CEO Pay and the Market for CEOs

Finance and Economics Discussion Series Divisions of Research & Statistics and Monetary Affairs Federal Reserve Board, Washington, D.C. CEO Pay and the Market for CEOs Antonio Falato, Dan Li, and Todd Milbourn 2012-39 NOTE: Staff working papers in the Finance and Economics Discussion Series (FEDS) are preliminary materials circulated to stimulate discussion and critical comment. The analysis and conclusions set forth are those of the authors and do not indicate concurrence by other members of the research staff or the Board of Governors. References in publications to the Finance and Economics Discussion Series (other than acknowledgement) should be cleared with the author(s) to protect the tentative character of these papers.

CEO Pay and the Market for CEOs Antonio Falato1 Dan Li Todd Milbourn Federal Reserve Board Federal Reserve Board Washington University in St Louis April 2012 1CorrespondingAuthor: AntonioFalato, FederalReserveBoard-DivisionofResearchandStatistics, Washington DC. Phone: (202) 452-2861. Email: antonio.falato@frb.gov. Comments from Joseph Altonji, Lucian Bebchuk, E¢ Benmelech, Nick Bloom, Claudia Custodio (Discussant at the 2011 WFA Meetings), Alex Edmans, Carola Frydman (Discussant at the 2011 Duke-UNC Corporate Finance Conference), Xavier Gabaix, Gerald Garvey, Radakrishnan Gopalan, Nellie Liang, Kevin Murphy, Thomas Noe (Discussant at the 2010 Conference on Labor and Finance, Oxford University), Teodora Paligorova, Francisco Perez-Gonzalez, Raven Saks, Steve Sharpe, Josef Zechner (Discussant at the 2010 EFA Meetings), and seminar participants at the Federal Reserve Board and Blackrock are gratefully acknowledged. Suzanne Chang, Lindsay Relihan, Nicholas Ryan, and Lincoln Verlander provided excellent research assistance. All remaining errors are ours. The analysis, conclusions, and discussion in this paper are those of the authors and do not indicate concurrence by other members of the research sta⁄ or the Board of Governors of the Federal Reserve System. References to the Finance and Economics Discussion Series(otherthanacknowledgement)shouldbeclearedwiththeauthorstoprotectthetentativecharacterofthese papers.

Abstract Competitive sorting models of the CEO labor market (e.g., Edmans, Gabaix and Landier (2009)) predict thatdi⁄erencesinCEOproductiveabilities,or(cid:147)talent(cid:148),shouldbeanimportantdeterminantofCEOpay. However, measuring CEO talent empirically represents a major challenge. In this paper, we document reliable evidence of pay for CEO credentials and argue that the evidence is consistent with models of the CEO labor market. Our main (cid:133)nding is that boards(cid:146)compensation decisions reward several reputational, career, and educational credentials of CEOs, with newly-appointed CEOs earning a 5 percent ($280,000) total pay premium for each decile improvement in the distribution of these credentials. Consistent with boards using credentials as publicly-observable signals of CEO abilities, we show that pay for credentials displays key cross-sectional features predicted by theory, such as convexity in credentials and complementarity with (cid:133)rm size. Our main (cid:133)nding is robust to a battery of identi(cid:133)cation tests that address selectivity and endogeneity concerns, including instrumental variables estimates and controlling for (cid:133)rm and CEO (cid:133)xed e⁄ects. We also show that credentials capture variation in CEO human capital that is di⁄erent from lifetime work experience, and are positively related to long-term (cid:133)rm performance and board monitoring, which helps to distinguish our results from alternative stories based on CEO general human capital, hype, and entrenchment. Overall, our (cid:133)ndings suggest that sorting considerations in the CEO labor market are an important determinant of CEO pay. Our results also suggest that the rise in CEO pay over the last decades may owe at least in part to a rise in the CEO talent premium.

1 Introduction Public corporations invest considerable resources in the search for top executive talent. Recent theories, such as Edmans, Gabaix, and Landier (2009), Gabaix and Landier (2008), and Tervio (2008), argue that competition for talent in the CEO labor market is an important determinant of CEO pay. However, while some recent empirical studies point to an increased importance of the labor market for CEOs over the last two decades,1 we know relatively little about whether di⁄erences in CEO productive abilities are an important empirical determinant of CEO pay. That is, the existing literature on the CEO labor market is mostly theoretical, the evidence we do have is indirect, and ultimately we know relatively little about the extent to which di⁄erences in CEO abilities matter for pay. In order to (cid:133)ll this gap, we explore the empirical relation between several observable CEO credentials and pay (cid:150)which we denote as pay for CEO credentials (cid:150)and examine whether this relation is consistent with theory. We develop and test the cross-sectional implications of a stylized competitive assignment model of the market for CEOs. If observed credentials provide valuable signals of CEOs(cid:146)productive abilities, then we expect that pay packages should reward credentials. Theory also suggests that pay for credentials should be convex in credentials and complementary with (cid:133)rm size. Most importantly, better credentials should be positively associated with (cid:133)rm performance. We explore these predictions using a large hand-collected sample of 2,195 CEO successions between 1993 and 2005. The critical step is to construct measures of CEO credentials that plausibly re(cid:135)ect public information available to boards at the time they make pay decisions. We consider three such measures of credentials based on each CEO(cid:146)s resume: her industry reputation, labor market status, and educational pedigree. First, the CEO reputational signal, Press, measures outside perceptions of CEO abilities and is constructed by counting the number of articles containing the CEO(cid:146)s name that appear in the major business newspapers in the year prior to the CEO(cid:146)s appointment (as in Milbourn (2003)). The basic idea is that to have been previously recognized by the business press should be perceived by boards as a good signal. Second, the labor market signal, Fast-Track Career, measures the quality of the CEO(cid:146)s career record and is de(cid:133)ned as the age at which the executive (cid:133)rst took a CEO job. Intuitively, if the market for CEOs is at least in part meritocratic (see Kaplan, Klebanov, and Sorensen (2011) for evidence of such), the younger an executive is when she gets her (cid:133)rst CEO job, the more positive is the signal of her abilities. Third, the schooling signal, Selective College, measures the quality of CEO educational background and is constructed using Barron(cid:146)s rankings of the selectivity of the CEO(cid:146)s undergraduate 1MurphyandZÆbojn(cid:237)k(2007)showthatthereisasigni(cid:133)canttrendtowardmoreexternalhiresoverthelastthreedecades, which has been accompanied by an upward trend in pay. 1

college. Based on signaling models of education (Spence (1973)), we expect attendance at more selective colleges to be a stronger signal about CEO abilities. We further re(cid:133)ne these basic de(cid:133)nitions of Press and Fast-Track Career to address any concern that theymightcapturevariationunrelatedtoreputationalormarketsignals. OneconcernisthatPress might re(cid:135)ect bad press. In robustness tests, we ensure that the number of articles is not merely a re(cid:135)ection of CEO infamy by screening for the tone of each article and netting out negative press coverage, or Bad Press, from Press. A second concern is that the article count simply re(cid:135)ects luck or characteristics of the (cid:133)rm that previously employed the CEO. We address this by screening the tone of each article to re(cid:135)ect only positive personal traits of the CEO based on Kaplan, Klebanov, and Sorensen (2011) and only count articles that contain mention of such traits, which we denote as Good Press.2 Finally, we ensure that Fast-Track Career does not simply re(cid:135)ect common circumstances of the (cid:133)rst CEO job (see Malmendier, Tate, and Yan (2011) and Schoar (2007)) by using a cohort-adjustment aimed at capturing only variation beyond factors common across the same age cohort of executives. OurthreemeasuresofPress, Fast-Track Career andSelective College provideauniqueopportunityto assesswhetherandwhyCEOcredentialsmatterforCEOpay.Themain(cid:133)ndingofourstudyisthatthere is reliable evidence of pay for CEO credentials for newly-appointed CEOs. In particular, we show that across our three measures, CEOs with better credentials earn signi(cid:133)cantly higher total compensation. Our estimates imply an empirical sensitivity of (cid:133)rst-year total CEO pay per credentials decile ranging from about 5% for Press and Fast-Track Career to about 2% for Selective College. These estimates are alsoeconomicallysigni(cid:133)cant(cid:150)CEOswhoareonedecilehigherinthecredentialsdistributionearnabouta $280,000 premium.3 Results for a nearest-neighbor matching estimator (Abadie and Imbens (2007)) and a standard Heckman (1979) selection analysis con(cid:133)rm these baseline estimates, suggesting that selection on observables and the non-random nature of our CEO succession sample are not to blame. Since theory predicts that total compensation should be increasing in CEO talent, the positive relation between pay and CEO credentials o⁄ers a (cid:133)rst indication consistent with boards(cid:146)relying on credentials as signals of CEO productive abilities. Next, we document key cross-sectional features of pay for CEO credentials (cid:150)convexity and complementarity with (cid:133)rm size (cid:150)and argue that they are consistent with models of the labor market for CEO talent. We (cid:133)rst use a piece-wise linear speci(cid:133)cation to allow for heterogeneity in the relation between total CEO pay and credentials at di⁄erent levels of the credentials distribution. We show that there is a 2We also consider ratios of these (cid:133)ner press counts to control for (cid:133)rm-related press. 3In a battery of robustness checks, we show that these estimates are robust to alternative de(cid:133)nitions of the proxies, as well as adjustments at the (cid:133)rm and industry level. 2

convexrelationbetweenpayandcredentialsthatisstatisticallyandeconomicallysigni(cid:133)cant. Forthetop decileofthecredentialsdistributionforPress andFast-Track Career, weestimateanempiricalsensitivity of (cid:133)rst-year total CEO pay to credentials over twenty times larger than the average, a similar result also holds for Selective College. Among these top-ranked CEOs, the implied premium is the equivalent of about $600,000 for each percentile improvement in the distribution of credentials. We also document a complementary relation between pay for CEO credentials and (cid:133)rm size. For newly-appointed CEOs at (cid:133)rms in the top tercile of the size distribution, we estimate an empirical sensitivity of total pay to credentials more than double the average for Press, Fast-Track Career and Selective College. In dollar terms, this premium is the equivalent of up to $770,000 for each percentile increase. Both convexity and complementarity are consistent with our theory that predicts that more talented CEOs be matched to larger (cid:133)rms where they are more valuable. This complementarity ultimately leads to proportionally larger rewards for more talented CEOs, which Rosen (1981) coins the (cid:147)superstar e⁄ect(cid:148). We develop three main batteries of identi(cid:133)cation tests to show that our results are not biased by selectivity and endogeneity issues that arise from the non-random sorting on unobservable CEO and (cid:133)rm characteristics. In addition to dealing with measurement error, we address endogeneity issues by estimating a speci(cid:133)cation in changes, controlling for (cid:133)rm and CEO (cid:133)xed e⁄ects, and combining (cid:133)rm (cid:133)xed e⁄ectswithaninstrumentalvariable(IV)approach. Weusetheinformationcontainedinthethreeproxies jointly to address measurement error by constructing a single CEO Talent Factor as a linear combination of the three proxies.4 This factor delivers an estimated pay premium in line with our baseline estimates. Our (cid:133)rst identi(cid:133)cation test uses a speci(cid:133)cation in changes, rather than levels, which gives estimates that are very close to our baseline ones. Second, we estimate speci(cid:133)cations with (cid:133)rm (cid:133)xed e⁄ects using the entire ExecuComp panel. By looking at changes over time, these speci(cid:133)cations control for permanent unobserved characteristics of (cid:133)rms that might bias our simpler cross-sectional speci(cid:133)cation due to the initial selection of CEOs with di⁄erent credentials into (cid:133)rms that di⁄er along unobservable dimensions. We also address the potential concern that credentials are simply picking up unobservable CEO traits that are not necessarily related to talent by presenting results for speci(cid:133)cations with CEO (cid:133)xed e⁄ects that examine how pay for credentials changes in response to several industry shocks, including shocks to technology (Juhn, Murphy, and Pierce (1993), growth opportunities (Harford (2005)), organizational capital(CaroliandVanReenen(2001)), andproductmarketcompetition(Guadalupe(2007)), thatonan a priori ground we would expect should increase the returns to CEO talent. Industry shocks allow us to estimate a speci(cid:133)cation with CEO (cid:133)xed e⁄ects that examines time-series variation in the cross-sectional 4Factor loadings are derived using data for the entire ExecuComp sample. 3

estimates of pay of credentials and, thus, derive estimates of the change in the credentials premium that control for time-invariant unobservable CEO characteristics. Our (cid:133)nding of a signi(cid:133)cant premium for CEO credentials holds robustly across speci(cid:133)cations with either (cid:133)rm or CEO (cid:133)xed e⁄ects. Finally, although our speci(cid:133)cations with either (cid:133)rm or CEO (cid:133)xed e⁄ects control for time-invariant unobserved (cid:133)rm or CEO characteristics, to further corroborate the validity of our baseline estimates we address the residual endogeneity concern that time-varying (cid:133)rm characteristics, say for example productivity shocks that are unrelated to CEO talent, may be correlated with CEO credentials, thus still potentially leading to selection bias in our results. To lessen any fear that CEO credentials are correlated withtime-varyingunobservedoromittedfactors, weuseanapproachthatcombines(cid:133)rm(cid:133)xede⁄ectsand instrumental variables. IV estimates with (cid:133)rm (cid:133)xed e⁄ects insure that our source of identi(cid:133)cation is from time-serieschangesratherthanpurelycross-sectionalvariation. Wepresentresultsforthreesetsofinstruments that exploit di⁄erent sources of exogenous variation in CEO credentials: geographic instruments (see, for example, Becker, Cronqvist, and Fahlenbrach (2010)), which measure average CEO credentials for all (cid:133)rms in the state where a (cid:133)rm is headquartered; instruments that use characteristics of UK CEOs to capture exogenous variation in the characteristics of their US counterparts (see, for example, Ellison, Glaeser, and Ker (2010)); and instruments that exploit exogenous variation in the relative demand for talented CEOs across-industries, an approach that is widely-employed in the labor literature (see, for example, Katz and Murphy (1992)). Robustly across these di⁄erent sets of instruments we document evidence of a signi(cid:133)cant credentials premium, suggesting that unobserved heterogeneity is not driving our results. Overall, the (cid:133)rst part of our analysis suggests that boards rely on several CEO credentials in making compensation decisions of newly-appointed CEOs, and that more current reputational and market signals tend to be relied upon more as compared to the more lagging school ranking. In the second part of our analysis, we assess the importance of our (cid:133)ndings for the literature and validateatalentinterpretationofpayforcredentialsbyrulingoutalternativeexplanations,includingCEO lifetime experience, hype, and CEO power. We argue that there is much to learn from our analysis about fundamental issues in executive compensation. In particular, we show evidence of a rising credentials premium in CEO pay over the last two decades and argue that this (cid:133)nding o⁄ers a novel perspective over key stylized facts of the overall trend on CEO pay (see Jensen, Murphy, and Wruck (2012) for a recent detailed discussion of these well-established trends). First, we replicate in our sample the wellknown result that, even after controlling for (cid:133)rm, succession, and other CEO characteristics, there was a strong upward trend in CEO pay over the 1990s and 2000s. We then show that the upward trend was about twice as large in magnitude for CEOs at the top of the credentials ladder relative to those at 4

the bottom. Strikingly, for recently-appointed CEOs there is no signi(cid:133)cant trend among those with the lowest credentials. Thus, especially among newly appointed CEOs, a rising premium for CEO credentials can help to explain the overall trend. The rising premium does a particularly good job at explaining the overall trend among outside hires and at the very top of the distribution of pay. Finally, when we repeat the analysis by broad industry groups, we see that a rising talent premium is especially relevant for understanding the stylized developments in CEO pay for the manufacturing, services, and hi-tech sectors. Turningtoalternativestories, MurphyandZÆbojn(cid:237)k(2007)andCustodio, Ferreira, andMatos(2011) show evidence of a premium to general CEO human capital. To the extent that our baseline speci(cid:133)cation does not control for these other features of CEO human capital, a potential concern with our results is thatpayforcredentialsmaysimplybeare(cid:135)ectionofpayfor(omitted)CEOgeneralhumancapital. Using standard measures of CEO general human capital based on CEO lifetime experience (whether the new CEO previously held a CEO position, the number of di⁄erent positions held in the past by the new CEO, andthenumberofdi⁄erentindustriesthenewCEOhasworkedinthepast),weshowthatcredentialsand general experience are clearly distinct, though both important, features of CEO human capital. In fact, we replicate the results of the previous literature in our sample, as robustly across the di⁄erent controls there is a signi(cid:133)cant premium for general CEO human capital. However, controlling for this premium does not meaningfully change the relation between total CEO pay and credentials of newly-appointed CEOs, which remains positive and strongly statistically signi(cid:133)cant, with an implied sensitivity of about 0.4 in percentage terms. In addition, we show evidence consistent with a substitutes relation between credentials and general experience in pay, in that the positive relation between pay and credentials is signi(cid:133)cantly stronger for CEOs that have less work experience or less general human capital. Overall, our evidence shows that both lifetime work experience and credentials represent important, though distinct, features of CEO human capital that carry an equally signi(cid:133)cant premium in CEO pay. The work of Khurana (2002) and Malmendier and Tate (2011) might suggest that CEOs with better credentials are (cid:147)hyped up(cid:148)CEOs who initially attract boards(cid:146)attention and pay for credentials is simply an indication of temporary luck that will ultimately lead to disappointing performance. We address this concern in two ways. First, we document that the pay for credentials relation is not temporary, but insteadissustainedovertheCEO(cid:146)sentirecareer.Second,weassesswhethercredentialsbearthehallmark of hype by exploring whether they ultimately lead to subpar or superior long-term (cid:133)rm performance. We analyze a wide array of operating performance measures subsequent to CEO appointments and document that (cid:133)rms run by CEOs with superior credentials perform signi(cid:133)cantly better in the long term. 5

Our estimates of the sensitivity of operating returns to CEO credentials range between 2% and 3%, in line with the 1.7% impact of CEO deaths in Bennedsen, Perez-Gonazalez, and Wolfenzon (2008).5 Lastly, we document that CEOs with better credentials are more likely to cut expenditures, shed excess capacity, cut leverage, increase cash, and increase (cid:133)rm focus. Overall, this evidence is inconsistent with myopic, hyped-up CEOs intent on milking their (cid:133)rms, but rather consistent with a talent view of CEO credentials as initial signals used by boards to learn about CEO turnaround abilities and subsequent (cid:133)rm performance.6 Next, we consider and refute a CEO power view (see Bebchuk, Fried, and Walker (2002)) whereby credentials are proxies of CEOs(cid:146)power in setting their own pay and pay for credentials is a re(cid:135)ection of entrenchment or a combination of entrenchment and CEO connections.7 We show that our estimates of the credentials premium are robust to controlling for both internal and external (cid:133)rm governance (including the GIM index of Gompers, Ishii, and Metrick (2003), board size and independence) and for CEO education and corporate networks, and are signi(cid:133)cantly higher for (cid:133)rms with better governance and those that hire external CEOs, both inconsistent with a power story. Finally, CEOs with better credentials are subject to signi(cid:133)cantly more aggressive performance-related board monitoring, which is consistent with a talent story whereby it is more e⁄ective for boards to more closely tie the threat of dismissal to performance for more talented CEOs. This result is again inconsistent with credentials being a proxy for powerful CEOs extracting rents from captive boards. In conclusion, our paper is most closely related to recent work by Edmans, Gabaix, and Landier (2009) and others on competitive sorting models of the CEO labor market. To date, this literature has been mostly theoretical. Our contribution is to bring these models closer to the data by developing new measures of CEO credentials and documenting their empirical relation with pay. Thus, our study o⁄ers the (cid:133)rst direct empirical evidence consistent with competitive sorting models of CEO pay.8 Our evidence 5Also contrary to investors(cid:146)hype, we show that investors(cid:146)initial reaction to CEO appointment announcements predicts subsequent operating performance signi(cid:133)cantly better for CEOs with better credentials. 6WhilewecanclearlyrefuteCEOhypeasanexplanationforourresults,our(cid:133)ndingsarenotinconsistentwiththeactual evidence in Malmendier and Tate (2011). They (cid:133)nd that CEOs tend to underperform subsequent to receiving a business award. Incontrasttoourcareerandschoolingproxies,whicharewell-understoodtobe(cid:147)hard(cid:148)labormarketsignalsandfor whichthereissoundevidencethattheymatterforearningsofemployeesbelow theexecutivelevel(seeFarberandGibbons (1996) and Altonji and Pierret (2001) for evidence in the labor literature), awards are typically ex post recognitions and thus, represent (cid:147)soft(cid:148)signals which are more likely to be subject to hype issues. 7Gabaix and Landier (2008) and Edmans, Gabaix, and Landier (2009) emphasize that the relation between CEO pay and(cid:133)rmsizeisconsistentwiththetalentview. However,FrydmanandSaks(2010)(cid:133)ndthattheempiricalpay-sizerelation is actually weak prior to the 1980s even though (cid:133)rms grew at roughly the same rate from the 1980s onward. Bebchuk and Fried (2003) argue that the recent thirty years of the pay-size relation is consistent with a rent-extraction story. 8There is also a related literature that links CEO traits to pay. Graham, Li and Qiu (2009) and Coles and Li (2011) presentevidencethatCEO(cid:133)xed-e⁄ectsmatterforpay. GarveyandMilbourn(2003,2006),Milbourn(2003),andRajgopal, Shevlin,and Zamora(2006)linkCEO pay,pay-performancesensitivities,and thelackofrelativeperformanceevaluation to 6

strongly suggests that the growth in the high CEO talent market is an important factor behind recent trends in CEO pay, consistent with Murphy and ZÆbojn(cid:237)k (2007). Our evidence is complementary to recent work by Kaplan, Klebanov, and Sorensen (2011), who link several CEO traits to (cid:133)rm performance but not pay.9 Overall, our results have important implications for the recent policy debate on CEO pay andsuggestthattherelationbetweenpayandcredentialsisinfactconsistentwithoptimalcontracting. In contrast to the standard criticism of boards not prudently rewarding and monitoring CEOs, our evidence indicates that their compensation decisions are meritocratic.10 Theremainderofthepaperisorganizedasfollows. Section2outlinesasimplecompetitiveassignment model of the labor market for CEOs and develops its testable implications. Section 3 describes our new CEO succession dataset and our empirical measures of CEO credentials. Section 4 outlines our empirical strategy and presents our main results on pay for CEO credentials. Section 5 examines the implications of our (cid:133)ndings for key stylized facts of CEO pay and also considers alternative interpretations. Section 6 contains a battery of additional robustness checks and Section 7 concludes. 2 Model and Empirical Predictions In this section, we develop a simple model of the CEO labor market. Our model is based on recent work by Gabaix and Landier (2008) and Tervio (2008) and illustrates how equilibrium factors in the CEO labor market a⁄ect shareholders(cid:146)optimal CEO pay decisions. CEOs have observable productive abilities, or (cid:147)talent(cid:148), and are matched to (cid:133)rms competitively. The marginal impact of a CEO(cid:146)s talent is assumed to increase with the value of the assets under his control. The best CEOs go to the bigger (cid:133)rms, which maximizes their impact. We start with a simple benchmark case where incentive considerations do not matter and later introduce e⁄ort. Our analysis of this standard framework is aimed at developing new testable predictions for the link between CEO talent and pay that can be used to assess empirically whether boards(cid:146)pay decisions rely on CEO credentials as signals of CEO talent.11 executive characteristics such as age, wealth, and media cites. 9Baranchuk, MacDonald, and Yang (2011) add endogenous managerial e⁄ort and (cid:133)rm size to the model of Gabaix and Landier (2008) and show that their model can explain the recent increase in pay-(cid:133)rm size relation. 10Our results are silent about other aspects of the policy debate on CEO pay, such as, for example, whether the level of CEO pay is excessive in an absolute sense or relative to the pay of non-executive employees. 11See Sattinger (1979, 1993) for an earlier treatment of optimal assignment models of the labor maket and Himmelberg and Hubbard (2000) and Oyer (2004) for other models emphasizing the role of the CEO labor market. 7

2.1 Setup There is a continuum of (cid:133)rms and potential CEOs. Firms di⁄er in their size, k; and CEOs di⁄er in their productive abilities (talent), a: Let S(k) and T (a) denote the density functions of (cid:133)rms with respect to size and CEOs with respect to talent, respectively. Thus, k2S(x)dx will be the number of (cid:133)rms k1 with size between k and k . For simplicity, we assume thatRboth density functions take the Pareto 1 2 (exponential) form of T (a) = a (cid:11) and S(k) = k (cid:12), with (cid:11) 1 and (cid:12) 1. There is evidence that a (cid:0) (cid:0) (cid:21) (cid:21) Pareto distribution with coe¢ cient (cid:12) 1 (cid:133)ts the empirical (cid:133)rm size distribution well in the U.S. (Gabaix ’ and Landier (2008)). Both Gabaix and Landier and Tervio (2008) show that the key insights of our analysis generalize to a broader class of density functions for the distribution of CEO talent. The pro(cid:133)ts of a (cid:133)rm of size k that hires a CEO of ability a are given by revenues net of CEO pay: (cid:25)(a;k) = ak w(a);wherew isCEOpay. Shareholders, viatheboardofdirectors, decidewhichCEOto (cid:0) hire by maximizing pro(cid:133)ts net of CEO pay. We next derive the optimal allocation of CEO talent across (cid:133)rms and the equilibrium level of CEO pay, w (a); as implied by the assumptions of a competitive labor (cid:3) market for CEO talent and pro(cid:133)t-maximizing behavior. 2.2 Optimal Matching and CEO Pay Decisions AcompetitiveequilibriumintheCEOlabormarketconsistsofacompensationfunction,w(a);specifying the market pay of a CEO of talent a; and a matching function, k(a); specifying the size of a (cid:133)rm run by a CEO of talent a; such that shareholders of each (cid:133)rm maximize pro(cid:133)ts and the CEO labor market clears, giving each (cid:133)rm a CEO. 2.2.1 Optimal Matching In equilibrium, more talented CEOs work for larger (cid:133)rms. Technically, this competitive equilibrium is referred to as positive assortative matching. A su¢ cient condition for such matching is that CEO talent and (cid:133)rm assets are complements in that a talented CEO has a larger impact on her (cid:133)rm(cid:146)s pro(cid:133)ts when she has more assets under her control. This condition is satis(cid:133)ed in our model since the mixed partial derivative of (cid:133)rm revenues with respect to assets and CEO talent, @2(cid:25) = 1, is positive. Intuitively, if @a@k there are two (cid:133)rms with size k > k and two CEOs with talent a > a , the net surplus is higher by 1 2 1 2 putting CEO 1 at the helm of (cid:133)rm 1, and CEO 2 at the helm of (cid:133)rm 2. Formally, this is expressed as: a k +a k > a k +a k ; which always holds given that (k k )(a a ) > 0. 1 1 2 2 2 1 1 2 1 2 1 2 (cid:0) (cid:0) Since positive assortative matching is e¢ cient in our model, CEO labor market clearing delivers the 8

optimal assignment function of CEO and (cid:133)rms via k(a): In fact, the market clearing condition requires that if k is the size of a (cid:133)rm run in equilibrium by a CEO with ability a; then the number of (cid:133)rms of size greater than k has to be equal to the number of CEOs with ability greater than a: Thus, competition 1 1 in the CEO labor market implies that x (cid:12)dx = x (cid:11)dx. Using this equation, we can derive the (cid:0) (cid:0) Z Z k a 1 equilibrium matching function, k(a) = (cid:30)a 1 1(cid:0) (cid:0) (cid:11) (cid:12); where (cid:30) = (cid:12) a(cid:0)1 1 1 (cid:0) (cid:12). It is immediately clear that in (cid:0) equilibrium, (cid:133)rm size is a strictly increasing function of CEO(cid:16)talen(cid:17)t since @k(a) > 0. @a 2.2.2 Equilibrium CEO Pay Pro(cid:133)t maximization by shareholders implies that optimal CEO pay satis(cid:133)es the following FOC: @w(a) = k: @a Thus, pro(cid:133)t-maximizing shareholders trade o⁄the marginal cost (higher pay) with the marginal bene(cid:133)t (higher revenues) of hiring a more talented CEO. Combining this equation with the equilibrium matching function, k(a); allows us to derive an implicit equation for equilibrium CEO pay, @w(a) = (cid:30)a 1 1(cid:0) (cid:11) (cid:12). @a (cid:0) Integrating this with respect to CEO talent, we obtain the following equilibrium CEO pay rate (up to a constant of integration equal to the pay of the least productive CEO and with (cid:18) = (cid:30) 1 (cid:12) ) of: 2 (cid:0)(cid:11) (cid:12) (cid:0) (cid:0) w(a) = (cid:18)a 1 1(cid:0) (cid:11) (cid:12) +1 : (1) (cid:0) @w(a) Clearly, equilibrium CEO pay is a strictly increasing function of CEO talent, i.e., > 0. But, is @a equilibrium CEO pay a convex function of CEO talent, reminiscent of Rosen(cid:146)s (1981) so-called superstar e⁄ect? The answer to this question is yes. To see this, consider that given equation (1); a su¢ cient @2w(a) @k(a) condition for > 0 is that > 0; which is exactly what the e¢ cient allocation of CEO talent @a2 @a (assortativematching)implies. Thus,e¢ cientsortingintheCEOlabormarketimpliesthatmoretalented CEOsarematchedtolarger(cid:133)rmswheretheyaremorevaluable,leadingtoconvexrewardsforCEOtalent. This complementarity between CEO talent and (cid:133)rm size also leads to rewards for CEO talent that are @2w(a) larger for larger (cid:133)rms, i.e., > 0. In summary, our model makes the following testable predictions @a@k for the joint variation of CEO talent and CEO pay: Prediction T1 (Talent Premium in CEO Pay): CEOs with more productive abilities receive 9

higher total compensation. Prediction T2 (Cross-Sectional Properties of the Talent Premium): The relation between CEO pay and productive abilities is convex, in that the talent premium in increasing in talent. In addition, there is a complementarity between pay for talent and (cid:133)rm size, in that the talent premium is increasing in (cid:133)rm size. 2.2.3 Shareholder Returns An obvious question is how large is the impact of CEO talent on shareholder value? The answer to this will prove important to distinguish empirically between talent and hype explanations for our results. As in Gabaix and Landier (2008) and Tervio (2007), we study the following counterfactual. We consider a (cid:133)rm that at no additional cost can replace its current CEO with ability a with a more talented CEO 0 of ability a > a . We abstract from the additional wage cost of hiring a more talented CEO, and (cid:133)rst 1 0 focus on gross pro(cid:133)ts in order to derive an upper bound on the impact of CEO talent di⁄erences. Annual shareholder returns subsequent to CEO appointment, Ret; are given by (cid:25)(a ;k) (cid:25)(a ;k) a 1 0 1 Ret(a ;a ) = (cid:0) = : 1 0 (cid:25)(a ;k) a 0 0 Someinterestingfeaturesofthisexpressionimmediatelyobtain. First, shareholderreturnsareincreasing in the talent of the incoming CEOs given the fact that Ret 0 > 0: However, given that Ret00 = 0, we see that although it is optimal for shareholders to set convex pay, shareholder returns need not be a convex function of CEO talent. In other words, although superstar pay is consistent with shareholder maximization, shareholder returns are less sensitive to CEO talent than they are to pay. That said, our modelmakesthefollowingtestablepredictionforthejointvariationofCEOtalentand(cid:133)rmperformance: Prediction T3 (Firm Performance): Appointments of CEOs with more productive abilities are more likely to bene(cid:133)t shareholders (cid:150)that is, the impact of CEO appointments on shareholder value is more likely to be positive for relatively more talented incoming CEOs. 2.2.4 Equilibrium CEO E⁄ort In order to help distinguish empirically between talent and CEO power explanations for our results, we develop implications for board monitoring by introducing e⁄ort as in standard multitask, moral hazard models (Holmstrom and Milgrom (1992)). We assume that CEOs di⁄er not only with respect to their talent, a; but also with respect to their e⁄ort, e: E⁄ort is distributed independently from talent and E(e) 10

denotes the density functions of CEOs with respect to e⁄ort, which for simplicity we assume to take the Pareto (exponential) form of E(e) = e ": The pro(cid:133)ts of a (cid:133)rm of size k that hires a CEO of ability a (cid:0) willing to put in e⁄ort e are given by revenues net of CEO pay: (cid:25)(a;e;k) = aek w(a;e). This section (cid:0) shows that incentive devices aimed at increasing e⁄ort are more valuable to (cid:133)rms that hire more talented CEOs. Thus, we o⁄er a sorting-rationale for incentive provision. In equilibrium, it is e¢ cient for (cid:133)rms that hire more talented CEOs to make them work harder. Technically, this is again positive assortative matching. A su¢ cient condition for such is that CEO talent and e⁄ort are complements in the sense that a talented CEO has a larger impact on (cid:133)rm pro(cid:133)ts when she works harder and this is satis(cid:133)ed in our model since the mixed partial derivative of (cid:133)rm revenues with respect to CEO talent and e⁄ort, @2(cid:25) = k; is positive. For any given (cid:133)rm, if there are two CEOs with @a@e talent a > a and two possible contracts that induce e⁄ort e > e , the net surplus is higher by o⁄ering 1 2 1 2 to CEO 1 the contract that induces e⁄ort 1, and to CEO 2 the contract that induces e⁄ort 2. Formally, this is expressed as a ke +a ke > a ke +a ke ; which obtains since (e e )(a a ) > 0. 1 1 2 2 2 1 1 2 1 2 1 2 (cid:0) (cid:0) Since positive assortative matching is e¢ cient in our model, the assumption of CEO labor market clearing delivers the optimal assignment function of CEO talent and e⁄ort, e(a): In fact, the CEO labor market clearing condition requires that, if e is e⁄ort in equilibrium by a CEO with ability a; then the numberofCEOswithe⁄ortgreaterthanehastobeequaltothenumberofCEOswithabilitygreaterthan 1 1 a: Thus, competition in the CEO labor market implies that x "dx = x (cid:11)dx. Using this equation, we (cid:0) (cid:0) Z Z e a 1 can derive the equilibrium matching function, e(a) = (cid:17)a 1 1(cid:0) (cid:0) (cid:11) "; where (cid:17) = a " (cid:0) 1 1 1 (cid:0) ". Clearly, equilibrium (cid:0) e⁄ort is a strictly increasing function of CEO talent, i.e., @e(a) > 0. In t(cid:16)his se(cid:17)nse, it is e¢ cient to o⁄er @a to more talented CEOs contracts that induce higher e⁄ort. This is the case since shareholders that hire more talented CEOs also derive the most value from their e⁄ort. Thus, they bene(cid:133)t the most from an incentive provision such as performance-based dismissals. With this, our model makes the following testable prediction for the joint variation of CEO talent and CEO turnover: Prediction T4 (CEO Turnover): Boards should more aggressively monitor talented CEOs (cid:150)that is, the sensitivity of turnover to performance is increasing in CEO talent. 3 Data To assess the empirical relation between CEO pay and credentials, we construct a database of the CEO labor market that contains detailed information on CEO successions, as well as three empirical proxies 11

for CEO reputational, career, and schooling credentials at the time the initial terms of the compensation contractaresetbytheboard. Thissectiondetailshowweconstructthedatasetandthecollectionprocess for each of our variables. Details on variable de(cid:133)nitions are in Appendix C. 3.1 Selection of the CEO Successions Sample We hand-collect our CEO succession data for the universe of all (cid:133)rms in the ExecuComp from 1993 to 2005. ExecuComp contains information on the top executives of all S&P 1500 (cid:133)rms. We recognize a CEO turnover for each year in which the identi(cid:133)ed CEO changes (Parrino (1997), Huson, Parrino, and Stark (2001), and Huson, Malatesta, and Parrino (2004) use Forbes surveys; Jenter and Kanaan (2006) also use ExecuComp but only study departing CEOs for the 1993-2001 period). This gives us a (cid:133)rst sample of 2,357 candidate CEO succession events. We then search the Factiva news database in order to collect information about the circumstances around each succession. We exclude 67 successions that are directly related to a takeover and 95 successions involving interim CEOs. The (cid:133)nal sample contains 2,195 CEO succession events for a total of 20,904 (cid:133)rm-year observations. We classify each CEO turnover according to whether it was forced or voluntary and whether the incoming CEO is an insider or an outsider to the (cid:133)rm. Here we follow standard criteria in the literature that began with Parrino (1997). Departures for which the press reports state that the CEO has been (cid:133)red, forcedout, orretired/resignedduetopolicydi⁄erencesorpressureareclassi(cid:133)edasforced. Allother departures for CEOs age 60 above are classi(cid:133)ed as not forced. All departures for CEOs below age 60 are reviewed further and classi(cid:133)ed as forced if either the article does not report the reason as death, poor health, or the acceptance of another position (including the chairmanship of the board), or the article reports that the CEO is retiring, but does not announce the retirement at least six months before the succession.12 This careful classi(cid:133)cation scheme is necessary since CEOs are rarely openly (cid:133)red from their positions. We classify as outsiders those successor CEOs who had been with their (cid:133)rms for one year or less at the time of their appointments. All other new CEOs are classi(cid:133)ed as insiders. Finally, for each succession we determine exact announcement dates, which are the earliest dates of the news about incumbent CEO departure and successor CEO appointment. Table 1 presents an overview of our CEO succession dataset with descriptive statistics on total and forced turnover. Panel A summarizes successor type for each year, and Panel B contains the three subperiodscoveredbyoursample,whicharethe(cid:133)rstandsecondhalfofthe1990(cid:146)sand(cid:133)rsthalfofthe2000(cid:146)s. 12The cases classi(cid:133)ed as forced can be reclassi(cid:133)ed as voluntary if the press reports convincingly explain the departure as due to previously undisclosed personal or business reasons that are unrelated to the (cid:133)rm(cid:146)s activities. 12

We are able to give a more comprehensive picture of the CEO labor market than previous studies since our sample includes a more detailed collection and larger cross-section of (cid:133)rms than has been standard.13 These statistics suggest that the nature of the CEO labor market has changed signi(cid:133)cantly with respect to the 1970s and 1980s. Both the likelihoods that a turnover is forced and that the new CEO comes from outside the (cid:133)rm increase over time and are higher than in previous decades. These two trends are particularly evident when viewed across the sub-periods in Panel B, which (cid:133)rst shows that the frequency of forced turnover is higher in the later part of our sample. Forced turnovers represent about 22 percent of all turnovers in the 1993 to 1995 sub-period and about 27 percent in the following sub-periods, an increase of almost 25%. Irrespective of the sub-samples, forced turnovers are higher than in previous decades. For example, Huson, Parrino, and Starks (2001) report that forced turnovers represented only about 10 percent of all turnovers in the 1970(cid:146)s, and about 17% in the 1980(cid:146)s. Panel A shows that there is signi(cid:133)cant time-variation in both forced and voluntary turnover. Forced turnover (percentage of (cid:133)rms with forced CEO turnovers) is as low as 1.9% in 1993 and as high as 4.1% in 2002. These trends and the overall frequency of forced (2.8%) and voluntary (10.4%) CEO turnovers in our sample are in line with recent studies (e.g., see Huson, Parrino, and Starks (2001) who report 23.4% of forced to total turnovers for the 1989-1994 period). Panel B shows a second important trend in the CEO labor market: the percentage of outside successions increases monotonically across the three sub-periods. The increasing prevalence of (cid:133)lling CEO openings through external hires rather than through internal promotions suggests that there has been a material change in the CEO selection process in the 1990s. About 30% of the departing CEOs in the 1993 to 1995 sub-period are replaced by executives who have been employed at the (cid:133)rm for one year or less. In contrast, the frequency of outside appointments is about 40% percent in the 2000 to 2005 sub-period. Moreover, as shown in Panel A, while there is some time-variation (a peak of 41.8% in 2005 and a dip of 34.3% in 2003), the frequency of outside hires has been consistently around 40% since 1998. These (cid:133)gures are even more striking if contrasted against earlier decades. Murphy and Zabojnik (2007) and Huson, Parrino, and Starks (2001) report that during the 1970s and 1980s, outside hires accounted for only 15% to 17% of all CEO replacements, less than half as large as our (cid:133)gures since 1998. It is tempting to attribute this outsider trend to the higher incidence of forced turnovers. However, 13Studies covering earlier periods use Forbes Compensation Surveys, which roughly include S&P 500 and S&P MidCap 400 (cid:133)rms. Denis and Denis (1995) cover a sample of 908 CEO successions between 1985 and 1988. Huson, Parrino, and Starks (2001) and Huson, Malatesta, and Parrino (2004) have 1,316 and 1,344 CEO successions, respectively, between 1971 and1994. MurphyandZabojnik(2007)have2,783appointmentsbetween1970and2005,whichisalarger,butsigni(cid:133)cantly less detailed dataset than ours. 13

this is not the case since the trend holds for both voluntary and forced successions. While not reported, we (cid:133)nd that the percentage of voluntary (forced) successions in which an outsider is appointed increased from about 30 (33) percent in the (cid:133)rst sub-period to about 38 (44) percent in the last subperiod. Finally, notice that the percentage of outside hires over 2001 to 2005 in our data is higher than the 32.7% (cid:133)gure reported by Murphy and Zabojnik (2007). This is because their sample only includes S&P 500 and S&P MidCap 400 (cid:133)rms, which tend to rely more on inside hires (32.8% in our sample). 3.2 Construction of Proxies for CEO Credentials Our key explanatory variables are measures of CEO credentials that can plausibly represent publicly observablesignalsofCEOabilities. Weconstructthreemainempiricalproxiesforreputation, labormarket, and schooling credentials. The (cid:133)rst proxy, Press, is a reputational signal based on the number of articles containing the CEO(cid:146)s name and company a¢ liation that appear in the major US and global business newspapers in the calendar year prior to the CEO appointment. We expect that previous recognition by the business press should be perceived by boards as a good signal about CEO reputation. The second, Fast-Track Career, is a labor market signal based on the speed with which an executive becomes CEO. Intuitively, if the market for CEOs is at least in part meritocratic, the younger an executive is when she gets her (cid:133)rst CEO job, the more positive a signal boards should take about her productive abilities. The third, Selective College, is a schooling signal based on the selectivity of the CEO(cid:146)s undergraduate college. Based on signalling models of education (Spence (1973)), we expect attendance of more selective colleges to be a better signal about CEO abilities. We detail these measures next. Our reputational signal, Press, is intended to capture external parties(cid:146)perceptions of CEO reputation. We construct Press by counting the number of articles containing the CEO(cid:146)s name and company a¢ liation that appear in the major U.S. and global business newspapers in the calendar year prior to CEO appointment. The choice of pre-appointment press is important in order to mitigate simultaneity concerns, as well as the concern that the press count might be capturing characteristics of the current (cid:133)rm employing the CEO, rather than CEO-speci(cid:133)c characteristics. In robustness tests, we also consider anaverageoftheannualpresscountinthethreeyearspriortothetransition. Thenewspapersconsidered and the search criteria are analogous to previous studies in the literature and listed in Appendix A. Our text search uses both the CEO(cid:146)s last name and company name (e.g., Akers and International Business Machines or IBM). We include an article only once, irrespective of how many times the CEO(cid:146)s name appears in the article. We classify CEOs with larger values of press coverage as more reputable. With respect to the literature, we construct our reputation measure for a signi(cid:133)cantly larger cross- 14

section of (cid:133)rms and longer time-series.14 For robustness, we develop a novel approach to overcome two potential concerns with Press. First, not all press is necessarily good press, and thus we screen articles to only include nonnegative press coverage. To screen for each article(cid:146)s tone, we check whether it includes words with a negative connotation. Appendix A contains a list of the precise words we use. The list was compiled by randomly sampling 50 CEOs and reading articles about them. We then return to our full sample and count the number of articles containing the CEO(cid:146)s name, company a¢ liation, and any of the words with a negative connotation that appear in the major U.S. and global business newspapers. This gives us a proxy for Bad Press, which we can use to construct Press (cid:150)Bad Press. AsecondconcernisthatPress mightsimplyre(cid:135)ectcoverageofthe(cid:133)rmratherthantheCEO.Inorder to ensure that the number of articles is not merely a re(cid:135)ection of luck or characteristics of the previous employer, we again screen the tone of each article to re(cid:135)ect positive personal traits of the CEO using the word list in Appendix A. The list was also compiled by randomly sampling 50 CEOs and reading articles about them, as well as based on the CEO abilities that are shown to matter in Kaplan, Klebanov, and Sorensen (2011). Good Press is a count of the number of articles that contain the CEO(cid:146)s name, company a¢ liation, and any of these positive words. We also consider ratios of (Press - Bad Press) and Good Press to the total Press count, which measure the share of good press in total press and are more likely re(cid:135)ect a CEO(cid:146)s own reputation rather than a (cid:133)rm(cid:146)s. Our Bad and Good Press proxies are novel to the literature. The standard approach is to verify whether the Press variable is highly correlated with (Press - Bad Press) and Good Press only for a small, randomly-selected sample of CEOs. Our strategy allows us to construct the Good and Bad Press for the entire sample so as to test directly their role in the CEO labor market. Another advantage of our approach is that we can o⁄er a large sample validation of simple count measures (e.g., Press) typically used in the literature. The good news for the previous literature is that in our large sample, (Press - Bad Press) and Good Press are highly correlated (0.9 and 0.6, respectively) with Press since few negative articles apparently appear in print. Our second proxy for CEO talent, Fast-Track Career, is also novel to the literature and is intended to capture a labor market signal about CEO abilities. We conjecture that whether CEOs have a faster career path to the top might constitute a valuable signal of their abilities. If the selection process of corporate elites is meritocratic, the executive(cid:146)s age as of her (cid:133)rst CEO 14Milbourn (2003) considers all ExecuComp (cid:133)rms as we do, but only covers a six-year period (1993-1998). Rajgopal, Shevlin, and Zamora (2006)) consider a nine-year time period (1993-2001), but focus only on S&P 500 (cid:133)rms. Likely due to these di⁄erences, in our sample the median CEO gets about 7 mentions in the press in a year. This is in line with previous studies,butsomewhatlowerthanRajgopal,Shevlin,andZamora. However,whenweconsideronlytheS&P500subsample, we are closer to their median number of articles (13 in our sample vs. 11 in theirs). 15

appointment should be indicative of her talent. The intuition is that more talented executives will need to spend less time on the corporate ladder and will sooner clear the CEO hurdle. A related spin would be that the hurdle for appointing a young CEO is higher since younger executives have less experience.15 To construct our labor market signal, we collect detailed information about the complete career histories of CEOs from the following sources: (1) Dun & Bradstreet Reference Book of Corporate Managements (various years); (2) Standard & Poor(cid:146)s Register of Corporations, Directors and Executives; (3) Marquis Who(cid:146)s Who in Finance and Industry; (4) Biography Resource Center by Thomson Gale; (5) Lexis-Nexis, Factiva, and (6) various web searches. Given the evidence of higher job mobility over the last two decades, an important concern with this Fast-Track Career proxy is that it might simply captureacohort-e⁄ect, withyoungercohortsofexecutivesbeingabletogettheir(cid:133)rstCEOjobsooner, or common circumstances of the (cid:133)rst CEO job (see Malmendier, Tate, and Yan (2011) and Schoar (2007)). To address this concern, we use a cohort-adjusted version of our measure where we divide our sample of CEOs into three age cohorts and here de(cid:133)ne Fast-Track Career as the di⁄erence between age of the (cid:133)rst CEO job and median (cid:133)rst CEO job age in that age cohort. Ultimately, this re(cid:133)ned proxy classi(cid:133)es executives that got their (cid:133)rst CEOjob soonerthan other executives in theirage cohort as a more positive signal ability. Our third and (cid:133)nal proxy is a schooling signal based on CEO educational background. Using the same (cid:133)ve sources employed to collect information on career histories, we compile information on CEO academichistoriesandcollegeattendance. WeuseBarron(cid:146)sPro(cid:133)les of American Colleges (1980)rankings to sort CEOs into six groups depending on the selectivity of their undergraduate institution. Barron(cid:146)s assigns colleges to one of the following six bins: Most Competitive, Highly Competitive, Very Competitive, Competitive, Less Competitive, or Noncompetitive. Thus, our proxy is de(cid:133)ned as a numerical rank that takes values between 1 (worst) and 6 (best) depending on Barron(cid:146)s ranking of the undergraduate institution.16 We verify that our results are robust to classifying CEOs with missing college information as less selective college CEOs, since CEOs are arguably more likely to disclose their alma mater when they attended prominent colleges. Since there are no available comprehensive rankings of foreign undergraduate institutions, in our main analysis we exclude these CEOs and classify them as less selective college CEOs in robustness tests. While the schooling proxy has been used previously in the literature 15The motivation for this measure comes from the evidence by sociologists and work by Kaplan, Klebanov, and Sorensen (2011)thattheselectionprocessofcorporateelitesintheUShasbeenrelativelymeritocratic. SeealsoFriedmanandTedlow (2003) for a comprehensive review of the literature, and Capelli and Hamori (2005) for evidence. 16Thetopthreeclassi(cid:133)cationsinBarron(cid:146)s(1980)are(cid:147)MostCompetitive,(cid:148)(cid:147)HighlyCompetitive,(cid:148)and(cid:147)VeryCompetitive,(cid:148) which include 33, 52 and 104 undergraduate institutions, respectively. We were able to (cid:133)nd information on the college attended in 95 percent of the cases. 16

(see, for example, Perez-Gonzalez (2008) and Palia (2000)), our study is, to the best of our knowledge, the (cid:133)rst to employ it for a large cross-section of CEOs as a signal of CEO abilities. In summary, we use three measures of CEO credentials, based on CEO reputation, career, and educational background. An advantage of having multiple proxies is that we can validate them by checkingtheirpairwisecorrelations. PanelAofTable2displayspairwisecorrelationsamongourvariables for di⁄erent sub-samples of our dataset. The correlations are positive and all statistically signi(cid:133)cant, suggesting that indeed the variables may capture signals of CEO abilities. However, the correlations are far from one, suggesting that they likely capture di⁄erent CEO abilities and are noisy. The di⁄erence between each of our proxy variables and latent CEO abilities is measurement error.17 PanelBcontainssummarystatisticsforboththeoutgoingCEOandhersuccessor,aswellassome(cid:133)rm characteristics. These are additionally sorted by whether the departing CEO is forced out, and whether theincomingCEOisaninsideroroutsider. Particularlyforoutsidehiresandforcedsuccessions,outgoing CEOs tend to rank lower than successor CEOs in terms of our credentials measures. For example, for outside successions, the median outgoing CEO has 6 press articles (5 good articles) versus 9 articles (7 good articles) for the median outside successor and has a somewhat worse schooling record (2.4 vs. 2.9). For forced successions, the median outgoing CEO got his (cid:133)rst CEO job at age 46 and has a schooling rank of 2.6, while the median successor CEO got his (cid:133)rst CEO job at age 45 and has a schooling rank of 3.2. Moreover, among successor CEOs, outside hires have higher press coverage (9 vs. 7 articles), and were younger when they got their (cid:133)rst CEO job (48 years old vs. 50) as compared to inside hires. These di⁄erences are even larger when considering incoming CEOs after forced successions.18 Finally,PanelB.3showsthataveragestockreturnsinthe12monthsbeforeaforcedCEOturnoverare aboutnegative28%. Theaverageequally-weighted(2-SIC)industryreturnbeforeforcedturnoversisalso lower than before voluntary turnovers. This is consistent with the results in Kaplan and Minton (2008) and Jenter and Kanaan (2006) that CEO dismissals are more common in underperforming (cid:133)rms and industries. Panel B.3 also shows that our sample (cid:133)rms are relatively large, and tend to have outsiderdominated boards (65% of the directors on the median board are outsiders). However, (cid:133)rm size and governancecharacteristicsarenotstatisticallysigni(cid:133)cantlydi⁄erentfromthemedian(cid:133)rminExecuComp. 17Laterwedevelopasimpleempiricalstrategythatdirectlyaddressestheclassicproblemofnoisyproxiesandmeasurement error (see Wooldridge (2002)). 18These univariate results are consistent with Prediction T3. 17

4 Empirical Strategy and Main Findings Our research setting allows us to implement direct tests of the relation between CEO pay and credentials and the economic mechanisms behind this relation. In particular, we assess a talent interpretation, which suggests that credentials serve as valuable signals of CEO productive abilities for boards(cid:146)pay decisions. This section outlines our empirical strategy and then reports the results of the main analysis of pay for CEO credentials, the cross-sectional analysis to test for whether pay for credentials is consistent with the predictions of competitive sorting models of the CEO labor market, and identi(cid:133)cation tests to address potential biases from measurement error, endogenous selection, and unobserved (cid:133)rm and CEO heterogeneity. 4.1 Empirical Strategy Our baseline empirical speci(cid:133)cation is as follows: ln(CEO pay ) = (cid:11)+(cid:12) CEO Credentials +(cid:13) Controls +(cid:14) +" , (2) ijt it ijt t ijt (cid:3) (cid:3) where executive i works at (cid:133)rm j in year t; the dependent variable, CEO pay ; is the natural logarithm ijt of total CEO pay. In our baseline analysis, we consider only newly-appointed CEOs whose credentials are more likely to be a valuable external signal of ability since they do not yet have a performance record at the new job. In addition, appointment-year pay is closest to contractual pay set by boards at the time the initial terms of the pay packages are contracted upon, and thus represent the closest empirical counterpart to the predictions of our model.19 The key explanatory variable is CEO Credentials as proxied iteratively by Press, Fast-Track Career, and Selective College. To facilitate intuitive interpretations of the economic signi(cid:133)cance of the results, we follow Aggarwal and Samwick (1999) and construct the cumulative distribution functions (CDFs) of our proxies. In our baseline speci(cid:133)cation we include controls for (cid:133)rm, CEO, and succession characteristics, such as (cid:133)rm size, CEO age, and inside succession, that have been found to be important covariates of pay in previousstudies. Theroleof(cid:133)rmsizeintheCEOlabormarketisanimportantimplicationofcompetitive models such as ours. Previous research also suggests that CEO pay and turnover rates are a function of CEO age. Our controls also include observables that are likely to be selection variables, such as prior performance. All measures are at calendar year-end, and details on their de(cid:133)nitions are in Appendix C. 19Toaddressalternativeexplanationsofourresults,laterwecomplementthisbaselineanalysiswithestimatesofequation (2) for the entire ExecuComp, which includes years subsequent to CEO appointments. 18

Finally, all our speci(cid:133)cations include year e⁄ects and 48 (Fama-French) industry (cid:133)xed e⁄ects. We assess statistical signi(cid:133)cance using clustered standard errors adjusted for non-independence of observations by executive. Wewilluseourestimatesof(cid:12) toderiveanimplieddollarsensitivityofCEOpaytocredentials. We also consider two more inclusive speci(cid:133)cations. In one of them we address the potential concern that other (cid:133)rm characteristics that are omitted from our baseline speci(cid:133)cation may be correlated to both pay and credentials, thus confounding our inference. In order to address this concern, we saturate our baseline speci(cid:133)cation with additional (cid:133)rm-level controls for capital structure, liquidity and payout policy (leverage, dividend payout, and cash holdings), additional performance measures (Tobin(cid:146)s Q, ROA, and cash (cid:135)ow), and controls for investment and operating decisions (sales growth, R&D, and capital expenditures). We also consider a second additional speci(cid:133)cation that adds CEO pay in his prior positiontothefulllistof(cid:133)rm-levelcontrols. Byincludingthisadditionalcontrolweaddressthepotential concern that CEO pay in his prior position may also be considered a signal of CEO ability and, as such, raises the question of whether credentials are an informative signal of CEO ability over and above prior pay. In our baseline tests, estimates of pay for credentials are derived from equation (2) using ordinary least squares (OLS). However, we also address directly the potential identi(cid:133)cation issues of measurement error and imperfect proxies that arise from the fact that our credential proxies are likely to be noisy.20 In ordertoaddressthefundamentalidenti(cid:133)cationproblemthatariseswhenusingproxyvariables,wepursue a strategy aimed at combining our di⁄erent proxies to obtain more reliable estimates. In particular, we estimate the following more general model: ln(CEO pay ) = (cid:11)+(cid:12) CEO Talent +(cid:13) Controls +(cid:14) +" , (3) ijt (cid:3)it ijt t ijt (cid:3) (cid:3) where all variables are the same as in (2) except for CEO Talent , which we now treat as a latent (cid:3)it CEO talent variable. Since we do not measure CEO talent directly, we specify the following classic measurement error equation: CEO Credentials = CEO Talent +u , kit (cid:3)it kjt whereu ismeasurementerrorthatweassumeisuncorrelatedwithbothCEOTalent andControls .21 kjt (cid:3)jt ijt 20It is well known that in the presence of classic measurement error, OLS estimates will be attenuated (see Wooldridge (2002)). Black and Smith (2006) conclude that OLS estimates may actually be biased upward despite attenuation. 21Observe that by including a rich set of controls, we are likely to exacerbate the attenuation bias because the controls explain a portion of CEO Talent(cid:3)it but none of the error term (see Griliches and Hasuman (1986)). 19

We estimate this more general speci(cid:133)cation using factor analysis.22 Intuitively, factor analysis allows us to aggregate our multiple measures of credentials into a single CEO Talent or T-Factor, which is a linear combination of the underlying measures with weights chosen in such a way that leans more heavily on proxies that more accurately re(cid:135)ect latent CEO abilities. To implement the model, we (cid:133)rst derive the CEO T-Factor using our three proxies, Press, Fast-Track Career, and Selective College. After obtaining the factor loadings using data for the entire ExecuComp sample,23 we estimate equation (2) using OLS with the CEO T-Factor included as the main explanatory variable. This factor analysis approach has several advantages: it is intuitive, easy to implement, and generates a simple one-dimensional variable that ranks CEOs based on a summary measure of their credentials. Finally, there is a second important set of identi(cid:133)cation issues stemming from unobserved (cid:133)rm and CEOheterogeneitythatmaya⁄ectbothpayandourcredentialsmeasuresduetothenon-randomsorting of (cid:133)rms and CEOs. We address these issues in three distinct ways: estimating a speci(cid:133)cation in changes, controlling for (cid:133)rm and CEO (cid:133)xed e⁄ects, and combining (cid:133)rm (cid:133)xed e⁄ects with an instrumental variable (IV) approach. First, we estimate equation (2) in changes, rather than levels, as: (cid:1)ln(CEO pay ) = (cid:11)+(cid:12) (cid:1) CEO Credentials +(cid:13) (cid:1) Controls +(cid:14) +" , ijt it ijt t ijt (cid:3) (cid:3) where changes in each variable are de(cid:133)ned with respect to its respective value in the year prior to transition. For credentials, this speci(cid:133)cation considers changes between the credentials of the incoming CEO and those of the outgoing CEO. Di⁄erencing ensures that time-invariant (cid:133)rm e⁄ects are not biasing our results. Second, to address unobserved (cid:133)rm heterogeneity we estimate equation (2) with (cid:133)rm (cid:133)xed e⁄ects using the entire ExecuComp panel. By looking at changes over time, these speci(cid:133)cations control for permanent unobserved characteristics of (cid:133)rms that might bias our simpler cross-sectional speci(cid:133)cation due to the initial selection of CEOs with di⁄erent credentials into (cid:133)rms that di⁄er along unobservable dimensions. We also address the potential concern that credentials are simply picking up unobservable CEO traits that are not necessarily related to talent by analyzing how pay for credentials changes in response to several industry shocks, including shocks to technology (Juhn, Murphy, and Pierce (1993), growth opportunities (Harford (2005)), organizationalcapital (Caroliand Van Reenen (2001)), and product market competition (Guadalupe (2007)), that on an a priori ground we would expect should increase the returns to CEO talent. Industry shocks allow us to estimate a speci(cid:133)cation with CEO (cid:133)xed e⁄ects 22See Harman (1976) for details on factor analysis. Joreskog and Goldberger (1975) is an early study and Heckman, Stixrud, and Urzua (2006) and Black and Smith (2006) are recent papers using factor analysis to address measurement error. We o⁄er details on why this approach is e⁄ective in Appendix B. 23The values of the factor loading are 0.646 for Fast-Track Career, 0.638 for Press, and 0.465 for Selective College. 20

that examines time-series variation in the cross-sectional estimates of pay of credentials and, thus, derive estimates of the change in the credentials premium that control for time-invariant unobservable CEO characteristics. As it is not obvious why potential omitted variables would have a stronger systematic effect on the credentials premium across various industry groups over time, cross-industry contrasts should further limit the risk of spurious correlation. Finally, although our speci(cid:133)cations with either (cid:133)rm or CEO (cid:133)xed e⁄ects control for time-invariant unobserved (cid:133)rm or CEO characteristics, to further corroborate the validity of our baseline estimates we need to address the residual endogeneity concern that time-varying (cid:133)rm characteristics, say for example productivity shocks that are unrelated to CEO talent, may be correlated with CEO credentials, thus still potentially leading to selection bias in our results. To lessen any fear that CEO credentials are correlated withtime-varyingunobservedoromittedfactors, weuseanapproachthatcombines(cid:133)rm(cid:133)xede⁄ectsand instrumental variables. IV estimates with (cid:133)rm (cid:133)xed e⁄ects insure that our source of identi(cid:133)cation is from time-series changes rather than purely cross-sectional variation. For an instrument to be valid, it must be exogenous and satisfy the exclusion restriction. In other words, we need variables that are potentially correlated to CEO credentials (relevancy condition) but a⁄ect any given CEO(cid:146)s pay only through its e⁄ect on CEO credentials (exclusion restriction), i.e., a variables that are orthogonal to (unobserved) (cid:133)rm characteristics. We propose three sets of instrumental variables, based on three distinct sources of exogenous variation. First, we consider a set of geographic instruments (see, for example, Becker, Cronqvist, and Fahlenbrach (2010)), which measure average CEO credentials for all (cid:133)rms in the state wherea(cid:133)rmisheadquartered,excludingthose(cid:133)rmsthatareinthesame(FF-48)industrygroups. Tothe extent that changes in local factors drive the demand for CEO talent, we expect that these instruments should be correlated with any given local CEO(cid:146)s credentials, but should otherwise be unlikely to capture (cid:133)rm-speci(cid:133)c characteristics since we are excluding (cid:133)rms in the same industry. However, one may be concerned that local shocks may be correlated with industry shocks, thus making the exclusion restriction unlikely to hold. Our second set of instruments directly addresses this concern by considering (FF-48) industry-wide averages of CEO credentials calculated for (cid:133)rms that are headquarteredintheUnitedKingdom(see,forexample,Ellison,Glaeser,andKer(2010)). Thisapproach uses characteristics of UK CEOs as instruments for the characteristics of their US counterparts. The identifying assumption is that, to the extent that the same industries in the U.S. and the U.K. share common fundamental factors such as technology and barriers to entry, changes in the observed CEO credentials rankings across industries in the U.K. should be predictive of those in the U.S., but are orthogonaltoanyendogenousindustryinter-dependenciespresentintheU.S.datathatarisefromreverse 21

causality. A residual potential concern with this second set of instruments is that average CEO credentials in eachindustrymayhaveanindependente⁄ectonCEOpay,perhapsbecausetheyproxyforcompetitionfor CEO talent, and thus the exclusion restriction may again not hold. Our third and (cid:133)nal set of instruments addresses this concern by considering cross-industry variation in the relative demand for talented CEOs, an approach that is widely-employed in the labor literature (see, for example, Katz and Murphy (1992)). To capture this exogenous variation, we construct CEO labor market instruments as weighted-averages of CEO credentials among all ExecuComp (cid:133)rms in each year, with weights re(cid:135)ecting the industry-speci(cid:133)c CEO labor market share. In particular, weights are de(cid:133)ned as the share of (cid:133)rms in any given (Fama- French 48) industry group in 1990 with respect to the total number of (cid:133)rms in Compustat. If demand for CEO credentials increases (decreases) nationally in any given year, industries that employ a larger share of CEOs will experience a positive (negative) relative shock to the demand for high credentials CEOs. 4.2 Baseline Analysis of Pay for CEO Credentials We now present our main (cid:133)ndings. Before discussing regression results, we plot evidence of pay for CEO credentials for newly-appointed CEOs in Figure 1. The (cid:133)gure plots the relationship between (the logarithmof)totalpayofnewly-appointedCEOsandPress.24 Whatemergesisapatternthatisstrikingly consistentwithatalentinterpretationofboards(cid:146)payforcredentialsdecisions: therelationbetweenCEO pay and reputational credentials is (cid:135)at for relatively low credentials, and then increasing and convex, as predicted by competitive assignment models of the CEO labor market (Predictions T1 and T2). Table 3 presents results of our baseline regression analysis as well as of the two more inclusive speci(cid:133)cations with additional (cid:133)rm-level controls and CEO(cid:146)s pay in his prior position. We estimate equation (2), where the log of total dollar CEO compensation is regressed iteratively on our three measures of credentials, controlling for (cid:133)rm, CEO, and succession characteristics and include (cid:133)rm size, performance in the year prior to succession, and dummies that take the value of one, respectively, if the incoming CEO is an insider and whether the succession involves a forced departure of the outgoing CEO. All speci(cid:133)cations include year and industry (cid:133)xed e⁄ects. In Columns (1), (4), and (7), we report results for each of the three measures of credentials in this baseline speci(cid:133)cation, while results for the speci(cid:133)cation with the fuller set of (cid:133)rm-level controls are in Columns (2), (5), and (8), and results for the speci(cid:133)cation that also controls for CEO(cid:146)s pay in his prior position are in Columns (3), (6), and (9). The estimates in Table 3 show that total compensation of newly-appointed CEOs is positively and signi(cid:133)cantly associated 24Fast-Track Career and Selective College deliver qualitatively similar results. 22

with our three credentials measures, and this is the case both in the baseline speci(cid:133)cation and in those with additional controls. The magnitude of the coe¢ cient estimate for each measure is stable across speci(cid:133)cations, suggesting that CEO credentials constitute an informative signal over and above observable characteristics of the newly employing (cid:133)rm or CEO(cid:146)s pay in his prior position. Depending on which measure is used, our estimates imply an empirical sensitivity of (cid:133)rst-year total CEO pay to credentials ranging from about 0.5 for Press and Fast-Track Career to about 0.2 for Selective College. This evidence suggests that better credentials carry a pay premium for CEOs as predicted by our model. How economically important is our (cid:133)nding of pay for credentials? Our estimates imply that CEOs who are one decile higher in the distribution of credentials earn up to 5 percent higher total pay. Given our semi-log speci(cid:133)cation of (2), we can write the implied expected change in dollar compensation as: dE(CEO pay) dexp (cid:11)+(cid:12) CEO Credentials +(cid:13) Controls +(cid:14) it ijt t = f (cid:3) (cid:3) g. (4) dCEO Credentials dCEO Credentials Using our estimates in Table 3 and the average CEO pay of $5.2 million, we can calculate the dollar comparative static for going from the worst to the best of each of our credentials as: dE(CEO pay) = E(W) (cid:12) = E(W) 0:544 = $2:8M dPress (cid:3) (cid:3) dE(CEO pay) = E(W) (cid:12) = E(W) 0:459 = $2:4M dFast Track Career (cid:3) (cid:3) dE(CEO pay) = E(W) (cid:12) = E(W) 0:201 = $1:1M. dSelective College (cid:3) (cid:3) Therefore, an improvementofone decile(10%)inPress carries aninitialpaypremiumofabout$280,000, whichiscertainlyeconomicallysigni(cid:133)cant. Overall,thepositiverelationbetweenpayandCEOcredentials o⁄ersa(cid:133)rstindicationconsistentwithboards(cid:146)relyingoncredentialsassignalsofCEOtalentsincetheory predicts that total compensation should be increasing in CEO talent. Next, we further corroborate this talent interpretation of the evidence by considering our model(cid:146)s second prediction. 4.3 Cross-Sectional Variation in Pay for CEO Credentials In this section, we document key cross-sectional features of pay for CEO credentials (cid:150)convexity and complementarity with (cid:133)rm size (cid:150)and argue that they are as predicted by our model (Prediction T2). We consider a variant of our baseline framework that includes a piece-wise linear speci(cid:133)cation of the credentials measures. We use this speci(cid:133)cation to examine if pay for credentials is stronger for CEOs in the highest brackets of the empirical distribution of each of the credentials measures and for larger (cid:133)rms. 23

Table 4 presents results of our test of convexity in pay for credentials. The full set of controls are included in the estimation but unreported. In Columns (1), (4), and (7), we report results for piece-wise linearsplinesofeachofthethreemeasuresofcredentialsinthebaselinespeci(cid:133)cation, whileresultsforthe speci(cid:133)cation with the fuller set of (cid:133)rm-level controls are in Columns (2), (5), and (8), and results for the speci(cid:133)cation that also controls for CEO(cid:146)s pay in his prior position are in Columns (3), (6), and (9). The estimates in Table 4 show that the relation between total compensation of newly-appointed CEOs and each of our three credentials measures is positive and convex. Our estimates for newly-appointed CEOs whose credentials are in the top 10% imply an empirical pay-to-credentials sensitivity of more than 10 for Press and Fast-Track Career (and about 1 for above-median CEOs based on Selective College, which is a coarser variable that does not allow for a richer spline). The magnitude of these coe¢ cient estimates for any given measure is quite stable across speci(cid:133)cations. Using the same dollar comparative statics calculation as in (4), these estimates imply that for the top-decile CEOs, each percentile improvement in thecredentialsdistributioncarriesapremiumof$600,000. Incontrasttotheselargesensitivitiesatthetop of the distribution of credentials, our coe¢ cient estimates imply negligible, albeit positive, sensitivities for CEOs with poorer credentials. Taken together, this cross-sectional feature of the empirical paycredential relation is consistent with a talent interpretation from competitive sorting models predicting that compensation is increasing and convex in CEO talent a lÆ Rosen(cid:146)s (1981) (cid:147)superstar e⁄ect(cid:148)and our Prediction T2. Testing the second part of Prediction T2, Table 5 presents results of the analysis of cross-sectional variation with (cid:133)rm size. Here we use piece-wise linear versions of each of the three credentials measures interactedwithdummiesfor(cid:133)rmsizetercilestotestwhetherthereisheterogeneityintherelationbetween thetalentpremiumand(cid:133)rmsize. InColumns(1),(4),and(7),wereportresultsforinteractionsofeachof the three measures of credentials in the baseline speci(cid:133)cation, while results for the speci(cid:133)cation with the fuller set of (cid:133)rm-level controls are in Columns (2), (5), and (8), and results for the speci(cid:133)cation that also controls for CEO(cid:146)s pay in his prior position are in Columns (3), (6), and (9). The results show that the positive relation between pay and CEO credentials is signi(cid:133)cantly stronger for larger (cid:133)rms (middle and top terciles). In other words, there is a complementary relation between pay for credentials and (cid:133)rm size. Fornewly-appointedCEOsat(cid:133)rmsinthetopsizetercile,weestimateanempiricalsensitivityoftotalpay to credentials ranging from about 1 for Press and Fast-Track Career to about 0.5 for Selective College, with coe¢ cient estimates for each measure that are little changed across speci(cid:133)cations. In dollar terms, thecredentialspremiumimpliedis$77,000percredentialpercentileforCEOsrunninglarger(cid:133)rms. While still positive, the credentials premium is small and insigni(cid:133)cant for the smallest (cid:133)rms (bottom tercile). 24

This evidence suggests that better credentials carry a much higher pay premium for CEOs who run larger (cid:133)rms. This result supports a talent interpretation that boards relying on credentials as signals of productive abilities (cid:133)nd it e¢ cient for more talented CEOs to be matched to larger (cid:133)rms, leading to a complementary relation between pay for talent and (cid:133)rm size. 4.4 Identi(cid:133)cation Issues: Firm and CEO Fixed E⁄ects and Instrumental Variables (IV) Estimates This section shows that measurement error and unobserved (cid:133)rm and CEO heterogeneity are not driving our results. To address measurement error, we use the information from our three credential measures jointly, rather than iteratively, and aggregate the three proxies into a single CEO Talent Factor. To address unobserved (cid:133)rm heterogeneity, we analyze a speci(cid:133)cation in changes of pay and CEO credentials, rather than levels, that di⁄erences out (cid:133)rm e⁄ects and a speci(cid:133)cation with long-term pay for CEO credentials for the full ExecuComp that controls for time-invariant unobservable (cid:133)rm characteristics by including (cid:133)rm (cid:133)xed e⁄ects. Finally in order to address potentially time-varying unobservable (cid:133)rm characteristics, we use an instrumental variables (IV) approach. Results for these (cid:133)rst three sets of identi(cid:133)cation tests are reported in Table 6. In Columns (1) and (2), we report results for the CEO Talent Factor and our baseline speci(cid:133)cation in levels and changes, respectively, while results for the speci(cid:133)cation with (cid:133)rm (cid:133)xed e⁄ects for the entire ExecuComp are in Columns (3) and (4), and results for the instrumental variables (IV) analysis with (cid:133)rm (cid:133)xed e⁄ects are in Columns (5), (6), and (7). The bottom panel displays for each column estimated coe¢ cient for the instruments in the (cid:133)rst-stage regression and IV estimation diagnostic statistics for joint excluded instrument signi(cid:133)cance (F-test statistic) and instrument over-identi(cid:133)cation restrictions (p-values of Hansen J-statistic). The estimate for the Talent Factor in Column (1) con(cid:133)rms our main (cid:133)nding that there is a signi(cid:133)cant positive relation between pay of newly-appointed CEOs and their credentials. The sensitivity of pay for credentials decile implied by the factor estimates is about $250,000, which is in line with our baseline estimates. Also estimates in changes from Column (2) con(cid:133)rm that there is a signi(cid:133)cant pay-tocredentials sensitivity of about $220,000, suggesting that time-invariant unobserved (cid:133)rm heterogeneity is unlikely to be driving our results. The results for speci(cid:133)cations with (cid:133)rm (cid:133)xed e⁄ects in Columns (3) and (4) o⁄er additional evidence that time-invariant unobserved (cid:133)rm heterogeneity is unlikely to be driving our results. The estimates in Column (3) reveal that total CEO compensation remains positively and signi(cid:133)cantly associated with credentials throughout CEO tenure and imply a long-term sensitivity of total CEO pay to credentials of 25

about 0.29, which is economically signi(cid:133)cant and correspond to about $130,000 premium per credentials decile. Column (4) reports results for a speci(cid:133)cation that adds an interaction term between the CEO Talent Factor and CEO tenure to allow for heterogeneity in pay for credentials depending on CEO tenure. Here we see that the sensitivity of pay to credentials declines signi(cid:133)cantly over the CEO(cid:146)s tenure, consistent with our talent interpretation since presumably boards observe additional private and public signals of CEO abilities, including (cid:133)rm performance subsequent to the CEO appointment. However, the sensitivityisnotapurelytemporaryphenomenonasthecredentialspremiumremainssigni(cid:133)cantatabout $100,000 even for CEOs with above-median tenure.25 The IV estimates with (cid:133)rm (cid:133)xed e⁄ects in Columns (5), (6) and (7) suggest that time-varying unobserved (cid:133)rm heterogeneity is also unlikely to be driving our OLS estimates which may actually be downward biased by this source of endogeneity. The estimates refer to the CEO Talent Factor instrumentedinturnbythreedi⁄erentsetsofgeographic,industry-UK,andCEOlabormarketvariables,which are listed in the bottom panel with their respective (cid:133)rst-stage regression coe¢ cients. Robustly across the three di⁄erent sets of instruments, the IV estimates reveal that total CEO compensation remains positively and signi(cid:133)cantly associated with credentials and imply a long-term sensitivity of total CEO pay to credentials of at least 0.41, which is economically signi(cid:133)cant and correspond to about $220,000 premium per credentials decile. The fact that the IV estimates are somewhat larger than their OLS counterparts suggests that unobserved (cid:133)rm heterogeneity may actually lead to OLS estimates that are biaseddownwardand, thus, understatepayforcredentials. Turningtothe(cid:133)rststageregressionestimates in the bottom panel, all the instruments are positively and statistically signi(cid:133)cantly related to the Talent Factor and have strong predictive power as the large R2 suggests that the instrumental set explains a sizeable fraction of the variation in the Talent Factor thus lessening the possibility that weak instruments contaminateourinference. Anadvantageofusingmultipleinstrumentsisthattheoveridentifyingrestrictionscanbetestedusingdi⁄erentsourcesofvariationintheTalentFactor. Robustlyacrossthethreesets of instruments, the Hansen-Sargan overidenti(cid:133)cation test cannot reject the joint null hypothesis that the instruments are valid (for example, in Column (7) the Hansen J-statistic has a p-value of 0.24) and the classic F-test for the joint signi(cid:133)cance of the excluded instruments shows that they are highly signi(cid:133)cant jointly, lending further support to our choice of instruments. Results for our (cid:133)nal battery of identi(cid:133)cation tests are reported in Table 7, which shows that pay for 25The magnitude of our estimates lends support to values of approximately 1/3 that are commonly used to calibrate the empiricaldistributionofCEOtalent(e.g.,GabaixandLandier(2008)). Inunreportedresults,weuseanapproachanalogous to theirs and (cid:133)tan empiricalPareto distribution to ourcredentials proxies, which deliversestimates ofthe Pareto exponent ranging between 0.28 and 0.33. 26

credentials increases signi(cid:133)cantly in response to several industry shocks, including shocks to technology (Juhn, Murphy, and Pierce (1993), growth opportunities (Harford (2005)), organizational capital (Caroli and Van Reenen (2001)), and domestic and foreign product market competition (Guadalupe (2007)). Since theory suggests that these shocks should increase the returns to CEO talent, the evidence from industry shocks lends further support to a talent interpretation of pay for credentials. The estimates are particularly strong for shocks to organizational capital in Columns (5) and (6), for which the sensitivity of total CEO pay to credentials increases by about 0.34 on impact, which is an economically signi(cid:133)cant e⁄ect and corresponds to a cumulative dollar e⁄ect of about $320,000 higher premium per credentials decile. An additional advantage of considering industry shocks is that we estimate speci(cid:133)cations with CEO (cid:133)xed e⁄ects that controls for time-invariant unobservable CEO characteristics. As it is not obvious why potential unobserved CEO characteristics would have a stronger systematic e⁄ect on the credentials premium across various industry groups over time, the evidence of signi(cid:133)cant pay for credentials in these speci(cid:133)cations further limit the risk that credentials are simply picking up unobservable CEO traits that are unrelated to talent. 5 Assessing and Interpreting Pay for CEO Credentials Above, we document reliable evidence of a (cid:133)rst-year sensitivity of CEO pay to credentials of about 0.5, which increases for CEOs with better credentials and those who run larger (cid:133)rms. These results suggests thatboardsrelyonseveralCEOcredentialsinmakingcompensationdecisionsofnewly-appointedCEOs, andthatmorecurrentcredentials,suchasthereputationalandmarketonesaremostimportant. However, these (cid:133)ndings leave two major questions still open. First, why are the (cid:133)ndings important? In order to address this question, we assess whether our analysis o⁄ers useful insights into the key stylized facts of the recent growth in CEO pay. Second, are these (cid:133)ndings the results of a well functioning CEO labor market, or are there alternative explanations at play, such as CEO lifetime work experience, hype, CEO power and connections? A less benevolent interpretation of our (cid:133)ndings is that CEOs with apparent high abilityaresimplyexecutivesthatperhapshavemoregeneralistskills, orthosethatareinitiallyhypedup, but whose hype will fade over time as her (cid:133)rm ultimately underperforms. Alternatively, perhaps these CEOs wield their power and use their (cid:133)rms(cid:146)resources to manage their own press and milk their (cid:133)rms. Lastly, perhaps these CEOs are better connected and can extract higher rents because of their education or corporate ties. We take up each of these in turn. 27

5.1 Assessing Pay for CEO Credentials: Implications for Stylized Facts of Trend in CEO Pay Is pay for credentials an important new result? If so, how does it contribute to the literature? What is there to learn from our analysis about fundamental issues in executive compensation? In this section, we show evidence of a rising credentials premium in CEO pay over the last two decades and argue that this (cid:133)nding o⁄ers a novel perspective over key stylized facts of the overall trend on CEO pay (see Jensen, Murphy, and Wruck (2012) for a recent detailed discussion of these well-established trends). The results presented in Panels A and B of Table 8 consider these trends in turn for the entire ExecuComp sample and for a sub-sample of freshly-appointed CEOs, respectively. For any given stylized fact, we present (cid:133)rst estimates of speci(cid:133)cations with time trend indicator variables that refer to three sub-partitions of our overall time period, 1993-1995, 1996-2000, and 2001-2005. We then present results for speci(cid:133)cations that add interactions of these time dummies with our CEO Talent Factor variable, to explore di⁄erential trends depending on the level of CEO credentials. All speci(cid:133)cations include (cid:133)rm (cid:133)xed e⁄ects, as well as controls for the same set of (cid:133)rm, successions, and other CEO characteristics that are included in our baseline speci(cid:133)cation (Table 3). Estimates for the time dummies in Column (1) replicate the well-known result that, even after controlling for (cid:133)rm, succession, and other CEO characteristics, there was a strong upward trend in CEO pay over the 1990s and 2000s. Column (2) shows that the upward trend was about twice as large in magnitude for CEOs at the top of the credentials ladder relative to those at the bottom. Strikingly, looking at the results for recently-appointed CEOs in Panel B, there is no signi(cid:133)cant trend for CEOs with the lowest credentials. Thus, especially among newly appointed CEOs, a rising premium for CEO credentials can help to explain the overall trend. Column (3) and (4) show that the trend was somewhat more pronounced among outside hires and that a rising credentials premium does a particularly good job at explaining the overall trend among these CEOs. Since outside hires are those that are typically most active in the CEO labor market, this result lends further support to a labor market interpretation of our (cid:133)ndings. Columns (5) to (8) use quantile regression analysis to examine the trend at the top and a the very top of the distribution of pay. The results show that the overall trend was even more pronounced at the top and that is exactly where the rise in the credentials premium was also most pronounced. These resultsarethetime-seriescounterpartofthe"superstare⁄ect"wedocumentedinTable4andlendfurther support to Prediction T2 of our model. Finally, Columns (9) and (10) show that the upward trend was more pronounced for the equity component of CEO pay, especially among recently-appointed CEOs and that again that(cid:146)s where the credentials premium rose the most. 28

Panel C repeats the analysis by broad industry groups, with Columns (1) and (2) reporting results for the manufacturing sector, Columns (3) and (4) for retail, Columns (5) and (6) for services, Columns (7) and (8) for hi-tech sectors (such as biotech, computing, computer equipment, electronics, medical equipment, pharmaceuticals, software), and Columns (9) and (10) for regulated sectors ((cid:133)nancials and utilities). The results show that the upward trend in CEO pay holds across the board of a wide array of di⁄erent industrial sectors, though the trend in the 1990s was more pronounced in hi-tech and services, while regulated had a stronger rise in the 2000s. The rising credentials premium is not con(cid:133)ned to any one particular industry, as it holds signi(cid:133)cantly for manufacturing, services, and hi-tech. However, it appears to o⁄er less of a compelling explanation for the overall upward trend in retail and regulated industries. Overall, this evidence broadly suggests that a rising talent premium o⁄ers an important and novel perspective over key recent stylized developments in CEO pay. 5.2 Talent vs. Lifetime Work Experience: Pay for Credentials and Generalist CEO Human Capital Inthis section, weshowthatpayforCEOcredentialsisnotare(cid:135)ectionof otherimportantcharacteristics of CEO human capital that have been previously recognized in the literature, such as previous experience of the CEO and generalist vs. specialist features of his human capital. Murphy and ZÆbojn(cid:237)k (2007) and Custodio, Ferreira, and Matos (2011) show evidence that there is a trend toward appointing more generalist CEOs among publicly traded (cid:133)rms in the U.S. in the last decades. In addition, these papers present evidence of a premium to generalist CEO human capital. To the extent that our baseline speci(cid:133)cation does not control for these other features of CEO human capital, a potential concern with our results is that pay for credentials may simply be a re(cid:135)ection of pay for (omitted) CEO general human capital. The results in Table 9 show that pay for credentials and generalist experience are clearly distinct, though both important, features of CEO human capital. Columns (1) to (3) present estimates for a speci(cid:133)cation that adds controls for standard measures of CEO general human capital based on CEO lifetime experience: whether the new CEO previously held a CEO position, the number of di⁄erent positions held in the past by the new CEO, and the number of di⁄erent industries the new CEO has worked in the past. Column (4) shows results when we control for a measure that aggregates these lifetime experience variables into a CEO General Ability Factor extracted using principal component analysis from the three underlying experience proxies as in Custodio, Ferreira, and Matos (2011)). Here we see that we can replicate the results of the previous literature in our sample, as robustly across the di⁄erent controls there is a signi(cid:133)cant premium for general CEO human capital. However, controlling 29

for this premium does not meaningfully change the relation between total CEO pay and credentials of newly-appointed CEOs, which remains positive and statistically signi(cid:133)cant, with an implied sensitivity of about 0.4 in percentage terms. These estimates of the credentials premium are a bit lower but little changed in therms of their economic signi(cid:133)cance with respect to a speci(cid:133)cation without CEO lifetime experience controls (Column (4) of Table 6). Columns (5) to (7) o⁄er additional analysis of the relation between pay for credentials and pay for general human capital. Here, rather than taking CEO credentials and CEO lifetime work experience as two separate groups of variables, we present results for speci(cid:133)cations that includes two CEO Human Capital Factors, "Experience" and "Talent," which are the (cid:133)rst two principal components extracted from using our three CEO credentials proxies jointly with the three CEO lifetime work experience proxies. The fact that factor analysis gives us two orthogonal principal components, one of which is more highly correlatedwiththeexperienceproxiesandtheotherwhichismorecorrelatedwiththecredentialsproxies, o⁄ers additional evidence supporting the notion that credentials and work experience pick up di⁄erent characteristics of CEO human capital. Estimates in Column (5) show that both the "Experience" and the "Talent" factors are signi(cid:133)cantly positively associated with total CEO pay, suggesting that there is both a CEO credentials premium and a CEO general human capital premium in pay. In addition, Columns (6) and (7) show evidence consistent with a substitutes relation between credentials and general experience in pay. Here we consider interactions between the two CEO Human Capital Factors to allow for heterogeneity in pay for CEO credentials depending on CEO experience and viceversa. We (cid:133)nd that the positive relation between pay and credentials is signi(cid:133)cantly stronger for CEOs that have less work experienceorlessgeneralhumancapital. Viceversa,thepremiumtogeneralhumancapitalissigni(cid:133)cantly higher for CEOs with less credentials. This evidence suggests that boards(cid:146)pay decisions load relatively more heavily on credentials when hiring CEOs with shorter work histories, which presumably o⁄er fewer other observable signals of CEO ability. Overall, based on this evidence we conclude that both lifetime work experience and credentials represent important, though distinct, features of CEO human capital and both carry an equally signi(cid:133)cant premium in CEO pay. 5.3 Talent vs. Hype: Pay for Long-Term Credentials, Firm Performance and Corporate Policies In this section, we use the predictions of our competitive sorting model to distinguish between interpretations based on talent versus those based on hype. While a talent interpretation considers CEO credentials valuable signals of CEO abilities, the hype view (Khurana (2002) and Malmendier and Tate 30

(2011))wouldconsiderCEOswithbettercredentialsascharismatic,(cid:147)hypedup(cid:148)CEOswhoattractattentioninitially,butsubsequentlyunderwhelm. Ifcredentialsareanindicationoftemporaryhype,weshould see disappointing subsequent performance and a disappearing pay-for-credentials premium. By contrast, if credentials are signals of productive abilities, premium pay for credentials should remain signi(cid:133)cant in the long-run and be associated with superior long-term operating performance (see Prediction T3). Examining long-term pay for credentials and the relation between credentials and long-term operating (cid:133)rm performance allows us to distinguish between the two alternative stories. Overall, long-termfeaturesofpayforcredentialsinTable6appearmoreconsistentwithatalentstory of boards learning from multiple signals of CEO abilities rather than being the decision of passive boards hypnotized by CEO hype. There we saw that the sensitivity of pay to credentials declines signi(cid:133)cantly over the CEO(cid:146)s tenure, but it is not a purely temporary phenomenon as the hype story predicts. Before presenting the results of our formal tests of the relation between credentials and long-term (cid:133)rm operating performance, we plot univariate evidence in Figure 2. The (cid:133)gure plots sample median OROA over the period from four years before to four years after CEO succession for our entire succession sample. The dotted line represents median OROA for the entire sample, while the bold line represents median OROA for new CEOs with better reputational credentials (top quartile of Press),26 and the thin line represent median OROA for bottom-quartile CEOs. The OROA (cid:147)smile(cid:148)suggests that, on average, CEO turnover follows a period of deteriorating (cid:133)rm performance which tends to be reversed subsequently. A striking feature that emerges is that the smile is an artifact of averaging out performance in a sample that pools CEOs with good credentials together with relatively less accomplished ones. Panel A of Table 10 presents results of our regression analysis of long-term operating (cid:133)rm performance. We estimate a version of equation (2) where now the dependent variables are changes around CEO successions in various industry-adjusted measures of long-term operating (cid:133)rm performance. The changes in these measures are regressed on the CEO Talent factor and controls. In order to control for mean-reversion, we include in all speci(cid:133)cations prior performance measured as average annual performance in the three years prior to transition. In Columns (1), we examine short-run cumulative abnormal returns (CARs) around CEO appointments and see that investors anticipate subsequent performance improvements, which corresponds to them reacting more favorably to the news of successions that involve incoming CEOs with better credentials. Columns (2)-(7) report our main results, with long-term operating performance measured by net income to assets (ROA), operating return on assets (OROA), operating 26We uncover qualitatively similar results using Fast-Track Career and Selective College, as well as when we measure performance using OROS and ROA. 31

return on sales (OROS), return on equity (ROE), stock market returns, and cash (cid:135)ows, respectively. For every performance measure, we uncover estimates of the sensitivity of shareholder returns to CEO credentials that are positive and strongly statistically signi(cid:133)cant, ranging between 2% and 3%.27 Finally, Column (8) examines ROA in a speci(cid:133)cation that adds appointment CARs and an interaction term between them and the CEO Talent Factor (estimate of the interaction term reported) to allow for heterogeneity in the predictive power of short-term CARs depending on CEO credentials. Here we see that investors(cid:146)reaction is a better predictor of subsequent long-term performance for CEOs with better credentials. ThelatterresultisinconsistentwithinvestorsoverreactingtotheappointmentofaCEOwith better credentials and suggests that credentials are in fact an informative signal of future performance. Overall, our estimates of the credentials premium for shareholder returns are consistent with models of competitive sorting in the CEO market (Prediction T3), rather than CEO hype which predict that the performance impact of CEO talent should be an order of magnitude smaller than the pay impact. To buttresstheseperformanceresults,PanelBofTable10presentsresultsofourregressionanalysisofactual CEO decisions. We estimate a version of equation (2), where now the dependent variables are changes aroundCEOsuccessionsinvariousindustry-adjusted(cid:133)rmpolicies,whichareregressedontheCEOTalent factor and our standard controls. We report results on investment policy in Columns (1)-(3), (cid:133)nancial policy in Columns (4)-(6), and on organizational strategy in Columns (7) and (8). Our estimates show that CEOs with better credentials are signi(cid:133)cantly more likely to cut capital and M&A expenditures, shed excess-capacity (existing divisions), cut leverage and increase internal (cid:133)nancing (cash), and increase (cid:133)rm focus. Overall, this evidence is inconsistent with myopic, hyped-up CEOs intent on milking their (cid:133)rms, and instead consistent with a talent view that credentials are signals of CEO turnaround abilities re(cid:135)ected in long-term performance. 5.4 Talent vs. CEO Power: Pay for Credentials, CEO Connections, and Firm Governance In this section, we use the predictions of competitive sorting models to distinguish between a talent interpretation and one based on CEO power (Bebchuk and Fried (2003)). If credentials are proxies for CEO power in setting their own pay, then pay for credentials is actually a re(cid:135)ection of entrenchment issues and thereby we should see signi(cid:133)cantly higher premiums for (cid:133)rms with worse governance and even more so if their CEOs are more connected (e.g., Fracassi and Tate (2011)). Also, if better credentials proxy for power, then we should see weaker board monitoring of these CEOs. By contrast, if credentials 27Our estimates are in line with the 1.7% impact of CEO deaths in Bennedsen, Perez-Gonazalez, and Wolfenzon (2008). 32

are signals of productive abilities, we should see higher premiums at better governed (cid:133)rms to go along with the better (cid:133)rm performance documented above. In addition, Prediction T4 suggests that we should see tougher board monitoring of CEOs with better credentials. Columns(1)-(6)ofTable11presentsresultsofouranalysisoftheimpactof(cid:133)rmgovernanceandCEO networks on pay for credentials. Column (1) presents estimates for a speci(cid:133)cation that adds controls for standard measures of (cid:133)rm governance, the GIM Index, board size, and board independence, and Column (2) shows a speci(cid:133)cation that also adds controls for standard measures of CEO networks, the intensity of CEOeducationandcorporateties. HereweseethattherelationbetweentotalCEOpayandcredentialsof newly-appointed CEOs remains positive and statistically signi(cid:133)cant after controlling for (cid:133)rm governance and CEO connections, with an implied sensitivity of about 0.5 in percentage terms. These estimates are little changed with respect to a speci(cid:133)cation without governance and CEO connections controls (Column (4) of Table 6). Columns (3)-(6) individually add interactions between the CEO Talent Factor and the three governance variables (Columns (3), (5), and (6)) to allow for heterogeneity in pay for CEO credentialsdependingonthequalityof(cid:133)rmgovernance,aswellasinteractionsbetweentheTalentFactor, the GIM index, and CEO connections to explore whether governance issues have a di⁄erential impact on pay for credentials depending on CEO networks, since the evidence in Fracassi and Tate (2011) suggests that governance issues are particularly important for (cid:133)rms whose CEOs are well-connected. We (cid:133)nd that thepositiverelationbetweenpayandcredentialsissigni(cid:133)cantlystrongerfor(cid:133)rmswithbettergovernance andforexternally-hiredCEOswhichareobviouslytheleastlikelytobeentrenched. Inaddition,wedonot (cid:133)nd any evidence of stronger e⁄ects of governance on pay for credentials depending on CEO connections. Overall, these results are inconsistent with an entrenchment view of more accomplished CEOs. Columns (7) and (8) present results of the relation between credentials and board monitoring. All speci(cid:133)cations are for probit regressions of the likelihood of forced CEO turnover on measures of CEO credentials for the entire ExecuComp, where the dependent variable is a dummy that takes value of one in any given (cid:133)rm-year when a forced CEO turnover occurs.28 We present estimates for two di⁄erent subsamples of underperforming (cid:133)rms, which are de(cid:133)ned as (cid:133)rms whose performance in the prior year was below median (Column (7)) and in the bottom quintile (Column (8)) of performance in their industry, respectively. CEOs with better credentials are subject to signi(cid:133)cantly more aggressive board monitoring asmeasuredbythelikelihoodofbeing(cid:133)rediftheyunderperform,ane⁄ectthatinterestinglyismonotonic 28Werunastandardcross-sectionalprobitregression(e.g.,JenterandKanaan(2006)): Prob(Forced CEO Turnover )= jt (cid:11)+(cid:12) Firm Return +(cid:12) Firm Return CEO Credentials +(cid:12) Firm Return Controls +(cid:12) CEO 1 (cid:3) jt 2 (cid:3) jt (cid:3) jt 3 (cid:3) jt (cid:3) jt 4 (cid:3) Credentials +(cid:12) Controls +" ; where Controls include (cid:133)rm size, CEO age, tenure, and insider dummy, and all jt 5(cid:3) jt jt jt speci(cid:133)cations include year and (Fama French 48) industry dummies. 33

inthestrengthofunderperformance. Thisresultisinconsistentwithcredentialsbeingaproxyforpowerful CEOs who extract higher rents from captive boards, and consistent with a talent story whereby tying the threat of dismissal more closely to performance is more e⁄ective for more talented CEOs (Prediction T4 of our model). In summary, the evidence in Table 11 is inconsistent with a power interpretation and more in line with our CEO labor market view of pay for credentials. 6 Additional Robustness Checks We conduct several additional tests to con(cid:133)rm that our main result is robust. In particular, we o⁄er additional evidence that selection issues are unlikely to be driving our results and implement robustness checks for each of the credentials measures used in our baseline regression analysis in Table 3. 6.1 Matched Sample and Heckman Analyses We address two additional selection concerns. First, a selection story would attribute pay for credentials to the ability of CEOs with better credentials to (cid:147)cherry pick(cid:148)prospective (cid:133)rms that are easier to turn around. Cherry picking is indicative of a broader range of issues related to selection on observable (cid:133)rm characteristics that arise due to the non-random assignment of CEOs to (cid:133)rms. Economically, this selection issue re(cid:135)ects the endogeneity of CEO succession decisions. For example, since large (cid:133)rms are more likely to hire talented CEOs based on our model, it might be that part of the credentials premium is simply due to CEOs with better credentials being appointed to run larger (cid:133)rms. Panel 1.A of Table 12 presents results of a matched-sample analysis that addresses this (cid:133)rst selection concern. Here, we use a nearest-neighbor matching estimator (Abadie and Imbens (2007)). Ideally, we would like to compare CEO pay of a (cid:133)rm that appoints a CEO with good credentials to the same (cid:133)rm(cid:146)s pay had it appointed a CEO with worse credentials. Since the counterfactual is not observed, we construct a hypothetical one by estimating a (cid:133)rst-stage probit regression of the likelihood that a (cid:133)rm appoints a CEO with good credentials (top quartile of the CEO Talent Factor) using a speci(cid:133)cation that includes observable pre-transition (cid:133)rm characteristics (size, performance, and forced turnover) related to cherry picking. First-stage estimation results are reported in Column 2. There is a signi(cid:133)cant and positive relation betweenthelikelihoodofappointingaCEOwithgoodcredentialsand(cid:133)rmsize. Forcedturnoversarealso more likely to be associated with subsequent appointments of CEOs with better credentials. By contrast, controllingforthesevariables,we(cid:133)ndanegativebutstatisticallyinsigni(cid:133)cantrelationwithpre-transition (cid:133)rmperformanceandthelikelihoodofappointingatalentedCEO.Column1reportsresultsofthesecond 34

stage, where we take the di⁄erence between total CEO pay for successions involving CEOs with good credentials(thetreatedgroup)andmatchedsuccessionswiththeclosestpredictedprobabilityofinvolving CEOs with good credentials (the control group). We estimate a pay-credential sensitivity of 0.6, which remainssigni(cid:133)cantandinlinewithourbaselineresults, suggestingthattheendogeneityofCEOselection is unlikely to be driving our main (cid:133)nding. Panel 1.B of Table 12 addresses a second selection concern that our baseline estimates for newlyappointed CEOs may be driven by the non-random selection of (cid:133)rms into the CEO appointment sample. Since (cid:133)rm characteristics, such as size and performance, are signi(cid:133)cant determinants of the likelihood of a CEO succession, our sample is clearly not randomly selected from the ExecuComp population and thereby our previous estimates may su⁄er from sample selection bias. We address this issue using a standard Heckman (1979) selection approach that estimates pay for CEO credentials jointly in a system of two equations that adds a probit regression of CEO succession likelihood for the entire ExecuComp sample. The (cid:133)rst-stage selection equation includes an indicator variable for CEO death or retirement, which clearly should a⁄ect the likelihood of a succession but not the subsequent pay of the new CEO, and is thus excluded from the second-stage. Using a standard two-step procedure based on the probit estimates in Column 3, we construct estimated inverse Mills ratios and use them to augment our baseline pay equation (2) in the second step. The standard errors in the second stage regression are corrected for the fact that the inverse Mills ratio is estimated (Wooldridge (2002)). Column 4 reports results of the (cid:133)rst-stage probit regression. Not surprisingly, (cid:133)rms whose CEO died recently or reached "retirement" age are signi(cid:133)cantly more likely to experience a CEO succession, and so arelargerandunderperforming(cid:133)rms. Column3reportsresultsfortheHeckmantwo-stepselectionmodel of total CEO pay. The inverse Mills ratio has a signi(cid:133)cant positive coe¢ cient, con(cid:133)rming that sample selection is a relevant concern in our study and tends to increase pay. However, even after controlling for the inverse Mills ratio, there is a positive and signi(cid:133)cant relation between pay and CEO credentials. Finally, the two-step procedure leads estimates of the sensitivity of pay for credentials that are a bit larger than our OLS ones (Column 4 of Table 6). Thus, non-random selection of the CEO succession sample is unlikely to be driving our main (cid:133)nding. 6.2 Additional Controls and Di⁄erent De(cid:133)nitions of the CEO Credentials Proxies Turning to Panel 2 of Table 12, the results in Rows (1)-(4) address the potential concern that Press might capture variation unrelated to CEO reputation, such as bad press or simply coverage of the (cid:133)rm. We show that our results are robust to using a measure that nets out negative press coverage, or Bad 35

Press, from Press (Row (1)). A second concern is that the article count might simply re(cid:135)ect luck or characteristics of the (cid:133)rm that previously employed the CEO, which we address by screening the tone of each article to re(cid:135)ect positive personal traits of the CEO based on Kaplan, Klebanov, and Sorensen (2011) and only count articles that contain mention of such traits, or Good Press (Row (3)). Notably, the sensitivity of pay to this re(cid:133)ned measure or reputation is even larger than our baseline estimate for the total press count. Next, we show that our results are robust to using (Press - Bad Press)/Press (Row (2)) and Good Press/Press (Row (4)). These ratios measure the share of good press out of total press and more likely re(cid:135)ect CEO personal reputation rather than (cid:133)rm characteristics. We also address the concern that Press may re(cid:135)ect (cid:133)rm size, by showing robustness to a (cid:133)rm-adjusted Press measure that subtracts from the total Press count for each CEO the median Press of CEOs at (cid:133)rms with similar size (Row (6)). Finally, Row (5) shows robustness to using an average of Press in the three years prior to appointment. Row (7) addresses the concern that Fast-Track Career is mechanically correlated with age for CEOs whose current appointment is also their (cid:133)rst CEO job (797 successions). Excluding these CEOs only strengthens our results. Rows (8) and (9) show that our sensitivity estimates for Selective College are robust to using a dummy approach that only classi(cid:133)es as selective those colleges that are in the top Barron(cid:146)s rank and to including CEOs that did not attend college or attended a foreign institutions as least selective, as done by Perez-Gonzalez (2006). In the last battery of checks, we show that our baseline estimates for each of the three credentials proxies are robust to using industry-adjusted measures (Row (10)) to address the concern that there may be common industry factors correlated with our proxies. We also show that the estimates are robust to controlling for graduate education using a dummy for whether CEOs have an MBA (Row (11)), which addresses the standard (cid:133)nding that MBA education is related to pay (Murphy and ZÆbojn(cid:237)k (2007), Frydman (2005)). Finally, we show that our baseline estimates are robust to controlling for size in a less parametric way which includes polynomials up to the 3rd order of the size variable (Row (12)) and to including controls for (cid:133)rms(cid:146)headquarter location to address the potential concern that local CEO labor market factors ma be driving our results (Raw (13)). 7 Conclusion This paper argues that focusing on the labor market for CEOs can augment our understanding of the empirical determinants of top executive pay. To that end, we have documented reliable evidence of pay for several CEO credentials, which include reputational, career, and educational ones. We have shown 36

that the credentials premium is larger for the most accomplished CEOs and for larger (cid:133)rms, which is consistent with competitive sorting models of the market for CEOs. Finally, the premium remains signi(cid:133)cant in years subsequent to appointment, is robust to controlling for (cid:133)rm and CEO (cid:133)xed e⁄ects as well as using an instrumental variable (IV) approach to address endogeneity, and is larger for (cid:133)rms with better governance. In addition, credentials carry a signi(cid:133)cant performance premium for shareholders. Overall, these results strongly support an interpretation of pay for credentials based on the market for CEO talent and are inconsistent with alternative stories based on CEO lifetime experience, hype, or entrenchment. In sum, our work represents the (cid:133)rst direct evidence that sorting considerations in the CEO labor market are an important determinant of CEO pay. Our results have important implications for the recent debate on the rise in CEO pay and suggest that a rising CEO talent premium may have contributed to the recent rise in CEO pay. There are, of course, other important aspects of the policy debate on CEO pay about which our results are silent. For example, some have decried the level of CEO pay as being excessive in an absolute sense or relative to the pay of non-executive employees. An interestingavenueforfutureresearchwouldbetoexploretheseissuebyconsideringtheinterplaybetween credentials and di⁄erences in responsibility along the corporate hierarchy. 37

References [1] Abadie, A. and G. W. Imbens, 2007, (cid:147)Bias Corrected Matching Estimators for Average Treatment E⁄ects,(cid:148) Mimeo, 2007. [2] Aggarwal, R. and A. Samwick, 1999, (cid:147)The Other Side of the Trade-O⁄: The Impact of Risk on Executive Compensation,(cid:148)Journal of Political Economy 107, 65-105. [3] Altonji, J.J. and C.R. Pierret, 2001, (cid:147)Employer Learning and Statistical Discrimination,(cid:148)Quarterly Journal of Economics, 116, pp.313-350. [4] Becker, B., H. Cronqvist, and R. Fahlenbrach, 2010, "Estimating the E⁄ects of Large Shareholders Using a Geographic Instrument," forthcoming, Journal of Financial and Quantitative Analysis. [5] Baranchuk,N.,G.MadDonald,andJ.Yang,2011,(cid:147)TheEconomicsofSuperManagers(cid:148),forthcomingReview of Financial Studies. [6] Barber, B.M., and J.D. Lyon, 1996, (cid:147)Detecting Abnormal Operating Performance: The Empirical Power and Speci(cid:133)cation of Test Statistics,(cid:148)Journal of Financial Economics 41, pp. 359(cid:150)399. [7] Barron(cid:146)s Educational Series, Inc. Barron(cid:146)s Pro(cid:133)les of American Colleges. Great Neck, N.Y: 1980. [8] Bebchuk,L.A.andJ.M.Fried,2003,(cid:147)ExecutiveCompensationasanAgencyProblem,(cid:148)JournalofEconomic Perspectives 17(3), pp.71-91. [9] Bebchuk, L. A., J. M. Fried and D. I. Walker, 2002, (cid:147)Managerial Power and Rent Extraction in the Design of Executive Compensation,(cid:148)University of Chicago Law Review, 69, pp.751-846. [10] Bennedsen, M., F. Perez-Gonzalez, and D. Wolfenzon, 2008, (cid:147)Do CEOs Matter?(cid:148)mimeo, NYU Stern. [11] Black, D.A., and J.A. Smith, 2006, (cid:147)Estimating the Returns to College Quality with Multiple Proxies for Quality,(cid:148)Journal of Labor Economics, 24(3), pp.701-728. [12] Capelli, P. and M. Hamori, 2005, (cid:147)The New Road to the Top(cid:148), Harvard Business Review 83-1, 25-32. [13] Caroli, E. and J. van Reenen, 2001, "Skill-Biased Organizational Change?" Quarterly Journal of Economics, 116(4), pp.1449-92. [14] Coles, J.L., and Z.F. Li, 2001, (cid:147)Managerial Attributes, Incentives, and Performance,(cid:148)mimeo, Arizona State University. [15] Custodio, C., M. A. Ferreira, and P. Matos, 2011, "Generalists vs. Specialists: Lifetime Work Experience and CEO Pay," mimeo, Arizona State University [16] Denis, David J. and Diane K. Denis, 1995, (cid:147)Firm Performance Changes Following Top Management Dismissals,(cid:148)Journal of Finance, 50, pp.1029-1057. [17] Dun & Bradstreet, Inc. Dun & Bradstreet Reference Book of Corporate Managements. New York: various years. [18] Edmans, A., G.Xavier, andA.Landier, 2009, (cid:147)AMultiplicativeModelofOptimalCEOIncentivesinMarket Equilibrium ,(cid:148)Review of Financial Studies, 22(12), pp. 4881-4917. [19] Ellison, G, E.L. Glaeser, and W. Kerr, 2010, "What Causes Industry Agglomeration?" American Economic Review, 100(3), pp. 1195-1213. 38

[20] Fama, E.F. and K.R. French, 1993, (cid:147)Common risk factors in the returns on stocks and bonds(cid:148), Journal of Financial Economics 33, 3-56. [21] Farber,Henry,S.andRobert,Gibbons,1996,(cid:147)LearningandWageDynamics,(cid:148)QuarterlyJournalofEconomic, 111, pp.1007-1047. [22] Feenstra, R. C., J. Romalis, and P. K. Schott, 2002, "U.S. Imports, Exports and Tari⁄Data, 1989 to 2001," NBER Working Paper 9387. [23] Fracassi,C.andG.Tate,2012,"ExternalNetworkingandInternalFirmGovernance,"TheJournalofFinance 67(1), pp. 153-94. [24] Friedman,W.A.andR.S.Tedlow,2003,(cid:147)StatisticalPortraitsofAmericanBusinessElites: AReviewEssay,(cid:148) Business History 45 (October): 89-113 [25] Frydman,C.,2007,(cid:147)RisingThroughtheRanks: TheEvolutionoftheMarketforCorporateExecutives,1936- 2003,(cid:148)working paper. [26] Frydman, C. and R. Saks, 2010, (cid:147)Executive Compensation: A New View from a Long-Run Perspective, 1936-2005,(cid:148)Review of Financial Studies 23-5, 2099-2138. [27] Gabaix, X., andA.Landier, 2008, (cid:147)WhyhasCEOPayIncreasedsoMuch?(cid:148)QuarterlyJournalofEconomics, 123(1), pp. 49-100. [28] Garvey, G. and T. Milbourn, 2003, (cid:147)Incentive Compensation When Executives Can Hedge the Market: Evidence of Relative Performance Evaluation in the Cross Section,(cid:148)Journal of Finance 58, 1557-1581. [29] Garvey, G. and T. Milbourn, 2006, (cid:147)Asymmetric Benchmarking in Compensation: Executives are Rewarded for Good Luck but not Penalized for Bad,(cid:148)Journal of Financial Economics 82, 197-225. [30] Gompers,P.A.,J.L.IshiiandA.Metrick,2003,(cid:147)CorporateGovernanceandStockPrices,(cid:148)QuarterlyJournal of Economics (forthcoming). [31] Graham,JohnR.,SiLi,andJiapingQiu,2009,(cid:147)ManagerialAttributesandExecutiveCompensation,(cid:148)mimeo, Duke University. [32] Griliches, Z., and J. Hausman, 1986, (cid:147)Error in Variables in Panel Data,(cid:148)Journal of Econometrics, 36(1), pp.93-118. [33] Guadalupe,M.,2007,"ProductMarketComeptition,ReturnstoSkill,andWageInequality,"JournalofLabor Economics, 25(3), pp. 439-74. [34] Harford, J, 2005, (cid:148)What Drives Merger Waves?(cid:148)Journal of Financial Economics 77, pp.529-560. [35] Harman, H., 1976, Modern Factor Analysis, Chicago University Press. [36] Heckman, J., J. Stixrud, and S. Urzua, 2006, (cid:147)The E⁄ects of Cognitive and Noncognitive Abilities on Labor Market Outcomes and Social Behavior,(cid:148)Journal of Labor Economics, 24(3), pp.411-482. [37] Himmelberg, C. P., and R. G. Hubbard, 2000, (cid:147)Incentive Pay and the Market for CEOs: An Analysis of Pay-for-Performance Sensitivity,(cid:148)mimeo, Columbia University. [38] Holmstrom, B. and Milgrom, P., 1992, (cid:147)Multitask Principal Agent Analysis(cid:151)Incentive Contracts, Asset Ownership, and Job Design,(cid:148)Journal of Law Economics and Organization, 7, pp. 24(cid:150)52. 39

[39] Huson, M., P. H. Malatesta, and Parrino, R, 2004, (cid:147)Managerial Succession and Firm Performance,(cid:148)Journal of Financial Economics 74, pp.237-275. [40] Huson, M., Parrino, R., and Starks, L., 2001, (cid:147)Internal Monitoring Mechanisms and CEO Turnover: A Long Term Perspective,(cid:148)Journal of Finance, 56, pp. 2265-2297. [41] Jensen,M.C.,K.J.Murphy,andE.G.Wruck,2012,CEOPayandWhattoDoAboutit: RestoringIntegrity tobothExecutiveCompensationandCapital-MarketRelations,forthcoming,HarvardBusinessSchoolPress. [42] Jenter, D. and F. Kanaan, 2006, (cid:147)CEO Turnover and Relative Performance Evaluation,(cid:148)MIT WP 4594-06 [43] Joreskog, K.G., and A.S. Goldberger, 1975, (cid:147)Estimation of a Model with Multiple Indicators and Multiple Causes of a Single Latent Variable,(cid:148)Journal of theAmerican Statistical Association, 70(351), pp.631-639. [44] Juhn, C., K. Murphy, and B. Pierce, 1993, "Wage Inequality and the Rise in Returns to Skills," Journal of Political Economy, 101(3), pp.410-42. [45] Kaplan, Steven N., Klebanov, Mark M., and Sorensen, Morten, 2011, "Which CEO Characteristics and Abilities Matter?(cid:148)forthcoming, Journal of Finance. [46] Kaplan, S. and A. Minton, 208, (cid:147)How Has CEO Turnover Changed? Increasingly Performance Sensitive Boards and Increasingly Uneasy CEOs(cid:148)mimeo, University of Chicago. [47] Kaplan, S., andJ.D.Rauh, 2010, (cid:147)WallStreetandMainStreet: WhatContributestotheRiseintheHighest Incomes? (cid:148)forthcoming, Review of Financial Studies. [48] Katz, L. F. and K. L. Murphy, 1992, "Changes in Relative Wages, 1963-1987: Supply and Demand Factors," Quarterly journal of Economics, 107(1), pp.35-78. [49] Khurana, R. 2002, Searching for a Corporate Savior: The Irrational Quest for Charismatic CEOs, Princeton University Press. [50] Loughran, T. and J. R. Ritter, 2004, (cid:148)Why has IPO Underpricing Changed Over Time?(cid:148)Financial Management 33, 5-37. [51] MacKinlay,A.C.,1997,(cid:147)EventStudiesinEconomicsandFinance(cid:148),JournalofEconomicLiterature35,13-39. [52] Malmendier,U.andG.Tate,2009,(cid:147)SuperstarCEOs,(cid:148)QuarterlyJournalofEconomics,124(4),pp.1593-1638. [53] Malmendier, U., G. Tate, and J. Yan, 2011, (cid:147)Overcon(cid:133)dence and Early-life Experiences: The E⁄ect of Managerial Traits on Corporate Financial Policies,(cid:148)forthcoming, Journal of Finance. [54] Marquis Who(cid:146)s Who, Inc. Who(cid:146)s Who in Finance and Business. Chicago: various years [55] Milbourn, T., 2003, (cid:147)CEO Reputation and Stock-Based Compensation,(cid:148)Journal of Financial Economics, 68, pp. 233-262. [56] Murphy, K. J. and J. ZÆbojn(cid:237)k, 2007, (cid:147)Managerial Capital and the Market for CEOs,(cid:148)Working Paper, USC. [57] Oyer, P., 2004, (cid:147)Why do Firms use Incentives that have no Incentive E⁄ects?(cid:148)Journal of Finance, 59, pp. 1619-1649. [58] Palia, D., 2000, (cid:147)The Impact of Regulation on CEO Labor Markets,(cid:148)RAND Journal of Economics, 31(1), pp.165-179. 40

[59] Parrino, R., 1997, (cid:147)CEO Turnover and Outside Succession: A Cross-sectional Analysis,(cid:148)Journal of Financial Economics, 46, pp. 165-197. [60] Perez-Gonzalez, F., 2006, (cid:147)Inherited Control and Firm Performance,(cid:148)American Economic Review, 96(5), pp.1559-1588. [61] Rajgopal, S., T. Shevlin, and V. Zamora, 2006, (cid:147)CEOs(cid:146)Outside Employment Opportunities and the Lack of Relative Performance Evaluation in Compensation Contracts,(cid:148)Journal of Finance, 61, pp. 1813-1844. [62] Rosen, S., 1981, (cid:147)The Economics of Superstars.(cid:148)American Economic Review, 71(3), pp. 845-858. [63] Sattinger, M., 1979, (cid:147)Di⁄erential Rents and the Distribution of Earnings,(cid:148)Oxford Economic Papers, 31, pp. 60(cid:150)71. [64] Sattinger, M., 1993, (cid:147)Assignment Models of the Distribution of Earnings.(cid:148)Journal of Economic Literature, 31, pp. 831(cid:150)880. [65] Schoar, Antoinette, 2007, (cid:147)CEO Careers and Style,(cid:148)working paper, MIT. [66] Spence, M.A., 1973, Market Signaling: Informational Transfer in Hiring and Related Processes, Cambridge, Harvard University Press. [67] Standard & Poor(cid:146)s Corporation. Standard & Poor(cid:146)s Register of Corporations, Directors and Executives. New York: various years. [68] Stiroh, K.J., 2002, "Information Technology and the U.S. Productivity Revival: What do the Industry Data Say?" American Economic Review 92, pp.1559-1576. [69] Tervio, M., 2008, (cid:147)The Di⁄erence that CEOs Make: An Assignment Model Approach,(cid:148)American Economic Review, 98-3, 642-668. [70] Thomson, Gale. Biography Resource Center. Farmington Hills, MI: Gale Group. [71] Wooldridge, J., 2002, Econometric Analysis of Cross Section and Panel Data, Cambridge, MA: MIT Press. 41

8 Appendix A: Details on the Article-Based Proxies To construct our Press, Bad Press, and Good Press proxies, we include the following publications in our search: BusinessWeek, Dow Jones News Service, Financial Times, Forbes, Fortune, International Herald Tribune, Los Angeles Times, The Economist, The New York Times, The Wall Street Journal, The Wall Street Journal Asia, The Wall Street Journal Europe, The Washington Post, USA Today. Our Bad Press proxy is the total count of articles containing the following keywords: scandal or investigat* or (cut w/2 jobs) or resign* or (force* w/3 quit) or dismiss* or demote* or demotion or accuse* or critici* or allegation* or indict* or arrest* or guilty or fraud or litigation or abrasive or excessive pay or overpaid or perquisites or (force* w/3 step down) or under (cid:133)re or under scrutiny or under pressure or law suit or class action or in trouble. Our Good Press proxy is the total count of articles containing the following keywords: leaderorleadershiporreputableorrecognitionordistinguishedorgoodreputationorgreatreputationorhuge reputation or visionary or skillful or personable or talent* or aggressive or (cid:135)exible or adaptable or respectful or fair or integrity or focused or organizer or planner or calm or doer or brainpower or communicator or creative or motivational or enthusiasm or enthusiastic or persisten* or attentive or proactive or tenacity or work* hard or thinker or long hours or persuasive or team play* or teamwork or coaching out or listener or persuas* or persuade or moves fast. 9 Appendix B: Factor Analysis and Measurement Error Factor analysis allows us to combine our various proxies of CEO talent to obtain a more reliable measure of the latent CEO talent variable (our discussion is based on Black and Smith (2006), but see Harman (1976) for details on factor analysis). Formally, suppose that across all CEOs E(CEO Talent )= 0, which is a harmless (cid:3)it normalizationthatkeepsnotationsimple. LetT = (T ;:::;T )beaK-vectorofnoisysignalsofCEOtalent, such 1 k thatforaCEOwithtalentCEOTalent ,thevalueofeachsignalisT = CEOTalent +u withE(T )= 0, (cid:3)it ki (cid:3)i ki ki E u2 = (cid:27)2, E(u u )= 0, j = h; E(u u )= 0, k = k; and E(CEO Talent u )= 0 and the time kit k kj kh 8 6 kj lj 8 6 (cid:3)it ki sub(cid:0)scrip(cid:1)ts are omitted to save on notation. We construct a measure of CEO talent by taking a linear combination of the signals. De(cid:133)ne T= K (cid:28) T (where there is no need for an intercept term because the expected value of k=1 k k CEO Talent (cid:3)i is norm b aliPzed to zero). We select the (cid:28) k (cid:146)s to minimize the expected squared distance between T and CEO Talent , or (cid:3) b 2 min E CEO Talent T : (cid:28);:::;(cid:28)k (cid:3) (cid:0) (cid:16) (cid:17) b 42

The necessary conditions for minimization are K Var(CEO Talent ) (cid:28) Var(CEO Talent ) (cid:28) (cid:27)2= 0; k (cid:3) l (cid:3) k k (cid:0) (cid:0) 8 l=1 X or 1 K (cid:28) (cid:28) r = 0; k, where r = (cid:27)2=Var(CEO Talent ) is the noise-to-signal ratio. For k = 1 and (cid:0) l=1 l (cid:0) k k 8 k k (cid:3) k = l;Pwehavethat(cid:28) l = (cid:28) 1 r r 1 l . Thus,wemaysolvefor(cid:28) 1 toobtain(cid:28) 1 = 1+ r K l 1(cid:0) = 1 1 r l(cid:0) 1 .Theremaining(cid:28)(cid:146)shavesimilar formulae. Thus, (cid:28) k decreases in the variance of the idiosyncratic error u Pk , so that signals that more accurately re(cid:135)ect latent CEO talent receive more weight in the forecast. 10 Appendix C: Variable De(cid:133)nitions Thevariablesusedinthispaperareeitherhand-collectedorextractedfrom(cid:133)vemajordatasources: EXECUCOMP, COMPUSTAT, CRSP, IRRC, BoardEx. For each data item, we indicate the relevant source in square brackets. The speci(cid:133)c variables used in the analysis are de(cid:133)ned as follows: CEO Credentials Proxies: Press: the number of articles containing the CEO(cid:146)s name and company a¢ liation that appear in the major (cid:15) U.S. and global business newspapers in the calendar year prior to succession. For the analysis of the entire ExecuComp sample, we use one-year-lagged count, which measured as of (cid:133)scal year end prior. We also construct Bad Press and Good Press. Bad Press is the number of articles containing the CEO(cid:146)s name, company a¢ liation, and any of the words with a negative connotation that appear in the major U.S. and global business newspapers in the calendar year prior to succession. Good Press is the number of articles containing the CEO(cid:146)s name, company a¢ liation, and any of the words with a positive connotation about CEO talent that appear in the major U.S. and global business newspapers in the calendar year prior to succession. Our text search uses both the CEO(cid:146)s last name and company name. Appendix A contains the detailed list of newspapers used in our Factiva search as well as of the negative and positive words used to construct Bad and Good Press, respectively. All speci(cid:133)cations use the cumulative distribution function of Press, CDF(Press). [Factiva searches] Fast-Track Career: age of the CEO when he took his (cid:133)rst CEO job. We use a cohort-adjusted version of (cid:15) this measure, where we divide our sample of CEOs into three age cohorts and de(cid:133)ne Fast-Track Career as the di⁄erence between age of the (cid:133)rst CEO job and median (cid:133)rst CEO job age in the age cohort. To ease comparisonwiththeotherproxies(sincelowerageof (cid:133)rstCEOjobrepresentsabetterjobmarketcredential), all speci(cid:133)cations use the complement to one of the cumulative distribution function of Fast-Track Career, 1-CDF(Fast-Track Career). [Dun & Bradstreet Reference Book of Corporate Managements (various years); Standard & Poor(cid:146)s Register of Corporations, Directors and Executives; Marquis Who(cid:146)s Who in Finance and Industry; Biography Resource Center by Thomson Gale; Lexis-Nexis, Factiva, and web searches] Selective College: is a numerical rank that takes values between 1 and 6 based on Barron(cid:146)s Pro(cid:133)les of (cid:15) American Colleges (1980) rankings of the undergraduate institution attended by the CEO. In Barron(cid:146)s (1980)rankings, collegesareassignedoneofthefollowingsixranks: MostCompetitive, HighlyCompetitive, Very Competitive, Competitive, Less Competitive, or Noncompetitive. All speci(cid:133)cations use the cumulative distributionfunctionofSelectiveCollege,CDF(SelectiveCollege).[Dun&BradstreetReferenceBookofCorporateManagements(variousyears); Standard&Poor(cid:146)sRegisterofCorporations,DirectorsandExecutives; 43

Marquis Who(cid:146)s Who in Finance and Industry; Biography Resource Center by Thomson Gale; Lexis-Nexis, Factiva, and web searches] CEO Talent Factor: linear combination of Press, Fast-Track Career, and Selective College, with weights (cid:15) calculated using factor analysis for the entire ExecuComp sample. The values of the factor loading are as follows: 0.646 for Fast-Track Career, 0.638 for Press, and 0.465 for Selective College. Press Splines: Press (<50%) equals CDF(Press) if 0.00 CDF(Press) < 0.5 and 0.5 if CDF(Press) 0.5; (cid:15) (cid:20) (cid:21) Press(50%<X<90%)equals CDF(Press)-0.5if 0.5< CDF(Press) < 0.9, 0.0 if CDF(Press) 0.5, and 0.4.if (cid:20) CDF(Press) 0.9; Press (>10%) equals CDF(Press)-0.9 if 0.9 < CDF(Press) < 1.0, 0.0 if CDF(Press) (cid:21) (cid:20) 0.9, where CDF(Press) is the cumulative distribution function of Press. Fast-Track Career Splines: Fast-Track Career (<50%) equals CDF(Fast-Track Career) if 0.00 CDF(Fast- (cid:15) (cid:20) Track Career) < 0.5 and 0.5 if CDF(Fast-Track Career) 0.5; Fast-Track Career (50%<X<90%) equals (cid:21) CDF(Fast-Track Career)-0.5 if 0.5 < CDF(Fast-Track Career) < 0.9, 0.0 if CDF(Fast-Track Career) 0.5, (cid:20) and 0.4.if CDF(Fast-Track Career) 0.9; Fast-Track Career (>10%) equals CDF(Fast-Track Career)-0.9 if (cid:21) 0.9< CDF(Fast-TrackCareer)< 1.0, 0.0ifCDF(Fast-TrackCareer) 0.9, whereCDF(Fast-TrackCareer) (cid:20) is the cumulative distribution function of Fast-Track Career. Selective College Splines: Selective College (<50%) equals CDF (Selective College) if 0.00 CDF (Selective (cid:15) (cid:20) College) < 0.5 and 0.5 if CDF (Selective College) 0.5; Selective College (X>50%) equals CDF (Selective (cid:21) College)-0.5 if 0.5 < CDF (Selective College) 1.0, 0.0 if CDF (Selective College) 0.5 where CDF (cid:20) (cid:20) (Selective College) is the cumulative distribution function of Selective College. Size-Adjusted Press: calculated by subtracting median Press of a control group of (cid:133)rms with similar (cid:133)rm (cid:15) size. ThecontrolgroupsarecreatedbydividingExecuComp(cid:133)rmsintodecilesbasedon(cid:133)rmsize. Theyearly median Press of the relevant group of (cid:133)rms is then used as the control for each (cid:133)rm-year observation (see Barber and Lyon (1996)). Industry-AdjustedPress,Fast-TrackCareer,andSelectiveCollege: arecalculatedbysubtractingthemedian (cid:15) of (Fama-French 48) industry and year of the respective measure. Instrumental Variables for CEO Credentials: Geographic instruments (Average State Press, Average State Fast-Track Career, Average State Selective (cid:15) College): mean of the respective credential proxy among all (cid:133)rms whose headquarters are located in the (cid:133)rm(cid:146)s same state in each year, excluding those (cid:133)rms that are in the (cid:133)rm(cid:146)s same (Fama-French 48) industry group. All speci(cid:133)cations use the cumulative distribution function (CDF) of the underlying instrumental variable. Industry-UKinstruments(AverageUKIndustryFast-TrackCareer,AverageUKIndustrySelectiveCollege): (cid:15) mean of the respective credential proxy among all UK (cid:133)rms that are in the same (Fama-French 48) industry group. Selective College for the UK is de(cid:133)ned based on the list of the most prestigious (so called "ancient") such institutions which we complement with those institutions that are consistently ranked in the top ten basedonthemostpopularpublications(TheTimes,TheGuardian). Theincludedinstitutionsareasfollows: University of Cambridge, University of Oxford, University of St Andrews, London School of Economics, University College London, Durham University. All speci(cid:133)cations use the cumulative distribution function (CDF) of the underlying instrumental variable. [BoardEx, WorldScope] CEO labor market instruments (Average Labor Market Press, Average Labor Market Fast-Track Career, (cid:15) Average Labor Market Selective College): weighted-average of the respective credential proxy among all 44

ExecuComp (cid:133)rms in each year, excluding those (cid:133)rms that are in the (cid:133)rm(cid:146)s same (Fama-French 48) industry group, with weights re(cid:135)ecting the industry-speci(cid:133)c CEO labor market share. In particular, weights are de(cid:133)ned as the share of (cid:133)rms in any given (Fama-French 48) industry group in 1990 with respect to the total number of (cid:133)rms in Compustat. All speci(cid:133)cations use the cumulative distribution function (CDF) of the underlying instrumental variable. CEO Pay and Turnover: CEO pay: log total compensation (TDC1), which is de(cid:133)ned as the sum of short-term compensation (salary (cid:15) and bonus) and long-term compensation (long-term incentive plans, restricted stock, and stock appreciation rights), de(cid:135)ated by CPI in 1990. [EXECUCOMP] Insider: dummy which equals zero when successor CEOs has been with their (cid:133)rms for one year orless at the (cid:15) time of their appointments, and one for all other new CEOs. [Factiva searches] Forced: dummy de(cid:133)ned as in Parrino (1997). It equals one for CEO departures for which the press reports (cid:15) that the CEO has been (cid:133)red, forced out, or retired/resigned due to policy di⁄erences or pressure. It equals zero for departing CEOs above and including age 60. All departures for CEOs below age 60 are reviewed further and classi(cid:133)ed as forced if either the article does not report the reason as death, poor health, or the acceptance of another position (including the chairmanship of the board), or the article reports that the CEO is retiring, but does not announce the retirement at least six months before the succession. [Factiva searches] Firm Performance: AnnouncementCARsforCEOAppointments: cumulativeabnormalreturntotheappointing(cid:133)rm(cid:146)sstockfor (cid:15) trading days (-2, +2) relative to the date of the (cid:133)rst article covering the news of a new CEO appointment. Abnormal returns are calculated using the capital asset pricing model (CAPM) and standard event study methodology (see MacKinlay (1997) for a detailed review). We use the market model and CRSP equallyweightedreturnasthemarketreturntoestimatethemarketmodelparametersfromeventday-210toevent day -11. [CRSP] ROA: ratio of operating income after depreciation (item 178) to book value of assets (item 6). Industry- (cid:15) adjusted ROA is calculated by subtracting the median of (Fama-French 48) industry and year ROA. [COM- PUSTAT] OROA: ratio of net income (item 172) to the book value of assets (item 6). Industry-adjusted OROA is (cid:15) calculated by subtracting the median of (Fama-French 48) industry and year OROA. [COMPUSTAT] OROS:ratioofnetincome(item172)tosales(item12). Industry-adjustedOROSiscalculatedbysubtracting (cid:15) the median of (Fama-French 48) industry and year OROS. [COMPUSTAT] ROE: ratio of net income (item 172) to common equity (item 60). Industry-adjusted ROE is calculated by (cid:15) subtracting the median of (Fama-French 48) industry and year ROE. [COMPUSTAT] Stock returns: annual stock return ((cid:133)scal year-end). [COMPUSTAT] (cid:15) Tobin(cid:146)s Q: ratio of the market value of assets to the book value of assets (item 6). Market value of assets is (cid:15) the book value of assets plus the market value of common equity less the sum of the book value of common equity (item 60) and balance sheet deferred taxes (item 74). [Compustat] Firm Controls & Policies: 45

Size: logofthebookvalueofassets(item6),de(cid:135)atedbyCPIin1990. SmallFirm,MediumFirm,andLarge (cid:15) Firm are three dummies that take value of one for (cid:133)rms in the bottom, intermediate, and top tercile of the sample (cid:133)rm size distribution. [COMPUSTAT] Capital expenditures: capital expenditures (item 128) over total assets at the beginning of the (cid:133)scal year (cid:15) (item 6). [COMPUSTAT] M&As: total number of takeover bid o⁄ers that are classi(cid:133)ed as mergers (successful and unsuccessful) and (cid:15) are announced in a given year. To be included in the count, we require that the merger is material to the acquirer, as standard in the literature, and limit the sample to deals whose value is at least $1 million and at least 1% of the market value of the assets of the acquirer. Finally, we require that the target is a U.S. public or private (cid:133)rm, or a subsidiary, division, or branch of a U.S. (cid:133)rm and that the acquirer controls less than 50% of the shares of the target prior to the acquisition announcement and obtains 100% of the target shares as a result of the transaction. [SDC Platinum, U.S. Mergers and Acquisitions database] Divestitures: totalnumberofassetsales,suchassalesofdivisions,brunches,andproductlines(successfuland (cid:15) unsuccessful) that are announced in a given year [SDC Platinum, U.S. Mergers and Acquisitions database] Diversifying M&As: total number of takeover bid o⁄ers that are classi(cid:133)ed as mergers and involve a target in (cid:15) the same (3-SIC) industry (successful and unsuccessful) and are announced in a given year [SDC Platinum, U.S. Mergers and Acquisitions database] Leverage (book): long term debt (item 9) plus debt in current liabilities (item 34) over the book value of (cid:15) assets (item 6). [COMPUSTAT] Cash holdings: cash (item 1) over book value of assets (item 6). [COMPUSTAT] (cid:15) Dividend Payouts: dividends (item 21) over book value of assets (item 6). [COMPUSTAT] (cid:15) R&D:ratioofR&Dexpenditures(item46,or0ismissing)overbookvalueofassets(item6). [COMPUSTAT] (cid:15) CashFlow: sumofearningsbeforeextraordinaryitems(item18)anddepreciation(item14)overbookvalue (cid:15) of assets (item 6). [COMPUSTAT] Sales Growth: log of the ratio of sales (item 12) in year t to sales in year t 1. [COMPUSTAT] (cid:15) (cid:0) Industry Shocks: For each of the following industry shocks variables, we take the (Fama-French 48) industry median of the absolutevalueofthechangeinthevariableovertheyear. Wethenrank(z-score)eachindustry-yearshockrelative to the 10-year time series of shock observations for the industry. The shock dummy variable takes value of one for increases that are one standard deviation or more above the sample mean. Technology shocks: change in the intensity of investment in information technology (IT) capital. Industry (cid:15) IT intensity in year t is its stock of IT capital relative to other capital. Following Stiroh (2002), we de(cid:133)ne IT capital as seven classes of computer hardware (mainframe computers, personal computers, direct access storagedevices,computerprinters,computerterminals,computertapedrives,andcomputerstoragedevices) and three classes of software (pre-packaged, custom, and own-account software). Investment expenditure in eachofthe61classesareconvertedintoacapitalstockusingstandardperpetualinventorymethod. [Bureau of Economic Analysis (BEA) Fixed Reproducible Tangible Wealth (FRTW)] Growth opportunities shocks: the (cid:133)rst principal component of changes in seven industry growth variables (cid:15) (median ROA, pro(cid:133)tability, asset turnover, R&D, capital expenditures, sales growth, and employee growth) (Harford (2005)).[COMPUSTAT] 46

Organizational capital shocks: change in selling, general, and administrative expenses (SG&A) (item 189). (cid:15) [COMPUSTAT] Domestic competition shocks: change in Her(cid:133)ndahl-Hirschman index (HHI) of sales of all (cid:133)rms in the same (cid:15) industry, where the HHI index is computed using all (cid:133)rms in Compustat. [COMPUSTAT] Foreign competition shocks: change in import penetration, which is de(cid:133)ned as total value of annual imports (cid:15) divided by the sum of total import and domestic production. [Feenstra et al. (2002)] CEO Controls: CEO age: current age of the CEO (years since year of birth). [EXECUCOMP and Dun & Bradstreet (cid:15) Reference Book of Corporate Managements (various years); Standard & Poor(cid:146)s Register of Corporations, Directors and Executives; Marquis Who(cid:146)s Who in Finance and Industry; Biography Resource Center by Thomson Gale; Lexis-Nexis, Factiva, and web searches] CEOtenure: numberofyearsino¢ ceasaCEOatthecurrent(cid:133)rm. [EXECUCOMPandDun&Bradstreet (cid:15) Reference Book of Corporate Managements (various years); Standard & Poor(cid:146)s Register of Corporations, Directors and Executives; Marquis Who(cid:146)s Who in Finance and Industry; Biography Resource Center by Thomson Gale; Lexis-Nexis, Factiva, and web searches] MBA: dummy which equals one if the CEO has an MBA degree. [Dun & Bradstreet Reference Book (cid:15) of Corporate Managements (various years); Standard & Poor(cid:146)s Register of Corporations, Directors and Executives; Marquis Who(cid:146)s Who in Finance and Industry; Biography Resource Center by Thomson Gale; Lexis-Nexis, Factiva, and web searches] Past CEO position: Dummy variable that takes the value of one if a CEO held a CEO position at another (cid:15) publicly-traded company prior to the current position.[BoardEx] Past Number of Jobs: Number of di⁄erent positions a CEO worked in at publicly-traded (cid:133)rms prior to the (cid:15) current position.All speci(cid:133)cations use the cumulative distribution function (CDF) of Past Number of Jobs. [BoardEx] PastNumberofIndustries: Numberof(Fama-French48)industrieswhereaCEOworkedpriortothecurrent (cid:15) position. All speci(cid:133)cations use the cumulative distribution function (CDF) of Past Number of Industries. [BoardEx] CEO General Ability Factor: factor extracted using principal component analysis from the three underlying (cid:15) experience proxies, Past CEO position, Past Number of Jobs, and Past Number of Industries. (Custodio, Ferreira, and Matos (2011)) [BoardEx] CEO Human Capital Factors, #1 ("Experience") & #2 ("Talent"): the (cid:133)rst two principal components (cid:15) extracted from using our three CEO credentials proxies (Press, Fast-Track Career, and Selective College) jointly with the three CEO lifetime work experience proxies (Past CEO position, Past Number of Jobs, and Past Number of Industries). [Dun & Bradstreet Reference Book of Corporate Managements (various years); Standard & Poor(cid:146)s Register of Corporations, Directors and Executives; Marquis Who(cid:146)s Who in Finance and Industry; Biography Resource Center by Thomson Gale; Lexis-Nexis, Factiva, and web searches; BoardEx] Governance & Connections Controls: GIM-index ( 11) dummy variable that takes value of one for (cid:133)rms with 11 of more of the 24 antitakeover (cid:15) (cid:21) provisions includes in the GIM index of Gompers, Ishii, and Metrick (2003). [IRRC]. 47

Board size: total number of directors on the board in a given (cid:133)rm-year. [IRRC] (cid:15) Board independence: dummy variable that takes value of one for (cid:133)rms whose ratio of the number of in- (cid:15) dependent directors to overall number of directors in a given (cid:133)rm-year above median (larger than 0.67). [IRRC] CEO Education Network: number of education ties of the CEO, as measured by the number of individuals (cid:15) (top executives and directors) in BoardEx who attended the same school of the CEO at the same time. All speci(cid:133)cations use the cumulative distribution function (CDF) of CEO Education Network. [BoardEx] CEO Corporate Network: numberof corporate ties of the CEOas measured by the sumof CurrentEmploy- (cid:15) ment Network and Prior Employment Network. Current Employment Network is the number of individuals in BoardEx who currently serve in another common publicly traded company with the CEO. Prior Employment Network is the number of individuals in BoardEx who served in at least one common publicly traded company with the CEO in the past, excluding prior roles in the company in question. All speci(cid:133)cations use the cumulative distribution function (CDF) of CEO Corporate Network. [BoardEx] 48

Table 1 Sample Distribution by Year The sample consists of 2,195 CEO successions between 1993 and 2005 for (cid:133)rms whose CEOs are covered by the ExecuCompdatabase. Thistablepresentsanoverviewofthedatasetbyshowingthenumberandthefrequencyof forced, voluntary, and outside successions in the sample. Classi(cid:133)cation of each succession into forced or voluntary is based on the Factiva news database search following Parrino (1997). Successions are classi(cid:133)ed as internal when incoming CEOs were hired by the (cid:133)rm earlier than a year before succession, and external otherwise. Successions due to mergers and spin-o⁄s are excluded. Panel A: Sample Distribution by Year Number Number of Percent Firms Percent Firms Percent Firms Number of Year of forced outsiders with with forced with outsiders successions successions appointed successions successions appointed 1993 110 22 (20.0%) 31 (28.1%) 9.6% 1.9% 2.7% 1994 125 31 (24.8%) 38 (30.4%) 8.1% 2.0% 2.5% 1995 158 32 (20.5%) 52 (32.9%) 10.0% 2.0% 3.3% 1996 155 45 (29.0%) 52 (33.5%) 9.5% 2.7% 3.1% 1997 185 46 (24.9%) 63 (34.1%) 11.1% 2.8% 3.8% 1998 186 49 (26.3%) 74 (39.8%) 10.8% 2.8% 4.2% 1999 224 67 (29.9%) 85 (38.0%) 12.5% 3.7% 4.7% 2000 244 59 (24.2%) 93 (38.1%) 13.6% 3.3% 5.2% 2001 173 49 (28.3%) 67 (38.7%) 10.4% 2.9% 4.0% 2002 195 68 (34.9%) 77 (39.5%) 11.8% 4.1% 4.6% 2003 166 40 (24.1%) 65 (34.3%) 9.9% 2.4% 3.9% 2004 152 37 (24.3%) 62 (40.8%) 9.8% 2.2% 3.7% 2005 122 30 (24.6%) 51 (41.8%) 9.5% 2.3% 3.9% Total 2195 575 (26.2%) 810 (36.9%) 10.5% 2.8% 3.9% Panel B: Annual Averages by Sub-Period Number Number of Percent Firms Percent Firms Percent Firms Number of Period of forced outsiders with with forced with outsiders successions successions appointed successions successions appointed 1993-95 131 28 (21.8%) 40 (30.5%) 9.2% 2.0% 2.8% 1996-00 199 53 (26.9%) 73 (36.7%) 11.5% 3.1% 4.2% 2001-05 162 45 (27.2%) 64 (39.0%) 10.3% 2.8% 4.0% 49

Table 2 Summary Statistics The sample consists of 2,195 CEO successions between 1993 and 2005 for (cid:133)rms whose CEOs are covered by the ExecuComp database. This table reports summary statistics of the key variables used in our analysis. Panel A shows pairwise correlations between our three measures of CEO credentials. Panel B shows summary statistics for CEO credentials, (cid:133)rm characteristics, and other CEO controls by CEO succession type. The three measures of CEO credentials are: Press, which is the number of articles containing the CEO(cid:146)s name and company a¢ liation that appear in the major U.S. and global business newspapers in the calendar year prior to succession; Fast-Track Career, which is the age of CEO when he took his (cid:133)rst CEO job; Selective College, which is the standing in the Barron(cid:146)s (1980) rankings of the undergraduate institution attended by the CEO. Classi(cid:133)cation of each succession into forced or voluntary is based on the Factiva news database search following Parrino (1997). Successions are classi(cid:133)edasinternalwhenincomingCEOswerehiredbythe(cid:133)rmearlierthanayearbeforesuccession,andexternal otherwise. See Appendix C for additional details on the three measures of CEO credentials and for de(cid:133)nitions of the controls. Panel A: Pairwise Correlations Among CEO Credentials Press Fast-Track Career Selective College A.1: All Successions [N=2,195] Press 1.000 Fast-Track Career 0.144 1.000 (cid:3)(cid:3)(cid:3) Selective College 0.075 0.065 1.000 (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) A.2: All Successions, Top Quartile Press [N=548] Press 1.000 Fast-Track Career 0.243 1.000 (cid:3)(cid:3)(cid:3) Selective College 0.137 0.182 1.000 (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) Panel B: CEO Credentials by Succession Type Type of Succession All Forced Outside Inside N=2195 N=581 N=810 N=1385 B.1: Outgoing CEO CEO Credentials: Press 7.2 7.7 6 7.4 Fast-Track Career (years) 49 46 48 49 Selective College (rank) 2.4 2.6 2.4 2.4 B.2: Successor CEO CEO Credentials: Press 7.9 10.8 9.1 6.9 Fast-Track Career (years) 49 45 48 50 Selective College (rank) 2.9 3.2 2.9 2.9 CEO Pay: Total CEO Pay (log tdc1, $000) 7.8 7.8 7.9 7.6 B.3: Firm Variables (year prior to transition) Size (log total assets, $mil) 7.4 7.3 7.1 7.6 Firm Stock Return -14.1% -28.3% -21.4% -10.1% Industry Stock Return (EW) 13.9% 13.0% 14.7% 13.4% Industry-Adjusted OROA 0.014 -0.022 -0.015 0.023 GIM index 9 9 9 9 Board Independence 65% 64% 66% 64% 50

3 elbaT sisylanA noissergeR enilesaB :slaitnederC OEC rof yaP ehT .sOEC detnioppa ylwen rof 5002 ot 3991 morf slaitnederc OEC fo serusaem no yap OEC latot fo snoisserger SLO fo setamitse stroper elbat sihT ,reeraC kcarT-tsaF ,sserP slaitnederc OEC fo serusaem eerht eht yolpme ylevitareti eW .)1cdt( yap latot fo mhtiragol eht si elbairav tnedneped llew sa ,stce⁄e dex(cid:133)-yrtsudni )84 hcnerF-amaF( dna -raey htiw noitac(cid:133)iceps enilesab a :snoitac(cid:133)iceps tnere⁄id eerht ni hcae - egelloC evitceleS dna ,)1( snmuloC( yap OEC latot tce⁄a ot hcraeser suoiverp ni nwohs neeb evah taht scitsiretcarahc OEC rehto dna ,snoisseccus ,mr(cid:133) rof slortnoc sa latipac dna ,D&R ,htworg selas ,sgnidloh hsac ,wo(cid:135) hsac ,AOR ,Q s(cid:146)niboT ,tuoyap dnedivid ,egarevel )koob( mr(cid:133) sdda taht noitac(cid:133)iceps a ;))7( ,)4( ,)3( snmuloC( tnemtnioppa hcae ot roirp boj eht ni yap latot )gol( OEC sdda rehtruf taht noitac(cid:133)iceps a dna ;))8( ,)5( ,)2( snmuloC( serutidnepxe era evitucexe yb snoitavresbo fo ecnednepedni-non rof detsujda srorre dradnats deretsulc tsuboR .C xidneppA ni era snoitin(cid:133)ed elbairaV .))9( ,)6( .ylevitcepser ,level %01 dna ,%5 ,%1 eht ta ecnac(cid:133)ingis lacitsitats rof dna , , yb detoned era ecnac(cid:133)ingis fo sleveL .sesehtnerap ni detroper (cid:3) (cid:3)(cid:3) (cid:3)(cid:3)(cid:3) .yap fo naem elpmas eht ta detaulave si ytivitisnes deilpmI ylno raey tnemtnioppa ;noitasnepmoc launna latot gol :elbairav tnednepeD )9( )8( )7( )6( )5( )4( )3( )2( )1( rof lortnoC mriF eroM enilesaB rof lortnoC mriF eroM enilesaB rof lortnoC mriF eroM enilesaB yaP roirP slortnoC yaP roirP slortnoC yaP roirP slortnoC :slaitnederC OEC ***914.0 ***905.0 ***445.0 sserP )811.0( )290.0( )980.0( 745.0 764.0 954.0 reeraC kcarT-tsaF (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) )981.0( )171.0( )761.0( **162.0 **642.0 **102.0 egelloC evitceleS )131.0( )011.0( )980.0( ,noisseccuS ,mriF :slortnoC OEC & 471.0 731.0 541.0 740.0 970.0- 330.0- 911.0 830.0 221.0 nruteR kcotS (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3) (cid:3)(cid:3) 1 t )850.0( )540.0( )950.0( )870.0( )550.0( )840.0( )870.0( )940.0( )550.0( (cid:0) ***803.0 ***604.0 ***393.0 ***873.0 ***014.0 ***524.0 ***443.0 ***793.0 ***973.0 eziS mriF )630.0( )320.0( )810.0( )320.0( )910.0( )510.0( )130.0( )710.0( )610.0( **010.0- **010.0- **110.0- **120.0- ***620.0- ***910.0- ***610.0- ***610.0- ***310.0egA OEC )400.0( )400.0( )500.0( )900.0( )900.0( )700.0( )500.0( )400.0( )500.0( 501.0- 751.0- 741.0- 101.0- ***424.0- ***184.0- 011.0- ***582.0- ***563.0noisseccuS redisnI )380.0( )011.0( )901.0( )280.0( )670.0( )950.0( )880.0( )740.0( )840.0( 801.0 831.0 360.0 961.0 360.0 721.0 180.0 350.0 670.0 noisseccuS decroF (cid:3) (cid:3) )931.0( )890.0( )070.0( )090.0( )470.0( )170.0( )190.0( )850.0( )360.0( 912.0 401.0 151.0 yaP roirP OEC (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) )370.0( )630.0( )340.0( seY seY oN seY seY oN seY seY oN slortnoC mriF eroM seY seY seY seY seY seY seY seY seY .E.F raeY seY seY seY seY seY seY seY seY seY .E.F yrtsudnI %3.15 %5.64 %1.44 %8.64 %4.14 %1.14 %7.93 %5.43 %7.23 2R 977,1 977,1 977,1 869 828,1 828,1 250,1 221,2 221,2 snoitavresbO :)slaitnederC %1-yap 000$( ytivitisneS laitnederC-yaP deilpmI 4.82 sserP 0.42 reeraC kcarT-tsaF 5.01 egelloC evitceleS 51

4 elbaT ytixevnoC :slaitnederC OEC rof yaP ehT .sOEC detnioppa ylwen rof 5002 ot 3991 morf slaitnederc OEC fo serusaem no yap OEC latot fo snoisserger SLO fo setamitse stroper elbat sihT dna ,reeraC kcarT-tsaF ,sserP - slaitnederc OEC fo serusaem eerht eht yolpme ylevitareti eW .)1cdt( yap latot fo mhtiragol eht si elbairav tnedneped slaitnedercOECrofyapniytienegoretehrofwollaotserusaemgniylrednuehtfosenilpssesutahtnoitac(cid:133)icepsraenil-esiweceipani-egelloC evitceleS slaitnederc OEC eht fo senilps raenil esiw-ecip eht rof stluser tneserp eW .slaitnederc OEC fo noitubirtsid eht fo segnar tnere⁄id no gnidneped slortnoc sa llew sa ,stce⁄e dex(cid:133)-yrtsudni )84 hcnerF-amaF( dna -raey htiw noitac(cid:133)iceps enilesab a :snoitac(cid:133)iceps tnere⁄id eerht ni hcae selbairav a ;))7( ,)4( ,)1( snmuloC( yap OEC latot tce⁄a ot hcraeser suoiverp ni nwohs neeb evah taht scitsiretcarahc OEC rehto dna ,snoisseccus ,mr(cid:133) rof serutidnepxe latipac dna ,D&R ,htworg selas ,sgnidloh hsac ,wo(cid:135) hsac ,AOR ,Q s(cid:146)niboT ,tuoyap dnedivid ,egarevel )koob( mr(cid:133) sdda taht noitac(cid:133)iceps llA .))9( ,)6( ,)3( snmuloC( tnemtnioppa hcae ot roirp boj eht ni yap latot )gol( OEC sdda rehtruf taht noitac(cid:133)iceps a dna ;))8( ,)5( ,)2( snmuloC( taht scitsiretcarahc OEC rehto dna ,snoisseccus ,mr(cid:133) rof slortnoc sa llew sa ,stce⁄e dex(cid:133)-yrtsudni )84 hcnerF-amaF( dna -raey edulcni snoitac(cid:133)iceps detsujda srorre dradnats deretsulc tsuboR .C xidneppA ni era snoitin(cid:133)ed elbairaV .yap OEC latot tce⁄a ot hcraeser suoiverp ni nwohs neeb evah lacitsitats rof dna , , yb detoned era ecnac(cid:133)ingis fo sleveL .sesehtnerap ni detroper era evitucexe yb snoitavresbo fo ecnednepedni-non rof (cid:3) (cid:3)(cid:3) (cid:3)(cid:3)(cid:3) .yap OEC latot fo naem elpmas eht ta detaulave si ytivitisnes deilpmI .ylevitcepser ,level %01 dna ,%5 ,%1 eht ta ecnac(cid:133)ingis ylno raey tnemtnioppa ;noitasnepmoc launna latot gol :elbairav tnednepeD )9( )8( )7( )6( )5( )4( )3( )2( )1( rof lortnoC mriF eroM enilesaB rof lortnoC mriF eroM enilesaB rof lortnoC mriF eroM enilesaB yaP roirP slortnoC yaP roirP slortnoC yaP roirP slortnoC 140.0 123.0 641.0 )%05<( sserP (cid:3) )991.0( )171.0( )751.0( 070.2 688.2 ***869.2 )%09<X<%05( sserP (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) )243.0( )582.0( )042.0( 699.9 403.11 ***891.31 )%09>( sserP (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) )575.3( )724.2( )009.1( 172.0 333.0 661.0 )%05<( reeraC kcarT-tsaF )283.0( )423.0( )312.0( 019.1 884.1 587.1 )%09<X<%05( reeraC kcarT-tsaF (cid:3)(cid:3) (cid:3)(cid:3) (cid:3)(cid:3) )469.0( )447.0( )747.0( 544.41 592.11 026.11 )%09>( reeraC kcarT-tsaF (cid:3)(cid:3) (cid:3)(cid:3) (cid:3)(cid:3) )908.6( )165.4( )478.4( 340.0 990.0 420.0 )%05<( egelloC evitceleS )981.0( )561.0( )851.0( 190.1 330.1 811.1 )%05>( egelloC evitceleS (cid:3)(cid:3) (cid:3)(cid:3) (cid:3)(cid:3)(cid:3) )644.0( )234.0( )773.0( ,noisseccuS ,mriF seY seY seY seY seY seY seY seY seY slortnoC OEC & seY seY oN seY seY oN seY seY oN slortnoc mr(cid:133) eroM seY seY seY seY seY seY seY seY seY .E.F raeY seY seY seY seY seY seY seY seY seY .E.F yrtsudnI %3.35 %9.94 %1.64 %2.45 %1.94 %8.54 %1.64 %3.04 %4.63 2R 977,1 977,1 977,1 869 828,1 828,1 250,1 221,2 221,2 snoitavresbO :)slaitnederC %1-yap 000$( tekcarB laitnederC poT ni sOEC rof ytivitisneS laitnederC-yaP deilpmI 9.986 sserP 4.706 reeraC kcarT-tsaF 4.85 egelloC evitceleS 52

5 elbaT eziS mriF htiw ytiratnemelpmoC :slaitnederC OEC rof yaP ehT .sOEC detnioppa ylwen rof 5002 ot 3991 morf slaitnederc OEC fo serusaem no yap OEC latot fo snoisserger SLO fo setamitse stroper elbat sihT dna ,reeraC kcarT-tsaF ,sserP slaitnederc OEC fo serusaem eerht eht yolpme ylevitareti eW .)1cdt( yap latot fo mhtiragol eht si elbairav tnedneped dna ,muidem ,llams rof seimmud eerht htiw serusaem gniylrednu eht fo snoitcaretni sesu taht noitac(cid:133)iceps raenil-esiweceip a ni egelloC evitceleS stluser tneserp eW .ezis mr(cid:133) fo noitubirtsid eht fo segnar tnere⁄id no gnidneped slaitnederc OEC rof yap ni ytienegoreteh rof wolla ot smr(cid:133) egral )84 hcnerF-amaF( dna -raey htiw noitac(cid:133)iceps enilesab a :snoitac(cid:133)iceps tnere⁄id eerht ni hcae selbairav slaitnederc OEC eht fo noitcaretni eht rof tce⁄a ot hcraeser suoiverp ni nwohs neeb evah taht scitsiretcarahc OEC rehto dna ,snoisseccus ,mr(cid:133) rof slortnoc sa llew sa ,stce⁄e dex(cid:133)-yrtsudni ,sgnidloh hsac ,wo(cid:135) hsac ,AOR ,Q s(cid:146)niboT ,tuoyap dnedivid ,egarevel )koob( mr(cid:133) sdda taht noitac(cid:133)iceps a ;))7( ,)4( ,)1( snmuloC( yap OEC latot ot roirp boj eht ni yap latot )gol( OEC sdda rehtruf taht noitac(cid:133)iceps a dna ;))8( ,)5( ,)2( snmuloC( serutidnepxe latipac dna ,D&R ,htworg selas fo ecnednepedni-non rof detsujda srorre dradnats deretsulc tsuboR .C xidneppA ni era snoitin(cid:133)ed elbairaV .))9( ,)6( ,)3( snmuloC( tnemtnioppa hcae ,%5 ,%1 eht ta ecnac(cid:133)ingis lacitsitats rof dna , , yb detoned era ecnac(cid:133)ingis fo sleveL .sesehtnerap ni detroper era evitucexe yb snoitavresbo (cid:3) (cid:3)(cid:3) (cid:3)(cid:3)(cid:3) .yap OEC latot fo naem elpmas eht ta detaulave si ytivitisnes deilpmI .ylevitcepser ,level %01 dna ylno raey tnemtnioppa ;noitasnepmoc launna latot gol :elbairav tnednepeD )9( )8( )7( )6( )5( )4( )3( )2( )1( rof lortnoC mriF eroM enilesaB rof lortnoC mriF eroM enilesaB rof lortnoC mriF eroM enilesaB yaP roirP slortnoC yaP roirP slortnoC yaP roirP slortnoC 060.0 071.0 841.0 mriF llamS *sserP )682.0( )291.0( )991.0( 447.0 094.0 ***065.0 mriF muideM*sserP (cid:3)(cid:3) (cid:3)(cid:3) )692.0( )222.0( )081.0( 600.1 910.1 ***931.1 mriF egraL*sserP (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) )323.0( )532.0( )732.0( 350.0 221.0 890.0 mriF llamS *reeraC kcarT-tsaF )351.0( )711.0( )111.0( 334.0 014.0 263.0 mriF muideM*reeraC kcarT-tsaF (cid:3)(cid:3)(cid:3) (cid:3)(cid:3) (cid:3)(cid:3) )441.0( )971.0( )461.0( 607.1 584.1 374.1 mriF egraL*reeraC kcarT-tsaF (cid:3)(cid:3)(cid:3) (cid:3)(cid:3) (cid:3)(cid:3)(cid:3) )553.0( )147.0( )283.0( 820.0 700.0 390.0 mriF llamS *egelloC evitceleS )990.0( )121.0( )021.0( 460.0 160.0 831.0 mriF muideM*egelloC evitceleS )021.0( )911.0( )031.0( 546.0 744.0 **474.0 mriF egraL*egelloC evitceleS (cid:3)(cid:3) (cid:3)(cid:3) )582.0( )291.0( )291.0( ,noisseccuS ,mriF seY seY seY seY seY seY seY seY seY slortnoC OEC & seY seY oN seY seY oN seY seY oN slortnoc mr(cid:133) eroM seY seY seY seY seY seY seY seY seY .E.F raeY seY seY seY seY seY seY seY seY seY .E.F yrtsudnI %5.55 %1.05 %1.54 %8.25 %1.84 %3.34 %9.44 %0.14 %4.53 2R 977,1 977,1 977,1 869 828,1 828,1 250,1 221,2 221,2 snoitavresbO :)slaitnederC %1-yap 000$( smriF egraL rof ytivitisneS laitnederC-yaP deilpmI 5.95 sserP 0.77 reeraC kcarT-tsaF 8.42 egelloC evitceleS 53

6 elbaT setamitsE )VI( selbairaV latnemurtsnI dna stce⁄E dexiF mriF :slaitnederC OEC rof yaP gniyfitnedI fo erusaem a no yap OEC latot fo snoisserger ))7(-)5( snmuloC( selbairaV latnemurtsnI dna ))4(-)1( snmuloC( SLO fo setamitse stroper elbat sihT tnelaT OEC slaitnederc OEC fo erusaem ehT .)1cdt( yap latot fo mhtiragol eht si elbairav tnedneped ehT .5002 ot 3991 morf slaitnederc OEC raey edulcni snoitac(cid:133)iceps llA .egelloC evitceleS dna ,reeraC kcarT-tsaF ,sserP morf sisylana tnenopmoc lapicnirp gnisu detcartxe rotcaf a si rotcaF OEC rehto dna ,snoisseccus ,mr(cid:133) rof slortnoc sa llew sa ,stce⁄e dex(cid:133) ]yrtsudni )84 hcnerF-amaF( edulcni hcihw )2(-)1( snmuloC rof tpecxe[ mr(cid:133) dna detnioppaylwenrofsetamitseSLOenilesabtneserp)2(-)1(snmuloC .yapOEClatottce⁄aothcraesersuoiverpninwohsneebevahtahtscitsiretcarahc htiw,pmoCucexEnisOECllarofstce⁄e-dex(cid:133)mr(cid:133)htiwsetamitseSLOera)4(-)3(snmuloC .ylevitcepser,segnahcdnaslevelnisnoitac(cid:133)icepsnisOEC snmuloC .erunet OEC no gnidneped slaitnederc OEC rof yap ni ytienegoreteh rof wolla ot erunet OEC htiw mret noitcaretni na gnidda )4( nmuloC OEC dna ,KU-yrtsudni ,cihpargoeg fo stes tnere⁄id eerht yb nrut ni detnemurtsni si rotcaF tnelaT OEC eht erehw ,setamitse VI eht troper )7(-)5( stroper osla lenap ehT .noitamitse pets-tsr(cid:133) eht ni stneic ¢eoc evitcepser rieht htiw selbairav eseht stsil lenap mottob ehT .selbairav tekram robal seulav-p( snoitcirtser noitac(cid:133)itnedi-revo tnemurtsni dna )citsitats tset-F( ecnac(cid:133)ingis tnemurtsni dedulcxe tnioj rof scitsitats citsongaid noitamitse VI yb snoitavresbo fo ecnednepedni-non rof detsujda srorre dradnats deretsulc tsuboR .C xidneppA ni era snoitin(cid:133)ed elbairaV .)citsitats-J nesnaH fo ,level %01 dna ,%5 ,%1 eht ta ecnac(cid:133)ingis lacitsitats rof dna , , yb detoned era ecnac(cid:133)ingis fo sleveL .sesehtnerap ni detroper era evitucexe (cid:3) (cid:3)(cid:3) (cid:3)(cid:3)(cid:3) .yap OEC latot fo naem elpmas eht ta detaulave si ytivitisnes deilpmI .ylevitcepser noitasnepmoc launna latot gol :elbairav tnednepeD sisylanA selbairaV latnemurtsnI pmoCucexE llA ylno raey tnemtnioppA )7( )6( )5( )4( )3( )2( )1( tekraM robaL yrtsudnI-KU cihpargoeG htiw noitcaretnI .E.F mriF enilesaB enilesaB stnemurtsnI stnemurtsnI stnemurtsnI eruneT OEC )1cdt(gol (cid:1) 314.0 694.0 424.0 844.0 982.0 914.0 074.0 rotcaF tnelaT OEC (cid:3)(cid:3)(cid:3) (cid:3)(cid:3) (cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) )901.0( )152.0( )391.0( )770.0( )150.0( )151.0( )990.0( *rotcaF tnelaT OEC 810.0eruneT OEC (cid:3)(cid:3)(cid:3) )600.0( ,noisseccuS ,mriF seY seY seY seY seY seY seY slortnoC OEC & seY seY seY seY seY seY seY .E.F raeY seY seY seY seY seY oN oN .E.F mriF %5.27 %8.18 %1.17 %5.76 %1.07 %4.62 %3.14 2R 747,21 832,6 237,21 747,21 747,21 963,1 177,1 snoitavresbO rotcaF tnelaT OEC :elbairav tnednepeD - )sisylanA VI( noitamitsE egats-tsriF 740.0 sserP etatS egarevA (cid:3)(cid:3)(cid:3) )210.0( 730.0 reeraC kcarT-tsaF etatS egarevA (cid:3)(cid:3)(cid:3) )010.0( 170.0 egelloC evitceleS etatS egarevA (cid:3)(cid:3)(cid:3) )020.0( 230.0 reeraC kcarT-tsaF yrtsudnI KU egarevA (cid:3)(cid:3) )410.0( 490.0 egelloC evitceleS yrtsudnI KU egarevA (cid:3)(cid:3)(cid:3) )810.0( 951.0 sserP tekraM robaL egarevA (cid:3)(cid:3)(cid:3) )350.0( 194.0 reeraC kcarT-tsaF tekraM robaL egarevA (cid:3)(cid:3)(cid:3) )071.0( 961.0 egelloC evitceleS tekraM robaL egarevA (cid:3)(cid:3) )280.0( %8.28 %1.28 %2.57 2R 4.12 97.8 25.7 stnemurtsni .lcxe fo tset-F (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) 42.0 45.0 25.0 )eulav-p( citsitats-J nesnaH :)slaitnederC %1-yap 000$( ytivitisneS laitnederC-yaP deilpmI 7.12 8.42 1.22 7.21 9.12 6.42 rotcaF tnelaT OEC )6.9( 2.42 ))5>( 1=eruneT(*rotcaF tnelaT OEC 54

7 elbaT stce⁄E dexiF OEC dna skcohS yrtsudnI :slaitnederC OEC rof yaP gniyfitnedI ediw-yrtsudni fo yteirav a htiw noitcaretni sti dna slaitnederc OEC fo erusaem a no yap OEC latot fo snoisserger SLO fo setamitse stroper elbat sihT OEC fo erusaem ehT .)1cdt( yap latot fo mhtiragol eht si elbairav tnedneped ehT .pmoCucexE ni sOEC lla rof 5002 ot 3991 morf skcohs cimonoce llA .egelloC evitceleS dna ,reeraC kcarT-tsaF ,sserP morf sisylana tnenopmoc lapicnirp gnisu detcartxe rotcaf a si - rotcaF tnelaT OEC - slaitnederc sa llew sa ,stce⁄e dex(cid:133) ))01( dna ,)8( ,)6( ,)4( ,)2( snmuloC( OEC ro ))9( dna ,)7( ,)5( ,)3( ,)1( snmuloC( mr(cid:133) rehtie dna raey edulcni snoitac(cid:133)iceps troper)2(-)1(snmuloC .yapOEClatottce⁄aothcraesersuoiverpninwohsneebevahtahtscitsiretcarahcOECrehtodna,snoisseccus,mr(cid:133)rofslortnoc tnemtsevni fo ytisnetni eht ni htworg hgih htiw sraey-yrtsudni esoht ni eno slauqe taht ymmud a sa den(cid:133)ed era hcihw ,skcohs ygolonhcet rof stluser slauqe taht ymmud a sa den(cid:133)ed era hcihw ,seitinutroppo htworg ot skcohs yrtsudni ot refer )4(-)3( snmuloC .latipac )TI( ygolonhcet noitamrofni ni selbairav htworg yrtsudni neves ni segnahc fo tnenopmoc lapicnirp tsr(cid:133) eht yb deixorp sa seitinutroppo htworg hgih htiw sraey-yrtsudni esoht ni eno troper )6(-)5( snmuloC .))5002( drofraH( )htworg eeyolpme dna ,htworg selas ,serutidnepxe latipac ,D&R ,revonrut tessa ,ytilibat(cid:133)orp ,AOR naidem( lanoitazinagro ni htworg hgih htiw sraey-yrtsudni esoht ni eno slauqe taht ymmud a sa den(cid:133)ed era hcihw ,skcohs latipac lanoitazinagro rof stluser noititepmoc citsemod rof stluser troper )8(-)7( snmuloC .)A&GS( sesnepxe evitartsinimda dna ,lareneg ,gnilles naidem yrtsudni yb seixorp sa latipac snmuloC .)IHH( xedni lahdn(cid:133)reH yrtsudni ni sesaerced egral htiw sraey-yrtsudni esoht ni eno slauqe taht ymmud a sa den(cid:133)ed era hcihw ,skcohs ni sesaercni egral htiw sraey-yrtsudni esoht ni eno slauqe taht ymmud a sa den(cid:133)ed era hcihw ,skcohs noititepmoc ngierof rof stluser troper )01(-)9( .raey eht revo elbairav eht ni egnahc eht fo eulav etulosba eht fo naidem yrtsudni eht ekat ew ,selbairav skcohs eseht fo hcae roF .noitartenep tropmi elbairav ymmud kcohs ehT .yrtsudni eht rof snoitavresbo kcohs fo seires emit raey-01 eht ot evitaler kcohs raey-yrtsudni hcae )erocs-z( knar neht eW tsuboR .C xidneppA ni era snoitin(cid:133)ed elbairaV .naem elpmas eht evoba erom ro noitaived dradnats eno era taht sesaercni rof eno fo eulav sekat detoned era ecnac(cid:133)ingis fo sleveL .sesehtnerap ni detroper era evitucexe yb snoitavresbo fo ecnednepedni-non rof detsujda srorre dradnats deretsulc latot fo naem elpmas eht ta detaulave si ytivitisnes deilpmI .ylevitcepser ,level %01 dna ,%5 ,%1 eht ta ecnac(cid:133)ingis lacitsitats rof dna , , yb (cid:3) (cid:3)(cid:3) (cid:3)(cid:3)(cid:3) .yap OEC pmoCucexE lla ;noitasnepmoc launna latot gol :elbairav tnednepeD ngieroF citsemoD lanoitazinagrO htworG ygolonhceT noititepmoC noititepmoC latipaC seitinutroppO )01( )9( )8( )7( )6( )5( )4( )3( )2( )1( EF OEC EF mriF EF OEC EF mriF EF OEC EF mriF EF OEC EF mriF EF OEC EF mriF 772.0 032.0 202.0 342.0 261.0 851.0 251.0 002.0 991.0 051.0 rotcaF tnelaT OEC (cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3) (cid:3)(cid:3) )141.0( )780.0( )080.0( )060.0( )960.0( )650.0( )760.0( )060.0( )580.0( )470.0( *rotcaF tnelaT OEC 470.0 544.0 411.0 711.0 813.0 343.0 311.0 911.0 480.0 470.0 kcohS yrtsudnI (cid:3)(cid:3)(cid:3) (cid:3)(cid:3) (cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) 1 t )281.0( )271.0( )550.0( )050.0( )780.0( )480.0( )240.0( )540.0( )860.0( )260.0( (cid:0) *rotcaF tnelaT OEC 711.0 892.0 921.0 411.0 441.0 332.0 320.0 020.0 951.0 ***481.0 kcohS yrtsudnI (cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3) 2 t )522.0( )971.0( )050.0( )440.0( )080.0( )870.0( )140.0( )440.0( )660.0( )850.0( (cid:0) *rotcaF tnelaT OEC 722.0 700.0 360.0 950.0 801.0 941.0 600.0 330.0 712.0 ***471.0 kcohS yrtsudnI (cid:3)(cid:3) (cid:3)(cid:3) (cid:3)(cid:3)(cid:3) 3 t )211.0( )121.0( )550.0( )050.0( )670.0( )570.0( )040.0( )640.0( )960.0( )660.0( (cid:0) ,noisseccuS ,mriF seY seY seY seY seY seY seY seY seY seY slortnoC OEC & oN seY oN seY oN seY oN seY oN seY .E.F mriF seY oN seY oN seY oN seY oN seY oN .E.F OEC seY seY seY seY seY seY seY seY seY seY .E.F raeY seY oN seY oN seY oN seY oN seY oN .E.F yrtsudnI %9.37 %6.96 %7.17 %7.86 %3.17 %1.86 %7.27 %3.17 %7.37 %7.17 2R 421,6 421,6 747,21 747.21 747,21 747.21 747,21 747.21 761,6 761,6 snoitavresbO :)slaitnederC %1-yap 000$( ytivitisneS laitnederC-yaP no skcohS yrtsudnI fo tce⁄E deilpmI 7.23 2.01 9.13 2.5 7.51 rotcaF tnelaT OEC 55

8 elbaT yaP OEC ni dnerT fo stcaF dezilytS rof snoitacilpmI :slaitnederC OEC rof yaP gnissessA dnert emit htiw noitcaretni sti dna slaitnederc OEC fo erusaem a no yap OEC latot fo snoisserger elitnauq dna SLO fo setamitse stroper elbat sihT sa den(cid:133)ed era hcihw ,pmoCucexE ni sOEC detnioppa yltnecer rof dna )A lenaP( pmoCucexE ni sOEC lla rof 5002 ot 3991 morf selbairav rotacidni mhtiragolehtdna)8(-)1(snmuloCni)1cdt(yaplatotfomhtiragolehtsielbairavtnednepedehT.)BlenaP(sselrosraeyowtfoerunethtiwsOECesoht sisylana tnenopmoc lapicnirp gnisu detcartxe rotcaf a si - rotcaF tnelaT OEC - slaitnederc OEC fo erusaem ehT .)01(-)9( snmuloC ni yap ytiuqe fo dna0002ot6991sraeynienofoeulavekattahtseimmuderaselbairavrotacidnidnertemitehT .egelloCevitceleSdna ,reeraCkcarT-tsaF ,sserPmorf evahtahtscitsiretcarahcOECrehtodna,snoisseccus,mr(cid:133)rofslortnocsallewsa,stce⁄edex(cid:133)mr(cid:133)edulcnisnoitac(cid:133)icepsllA .ylevitcepser,5002ot1002 troper )4(-)3( snmuloC .yap OEC ni dnert llarevo eht rof stluser troper )2(-)1( snmuloC .yap OEC latot tce⁄a ot hcraeser suoiverp ni nwohs neeb noitubirtsid eht fo pot eht ta dnert eht enimaxe )6(-)5( snmuloC .stnemtnioppa OEC edistuo fo elpmas-bus eht ni yap OEC ni dnert eht rof stluser ,yap OEC fo noitubirtsid laciripme eht fo eliced pot eht ni si noitasnepmoc latot esohw sOEC rof snoisserger elitnauq fo stluser stroper dna yap fo .yap OEC fo tnenopmoc ytiuqe eht ni dnert eht rof stluser troper )01(-)9( snmuloC .elitniuq pot eht ni sOEC rof stluser troper )8(-)7( snmuloC dna ni detroper era evitucexe yb snoitavresbo fo ecnednepedni-non rof detsujda srorre dradnats deretsulc tsuboR .C xidneppA ni era snoitin(cid:133)ed elbairaV .ylevitcepser ,level %01 dna ,%5 ,%1 eht ta ecnac(cid:133)ingis lacitsitats rof dna , , yb detoned era ecnac(cid:133)ingis fo sleveL .sesehtnerap (cid:3) (cid:3)(cid:3) (cid:3)(cid:3)(cid:3) pmoCucexE lla ;noitasnepmoc launna latot gol si elbairav tnednepeD :A lenaP OEC ni dnerT pot rof dnerT pot rof dnerT rof dnerT ni dnerT yap ytiuqe yap OEC %5 yap OEC %01 sOEC edistuO yap OEC )01( )9( )8( )7( )6( )5( )4( )3( )2( )1( snoitcaretnI dnerT snoitcaretnI dnerT snoitcaretnI dnerT snoitcaretnI dnerT snoitcaretnI dnerT ]747;21[ ]747;21[ ]747;21[ ]747;21[ ]747;21[ ]747;21[ ]385;2[ ]385;2[ ]747;21[ ]747;21[ snoitavresbO 010.0- 262.0 422.0 624.0 592.0 134.0 870.0 453.0 891.0 323.0 I (cid:3)(cid:3)(cid:3) (cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) 0002 6991 )170.0( )840.0( )321.0( )630.0( )660.0( )530.0( )131.0( )870.0( )930.0( )520.0( (cid:0) 832.0 894.0 583.0 794.0 774.0 665.0 922.0 625.0 983.0 805.0 I (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) 5002 1002 )870.0( )260.0( )080.0( )450.0( )660.0( )140.0( )241.0( )090.0( )340.0( )920.0( (cid:0) *rotcaF tnelaT OEC 254.0 254.0 192.0 644.0 152.0 I (cid:3)(cid:3)(cid:3) (cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3) (cid:3)(cid:3)(cid:3) 0002 6991 )211.0( )412.0( )401.0( )302.0( )160.0( (cid:0) *rotcaF tnelaT OEC 912.0 182.0 722.0 463.0 102.0 I (cid:3) (cid:3)(cid:3) (cid:3)(cid:3) (cid:3) (cid:3)(cid:3)(cid:3) 5002 1002 )621.0( )431.0( )390.0( )512.0( )930.0( (cid:0) )2 erunet( sOEC detnioppa yltnecer ;noitasnepmoc launna latot gol si elbairav tnednepeD :B lenaP (cid:20) ]831;3[ ]831;3[ ]831;3[ ]831;3[ ]831;3[ ]831;3[ ]631;1[ ]631;1[ ]831;3[ ]831;3[ snoitavresbO 720.0- 772.0 151.0 793.0 813.0 054.0 371.0- 103.0 400.0 652.0 I (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3) (cid:3)(cid:3)(cid:3) 0002 6991 )411.0( )270.0( )011.0( )011.0( )840.0( )460.0( )961.0( )551.0( )380.0( )350.0( (cid:0) 702.0 205.0 493.0 224.0 815.0 345.0 802.0- 915.0 690.0 383.0 I (cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) 5002 1002 )421.0( )390.0( )270.0( )621.0( )250.0( )680.0( )981.0( )961.0( )980.0( )060.0( (cid:0) *rotcaF tnelaT OEC 405.0 426.0 792.0 648.0 155.0 I (cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) 0002 6991 )312.0( )102.0( )390.0( )523.0( )061.0( (cid:0) *rotcaF tnelaT OEC 463.0 592.0 652.0 412.1 896.0 I (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) 5002 1002 )432.0( )980.0( )190.0( )553.0( )971.0( (cid:0) ,noisseccuS ,mriF seY seY seY seY seY seY seY seY seY seY slortnoC OEC & seY seY seY seY seY seY seY seY seY seY .E.F mriF seY seY seY seY seY seY seY seY seY seY .E.F raeY 56

)deunitnoC( 8 elbaT yaP OEC ni dnerT fo stcaF dezilytS rof snoitacilpmI :slaitnederC OEC rof yaP gnissessA yrtsudnI yb sisylanA rotacidni dnert emit htiw noitcaretni sti dna slaitnederc OEC fo erusaem a no yap OEC latot fo snoisserger SLO fo setamitse stroper elbat sihT .)1cdt( yap latot fo mhtiragol eht si elbairav tnedneped ehT .pmoCucexE ni sOEC lla rof 5002 ot 3991 morf spuorg yrtsudni daorb yb selbairav dna ,reeraC kcarT-tsaF ,sserP morf sisylana tnenopmoc lapicnirp gnisu detcartxe rotcaf a si rotcaF tnelaT OEC - slaitnederc OEC fo erusaem ehT llA .ylevitcepser ,5002 ot 1002 dna 0002 ot 6991 sraey ni eno fo eulav ekat taht seimmud era selbairav rotacidni dnert emit ehT .egelloC evitceleS suoiverp ni nwohs neeb evah taht scitsiretcarahc OEC rehto dna ,snoisseccus ,mr(cid:133) rof slortnoc sa llew sa ,stce⁄e dex(cid:133) mr(cid:133) edulcni snoitac(cid:133)iceps )4(-)3( snmuloC .)9993 dna 0002 neewteb sedoc CIS( rotces gnirutcafunam eht rof stluser troper )2(-)1( snmuloC .yap OEC latot tce⁄a ot hcraeser 0007 neewteb sedoc CIS( rotces secivres eht rof stluser troper )6(-)5( snmuloC .)9995 dna 0005 neewteb sedoc CIS( rotces liater eht rof stluser troper ,tnempiuqe lacidem ,scinortcele ,tnempiuqe retupmoc ,gnitupmoc ,hcetoib sa hcus( srotces hcet-hgih eht rof stluser troper )8(-)7( snmuloC .)9997 dna dna narhguoL( 478 dna ,378 ,737 ,483 ,383 ,283 ,183 ,763 ,663 ,753 ,382 :sedoc CIS-3 gniwollof eht ot dnopserroc hcihw ,erawtfos ,slacituecamrahp 0094 neewteb dna 9996 dna 0006 neewteb sedoc CIS ,seitilitu dna slaicnan(cid:133)( srotces detaluger rof stluser troper )01(-)9( snmuloC .))4002( rettiR era evitucexe yb snoitavresbo fo ecnednepedni-non rof detsujda srorre dradnats deretsulc tsuboR .C xidneppA ni era snoitin(cid:133)ed elbairaV .)9994 dna .ylevitcepser ,level %01 dna ,%5 ,%1 eht ta ecnac(cid:133)ingis lacitsitats rof dna , , yb detoned era ecnac(cid:133)ingis fo sleveL .sesehtnerap ni detroper (cid:3) (cid:3)(cid:3) (cid:3)(cid:3)(cid:3) pmoCucexE lla ;noitasnepmoc launna latot gol si elbairav tnednepeD :C lenaP detalugeR hceT-hgiH secivreS liateR gnirutcafunaM )01( )9( )8( )7( )6( )5( )4( )3( )2( )1( snoitcaretnI dnerT snoitcaretnI dnerT snoitcaretnI dnerT snoitcaretnI dnerT snoitcaretnI dnerT ]165;1[ ]165;1[ ]486;1[ ]486;1[ ]558[ ]558[ ]453;1[ ]453;1[ ]826;5[ ]826;5[ snoitavresbO 404.0 044.0 890.0 244.0 541.0 474.0 242.0 503.0 561.0 003.0 I (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) 0002 6991 )331.0( )250.0( )641.0( )280.0( )122.0( )441.0( )421.0( )760.0( )540.0( )820.0( (cid:0) 616.0 817.0 112.0 405.0 550.0 054.0 754.0 415.0 853.0 774.0 I (cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) 5002 1002 )82.0( )780.0( )641.0( )690.0( )612.0( )951.0( )831.0( )970.0( )150.0( )330.0( (cid:0) *rotcaF tnelaT OEC 790.0 307.0 976.0 642.0 772.0 I (cid:3)(cid:3)(cid:3) (cid:3)(cid:3) (cid:3)(cid:3)(cid:3) 0002 6991 )732.0( )242.0( )313.0( )961.0( )370.0( (cid:0) *rotcaF tnelaT OEC 573.0 296.0 117.0 702.0 171.0 I (cid:3)(cid:3)(cid:3) (cid:3)(cid:3) (cid:3)(cid:3) 5002 1002 )567.0( )952.0( )643.0( )881.0( )480.0( (cid:0) ,noisseccuS ,mriF seY seY seY seY seY seY seY seY seY seY slortnoC OEC & seY seY seY seY seY seY seY seY seY seY .E.F mriF seY seY seY seY seY seY seY seY seY seY .E.F raeY 57

9 elbaT ?ecneirepxE kroW emitefiL ro tnelaT :slaitnederC OEC rof yaP gniterpretnI latipaC namuH OEC tsilaicepS .sv tsilareneG htiw noitairaV ot 3991 morf ecneirepxe krow emitefil OEC dna slaitnederc OEC fo serusaem no yap OEC latot fo snoisserger SLO fo setamitse stroper elbat sihT )84 hcnerF-amaF( dna -raey edulcni snoitac(cid:133)iceps llA .)1cdt( yap latot fo mhtiragol eht si elbairav tnedneped ehT .sOEC detnioppa ylwen rof 5002 latot tce⁄a ot hcraeser suoiverp ni nwohs neeb evah taht scitsiretcarahc OEC rehto dna ,snoisseccus ,mr(cid:133) rof slortnoc sa llew sa ,stce⁄e dex(cid:133)-yrtsudni lapicnirp gnisu detcartxe rotcaf a si hcihw - rotcaF tnelaT OEC slaitnederc OEC fo erusaem niam ruo rof stluser tneserp )4(-)1( snmuloC .yap OEC krow emitefil OEC fo seixorp tnere⁄id eerht rof ylevitareti lortnoc ew nehw ,egelloC evitceleS dna ,reeraC kcarT-tsaF ,sserP morf sisylana tnenopmoc eerht eht morf sisylana tnenopmoc lapicnirp gnisu detcartxe rotcaf a si hcihw rotcaF ytilibA lareneG OEC a rof dna ))3(-)1( snmuloC( ecneirepxe namuH OEC owt sedulcni taht noitac(cid:133)iceps a rof stluser stneserp )5( nmuloC .))1102( sotaM dna ,arierreF ,oidotsuC( seixorp ecneirepxe gniylrednu slaitnederc OEC eerht ruo gnisu morf detcartxe stnenopmoc lapicnirp owt tsr(cid:133) eht era hcihw ,)"tnelaT" ,2# dna "ecneirepxE" ,1#( srotcaF latipaC latipaC namuH OEC owt eht neewteb snoitcaretni redisnoc )7(-)6( snmuloC .seixorp ecneirepxe krow emitefil OEC eerht eht htiw yltnioj seixorp noisserger enilesab ruo nur ew ,os od oT .asreveciv dna ecneirepxe OEC no gnidneped slaitnederc OEC rof yap ni ytienegoreteh rof wolla ot srotcaF woleb si ",ecneirepxE" ,1# rotcaF latipaC namuH esohw sOEC esoht( ecneirepxe wol htiw sOEC detnioppa ylwen fo elpmas-bus eht ni yletarapes ;naidem woleb si ",tnelaT" ,2# rotcaF latipaC namuH esohw sOEC esoht( slaitnederc wol htiw sOEC detnioppa ylwen dna ))6( nmuloC ;naidem yb snoitavresbo fo ecnednepedni-non rof detsujda srorre dradnats deretsulc tsuboR .C xidneppA ni era snoitin(cid:133)ed elbairaV .ylevitcepser ,))7( nmuloC ,level %01 dna ,%5 ,%1 eht ta ecnac(cid:133)ingis lacitsitats rof dna , , yb detoned era ecnac(cid:133)ingis fo sleveL .sesehtnerap ni detroper era evitucexe (cid:3) (cid:3)(cid:3) (cid:3)(cid:3)(cid:3) ylevitcepser ylno raey tnemtnioppa ;noitasnepmoc launna latot gol :elbairav tnednepeD snoitcaretnI ecneirepxE kroW OEC rof gnillortnoC )7( )6( )5( )4( )3( )2( )1( slaitnederC woL ecneirepxE woL ylnO sOEC ylnO sOEC :slaitnederC OEC 373.0 273.0 873.0 493.0 rotcaF tnelaT OEC (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) )101.0( )201.0( )101.0( )990.0( 995.0 143.0 )"tnelaT"( 2# rotcaF latipaC namuH OEC (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) )261.0( )611.0( :ecneirepxE kroW OEC 471.0 noitisop OEC tsaP (cid:3)(cid:3) )070.0( 323.0 sboj fo rebmun tsaP (cid:3)(cid:3)(cid:3) )790.0( 373.0 seirtsudni fo rebmun tsaP (cid:3)(cid:3)(cid:3) )690.0( 473.0 rotcaF ytilibA lareneG OEC (cid:3)(cid:3)(cid:3) )690.0( 004.0 782.0 )"ecneirepxE"( ,1# rotcaF latipaC namuH OEC (cid:3)(cid:3) (cid:3)(cid:3)(cid:3) )171.0( )080.0( seY seY seY seY seY seY seY slortnoC OEC & ,noisseccuS ,mriF seY seY seY seY seY seY seY .E.F raeY seY seY seY seY seY seY seY .E.F yrtsudnI %3.44 %7.44 %1.34 %9.24 %7.24 %9.24 %5.24 2R 909 909 818,1 818,1 818,1 818,1 818,1 snoitavresbO 58

01 elbaT ?epyH ro tnelaT :slaitnederC OEC rof yaP gniterpretnI snoisiceD OEC dna ecnamrofreP mriF mreT-gnoL fo sisylanA slaitnederc OEC fo serusaem no ecnamrofrep mr(cid:133) gnitarepo mret-gnol fo serusaem fo snoisserger SLO fo setamitse stroper elbat siht fo A lenaP mr(cid:133) gnitarepo mret-gnol detsujda-yrtsudni ni segnahc era )7(-)2( snmuloC ni selbairav tnedneped llA .sOEC detnioppa ylwen rof 5002 ot 3991 morf OEC eht ot tneuqesbus sraey eerht eht ni ecnamrofrep detsujda-yrtsudni launna egareva neewteb ecnere⁄id eht sa detaluclac era hcihw ,ecnamrofrep erusaem a yolpme eW .)dedulcxe era sraey tnemtnioppa( noitisnart eht ot roirp raey eht ni ecnamrofrep detsujda-yrtsudni launna dna tnemtnioppa evitceleS dna ,reeraC kcarT-tsaF ,sserP morf sisylana tnenopmoc lapicnirp gnisu detcartxe rotcaf a si hcihw - rotcaF tnelaT OEC - slaitnederc OEC fo rehto dna ,snoisseccus ,mr(cid:133) rof slortnoc emas eht sa llew sa ,stce⁄e dex(cid:133)-yrtsudni )84 hcnerF-amaF( dna -raey edulcni snoitac(cid:133)iceps llA .egelloC edulcni osla snoitac(cid:133)iceps lla ,noisrever-naem rof lortnoc ot redro nI .)3 elbaT( yap OEC fo sisylana noisserger enilesab eht ni sa scitsiretcarahc OEC snruter lamronba evitalumuc nur-trohs si )1( nmuloC ni elbairav tnedneped ehT .noitisnart ot roirp sraey eerht eht ni ecnamrofrep launna egareva sisylana fo wodniw )2+,2-( ehT .)MPAC( ledom gnicirp tessa latipac eht gnisu detaluclac era snruter lamronbA .stnemtnioppa OEC dnuora )sRAC( ni selbairav tnedneped ehT .tnemecnuonna eht fo yad eht si 0=t erehw ,)syad ni( stnemtnioppa OEC fo setad tnemecnuonna lautca ot evitaler si ,)EOR( ytiuqe no nruter ,)SORO( selas no nruter gnitarepo ,)AORO( stessa no nruter gnitarepo ,)AOR( stessa ot emocni ten era )7(-)2( snmuloC dna sRAC tnemtnioppa neewteb mret noitcaretni na dna sRAC tnemtnioppa sdda )8( nmuloC .ylevitcepser ,swo(cid:135) hsac dna ,snruter tekram kcots .C xidneppA ni era snoitin(cid:133)ed elbairaV .mret noitcaretni eht fo etamitse eht stroper dna )2( nmuloC ni noitac(cid:133)iceps eht ot rotcaF tnelaT OEC eht era ecnac(cid:133)ingis fo sleveL .sesehtnerap ni detroper era evitucexe yb snoitavresbo fo ecnednepedni-non rof detsujda srorre dradnats deretsulc tsuboR eht ot tcepser htiw detaulave si ytivitisnes deilpmI .ylevitcepser ,level %01 dna ,%5 ,%1 eht ta ecnac(cid:133)ingis lacitsitats rof dna , , yb detoned (cid:3) (cid:3)(cid:3) (cid:3)(cid:3)(cid:3) .noitisnart eht ot roirp raey eht ni erusaem ecnamrofrep gnitarepo evitcepser eht fo naem elpmas )erofeb raey 1 -retfa egareva sraey 3( ecnamrofreP mriF mreT-gnoL fo sisylanA :A lenaP )8( )7( )6( )5( )4( )3( )2( )1( sRAC-AOR hsaC kcotS EOR SORO AORO AOR -tnioppA noitalerroC swolF snruteR sRAC tnem :slaitnederC OEC 304.0 502.0 890.0 940.0 440.0 240.0 430.0 810.0 rotcaF tnelaT OEC (cid:3)(cid:3)(cid:3) (cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3) (cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3) )251.0( )990.0( )930.0( )420.0( )020.0( )410.0( )210.0( )900.0( & ,noisseccuS ,mriF seY seY seY seY seY seY seY seY slortnoC OEC rehtO seY seY seY seY seY seY seY seY .E.F raeY seY seY seY seY seY seY seY seY .E.F yrtsudnI %2.61 %2.7 %7.51 %1.9 %7.8 %7.9 %2.11 %5.7 2R 178 817 677 418 788 198 178 1771 snoitavresbO :)slaitnederC %1-nruter naem %( ytivitisneS laitnederC-ecnamrofreP deilpmI 8.0 5.1 9.1 1.2 9.2 0.2 rotcaF tnelaT OEC 59

)deunitnoC( 01 elbaT ?epyH ro tnelaT :slaitnederC OEC rof yaP gniterpretnI snoisiceD OEC dna ecnamrofreP mriF mreT-gnoL fo sisylanA ylwen rof 5002 ot 3991 morf slaitnederc OEC fo serusaem no seicilop mr(cid:133) fo serusaem fo snoisserger SLO fo setamitse stroper elbat siht fo B lenaP ecnere⁄id eht sa detaluclac era hcihw ,seicilop mr(cid:133) detsujda-yrtsudni ni segnahc era )8(-)1( snmuloC ni selbairav tnedneped llA .sOEC detnioppa ni ycilop detsujda-yrtsudni launna dna tnemtnioppa OEC eht ot tneuqesbus sraey eerht eht ni ycilop mr(cid:133) detsujda-yrtsudni launna egareva neewteb rotcaf a si hcihw rotcaF tnelaT OEC - slaitnederc OEC fo erusaem a yolpme eW .)dedulcxe era sraey tnemtnioppa( noitisnart eht ot roirp raey eht hcnerF-amaF( dna -raey edulcni snoitac(cid:133)iceps llA .egelloC evitceleS dna ,reeraC kcarT-tsaF ,sserP morf sisylana tnenopmoc lapicnirp gnisu detcartxe OECfosisylananoissergerenilesabehtnisascitsiretcarahcOECrehtodna,snoisseccus,mr(cid:133)rofslortnocemasehtsallewsa,stce⁄edex(cid:133)-yrtsudni)84 ehT .noitisnartotroirpsraeyeerhtehtniycilopmr(cid:133)launnaegarevaedulcnioslasnoitac(cid:133)icepslla,noisrever-naemroflortnocotredronI .)3elbaT(yap rebmun eht ,reriuqca na sa detelpmoc sah mr(cid:133) eht snoitcasnart A&M fo rebmun eht ,serutidnepxe latipac era )8(-)1( snmuloC ni selbairav tnedneped sah mr(cid:133) eht snoitcasnart A&M gniyfisrevid fo rebmun eht ,sdnedivid ,sgnidloh hsac ,egarevel koob ,mr(cid:133) eht yb detelpmoc snoitcasnart erutitsevid fo fo ecnednepedni-non rof detsujda srorre dradnats deretsulc tsuboR .C xidneppA ni era snoitin(cid:133)ed elbairaV .ylevitcepser ,reriuqca na sa detelpmoc ,%5 ,%1 eht ta ecnac(cid:133)ingis lacitsitats rof dna , , yb detoned era ecnac(cid:133)ingis fo sleveL .sesehtnerap ni detroper era evitucexe yb snoitavresbo (cid:3) (cid:3)(cid:3) (cid:3)(cid:3)(cid:3) .ylevitcepser ,level %01 dna )erofeb raey 1 -retfa egareva sraey 3( snoisiceD OEC fo sisylanA :B lenaP ygetartS lanoitazinagrO yciloP laicnaniF yciloP tnemtsevnI )8( )7( )6( )5( )4( )3( )2( )1( D&R gniyfisreviD sdnediviD hsaC egareveL serutitseviD sA&M XEPAC sA&M sgnidloH :slaitnederC OEC 300.0 380.0- 500.0- 830.0 240.0- 101.0 141.0- 310.0rotcaF tnelaT OEC (cid:3)(cid:3) (cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3) )800.0( )230.0( )200.0( )210.0( )610.0( )940.0( )350.0( )600.0( & ,noisseccuS ,mriF seY seY seY seY seY seY seY seY slortnoC OEC rehtO seY seY seY seY seY seY seY seY .E.F raeY seY seY seY seY seY seY seY seY .E.F yrtsudnI %4.8 %7.22 %4.52 %1.7 %3.71 %2.91 %4.91 %5.22 2R 878 878 367 878 878 878 878 878 snoitavresbO 60

11 elbaT ?rewoP ro tnelaT :slaitnederC OEC rof yaP gniterpretnI snoisiceD gnirotinoM draoB morf ecnedivE dna ecnanrevoG htiw noitairaV sOEC detnioppa ylwen rof 5002 ot 3991 morf slaitnederc OEC fo erusaem a no yap OEC latot fo snoisserger SLO fo setamitse stroper elbat sihT 5002 ot 3991 morf slaitnederc OEC fo erusaem a no revonrut OEC decrof fo doohilekil eht fo snoisserger tiborp fo setamitse dna ,))6(-)1( snmuloC( elbairav ymmud a dna )6(-)1( snmuloC ni )1cdt( yap latot fo mhtiragol eht si elbairav tnedneped ehT .))8(-)7( snmuloC( pmoCucexE eritne eht rof fo erusaem a yolpme ew snoitac(cid:133)iceps lla nI .)8(-)7( snmuloC rof srucco revonrut OEC decrof a nehw raey-mr(cid:133) nevig yna ni eno fo eulav sekat taht evitceleS dna ,reeraC kcarT-tsaF ,sserP morf sisylana tnenopmoc lapicnirp gnisu detcartxe rotcaf a si hcihw rotcaF tnelaT OEC slaitnederc OEC taht scitsiretcarahc OEC rehto dna ,snoisseccus ,mr(cid:133) rof slortnoc sa llew sa ,stce⁄e dex(cid:133)-yrtsudni )84 hcnerF-amaF( dna -raey edulcni dna ,egelloC rof slortnoc sedulcni taht noitac(cid:133)iceps a rof setamitse enilesab stneserp )1( nmuloC .yap OEC latot tce⁄a ot hcraeser suoiverp ni nwohs neeb evah draob dna ,ezis draob ,))3002( kcirteM dna ,iihsI ,srepmoG( sesnefed revoekat-itna fo xednI MIG eht edulcni taht scitsiretcarahc ecnanrevog mr(cid:133) eht neewteb snoitcaretni dda ylevitareti )4(-)3( snmuloC .snoitcennoc etaroproc dna noitacude OEC rof slortnoc sdda )2( nmuloC .ecnednepedni OEC rof yap ni ytienegoreteh rof wolla ot selbairav snoitcennoc OEC eht htiw snoitcaretni rieht sa llew sa xedni MIG eht dna rotcaF tnelaT OEC rehto htiw snoitcaretni redisnoc )6(-)5( snmuloC .snoitcennoc OEC fo ytisnetni eht dna ecnanrevog mr(cid:133) fo ytilauq eht no gnidneped slaitnederc hcihw ,smr(cid:133) gnimrofreprednu fo selpmas-bus tnere⁄id rof doohilekil revonrut OEC decrof fo setamitse tneserp )8(-)7( snmuloC .selbairav ecnanrevog ni ecnamrofrep fo ))8( nmuloC( elitniuq mottob eht ni ro ,))7( nmuloC( naidem woleb saw raey roirp eht ni ecnamrofrep esohw smr(cid:133) sa den(cid:133)ed era evitucexe yb snoitavresbo fo ecnednepedni-non rof detsujda srorre dradnats deretsulc tsuboR .C xidneppA ni era snoitin(cid:133)ed elbairaV .yrtsudni rieht .ylevitcepser ,level %01 dna ,%5 ,%1 eht ta ecnac(cid:133)ingis lacitsitats rof dna , , yb detoned era ecnac(cid:133)ingis fo sleveL .sesehtnerap ni detroper era (cid:3) (cid:3)(cid:3) (cid:3)(cid:3)(cid:3) decroF fo sisylanA launna latot gol :elbairav tnednepeD revonruT OEC ylno raey tnemtnioppa ;noitasnepmoc )8( )7( )6( )5( )4( )3( )2( )1( sOEC gnimrofreprednU mottoB woleB elitniuQ naideM 170.0 320.0 167.0 121.0 408.0 408.0 305.0 025.0 rotcaF tnelaT OEC (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) )720.0( )700.0( )231.0( )051.0( )592.0( )592.0( )211.0( )031.0( :snoitcennoC OEC & ecnanrevoG 322.0 krowteN noitacudE OEC )831.0( 145.0 krowteN etaroproC OEC (cid:3)(cid:3)(cid:3) )411.0( (cid:3) 008.0- (cid:3)(cid:3)(cid:3) 777.0- )11>(MIG*rotcaF tnelaT )434.0( )883.0( 590.0krowteN noitacudE*)11>(MIG*rotcaF tnelaT )504.0( 483.0 krowteN etaroproC*)11>(MIG*rotcaF tnelaT )164.0( 705.0 ecnednepednI draoB*rotcaF tnelaT (cid:3)(cid:3)(cid:3) )181.0( 405.0tnemtnioppA edisnI*rotcaF tnelaT (cid:3)(cid:3)(cid:3) )331.0( seY seY seY seY seY seY seY seY slortnoC ecnanrevoG oN oN oN oN seY oN seY oN slortnoC krowteN seY seY seY seY seY seY seY seY slortnoC OEC & ,noisseccuS ,mriF seY seY seY seY seY seY seY seY .E.F raeY seY seY seY seY seY seY seY seY .E.F yrtsudnI %2.02 %6.81 %3.24 %4.34 %9.84 %5.94 1.04 %2.14 2R 945,2 373,6 523,1 523,1 495 266 490,1 523,1 snoitavresbO 61

21 elbaT stseT ssentsuboR lanoitiddA :slaitnederC OEC rof yaP sesylanA noitceleS namkceH dna elpmaS dehctaM tnioj dna )A lenaP( sOEC detnioppa ylwen rof 5002 ot 3991 morf slaitnederc OEC rof yap fo sisylana elpmas-dehctam fo stluser stroper elbat sihT htob nI .)B lenaP( pmoCucexE eritne eht rof )sisylana noitceles namkceH( doohilekil noisseccus OEC dna slaitnederc OEC rof yap fo noitamitse slaitnederc OEC fo erusaem eht ,A.1 lenaP nI .)1cdt( yap latot fo mhtiragol eht si noitamitse egats dnoces eht fo elbairav tnedneped eht ,slenap gnisu detcartxe rotcaf a si hcihw ,rotcaf tnelat OEC eht fo elitrauq pot eht ni era slaitnederc esohw sOEC rof eno fo eulav sekat taht ymmud a si noisserger tiborp egats-tsr(cid:133) eht fo stluser eht stroper 2 nmuloC .egelloC evitceleS dna ,reeraC kcarT-tsaF ,sserP morf sisylana tnenopmoc lapicnirp dna ,noisseccus ,mr(cid:133) rof slortnoc emas eht htiw hctam erocs ytisneporp robhgien-tseraen a gnisu enod si hcihw ,elpmas lortnoc eht tcurtsnoc ot desu dex(cid:133)raeydna ,-yrtsudni)84hcnerF-amaF(dna ,)ytiverbrofdettimostneic ¢eocemos ,3elbaT(sisylananoissergerniamehtnisascitsiretcarahcOEC neewteb secnere⁄id rof tnuocca ot detsujda-saib ,puorg lortnoc )dehctam( eht dna tnemtaert eht neewteb ecnere⁄id eht stroper 1 nmuloC .stce⁄e fo erusaem eht ,B.1 lenaP nI .hctam tseraen rieht dna rotcaf tnelat OEC eht fo elitrauq pot eht ni sOEC detnioppa ylwen fo serocs ytisneporp eht pmoCucexE ni mr(cid:133) a taht doohilekil eht fo noisserger tiborp egats-tsr(cid:133) eht fo stluser stroper 4 nmuloC .rotcaf tnelat OEC eht si slaitnederc OEC egats-tsr(cid:133) eht erehw ,yap OEC latot fo ledom noitceles pets-owt namkceH a rof stluser stroper 3 nmuloC .raey nevig a ni noisseccus OEC a seogrednu emos ,3 elbaT( sisylana noisserger niam eht ni sa slortnoc emas eht ot noitidda nI .4 nmuloC morf setamitse tiborp eht yb nevig si noitauqe noitceles egats-tsr(cid:133) eht ,stce⁄e dex(cid:133) raey dna ,-yrtsudni )84 hcnerF-amaF( dna ,)egats-tsr(cid:133) eht morf dettimo revonrut decrof ,ytiverb rof dettimo stneic ¢eoc ot roirp sraey owt eht ni OEC eht fo htaed ro )56 ega OEC( tnemeriter a saw ereht rehtehw rof elbairav rotacidni na sedulcni noitauqe noitceles (cid:21) dradnats deretsulc tsuboR .C xidneppA ni era snoitin(cid:133)ed elbairaV .noisserger egats-dnoces eht morf dedulcxe si elbairav sihT .raey lacs(cid:133) tnerruc eht rof dna , , yb detoned era ecnac(cid:133)ingis fo sleveL .sesehtnerap ni detroper era evitucexe yb snoitavresbo fo ecnednepedni-non rof detsujda srorre (cid:3) (cid:3)(cid:3) (cid:3)(cid:3)(cid:3) .ylevitcepser ,level %01 dna ,%5 ,%1 eht ta ecnac(cid:133)ingis lacitsitats ylno raey tnemtnioppa ;noitasnepmoc launna latot gol :elbairav tnednepeD ]rotcaF tnelaT OEC fo noitceleS[ sisylanA elpmaS dehctaM : A.1 lenaP )2( noitauqE noitceleS )1( setamitsE egatS-dnoceS 260.0 eziS mriF 106.0 rotcaF tnelaT OEC (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) )900.0( )531.0( )ymmuD elitrauQ poT( 320.0ecnamrofreP mriF roirP )820.0( 031.0 noisseccuS decroF (cid:3)(cid:3)(cid:3) )040.0( ,noisseccuS ,mriF seY slortnoC OEC & seY .E.F raeY seY .E.F yrtsudnI %6.21 2R 134 .sbO detaerT 177,1 snoitavresbO 424 .sbO lortnoC ]elpmaS tnemtnioppA fo noitceleS[ sisylanA noitceleS namkceH : B.1 lenaP )4( noitauqE noitceleS )3( setamitsE egatS-dnoceS 473.0 OEC deriteR ro desaeceD 545.0 rotcaF tnelaT OEC (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) )950.0( ]elbairaV dedulcxE[ )011.0( 230.0 eziS mriF 834.0 oitaR slliM esrevnI (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) )010.0( )211.0( 762.0ecnamrofreP mriF roirP (cid:3)(cid:3)(cid:3) )740.0( 820.0 ega OEC (cid:3)(cid:3)(cid:3) )400.0( ,noisseccuS ,mriF seY slortnoC OEC & seY .E.F raeY seY .E.F raeY seY .E.F yrtsudnI seY .E.F yrtsudnI %6.51 2R 747,21 snoitavresbO 177,1 snoitavresbO 62

Table 12 (Continued) Pay for CEO Credentials: Additional Robustness Tests Additional Controls and Di⁄erent De(cid:133)nitions of CEO Credentials Proxies This table reports estimates of OLS regressions of total CEO pay on measures of CEO credentials from 1993 to 2005 for newly appointed CEOs. The dependent variable is the logarithm of total pay (tdc1). We iteratively employ the three measures of CEO credentials - Press, Fast-Track Career, and Selective College - in a series of robustness tests. All speci(cid:133)cations include year- and (Fama-French 48) industry-(cid:133)xed e⁄ects, as well as controls for (cid:133)rm, successions, and other CEO characteristics that have been shown in previous research to a⁄ect total CEO pay. Variable de(cid:133)nitions are in Appendix C. Robust clustered standard errors adjusted for non-independence of observations by executive are reported in parentheses. Levels of signi(cid:133)cance are denoted by , , and for (cid:3)(cid:3)(cid:3) (cid:3)(cid:3) (cid:3) statistical signi(cid:133)cance at the 1%, 5%, and 10% level, respectively. Panel 2: Dependent variable: log total annual compensation; appointment year only (1) (2) (3) Press Fast-Track Selective Career College [1] Press-Bad Press 0.614 (cid:3)(cid:3)(cid:3) (0.100) [2] (Press-Bad Press)/Press 0.411 (cid:3)(cid:3) (0.181) [3] Good Press 0.828 (cid:3)(cid:3)(cid:3) (0.167) [4] Good Press/Press 0.870 (cid:3)(cid:3)(cid:3) (0.260) [5] Past 3 Yrs Mean Press 0.561 (cid:3)(cid:3)(cid:3) (0.112) [6] Firm Size-Adjusted Press 0.524 (cid:3)(cid:3)(cid:3) (0.086) [7] First CEO job is not 0.520 (cid:3)(cid:3) current CEO appointment (0.204) [8] Selective is Most Competitive Colleges Only 0.190 (cid:3)(cid:3)(cid:3) (33 Institutions) (0.070) [9] Includes no college & 0.172** foreign institutions (0.078) [10] Industry-Adjusted 0.526 0.430 0.181** (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (0.090) (0.158) (0.089) [11] Controlling for MBA 0.546 0.435 0.201 (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3) (0.089) (0.164) (0.089) [12] Controlling for higher (3rd) 0.550 0.515 0.200** (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) order (cid:133)rm size splines (0.093) (0.176) (0.089) [13] Controlling for headquarter 0.512 0.535 0.191 (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3) location (state) (cid:133)xed e⁄ects (0.095) (0.179) (0.095) 63

Figure 1 Pay for CEO Credentials: New CEOs(cid:146)Pay and Press Coverage This (cid:133)gure plots the logarithm of total CEO pay (TDC1) against the distribution of Press quantiles for newlyappointed CEOs from 1993 to 2005. Variable de(cid:133)nitions are in Appendix C. 5.9 9 5.8 8 5.7 7 0 .2 .4 .6 .8 1 New CEO Press 95% CI Log Total Compensation Figure 2 CEO Credentials and Firm Performance This (cid:133)gure plots median industry-adjusted operating return on assets (OROA) around CEO succession events from 1993 to 2005. The dotted line refers to the entire sample, while the thin (bold) line is for the sub-sample of successions involving newly-appointed CEOs in the top (bottom) quartile of Press. Variable de(cid:133)nitions are in Appendix C. 64