Self-selection, endogeneity, and the relationship between CEO duality and firm performance

Strategic Management JournalStrat. Mgmt. J., 30: 1092–1112 (2009)

Published online EarlyView in Wiley InterScience (www.interscience.wiley.com) DOI: 10.1002/smj.776

Received 27 December 2006; Final revision received 9 March 2009

SELF-SELECTION, ENDOGENEITY, AND THERELATIONSHIP BETWEEN CEO DUALITY AND FIRMPERFORMANCE

RAGHAVAN J. IYENGAR1 and ERNEST M. ZAMPELLI2*1 School of Business, North Carolina Central University, Durham, North Carolina,U.S.A.2 Department of Business and Economics, The Catholic University of America,Washington, District of Columbia, U.S.A.

This study focuses explicitly on the methodological implications of the endogenous theory ofgovernance as applied to firm performance. In particular, if firms choose their governancestructures as part of a constrained performance maximization process, then application of anappropriate empirical methodology should reveal statistical evidence of such behavior. In thisstudy we take advantage of the endogenous switching regression model framework to determinewhether such predicted optimizing behavior can be corroborated by the data. The model allowsus to test explicitly for selection behavior in accordance with comparative advantage and,concomitantly, the presence of selectivity bias, in estimating the impact of CEO duality on firmperformance. The selection and performance equations are modeled in accordance with theextant accounting, economics, and management literature on the impact of the dual governancestructure on firm performance. Overall, we tested four performance measures for the entiresample of firm-year observations as well as for the largest three industries in terms of samplesizes. The major finding, robust in all cases, is that there is no evidence to support a contentionthat CEO duality is a structure purposefully chosen for optimizing performance. If firms areindeed choosing the dual leadership structure, they are doing so for reasons other than improvingperformance from what it would be otherwise. In fact, for performance measured as marketreturn and earnings per share, there is evidence of a significant selectivity bias that acts tolower performance below what it would have been under random assignment. For performancemeasured by Tobin’s q and return on assets, we found neither evidence of selectivity bias, norany significant marginal performance impacts of CEO duality. Such findings are inconsistentwith an endogenous governance theory, at least when applied to firm performance. Copyright 2009 John Wiley & Sons, Ltd.

INTRODUCTION

Should a firm’s CEO also serve as the chair of itsboard of directors? The corporate scandals of thevery recent past have placed this question squarelyat the forefront of the public discourse on the

Keywords: duality; corporate governance; firm perfor-mance∗ Correspondence to: Ernest M. Zampelli, Department of Busi-ness and Economics, The Catholic University of America, 620Michigan Ave. NE., Washington, DC 20064, U.S.A.E-mail: [email protected]

nature and structure of corporate governance in theUnited States. The academic research on this issue,both theoretical and empirical, has provided mixedresults. Fama and Jensen (1983) and Jensen (1993)argue that CEO duality, that is, a firm’s CEO alsoserving as board chair, violates the rubric of theseparation of decision-management from decision-control. This in turn impedes a board’s abilityto monitor effectively a CEO’s decisions, leavinggreater opportunities for CEOs to advance theirown personal interests to the (possible) detrimentof the firm’s shareholders. However, Stoeberl and

Copyright 2009 John Wiley & Sons, Ltd.

Self-Selection, CEO Duality, and Firm Performance 1093

Sherony (1985) and Anderson and Anthony (1986)contend that CEO duality provides a single focalpoint for company leadership, thereby creating animage of firm stability, instilling confidence in thefirm’s management, and fostering better commu-nication between management and the board ofdirectors. This argument is echoed in Donaldsonand Davis (1991), Finkelstein and D’Aveni (1994),Dahya, Lonie, and Power (1996), Brickley, Coles,and Jarrell (1997), and Bhagat and Black (2001).In addition, Brickley et al. (1997) point out thatemploying a leadership structure that separates theroles of CEO and board chair is not costless. Theauthors identify information, agency, and incentivecosts of such a structure that are, in fact, ame-liorated by a structure where the CEO serves inthe dual capacities. The reduction in these costs asa consequence of CEO duality may outweigh thebenefits from a structure that separates the roles ofCEO and board chair.

The empirical evidence on the impact of CEOduality on firm performance is similarly mixed.In an analysis of Fortune 500 companies, Rechnerand Dalton (1991) find that firms with ‘indepen-dent’ leadership structures consistently outperformthose with a duality structure with respect to returnon investment (ROI), return on equity (ROE),and profit margins. In their study of the bankingindustry, Pi and Timme (1993) provide evidenceof higher accounting returns for banks that sepa-rate the roles of CEO and board chair. Consistentwith these results, Daily and Dalton (1994), inan examination of the relationship between corpo-rate governance/board composition variables andbankruptcy filings, report a strong and robust pos-itive association between CEO duality and firmbankruptcies. In contrast, Boyd (1995) estimatesthat after controlling for the interactions betweenduality and different organizational environments,duality has an independent and positive impacton subsequent firm performance. Sridharan andMarsinko (1997) investigate the impact of CEOduality and firm value in the paper and forest prod-ucts industry and find that firms with dual CEOsexhibit higher market values than their non-dualcounterparts. And Peng (2004), in a study of 405firms listed on the Shanghai and Shenzhen StockExchanges, reports a statistically robust positiveimpact of CEO duality on firm performance asmeasured by either ROE or sales growth. Baliga,Moyer, and Rao (1996) only add to the ambigu-ity in their analysis of the announcement effects

of changes in duality status. Specifically, theirresults indicate that the market does not respondto changes in duality status, that changes in dual-ity status do not affect the operating performanceof firms, and that there is, at best, only weak evi-dence of a link between CEO duality and long-termfirm performance. Finally, Faleye (2007) exam-ines the impact of the dual structure for a largesample of COMPUSTAT firms in 1995 on per-formance, accounting for the interactions of CEOduality with the mediating factors of firm com-plexity, CEO reputation, and governance structure.His findings, similar to but more compelling thanBoyd’s (1995), provide evidence that the impact ofduality depends on firm and CEO characteristics.

The overall lack of consensus among researcherscertainly provides some of the motivation forundertaking this project. In addition, however, isour concern that the prior empirical literature maysuffer from a serious methodological problem.Specifically, the typical empirical analysis incorpo-rates CEO duality into a firm performance equationthrough the use of a dummy (binary) variable indi-cating whether a firm’s CEO is board chair or not.The estimation technique, in turn, treats the vari-able as exogenous, implying that a firm’s decisionto institute (or not) a dual governance structure iseither a function of forces external to and out ofthe control of the firm or is simply the result ofrandom choice.1 This implicit ‘exogenous’ viewof the nature of the governance structure choiceis challenged, however, by the work of Herma-lin and Weisbach (1998) who formulate a modelin which a firm’s decisions regarding the struc-ture and composition of its board of directorsare looked upon as the solution to an organiza-tional design problem subject to, of course, legaland political constraints. In our particular context,then, this suggests that a firm’s selection of a dualgovernance structure results from rational choiceconsistent with constrained optimization behavior.More intuitively, the literature identifies both costsand benefits to firms of a dual leadership structureand, as noted by Faleye, ‘the appropriateness ofa particular leadership structure for a given firmdepends on how the firm’s characteristics influ-ence the balance between these costs and benefits

1 The exogeneity assumption is true even in Faleye (2007), whereprior to his analysis of duality on firm performance, he conductsprobit regressions to examine the relationship between firmcharacteristics and the choice of the dual governance structure.

Copyright 2009 John Wiley & Sons, Ltd. Strat. Mgmt. J., 30: 1092–1112 (2009)DOI: 10.1002/smj

1094 R. J. Iyengar and E. M. Zampelli

at the margin’ (Faleye, 2007: 242). Within suchan ‘endogenous’ framework, the results from priorempirical studies of the impact of CEO duality onfirm performance may be tainted by self-selectionbias. It is the intention of this study to exam-ine the causal relationship between CEO dualityand firm performance with an empirical method-ology that explicitly tests for the presence of aselection bias problem. The framework will allowus to first test explicitly for whether or not firmspurposefully choose a governance structure, and,secondly, if that choice is consistent with improvedperformance.

The remainder of the article is organized as fol-lows. We provide the theoretical background anddevelop the hypotheses to be tested, then describethe characteristics of a general self-selection modelthat may be usefully applied to the ‘endogenous’view of the duality decision. The model’s econo-metric specification and variable definitions arethen presented and the sample design and descrip-tive statistics follow. Next, we detail the multi-variate analysis, report the empirical results anddiscuss their implications. Results from a numberof robustness tests are also discussed. Finally, weoffer a brief summary and concluding remarks.

THEORETICAL BACKGROUND ANDHYPOTHESIS DEVELOPMENT

As noted in the introduction, there are two pri-mary theoretical perspectives on the relationshipbetween firm performance and the dual corpo-rate governance structure. From an agency theoryperspective, CEOs (the agents) are cast as utility-maximizing or risk-minimizing individuals whomake decisions that enhance their own personalwelfare at the expense of shareholders (the princi-pals). An effective counterbalance to such behaviorrests in an independent board of directors that mon-itors the decisions of CEOs to ensure that they arein the best interests of shareholders, that is, in agovernance structure that separates decision man-agement from decision control. When the CEOserves as the chair of the board of directors, thisseparation is breached and the board becomes lesseffective in its monitoring function. Consequently,a dual governance structure makes it more proba-ble that CEOs will be able to implement decisionsthat enhance their personal welfare at the expenseof shareholders’ wealth. This then naturally leads

to the hypothesis that CEO duality will negativelyimpact a firm’s performance.

In contrast, other research by organizationaleconomists and management theorists contendsthat agency theory is excessively restrictive in itsview of the CEO as an opportunistic personal wel-fare maximizer. Specifically, it argues that agencytheory ignores a vast array of alternative moti-vations, for example, achievement, recognition,respect, reputation, altruism, and so on, that pro-vide a CEO with the incentive to do the besthe/she can in the responsible stewardship of theassets of the firm. Additionally, organizational andmanagement theorists have argued that the dualgovernance structure provides focused leadership,greater decision accountability, the flexibility tomake strategic decisions more quickly in responseto market/environmental changes, and an effec-tive counterweight to the power of special inter-ests. From this theoretical perspective, the logicalhypothesis is that CEO duality will have a positiveeffect on firm performance.

The major empirical literature on the duality-performance relationship over the last 25 years orso has been motivated by one or both of these theo-retical perspectives and has yielded equivocal andconflicting results. For example, Berg and Smith(1978) analyzed the relationship from the agencyperspective for a sample of Fortune 200 firmsand reported both positive and negative significanteffects of duality on firm performance. Chaganti,Mahajan, and Sharma (1985) examined a sampleof 21 matched pairs of bankrupt and non-bankruptretail firms in the United States using the agencyperspective and reported no significant relation-ship between CEO duality and performance. Rech-ner and Dalton (1991) and Pi and Timme (1993)employ both perspectives and find significant nega-tive relationships between duality and firm perfor-mance. In contrast, relying on both perspectives,a study by Sridharan and Marsinko (1997) of 18firms in the U.S. paper and forest products indus-try yielded a significant positive impact of CEOduality on firm performance.

A number of other studies might also be citedas evidence that the empirical work thus far hasfailed to provide compelling evidence regardingthe duality-performance relationship from eitherthe agency or stewardship perspectives. Referencesfor 30 such papers can be found in Kang and



Zardkoohi (2005).2 But more importantly, Kangand Zardkoohi (2005) identify a critical conceptualissue that has been generally ignored by the afore-mentioned empirical research and that is consistentwith the ‘endogenous’ framework of Hermalin andWeisbach (1998). Specifically, ‘duality is not a ran-dom phenomenon, but an organizational practicethat is adopted under appropriate or inappropri-ate conditions’ (Kang and Zardkoohi, 2005: 786).In this regard, any examination of the duality-performance relationship must at least attempt tointegrate the literature on the antecedents of dual-ity, that is, those factors that help to explain thefirm’s choice of a dual governance structure and, inturn, may moderate the impact of duality on firmperformance.3 As noted by Kang and Zardkoohi(2005), this literature identifies five antecedentsof duality: institutional, power, social reciprocity,reward, and organizational. The institutional expla-nation suggests that a firm’s board may choosethe dual structure as a response to industry insti-tutional pressures, specifically because the prac-tice is commonplace in the industry. The powerantecedent explains the choice as the result ofa powerful CEO exercising his/her will over arelatively weak and less vigilant board of direc-tors. Social reciprocity implies that top executivesfrom firms with a dual governance structure willlikely favor such a structure in their capacitiesas directors on the boards of other firms. Thereward explanation simply suggests a CEO maybe appointed as board chair as a reward for goodperformance and to demonstrate the board’s confi-dence in the competence and stewardship of its topexecutive. And finally, the dual structure may bechosen as an organizational solution for firms thatare internally complex and/or for firms that oper-ate in environments characterized by high levelsof uncertainty where speedy decision making andflexibility are critical to competitiveness, growth,and survival.

The general hypothesis emanating from the con-ceptual framework underlying the works of Her-malin and Weisbach (1998), Kang and Zardkoohi(2005), Faleye (2007), and others is clear: afirm’s choice of a dual governance structure is

2 A useful review of the empirical literature on the more generaltopic of board composition and firm performance is offered byHermalin and Weisbach (2003).3 This literature includes Harrison, Torres, and Kukalis (1988),Finkelstein and D’Aveni (1994), Westphal and Zajac (1997),Vancil (1987), and Boyd (1995).

not random but strategic and is determined aspart of a broader constrained optimization pro-cess. To test the hypothesis, of course, requiresthe specification of an objective function, thatis, the metric to be optimized. For our pur-poses, that metric is firm performance. Corre-spondingly, a major hypothesis to be tested in thisstudy is:

Hypothesis 1: Ceteris paribus, a firm chooses itsgovernance structure according to comparativeadvantage in performance.

If the hypothesis is rejected, the study willsubsequently test the following:

Hypothesis 2: There is no selection bias affect-ing the relationship between a firm’s choice ofgovernance structure and its performance.

Tests of Hypothesis 1 and Hypothesis 2 will bebased on an econometric specification that explic-itly recognizes the (potential) nonrandomness ofthe governance structure decision.

Before proceeding to a description of our model,we note that there is one recent study that exam-ines the relationship between CEO duality and firmperformance within the selection bias framework.Chen, Lin, and Yi (2008) examines the relation-ship between CEO duality and Tobin’s q for asample of firms over the 1999–2003 period usinga standard two-step treatment effects model. Forthe full sample, their results indicate no indepen-dent effect of CEO duality on firm performance,but do show the presence of a positive selectionbias, supporting the hypothesis that firms choosetheir leadership structure to improve their perfor-mance. For a fairly small subsample of firms thatchanged leadership structure from dual to non-dualor vice versa, they continue to find no indepen-dent effect of duality on Tobin’s q. Surprisingly,however, for this subsample of firms they find anegative selection bias. Differences between theirand our results are likely due to methodologi-cal and/or sample differences that will becomeclear from the discussions and descriptions thatfollow.



A SELF-SELECTION MODEL OF FIRMPERFORMANCE AND DUALITY4

Consider a situation in which one wishes to modelthe ‘benefits’ to firms of choosing a dual gover-nance structure. The general model that is oftenused is:

Yj = Xjβ + δIj + εj (1)

where Y is firm performance, X is a vector offirm, industry, and/or environmental characteris-tics, I is a dummy variable equal to one for firmswith a dual governance structure and zero other-wise, and ε is a random error term. The ceterisparibus effect of CEO duality on firm perfor-mance is given by δ. Effectively, the dual gov-ernance structure is assumed to shift exogenouslythe firm’s performance function. Other forms ofEquation (1), as found in Faleye (2007), permitthe slope coefficients to be affected by the dualgovernance structure as well by including the inter-action of I with the X vector. This latter form ofthe Equation (1) model is more general, of course,and in accordance with the theoretical literaturethat suggests certain firm, industry, and/or environ-mental factors may ‘modify’ the impact of dualityon firm performance. It is instructive at this pointto think of the more general alternative form ofEquation (1) as tantamount to the two equationsystem given by:

Y1j = Xjβ1 + ε1j (performance of firms with

dual governance structure) (2a)

Y2j = Xjβ2 + ε2j (performance of firms with

non - dual governance structure) (2b)

OLS estimation of Equations (2a) and (2b), orequivalently of Equation (1) expanded to includeinteractions of Ij with Xj , will yield unbiased andconsistent estimates if Ij is exogenous. If, however,the choice of governance structure is not randombut is motivated by firm-specific and other factorsimportant to the constrained maximization of firmperformance, then Ij will not be exogenous andthe ordinary least squares (OLS) estimates willbe subject to selectivity bias. In such a case, a

4 The discussion in this section borrows liberally from chapters5, 8, and 9 of Maddala (1983).

more appropriate specification is offered by thefollowing endogenous switching regression model:

Y1j = Xjβ1 + ε1j (firm performance with

dual governance structure)

Y2j = Xjβ2 + ε2j (firm performance with

non - dual governance structure)

I ∗j = Zjγ − uj (governance structure

selection function)

Ij = 1 iff I ∗j > 0

Ij = 0 iff I ∗j ≤ 0. (3)

I ∗j is an unobserved latent variable measuring the

benefits of a dual governance structure and Zj isa vector of observable firm, industry, and/or envi-ronmental variables that affect the firm’s choiceof governance structure. The observed Yj aredefined as:

Yj = Y1j iff Ij = 1 (4)

Yj = Y2j iff Ij = 0. (5)

The random errors are assumed to be trivariate nor-mally distributed with zero means and covariancematrix:

� =[

σ11 σ12 σ1u

σ12 σ22 σ2u

σ1u σ2u 1

](6)

The model permits the interaction of both observ-able and unobservable firm characteristics with thechoice of governance structure and, unlike the con-ventional formulation represented by Equations (1)[or (2a) and (2b)], is suitable for determiningwhether or not there is statistical evidence that afirm’s choice of governance structure is part ofa rational constrained optimization process whoseobjective is maximum performance.

The expectations of firm performance condi-tional on the chosen governance structures can bewritten as:

E(Y1j |Ij = 1) = Xjβ1 + E(ε1j |Ij = 1)

= X1jβ1 − σ1u

(ϕ(Zjγ )

�(Zjγ )

)(7)

E(Y2j |Ij = 0) = Xjβ2 + E(ε2j |Ij = 0)



= X2jβ2 + σ2u

(ϕ(Zjγ )

1 − �(Zjγ )

)(8)

where ϕ(Zjγ ) and �(Zjγ ) are the standard nor-mal density and cumulative distribution functions,respectively. Define H1j = ϕ(Zjγ )/�(Zjγ ) andH2j = ϕ(Zjγ )/[1 − �(Zjγ )], and rewrite (7) and(8) as:

E(Y1j |Ij = 1) = Xjβ1 − σ1uH1j (9)

E(Y2j |Ij = 0) = Xjβ2 + σ2uH2j (10)

Equations (9) and (10) make it clear that the selec-tion bias that results from the OLS estimation ofEquations (2a) and (2b) is, in effect, an omit-ted variables problem where the omitted variablesare H1j and H2j . In other words, the endogene-ity arises from a stochastic sorting process, thenature of which is embedded in the covariancesbetween the unobservables in the performance andgovernance structure selection equations. To exam-ine the nature of self-selection and its implicationsfor the parameters of Equations (9) and (10), onemust consider as well the expected value of firmperformance conditioned on the governance struc-ture not chosen by the firm. In other words, oneneeds to know the expected performance of firmswith a dual (non-dual) structure had they chosen anon-dual (dual) structure. These conditional expec-tations are given by:

E(Y2j |Ij = 1) = Xjβ2 − σ2uH1j (11)

E(Y1j |Ij = 0) = Xjβ1 + σ1uH2j (12)

where (11) represents the mean performance ofdual governance structure firms had they chosena non-dual governance structure and (12) repre-sents the mean performance of non-dual gover-nance structure firms had they chosen a dual gover-nance structure. The Xjβi terms on the right-handsides of Equations (9)–(12) can be interpreted asthe mean firm performance under a random assign-ment of governance structures among the pop-ulation of firms. Subtracting Equation (11) from(9) and Equation (12) from (10) yields:

E(Y1j |Ij = 1) − E(Y2j |Ij = 1) = Xj(β1

− β2) + (σ2u − σ1u)H1j (13)

E(Y2j |Ij = 0) − E(Y1j |Ij = 0) = Xj(β2

− β1) + (σ2u − σ1u)H2j (14)

Equations (13) and (14) demonstrate that if theself-selection process is based on comparativeadvantage in performance, then (σ2u − σ1u) mustbe greater than zero, that is, performance willbe higher with self-selection than with randomassignment.

Keeping in mind that (σ2u − σ1u) is a cru-cial parameter to be estimated, we consider theunconditional expected performance of a randomlyselected firm given by:

E(Yj) = E(Yj |Ij = 1) × P(Ij = 1)

+ E(Yj |Ij = 0) × P(Ij = 0) (15)

where P(Ij = 1) = �(Zjγ ) and P(Ij = 0) = 1 −�(Zjγ ). Multiplying and recognizing that the Xj

vector is common to all firms, Equation (15) canbe rewritten as:

E(Yj) = Xjβ2 + Xj�(Zjγ )(β1 − β2)

+ ϕ(Zjγ )(σ2u − σ1u) (16)

Using all observations, Equation (16) can be esti-mated in two stages. In the first stage, probitmaximum likelihood (ML) is used to obtain anestimate of γ , γ , which in turn can be used toobtain estimates of � and ϕ. These are insertedinto Equation (16), which can then be estimated byOLS. If (β1 − β2) �= 0 and (σ2u − σ1u) �= 0, thenthe correct model is given by Equation (3), that is,a two-regime performance model with endogenousswitching. In this case, the parameters β1, β2, σ2u,and σ1u can be obtained from the two-stage estima-tions of Equations (9) and (10). In the first stage,probit ML is used to obtain γ , which in turn allowsus to calculate H1j and H2j . In the second stage,Equations (9) and (10) can be estimated separatelyby OLS, substituting H1j and H2j for H1j and H2j .Or equivalently, they can be estimated as a singleequation by including Ij as an endogenous shiftdummy, the Xj variables, interactions of Ij withXj , the variable Hj defined as Ij × H1j + (1 −Ij ) × H2j , and the interaction of Ij with Hj . If,however, (β1 − β2) = 0 and (σ2u − σ1u) = 0, thenthe correct specification is a single regime perfor-mance model similar to Equation (1) with Ij as anendogenous shift dummy and including the vari-able Hj to account for the selectivity problem.

Two remaining possibilities from the estimationof Equation (16) are worth mentioning: [(β1 −β2) = 0 and (σ2u − σ1u) �= 0] and [(β1 −β2) �= 0



and (σ2u − σ1u) = 0]. In both instances, the two-regime performance model with endogenous swit-ching still applies. In the former case, Equations(9) and (10) can be estimated as a single equationwith Ij as an endogenous shift dummy, the Xj

variables, the variable Hj , and the interaction of Ij

with Hj . In the latter case, the estimating equationincludes Ij as an endogenous shift dummy, the Xj

vector, interactions of Ij with Xj , and the variableHj .

For our purposes, a finding that (σ2u − σ1u) �= 0from the two-stage estimation of Equation (16)would provide evidence in support of our hypothe-sis that firms self-select into the governance struc-ture in accordance with comparative advantage.In principle, subsequent estimates of σ2u and σ1u

obtained from the estimation of Equations (9) and(10) would be unnecessary, though they wouldprovide additional insight into the self-selectionprocess. For example, if both σ2u and σ1u werefound to be less than zero, this would suggest thatdual (non-dual) structured firms would performbetter (worse) than average under either gover-nance structure, but do relatively better under thedual (non-dual) governance structure. In contrast,if the estimation of Equation (16) leads to a con-clusion that (σ2u − σ1u) = 0, our hypothesis wouldnot be supported but subsequent estimations ofEquations (9) and (10) would be imperative. Thereason is straightforward—if subsequent estima-tions cannot reject the hypothesis that σ1u = σ2u =0, then there is no evidence of selection bias andthe conventionally used methodologies to deter-mine the performance impact of the dual gover-nance structure are appropriate.

ECONOMETRIC SPECIFICATION ANDVARIABLE DEFINITIONS

The operational version of the endogenous switch-ing regression model described previously byEquation (3) relies mainly on the extant literatureon the antecedents of duality and the relationshipbetween CEO duality and firm performance.

Selection equation specification

The dependent variable, DUAL, is dichotomousand equal to one for firms with a dual gover-nance structure and equal to zero for firms with

a non-dual governance structure. Variables influ-encing the dual/non-dual decision were chosen inaccordance with the antecedents to duality litera-ture. In particular, explanatory variables associatedwith duality as a solution to internal organiza-tional complexity include the size of the firm asproxied by the natural log of sales (LNSALES ),the total number of employees (EMP ), the num-ber of business segments (SEGNUM ), the ratio ofnet property, plant, and equipment to total assets(PPEAT ), and the growth opportunities of the firmas proxied by the realized growth of sales revenue(REVGWTH ).5

Similar to Boyd (1995), we include measuresof growth in industry sales (INDSLG5 ) and thevolatility in industry sales (INDSLV5 ) over themost recent previous five-year period to accountfor the possibility that duality is chosen as anorganizational solution to environmental uncertain-ties. More specifically, the two variables, respec-tively, are proxies for munificence, which refersto the abundance of resources in the environ-ment, and dynamism, which measures environ-mental volatility. Low munificence/high dynamismenvironments lead to greater uncertainties that, inturn, may require the faster response, greater flex-ibility, and greater accountability afforded by thedual governance structure.

To account for the ‘power’ explanation of CEOduality, we include the percentage of the firm’scommon stock outstanding that is owned by theCEO (OWN ), board size (BDSIZE ), and boardindependence as proxied by the percent of boarddirectors who are unaffiliated with the firm(INDBD).6 The predicted impact of OWN on dual-ity is a priori ambiguous in that it can mea-sure both the degree of entrenchment and thedegree of alignment of the CEO’s interests withshareholders’ interests. In line with Jensen (1993),the expected sign on board size is negative aslarger boards are probably less effective in com-municating and reaching consensus and, hence,more susceptible to manipulation by the CEO.

5 PPEAT is assumed to be inversely related to the ratio ofintangible assets to total assets. Hence, firms with smaller PPEATwould be considered more complex. Though a common measureof growth opportunities is the market-to-book ratio, we chooseto exclude it since it could be also be interpreted as measure offirm performance.6 CEO tenure is not included here as it is likely endogenous withrespect to duality as noted by Faleye (2007)—CEOs who serveas board chairs are more difficult to remove and hence are likelyto have longer tenures.



Board independence is expected to increase theprobability of CEO duality. Since independentboards are more effective at monitoring CEOs, theadverse consequences that could result from theseparation of decision-management from decision-control become less probable. We also includethree other board ‘demographic’ composition vari-ables to account for the potential influences on theselection process of unique perspectives on orga-nization and management brought to the boardby different subgroups. These include the num-ber of employee directors (EMPDIRS ), the numberof directors who are members of a racial/ethnicminority group (MINONBRD), and the number offemale board directors (FEMONBRD). We makeno a priori claims regarding the direction of theirimpacts on the choice of a dual governance struc-ture. Finally, industry- and time-specific effectsare controlled for with industry and year dummyvariables.7

Performance equation specification

Because there exists no single, unique measureof firm performance, we employ four alternatives.These include the one-year total market return toshareholders, including the reinvestment of div-idends (MKTRET ), Tobin’s q calculated as theratio of the sum of the market value of commonequity and the book values of preferred equityand long-term debt to the book value of assets(TOBINSQ), return on assets (ROA), and earn-ings per share deflated by beginning-of-year stockprice (EPS ). We assume that an individual firm’s

7 Firm-specific dummy variables are not included for two rea-sons. First, severe to near-fatal colinearity problems are likelyto cause serious convergence/computational problems, since formost firms in the sample the binary variable DUAL is timeinvariant and other variables, such as board size (BDSIZE ),board independence (INDBD), the number of employees (EMP ),and the number of distinct business segments (SEGNUM ) arequasi-time invariant. Second, even in the absence of (quasi-)time invariant variables, fixed effects estimation of the currentmodel would be susceptible to the incidental parameters prob-lem that plagues many nonlinear panel data models, especiallythose including probit selection equations. In particular, whenthe number of time periods (T) is small and the number of firms(N) is large, as in the present case (T ≤ 5 and N = 531), theestimation of firm-specific effects will render estimates of themodel’s structural parameters inconsistent. Additionally, therewould probably be a small T sample bias in the estimates aswell. Since predictions from the probit selection equation areused to construct estimates of the hazard variables for inclusionin the performance equations, we think it best that such problemsbe avoided. See Greene (2003) and Wooldridge (2002).

performance can be explained in part by the over-all performance of the industry of which it isa part. These are labeled INDMKTRET, INDTO-BINSQ, INDROA, and INDEPS, respectively, andtheir marginal performance impacts are expectedto be positive. We also assume that past perfor-mance may be related to current performance andinclude the one-year lagged values of the perfor-mance measures labeled MKTRETLAG, TOBIN-SQLAG, ROALAG, and EPSLAG. Other explana-tory variables are taken from the past literatureand include firm size as measured by the natu-ral log of sales (LNSALES ), future growth oppor-tunities as measured by realized revenue growth(REVGWTH ), leverage as measured by the ratioof long-term debt to total assets (DAT ), man-agerial ownership (OWN ), board size and boardindependence (BDSIZE and INDBD), and div-idends paid out as a percent of net operatingincome (DVPOR). DAT and DVPOR are includedsince both can be used to limit managerial dis-cretion by reducing the size of free cash flow.To account for the potential impact of past per-formance on current performance, the right-handsides also include the one-year lags of the perfor-mance variables, labeled as MKTRETLAG, TOBIN-SQLAG, ROALAG, and EPSLAG. Industry andyear dummies are included in all performanceequations.

Structural model specification

Given the discussion of subsections Selectionequation specification and Performance equationspecification and in accordance with the modelspecified in Equation (3), the probit selection andperformance equations can be written as:

P(DUALj = 1|Zj) = �(γ0

+ γ1LNSALESj + γ2PPEATj

+ γ3EMPj + γ4SEGNUMj

+ γ5REV GWT Hj + γ6OWNj

+ γ7BDSIZEj + γ8INDBDj

+ γ9EMPDIRS + γ10MINONBRDj

+ γ11FEMONBRDj + γ12INDSLG5j

+ γ13INDSLV 5j − uj ) (17)

PERFij = βi0 + βi1PERFLAGj

+ βi2INDPERFj + βi3LNSALESj



Table 1. Selection of sample of firm years 1995–2003

Number of firm years

Number of nonfinancial, nonutility, firm-year observations from ExecuComp database forthe sample period

3,153

Less: firm years1) with insufficient financial data in Compustat database (1,039)2) with insufficient governance data in IRRC database or in proxy statements (127)3) with insufficient data on CEO stock ownership in ExecuComp database (107)

Number of firm-year observations in the final sample 1,880

+ βi4DATj + βi5REV GWT Hj

+ βi6DV PORj + βi7OWNj

+ βi8BDSIZEj + βi9INDBDj + εij (18)

where PERF = {MKT RET or T OBINSQ orROA or EPS}, subscript i = one for dual firmsand two for non-dual firms, and where γ0 and βi0

are assumed to be linear functions of industry- andyear-specific dummy variables.

RESEARCH DESIGN ANDDESCRIPTIVE STATISTICS

Sample

To compile the sample of firms, we began with thepopulation of nonfinancial, nonutility firms in theExecuComp database for the period 1995–2003.The sample was restricted to those firms in whichthere was no change in the duality structure overthe sample period.8 This initial sample consistedof 3,153 firm-year observations. From this popu-lation, we deleted 1,039 observations with insuf-ficient financial data in the Compustat database.We then discarded 127 observations due to lackof governance data in the Investor ResponsibilityResearch Center (IRRC ) database. An additional107 firm-year observations were excluded becauseof insufficient data about CEO stock ownership inthe ExecuComp database. Table 1 details how theselection criteria resulted in a final sample of 1,880firm-year observations.

8 The model as structured is inappropriate for dealing with firmsthat change their leadership structures during the sample period.For a study that examines the relationship between performancechanges and changes in leadership structure within a selectionbias framework, the reader is referred to Chen et al. (2008).

Descriptive statistics

Table 2 presents the distribution of our sample byindustry and median values of main variables inour study. Table 2, Panel A, reports the medianvalues of the dependent variables, while Table 2,Panel B, presents the median values of the inde-pendent variables. The industry distribution of oursample is similar to prior studies using comparablesample evidence (Frankel, Johnson, and Nelson,2002; Whisenant, Sankaraguruswamy, and Raghu-nandan, 2003).

On average, companies in the pharmaceuticaland extractive industries have performed best witha median Tobin’s q of 3.15 and a median marketreturn of over 18 percent, respectively, while firmsin the mining and computer industries have donethe worst with a Tobin’s q of 0.97 and marketreturn of −3.54 percent, respectively. In terms ofaccounting performance, retail industry firms haveperformed exceedingly well with an average (i.e.,median) ROA of 7.12 percent while the computerindustry has been the laggard with an averageROA of 4.40 percent and EPS of 0.02. Companiescomprising the chemical industry have the largestboard (10.00), headed by executives who havethe lowest median stock ownership (0.00), but thehighest median dividend payout ratio (33.01%).On the other hand, companies in the computerindustry have the smallest board (7.00), the lowestleverage (0.03), and the lowest median dividendpayout ratio (0%). Also, pharmaceutical firms havethe lowest sales on average ($507 million), but thehighest growth in revenue of 16.12 percent.

We employ a Wilcoxon rank-sum test9 to testfor significant differences in corporate governance,

9 Wilcoxon rank-sum test is a nonparametric two-sample test thatis solely based on the order (rank) in which the observations fromthe two samples fall.



Tabl

e2.

Des

crip

tive

stat

istic

sPa

nel

A—

Med

ians

ofde

pend

ent

vari

able

s

Indu

stry

Num

ber

offir

mye

ars

MK

TR

ET

TO

BIN

SQR

OA

EPS

Min

ing/

cons

truc

tion

5010

.09

0.97

6.27

0.10

Text

iles

prin

t/pu

blis

h15

44.

841.

145.

740.

06C

hem

ical

s97

10.0

71.

154.

930.

05Ph

arm

aceu

tical

s75

7.55

3.15

6.44

0.03

Ext

ract

ive

9418

.07

1.23

5.19

0.05

Dur

able

607

11.6

91.

265.

020.

05C

ompu

ters

343

−3.5

41.

954.

400.

02R

etai

l27

08.

091.

447.

120.

06Se

rvic

es19

04.

341.

285.

060.

04O

vera

ll18

808.

881.

385.

380.

05

Pane

lB

—M

edia

nsof

inde

pend

ent

vari

able

s

Indu

stry

IND

MK

TR

ET

IND

TO

BIN

SQIN

DR

OA

IND

EPS

SAL

ES

RE

VG

WT

HO

WN

BD

SIZ

ED

AT

DV

POR

Min

ing/

cons

truc

tion

27.7

51.

016.

850.

1126

8314

.64

0.00

9.50

0.32

3.61

Text

iles

prin

t/pub

lish

9.07

1.32

6.06

0.05

1571

2.67

1.20

9.00

0.29

22.3

1C

hem

ical

s9.

891.

314.

200.

0418

062.

860.

0010

.00

0.28

33.0

1Ph

arm

aceu

tical

s30

.54

3.89

1.67

−0.0

050

716

.12

1.90

8.00

0.19

0.00

Ext

ract

ive

34.9

81.

364.

730.

0498

614

.50

1.15

8.00

0.26

7.33

Dur

able

9.75

1.39

4.54

0.04

1073

5.31

1.70

9.00

0.22

8.19

Com

pute

rs−5

.94

2.36

1.48

−0.0

064

18.

572.

107.

000.

030.

00R

etai

l10

.07

1.74

7.95

0.06

2737

10.1

71.

509.

000.

190.

00Se

rvic

es8.

301.

544.

540.

0485

86.

312.

758.

000.

150.

00O

vera

ll9.

461.

554.

610.

0411

937.

081.

708.

000.

200.

00

Indu

stry

mem

bers

hip

isde

term

ined

bySI

Cco

deas

follo

ws:

min

ing

and

cons

truc

tion

(100

0–

1999

,ex

clud

ing

1300

–13

99),

text

iles

and

prin

ting/

publ

ishi

ng(2

200

–27

99),

chem

ical

s(2

800

–28

24,

2840

–28

99),

phar

mac

eutic

als

(283

0–

2836

),ex

trac

tive

(290

0–

2999

,13

00–

1399

),du

rabl

em

anuf

actu

rers

(300

0–

3999

,ex

clud

ing

3570

–35

79an

d36

70–

3679

),co

mpu

ters

(737

0–

7379

,35

70–

3579

,36

70–

3679

),re

tail

(500

0–

5999

),an

dse

rvic

es(7

000

–89

99),

excl

udin

g73

70–

7379

).Fi

nanc

ial

serv

ices

(600

0–

6999

)fir

ms,

tran

spor

tatio

n(4

000

–47

99),

utili

ties

(490

0–

4999

),fo

od(2

000

–21

99)

and

othe

rs(0

00–

0999

,90

00–

9999

)ar

eex

clud

edfr

omth

esa

mpl

e.T

his

clas

sific

atio

nis

follo

wed

inth

epr

ior

liter

atur

e[s

ee,

Whi

sena

ntet

al.

(200

3)].

TOB

IN’s

qis

defin

edas

follo

ws

=(m

arke

tva

lue

ofeq

uity

+bo

okva

lue

ofpr

efer

red

stoc

k+

book

valu

eof

debt

)/bo

okva

lue

ofto

tal

asse

ts;

MK

TR

ET

ison

eye

arto

tal

retu

rnto

shar

ehol

ders

,in

clud

ing

rein

vest

men

tof

divi

dend

s;R

OA

(%):

retu

rnon

asse

tsde

fined

asin

com

ebe

fore

extr

aord

inar

yite

ms

divi

ded

byto

tal

asse

ts;

EP

S:ea

rnin

gspe

rsh

are

(sca

led

bybe

ginn

ing-

of-p

erio

dsh

are

pric

e);

IND

MK

T,

IND

TOB

INSQ

,IN

DR

OA

,IN

DE

PS

are,

resp

ectiv

ely,

the

aver

age

(i.e

.,m

ean)

mar

ket

retu

rn,

TOB

IN’S

q,R

OA

,an

dE

PS

ofal

lfir

ms

with

inth

esa

me

four

-dig

itSI

Cco

deas

the

expe

rim

enta

lfir

m;

RE

VG

WT

His

the

aver

age

grow

thin

sale

sov

erth

epr

eced

ing

five-

year

peri

od;

OW

Nis

the

perc

enta

geof

com

mon

stoc

kshe

ldby

CE

O;

BD

SIZ

Eis

the

num

ber

ofbo

ard

mem

bers

ofth

efir

m;

DA

Tis

tota

lde

btov

erto

tal

asse

ts;

and

DV

PO

Ris

the

divi

dend

payo

utra

tiode

fined

asth

epe

rcen

tage

ofin

com

ebe

fore

extr

aord

inar

yite

ms

dist

ribu

ted

asca

shdi

vide

nds.



Table 3. Descriptive statistics and two-sample tests on DUAL and NON-DUAL firms for the period 1995–2003

Independent variable NON-DUAL firms (n = 390)Mean (Median)

DUAL firms (n = 1490)Mean (Median)

Wilcoxon Z(Prob > |Z|)

MKTRET (%) 22.361 17.447 1.083(12.218) (8.436) (0.279)

TOBIN’S q 2.215 1.914 2.275∗∗

(1.454) (1.375) (0.023)∗∗

ROA (%) 4.240 4.019 0.910(5.775) (5.275) (0.363)

EPS 0.036 0.027 −0.485(0.043) (0.047) (0.628)

INDMKTRET (%) 24.982 23.874 −0.516(7.735) (9.545) (0.606)

INDTOBINSQ 2.087 1.881 2.841∗

(1.701) (1.521) (0.005)∗

INDROA (%) 3.974 3.650 0.771(4.780) (4.538) (0.441)

INDEPS 0.031 0.024 0.320(0.036) (0.039) (0.749)

SALES 5534 4422 −5.324∗

(762) (1303) (0.000)∗

REVGWTH (%) 12.474 9.330 2.276∗∗

(9.513) (6.313) (0.023)∗∗

OWN (%) 2.061 5.675 −7.620∗

(1.10) (1.800) (0.000)∗

BDSIZE 8.803 8.683 0.794(9.00) (8.00) (0.427)

DAT 0.183 0.217 −3.368∗

(0.164) (0.216) (0.001)∗

DVPOR (%) 12.878 54.23 −2.840∗

(0.000) (0.000) (0.005)∗

Notes:∗∗ , ∗ indicate two-tailed significance at the 0.05 and 0.01 levels respectively.TOBIN’s q is defined as follows = (market value of equity + book value of preferred stock + book value of debt)/book value oftotal assets; MKTRET is one year total return to shareholders, including reinvestment of dividends; ROA(%): return on assets definedas income before extraordinary items divided by total assets; EPS: earnings per share (scaled by beginning-of-period share price);INDMKT, INDTOBINSQ, INDROA, INDEPS are, respectively, the average (i.e., mean) market return, TOBIN’S q, ROA, and EPS ofall firms within the same four-digit SIC code as the experimental firm; REVGWTH is the average growth in sales over the precedingfive-year period; OWN is the percentage of common stocks held by CEO; BDSIZE is the number of board members of the firm;DAT is total debt over total assets; and DVPOR is the dividend payout ratio defined as the percentage of income before extraordinaryitems distributed as cash dividends

financial, and all other explanatory variablesbetween firms whose CEOs are not the chairof the board (DUAL = 0) and whose CEOs are(DUAL = 1). Table 3 offers a glimpse of the dif-ferences in these variables between non-dual (n =390) and dual firm-year observations (n = 1, 490).Results from the two-sample difference test sug-gest that dual and non-dual firms are significantlydifferent along several dimensions. For instance,dual firms, on average, have, lower Tobin’s q,higher leverage, lower revenue growth and lowerdividend payout ratio, and are managed by CEOswith substantially larger stock ownership. Despitethese differences, we need to be cautious before

drawing conclusions based on these univariatestatistics.

MULTIVARIATE ANALYSIS

As noted above, the estimation of the model’sstructural parameters will proceed in a number ofstages. In the first stage, Equation (17) is estimatedby ML with standard errors robust to heteroskedas-ticity and autocorrelation. The results are reportedin Table 4. The positive and significant coefficientson LNSALES and PPEAT are consistent with thehypothesis that larger, more complex firms benefit



from focus, stability, and better communicationoffered by the dual governance structure. Incon-sistent with this hypothesis are the negative andsignificant coefficients for the number of employ-ees, EMP, and future growth opportunities as prox-ied by REVGWTH. The results also indicate apositive and significant impact of CEO stock own-ership, OWN, on the likelihood of a dual struc-ture. This finding provides some support for thecontention that a CEO’s stock ownership alignsmanagerial interest with shareholder interest andhelps to lessen the agency costs that might arisefrom appointing the CEO as board chair. Con-versely, it is also consistent with the power hypoth-esis—greater ownership provides the CEO withthe leverage necessary to win appointment as theboard’s chair. Also, as expected, larger boards(BDSIZE ) reduce the probability of selecting adual governance structure since larger boards maybe more easily manipulated by a CEO who alsoserves as board chair. Consistent with this is thefinding that more independent boards (INDBD)increase the likelihood of a dual structure sincethey more effectively monitor the actions of theCEO. More minority directors (MINONBRD) arefound to raise the probability of the dual structure,while interestingly, a greater number of employeedirectors (EMPDIRS ) are shown to reduce it. Thelatter result may signal a reluctance on the partof a firm’s employees to confer ‘too much’ powerand authority upon the firm’s executives. Finally,environmental munificence (INDSLG5 ) and uncer-tainty (INDSLV5 ) seem to play no significant rolein the governance structure selection decision.

With the estimated parameters from Table 4,the cumulative standard normal distribution, �,and standard normal density, ϕ, are evaluated atZj γ for all observations. Using least squares withstandard errors robust to heteroskedasticity andautocorrelation, we then estimate the operationalversions of Equation (16) given by:

PERFij = β20 + β21PERFLAGj

+ β22INDPERFj + β23LNSALESj

+ β24DATj + β25REV GWT Hj

+ β26DV PORj + β27OWNj + β28BDSIZEj

+ β29INDBDj + (β10 − β20)�j

+ (β11 − β21)�jPERFLAGj + (β12

− β22)�j INDPERFj + (β13 − β23)

Table 4. Maximum likelihood estimates of selectionequation

Variable Est.coef.

Std.err.

Z P-value

LNSALES 0.248 0.065 3.80 0.000∗∗∗

PPEAT 0.640 0.389 1.64 0.100∗

EMP −0.002 0.001 −2.88 0.004∗∗∗

SEGNUM −0.005 0.046 −0.11 0.908REVGWTH −0.003 0.001 −2.33 0.020∗∗

OWN 0.090 0.025 3.61 0.000∗∗∗

BDSIZE −0.084 0.040 −2.09 0.036∗∗

INDBD 0.013 0.005 2.85 0.004∗∗∗

EMPDIRS −0.225 0.081 −2.77 0.006∗∗∗

MINONBRD 0.293 0.113 2.60 0.009∗∗∗

FEMONBRD 0.045 0.096 0.47 0.636INDSLG5 0.008 0.006 1.35 0.176INDSLV5 0.216 0.854 0.25 0.800constant −0.725 0.754 −0.96 0.336

Number of obs 1880Wald χ 2(25) 104.83P-value 0.000∗∗∗

Pseudo r2 0.174

∗ , ∗∗ , ∗∗∗ indicate significance at the 0.10, 0.05, and 0.01 levelsfor a two-tailed test. Estimated coefficients for year and industrydummy variables not reported. Standard errors are robust toheteroskedasticity and autocorrelation.

�jLNSALESj + (β14 − β24)�jDATj

+ (β15 − β25)�jREV GWT Hj

+ (β16 − β26)�jDV PORj

+ (β17 − β27)�jOWNj + (β18 − β28)

�jBDSIZEj + (β19 − β29)�j INDBDj

+ (σ2u − σ1u)ϕj + εij (19)

for the four performance measures.10 Recall thatthe purpose of this estimation is to determinespecifically whether there is evidence that firmschoose a dual or non-dual structure according tocomparative advantage, and whether there are dif-ferences in the marginal performance impacts ofthe independent variables between the two gover-nance structures. Results are reported in Table 5.

10 Year and industry dummies along with the associated inter-action terms are included in the estimation but are not shownin the text of Equation (19) in the interest of space. The readerwill also note that all independent variables have been mean-centered in accordance with the work of Irwin and McClelland(2001) on moderated multivariate regression models. This is truefor all estimations in this study that include interaction terms.We thank an anonymous referee for this suggestion.



Table 5. Parameter estimates—Equation (19)—testing for two regimes and comparative advantage

Dependent variable

MKTRET TOBINSQ ROA EPS

Variable Est. coef. P-value Est. coef. P-value Est. coef. P-value Est. coef. P-value

MKTRETLAG −0.046 0.670INDMKTRET 0.438 0.054∗

TOBINSQLAG −0.298 0.058∗

INDTOBINSQ 1.973 0.000∗∗∗

ROALAG 0.464 0.059∗

INDROA 1.005 0.021∗∗

EPSLAG 1.171 0.001∗∗∗

INDEPS −2.386 0.004∗∗∗

LNSALES 7.991 0.041∗∗ 0.094 0.590 1.812 0.171 0.024 0.008∗∗∗

DAT −0.099 0.781 0.029 0.123 −0.068 0.603 0.000 0.697REVGWTH 0.309 0.359 0.008 0.202 0.044 0.469 0.000 0.453DVPOR −0.039 0.035∗∗ 0.000 0.612 0.004 0.179 0.000 0.485OWN 0.166 0.932 0.091 0.135 −0.714 0.080∗ 0.001 0.817BDSIZE −4.079 0.117 −0.259 0.009∗∗∗ −1.381 0.056∗ −0.015 0.001∗∗∗

INDBD −0.145 0.748 0.007 0.678 −0.117 0.381 −0.001 0.199� MKTRETLAG −0.134 0.332� INDMKTRET 0.216 0.490� TOBINSQLAG 0.785 0.001∗∗∗

� INDTOBINSQ −1.627 0.003∗∗∗

� ROALAG −0.211 0.503� INDROA −0.403 0.432� EPSLAG −1.108 0.009∗∗∗

� INDEPS 3.975 0.000∗∗∗

� LNSALES −9.555 0.049∗∗ −0.083 0.670 −0.968 0.541 −0.025 0.024∗∗

� DAT −0.141 0.758 −0.047 0.033∗∗ −0.028 0.854 −0.001 0.258� REVGWTH 0.049 0.909 −0.002 0.819 0.042 0.565 0.000 0.824� DVPOR 0.039 0.040∗∗ 0.000 0.633 −0.004 0.172 0.000 0.399� OWN −0.020 0.992 −0.091 0.142 0.727 0.075∗ −0.001 0.784� BDSIZE 4.081 0.206 0.277 0.015∗∗ 1.677 0.043∗∗ 0.020 0.001∗∗∗

� INDBD 0.262 0.632 −0.003 0.845 0.124 0.423 0.001 0.197� 43.686 0.458 0.795 0.598 2.689 0.799 −0.431 0.006∗∗∗

ϕ 5.090 0.874 0.172 0.875 1.176 0.877 0.028 0.647Constant −39.651 0.457 −0.516 0.712 −7.892 0.416 0.366 0.007∗∗∗

Number of obs. 1880 1880 1880 1880F (44, 530) 17.64 38.95 22.75 14.03P-value 0.000∗∗∗ 0.000∗∗∗ 0.000∗∗∗ 0.000∗∗∗

r2 0.321 0.551 0.480 0.707

∗ , ∗∗ , ∗∗∗ indicate significance at the 0.10, 0.05, and 0.01 levels for a two-tailed test. Coefficients for year and industry dummyvariables not reported. Standard errors are robust to heteroskedasticity and autocorrelation.

Selection under comparative advantage requiresthe estimate of (σ2u − σ1u) to be positive or σ2u >

σ1u. Examining the results in the row labeled ϕ,all parameter estimates are positive, but all arehighly insignificant with p-values well in excessof the standard significance levels. Hence, we can-not reject the null hypothesis that (σ2u − σ1u) = 0,or equivalently, there is no evidence that firmschoose their governance structure based on com-parative advantage. Hypothesis 1 is rejected. The

result is consistent with the sample means reportedin Table 3, which show that the mean performancelevels over the sample period were lower for firmswith a dual governance structure. Table 5 alsoshows a number of significant interaction termssuggesting that the marginal performance impactsof at least some variables may differ for dualversus non-dual firms.

For each of the performance measures, a sin-gle equation is now estimated by least squares



with robust standard errors. The dummy variableDUAL is included as a shift parameter and is inter-acted with those explanatory variables identifiedin Table 5 as possibly having differential marginalperformance impacts for dual versus non-dualfirms. To account for selection bias, the equationsinclude the hazard variable, Hj , henceforth labeledas HAZARD.

The results, reported in Table 6, indicate thatselection bias is present in the MKTRET andEPS equations as the estimated coefficients onHAZARD are positive and significant at a fivepercent level, that is, Hypothesis 2 is rejected.11

Selection bias does not seem to be problematicwhen performance is measured by TOBINSQ orROA. In addition, the dual governance structureshifts the performance relationship significantlyonly when performance is measured as MKTRET.Ceteris paribus, dual governance increases a firm’smarket return by about 17.9 percent relative tothe non-dual structure, excluding the quantitativeimpact of the selection bias.

Before returning to the selection bias issuebelow, we first offer a brief discussion of someother results in Table 6. Results that are robustacross the equations are that the previous levelof the firm’s performance (PERFLAG), the over-all industry level of performance (INDPERF ), anda firm’s revenue growth (REVGWTH ), are highlysignificant in explaining current performance lev-els. Moreover, there is a positive and significantmoderating influence of the dual structure on theperformance impact of industry performance forthe EPS equation with a fairly large point esti-mate of 0.877 on DUAL INDEPS. The positivecoefficients for LNSALES imply that larger firmsperform better than smaller firms, though the esti-mates are significant only for ROA and EPS. Thenegative coefficients on DUAL LNSALES suggestthat larger firms do poorer under a dual struc-ture, though none are statistically significant. Moreleverage (DAT ) has a significant adverse impact

11 The variables PPEAT, EMP, SEGNUM, EMPDIRS, MINON-BRD, FEMONBRD, INDSLG5, and INDSLV5 are treated asinstrumental variables, included in the selection equation butexcluded from the performance equations. A test of weak instru-ments strongly rejected the null hypothesis that all instrumentalvariables coefficients are jointly zero in the selection equation(p-value = 0.0019). Subsequent, Sargan tests for instrumentvalidity do not reject the null hypotheses that all of the overiden-tifying restrictions (surplus moment conditions) are valid withp-values well above standard levels. Details are available uponrequest.

on a firm’s MKTRET, its ROA, and its EPS.Dividends as a proportion of operating income(DVPOR) have a significant negative impact onperformance when measured by market return,which is contrary to expectations according tothe free cash flow/managerial alignment argument.However, for dual firms this is almost entirelyoffset by a positive and significant moderatinginfluence of the dual governance structure as evi-denced by the point estimate on DUAL DVPOR.In contrast, the free cash flow/managerial argu-ment for a positive performance impact of DVPORseems to be supported for performance measuredas TOBINSQ and EPS, though the point estimatesare quite small. CEO stock ownership (OWN ) issignificant only for EPS with more ownershipleading to poorer performance. Larger boards areestimated to have a negative impact on MKTRET,ROA, and EPS, though the point estimate is sig-nificant only for ROA. Dual structured firms withlarger boards seem to have higher EPS ratios thantheir non-dual counterparts according to the posi-tive and significant coefficient on DUAL BDSIZE.Board independence is insignificant in three outof the four equations, but is positively and signif-icantly associated with a firm’s EPS.

More on selection bias

In extending our discussion of selection bias, itis useful to recall Equations (9) and (10), whichshow the expected values of firm performance con-ditional on governance structure. The difference inexpected performance between dual and non-dualfirms can be derived by subtracting Equation (10)from Equation (9). Since the hypothesis that (σ2u −σ1u) = 0 could not be rejected, we set σ1u = σ2u =σ and write this difference as:

E(Y1j |Ij = 1) − E(Y2j |Ij = 0) = Xj(β1 − β2)

− σ

(ϕ(Zjγ )

�(Zjγ )[1 − �(Zjγ )]

)(20)

where the ratio in parentheses on the right-handside is equal to (H1j + H2j ), henceforth labeledas SELECT. With σ > 0, Equation (20) impliesthat firms that select the dual (non-dual) structureperform poorer (better), on average, than underrandom assignment due to the correlation betweenthe nonobservables of the selection equation andthe nonobservables of the performance equation,



Tabl

e6.

Perf

orm

ance

equa

tion

estim

ates

—te

stin

gfo

rse

lect

ion

bias

Dep

ende

ntva

riab

le

MK

TR

ET

TOB

INSQ

RO

AE

PS

Var

iabl

eE

st.c

oef.

P-v

alue

Var

iabl

eE

st.c

oef.

P-v

alue

Var

iabl

eE

st.c

oef.

P-v

alue

Var

iabl

eE

st.c

oef.

P-v

alue

MK

TR

ET

LA

G−0

.150

0.00

0∗∗∗

TOB

INSQ

LA

G0.

272

0.05

9∗R

OA

LA

G0.

312

0.00

0∗∗∗

EP

SLA

G0.

293

0.00

8∗∗∗

IND

MK

TR

ET

0.60

40.

000∗∗

∗D

UA

LTO

BIN

SQL

AG

−0.0

740.

650

IND

RO

A0.

684

0.00

0∗∗∗

DU

AL

EP

SLA

G−0

.089

0.53

6L

NSA

LE

S1.

173

0.43

3IN

DTO

BIN

SQ1.

125

0.00

8∗∗∗

LN

SAL

ES

1.28

10.

000∗∗

∗IN

DE

PS

0.42

70.

020∗∗

DU

AL

LN

SAL

ES

−1.0

140.

560

DU

AL

IND

TOB

INSQ

−0.4

450.

291

DA

T−0

.087

0.00

1∗∗∗

DU

AL

IND

EP

S0.

877

0.01

3∗∗

DA

T−0

.186

0.00

8∗∗∗

LN

SAL

ES

0.00

30.

948

RE

VG

WT

H0.

066

0.00

1∗∗∗

LN

SAL

ES

0.00

70.

053∗

RE

VG

WT

H0.

352

0.00

0∗∗∗

DA

T−0

.004

0.38

8D

VP

OR

2E-0

50.

605

DU

AL

LN

SAL

ES

−0.0

050.

295

DV

PO

R−0

.210

0.00

0∗∗∗

DU

AL

DA

T−0

.007

0.28

3O

WN

−0.1

950.

172

DA

T6E

-04

0.00

0∗∗∗

DU

AL

DV

PO

R0.

209

0.00

0∗∗∗

RE

VG

WT

H0.

008

0.00

0∗∗∗

DU

AL

OW

N0.

194

0.16

4R

EV

GW

TH

3E-0

40.

006∗∗

∗

OW

N0.

096

0.48

3D

VP

OR

−1.3

E-0

50.

041∗∗

BD

SIZ

E−0

.475

0.04

5∗∗D

VP

OR

3E-0

60.

030∗∗

BD

SIZ

E−0

.887

0.16

7O

WN

−0.0

050.

221

DU

AL

BD

SIZ

E0.

393

0.13

8O

WN

−0.0

010.

084∗

IND

BD

0.02

80.

700

BD

SIZ

E0.

001

0.99

0IN

DB

D−0

.015

0.40

1B

DSI

ZE

−0.0

030.

126

DU

AL

17.8

990.

053∗

DU

AL

BD

SIZ

E−0

.036

0.47

7D

UA

L3.

965

0.15

1D

UA

LB

DSI

ZE

0.00

50.

048∗∗

HA

ZA

RD

10.7

650.

049∗∗

IND

BD

0.00

10.

786

HA

ZA

RD

1.43

40.

363

IND

BD

3E-0

40.

083∗

Con

stan

t−5

.520

0.57

2D

UA

L0.

243

0.46

8co

nsta

nt0.

931

0.76

3D

UA

L−0

.055

0.19

1H

AZ

AR

D0.

149

0.45

4H

AZ

AR

D0.

029

0.04

4∗∗

cons

tant

1.89

80.

000∗∗

∗co

nsta

nt0.

035

0.34

0

Num

ber

ofob

s18

80N

umbe

rof

obs

1880

Num

ber

ofob

s18

80N

umbe

rof

obs

1880

F(2

5,53

0)24

.99

F(2

8,53

0)48

.46

F(3

2,53

0)23

.01

F(3

7,53

0)11

.80

P-V

alue

0.00

0∗∗∗

P-V

alue

0.00

0∗∗∗

P-V

alue

0.00

0∗∗∗

P-V

alue

∗0.

000∗∗

∗

R2

0.32

0R

20.

523

R2

0.46

6R

20.

658

∗ ,∗∗

,∗∗

∗in

dica

tesi

gnifi

canc

eat

the

0.10

,0.

05,

and

0.01

leve

lsfo

ra

two-

taile

dte

st.

Coe

ffici

ents

for

year

and

indu

stry

dum

my

vari

able

sno

tre

port

ed.

Stan

dard

erro

rsar

ero

bust

tohe

tero

sked

astic

ityan

dau

toco

rrel

atio

n.



that is, the correlation of the disturbances. Thiscan be seen more clearly from a simple rewritingof Equations (7) and (8) with σ in place of bothσ1u and σ2u

E(Y1j |Ij = 1) = Xjβ1 + E(ε1j |Ij = 1)

= X1jβ1 − σ

(ϕ(Zjγ )

�(Zjγ )

)

E(Y2j |Ij = 0) = Xjβ2 + E(ε2j |Ij = 0)

= X2jβ2 + σ

(ϕ(Zjγ )

1 − �(Zjγ )

)

A quantitative assessment of Equation (20) atthe sample means is straightforward. Since themean-centered interaction terms, by construction,will equal zero at their sample means, the predicteddifference in the mean performance levels can bewritten simply as:

DIFF = δ0 − σ × SELECT (21)

where δ0 is the estimated coefficient on DUALand σ is the estimated coefficient on HAZARD.The sample mean of SELECT is calculated as thesample mean of:

(ϕ(Zj γ )

�(Zj γ )[1 − �(Zj γ )]

)

where the values of the vector γ are taken fromthe maximum likelihood estimates presented inTable 4.

For the market return equation, δ0 = 17.9, σ =10.8, and the sample mean of SELECT = 1.96.Using Equation (21), this yields a selection bias of−21.1 and a predicted difference in mean returnsof −3.2 percent. The actual difference in the sam-ple mean market returns is about −4.9 percent.A similar calculation for the difference in meanEPS ratios yields a selection bias of −0.057 anda predicted difference in means of −0.110. Theactual sample mean EPS difference is approxi-mately −0.009.

Does accounting for selection bias matter?

An obvious yet important question is whether ornot our approach in accounting for selection biasyields fundamentally different conclusions from

those that would be reached under (1) a conven-tional dummy variable approach and (2) the stan-dard treatment effects approach. To offer someinsight, we report the results from these twoapproaches in Tables 7 and 8, respectively.12

A comparison of the results from Table 6 withthose from Table 7 reveals a dramatic differ-ence when measuring performance by marketreturn. Specifically, the conventional dummy vari-able model shows a negative and statistically sig-nificant impact of CEO duality on market returnwith the point estimate suggesting that dual gover-nance firms, on average, yield a market return thatis almost six percentage points lower than non-dualfirms. Clearly, in this instance, neglecting selectionbias leads to very different conclusions regardingthe relationship between CEO duality and firm per-formance. This does not seem to extend to the casewhere EPS is taken as the measure of performance.Though the sign on DUAL from Table 7 is oppo-site to that from Table 6, neither of the coefficientsis significant at standard levels.

Additionally, we estimated the conventionaldummy variable model including all of the inter-action terms appearing in Table 6. Here, we dis-cuss selected results only.13 Using market returnas the measure of performance, the estimated coef-ficient on DUAL was 0.489 with a p-value of0.818, indicating in this case that the dual gover-nance structure has no impact on firm performanceexcept through a significant moderating effect onthe performance impact of DVPOR similar to thatreported in Table 6. Including the interaction termsin the conventional version of the EPS equationresulted in an estimated coefficient on DUAL equalto −0.099 with a p-value of 0.009, implying anegative independent impact of CEO duality onfirm performance. The estimated coefficients onthe interaction terms were virtually identical inmagnitude and significance to those reported inTable 6, suggesting that CEO duality may havepositive moderating influences on the performanceimpacts of other explanatory variables. Again, itseems clear that a failure to account for selectionbias can have serious impacts on the conclusionsto be drawn from such analyses.

12 Though we report the results for TOBINSQ and ROA, they arenot discussed as they are qualitatively the same as Table 6, andat this juncture, not very interesting.13 The entire set of results is available from the authors uponrequest.



Tabl

e7.

Perf

orm

ance

equa

tion

estim

ates

—co

nven

tiona

ldu

mm

yva

riab

leap

proa

ch

Dep

ende

ntva

riab

le

MK

TR

ET

TOB

INSQ

RO

AE

PS

Var

iabl

eE

st.

coef

.P

-val

ueV

aria

ble

Est

.co

ef.

P-v

alue

Var

iabl

eE

st.

coef

.P

-val

ueV

aria

ble

Est

.co

ef.

P-v

alue

MK

TR

ET

LA

G−0

.149

0.00

0∗∗∗

TOB

INSQ

LA

G0.

216

0.00

5∗∗∗

RO

AL

AG

0.31

50.

000∗∗

∗E

PSL

AG

0.22

20.

015∗∗

IND

MK

TR

ET

0.60

40.

000∗∗

∗IN

DTO

BIN

SQ0.

809

0.00

0∗∗∗

IND

RO

A0.

676

0.00

0∗∗∗

IND

EP

S1.

230

0.00

0∗∗∗

LN

SAL

ES

1.20

60.

201

LN

SAL

ES

0.01

70.

639

LN

SAL

ES

1.40

70.

000∗∗

∗L

NSA

LE

S0.

005

0.04

2∗∗

DA

T−0

.189

0.00

7∗∗∗

DA

T−0

.009

0.01

3∗∗D

AT

−0.0

870.

001∗∗

∗D

AT

−0.0

010.

001∗∗

∗

RE

VG

WT

H0.

340

0.00

0∗∗∗

RE

VG

WT

H0.

008

0.00

0∗∗∗

RE

VG

WT

H0.

063

0.00

1∗∗∗

RE

VG

WT

H3E

-04

0.03

4∗∗

DV

PO

R−0

.002

0.00

2∗∗∗

DV

PO

R6E

-06

0.40

4D

VP

OR

−2E

-05

0.62

2D

VP

OR

3E-0

60.

060∗

OW

N0.

238

0.05

0∗∗O

WN

−0.0

020.

575

OW

N0.

009

0.70

2O

WN

−0.0

010.

196

BD

SIZ

E−1

.395

0.02

1∗∗B

DSI

ZE

−0.0

290.

195

BD

SIZ

E−0

.216

0.16

3B

DSI

ZE

−2E

-04

0.86

8IN

DB

D0.

113

0.07

6∗IN

DB

D0.

002

0.44

7IN

DB

D−0

.004

0.77

4IN

DB

D−1

0E-0

50.

493

DU

AL

−5.8

570.

023∗∗

DU

AL

−0.0

740.

497

DU

AL

−0.2

920.

689

DU

AL

0.00

30.

711

cons

tant

−1.1

800.

898

cons

tant

0.37

10.

274

cons

tant

−5.7

850.

007∗∗

∗co

nsta

nt−0

.077

0.01

2∗∗

Num

ber

ofob

s18

80N

umbe

rof

obs

1880

Num

ber

ofob

s18

80N

umbe

rof

obs

1880

F(2

2,53

0)27

.78

F(2

2,53

0)27

.15

F(22

,530

)35

.31

F(22

,530

)15

.92

P-va

lue

0.00

0∗∗∗

P-va

lue

0.00

0∗∗∗

P-va

lue

0.00

0∗∗∗

P-va

lue

0.00

0∗∗∗

R2

0.31

8R

20.

503

R2

0.46

3R

20.

635

∗ ,∗∗

,∗∗

∗in

dica

tesi

gnifi

canc

eat

the

0.10

,0.

05,

and

0.01

leve

lsfo

ra

two-

taile

dte

st.

Coe

ffici

ents

for

year

and

indu

stry

dum

my

vari

able

sno

tre

port

ed.

Stan

dard

erro

rsar

ero

bust

tohe

tero

sked

astic

ityan

dau

toco

rrel

atio

n.



Tabl

e8.

Perf

orm

ance

equa

tion

estim

ates

—st

anda

rdtr

eatm

ent

effe

cts

appr

oach

Dep

ende

ntva

riab

le

MK

TR

ET

TOB

INSQ

RO

AE

PS

Var

iabl

eE

st.

coef

.P

-val

ueV

aria

ble

Est

.co

ef.

P-v

alue

Var

iabl

eE

st.

coef

.P

-val

ueV

aria

ble

Est

.co

ef.

P-v

alue

MK

TR

ET

LA

G−0

.150

0.00

0∗∗∗

TOB

INSQ

LA

G0.

216

0.00

5∗∗∗

RO

AL

AG

0.31

50.

000∗∗

∗E

PSL

AG

0.22

20.

015∗∗

IND

MK

TR

ET

0.60

60.

000∗∗

∗IN

DTO

BIN

SQ0.

809

0.00

0∗∗∗

IND

RO

A0.

675

0.00

0∗∗∗

IND

EP

S1.

228

0.00

0∗∗∗

LN

SAL

ES

0.46

10.

630

LN

SAL

ES

0.01

50.

681

LN

SAL

ES

1.33

80.

000∗∗

∗L

NSA

LE

S0.

003

0.20

7D

AT

−0.1

910.

006∗∗

∗D

AT

−0.0

090.

013∗∗

DA

T−0

.087

0.00

1∗∗∗

DA

T−0

.001

0.00

0∗∗∗

RE

VG

WT

H0.

353

0.00

0∗∗∗

RE

VG

WT

H0.

008

0.00

0∗∗∗

RE

VG

WT

H0.

064

0.00

1∗∗∗

RE

VG

WT

H3E

-04

0.01

7∗∗

DV

PO

R−0

.002

0.00

1∗∗∗

DV

PO

R−6

E-0

60.

402

DV

PO

R2E

-05

0.61

2D

VP

OR

3E-0

60.

058∗

OW

N0.

103

0.45

5O

WN

−0.0

030.

542

OW

N−0

.004

0.87

2O

WN

−0.0

010.

088∗

BD

SIZ

E−1

.020

0.11

0B

DSI

ZE

−0.0

280.

224

BD

SIZ

E−0

.180

0.23

4B

DSI

ZE

0.00

10.

656

IND

BD

0.03

40.

641

IND

BD

0.00

20.

536

IND

BD

−0.0

120.

516

IND

BD

−3E

-04

0.12

8D

UA

L10

.283

0.26

3D

UA

L−0

.027

0.92

8D

UA

L1.

224

0.62

0D

UA

L0.

040

0.08

4∗

HA

ZA

RD

10.0

660.

061∗

HA

ZA

RD

0.02

90.

861

HA

ZA

RD

0.94

60.

551

HA

ZA

RD

0.02

30.

089∗

cons

tant

−7.9

790.

452

cons

tant

0.35

10.

361

cons

tant

−6.4

110.

011∗∗

cons

tant

−0.0

920.

005∗∗

∗

Num

ber

ofob

s18

80N

umbe

rof

obs

1880

Num

ber

ofob

s18

80N

umbe

rof

obs

1880

F(2

3,53

0)26

.7F

(23,

530)

26.3

8F

(23,

530)

32.5

3F

(23,

530)

15.2

1P-

valu

e0.

000∗∗

∗P-

valu

e0.

000∗∗

∗P-

valu

e0.

000∗∗

∗P-

valu

e0.

000∗∗

∗

R2

0.31

9R

20.

503

R2

0.46

3R

20.

636

∗,

∗∗,

∗∗∗

indi

cate

sign

ifica

nce

atth

e0.

10,

0.05

,an

d0.

01le

vels

for

atw

o-ta

iled

test

.C

oeffi

cien

tsfo

rye

aran

din

dust

rydu

mm

yva

riab

les

not

repo

rted

.St

anda

rder

rors

are

robu

stto

hete

rosk

edas

ticity

and

auto

corr

elat

ion.



Turning now to Table 8, which presents theresults from the estimation of a standard treatmenteffects model, we see that for both the MKTRETand the EPS equations, selection bias is supportedwith estimated coefficients on HAZARD statisti-cally significant at less than 10 percent. Note alsothat the point estimates are almost identical to thevalues shown in Table 6. A difference, however,lies in the estimated coefficient on DUAL. In themarket return equation, it is statistically insignifi-cant, implying no independent impact of the dualstructure on firm performance. Conversely, in theearnings per share equation the coefficient is posi-tive and significant, indicating a beneficial inde-pendent impact of CEO duality on firm perfor-mance. These results suggest that at least withthis sample dataset, the standard treatment effectsmodel is equally capable of accounting for selec-tion bias. However, a formal testing of the endoge-nous switching regression model allows for theidentification of potentially significant interactionterms which, if not included, can bias the estimateof the independent performance effect of CEOduality. A lesson to be drawn from this is thataccounting for possible moderating influences ofCEO duality on the performance impact of otherrelevant variables is important for identifying theindependent performance impact of the dual gov-ernance structure.

Robustness tests

The durable goods industry comprises almost one-third of the entire sample of firm-year observa-tions. The model was estimated for this indus-try only. Qualitatively, the results for MKTRET,TOBINSQ, and ROA remained unchanged fromwhat is reported in Table 6. For EPS, however,selectivity bias was no longer significant. Wealso estimated the model for the computer andretail industries individually. No selectivity biaswas found for any of the performance measures,nor did we find any significant marginal per-formance impacts of CEO duality. The modelwas estimated with no interaction terms and witha complete set of possible interaction terms inthe four performance equations. The results forMKTRET and EPS again were consistent withthose reported in Table 6, both qualitatively andquantitatively. The same was true for TOBINSQ.For the ROA equation including the complete set

of interaction terms, we found a marginal selec-tivity bias (p-value = 0.105) with a positive andsignificant main effect of duality on ROA. Finally,each of the performance equations was estimatedincluding all of the variables appearing in theselection equation. In other words, there was noset of instrumental variables so that identifica-tion of the performance equations relies solelyon the nonlinearities of the hazard variable. Forall performance equations, no significant selec-tivity bias was detected, nor was any significantindependent performance impact of CEO duality.Results for all robustness tests are available uponrequest.

SUMMARY AND CONCLUDINGREMARKS

This study has focused explicitly on the method-ological implications of the endogenous theory ofgovernance as applied to firm performance. In par-ticular, if firms choose their governance structuresas part of a constrained performance maximizationprocess, then application of an appropriate empiri-cal methodology should reveal some statistical evi-dence of such behavior. Heretofore, no study hasemployed such a methodology. In this study, wehave taken advantage of the endogenous switchingregression model framework to do just that. Themodel allowed us to test explicitly for selectionbehavior in accordance with comparative advan-tage and, concomitantly, the presence of selectivitybias in estimating the impact of CEO duality onfirm performance. The selection and performanceequations were modeled in accordance with theextant accounting, economics, and management lit-erature on the impact of the dual governance struc-ture on firm performance. We tested four perfor-mance measures for the entire sample of firm-yearobservations as well as for the largest three indus-tries in terms of sample sizes. Comparisons weremade with a conventional dummy variable modelwith no accounting for selection bias, as well aswith the standard treatment effects model.

Overall, our results suggest that a firm’s selec-tion of the dual governance structure is not con-sistent with either comparative advantage or theobjective of maximizing firm performance. Indeed,with respect to performance as measured by mar-ket returns or by earnings per share, the selec-tion of the dual structure, on average, is clearly



suboptimal. In the terminology of Kang and Zard-koohi (2005), CEO duality seems to be an orga-nizational practice that is adopted under (nonob-servable) conditions that are inappropriate withregard to firm performance. Comparison with astandard treatment effects model corroborated thepresence of a significant selectivity bias in themarket return and earnings per share performanceequations. For particular firms, however, we haveprovided some evidence that the adverse perfor-mance consequences of such selection bias maybe mitigated by the moderating influences of CEOduality on the impacts of other performance-relatedvariables. We also find a significant independentpositive performance impact of the dual leadershipstructure for only market return, consistent witharguments that a dual structure provides a singlefocal point, firm stability, and better communica-tion between management and the board.

The findings of no selectivity bias when per-formance is measured by Tobin’s q and ROA areinconsistent with the theory that governance isendogenous with respect to performance. Specif-ically, the results of this study suggest that firmsdo not sort themselves between the two gover-nance structures strategically with the objective ofimproving performance, but rather do so in a waythat neither improves nor worsens performance rel-ative to what it would be under simple randomassignment.

Of course, no study is without its limitationsand ours is no exception. The selection of CEOduality is the sole governance choice that we con-sider. That, of course, is a simplification and itmight be that a model endogenizing a number ofother governance choices would reveal some sortof constrained performance maximizing behavior.Additionally, we measure performance using fouroften used metrics that certainly do not constitutean exhaustive list of possible performance mea-sures. Using other measures could possibly lead todifferent conclusions. It also could be that perfor-mance is not an objective itself, but acts as a con-straint on the optimization of some other objectiveor criterion, possibly only known to the manage-ment of the firm. Finally, unlike Chen et al. (2008),we exclude firms that changed leadership structureone or more times over the sample period. Closerexamination of these firms could certainly yielddifferent results.

Even with these limitations, however, we believethe study makes an important contribution that

should not be overlooked. In particular, if one sub-scribes to an endogenous theory of governance,then the empirical methodology of any associatedstudy of the impacts of corporate governance onfirm ‘objectives’ must address explicitly the issueof self-selection. Otherwise, within such a theo-retical context, the empirical results would haveto be considered unreliable because of their sus-ceptibility to selectivity bias. In this regard, ironi-cally, our study provides some retroactive supportto previous research using conventional economet-ric techniques to examine the relationship betweenCEO duality and firm performance measured byTobin’s q or ROA. Finally, even with the selec-tivity bias framework, researchers would be welladvised not to ignore potential moderating influ-ences that governance structures may have on theimpact of other performance-related factors, asthis may yield biased estimates of the governancestructure’s independent performance effect.

ACKNOWLEDGEMENTS

The authors thank two anonymous reviewers andEditor Will Mitchell for their very constructivecomments and suggestions throughout the reviewprocess. We would also like to thank Kevin Forbes,Samuel Kotz, Ibrahim Salama, and seminar par-ticipants at North Carolina Central University fortheir helpful comments on earlier drafts. Researchassistance of Julius Bradshaw is sincerely appre-ciated. Raghavan Iyengar wishes to acknowledgethe financial assistance provided by North CarolinaCentral University’s summer research grant andfaculty seed grant.

REFERENCES

Anderson CA, Anthony RN. 1986. The New CorporateDirectors. Wiley: New York.

Baliga BR, Moyer RC, Rao RS. 1996. CEO dualityand firm performance: what’s the fuss? StrategicManagement Journal 17(1): 41–53.

Berg SV, Smith SK. 1978. CEO and board chairman: aquantitative study of dual vs. unitary board leadership.Directors and Boards 3: 34–39.

Bhagat S, Black B. 2001. The non-correlation betweenboard independence and long-term firm performance.Journal of Corporation Law 27(2): 231–274.

Boyd BK. 1995. CEO duality and firm performance: acontingency approach. Strategic Management Journal16(4): 301–312.



Brickley JA, Coles JL, Jarrell G. 1997. Leadershipstructure: separating the CEO and chairman of theboard. Journal of Corporate Finance 3: 189–220.

Chaganti RS, Mahajan V, Sharma S. 1985. Corporateboard size, composition and corporate failures in theretailing industry. Journal of Management Studies 22:400–417.

Chen C, Lin JB, Yi B. 2008. CEO duality andfirm performance: an endogenous issue. CorporateOwnership and Control 6(1): 58–65.

Dahya J, Lonie AA, Power DM. 1996. The case forseparating the roles of chairman and CEO: an analysisof stock market and accounting data. CorporateGovernance: An International Review 4(2): 71–77.

Daily CM, Dalton DR. 1994. Corporate governance andbankrupt firm: an empirical assessment. StrategicManagement Journal 15(8): 643–654.

Donaldson L, Davis JH. 1991. Stewardship theory oragency theory: CEO governance and shareholderreturns. Australian Journal of Management 16(1):49–65.

Faleye O. 2007. Does one hat fit all? The case ofcorporate leadership structure. Journal of Managementand Governance 11(3): 239–259.

Fama EF, Jensen MC. 1983. Separation of ownershipand control. Journal of Law and Economics 26:301–325.

Finkelstein S, D’Aveni RA. 1994. CEO duality as adouble-edged sword: how boards of directors balanceentrenchment avoidance and unity of command.Academy of Management Journal 37(5): 1079–1108.

Frankel RM, Johnson MF, Nelson KK. 2002. Therelation between auditors’ fees for nonaudit servicesand earnings management. Accounting Review 77(S-1): 71–114.

Greene WH. 2003. Econometric Analysis (Fifth Edition).Prentice-Hall: Upper Saddle River, NJ.

Harrison JR, Torres DL, Kukalis S. 1988. The changingof the guard: turnover and structural change in thetop management positions. Administrative ScienceQuarterly 33: 211–232.

Hermalin BE, Weisbach MS. 1998. Endogenously chosenboards of directors and their monitoring of the CEO.American Economic Review 88: 96–118.

Hermalin BE, Weisbach MS. 2003. Boards of direc-tors as an endogenously determined institution: asurvey of the economic literature. Economic PolicyReview 9: 7–26.

Irwin JR, McClelland GH. 2001. Misleading heuristicsand moderated multiple regression models. Journal ofMarketing Research 38: 100–109.

Jensen MC. 1993. The modern industrial revolution, exit,and the failure of internal control systems. Journal ofFinance 48(3): 831–880.

Kang E, Zardkoohi A. 2005. Board leadership structureand firm performance. Corporate Governance 13(6):785–789.

Maddala GS. 1983. Limited Dependent and QualitativeVariables in Econometrics . Cambridge UniversityPress: Cambridge, UK.

Peng MW. 2004. Outside directors and firm performanceduring institutional transitions. Strategic ManagementJournal 25(5): 453–471.

Pi L, Timme SG. 1993. Corporate control and bankefficiency. Journal of Banking and Finance 17(2–3):515–530.

Rechner PL, Dalton DR. 1991. CEO duality andorganizational performance: a longitudinal analysis.Strategic Management Journal 12(2): 155–160.

Sridharan UV, Marsinko A. 1997. CEO duality in thepaper and forest products industry. Journal ofFinancial and Strategic Decisions 10(1): 59–65.

Stoeberl PA, Sherony BC. 1985. Board efficiency andeffectiveness. In Handbook for Corporate Directors ,Mattar E, Balls M (eds). McGraw-Hill: New York.

Vancil RF. 1987. Passing the Baton: Managing theProcess of CEO Succession . Harvard Business SchoolPress: Boston, MA.

Westphal JD, Zajac EJ. 1997. Defections from theinner circle: social exchange, reciprocity, and thediffusion of board independence in U.S. corporations.Administrative Science Quarterly 42: 161–183.

Whisenant S, Sankaraguruswamy S, Raghunandan K.2003. Evidence on the joint determination of audit andnon-audit fees. Journal of Accounting Research 41(4):721–744.

Wooldridge JM. 2002. Econometric Analysis of CrossSection and Panel Data . MIT Press: Cambridge, MA.


Self-selection, endogeneity, and the relationship between CEO duality and firm performance

Documents

Transcript of Self-selection, endogeneity, and the relationship between CEO duality and firm performance