Quantifying Managerial Ability

8/13/2019 Quantifying Managerial Ability

1/20

MANAGEMENT SCIENCEVol. 58, No. 7, July 2012, pp. 12291248ISSN 0025-1909 (print) ISSN 1526-5501 (online) http://dx.doi.org/10.1287/mnsc.1110.1487

2012 INFORMS

Quantifying Managerial Ability:

A New Measure and Validity TestsPeter Demerjian

Goizueta Business School, Emory University, Atlanta, Georgia 30322, [email protected]

Baruch LevStern School of Business, New York University, New York, New York 10012, [email protected]

Sarah McVayDavid Eccles School of Business, University of Utah, Salt Lake City, Utah 84112, [email protected]

We propose a measure of managerial ability, based on managers efficiency in generating revenues, which isavailable for a large sample of firms and outperforms existing ability measures. We find that our measureis strongly associated with manager fixed effects and that the stock price reactions to chief executive officer

(CEO) turnovers are positive (negative) when we assess the outgoing CEO as low (high) ability. We also findthat replacing CEOs with more (less) able CEOs is associated with improvements (declines) in subsequent firmperformance. We conclude with a demonstration of the potential of the measure. We find that the negativerelation between equity financing and future abnormal returns documented in prior research is mitigated bymanagerial ability. Specifically, more able managers appear to utilize equity issuance proceeds more effectively,illustrating that our more precise measure of managerial ability will allow researchers to pursue studies thatwere previously difficult to conduct.

Key words : managerial ability; managerial talent; managerial efficiencyHistory : Received November 18, 2009; accepted October 30, 2011, by Mary E. Barth, accounting. Published

online in Articles in Advance March 9, 2012.

1. Introduction

Quantifying managerial ability, or talent, is centralto many important research questions, such as thoseexamining managerial contributions to firm perfor-mance and investment decisions, executive compen-sation, corporate governance, economic effects ofcorporate ownership, and cross-country productiv-ity differences. Prior research indicates that manager-specific features (ability, talent, reputation, or style)affect economic outcomes and are therefore importantto economics, finance, accounting, and managementresearch as well as to practice.1 To infer manage-rial ability, researchers generally rely on proxies suchas firm size, past abnormal performance, compen-

sation, tenure, media mentions, education, or man-ager fixed effects. Researchers have also inferred theability of managers using data envelopment analy-sis (DEA) within specific industries (e.g., Leverty andGrace 2012). Most of these measures, however, alsoreflect significant aspects of the firm that are outside

1 For example, Bertrand and Schoar (2003) document that managersexhibit styles that are reflected in the underlying decisions of thecompany (e.g., aggressive R&D investment or merger and acquisi-tion activity); see 2 for additional studies.

of managements control. For example, media men-

tions are more prevalent for large firms, and abnormalstock returns are affected by many factors other thanmanagerial ability. Similarly, although manager fixedeffects are more directly attributable to management,they can be applied only to a relatively small sampleof firms and do not offer a stand-alone measure ofability.

We introduce a new measure of managerial abilitybased on managers efficiency, relative to their indus-try peers, in transforming corporate resources to rev-enues. We consider a multitude of revenue-generatingresources: cost of inventory, general and adminis-trative expenses, fixed assets, operating leases, past

research and development (R&D) expenditures, andintangible assets. We expect more able managers to

better understand technology and industry trends,reliably predict product demand, invest in highervalue projects, and manage their employees more effi-ciently than less able managers. In short, we expectmore able managers to generate higher revenue fora given level of resources or, conversely, to mini-mize the resources used for a given level of rev-enue (i.e., to maximize the efficiency of the resourcesused). Assessing managers based on the efficiency

1229


2/20

Demerjian, Lev, and McVay: Quantifying Managerial Ability1230 Management Science 58(7), pp. 12291248, 2012 INFORMS

with which they generate revenues, rather than bytheir pay or media mentions, is intuitively appeal-ing as it is more in line with the overarching goal ofprofit-maximizing firms.

We use DEA to create an initial measure of therelative efficiency of the firm within its industry.2

We form an efficient frontier by measuring theamount and mix of resources used to generate rev-enue by the firms within each industry. Firms oper-ating on the frontier are assigned a score of one;the lower the firms score, the further it is from thefrontier.

This firm efficiency measure, however, is affectedby both firm-specific factors and management char-acteristics, a limitation that also applies to othermanagerial ability proxies frequently used in the lit-erature, such as past stock returns, as well as toconventional efficiency measures, such as return onassets (ROA). For example, a mediocre manager of

a large company will be able to negotiate betterterms with suppliers than an outstanding managerof a small company. We therefore modify the DEA-generated firm efficiency measure by purging it ofkey firm-specific characteristics that we expect toaid or hinder managements efforts. We do this byremoving from the total firm efficiency measure theeffects of firm size, market share, positive free cashflow, and firm age (all aiding management), as wellas complex multi-segment and international opera-tions (challenges to management). We also removethe effects of industry and time in the estimation.After controlling for the above, we attribute the unex-

plained portion of firm efficiency to management.This unexplained portion may still contain otherunidentified drivers of firm efficiency, and we con-duct a number of validity tests to assess whether themeasure reflects managerial ability.3

Although our measure of managerial ability is pos-itively correlated with a number of alternative mea-sures of ability (historical industry-adjusted stockreturns, historical industry-adjusted return on assets,chief executive officer (CEO) pay, and CEO tenure),

2 DEA is an optimization procedure used to evaluate the relativeefficiency of decision-making units. In 3 we provide additional

detail on DEA and discuss the merits of DEA over conventionalratios (i.e., return on assets) and a regression analysis of the samevariables.3 We attribute our managerial ability measure to the managementteam, though in tests examining a specific manager, we focus on theCEO, who is the most powerful manager and thus, on average, themost likely to affect outcomes (Fee and Hadlock 2003). Although anumber of papers have used DEA to measure firm efficiency (e.g.,Thore et al. 1994; Murthi et al. 1996, 1997; Barr and Siems 1997;Berk and Green 2004; Berk and Stanton 2007; Leverty and Grace2012), we are the first to measure efficiency for a large cross sectionof firms, spanning most industries, and parse out key firm-specificdrivers of efficiency to focus on managerial ability.

we show that it dominates these alternative measuresin a number of ways. First, for a subset of 78 CEOswho switch firms within our sample, we examinethe explanatory power of manager fixed effects inexplaining managerial ability. We find that 60.5% ofthese manager fixed effects are statistically signifi-

cant in explaining managerial ability after controllingfor firm fixed effects, and that this is 31.4 percent-age points higher than the statistical significance offirm fixed effects (where only 29.1% are statisticallysignificant). These results indicate that our proposedability measure reflects, to a large extent, individ-ual managers, although firm fixed effects continueto have explanatory power, albeit weaker. Moreover,manager fixed effects are notably higher than firmfixed effects only for our ability score. For the alter-native measures, the range varies from 7.5% (histori-cal returns) to 36% (historical ROA). Second, for asample of 2,229 firms experiencing a CEO turnover

during our sample period, we correlate managerialability and announcement returns to CEO departures.We find that the turnover announcements of outgo-ing CEOs with low (high) ability are associated withpositive (negative) stock price reactions. We do notfind a similar association for any of the alternativemeasures of ability. Third, for CEOs switching firmswithin our sample, we examine performance changesat the CEOs new firm. Specifically, we documentthat when a firm hires a more able CEO (relativeto the outgoing CEO), firm performance improvesover the next three years (measured using changesin both industry-adjusted stock returns and industry-

adjusted return on assets). Our measure again outper-forms the alternative measures in this setting. Thus,although our proposed measure encompasses someaspects that are not directly attributable to managerialability, the partitioning is more precise than that ofexisting measures, and it appears to capture an eco-nomically significant manager-specific component ofability.

We conclude our study by demonstrating that ourmore precise measure of managerial ability can helpresolve extant research puzzles. We examine whethermanagerial ability plays a role in the new issue puzzledocumented by Loughran and Ritter (1995). Specif-

ically, seasoned equity offerings are associated withfuture negative abnormal returns, and Loughran andRitter (1997) suggest that this is, in part, a result ofoverly optimistic managers investing the proceeds innegative net present value projects. We hypothesizeand find evidence suggesting that more able man-agers utilize the proceeds from equity financing moreeffectively, thereby largely mitigating the previouslydocumented negative future abnormal returns.

In sum, we propose a measure of managerial abilitythat is based on easily obtainable financial data and


3/20


4/20


Regardless, it is clear from these studies that man-agers affect the firm.4

The contribution of this study is to advance amore precise measure of managerial ability; thus,we hypothesize that our proposed measure reflectsmanagerial ability and outperforms existing ability

measures. We test this hypothesis three ways. First,following Bertrand and Schoar (2003), we expect thatthe proposed measure will be economically and sig-nificantly associated with manager fixed effects. Sec-ond, in the vein of Hayes and Schaefer (1999), weexpect that the proposed measure will be negativelyassociated with the announcement returns to CEOturnovers. Specifically, we expect that the turnoverannouncement of low-ability (high-ability) managerswill lead to a positive (negative) price reaction. Third,along the lines of Bennedsen et al. (2010), we expectthat relative changes in ability will be associated withchanges in subsequent performance (e.g., Carter et al.

2010, Cazier and McInnis 2010). Moreover, we expectour managerial ability measure to outperform thealternative ability measures in each of these respects.

3. Generation of FirmEfficiency Measure UsingData Envelopment Analysis

3.1. OverviewIn this section we briefly describe the DEA method-ology (Charnes et al. 1978, Banker et al. 1984) andcompare it to more conventional efficiency measures.

We use DEA to generate a measure of firm effi-ciency; this is the first step in generating our man-agerial ability measure, which is the residual oftotal firm efficiency after removing a number offirm-specific characteristics. DEA has been used tomeasure efficiency across multiple disciplines. Forexample, Murthi et al. (1996) use DEA analysis toassess marketing efficiency, and Leverty and Grace(2012) use DEA to examine the relative efficiency ofinsurance companies; both of these studies attributethe resulting efficiency score (total firm efficiencyherein) to management quality.

4 Whereas Bertrand and Schoar (2003) examine the styles of theexecutives, other studies examine specific managerial traits. Forexample, Billet and Qian (2008) conclude that self-attribution biasleads to CEO overconfidence, and Malmendier and Tate (2005) findthat overconfident CEOs (i.e., those who repeatedly fail to exercisein-the-money options or habitually acquire their companys stock)cause distortions in corporate investment policies. Chatterjee andHambrick (2007) examine the effects of narcissistic CEOs using sixmeasures of narcissism, including the prominence of the CEOsphoto in the annual report and the length of the CEOs Whos Whoentry; they report that CEO narcissism is associated with organiza-tional strategy changes, a greater number and size of acquisitions,and extreme performance.

3.2. DEA Fundamentals

3.2.1. Framework. DEA is a statistical procedureused to evaluate the relative efficiency of separa-

ble entities, termed decision-making units (DMUs),where each DMU converts certain inputs (labor, capi-tal, etc.) into outputs (revenue, income, etc.). As with

the more widely used efficiency measures, such asthe return on assets and other profitability ratios,DEA efficiency is defined as the ratio of outputs overinputs: s

i=1 uiyikmj=1 vjxjk

k = 1 n (1)

In Equation (1), there are s outputs, m inputs, andn DMUs. In our study, we use firms as the DMUsand consider one output and seven inputs, all derivedfrom firms publicly available financial reports. Rev-enue is the sole output measure; we characterize anable management team as one that generates the high-

est level of revenue from a given set of inputs.5

We consider the following inputs into the revenueproduction process: Net Property, Plant, and Equip-ment(PP&E);Net Operating Leases;Net R&D;PurchasedGoodwill;Other Intangible Assets;Cost of Inventory; andSelling, General, and Administrative Expenses (SG&A).All these inputs contribute to the generation of rev-enue and are affected by managerial ability, as eachof the inputs is subject to managerial discretion.We motivate and describe each of these variablesin 4.

Each output and each input are assigned a weightin calculating the efficiency score, where the weights

are denoted by u and v for the outputs and inputs,respectively. The quantities of outputs and inputs aredenoted byy and x. The DEA optimization procedureinvolves the following steps:

1. We sort DMUs into groups (e.g., industries)within which the relative efficiency program is esti-mated. The groups are determined based on sim-ilarities in the underlying relations between input

5 Possible alternative (or additional) outputs are net income and themarket value of equity. Because we consider expenses as inputs, netincome is effectively an aggregation of our output and inputs (rev-enue less expenses), and thus we opt to use individual expenses asinputs rather than aggregating expenses and revenues and usingnet income as our output. Market value of equity is a noisy out-put because it is affected by many factors beyond managementscontrol. Clearly not all researchers would wish their output to berevenues. The model in this paper can be used as a basic frame-work and can be adapted to consider additional or different inputsor outputs. For example, a researcher wishing to examine manage-rial ability and R&D would clearly want to exclude R&D from theinput set, and a researcher wishing to consider nonfinancial per-formance metrics could include measures of customer satisfaction,or other metrics, in the output set. Researchers wishing to look atone-year changes following turnovers could also adapt the modelto consider only short-term investments, and instead control forprevious investments in Equation (3).


5/20

Demerjian, Lev, and McVay: Quantifying Managerial AbilityManagement Science 58(7), pp. 12291248, 2012 INFORMS 1233

and outputs so that DMUs within each group arecomparable.

2. Next, we maximize Equation (1) for each DMUby varying the weights u and v. This maximiza-tion uses all DMUs in the group and determines theweights that maximize Equation (1) for each DMU

relative to other DMUs in the group. The resultingweights are DMU specific.3. The derived optimal weights are then multiplied

by the corresponding output and input quantities andsummed across all outputs (in the numerator) andinputs (in the denominator). This yields a ratio-basedefficiency score for each DMU.

4. All efficiency scores are then scaled by the high-est efficiency score within the group, resulting in anordinal sorting of DMUs on relative efficiency wherethe most efficient DMUs have a value of one, indi-cating optimal efficiency. For example, if the highestunscaled efficiency score is 3.2, then that DMU would

have a score of one (3.2/3.2), whereas a firm with anunscaled efficiency of 2.2 would have a score of 0.6875(2.2/3.2).

5. The weights,u and v are constrained to be non-negative. This presumes that each input and outputis valuable. Because the quantity of each input andoutput is also nonnegative, the lower bound on theDEA efficiency score is zero.

3.2.2. Advantages of DEA. The DEA efficiencymethodology has two key advantages over conven-tional measures of efficiency. First, DEA provides anordinal ranking of relative efficiency compared to thePareto-efficient frontierthe best performance that

can be practically achieved. Parametric methods, suchas regression analysis and basic ratio comparisons,estimate efficiency relative to average performance,which is lowered disproportionately by inefficientindustry peers. To illustrate that this difference iseconomically important, we reestimate our efficiencymeasure using ordinary least squares (OLS) regres-sion, with revenue as the dependent variable and theinputs (resources), scaled by beginning of period totalassets to adjust for size effects, as the independentvariables. We estimate the regression by industry andinterpret the residual as the over- or underperfor-mance of a specific firm relative to its industry peergroup. Firms with positive residuals are more effi-cient, and those with negative residuals are less effi-cient, than other firms in the industry. We find that theregression residuals and the DEA firm efficiency mea-sure have a relatively low correlation of 0.079, whichincreases to only 0.109 for the rank (Spearman) cor-relation, suggesting that the low correlation is not ascale effect. Rather, the difference in efficiency mea-sures illustrates the multidimensional nature of DEArelative to OLS regression. DEA allows different firmsto optimize across different outputs and inputs, and

it compares each firm to the most efficient outcome,whereas the regression analysis benchmarks each firmagainst the average firm, resulting in a fundamentallydifferent efficiency ranking.6

The second key advantage of DEA is that widelyused efficiency measures, such as return on assets,

require that weights be explicitly set, often assumingthat all inputs and outputs are equally valuable acrossDMUs. The DEA procedure calculates efficiency with-out imposing an explicit, ad hoc weighting structure.If two firms produce the same output, but do so withdifferent mixes of inputs (even if the dollar valueof the inputs differs), both are considered efficient.With a sufficient number of observations, the frontieris formed using all possible combinations of inputs.Those DMUs using a less than optimal mix to reachthe same level of outputs receive efficiency scores ofless than one.

4. Data, Variable Definitions,and Descriptive Statistics

4.1. DataFor the empirical analysis, we obtain data from thefollowing sources: the 2009 annual Compustat file(for financial statement data), the Center for Researchin Security Prices (CRSP) file (for returns data), theExecucomp database (for executive compensation andtenure, and to track managers across firms, with cov-erage from 1992 to 2009), the Morningstar database(to track additional managers across firms, with cov-

erage from 1999 to 2009), and Audit Analytics (toobtain the 8-K filing dates of CEO turnover announce-ments, with coverage from 2000 to 2009). The sam-ple includes all firm-year observations from 1980 to2009 with the required data to calculate the firm effi-ciency measure, yielding 177,512 firm-year observa-tions. Our managerial ability measure (a residual fromthe firm efficiency measure, described in detail below)requires additional data, yielding a final sample of177,134 firm-year observations. The sample period

begins with 1980 because many of the variables weremissing in Compustat before 1980.

6 We empirically assess whether the multidimensionality of theDEA measure of managerial ability is more highly correlatedwith prior proxies of ability than the two-dimensional regressionapproach. We find the DEA measure is more highly correlated witheach of the alternative proxies. Specifically, the correlations betweenthe DEA-based measure of managerial ability and the five alterna-tive measures are 0.169, 0.065, 0.042, 0.060, and 0100, for histor-ical return, historical return on assets, compensation, tenure, andmedia citations, respectively (see Table 4), whereas the correlations

between the two-dimensional regression approach and these fivealternative measures are 0.039, 0.032, 0017, 0.020, and 0048 (nottabulated).


6/20


4.2. Measurement and Summary Statistics of theFirm Efficiency Measure

4.2.1. Inputs and Output of the Firm EfficiencyMeasure. We introduce the inputs and output of theefficiency measure here. All data were obtained fromCompustat, and the Xpressfeed data names are pro-

vided in quotations. To form the portfolio of inputs,we first consider acquired assets, both tangible andintangible, because the management team has a greatdeal of latitude in asset purchase and retirementdecisions, and a more capable management team isexpected to make more efficient purchasing decisions.

The first acquired asset, Net PP&E (PPENT), isreported on the balance sheet and reflects the unde-preciated portion of purchased fixed assets. An alter-native way to gain access to similar fixed assets isthrough the use of operating leases, used across manysectors (airlines, retailers, hotels). The structure ofoperating leases allows firms to exclude the asset

(and related debt) from their balance sheet, althoughthese assets will generate revenues; thus, we esti-mate their capitalized value. We calculate Net Operat-ing Leasesas the discounted present value of the nextfive years of required operating lease payments (avail-able in the firms footnotes to the financial statementsand on Compustat).7 The inclusion of Net OperatingLeases as an input increases the input comparabilityamong firms that effectively have the same opera-tions but either lease or buy their revenue-generatingequipment.

We include Net R&D as an input, expecting thatmore capable managers will be better able to deter-mine which R&D projects to pursue. To calculate NetR&D, which is not reported as an asset on the bal-ance sheet, we follow Lev and Sougiannis (1996),who use a five-year capitalization period of R&Dexpense (XRD), where the net value (net of amor-tization) is RDcap =

0t=41 + 02t RDexp. Thus, for

example, R&D expenditures from five years backreceive a weight of 0.2 (they were already amor-tized 80%), those from four years back receive aweight of 0.4 (amortized 60%), etc., with the prioryears R&D t = 1 receiving full weight. We nextinclude Purchased Goodwill, reported on the balance

sheet, which is the excess of the purchase price for

7 Note that capital leases are included in Net PP&E. The data itemsfor the five lease obligations are MRC1MRC5. We would alsolike to discount the thereafter payments; however, this line itemwas not collected by Compustat for the bulk of the sample period.We use a discount rate of 10% per year to calculate the presentvalue of the required operating lease payments following Ge (2006),who finds that results are substantively unchanged using alterna-tive discount rates of 8%, 12%, or the short-term average borrowingrate. Our empirical results are qualitatively and quantitatively sim-ilar if we exclude operating leases from the DEA estimation (nottabulated).

a business acquisition over the amounts allocatedto other separately identifiable assets and liabilities(GDWL). Goodwill generally reflects the value ofthe acquired intangible assets. We add to it otheracquired and capitalized intangibles (INTAN lessGDWL), also reported on the balance sheet, which

includes items such as client lists, patent costs, andcopyrights. We consider the beginning of period bal-ance for each of the five assets, because managerspast decisions regarding these assets are expected toaffect current period revenues.8

Inventory and advertising expenditures also con-tribute to the generation of revenues. For inventory,we consider the total amount of inventory sold dur-ing the periodthe cost of inventory, or cost of goodssoldto appropriately match the input to the rev-enues generated.9 Finally, advertising expendituresare often missing in Compustat, introducing a pro-hibitive data restriction. Instead, we include the cur-

rent period value ofSG&A(XSGA), which includesadvertising expenditures.10 This variable also capturesother assets that are not explicitly recognizable asaccounting assets, such as the quality of the sales force(training costs and information technology servicesare included in SG&A).

Specifically, we solve the following optimizationproblem:

maxv

= Sales v1CoGS+ v2SG&A + v3PPE

+ v4OpsLease + v5R&D

+ v6

Goodwill + v7

OtherIntan1

(2)

The five stock variables (Net PP&E, Net OperatingLeases, Net Research and Development, Purchased Good-will, and Other Intangible Assets) are measured at the

beginning of year t , and the two flow variables (Costof Inventory and SG&A) are measured over year t.These seven inputs capture, to a large degree, thechoices managers make in generating revenue. Weestimate DEA efficiency by industry (based on Famaand French 1997) to increase the likelihood that the

8 For example, Sunder (1980) finds that purchases of PP&E take an

average of two years to generate positive earnings.9 It is likely that more efficient managers might require less bufferstock or be more likely to use just-in-time inventory processes(Callen et al. 2005), which would also lower average inventory bal-ances, providing some support for considering the beginning bal-ance of inventory instead of total cost of inventory sold duringthe period. However, the beginning balance could be low for otherreasons, such as a cash shortage.10 Operating lease expense and research and development expenseare both components of SG&A expense; to avoid counting theseitems twice in the input vector, we subtract the current year oper-ating lease expense and research and development expense fromSG&A expense.


7/20


peer firms have similar business models and coststructures within the estimations.

Measurement Error in Accounting Variables. We useaccounting values to construct firm efficiency andhold the quality of financial reporting qualityconstant, which introduces several limitations tothis paper. First, there is known variation infinancial reporting quality because of intentionalmanipulationsespecially to revenues (e.g., Dechowet al. 1996). Although it is possible that earningsmanagement inflates the perceived efficiency of afirm, Demerjian et al. (2011) find that more able man-agers are associated with fewer subsequent earningsrestatements, indicating that earnings managementis not the primary driver of perceived efficiency.It remains, however, that earnings management willinflate perceived efficiency. Second, measurementerror also stems from our reliance on accountingnumbers formed using recognition and measurementrules under U.S. GAAP (generally accepted account-ing principles). For example, historical cost measure-ment is not comparable across firms (Curtis andLewis 2010); we must rely on researcher assump-tions to measure capitalized R&D and capitalizedoperating leases, and we must omit other impor-tant intangibles such as purchased R&D because ofdata constraints. Finally, we rely on imperfect indus-try groupings. We estimate Equation (1) by industryaccording to Fama and French (1997), but most firmsoperate in several industries, and even within indus-tries the relation between the accounting inputs andoutputs can vary substantially depending on firmsasset and operations mix. Although we do not expectthese measurement errors to systematically affect ourmanagerial ability score, they do introduce the poten-tial for confounding effects on the efficiency scoreand, thus, the inferences of the study.

4.2.2. Summary Statistics of Firm Efficiency.Panel A of Table 1 provides summary statistics ofthe firm efficiency measurethe starting point for thedevelopment of our managerial ability measureforthe full sample. Recall that DEA constrains the firmefficiency measure to be between zero and one. Themean value reported in Table 1 is 0.569, with a medianvalue of 0.588. A total of 4.5% of observations are on

the frontier (not tabulated).We estimate firm efficiency by industry group,because we expect firms in the same industry tohave similar technologies and business structures forconverting inputs into outputs. Because DEA requiresa sufficiently large number of observations to providea valid estimation, we partition the sample by indus-try and not by time; it is likely that an industrys busi-ness model will remain stable over time.11 We exclude

11 When there are too few firms, a large percentage of these firmswill be on the frontier, especially when there are multiple inputs.

from the sample financial services firms (banks, insur-ance, real estate, and finance companies) because ofthe uniqueness of their asset structure and earningsgenerating processes; we also exclude utilities becauseof regulation of the output price.

Panel B of Table 1 presents summary statistics of

firm efficiency by industry groups (based on Famaand French 1997). There is considerable firm effi-ciency variation across industries. The mean andmedian industry values are 0.672 and 0.674, respec-tively, whereas the range of scores is large, with a lowmean score of 0.271 (drugs) to a high of 0.942 (ships).12

Finally, because we estimate firm efficiency byindustry but not by year, we examine summarystatistics by year (not tabulated); the largest aver-age value is 0.623 (1980), and the smallest is 0.537(2001). Although there is not a lot of year-to-year vari-ation, we include year fixed effects in each of ourregressions.

4.3. Estimation of Managerial AbilityFirm efficiency could be used to assess manage-rial ability; however, this measure captures bothfirm-specific and manager-specific efficiency drivers,and thus it likely overstates or understates manage-rial ability, depending on the firm-specific efficiencydrivers. We parse out total firm efficiency into firmefficiency and managerial ability by regressing totalfirm efficiency on six firm characteristics that affectfirm efficiency: firm size, firm market share, cashavailability, life cycle, operational complexity, and for-eign operations.

First, we expect that managers of larger firms withmore market share will be more effective than othersin negotiating terms with suppliers and customers,holding their ability constant. Second, we expect man-agers in firms with available cash (measured with anindicator variable indicating positive free cash flows)to be able to pursue positive net present value projectsmore effectively, again holding ability constant. Third,we expect the life cycle of the firm to affect man-agements opportunity set of possible projects as well

The efficient frontier consists of, at a minimum, the sum of the num-ber of inputs and outputs, and because we use the variable returns-to-scale model in calculating DEA, there are additional points on

the frontier to accommodate different firm sizes. Therefore, werequire at least 100 observations to estimate DEA. Alternatively,Leverty and Qian (2011) estimate a very similar estimation byindustry and year, resulting in a higher average firm efficiencyscore (0.745 for the full sample; see their Table 1).12 The percentage of observations making up the efficient frontieralso varies across industries, ranging from 1.6% (business services)to 28.5% (guns) (not tabulated). This effect is likely driven in part

by the competitiveness of different industries, but is also a func-tion of the number of observations available to estimate the fron-tier (see also Footnote 11). This is a potential concern for users ofDEA; however, we estimate the second-stage regression by indus-try, removing systematic differences in efficiency across industries.


8/20


Table 1 Sample Statistics on Firm Efficiency

Obs. Mean Std. dev. 25% Median 75%

Panel A: Firm efficiency measure

Variable

Firm Efficiency 177512 0569 0273 0347 0588 0802

Panel B: Firm efficiency measure by industryIndustry

Agriculture 780 0778 0220 0617 0844 0984

Food 3672 0773 0156 0652 0800 0897

Soda 460 0896 0138 0848 0954 1000

Beer and liquor 733 0826 0143 0764 0853 0928

Smoking 268 0854 0158 0753 0905 1000

Toys 1868 0702 0194 0590 0700 0829

Fun 4038 0430 0251 0234 0381 0577

Books 2027 0796 0154 0697 0818 0913

Household products 4065 0680 0196 0513 0680 0846

Clothing 2803 0732 0145 0636 0719 0828

Health 3528 0737 0178 0627 0756 0866

Medical equipment 6274 0456 0231 0294 0417 0578

Drugs 8447 0271 0253 0088 0200 0345

Chemicals 3723 0717 0187 0609 0732 0845Rubber 2308 0834 0141 0780 0857 0926

Textiles 1422 0834 0108 0782 0842 0906

Building materials 4720 0608 0231 0458 0613 0785

Construction 2673 0669 0186 0578 0675 0784

Steel 3258 0700 0136 0610 0690 0785

Fabricated products 994 0874 0095 0815 0877 0954

Machinery 7085 0637 0234 0450 0676 0823

Electrical equipment 2129 0728 0186 0614 0735 0863

Utilities 1309 0616 0301 0375 0647 0897

Automobiles 3039 0783 0194 0736 0832 0911

Aerospace 1009 0860 0123 0773 0876 0973

Ships 433 0942 0075 0911 0969 1000

Guns 323 0878 0174 0810 0949 1000

Gold 1684 0342 0277 0121 0255 0518

Mining 1162 0281 0254 0116 0203 0349

Coal 328 0802 0187 0722 0835 0962Energy 10995 0328 0269 0128 0219 0485

Telecom 7799 0591 0229 0448 0603 0754

Personal services 1996 0685 0195 0574 0691 0820

Business services 21884 0381 0227 0226 0333 0500

Computers 9055 0459 0215 0314 0431 0571

Chips 11516 0472 0190 0360 0446 0553

Laborator y equi pment 4481 0577 0202 0447 0560 0697

Paper 3012 0807 0119 0746 0802 0875

Boxes 705 0933 0070 0898 0950 0991

Transportation 5752 0650 0225 0447 0691 0839

Wholesale 8919 0647 0203 0528 0676 0792

Retail 10453 0826 0119 0775 0842 0903

Restaurants 4383 0517 0246 0324 0464 0710

Average (43 industries) 4128 0672 0186 0553 0674 0799

Notes. The sample consists of 177,512 firm-year observations from 1980 to 2009. Firm Efficiency is measured

using DEA based on the vectors described in 4.2. Panel A presents statistics for the full sample. Panel B is sorted

by industry, based on Fama and French (1997).

as the required start-up costs of investments. We usethe number of years the firm has been listed as aproxy for a firms life-cycle stage (DeAngelo et al.2010). Finally, we consider the diversification of afirms operations, both operationally and geograph-ically. We expect that the greater the diversification,

the more challenging it is for the management team toefficiently allocate capital, because operating in mul-tiple industries requires a broader knowledge set andreduces the amount of attention management paysto any single industry (e.g., Stein 1997). We measureoperational complexity using the within-firm industry


9/20


concentration (e.g., Bushman et al. 2004) as well as anindicator variable signifying foreign operations. Thus,to form our estimate of managerial ability, we esti-mate the following Tobit regression by industry andinclude year fixed effects; we cluster standard errors

by firm and year to control for cross-sectional and

intertemporal correlation:

Firm Efficiencyi

= + 1 ln (Total Assets)i + 2Market Sharei

+ 3Free Cash Flow Indicatori + 4 lnAgei

+ 5Business Segment Concentrationi

+ 6Foreign Currency Indicatori + Yeari + i (3)

Because we estimate Equation (3) by industry, wedo not include industry-level drivers of efficiencysuch as competition. This estimation dampens vari-

ation in ability, for example, by controlling for firmsize, because better managers are more likely to behired by larger firms (Rosen 1982).13 Managerial abil-ity could also affect the variables we attribute to thefirm, such as market share (e.g., Vanhonacker andDay 1987). We opt to err on the side of attributingmanager characteristics to the firm, to maximize thelikelihood that the residual is largely attributable tothe manager.

We summarize the results from this estimation, aswell as the definitions of the variables, in Table 2.We present the average coefficient across the 43 indus-try estimations and note the significant percentage

and the percentage with the predicted sign. For exam-ple, the coefficient on firm size (the natural log oftotal assets) has a one-tailed p-value of less than 0.05in 90.7% of the industry estimations, and all 43 coef-ficients are positive as expected. The average coeffi-cients on each of the independent variables are sig-nificant in the predicted direction (using the standarderror of industry-level coefficients to test significancealong the lines of Fama and MacBeth 1973), support-ing our decision to remove these firm effects fromthe firm efficiency measure to create our managerialability measure. The importance of some of the vari-ables varies by industry; for example, while the free

cash flow indicator is statistically significant in thepredicted direction in 40 of the 43 industries, busi-ness concentration is significant in less than half ofthe industries.14

13 As an illustration of this limitation, Jack Welch is assigned a lowability score using this estimation because GEs size, market share,and age (each of which are in the highest percentile) predict a totalefficiency score of one for GE. If we exclude these variables fromEquation (3), Welchs ability score is high, as expected.14 To assess the economic significance of the firm-specific portionof total efficiency, we estimate Equation (3) by industry using OLS

Table 2 Managerial Ability

Dependent variable = Firm Efficiency

Average

coefficient Proportion Proportion with

Predicted (FamaMacBeth significant predicted

sign t-statistic) (%) sign (%)

Ln(Total Assets) + 0037 907 10001151

Market Share + 1599 651 767

561

Free Cash Flow + 0075 930 1000

Indicator 1112

Ln(Firm Age) + 0021 674 861

652

Business Segment 0029 419 674

Concentration 312

Foreign Currency 0014 674 721

Indicator 246

Intercept 0567

1671

Year fixed effects IncludedIndustry estimations 43

Notes. This table presents the averages from the Tobit estimation of

Equation (3) by industry; the residual from the estimation is Managerial Abil-

ity, described in 4.3. For illustrative purposes, we present the average of the

industry coefficients and calculate the Fama and MacBeth (1973) t-statistic

based on the standard error of these coefficients (in parentheses). The signif-

icant percentage is the number of coefficients that are statistically significant

at the 5% level (one-tailed) across the 43 industry regressions using stan-

dard errors clustered by firm and year. The Percentage with predicted sign is

the proportion of the 43 industry coefficients with the predicted sign. Vari-

ables are defined as follows (with Compustat identifiers in parentheses):Firm

Efficiencyis measured using DEA based on the vectors described in 3.2.

Total Assetsis Compustat (AT) at the end of year t.Market Shareis the per-

centage of revenues (SALE) earned by the firm within its Fama and French

(1997) industry in year t.Free Cash Flow Indicatoris coded to one when afirm has nonnegative free cash flow (defined as earnings before depreciation

and amortization (OIBDP) less the change in working capital (RECT + INVT +

ACO LCO AP) less capital expenditures (CAPX)) in year t.Firm Ageis the

number of years the firm has been listed on Compustat at the end of year t.

Business Segment Concentrationis the ratio of individual business segment

sales to total sales, summed across all business segments for year t. If the

firm is not in the segment file, it is assigned a concentration of one. Foreign

Currency Indicatoris coded to one when a firm reports a nonzero value for

foreign currency adjustment (FCA) in year t. Variables are winsorized at the

extreme 1%.5% and 1% statistical significance (two-tailed tests).

The residual from this estimation is our measureof managerial ability.15 Table 3 presents summarystatistics. The mean value of our managerial ability

regressions (Tobit regressions do not have a goodness-of-fit mea-sure analogous to adjusted R2. We find the average adjusted R2

is 37.4%, suggesting that, on average, over one-third of total firmefficiency is attributable to the firm features we have identified inEquation (3).15 The estimation of Equation (3) may not fully ameliorate theimpact of unidentified features, such as unions or investor base,that affect managers ability to utilize firm resources. One alterna-tive to better capture these unidentified drivers of efficiency would

be to include firm fixed effects. We opt to present the residual


10/20


Table 3 Descriptive Statistics

Standard

Variable Observations Mean deviation Min 1% 25% Median 75th 99% Max

Firm Efficiency 177512 0569 0273 0000 0024 0347 0588 0802 1000 1000

Managerial Ability 177134 0004 0149 0415 0349 0094 0013 0075 0451 0557

Fitted Value of Ability 22982 0000 0111 0374 0194 0074 0013 0065 0462 0936

Alternative measures ofability

Historical Return 83652 0224 2345 6274 3371 0954 0257a 0688 10281 22455

Historical ROA 125487 0346 1783 26847 7730 0286 0000 0204 0953 1243

CEO Cash Compensation 22982 1183610 1572510 0000 50000 535000 849277 1357200 5928740 77926000

CEO Tenure 22113 7465 7539 0000 0000 2167 5003 10005 35855 58786

Media Mentions 14943 213473 706688 0000 2000 46000 93000 182000 2016000 43152000

Notes. Firm Efficiencyis measured using DEA based on the vectors described in 3.2. Managerial Ability is the residual-based measure described in 4.3.

Fitted Value of Abilityis the fitted value of available manager fixed effects on firm efficiency described in 4.3. Historical Returnis the five-year historical

value-weighted industry-adjusted return (from year t 5 to yeart 1).Historical ROAis the five-year industry-adjusted return on assets (cumulative income

before extraordinary items (IBC) scaled by average total assets (AT) from year t 5 to yeart 1).CEO Cash Compensationis the salary and bonus of the firm

CEO (TOT_CURR from Execucomp; in thousands) for yeart.CEO Tenureis the number of years an executive has been listed as CEO by Execucomp at the end

of yeart.Media Mentionsis the number of articles mentioning the CEO over the preceding five-year period. Variables are winsorized at the extreme 1%.aNote that the industry-adjusted median is not zero because of the value-weighting procedure.

measure is 0004, and the median is 0013, with aninterquartile range of 0.169.16 The values range from0415 to 0.557 (see Table 3).

We also create an alternative measure of manage-rial ability, based on CEO fixed effects, for a sampleof 22,982 firm-years with available CEO identifiers.Specifically, we regress firm efficiency on CEO fixedeffects. We consider the fitted value of the CEO fixedeffects as the lower bound of the manager-specificcomponent of firm efficiency and correlate this pre-dicted value with our residual-based managerial abil-

ity measure. The correlation is over 0.80 (see Table 4),which provides a lower-bound estimate of the pro-portion of the residual that is CEO specific and sup-ports the notion that the residual from Equation (3) islargely attributable to the manager. We opt to focusour analyses on the residual-based measure of man-agerial ability because it is available for nearly everyfirm on Compustat, whereas the CEO fixed effectsmeasure is available only for firms covered by Execu-comp or Morningstar (limiting both the time and firmcoverage), and these fixed effects are uninformativewhen the executive does not change over the sample

estimate excluding firm fixed effects to maximize comparabilityacross firms, because including firm fixed effects creates a relativeability measure within the firm, but removes important variationacross firms (because each firm effectively must have a mean zeroefficiency).16 The mean is not zero because we estimate Equation (3) using aTobit regression; we subtract the predicted value from the Tobitestimation from the actual value ofFirm Efficiency to compute aTobit residual. Unlike OLS residuals, which must sum to zero

by definition, these computed differences need not. Empirical testresults are similar if we use truncated regression or OLS regressionto estimate managerial ability, and are also similar if we considerthe natural logarithm of total firm efficiency in the OLS estimation.

period (because firm fixed effects and manager fixedeffects are then indistinguishable).

4.4. Comparison with Alternative

Ability Measures

In this section, we correlate our proposed mea-sure with five alternative measures of managerial

ability used in prior research: historical industry-adjusted stock returns, historical industry-adjusted

ROA, CEO compensation, CEO tenure, and media

mentions (see 2). We present summary statistics inTable 3 and univariate correlations in Table 4. We also

include, in Table 4, correlations for ROA and firmsize. First, note the high positive correlation between

firm efficiency and managerial ability, which is rea-sonable given that managerial ability is a component

of total firm efficiency, but could also indicate thatour partition is imprecise (McNichols 2000). That firm

efficiency is also positively correlated with the fittedvalue of managerial ability suggests that better man-

agers are employed by better firms.Turning next to the correlations between the DEA-

based efficiency measures and the alternative mea-sures of ability, we see that the correlations are

reasonably low (all below 0.30), indicating that theDEA-based efficiency measures are different from the

alternative ability measures. We also see that thesecorrelations are lower for managerial ability than for

firm efficiency, consistent with the removal of manyfirm-specific characteristics from our managerial abil-

ity measure. OnlyMedia Mentionsdoes not exhibit theexpected positive correlation with managerial abil-

ity, although Media Mentions and total Firm Efficiency


11/20


Table 4 Univariate Correlations

Fitted

Firm Managerial Value of Historical Historical Ln(CEO Cash Ln(CEO Ln(Media

Efficiency Ability Ability Return ROA Compensation ) Tenure) Mentions) ROA Size

Firm Efficiency 0550 0410 0171 0136 0298 0011 0193 0338 0458

Managerial Ability 0548 0835 0169 0065 0042 0060 0100 0120 0047

Fitted Value of Ability 0389 0828 0147 0084 0033 0086 0088 0212 0069Historical Return 0279 0211 0168 0100 0039 0065 0025 0236 0176Historical ROA 0149 0156 0172 0313 0027 0015 0044 0523 0237

Ln(CEO Cash Compensation) 0417 0072 0054 0154 0008 0015 0312 0097 0382

Ln(CEO Tenure) 0004 0056 0086 0116 0071 0028 0009 0058 0030Ln(Media Mentions) 0196 0092 0079 0006 0016 0435 0030 0027 0511

ROA 0469 0336 0299 0481 0471 0149 0093 0102 0231Size 0459 0066 0095 0318 0405 0568 0014 0501 0335

Notes. This table presents univariate correlations between the main variables used in the tests. Firm Efficiency is measured using DEA based on the vectors

described in 3.2. Managerial Abilityis the residual-based measure described in 4.3. Fitted Value of Abilityis the fitted value of available manager fixed

effects on firm efficiency described in 4.3.Historical Returnis the five-year historical value-weighted industry-adjusted return (from year t 5 to yeart 1).

Historical ROA is the five-year industry-adjusted return on assets (cumulative income before extraordinary items (IBC) scaled by average total assets (AT)

from yeart 5 to yeart 1).CEO Cash Compensation is the salary and bonus of the firm CEO (TOT_CURR from Execucomp) for year t.CEO Tenureis the

number of years an executive has been listed as CEO by Execucomp.Media Mentionsis the number of articles mentioning the CEO over the preceding five-year

period.ROAis income before extraordinary items (IBC) scaled by average total assets (AT) for yeart.Sizeis the natural logarithm of the market value of equity

(PRCC_C CSHO) as of the end of year t. Pearson correlations are presented in the upper right and Spearman correlations are presented in the lower left.Variables are winsorized at the extreme 1%. Bold values indicate statistical significance at the 10% level.

are positively associated.17 Overall, the evidence indi-cates that our Managerial Abilitymeasure is positivelyassociated with prior measures of managerial ability.These positive correlations hold in the presence ofstandard control variables (not tabulated).

5. Validation TestsWe hypothesize that our proposed measure reflectsmanagerial ability. We test this hypothesis in three

ways. First, we explore whether the measure is eco-nomically and significantly associated with managerfixed effects (5.1). Second, we test whether the mea-sure is negatively associated with the announcementreturns to CEO turnovers (5.2). Third, we investi-gate whether appointing a relatively more or less ablemanager is systematically associated with changes insubsequent firm performance (5.3). For each of thesetests, we contrast our measure of managerial abilitywith the five alternative measures discussed in 4.4.

5.1. Economic Significance ofManager Fixed Effects

To investigate whether manager fixed effects explainan economically significant portion of our manage-rial ability measure, we conduct a series of tests fora subset of CEOs who switch employers within oursample. By examining CEOs who are present in atleast two firms, we can assess whether managerial

17 Note that the correlation is positive in Baik et al. (2011); however,their sample is constrained to large firms with Execucomp and FirstCall coverage, and it is likely that larger firms garner more mediaattention, holding managerial ability constant.

ability systematically differs across individual CEOs(see, for example, Bertrand and Schoar 2003, Levertyand Grace 2012). To identify this sample, we focuson firms that are both covered by Execucomp orMorningstar and had a change of CEO from 1993to 2009. We then identify those turnovers where theCEO moved from one sample firm to another dur-ing the period 1993 to 2009. We confirm that the CEOdid indeed switch firms, versus remaining the CEOof a firm that changed names. We also require thatthe CEO not switch firms as a result of a merger;following a merger, if the new entity retained eitherof the preexisting CEOs, we exclude this observation.Finally, we require that the CEO hold his or her postfor two years in each firm to allow the CEO time toaffect the firm. Our final sample contains 78 CEOsemployed by 125 unique firms.

We regress our managerial ability measure on man-ager and firm fixed effects and expect manager fixedeffects to explain a greater proportion of the varia-tion in our ability measure than firm fixed effects.We examine the proportion of fixed effects that arestatistically significant; we again cluster standarderrors by firm and year. Following Leverty and Grace(2012), we use the estimation with year fixed effectsas the benchmark specification. As Table 5 shows, wefind that 66.5% of the manager fixed effects are sta-tistically significant (p


12/20


Table 5 Manager and Firm Fixed Effects

Dependent variable = Ability Measure

Spread

Fixed Proportion (manager-

Ability measure effect significant (%) firm) (%)

Managerial Ability Manager 665 605 314

Firm 291

Historical Return Manager 481 269 75

Firm 194

Historical ROA Manager 473 290 36

Firm 326

Ln(CEO Cash Manager 6 97 427 42

Compensation) Firm 385

Ln(Tenure) Manager 529 756 32

Firm 724

Ln(Media Mentions) Manager 476 427 07

Firm 434

Notes. This table provides statistics on the economic significance of indi-

vidual managers on six ability measures. To be included, a CEO must be

employed as CEO by at least two companies in the sample from 1993

2009. There are a maximum of 78 CEOs from 125 firms, although for some

measures there will be fewer observations because of data availability. Each

regression includes year fixed effects and the natural logarithm of total assets

to control for size. The ability measure is the dependent variable, and we esti-

mate two specifications: with manager fixed effects (1st column) and with

manager and firm fixed effects (2nd column). The proportion significant is

determined based on robust t-statistics, with standard errors clustered by

firm and year (p


13/20


Table 6 Stock Price Reactions to CEO Turnovers

Panel A: Main analysis

Dependent variable = Announcement Return

Ability measure

Managerial Historical Historical Ln(CEO Cash

Ability Return ROA Compensation ) Ln(Tenure) Ln(Media Mentions)

(1 1) (5 1) (1 1) (5 1) (1 1) (5 1 (1 1) (5 1) (1 1) (5 1) (1 1) (5 1)

Return 0024 0028 0001 0001 0001 0000 0002 0004 0002 0000 0001 0001

(t-statistic) 213 205 163 158 047 014 206 203 145 005 071 047

Marginal Effect 00037 00044 00019 00020 00010 00002 00023 00044 00024 00002 00013 00015

Observations 2,229 1,604 1,883 879 855 454

Panel B: Subsample tests

Dependent variable = Announcement Return

Subsample

Ln(CEO Cash

Historical Return Historical ROA Compensation ) Ln(Tenure) Ln(Media Mentions)

(1 1) (5 1) (1 1) (5 1 (1 1) (5 1) (1 1) (5 1) (1 1) (5 1)

Managerial Ability 0032 0033 0028 0033 0002 0003 0003 0006 0002 0004

(t-statistic) 260 243 146 214 013 066 017 088 007 033

Marginal Effect 00050 00052 00044 00052 00003 00005 00005 00009 00003 00006

Observations 1,604 1,883 879 855 454

Notes. This table presents results on short-window returns following the announcement of CEO turnovers. Managerial Ability is the residual-based measure

described in 4.3. Historical Return is the five-year historical value-weighted industry-adjusted return (from year t 5 to year t 1).Historical ROA is the

five-year industry-adjusted return on assets (cumulative income before extraordinary items (IBC) scaled by average total assets (AT) from year t 5 to year

t 1).CEO Cash Compensationis the salary and bonus of the firm CEO (TOTAL CURR from Execucomp) for year t.Tenureis the number of years an executive

has been listed as CEO on Execucomp. Media Mentionsis the number of articles mentioning the CEO over the preceding five-year period. In panel A we

present the OLS regression coefficient on the ability measure on short-window returns of three and six days. We present the coefficient, robust t-statistics

with standard errors clustered by firm and year, and marginal significance (the difference in return for a one-standard-deviation change in ability measure).

In panel B we present regressions for subsamples where the specified variable is not missing; Managerial Abilityis the independent variable in each regression.

Variables are winsorized at the extreme 1%.5% and 1% statistical significance (two-tailed tests); 5% statistical significance (two-tailed tests) in the direction opposite to predictions.

results is estimated within the 1,604 observations thathave historical return data). We find similar resultswithin the two largest subsamples (historical returnsand historical ROA), suggesting that our ability scoredoes outperform these two alternative measures inthis setting. The results within the latter three sub-samples, however, are insignificant.

It is possible that these tests lack power, eitherbecause of the reduced sample size or because of sys-tematic differences between the subsamples and the

underlying population. Specifically, because the lat-ter three samples tend to contain larger firms, onaverage, than the full population for which we calcu-lated managerial ability, it may be that there is lessvariation in the measured managerial ability of thesesubsamples, leading to low power tests. We conductseveral additional analyses in an effort to determineif one or both of these explanations is valid. First,we reestimate the six-day announcement return onour managerial ability score for 100 random drawsof 879 and 454 observations from the sample of

2,229 observations in panel A. Although the bulk ofthese are insignificant, consistent with sample sizelowering the power of these tests, the average coef-ficients on these estimations are 0028 and 0029for the compensation and media mention samples,respectively. We compare these to the coefficients inpanel B of 0.003 (compensation), 0006 (tenure), and0.004 (media cites) and reject the null that they aredrawn from the same distribution with t-statistics of240 169, and 165, respectively. Thus, although

sample size is clearly playing a role, the insignificancein panel B is at least partly due to the underlyingpopulation examined. Thus, we cannot yet concludethat our measure dominates compensation, tenure,and media mentions, because it could be that thesesubsamples simply lack variation in ability. To inves-tigate this, we next reestimate managerial ability foronly those firms that have media mention data (byfirst estimating firm efficiency using DEA and thenreestimating Equation (3) to extract firm characteris-tics, again, only for the media mention subsample).


14/20


When we reexamine the relation between announce-ment returns and managerial ability, we obtain acoefficient of 0018 and a t-statistic of 135, bothof which are much more in line with the statisticsobtained in our bootstrapping analysis. Thus, usersof our managerial ability score will need to reesti-

mate DEA within the population of firms they wishto examine if their sample of interest is both suf-ficiently small and systematically different from thetotal population.

5.3. Transferability of Ability: Changes in FirmPerformance Following New Appointments

As a final validity test, we investigate whether anewly appointed CEOs prior ability score (esti-mated while employed at the prior firm) is corre-lated with the subsequent performance of the newfirm. We expect that firms hiring better (worse) man-agers experience improvements (declines) in firm per-

formance. For the 78 CEOs who were employed bymore than one firm in our sample, we calculate thedifference in CEO ability by subtracting the outgoingmanagers ability from the incoming managers ability(measured in their prior firm). We regress both subse-quent industry-adjusted stock returns and subsequentchanges in industry-adjusted ROA on the ability dif-ference. As before, we cluster standard errors by firmand year.

Table 7 presents the results. Consistent with ourexpectations, we find evidence of improved perfor-mance in the three years following the appointmentof a higher-ability CEO. For example, a one-standard-

deviation increase in the relative ability of a CEO isassociated with a 37.0% higher stock return and 3.2%higher ROA over the next three years; these changesrepresent 57% and 34% of the interquartile range forreturns and ROA, respectively. These results are espe-cially compelling because they are based on the newlyappointed CEOs ability from their prior firms, andthus are less subject to the concern that the residual-

based ability measure reflects other unidentified firmcharacteristics. Using this measure, there is strong evi-dence that managers assessed levels of ability trans-fer across firms and explain future performance attheir new firm.

We again replicate the analysis for the alterna-tive ability measures. None of the alternative mea-sures explains subsequent returns, but both historicalindustry-adjusted ROA and CEO tenure are posi-tively associated with changes in future industry-adjusted ROA. To assess the underlying reason forthe insignificant results for the alternative measures,we again reestimate our main results for each of thesubsamples in Table 7, panels C and D, respectively.In panel C, changes in our ability score are associ-ated with changes in industry-adjusted stock returns

for four of the five samples (all but media mentions),whereas in panel D, changes in our ability score areassociated with changes in industry-adjusted ROA inonly the two largest samples. As in Table 6, it is likelythat the power of the test is limited among the smallersamples because of both sample size and systemati-

cally different underlying populations. To investigatethis, we again reestimate the analysis using manage-rial ability estimated within the media mention sub-sample. As before, results strengthen considerably; wedocument positive coefficients on managerial abilityfor both estimations (with t -statistics of 2.20 and 1.61for industry-adjusted returns and industry-adjustedROA, respectively; not tabulated).

To summarize, we conclude that although it con-tains noise, our managerial ability measure offers acleaner depiction of managers ability than prior mea-sures. It is economically and significantly associatedwith manager fixed effects and is associated with both

the price reactions to CEO turnover announcementsand changes in firm performance following new CEOappointments.

6. Managerial Ability and theNew Issue Puzzle

In this section, we demonstrate the potential of ourmeasure to resolve extant research puzzles. Firmsissuing seasoned equity have inordinately low stockreturns during the five years after the offering(Loughran and Ritter 1995). Investigating this puz-zle, Loughran and Ritter (1997) find that firms that

rapidly increase either sales or capital expenditureshave lower subsequent stock returns than other firms.After controlling for growth, however, they continueto find that issuing firms substantially underperformnonissuers and suggest that the firms are investing inwhat the market views as positive net present valueprojects, when often these projects have negative netpresent values. Moreover, even in the face of dete-riorating performance, they find that managers con-tinue to invest heavily, suggesting that the managersare also overoptimistic about the issuing firms futureprofitability.

We conjecture that superior managers are bet-

ter able to effectively select and execute positivenet present value projects, and are less likely toallow unrealistic expectations to cloud their deci-sion making. Thus, we expect the negative relation

between equity financing and subsequent abnormalreturns to be mitigated in firms with better man-agers. Regression results are presented in panel A ofTable 8; as in prior estimations, we cluster standarderrors by firm and year. Similar to Bradshaw et al.(2006), we measure equity financing as the changein equity (Compustat: CEQ) plus the change in


15/20


Table 7 Performance Changes Following Management Changes

Panel A: Change in stock return

Dependent variable = Change in Industry-Adjusted Stock Return

Ability measure

Predicted Managerial Historical Historical Ln(CEO Cash

sign Ability Return ROA Compensation ) Ln(Tenure) Ln(Media Mentions)

Difference in Ability + 2817 0036 0047 0281 0066 0098276 039 043 205 050 088

Control variables for the CEOs new firmChange in ROA + 1646 0941 0764 0208 0942 2037

329 196 175 049 164 504

Change in Book-to-Market ? 0154 0457 0746 0979 0826 0039125 213 247 426 272 028

Change inLn(Market Value of Equity) ? 0501 0345 0325 0256 0382 0055249 316 232 120 195 021

Intercept 0057 0075 0116 0199 0102 0073042 058 105 112 076 043

Observations 78 68 74 50 69 31AdjustedR2 025 024 023 033 020 063

Panel B: Change in ROA

Dependent variable = Change in Industry-Adjusted ROA

Ability measure

Predicted Managerial Historical Historical Ln(CEO Cashsign Ability Return ROA Compensation ) Ln(Tenure) Ln(Media Mentions)

Difference in Ability + 0245 0011 0028 0033 0036 0019314 072 098 109 196 019

Control variables for the CEOs New FirmChange in Stock Return + 0042 0068 0051 0032 0041 0329

355 319 304 133 228 213

Change in Book-to-Market ? 0005 0023 0011 0035 0017 0019014 047 024 070 039 020

Change inLn(Market Value of Equity) ? 0024 0027 0025 0016 0034 0026091 095 096 020 141 038

Intercept 0002 0010 0001 0001 0024 0028013 053 006 003 120 045

Observations 78 68 74 50 69 31AdjustedR2 007 014 011 007 013 053

Panel C: Change in Stock Return

Dependent variable = Change in Industry-Adjusted Stock Return

Subsample

Predicted Historical Historical Ln(CEO Cashsign Return ROA Compensation ) Ln(Tenure) Ln(Media Mentions)

Difference in Ability + 1490 1699 1671 2910 1100

171 237 206 308 050Control variables for the CEOs new firm

Change in ROA + 0968 0980 0460 1483 2258

178 172 067 209 393

Change in Book-to-Market ? 0472 0598 0745 0524 0027246 294 315 207 010

Change inLn(Market Value of Equity) ? 0379 0348 0326 0413 0022358 307 231 295 010

Intercept 0068 0060 0168 0006 0140072 060 132 005 086

Observations 68 74 50 69 31AdjustedR2 023 028 035 030 061


16/20


Table 7 (Continued)

Panel D: Change in ROA

Dependent variable = Change in Industry-Adjusted ROA

Subsample

Predicted Historical Historical Ln(CEO Cash

sign Return ROA Compensation ) Ln(Tenure) Ln(Media Mentions)

Difference in Ability + 0226 0261 0134 0290 0668

252 277 080 150 082

Control variables for the CEOs New Firm

Change in Stock Return + 0066 0050 0028 0043 0315

313 260 116 271 366

Change in Book-to-Market ? 0010 0003 0046 0021 0021

028 009 115 061 021

Change inLn(Market Value of Equity) ? 0037 0026 0022 0037 0019

188 135 095 196 022

Intercept 0006 0000 0000 0005 0046

035 002 000 020 073

Observations 68 74 50 69 31

AdjustedR2 012 007 005 010 057

Notes. The two dependent variables are theChange in Industry-Adjusted Stock Returnand theChange in Industry-Adjusted ROA, where changes are measured

as the value in yeart+ 3 less the value in yeart 1, where both values are measured at the new firm. The differences in the six ability measures are calculated

across firms and are equal to the CEOs prior ability less the outgoing CEOs ability. The six ability measures considered are Managerial Ability, the residual-

based measure described in 4.3 (measured in year t 1);Historical Return, the five-year historical value-weighted industry-adjusted return (from year t 5

to yeart 1);Historical ROA, the five-year industry-adjusted return on assets (cumulative income before extraordinary items (IBC) scaled by average total

assets (AT) from yeart 5 to yeart 1);CEO Cash Compensation, the salary and bonus of the CEO (TOT_CURR from Execucomp) in year t 1;CEO Tenure,

the number of years an executive has been listed as CEO by Execucomp as of year t 1; andMedia Mentions, the number of articles mentioning the CEO from

year t 5 to yeart 1. Changes in control variables are measured as year t + 3 less yeart 1, where both values are measured at the new firm. Control

variables areROA, income before extraordinary items (IBC) scaled by average total assets (AT); Book-to-Market, the beginning of year book value of equity

(CEQ) divided byMarket Value of Equity; andMarket Value of Equity, the beginning of year equity capitalization (PRCC_C CSHO). Variables are winsorized at

the extreme 1%. Robustt-statistics, with standard errors clustered by firm and year, are presented in parentheses.10%, 5%, and 1% statistical significance (two-tailed tests); 5% statistical significance (two-tailed tests) in the direction opposite to predictions.

preferred stock (PSTK) minus net income (NI).We measure future abnormal returns as the value-weighted industry-adjusted annual return beginningfour months after year-end. We regress the future,post-equity-issue returns on the change in equityfinancing, our measure of managerial ability, and theinteraction of these two variables. Consistent withLoughran and Ritter (1995) and Bradshaw et al.(2006), we find a negative coefficient on the equityfinancing variable, indicating that higher levels ofequity financing result, on average, in lower futurereturns. The coefficient on managerial ability is pos-itive, indicating that more talented managers, as

defined by our measure, are associated with higherfuture returns. Finally, there is also a positive coef-ficient on the interaction term of equity financingand ability, our variable of interest, suggesting thatmore able managers mitigate the negative associa-tion between external financing and future abnormalreturns. The incremental effects are 0063 and 0.085for the best and worst manager, respectively, wherethe best (worst) manager has an ability score of 0.557(0415) (see Table 3). As in the prior tests, our man-agerial ability measure outperforms other proxies of

ability, with only historical returns loading weakly inthe predicted direction.18

In panel B of Table 8, we again investigate theunderlying reason for insignificance for the alterna-tive ability proxies by reestimating our main analy-sis for each of the subsamples. Consistent with theresults in panel A, we find that managerial abilitycontinues to mitigate the future abnormal returnsassociated with equity issuances for each of these sub-samples. This finding is important because it allowsus to conclude that our measure does dominate the

18 We further corroborate these results by exploring the associa-tion between managerial ability and growth opportunities (e.g.,Chemmanur and Paeglis 2005, Jones and Olken 2005). We expectmore able managers to successfully identify and capitalize oninvestment opportunities, and thus we expect our managerial abil-ity measure to be associated with Tobins q, a measure of invest-ment opportunities, defined as the market value of assets (marketvalue of equity plus the book value of long-term debt ((PRCC_C CSHO) + (DLTT + DLC)) divided by total assets (AT). In untabu-lated results, we find that more able managers have higher Tobinsq. Moreover, this association persists in the presence of the alterna-tive measures of managerial ability examined previously. In termsof economic significance, a one-standard-deviation increase in man-agerial ability is associated with a 10.6% higher Tobins qafter wecontrol for annual return, sales growth, and market value of equity.


17/20


Table 8 New Issue Puzzle

Panel A: Comparison of ability measures

Dependent variable = Future Returns

Ability measure

Predicted Managerial Historical Historical Ln (CEO Cash

sign Ability Return ROA Compensation ) Ln(Tenure) Ln(Media Mentions)

Change in Equity 0057 0088 0035 0329 0094 0575

122 189 210 307 165 322

Ability Measure + 0219 0042 0039 0040 0005 0017

714 1106 210 307 165 341

Change in Equity Ability Measure + 0153 0015 0011 0020 0060 0082

197 091 133 022 091 248

Ln(Market Value of Equity) 0022 0028 0026 0048 0035 0065

353 456 473 496 407 647

Book-to-market + 0074 0051 0071 0115 0109 0153

495 419 490 368 355 359

Intercept 0289 0214 0296 0205 0369 0519

979 715 1027 350 769 843

Observations 112,550 76,155 88,083 20,423 19,005 13,974

Year indicators Included Included Included Included Included Included

AdjustedR2 005 008 005 006 006 007

Panel B: Subsample Tests

Dependent variable = Future Returns

Predicted Historical Historical Compensation Tenure Media Mentions

sign Return Sample ROA Sample Sample Sample Sample

Change in Equity 0056 0036 0170 0166 0212

116 066 096 090 098

Managerial Ability + 0191 0181 0034 0023 0045

564 515 076 053 074

Change in Equity Managerial Ability + 0114 0113 0444 0440 0484

165 164 189 203 214

Ln(Market Value of Equity) 0023 0022 0038 0036 0060

366 363 440 412 639

Book-to-Market + 0066 0064 0115 0111 0153

466 427 352 343 348

Intercept 0283 0272 0518 0508 0522

915 847 773 767 834

Observations 76,155 88,083 20,423 19,005 13,974

Year indicators Included Included Included Included Included

AdjustedR2 005 005 006 006 007

Notes. This table presents results on the relation between external financing and managerial ability. Future Returns is the value-weighted industry-adjusted

return for the year following the measurement of change in equity and ability. Managerial Ability is the residual-based measure described in 4.3. Change

in Equityis the change in equity financing, defined as the change in equity (CEQ) plus the change in preferred stock (PSTK) less net income (NI) scaled by

average total assets (AT). Panel A presents regression results for Managerial Abilityand five other ability measures. Panel B presents regression results for

our managerial ability score restricting the sample to those used by the alternative ability measures in panel A. Variables are winsorized at the extreme 1%.

Robustt-statistics, with standard errors clustered by firm and year, are presented in parentheses.10%, 5%, and 1% statistical significance (two-tailed tests); 10%, 5%, and 1% statistical significance (two-tailed tests) in the direction opposite

to predictions.

alternative ability metricseven absent its calcula-tion within specific subsamplesif sample size issufficiently large. As noted previously, however, ifthe subsamples are too small and the underlyingdistribution differs from the population over whichour ability score is estimated, the scores explanatorypower is limited. Thus, in some analyses, researchers

may need to reestimate ability within their desiredsubsample.

To summarize, we demonstrate the usefulness ofour managerial ability measure in resolving researchpuzzles by focusing on the finding of underperfor-mance following seasoned stock issues. We documentthat firms that take on high levels of equity financing


18/20


tend to have lower returns in the future; however,these low returns are substantially mitigated whenthe equity is issued by more talented managers whoappear more able to identify and execute positive netpresent value projects and thus use the issue pro-ceeds more effectively. This illustrates one of many

potential applications for our measure of managerialability.

7. ConclusionResearch on managerial ability is active across manydisciplines and uses multiple proxies for talent, but allare affected to some extent by intervening factors suchas firm and industry attributes. We advance a moreprecise measure of managerial ability that is availablefor a large sample of firms and is comparable acrossindustries.

Using data envelopment analysis as a platform

to estimate firm efficiency, we quantify managerialability by distinguishing between managerial talentand a number of firm-driven effects on firm effi-ciency. We validate our managerial ability measure

by showing that (1) it has an economically signifi-cant association with manager fixed effects, (2) it isnegatively associated with the price reactions to CEOturnover announcements, and (3) it is positively asso-ciated with the subsequent performance at CEOs newappointments (where the score is measured in theirprior firms). Moreover, our ability measure outper-forms alternative ability measures on each of thesedimensions. Finally, to illustrate the usefulness of our

measure, we document that managerial ability miti-gates the negative association between equity financ-ing and future abnormal returns, a known puzzle inthe finance literature.

The measure of managerial ability, although animprovement over current measures, has limitations.First, there is measurement error in the inputs andoutput to firm efficiency; some accounting vari-ables contain measurement error, whereas others areunavailable. In addition, our second stage dampensvariation in ability, for example, by controlling forfirm size, because better managers are more likelyto be hired by larger firms (Rosen 1982). Future

researchers can expand the input and/or outputsets or the second-stage estimation to create a morerefined measure. Second, the measure is the residualfrom a model; however, a portion of this residualreflects factors that are not attributable to manage-rial ability. In robustness checks, we estimate the vari-able as a fitted value based on CEO fixed effects.This measure has a correlation of 0.835 with the man-agerial ability measure presented, providing a lower

bound of the proportion of the score attributable tothe manager. Another limitation is that our ability

score is estimated over a broad population, and thus,when the sample of interest has a different underly-ing distribution, the scores explanatory power is lim-ited. Thus, in some analyses, researchers may needto reestimate ability within their desired subsample.Finally, some of the DEA input variables, such as Net

PP&E, were determined by both the current managerand past managers, and thus the score is truly of boththe current and past management teams. Future stud-ies requiring an ability score for managers with feweryears of history might adapt the inputs to only thosethat can be changed in the short term, and insteadcontrol for the longer-term inputs that are more diffi-cult to change.

Despite these limitations, our measure of manage-rial ability is available for a broad cross section offirms, exhibits an economically significant manager-specific component, and contains less noise than exist-ing proxies of managerial ability. Future researchers

who use our measure or variations thereof may con-sider questions like whether better managers makesuperior dividend or share repurchase decisions,whether better managers select and execute higher-quality corporate acquisitions (e.g., Leverty and Qian2011) or issue higher quality disclosures (e.g., Baiket al. 2011), whether managers behave as if they areable to gauge their own ability, or whether the boardof directors and the market accurately price manage-rial ability (through compensation and stock price).And more fundamentally, what determines manage-rial ability (e.g., education, experience, social connec-

tions; Kaplan et al. 2012)? In sum, a more precise andmore widely available measure of managerial abilitywill allow us to expand our knowledge of the spe-cific role of management in the efficient allocation ofresources.

AcknowledgmentsThis paper has benefited greatly from comments and sug-gestions by three anonymous reviewers, an anonymousassociate editor, Mary Barth (the department editor), LindaBamber, Bryan Bonner, Larry Brown, Brian Cadman, MikeCooper, Asher Curtis, Dave Farber, Michal Goldberg, RachelHayes, Tyler Leverty, Melissa Lewis, Venky Nagar, Scott

Schaefer, and workshop participants at various conferences,Harvard University, the University of California at Berke-ley, the University of Indiana, the University of MissouriColumbia, the University of Oklahoma, and the Universityof Utah. The authors thank Sam Lee for providing themwith the media coverage data.

References

Baik, B., D. Farber, S. Lee. 2011. CEO ability and manage-ment earnings forecasts. Contemporary Accounting Res. 28(5)16451668.


19/20


Bamber, L., J. Jiang, I. Wang. 2010. Whats my style? The influenceof top managers on voluntary corporate financial disclosure.

Accounting Rev. 85(4) 11311162.

Banker, R., A. Charnes, W. Cooper. 1984. Some models for esti-mating technical and scale inefficiencies in data envelopmentanalysis. Management Sci. 30(9) 10781092.

Barr, R., T. Siems. 1997. Bank failure prediction using DEA to mea-

sure management quality. R. Barr, R. Helgason, J. Kennington,eds. Interfaces in Computer Science and Operations Research:Advances in Metaheuristics, Optimization and Stochastic ModelingTechnologies. Kluwer Acadmic Publishers, Boston, 341366.

Bennedsen, M., F. Perez-Gonzalez, D. Wolfenzon. 2010. Do CEOsmatter? Working paper, Copenhagen Business School, Copen-hagen.

Berk, J., R. Green. 2004. Mutual fund flows and performance inrational markets. J. Political Econom.112(6) 12691295.

Berk, J., R. Stanton. 2007. Managerial ability, compensation and theclosed-end fund discount. J. Finance 62(2) 529556.

Bertrand, M., S. Mullainathan. 2003. Enjoying the quiet life? Corpo-rate governance and managerial preferences. J. Political Econom.111(5) 10431075.

Bertrand, M., A. Schoar. 2003. Managing with style: The effect

of managers on firm policies. Quart. J. Econom. 118(4)11691208.

Billett, M., Y. Qian. 2008. Are overconfident managers bornor made? Evidence of self-attribution bias from frequentacquirers.Management Sci. 54(6) 10371051.

Bradshaw, M., S. Richardson, R. Sloan. 2006. The relation betweencorporate financing activities, analysts forecasts and stockreturns. J. Accounting Econom. 42(12) 5385.

Bushman, R., Q. Chen, E. Engel, A. Smith. 2004. Financial account-ing information, organizational complexity and corporate gov-ernance systems. J. Accounting Econom. 37(2) 167201.

Callen, J., M. Morel, C. Fader. 2005. Productivity measurement andthe relationship between plant performance and JIT intensity.Contemporary Accounting Res. 22(2) 271309.

Carter, M. E., F. Franco, I. Tuna. 2010. Premium pay for executive

talent: An empirical analysis. Working paper, Boston College,Boston.

Cazier, R., J. McInnis. 2010. Do firms contract efficiently on past per-formance when hiring external CEOs? Working paper, TexasChristian University, Fort Worth.

Charnes, A., W. Cooper, E. Rhodes. 1978. Measuring efficiency ofdecision making units. Eur. J. Oper. Res. 2(6) 429444.

Chatterjee, A., D. Hambrick. 2007. Its all about me: NarcissisticCEOs and their effects on company strategy and performance.

Admin. Sci. Quart.52(3) 351386.

Chemmanur, T., I. Paeglis. 2005. Management quality, certifica-tion, and initial public offerings. J. Financial Econom. 76(2)331368.

Curtis, A., M. Lewis. 2010. The comparability of accounting rates ofreturn under historical cost accounting. Working paper, Uni-

versity of Utah, Salt Lake City.DeAngelo, H., L. DeAngelo, R. M. Stulz. 2010. Seasoned equity

offerings, market timing, and the corporate lifecycle. J. Finan-cial Econom. 95(3) 275295.

Dechow, P., R. Sloan, A. Sweeney. 1996. Causes and consequencesof earnings manipulation: An analysis of firms subject to en-forcement actions by the SEC. Contemporary Accounting Res.13(1) 136.

Demerjian, P., B. Lev, M. Lewis, S. McVay. 2011. Managerial abil-ity and earnings quality. Working paper, Emory University,Atlanta.

Fama, E., K. French. 1997. Industry cost of equity. J. FinancialEconom.43(2) 153193.

Fama, E., J. MacBeth. 1973. Risk, return and equilibriumEmpiricaltests. J. Political Econom. 81(3) 607641.

Fee, C., C. Hadlock. 2003. Raids, rewards, and reputations inthe market for managerial talent. Rev. Financial Stud. 16(4)13151357.

Ge, W. 2006. Off-balance-sheet activities, earnings persistence andstock prices: Evidence from operating leases. Working paper,

University of Washington, Seattle.Ge, W., D. Matsumoto, J. Zhang. 2

Documents

Quantifying Managerial Ability