate denominator to calculate Votes, the percentage voting support for each company. We focus on observations for which both the voting support and the ISS recommendation are observed. We obtain the data on TSRs and rm characteristics from Compustat, the data on institutional ownership from Thomson Reuters 13F, and the data on executive compensation and insider ownership from GMI Ratings (formerly Corporate Library). 18 We rst match the ISS sample to Compustat and CRSP by ticker and company name. 19 We then merge the sample with GMI Ratings by ticker and name, and with Thomson Reuters 13F by CUSIP, name, and ticker. Appendix A contains the de nitions of all the variables used in the paper. For most of the analysis, except the falsi cation tests in Section 6.3, we focus on sayon-pay proposals voted on in 2010 and 2011 the years in which ISS applied its cuto methodology. Even though the cuto rule might have been used prior to 2010, we did not nd a formal mentioning of this rule in the ISS guidelines. We do not lose many observations by not including prior years because say-on-pay proposals were relatively rare before 2011, when they were made mandatory by the Dodd-Frank Act. 20 We also do not include 2012 because, as discussed in Section 6.3, ISS stopped using its cuto rule that year. Instead, we use the observations from 2012 (as well as the sample of director elections) as a falsi cation test to show the validity of our RD setting. Constructing the TSR cuto s According to the ISS guidelines, ISS utilizes S&P s Compustat database for TSR calculated values. The Total Return concepts are annualized rates of return re ecting price appreciation plus reinvestment of dividends (calculated monthly) and the compounding e ect of dividends 18 We use GMI Ratings and not Execucomp because Execucomp only tracks executive compensation in S&P 1500 rms, while our sample covers Russell 3000 rms. 19 We conduct two rounds of matching to ensure that the match is correct. The rst approach is to merge the data with Compustat to obtain GVKEY (and CUSIP) and then merge with the CRSP/COMPUSTAT Merged link table to get PERMNO (and CUSIP). The second approach is to rst merge the data with the stock names le from CRSP to get PERMNO, and then merge with the link table to obtain GVKEY. We cross-check the observations after the two rounds of matching. 20 There were only 404 say-on-pay proposals during the period, compared to 2,020 say-on-pay proposals in

13 paid on reinvested dividends. We therefore also use Compustat s Total Return concept to calculate TSRs. In particular, we multiply the current month s adjusted close price by the current month s Total Return Factor provided by Compustat, divide the result by the product of the adjusted close price multiplied by the Total Return Factor from the prior period (one or three prior years for T SR (1) and T SR (3), respectively), and annualize the three-year return. The Total Return Factor is a multiplication factor that re ects monthly price appreciation and reinvestment of monthly dividends and cash equivalent distributions and the compounding e ect of dividends paid on reinvested dividends. The TSR cuto for each rm is calculated as the median TSR of Russell 3000 rms in the same four-digit GICS industry. 21 ISS guidelines specify the following dates to compute TSRs. TSRs are downloaded at the end of each calendar quarter, i.e., on the last day of March, June, September, and December. For a given rm and year, the relevant download date is the last day of the calendar quarter closest to the scal-year end of the subject rm. For example, if a rm s scal-year-end is March 31, TSRs for rms in the same industry group are calculated on March Besides manually calculating the median industry TSRs, we obtain the list of median TSRs from the ISS website. For example, Appendix B presents a screenshot of the webpage with median TSRs for each four-digit GICS industry downloaded at the end of the four most recent calendar quarters. We obtain similar tables for most periods in our sample and nd that they mostly match our manually calculated cuto s. Because the medians from the ISS website capture the cuto s used by ISS more precisely, we use these medians for all quarters for which they are available on the ISS website and use our manually-calculated medians for the remaining quarters. For robustness, we have repeated the analysis using the manually-calculated medians for all quarters and obtained similar results. ISS guidelines do not specify whether a given rm s TSR, which is compared to the industry median cuto, is downloaded on the rm s scal-year-end date or, similarly to the corresponding industry median, at the end of the calendar quarter closest to its scalyear-end. 23 The two approaches coincide for rms whose scal-year-end falls on the end of 21 The list of Russell 3000 rms was obtained from Bloomberg. 22 See, e.g., the guidelines on the ISS website, which say ISS downloads TSR performance data at the end of March, June, September, and December for each of the four-digit industry GICS groups to determine the median applicable for Russell 3000 companies. Given that executive compensation is reported on a scal year basis, and that scal year ends vary by company, the applicable list will depend on the closest TSR performance download where a company s scal year falls. This methodology was also con rmed in authors communication with ISS representatives. 23 ISS uses TSRs downloaded on the rm s scal-year-end date for some of its other recommendations, such 13

14 the calendar quarter, which constitute 90% of the sample. To follow the ISS methodology precisely, we restrict our sample to these 90% of rms. We also note that regardless of the company s TSRs, ISS always gives it a positive say-on-pay recommendation if the total dollar value of CEO compensation is su ciently small, in particular, below the fth percentile of rms in that year. This implies that the fuzzy RD setting is not directly applicable for this group of rms, so we restrict our sample to observations with the total value of CEO compensation above the fth percentile in that year. Our results are very close if we keep these observations, and the only di erence is that at the rst stage, the discontinuity in the probability of a negative ISS recommendation around the cuto is slightly smaller. Descriptive statistics Our nal sample covers 1,932 rms and 2,020 say-on-pay proposals in 2010 and 2011: 106 in 2010, and 1,914 in 2011, when say-on-pay became mandatory for a large number of rms. 24 Panel A of Table 1 presents descriptive statistics of the sample rms. The average one- and three-year TSR is 29.86% and -0.8%, respectively. The number of observations below the ISS cuto (i.e., with MaxT SR < 0) is 613, which constitutes 30% of the sample. The average market capitalization is $4.35 billion, the average total value of executive compensation is $4.76 million, and the average say-on-pay voting support is 90.13%. Panel B presents voting outcomes depending on the ISS recommendation. The probability that ISS gives a negative recommendation is 12.7% (256 observations out of 2,020). The average voting support is 93.2% if ISS gives a positive recommendation, and 68.9% if ISS gives a negative recommendation. While all 1,764 proposals with a positive ISS recommendation received more than 50% voting support and thus passed, 29 out of 256 proposals with a negative recommendation failed. These statistics are consistent with the strong positive association between ISS recommendations and voting outcomes documented in prior studies. [INSERT TABLE 1 HERE] For illustrative purposes, Figure 1 plots the distribution of ISS recommendations (on the left) and voting support for say-on-pay proposals (on the right) on a relatively wide as evaluating pay-for-performance alignment. See, e.g., Evaluating Pay for Performance Alignment. 24 The fact that say-on-pay proposals were not mandatory in 2010 does not a ect our methodology: the factors that led a say-on-pay proposal to be included in the agenda do not invalidate the RD design as long as they are continuous around the cuto. 14

15 interval around the cuto. The x-axis presents the forcing variable M axt SR, measured in percentage points. Each dot in the left gure represents the average probability of a negative ISS recommendation measured in absolute values (from 0 to 1) in bins of 1%. Similarly, each dot in the right gure represents the average voting support, Votes, measured in percentage points, in bins of 1%. The solid lines in both gures represent the tted values of a cubic polynomial. Visual inspection of the graphs reveals a discontinuity in both variables around the cuto. We formally verify the presence of these discontinuities and quantify their exact magnitudes in Sections 4.1 and 4.2, by estimating local linear regressions in a more narrow bandwidth. [INSERT FIGURE 1 HERE] 4 E ect of ISS on voting outcomes In this section, we answer the rst question: What is the e ect of ISS recommendations on voting outcomes? Section 4.1 documents a discontinuity in the probability of a negative ISS recommendation around the cuto, and Section 4.2 uses this discontinuity to estimate the causal e ect of ISS recommendations on voting support. 4.1 Discontinuity of ISS recommendations around the cuto We start by showing that the probability of a negative ISS recommendation increases discontinuously when a rm s one- and three-year TSRs fall below the industry median TSRs. We present the graphical analysis in Figure 2 and the results of the rst-stage estimation in Table 2. The x-axis of Figure 2 presents the forcing variable MaxT SR in a 5% bandwidth around the cuto. Each dot shows the average probability of a negative ISS recommendation, measured in absolute values (from 0 to 1), in bins of 1%. The solid and dashed lines represent, respectively, the tted values and 95% con dence intervals of the rst-stage regression, estimated on the interval 5% < M axt SR < 5%. This subsample includes 403 observations, with 175 of them corresponding to rms below the cuto. Both visual inspection of the averages across bins and the local linear regression show that the probability of a negative recommendation is discontinuous and jumps from about 10% to about 25% once the company falls below the cuto. This increase is economically large given 15

16 that the sample average probability of a negative recommendation is 12.7%. Importantly, although the con dence intervals slightly overlap, the estimates on the left and right are signi cantly di erent at the 5% level as is formally shown in Table [INSERT FIGURE 2 HERE] Table 2 reports the results of the rst-stage regression for several speci cations estimated on a 5% bandwidth around the cuto. Model 1 restricts the slope of the linear control function to be the same on both sides of the cuto, while models 2-5 allow for di erent slopes around the cuto by including the interaction term BelowCuto MaxT SR. Model 3 extends model 2 by adding year and industry xed e ects. 26 In models 4 and 5, we add additional control variables that have been used in the literature on say-on-pay and proxy advisors (e.g., Ertimur, Ferri, and Oesch, 2013). Speci cally, in model 4, we control for corporate governance characteristics that are likely to a ect ISS recommendations and voting support on say-on-pay. They include various characteristics of the executive compensation package (the total dollar value of executive compensation, the percentage change in executive compensation from the previous year, and the proportion of stock-based compensation de ned as the proportion of total value coming from stock and option awards), as well as the percentage of institutional and insider ownership. 27 In addition, in model 5, we control for other rm characteristics such as rm size (measured as the logarithm of the market value of equity), return on assets (ROA), and market-to-book ratio. In all speci cations, the coe cient on BelowCuto is about 0.15, i.e., the probability of a negative ISS recommendation for rms just below the cuto is by 15 percentage points higher than for rms just above the cuto. This e ect is consistent with the visual estimate from Figure 2. The test for comparing coe cients between the models (Clogg, 1995) con rms that the coe cients on BelowCuto across all speci cations are not statistically signi cantly di erent from each other. [INSERT TABLE 2 HERE] 25 It is generally known in statistics that the test of non-overlapping 95% con dence intervals is more conservative and corresponds to a lower p-value than 0.05 (e.g., Goldstein and Healy (1995), Schenker and Gentleman (2001)). 26 The industry dummies are based on four-digit GICS classi cation to make the de nition of the industry consistent with that used by ISS. 27 Since the say-on-pay vote is to approve executive compensation over the preceding scal year, we use compensation variables from the scal year preceding the scal year of the shareholder meeting. 16

17 4.2 E ect of ISS recommendations on voting support Using the discontinuity in ISS recommendations as an instrument, we next analyze the e ect of ISS on voting outcomes. We start by presenting the graphical analysis in Figure 3. This gure is similar to Figure 2, but the y-axis now corresponds to Votes, voting support for the say-on-pay proposal, measured in percentage points. The gure shows a discontinuity in voting support around the cuto : Votes is by about four percentage points lower for rms just below the cuto relative to rms just above the cuto. Given the identi cation assumption, we attribute this discontinuity to the causal e ect of ISS recommendations. For comparison, we nd no similar discontinuity in votes for samples where the ISS cuto rule is not used (see Section 6.3 and Figure 5). [INSERT FIGURE 3 HERE] Table 3 presents the results of the second-stage regression. The outcome variable is Votes, measured in percentage points, and the main variable of interest is NegRec, the indicator of a negative ISS recommendation. We estimate several speci cations corresponding to those in Table 2. In particular, we control for year and industry xed e ects in models 3-5 and for governance and other rm characteristics in models 4-5. Importantly, our estimates are not biased by any omitted variables as long as these variables are continuous around the cuto. In models 1 and 2, the coe cient on NegRec is about -25 and is signi cant at the 5% level, suggesting that a negative ISS recommendation reduces the percentage of votes in favor of a say-on-pay proposal by 25 percentage points. As expected under the identi cation assumption, the e ect is quantitatively similar and remains signi cant once we include year and industry xed e ects and other control variables: the coe cient on NegRec in model 5 is and is signi cant at the 5% level. In Section 6.4, we verify the robustness of this e ect to the choice of the bandwidth and the inclusion of higher-order polynomial controls. [INSERT TABLE 3 HERE] The 25% e ect is economically important. Indeed, practitioners and prior academic studies consider voting dissent of above 20% to be an indication of strong shareholder dissatisfaction (e.g., Del Guercio, Seery, and Woidtke (2008), Ferri and Maber (2013)). According to a 2011 investor survey, 72% of investor respondents believe that voting dissent above 30% warrants an explicit response from the board regarding improvements in pay practices, and 20% 17

18 is the most commonly cited dissent level that should trigger such a response (see Policy Survey Summary of Results by ISS). Accordingly, say-on-pay dissent above 20%-30% prompts most rms to change their compensation policies. Ferri and Maber (2013) examine UK rms responses to say-on-pay votes and show that about 80% of rms with more than 20% voting dissent remove controversial pay practices, such as generous severance contracts and problematic performance-based vesting conditions in equity grants. Similarly, Ertimur, Ferri, and Oesch (2013) nd that more than 70% of US rms with at least 30% voting dissent change their compensation policies following the vote. Schwartz-Ziv and Wermers (2014) nd that companies with low say-on-pay support are likely to decrease excessive compensation and pick peer rms that more closely resemble their own company when ownership is concentrated. Why do rms respond to low say-on-pay support? In the past, low support levels, and especially failed say-on-pay votes, have led to shareholder lawsuits, negative media attention, and damage to rms reputation. In its guidelines, ISS notes that if a company fails to adequately respond to the company s previous say-on-pay proposal that received the support of less than 70 percent of votes cast, ISS may recommend to vote against the members of the compensation committee and potentially the full board, and Glass Lewis places extra scrutiny on rms with less than 75% approval. These consequences, together with the in uence of ISS on voting outcomes that we document in this paper, suggest that proxy advisors play an important role in rms governance practices. 5 Channels of ISS in uence In this section, we examine potential channels through which ISS may a ect voting outcomes. The rst is the informational channel: shareholders follow ISS because its recommendations and research reports provide new information to them. Because many institutions hold stock in thousands of rms, they may not be able to do the necessary research for all rms in their portfolios and may prefer to outsource this research to proxy advisors. According to Michelle Edkins, Managing Director at BlackRock, who represented the rm at the SEC roundtable on proxy advisors, There are days when we are voting 25, 30 meetings across our team. And so having that information synthesized and accessible is hugely important to us being able to take an informed decision. The informational role of ISS may be particularly important given that most investors are dissatis ed with the quality of information they receive through proxy 18

19 statements. According to a 2015 investor survey, only 38% of institutional investors believe that corporate disclosure about executive compensation is clear and easy to understand. Investors also complain that proxy statements are too long and claim to read only 32% of a typical proxy. 28 Proxy advisors can therefore help because they produce extremely detailed analyses of things that in many cases have been essentially rendered intentionally obscure in the proxy process (Damon Silvers, AFL-CIO representative at the SEC roundtable.) The other potential channel for the in uence of ISS is the certi cation e ect: following ISS recommendations may help an institutional investor prove that it voted in its clients best interests and thereby prevent criticism and litigation. The SEC commissioner Daniel Gallagher notes that relying on the advice from the proxy advisory rm became a cheap litigation insurance policy: for the price of purchasing the proxy advisory rm s recommendations, an investment adviser could ward o potential litigation over its con icts of interest (Gallagher, 2014). For example, some institutional investors may be reluctant to vote against management because they care about current or future business ties with the rm and the possibility of managerial retaliation (e.g., Davis and Kim (2007), Matvos and Ostrovsky (2010)). A positive say-on-pay recommendation from ISS could then play a certi cation role, allowing these investors to vote for management without the risk of being accused of con icts of interest. The certi cation role of ISS is further strengthened by its coordination role: following ISS helps institutions coordinate their votes with other shareholders and thereby decreases potential criticism both from their clients if they vote with management, and from management if they vote against it (Matvos and Ostrovsky, 2010). The certi cation and coordination e ects are closely related, so we refer to both as the certi cation channel and try to distinguish it from the informational channel. To understand which channel is more important, recall that if a rm falls above the cuto, ISS is likely to give a positive recommendation without conducting an in-depth qualitative review of the rm s compensation practices. In contrast, rms below the cuto are subject to an in-depth qualitative review and will only receive a positive recommendation if the review shows that their pay policies are appropriate. Consider two rms with a positive ISS recommendation and suppose that rm A is just below the cuto and rm B is just above 28 See the 2015 Investor Survey: Deconstructing Proxy Statements What Matters to Investors by Stanford University, RR Donnelley, and Equilar. See also the 2012 investor survey by Bew and Fields (2012), which further emphasizes proxy advisors role as data aggregators for asset managers. 19

20 the cuto. The ISS rule implies that the positive recommendation for rm A is based on an in-depth review, but the positive recommendation for rm B is not (or less likely to be). Under the assumption that the ISS in-depth review provides new valuable information to shareholders, the positive recommendation for rm A is a stronger positive signal about its pay practices than the positive recommendation for rm B. Thus, if the informational channel is important, i.e., investors change their voting decisions based on the information they receive from ISS, then voting support for rm A s say-on-pay proposal should be higher. In contrast, if investors primarily follow ISS recommendations for certi cation and not because of their actual information content, then voting support for rms A and B should be similar because investors do not care that the positive recommendation for rm A is more informative. Note that this argument does not rely on investors knowing about the ISS cuto methodology. Indeed, for every company in its portfolio, the client of ISS receives a research report that contains both ISS recommendations on each proposal and, importantly, the summary of the analysis underlying each recommendation. As long as the summary for a rm that received an in-depth review (i.e., a rm below the cuto ) is longer and more substantial than the summary for a rm that did not receive such a review (i.e., a rm above the cuto ), the informational channel would predict higher voting support for the rm below the cuto. Finally, it is quite likely that investors were aware of the ISS methodology and knew whether a given rm was above or below the cuto : ISS guidelines are widely publicized, and moreover, at the beginning of its research report for investors, ISS includes a table summarizing the rm s one- and three-year TSRs and the relevant cuto s, i.e., the median four-digit GICS group one- and three-year TSRs. We therefore focus on the subsample of observations with a positive ISS recommendation and compare voting support for rms just above and just below the cuto. Speci cally, we estimate the regression Votes = BelowCuto + 2 MaxT SR + 3 BelowCuto MaxT SR + X + " on a 5% bandwidth around the cuto. A positive and signi cant coe cient on BelowCuto would be evidence of the informational channel. Table 4 presents the results, with speci- cations 1-5 corresponding to those in Tables 2 and 3. In all speci cations, the coe cient on BelowCuto is insigni cant and has the opposite (negative) sign from the positive sign 20

21 predicted by the informational channel. The point estimate of 1 is about -2 (indicating that voting support just above the cuto is by 2 percentage points higher than just below the cuto ), with a standard error of 1.4. For comparison, the standard deviation of Votes in the subsample of rms with a positive ISS recommendation is 7.0 percentage points (and is 6.6 percentage points in the 5% bandwidth within this subsample). Thus, the coe cient on BelowCuto is statistically and economically insigni cant. The negative and insigni cant coe cient on BelowCuto is inconsistent with the predictions of the informational channel. This suggests two possibilities. One is that the in-depth review conducted by ISS does not provide new information to investors. The other is that the ISS review provides new information, but this information does not impact investors voting decisions. For example, if investors are reluctant to vote against management but do not want to be accused of con icts of interest, they are likely to follow a positive ISS recommendation regardless of what they think about its actual information content. Overall, our evidence suggests that, at least in our sample of say-on-pay proposals, the more likely channel of ISS in uence is the certi cation role of ISS recommendations. 29 [INSERT TABLE 4 HERE] 6 Validity of the RD design and robustness In this section, we perform several tests aimed to show the validity of our RD setting and the robustness of the results. 6.1 No manipulation of the forcing variable Our approach relies on the assumption that rms are randomly assigned to treatment, i.e., being just above or just below the cuto is random. In particular, we assume that rms cannot manipulate their TSRs in a way that pushes them just above the cuto. This assumption is plausible: rst, it is not that easy for a company to manipulate its TSR on a speci c date 29 Another potential way to test the informational channel would be to check whether the stock price reaction to a positive ISS recommendation is higher for rms just above than for rms just below the cuto. We cannot perform this test because ISS research reports are proprietary, and hence the date when they are sent to investors is not known to us. The only paper we know that has access to ISS proprietary reports and hence can look at the price reaction to the release of these reports is Ertimur, Ferri, and Oesch (2013). 21

22 given that it depends on stock price movements. Moreover, the TSR cuto for a given rm is a function of TSRs of all other rms in the industry, which is di cult to predict. To verify the assumption of no manipulation formally, we perform the procedure proposed by McCrary (2008), which tests for a discontinuity in the density of the forcing variable MaxT SR around the cuto. Figure 4 plots the estimated density of MaxT SR in a 5% bandwidth around the cuto and shows that the distribution is smooth. The absolute value of the McCrary test statistic for the 5% bandwidth is 0.84, which is not statistically signi cantly di erent from zero at any conventional level. 30 Thus, we cannot reject the null hypothesis that the density of the forcing variable is smooth around the cuto, suggesting that rms do not manipulate their TSRs to move above the cuto. [INSERT FIGURE 4 HERE] 6.2 Continuity of covariates To further test the assumption of random assignment to treatment, we compare the distribution of various rm characteristics just above and just below the cuto. Under the null hypothesis of random assignment, the distribution of characteristics una ected by ISS recommendations should be smooth around the cuto. We perform two sets of tests. First, we perform the RD analysis using each rm characteristic as the outcome variable. Speci cally, we regress each characteristic on BelowCuto, M axt SR, BelowCuto MaxT SR, and year and industry xed e ects using a 5% bandwidth around the cuto. The estimated coe cient and the standard error on the variable BelowCuto are reported in the second column of Table 5. The table shows that BelowCuto is not statistically signi - cant for any characteristic, consistent with the distribution being smooth around the cuto. We also repeat the graphical analysis similar to Figures 2-3 for each of these characteristics and do not nd any evidence of discontinuity either (these gures are omitted for brevity). Second, we compare the average value of each rm characteristic in two narrow intervals around the cuto : 5% < M axt SR < 0 and 0 < M axt SR < 5%. The results are summarized in Table 5. The p-values for the di erence in means test, presented in the last column of the table, con rm that the means of each rm characteristic on the two sides of 30 Both the gure and the McCrary test statistic were generated using the code provided by J. McCrary on his website: 22

23 the cuto are not statistically signi cantly di erent from each other. [INSERT TABLE 5 HERE] 6.3 Tests on alternative samples and placebo cuto s We further con rm the validity of our RD setting by repeating the analysis on two di erent samples. In particular, instead of looking at say-on-pay proposals in , which is our main sample, we study say-on-pay proposals in 2012 and director elections. (Other types of proposals are much less common and hence do not provide enough observations in a narrow bandwidth around the cuto.) The rationale for these falsi cation tests is that the ISS cuto rule does not apply to these two samples. Indeed, in 2012, ISS signi cantly changed its say-on-pay guidelines and, among other things, stopped using its cuto methodology. 31 We formally verify this by estimating the rst-stage regression on the 2012 sample and showing that the coe cient on BelowCuto is insigni cant (see Table D.1 in Appendix D). Similarly, the rst-stage regression on the sample of director elections in Table D.2 of Appendix D shows that the cuto rule does not a ect ISS recommendations on director elections either. 32 Thus, if our identi cation is valid and being above or below the cuto only a ects say-on-pay voting outcomes through ISS recommendations, then voting outcomes for say-on-pay proposals in 2012 and director elections should not exhibit any discontinuity around the cuto. Moreover, when examining director elections, we restrict attention to and to those rms in each year that had a say-on-pay proposal in that year and hence are included in the main sample. By restricting the sample this way, we ensure that our main sample of say-on-pay proposals in and the sample of director elections 31 For example, after discussing the cuto rule used in , the ISS 2012 white paper Evaluating Pay for Performance Alignment notes: This year,... ISS decided to re ne our approach to pay-forperformance evaluations and develop a more sophisticated methodology to drive the quantitative component of the analysis. The 2012 methodology is also aimed to screen rms before conducting an in-depth qualitative review, but the new screening rule does not allow us to perform the RD analysis for two reasons. First, the comparison group now consists of rms separately chosen by ISS for each company, which are not known to us. Second, the new screening is based on several characteristics of the compensation package (CEO s pay rank within a peer group, the multiple of the CEO s pay relative to the peer group, and the ve-year trend in CEO pay), which are easily manipulable. 32 Because each rm has many directors but only one say-on-pay proposal, we aggregate the director elections observations by rm-year-recommendation to make the sample of director elections consistent with our main sample. In particular, for each rm and year, we take all directors who received a positive (negative) recommendation from ISS and calculate the average proportion of votes in favor of these directors. 23

24 contain the same rms at the same points in time. Hence, continuity of voting outcomes in the second sample provides strong evidence that our RD design is valid. Figures 5a and 5b present the results of the analysis. Speci cally, we repeat Figure 3 on these two samples and show that in both samples, voting support is smooth around the cuto. The results of the regression analysis con rm the absence of discontinuity in votes at any conventional level of signi cance and are omitted for brevity. [INSERT FIGURE 5 HERE] Finally, we repeat the analysis on our original sample of say-on-pay proposals in , but using several placebo cuto s: MaxT SR = c for c 2 f 3%; 3%; 5%; 5%g. For each c, we rst verify that ISS recommendations exhibit no discontinuity around the cuto MaxT SR = c (these results are omitted for brevity) and then repeat the graphical analysis of Figure 3 but on the interval c 5% < MaxT SR < c + 5%. The results, presented in Figure 6, show no discontinuity in voting support around these placebo cuto s. [INSERT FIGURE 6 HERE] 6.4 Robustness tests In this section, we show the robustness of our main results in Tables 2 and 3 to alternative bandwidths and speci cations. Table 6 presents the summary of these results. The rst two rows in each panel present the estimate and standard error of the coe cient on BelowCuto in the rst-stage regression, and the third and fourth rows present the estimate and standard error of the coe cient on NegRec in the second-stage regression. Columns 1-8 of the table present the analysis of model 3 in Tables 2 and 3 on bandwidths ranging between 3% and 10%. 33 The results for both the rst and second stages are quantitatively similar across bandwidths. In addition, we calculate the optimal bandwidth using the cross-validation procedure described in Appendix C. We nd that the optimal bandwidth is between 4% and 5%, supporting our choice of 5% for the baseline analysis. An alternative to estimating a local linear regression on a narrow bandwidth is to use a larger sample but include higher-order polynomials of the forcing variable (e.g., Roberts and 33 The results for other speci cations are similar and are omitted for brevity. 24

25 Whited, 2012). In columns 9-11 of Table 6, we estimate regressions with higher-order polynomials of MaxT SR on a 20% bandwidth, allowing for di erent functional forms on the two sides of the cuto. For example, the regressors in column 9 include BelowCuto, MaxT SR, MaxT SR 2, BelowCuto MaxT SR, BelowCuto MaxT SR 2, and year and industry xed e ects. The table shows that using higher-order polynomials does not a ect our estimates. Finally, in column 12, we estimate the probit model for the rst-stage regression. The marginal e ect of the coe cient on BelowCuto is reported in square brackets to the right of the coe cient and equals 0.131, which is consistent with the estimates from the linear probability model. The coe cient on NegRec in the second-stage regression is strongly statistically signi cant and has a similar magnitude to that in other speci cations. [INSERT TABLE 6 HERE] 7 Variation in the e ect of ISS across rms The 25% estimate captures the impact of ISS on aggregate shareholder support and does not distinguish between di erent types of shareholders. ISS recommendations are likely to have a stronger impact on institutional investors (who are the main clients of ISS), and especially on those institutions who subscribe to ISS rather than to other proxy advisors. 34 In addition, ISS recommendations are likely to have a stronger impact on smaller institutions, which do not hold a concentrated stake in the rm. This is because institutions with a concentrated stake have stronger incentives to do their own independent research, and their research allows them both to vote in a value-increasing way and to rationalize their voting decisions, alleviating concerns about con icts of interest. Thus, institutions with a concentrated stake are less likely to follow ISS recommendations, regardless of whether they rely on these recommendations for certi cation or for information. In this section, we present evidence consistent with these hypotheses by studying how the impact of ISS on voting outcomes varies across rms. In particular, we examine the impact of ISS depending on the level of institutional ownership and the degree of institutional ownership 34 ISS controls about 61% of the market in terms of its clients equity assets. The second largest proxy advisor, Glass Lewis, controls 36% of the market, and all other proxy advisory rms have a much smaller market share (Government Accountability O ce report Corporate Shareholder Meetings: Issues Relating to Firms That Advise Institutional Investors on Proxy Voting ). 25

26 concentration (the data on ISS subscribers are not publicly available). Of course, because the rm s ownership structure is determined endogenously, we should be cautious about the interpretation of the cross-sectional results in this section. To study the e ect of institutional ownership, we restrict the sample to observations in a bandwidth around the cuto and calculate the median institutional ownership in the resulting sample. We next divide this sample into two subsamples, based on whether institutional ownership falls below or above the median, and refer to the rst (second) subsample as the subsample of rms with low (high) institutional ownership. We then repeat the RD analysis in Sections on each of the two subsamples. Because the sample size drops twice when we cut the sample into these subsamples, we focus on a 10% bandwidth to avoid losing power and to keep the size of each subsample similar to the sample size in our main tests. The results of the estimation are presented in model 1 of Table 7. The coe cient on NegRec in the second stage is and for the low and high institutional ownership subsamples, respectively. Thus, the magnitude of the coe cients is consistent with the hypothesis that the e ect of ISS is stronger in rms with higher institutional ownership, although the formal test cannot reject the hypothesis that the two coe cients are equal. 35 We next examine institutional ownership concentration by looking at the institutional ownership Her ndahl-hirschman index, de ned as the sum of squared share ownership over all institutional investors (the variable instown_hhi from Thomson Reuters 13F). As before, we divide the sample into subsamples depending on whether ownership concentration is above or below the sample median and conduct the analysis separately on the two subsamples. Models 2 and 3 of Table 7 present results for two speci cations, not controlling and controlling for the level of institutional ownership, respectively. As expected, the impact of ISS on voting outcomes is stronger when institutional ownership is less concentrated, although the di erence in coe cients is not signi cant. Ertimur, Ferri, and Oesch (2013) and Iliev and Lowry (2015). These results are in line with the ndings of [INSERT TABLE 7 HERE] 35 To test for the di erence in coe cients, we use the Stata command cmp, which estimates multiequation systems and can be applied in the 2SLS context. 26

27 8 Conclusion Proxy voting is a key channel of shareholder activism and has been playing an increasingly in uential role in corporate governance. It is therefore important to understand what factors a ect investors voting decisions. Proxy advisory rms, and ISS in particular, have emerged as prominent players in the proxy voting process. In this paper, we examine the role of ISS by analyzing the e ect of its recommendations on voting outcomes and the channels of its in uence. To study these questions, we use exogenous variation in ISS recommendations due to a cuto rule in the ISS voting guidelines. Speci cally, when giving recommendations on say-on-pay proposals, ISS used to conduct an initial screen of rms based on their one- and three-year TSRs and performed a deeper analysis of a rm s compensation policies if both TSRs fell below certain industry-related cuto s. This rule suggests that the probability of a negative ISS recommendation increases discontinuously for rms just below the cuto, the implication that we verify in the data. We therefore use a regression discontinuity design to identify the causal e ect of ISS recommendations on voting outcomes. An advantage of this approach is that because the TSR cuto s are based on industry medians, the sample of rms in a bandwidth around the cuto is likely to be representative of the sample as a whole. Our results are twofold. First, we nd that ISS has a strong e ect on voting outcomes: a negative ISS recommendation reduces the percentage of votes in favor of a say-on-pay proposal by about 25 percentage points. This e ect is economically large, con rming many commentators concerns about the in uence of proxy advisors. Second, we distinguish between two potential channels of ISS in uence: the informational role of ISS recommendations and their certi cation role. By comparing voting support for companies with positive ISS recommendations that received and did not receive in-depth review from ISS, we do not nd evidence that investors rely on ISS recommendations because of their information content. Hence, at least in our sample, the evidence is more consistent with the certi cation channel, whereby following ISS recommendations helps institutions protect themselves from criticism and legal liability. Our ndings contribute to the ongoing debate on the role and economic impact of proxy advisory rms. 27

28 References [1] Aggarwal, Reena, Isil Erel, and Laura Starks, 2014, In uence of public opinion on investor voting and proxy advisors, Working paper. [2] Aghion, Philippe, John Van Reenen, and Luigi Zingales, 2013, Innovation and institutional ownership, American Economic Review 103, [3] Alexander, Cindy R., Mark A. Chen, Duane J. Seppi, and Chester S. Spatt, 2010, Interim news and the role of proxy voting advice, Review of Financial Studies 23, [4] Almazan, Andres, Jay C. Hartzell, and Laura T. Starks, 2005, Active institutional shareholders and costs of monitoring: Evidence from executive compensation, Financial Management 34, [5] Appel, Ian R., Todd A. Gormley, and Donald B. Keim, 2015, Passive investors, not passive owners, Working paper. [6] Bakke, Tor-Erik, and Toni M. Whited, 2012, Threshold events and identi cation: A study of cash shortfalls, Journal of Finance 67, [7] Beshears, John, 2013, The performance of corporate alliances: Evidence from oil and gas drilling in the Gulf of Mexico, Journal of Financial Economics 110, [8] Bethel, Jennifer E., and Stuart L. Gillan, The impact of the institutional and regulatory environment on shareholder voting, Financial Management 31, [9] Bew, Robyn, and Richard Fields, 2012, Voting decisions at US mutual funds: How investors really use proxy advisers, IRRC Institute. [10] Boone, Audra, L., and Joshua T. White, 2015, The e ect of institutional ownership on rm transparency and information production, Journal of Financial Economics, forthcoming. [11] Brav, Alon, Wei Jiang, and Kim Hyunseob, 2009, Hedge fund activism: A review, Foundations and Trends in Finance 4, [12] Cai, Jie, Jacqueline L. Garner, and Ralph A. Walkling, 2009, Electing directors, Journal of Finance 64, [13] Cai, Jie, and Ralph A. Walkling, 2011, Shareholders say on pay: Does it create value? Journal of Financial and Quantitative Analysis 46, [14] Chava, Sudheer, and Michael Roberts, 2008, How does nancing impact investment? The role of debt covenant violations, Journal of Finance 63, [15] Choi, Stephen, Jill Fisch, and Marcel Kahan, 2010, The power of proxy advisors: myth or reality? Emory Law Journal 59,

29 [16] Clogg, Cli ord C., 1995, Statistical methods for comparing regression coe cients between models, American Journal of Sociology, 100, [17] Crane, Alan D., Sebastien Michenaud, and James Weston, 2014, The e ect of institutional ownership on payout policy: Evidence from index thresholds, Working Paper. [18] Cuñat, Vicente, Mireia Gine, and Maria Guadalupe, 2012, The vote is cast: the e ect of corporate governance on shareholder value, Journal of Finance 67, [19] Cuñat, Vicente, Mireia Gine, and Maria Guadalupe, 2013, Say pays! Shareholder voice and rm performance, Working Paper. [20] Davis, Gerald F., and E. Han Kim, 2007, Business ties and proxy voting by mutual funds, Journal of Financial Economics 85, [21] Del Guercio, Diane, Laura Seery, and Tracie Woidtke, 2008, Do boards pay attention when institutional investor activists just vote no? Journal of Financial Economics 90, [22] Ertimur, Yonca, Fabrizio Ferri, and Volkan Muslu, 2010, Shareholder activism and CEO pay, Review of Financial Studies 24, [23] Ertimur, Yonca, Fabrizio Ferri, and David Oesch, 2013, Shareholder votes and proxy advisors: Evidence from say on pay, Journal of Accounting Research 51, [24] Ferri, Fabrizio, 2012, Low cost shareholder activism: A review of the evidence, Research Handbook on the Economics of Corporate Law, Claire Hill & Brett McDonnell, eds., Elgar Publishers. [25] Ferri, Fabrizio, and David A. Maber, 2013, Say on pay votes and CEO compensation: Evidence from the UK, Review of Finance 17, [26] Gallagher, Daniel M., 2014, Outsized power & in uence: The role of proxy advisers, Critical Legal Issues Working Paper Series. [27] Gillan, Stuart L., and Laura T. Starks, 2007, The evolution of shareholder activism in the United States, Journal of Applied Corporate Finance 19, [28] Goldstein, Harvey, and Michael J. R. Healy, 1995, The graphical presentation of a collection of means, Journal of the Royal Statistical Society A 158, [29] Hartzell, Jay C., and Laura T. Starks, 2003, Institutional investors and executive compensation, Journal of Finance 58, [30] Iliev, Peter, 2010, The e ect of SOX Section 404: Costs, earnings quality, and stock prices, Journal of Finance 65, [31] Iliev, Peter, and Michelle Lowry, 2015, Are mutual funds active voters? Review of Financial Studies 28,

30 [32] Imbens, Guido W., and Thomas Lemieux, 2008, Regression discontinuity designs: A guide to practice, Journal of Econometrics 142, [33] Karpo, Jonathan M., 2001, The impact of shareholder activism on target companies: A survey of empirical ndings, University of Washington working paper. [34] Keys, Benjamin J., Tanmoy Mukherjee, Amit Seru, and Vikrant Vig, 2010, Did securitization lead to lax screening? Evidence from subprime loans, Quarterly Journal of Economics 125, [35] Larcker, David F., Allan L. McCall, and Gaizka Ormazabal, 2014, Outsourcing shareholder voting to proxy advisory rms, Stanford Graduate School of Business Research Paper No [36] Levit, Doron, and Nadya Malenko, 2011, Nonbinding voting for shareholder proposals, Journal of Finance 66, [37] Liberman, Andres, 2013, The value of a good credit reputation: Evidence from credit card renegotiations, Working Paper. [38] Matvos, Gregor, and Michael Ostrovsky, 2010, Heterogeneity and peer e ects in mutual fund proxy voting, Journal of Financial Economics 98, [39] McCrary, Justin, 2008, Manipulation of the running variable in the regression discontinuity design: A density test, Journal of Econometrics 142, [40] Morgan, Angela, Annette Poulsen, Jack Wolf, and Tina Yang, 2011, Mutual funds as monitors: Evidence from mutual fund voting, Journal of Corporate Finance 17, [41] Mullins, William, 2014, The governance impact of index funds: Evidence from regression discontinuity, Working Paper. [42] Rauh, Joshua D., 2006, Investment and nancing constraints: Evidence from the funding of corporate pension plans, Journal of Finance 61, [43] Reuter, Jonathan, and Eric Zitzewitz, 2013, How much does size erode mutual fund performance? A regression discontinuity approach, Working Paper. [44] Roberts, Michael R., and Toni M. Whited, 2012, Endogeneity in empirical corporate nance, in: George Constantinides, Milton Harris, and Rene Stulz (eds.), Handbook of the Economics of Finance, Volume 2 (Elsevier). [45] Schenker, Nathaniel, and Jane F. Gentleman, 2001, On judging the signi cance of differences by examining the overlap between con dence intervals, American Statistician 55, [46] Schwartz-Ziv, Miriam, and Russell Wermers, 2014, Do small and large shareholders have a say on pay? Working Paper. 30

31 Figures and tables Figure 1: Probability of a negative ISS recommendation and voting support The gure illustrates the fuzzy RD design by plotting the distribution of negative ISS recommendations and voting support around the cuto. The x-axis presents the forcing variable M axt SR, measured in percentage points, in a 10% bandwidth centered around the cuto. The y-axis of the left gure corresponds to the probability of a negative ISS recommendation, measured in absolute values (from 0 to 1). The y-axis of the right gure corresponds to Votes, the percentage of votes in favor of the say-on-pay proposal, measured in percentage points. The dashed vertical line represents the cuto MaxT SR = 0. Each dot in the left (right) gure represents the average probability of a negative ISS recommendation (percentage of votes in favor of the say-on-pay proposal) in bins of 1%. The solid lines represent the tted values of a third-degree polynomial of M axt SR. Variable de nitions are provided in Appendix A. 31

32 Figure 2: Probability of a negative ISS recommendation around the cuto Probability of a Negative ISS Recommendation MaxTSR The gure shows that the probability of a negative ISS recommendation increases discontinuously for rms that fall just below the ISS cuto. The x-axis presents the forcing variable M axt SR, measured in percentage points, in a 5% bandwidth centered around the cuto. The y-axis corresponds to the probability of a negative ISS recommendation, measured in absolute values (from 0 to 1). The dashed vertical line represents the cuto MaxT SR = 0. Each dot represents the average probability of a negative ISS recommendation in bins of 1%. The solid and dashed lines represent, respectively, the tted values and 95% con dence intervals of the rst-stage regression NegRec = BelowCuto + 2 MaxT SR + 3 BelowCuto MaxT SR + u, estimated on a 5% bandwidth around the cuto, i.e., on the interval 5% < MaxT SR < 5%. This subsample includes 403 observations. Variables de nitions are provided in Appendix A. 32

33 Figure 3: Voting support around the cuto Percent of Votes in Favor MaxTSR The gure shows that voting support for say-on-pay proposals decreases discontinuously for rms that fall just below the ISS cuto. The x-axis presents the forcing variable M axt SR, measured in percentage points, in a 5% bandwidth centered around the cuto. The y-axis corresponds to Votes, the percentage of votes in favor of the say-on-pay proposal, measured in percentage points. The dashed vertical line represents the cuto MaxT SR = 0. Each dot represents the average percentage of votes in favor of the say-on-pay proposal in bins of 1%. The solid and dashed lines represent, respectively, the tted values and 95% con dence intervals of the regression of Votes on BelowCuto, MaxT SR, and the interaction term BelowCuto MaxT SR, estimated on a 5% bandwidth around the cuto, i.e., on the interval 5% < MaxT SR < 5%. This subsample includes 403 observations. Variable de nitions are provided in Appendix A. 33

34 Figure 4: Density of the forcing variable Density of MaxTSR MaxTSR The gure con rms that companies do not manipulate their TSRs to push themselves above the ISS cuto by showing that the density of the forcing variable is smooth around the cuto. The x-axis presents the forcing variable M axt SR, measured in percentage points, in a 5% bandwidth centered around the cuto. The y-axis corresponds to the density of M axt SR, measured in absolute values. The solid vertical line represents the cuto MaxT SR = 0. The gure shows the histogram, estimated density, and 95% con dence intervals of MaxT SR in a 5% bandwidth around the cuto, generated using the code provided by J. McCrary on his website and based on McCrary (2008). This subsample includes 403 observations. The absolute value of the corresponding test statistic is 0.84, which is not statistically signi cantly di erent from zero at any conventional level. 34

35 Figure 5: Falsi cation tests on alternative samples (a) Say-on-pay proposals in 2012 (b) Director elections in Percent of Votes in Favor Percent of Votes in Favor MaxTSR MaxTSR The gures show that voting support exhibits no discontinuity around the cuto MaxT SR = 0 for two samples where the ISS cuto rule does not apply. Figure (a) considers the sample of say-on-pay proposals in Figure (b) considers the sample of director elections, where to match our main sample, we restrict attention to 2010 and 2011 and to those rms in each year that had a say-on-pay proposal in that year, and aggregate the observations by rm-year-recommendation. The x-axis presents the forcing variable M axt SR, measured in percentage points, in a 5% bandwidth centered around the cuto. The y-axis corresponds to Votes, the percentage of votes in favor of the say-onpay proposal (directors), measured in percentage points. The dashed vertical line represents the cuto MaxT SR = 0. Each dot represents the average percentage of votes in favor of the say-on-pay proposal (directors) in bins of 1%. The solid and dashed lines represent, respectively, the tted values and 95% con dence intervals of the regression of Votes on BelowCuto, MaxT SR, and the interaction term BelowCuto M axt SR, estimated on a 5% bandwidth around the cuto, i.e., on the interval 5% < MaxT SR < 5%. This subsample includes 289 and 470 observations for 2012 and director elections, respectively. Variable de nitions are provided in Appendix A. 35

36 Figure 6: Placebo tests using alternative cuto s (a) Cuto -3% (b) Cuto 3% Percent of Votes in Favor MaxTSR Percent of Votes in Favor MaxTSR (c) Cuto -5% (d) Cuto 5% Percent of Votes in Favor Percent of Votes in Favor MaxTSR MaxTSR The gure shows that voting support on the sample of say-on-pay proposals in exhibits no discontinuity around placebo cuto s MaxT SR = c for c 2 f 3%; 3%; 5%; 5%g. For each c, the x-axis presents the forcing variable M axt SR, measured in percentage points, in a 5% bandwidth centered around the cuto MaxT SR = c. The y-axis corresponds to Votes, the percentage of votes in favor of the say-on-pay proposal, measured in percentage points. The dashed vertical line represents the cuto MaxT SR = c. Each dot represents the average percentage of votes in favor of the say-on-pay proposal in bins of 1%. The solid and dashed lines represent, respectively, the tted values and 95% con dence intervals of the regression of Votes on BelowCuto c, MaxT SR, and BelowCuto c MaxT SR, estimated on c 5% < MaxT SR < c + 5%, where BelowCuto c is the indicator variable that equals one if MaxT SR < c and zero otherwise. This subsample includes 364, 453, 327, and 458 observations for c = 3%, 3%, 5%, and 5%, respectively. Variable de nitions are provided in Appendix A. 36

37 Table 1: Descriptive statistics Panel A Mean 25th 50th 75th Std Dev TSR (1) (%) TSR (3) (%) MaxTSR (%) Market Value of Equity (in $ billion) Market-to-Book ROA CEO Total Compensation (in $ million) Proportion of Stock-Based Compensation Institutional Ownership Insider Ownership Voting support (in percentage points) Panel B Number of observations ISS recommendation Voting Support by voting outcome Mean 10th 50th 90th Std Dev Fail Pass Total Against 68.9% 48.5% 68.9% 88.7% 15.2% For 93.2% 83.3% 95.6% 99.1% 7.0% 0 1,764 1,764 Total (Against and For) 90.1% 73.5% 94.9% 99.0% 11.7% 29 1,991 2,020 The table presents descriptive statistics for the sample of 1,932 companies and 2,020 say-on-pay proposals in 2010 and Panel A presents summary statistics (average, percentiles, and standard deviation) for the variables used in the study. Panel B presents voting outcomes depending on the ISS recommendation. Columns 2-6 of Panel B present summary statistics of the voting support for the subsample of Against recommendations, the subsample of For recommendations, and the whole sample. The next two columns present the number of observations with a given voting outcome (fail or pass) for a given ISS recommendation, where a proposal fails if it receives less than 50% support. The last column presents the total number of observations with a given ISS recommendation. Variable de nitions are provided in Appendix A. 37

38 Table 2: Probability of a negative ISS recommendation ( rst stage) (1) (2) (3) (4) (5) NegRec NegRec NegRec NegRec NegRec BelowCuto 0.145** 0.146** 0.146** 0.148** 0.142** (0.068) (0.069) (0.070) (0.067) (0.067) MaxTSR (0.012) (0.016) (0.016) (0.016) (0.016) BelowCuto MaxTSR (0.024) (0.024) (0.023) (0.023) CEO Total Compensation 0.007* 0.018*** (0.004) (0.006) Change in CEO Total Compensation 0.062*** 0.061*** (0.012) (0.012) Proportion of Stock-Based Compensation (0.093) (0.095) Institutional Ownership (0.096) (0.096) Insider Ownership 0.412*** 0.360*** (0.125) (0.127) Log(Market Value of Equity) ** (0.020) ROA (0.172) M/B (0.023) Year FE No No Yes Yes Yes Industry FE No No Yes Yes Yes Observations R The table shows that the probability of a negative ISS recommendation increases discontinuously if the company falls just below the ISS cuto and corresponds to the rst stage of the 2SLS procedure. All speci cations are estimated on a 5% bandwidth around the cuto, i.e., on the interval 5% < M axt SR < 5%. The outcome variable is NegRec, the indicator variable that equals one if ISS gives a negative recommendation, and zero otherwise. The main variable of interest is BelowCuto, which is an indicator variable that equals one if the rm is below the ISS cuto (MaxT SR < 0), and zero otherwise. We estimate a linear probability model, and hence the coe cient on BelowCuto can be interpreted as the di erence in the probability of a negative ISS recommendation between rms just below and just above the cuto. Variable de nitions are provided in Appendix A. Standard errors are reported in parentheses. *, **, and *** represent signi cance at the 10%, 5%, and 1% levels, respectively. 38

39 Table 3: Causal e ect of ISS recommendations on voting outcomes (second stage) (1) (2) (3) (4) (5) Votes Votes Votes Votes Votes NegRec ** ** ** *** ** (11.210) (11.160) (11.588) (10.655) (10.853) MaxTSR (0.358) (0.450) (0.531) (0.455) (0.463) BelowCuto MaxTSR (0.560) (0.585) (0.544) (0.534) CEO Total Compensation ** *** (0.120) (0.241) Change in CEO Total Compensation (0.720) (0.717) Proportion of Stock-Based Compensation (2.228) (2.218) Institutional Ownership ** *** (2.246) (2.186) Insider Ownership ** *** (5.292) (4.861) Log(Market Value of Equity) 1.658** (0.665) ROA (4.247) M/B (0.585) Year FE No No Yes Yes Yes Industry FE No No Yes Yes Yes Observations R The table shows that a negative ISS recommendation causes a signi cant decline in the proportion of votes in favor of a say-on-pay proposal and corresponds to the second stage of the 2SLS procedure. The outcome variable is Votes, the percentage of votes in favor of a say-on-pay proposal, measured in percentage points. The instrumented variable is NegRec, the indicator variable that equals one if ISS gives a negative recommendation, and zero otherwise. Estimation is conducted via 2SLS, where NegRec = BelowCuto + 2 MaxTSR+ 3 BelowCuto MaxTSR+X + u is the rst stage, estimated in Table 2, and Votes = NegRec+ 2 MaxTSR + 3 BelowCuto MaxTSR+X + " is the second stage, estimated in this table. The variable BelowCuto is an indicator variable that equals one if the rm is below the ISS cuto (MaxT SR < 0), and zero otherwise. All speci cations are estimated on a 5% bandwidth around the cuto, i.e., on the interval 5% < MaxT SR < 5%. Variable de nitions are provided in Appendix A. Standard errors are reported in parentheses. *, **, and *** represent signi cance at the 10%, 5%, and 1% levels, respectively. 39

40 Table 4: Voting outcomes for rms with a positive ISS recommendation (1) (2) (3) (4) (5) Votes Votes Votes Votes Votes BelowCuto (1.386) (1.415) (1.450) (1.431) (1.396) MaxTSR (0.238) (0.305) (0.309) (0.304) (0.298) BelowCuto MaxTSR (0.490) (0.504) (0.494) (0.486) CEO Total Compensation *** *** (0.082) (0.124) Change in CEO Total Compensation (0.567) (0.568) Proportion of Stock-Based Compensation (1.951) (1.972) Institutional Ownership (2.027) (1.981) Insider Ownership 7.976*** 9.179*** (2.916) (2.853) Log(Market Value of Equity) 1.428*** (0.425) ROA 7.896** (3.348) M/B (0.499) Year FE No No Yes Yes Yes Industry FE No No Yes Yes Yes Observations R The table shows that voting outcomes for say-on-pay proposals with a positive ISS recommendation for companies just below and just above the ISS cuto are not statistically signi cantly di erent from each other. We rst restrict the sample to observations with a positive ISS recommendation and then estimate all speci cations on a 5% bandwidth around the cuto, i.e., on the interval 5% < MaxT SR < 5%. The outcome variable is Votes, the percentage of votes in favor of a say-on-pay proposal, measured in percentage points. The main variable of interest is BelowCuto, which is an indicator variable that equals one if the rm is below the ISS cuto (MaxT SR < 0), and zero otherwise. The coe cient on BelowCuto can be interpreted as the di erence between voting support for rms just below the cuto and voting support for rms just above the cuto, conditional on a positive ISS recommendation. Variable de nitions are provided in Appendix A. Standard errors are reported in parentheses. *, **, and *** represent signi cance at the 10%, 5%, and 1% levels, respectively. 40

41 Table 5: Distribution of rm characteristics around the cuto Log(Market Value of Equity) (0.29) M/B (0.15) ROA (0.02) CEO Total Compensation 0.06 (0.96) Change in CEO Total Compensation Proportion of Stock-Based Compensation RD coe. on -5%<MaxTSR<0 0<MaxTSR<5% p-value BelowCuto Mean # Obs Mean # Obs (di. in means) 6.33 (29.38) 0.06 (0.04) Institutional Ownership (0.04) Insider Ownership 0.01 (0.03) The table shows that the distribution of rm characteristics is smooth around the cuto. For each rm characteristic in the rst column, column RD coe. on BelowCuto presents the results of a local linear regression of this characteristic on BelowCuto, M axt SR, BelowCuto M axt SR, and year and industry xed e ects using a 5% bandwidth around the cuto. The estimated coe cients on BelowCuto are reported in the rst row, and standard errors are reported in parentheses. Subsequent columns present the means of each rm characteristic in two narrow intervals around the cuto : 5% < MaxT SR < 0 and 0 < MaxT SR < 5%, as well as the corresponding number of observations in these intervals. The last column presents the p-values for the di erence in means test. Variable de nitions are provided in Appendix A. 41

42 Table 6: Robustness (1) (2) (3) (4) (5) (6) 3% 4% 5% 6% 7% 8% BelowCuto 0.176** 0.170** 0.146** 0.186*** 0.186*** 0.190*** (First Stage) (0.087) (0.075) (0.070) (0.063) (0.057) (0.053) NegRec * ** ** *** *** *** (Second Stage) (12.011) (10.261) (11.588) (8.330) (7.807) (7.084) Observations (7) (8) (9) (10) (11) (12) 9% 10% 20% Quadratic 20% Cubic 20% Quartic 5% Probit BelowCuto 0.199*** 0.178*** 0.194*** 0.190*** 0.205*** 0.641** [0.131] (First Stage) (0.050) (0.047) (0.051) (0.065) (0.078) (0.324) NegRec *** *** *** *** ** *** (Second Stage) (6.367) (6.931) (7.073) (9.164) (10.181) (11.565) Observations ,244 1,244 1, The table shows that the choice of the bandwidth and the degree of the fuzzy RD polynomial do not have a material e ect on the results of the paper. The rst two rows in each panel present the estimate and standard error of the coe cient on BelowCuto in the rst-stage regression, and the third and fourth rows present the estimate and standard error of the coe cient on NegRec in the second-stage regression of the 2SLS procedure. Columns 1-8 present the results for the linear speci cation corresponding to model 3 in Tables 2 and 3 (i.e., with regressors BelowCuto, M axt SR, BelowCuto M axt SR, and year and industry xed e ects) on alternative bandwidths, ranging between 3% and 10%. In columns 9, 10, and 11, respectively, we estimate second, third, and fourth-order polynomial functions of the forcing variable M axt SR, allowing for di erent functional forms on the two sides of the cuto, on a 20% bandwidth around the cuto. For example, the regressors in column 9 include BelowCuto, MaxT SR, MaxT SR 2, BelowCuto MaxT SR, Below- Cuto MaxT SR 2, and year and industry xed e ects. In column 12, we estimate the probit model for the rst-stage regression. The marginal e ect of the coe cient on BelowCuto is reported in square brackets to the right of the coe cient. Variable de nitions are provided in Appendix A. *, **, and *** represent signi cance at the 10%, 5%, and 1% levels, respectively. 42

43 Table 7: Variation in the e ect of ISS across rms (1) (2) (3) Institutional Ownership Ownership Concentration Ownership Concentration Low High Low High Low High BelowCuto 0.176** 0.187*** 0.164** 0.158** 0.158** 0.156** (First Stage) (0.068) (0.069) (0.071) (0.066) (0.071) (0.066) NegRec *** *** *** * *** * (Second Stage) (8.649) (10.190) (12.066) (11.044) (12.884) (10.799) IO control No No Yes Yes MaxTSR controls Yes Yes Yes Yes Yes Yes Year FE Yes Yes Yes Yes Yes Yes Industry FE Yes Yes Yes Yes Yes Yes Observations The table shows how the e ect of ISS on voting outcomes varies with the rm s ownership structure. Model 1 considers the e ect of institutional ownership, and models 2 and 3 consider the e ect of institutional ownership concentration. Institutional ownership is de ned as total institutional ownership in fraction of shares outstanding (instown_perc from Thomson Reuters 13F). Ownership concentration is measured by the institutional ownership Her ndahl-hirschman index, de ned as the sum of squared share ownership over all institutional investors (the variable instown_hhi from Thomson Reuters 13F). We rst restrict the sample to observations with MaxT SR within a 10% bandwidth around the cuto and calculate the median value of institutional ownership (ownership concentration) in the resulting sample. Next, we divide this sample into two subsamples, based on whether the rm s institutional ownership (ownership concentration) falls below or above the median, and refer to the rst and second subsample as the subsample of rms with low and high institutional ownership (ownership concentration), respectively. We then repeat the 2SLS procedure on each of the two subsamples. Models 1 and 2 present the results for the linear speci cation with regressors BelowCuto, MaxT SR, BelowCuto MaxT SR, and year and industry xed e ects, and in model 3, we also control for the level of institutional ownership. Variable de nitions are provided in Appendix A. *, **, and *** represent signi cance at the 10%, 5%, and 1% levels, respectively. 43

44 Appendix A. Variable de nitions Main variables Variable De nition Source T SR (1) (%) One-year TSR, de ned as the one year percentage change in the adjusted close price multiplied by the total return factor: T SR (1) =100[(PRCCM t TRFM t /AJEXM t ) / (PRCCM t 1 TRFM t 1 /AJEXM t 1 ) - 1]. CSRP/Compustat Merged T SR (3) (%) Three-year TSR, de ned as the annualized three year percentage change in the adjusted close price multiplied by the total return factor: T SR (3) =100[[(PRCCM t TRFM t /AJEXM t ) / (PRCCM t 3 TRFM t 3 /AJEXM t 3 )] 1=3-1]. CSRP/Compustat Merged MaxTSR (%) De ned as max(t SR (1) it MedianT SR (1) it ; T SR(3) it MedianT SR (3) it ), where T SR (n) it is the n-year TSR of rm i in year t, and MedianT SR (n) it is the median n-year TSR in year t computed across all Russell 3000 rms in the same four-digit GICS group as rm i. CSRP/Compustat Merged BelowCuto Indicator variable that takes the value of one if MaxTSR is negative, and zero otherwise. CSRP/Compustat Merged Votes (%) The percentage of votes in favor of a say-on-pay proposal. ISS Voting Analytics NegRec The indicator variable that takes the value of one if ISS gives a negative ("Against") recommendation, and zero otherwise. ISS Voting Analytics 44

45 Control variables Variable De nition Source Log(Market Value of Equity) M/B Natural logarithm of the market value of equity, which is calculated by multiplying the company s stock price by its number of outstanding shares as of its scal-year-end. The ratio of the market value of assets (equity market capitalization plus the book value of other liabilities) to the book value of assets. Compustat Compustat ROA Net income divided by total assets: ROA=NI/AT. Compustat CEO Total Compensation ($ million) Proportion of Stock-Based Compensation Change in CEO Total Compensation Institutional Ownership Insider Ownership The total compensation of the CEO (variable CEOTotSum- Comp from GMI Ratings) as reported in the company s proxy statement. It equals the aggregate total dollar value of each form of compensation quanti ed in the summary compensation table, including base salary, bonus, stock awards, option awards, non-equity incentive plan, change in pension value and nonquali ed deferred compensation, and all other compensation. The sum of stock and option awards divided by CEO total compensation: (CEOOptionAwards + CEOStock- Awards)/CEOTotSumComp from GMI Ratings. De ned as (CEO Total Compensation in year t - CEO Total Compensation in year t-1)/ceo Total Compensation in year t-1. Total institutional ownership in fraction of shares outstanding (INSTOWN_PERC from Thomson Reuters Institutional (13f) Holdings - Stock Ownership Summary). The estimated fraction of outstanding shares held by top management and directors, as reported in the company s most recent proxy statement (InsidersPctg from GMI Ratings). GMI Ratings GMI Ratings GMI Ratings Thomson Reuters 13F GMI Ratings 45

46 B. TSR cuto s from the ISS website The following picture presents a screenshot from the ISS website: It contains the list of ISS cuto s, i.e., one- and three-year median TSRs for each four-digit GICS group, downloaded at the end of the four most recent calendar quarters. (ISS provides these cuto s because it still uses them for some issues, e.g., CEOchairman separation.) We downloaded similar lists from the ISS website for most periods in our sample. The following picture presents an expanded view of the last table above, with the TSRs downloaded on March 31,