Balance Ms

Balancing Search and Stability:Interdependencies Among Elements of

Organizational Design

Jan W. Rivkin Nicolaj Siggelkow239 Morgan Hall, Harvard Business School, Boston, Massachusetts 02163

The Wharton School, 2017 Steinberg HallDietrich Hall, Philadelphia, Pennsylvania [email protected] [email protected]

Weexamine how and why elements of organizational design depend on one another. Anagent-based simulation allows us to model three design elements and two contextualvariables that have rarely been analyzed jointly: a vertical hierarchy that reviews propos-als from subordinates, an incentive system that rewards subordinates for departmental orrm-wide performance, the decomposition of an organizations many decisions into depart-ments, the underlying pattern of interactions among decisions, and limits on the ability ofmanagers to process information. Interdependencies arise among these features because ofa basic, general tension. To be successful, an organization must broadly search for good setsof decisions, but it must also stabilize around good decisions once discovered. An effectiveorganization balances search and stability. We identify sets of design elements that encour-age broad search and others that promote stability. The adoption of elements that encouragebroad search typically raises the marginal benet of other elements that provide offsettingstability. Hence, the need to balance search and stability generates interdependencies amongthe design elements. We pay special attention to interdependencies that involve the verticalhierarchy. Our ndings conrm many aspects of conventional wisdom about vertical hier-archies, but challenge or put boundary conditions on others. We place limits, for instance,on the received wisdom that rm-wide incentives and capable subordinates make top-leveloversight less valuable. We also identify circumstances in which vertical hierarchies can leadto inferior long-term performance.(Organizational Search; Stability; Interactions; Organizational Design; Organizational Structure;Vertical Hierachy; Agent-Based Simulation)

1. IntroductionIn the diverse literature on organizational design, atleast one proposition has gained widespread accep-tance: the many formal and informal structures, sys-tems, and processes that make up an organizationsdesign affect one another (e.g., Khandwalla 1973,Mintzberg 1979). Organizations are typically seen ascomplex entities . . . composed of tightly interdepen-dent and mutually supportive elements (Miller andFriesen 1984, p. 1) and as highly integrated system[s]

whose performance is determined by the degree ofalignment among the major elements (Nadler andTushman 1997, p. 23). The marginal costs and ben-ets associated with any design element depend onthe conguration of others. For instance, the efcacyof decentralized decision making may hinge on theincentives, information, and training given to middlemanagement. Management debacles are often inter-preted as failures to appreciate the systemic nature oforganizational design.

Management Science 2003 INFORMSVol. 49, No. 3, March 2003 pp. 290311

0025-1909/03/4903/0290$5.001526-5501 electronic ISSN

RIVKIN AND SIGGELKOWBalancing Search and Stability

Prior studies pinpoint specic interdependenciesamong elements of design, but they do not explainwhy interdependencies arise in general. In this paper, weuse an agent-based model of organizational decisionmaking to identify a general, underlying tension thatgives rise to interdependencies. Our results show thatsuccessful rms balance two opposing needs. On onehand, to be effective, a rm must search broadly forgood combinations of decisions; it must not lock inprematurely on the rst decent set of choices it dis-covers. On the other hand, a successful rm musthalt its search efforts and stabilize its decisions once itnds an outstanding set of choices. We identify spe-cic elements of organizational design that drive arm toward broad search and others that encouragestability. The need to balance search and stability cre-ates interdependencies among the elements. Often, arm that adopts an element that pushes it towardbroad search benets from a second element that pullsit toward stability. Prior formal models of organiza-tional search tend to overlook these interdependen-cies because they often grant stability for free; thatis, they assume that rms that discover good deci-sions through search can lock in on those decisionsforever. Contrary to this assumption, we illustrate thatorganizational elements that enable discovery mayundermine lock-in.Our model allows us to examine the relationships

among three prominent elements of formal organi-zational design: a vertical hierarchy embodied in aCEO, subordinate managers, and a ow of informa-tion among them; an incentive system that inuenceswhether managers act for the good of the overall rmor pursue the parochial interests of their departments;and the decomposition of a rms many decisions intodiscrete departments. We also pay careful attentionto two contextual variables: The underlying pattern ofinteraction among a rms decisions, and the limitson the ability of managers to process information andconsider alternatives. These ve features surely donot form an exhaustive list of the design elementsand contextual variables that organizational designershave explored. However, as we discuss in 2, they docover the important classes of considerations in theliterature on formal organizational design. Moreover,

this limited list of considerations is more than suf-cient to illustrate the need for balance between searchand stability.We model the ve features using an agent-based

simulation derived from research on complex adap-tive systems (see 3). This approach enables us tolook simultaneously at all ve features, distinguish-ing our work from prior models that have exam-ined only one or two at a time. Our effort joinsa growing set of agent-based simulations of humanorganizations (e.g., Carley and Lin 1997, Carley andSvoboda 1996, Levinthal 1997, Anderson et al. 1999,Axelrod and Cohen 1999, Chang and Harrington2000). In our model, rms with different organiza-tional designs face a long series of multidimensionaldecision problems. Decisions within each probleminteract with one another in a manner controlled bythe modeler. For each decision problem, each rmattempts to nd a good solution; that is, an effec-tive set of choices. The management team of eachrm consists of a simple hierarchy: a CEO and twosubordinate managers. Each subordinate manager haspurview over a subset of the organizations decisions,a department. Starting from a particular congura-tion of choices, each subordinate considers altering thedecisions under his command, evaluates the alterna-tives in light of an incentive system, and makes rec-ommendations to the CEO. The CEO reviews the pro-posals and accepts the pair of proposalsone fromeach subordinatethat will best serve the rm. Alter-natively, she may overrule her subordinates and main-tain the status quo for either or both departments.Modeled rms differ in their designs: how active

a role the CEO takes and how much informationshe receives from subordinates; whether subordinatesare rewarded for departmental success or the perfor-mance of the rm as a whole; and how rms decom-pose decisions into departments. Organizations alsodiffer in the cognitive abilities of their CEOs and man-agers. By comparing the performance levels of rmswith different designs across a large number of deci-sion problems, we can isolate how the distinct designelements depend on one another.To structure our analysis, we focus on the inter-

dependencies between the vertical hierarchy and theother organizational elements. The results of our

Management Science/Vol. 49, No. 3, March 2003 291


modeling effort (see 4) show that a hierarchy thatactively reviews subordinates proposals is often help-ful, but circumstances exist under which thoroughdecentralization produces superior performance. Thendings also pinpoint design elements and contex-tual variables that amplify or dampen the value ofan active hierarchy. The results conrm many aspectsof the conventional wisdom about hierarchies pro-duced by prior observers of organizations, but chal-lenge or show boundary conditions for other aspects.For instance, consistent with received wisdom, wend that an active vertical hierarchy tends to bemore valuable when interactions among decisions arepervasive. However, this benet arises only if theinformation ow in the hierarchy is rich enough.Without rich information ow, active hierarchies canlead rms to lock in on suboptimal solutions pre-maturely, leading to worse performance than a rmwith a purely passive hierarchy would achieve. Sim-ilarly, we bound conventional wisdom by showingthat rm-wide incentives and capable managers arecomplements to, not substitutes for, an active hier-archy when interactions among decisions are suf-ciently pervasive.To interpret our results (see 5), we use a land-

scape conceptualization that has become popularin certain formal models of organizational search(Kauffman 1995, Levinthal 1997, Levinthal andWarglien 1999, McKelvey 1999, Gavetti and Levinthal2000, Ghemawat and Levinthal 2000, Rivkin 2000,2001, Kauffman et al. 2000, Fleming and Sorenson2001). A mapping from rm decisions to payoffs cre-ates a landscape in the space of decisions. Firmscan be conceived of as trying to attain and sustaina high spot on such a landscapea combination ofdecisions that, together, yield a high payoff. Organi-zational design, we argue, affects rm performanceby altering rms search behavior on the landscapesthey face. A rm typically gravitates on its landscapetoward a sticking pointa conguration of choicesfrom which it will not change. Organizational designaffects long-term performance by two primary chan-nels. First, it alters the nature of a rms stickingpointsthe number of such points and the payoffsassociated with them. Second, it inuences the like-lihood that a rm will actually reach such a stable

conguration of choices. Organizations with the mosteffective designs, ones that balance search and stabil-ity, nd good points and stick to those points.

2. Organizational Design Elementsand Their Interdependencies

A rich heritage of qualitative studies identies a hostof design elements and contextual features that a for-mal model of organizational design might encompass.In this section, we rst explain why we focus ourmodeling effort on ve specic considerations. Wethen turn to conventional wisdom, drawn from thequalitative studies, about how these ve interact. Thiswisdom, focused on interdependencies between anactive vertical hierarchy and each of the other con-siderations, provides a backdrop for the results in 4.Finally, we discuss prior formal models of organiza-tional design. In contrast to the qualitative studies,only a few prior models have emphasized interdepen-dencies among design elements.

Common Considerations in Qualitative Studies.The qualitative literature on organizational design isextensive and diverse, encompassing grounded the-oretical work, eld studies, and numerous syntheses(e.g., Mintzberg 1979, Gibson et al. 2000). Nonethe-less, the literature is unied in what it perceives as thecentral challenge of organizational design: to dividethe tasks of a rm into manageable, specialized jobs,yet coordinate the tasks so that the rm reaps the ben-ets of harmonious action. Implicit in this challengeare two important assumptions. First, coordination isvaluable because the tasks of a rm interact with oneanother; that is, a decision made concerning one taskaffects the efcacy of performing another task oneway or another. Without such interactions, coordina-tion would be unnecessary. Accordingly, the literatureon organizational design (Thompson 1967, Galbraith1973, Mintzberg 1979 (especially Chapter 7), Nadlerand Tushman 1997) has repeatedly returned to theunderlying pattern of interaction among a rms tasks,the rst consideration we include in our model. Thesecond assumption is that the rms situation createsa demand for information processing that exceedsthe deliberative capacity of any individual manager

292 Management Science/Vol. 49, No. 3, March 2003


(Simon 1957). If this were not so, a single ber-manager could evaluate all of a rms alternativesand dictate the best, coordinated course of action.Hence, our model incorporates, as a second criticalfeature, limits on managerial ability.How can an organization in which the complex-

ity of decision problems exceeds the cognitive capac-ity of any single decision maker achieve coordinatedaction? Below, we focus on three design elementsthat have played a particularly prominent role inthe rich discussions found in the qualitative litera-ture: vertical hierarchies, incentive systems, and deci-sion decomposition.1 We emphasize these particularelements both because they are ubiquitous in realorganizations and because they exemplify the threemajor classes of organizational elements identiedby Nadler and Tushman (1997, p. 67): structurallinks (formal relationships among decision makersseparated by structural boundaries), systems and pro-cesses (formal guides to decision making), and group-ing (the aggregation of tasks into work units).A vertical hierarchy is perhaps the most common

mechanism employed to coordinate the decisions ofseparate decision makers. A CEO, for instance, maysit above a set of department heads, review the pro-posals of the departments, and try to integrate themin a way that achieves coordination. Other formalconnections across groups exist, e.g., lateral linkagessuch as liaison positions (Galbraith 1973), but for thesake of parsimony, and because it exists in virtuallyevery organization, we focus our modeling effort onthe vertical hierarchy.Coordination can also be enhanced by systems

and processes that span group borders. Designedwell, such systems and processes can make managers

1 We focus on formal design elements rather than informal elementssuch as casual communication systems (Mintzberg 1979) or corpo-rate culture (Camerer and Vepsalainen 1988). We do so not becauseinformal elements are unimportant or lack interesting interdepen-dencies, but rather for other reasons. The formal elements aloneare more than sufcient to fuel a complex analysis as the rest ofthe paper shows. Moreover, the formal elements can be modeledmore precisely than the informal. Precise modeling of each indi-vidual element is especially important if one wants, as we do, toexamine the interdependencies among them. That said, we considerthe analysis of informal elements of organizational design to be animportant topic for future research.

aware of and responsive to what happens beyondtheir own domains. Nadler and Tushman (1997) iden-tify a rich variety of coordinating systems and pro-cesses: strategic planning efforts, resource allocationprograms, information management systems, and soforth. Among these, we choose to model the incen-tive system, the system that arguably has received thegreatest attention. In particular, we explore a systemin which managers may be rewarded on the basis ofoverall rm performance rather than on the perfor-mance of their individual departments.Besides vertical hierarchies and incentive systems,

treatises on organizational design consider the waytasks and decisions are grouped together to be a fun-damental means to coordinate work in the organiza-tion (Mintzberg 1979, p. 106). Theoretical analysesemphasize how interactions among tasks inuencethe way in which decisions should be grouped.Thompson (1967), for instance, argues that decisionsshould be grouped so that the most intensive inter-actions are internalized, while Simon (1973) stressesthat interactions across decision makers should beminimized. Consequently, groups should be formedso that, as nearly as possible, the rm is decom-posed (Simon 1962) into independent entities. Giventhe prominence of grouping in the qualitative liter-ature on organizational design, our model includes,as its fth feature, the notion of decomposition. Thatis, modeled rms are able to try to achieve coordina-tion by assigning decision rights in a way that placesrelated decisions under a single manager.

Conventional Wisdom About Interdependencies.The qualitative literature not only emphasizes thatdesign elements in general have profound effects oneach other, but it also identies some patterns in theseinterdependencies. Many of these patterns addressa critical question about vertical hierarchies: Underwhat circumstances and in which combination withother design elements should a rm employ seniormanagers who actively review subordinates propos-als and retain the right to veto changes? Conversely,when should a rm delegate the right to make deci-sions to subordinate managers? We summarize thereceived wisdom on these questions as follows: Active vs. passive vertical hierarchy and degree of

interaction. In general, qualitative studies have argued



that the value of active vertical oversight grows asthe degree of interaction among decisions increases.Khandwalla (1973, p. 521), for example, notes thatthe greater the interface between functionally orga-nized departments, the greater is the need for coor-dinative mechanisms such as a common boss.However, the value of active vertical oversight is cru-cially limited by the specter that senior managers maybecome overloaded by the information owing up tothem, causing decision making to grind to a halt. AsChild (1984, p. 148) points out: If top executives areoverloaded then the effective control they can exercisewill be diminished and they will tend to sit on deci-sions which may require speedy attention. Decen-tralized decision making allows a rm to respondmore quickly to (changed) local conditions (Mintzberg1979) and is therefore more appropriate in volatileenvironments (Mintzberg 1981). Active vertical hierarchy and managerial ability. The

presence of an active hierarchy and the ability of sub-ordinate managers are usually seen as substitutes.Child (1984, p. 71) notes: The greater the compe-tence of subordinates, the less closely they need to besupervised and the less often does their work requirereview. Therefore as the competence of subordinatesrises so it becomes feasible to widen spans of con-trol and to reduce levels of management. Active vertical hierarchy and rm-wide incentives.

An incentive system that rewards overall rm goals,rather than departmental goals, has also been consid-ered a partial substitute for an active, coordinatinghierarchy. For instance, Galbraith (1973, p. 14) pointsout: Goal setting helps coordinate interdependentsubtasks and still allows discretion at the local sub-task level, while Child (1984, p. 149) observes thatattention to developing a strong identication withtop management objectives permits delegated deci-sions to be made. Active vertical hierarchy and decision decomposi-

tion. If decisions can be decomposedparsed outto departmentsin such a way that few cross-departmental interactions remain, the value of anactive vertical hierarchy declines. [G]roups that areonly minimally interdependent have relatively littleneed for coordination and therefore do not requireactive oversight by a hierarchy (Nadler and Tushman

1997, p. 92). Moreover, decisions that interact shouldbe grouped together as much as possible and assignedto a single decision maker, regardless of the absenceor presence of a vertical hierarchy (Thompson 1967,Simon 1973).Below, we revisit each of these pieces of received

wisdom in light of our simulation ndings.

Prior Formal Modeling Efforts. Formal model-ing of organizational design has burgeoned in thepast decade as economists have sought to pryopen the black box of the rm, yet few stud-ies have considered multiple organizational elements.As a result, interdependencies among elements havereceived little modeling attention. Insofar as interde-pendencies have been noted (often as a by-productrather than the focus of analysis), results involv-ing vertical hierarchies conrm the conventionalwisdom outlined above. For instance, prior mod-els have noted that rm-wide incentives and anactive hierarchy may serve as substitutes. Firmswith compensation schemes that reward company-wide performance are more likely to have highlydecentralized organization structures (Harris andRaviv 2002, p. 864). Further, in line with conventionalwisdom, Aghion and Tirole (1997) observe that thevalue of an active hierarchy increases when decisionsare not fully decomposed. They show that a CEO ismore likely to intervene when a division managerspreferred decisions are likely to be suboptimal for therm. This situation tends to arise when there aresubstantial externalities on other divisions, on futuremanagers of the division, or on the rm as whole(p. 14).

3. A Model of OrganizationalDesign and Search

Our goal is to test and extend conventional wis-dom about interdependencies among organizationaldesign elements by probing the roots of such connec-tions. To do so, we develop a simulation model inwhich the modeler dictates the underlying pattern ofinteraction among decisions; a computer generates aset of particular decision problems that follow that



pattern; and large numbers of modeled rms with dif-ferent organizational characteristics tackle these deci-sion problems. In this section, we describe each ofthese steps in detail.Though a simulation model does not yield exact,

closed-form solutions as an algebraic approach might,the number and nature of the features we feel areimportant in a model of interdependencies (vary-ing degrees of interactions among decisions, limitedability of decisions makers, a vertical hierarchy withinformation ows, changeable incentive systems, anddifferent types of decision decompositions) make analgebraic approach infeasible. Take, for instance, inter-actions among decisions. In our model, as in real-ity, pairs of decisions may interact as complementsor as substitutes (Porter and Siggelkow 2002). Asa result, the closed-form analysis of supermodularfunctions, which has been used to study comple-mentary interactions (e.g., Milgrom and Roberts 1990,1995, Holmstrm and Milgrom 1994), cannot be read-ily applied. Moreover, given the limited cognitiveabilities of our agents, the long-term outcomes of themodel can be fully understood only by taking thesearch process, i.e., the dynamics of the model, intoaccount. Characterizing such paths is straightforwardwith simulations and extremely difcult with ana-lytical models, which tend to focus on equilibriumoutcomes and tend to ignore how, indeed whether,such equilibria can be reached.

3.1. Setting the Pattern of InteractionThe management team of each modeled rm mustmake N binary decisions about how to congureits activities. N reects the fact that a real rmmust make numerous decisions. It must choose, forinstance, whether to have its own sales force or tosell through third parties, whether to eld a broadproduct line or a narrow one, whether to pursue basicR&D or not, and so forth. An N -digit string of zeroesand ones summarizes all the decisions a rm makesthat affect its performance. We represent this choiceconguration as d = d1d2 dN with each di either0 or 1. Two rms that arrive at the same congu-ration of choices are presumed to achieve the samelevel of performance even if different organizationalstructures guided them to this common set of choices.

Figure 1 Examples of Inuence Matrices (N = 6) A. Independence B. Full interaction

x

x

x

x

x

x

xxxxxx

xxxxxx

xxxxxx

xxxxxx

xxxxxx

xxxxxx

C. Block-diagonal D. Random influence (K = 2)

xxx

xxx

xxx

xxx

xxx

xxx

xxx

xxx

xxx

xxx

xxx

xxx

Put differently, organizational arrangements have nodirect costs or benets. They inuence performanceonly through the operational choices they evoke.2

The efcacy of each decision is affected not onlyby the choice (0 or 1) made concerning that deci-sion, but also by the choices regarding other decisions.In the model, each decision i makes a contributionCi to overall rm performance. Ci depends not onlyon di, but also on how other decisions are resolved:Ci = Cidi; other dj s). An N N inuence matrix,I, records the relationships among decisions. Figure 1gives some examples of inuence matrices for N = 6.The i jth entry of I is marked by an x if columndecision j inuences the contribution of row decisioni and is blank otherwise.3

In the simulations reported below, we set inu-ence matrices in two ways. In some cases, we fully

2 The model also does not incorporate interplay among rms. Thatis, a rms payoff from a conguration is independent of otherrms congurations.3 Our inuence matrix is closely related to the design and taskstructure matrices pioneered by Steward (1981), Eppinger et al.(1994), and Baldwin and Clark (2000) in the context of productdesign.



specify the matrix by hand (Ghemawat and Levinthal2000). We might, for instance, dictate a block-diagonalmatrix of inuences as shown in Figure 1C. With sucha pattern of inuence, one can compare, say, the per-formance of a rm that allocates decisions 13 to onesubordinate and decisions 46 to the other to the per-formance of a rm that divides responsibility in someother way. In other simulations, we simply specifyK, the number of decisions that inuence each deci-sion (Kauffman 1993). The computer then randomlydetermines the identity of the K other decisions thataffect each focal decision. With N = 6, K = 2, the inu-ence matrix might appear as shown in Figure 1D.There, the contribution of each decision to rm perfor-mance is inuenced by the resolution of that decisionitself and the choices made concerning two randomlyassigned other decisions. (Thus, each row containstwo off-diagonal xs.) More generally, K can rangefrom 0 to N 1. K = 0 implies that decisions are inde-pendent. K =N 1 captures a situation in which eachdecision is inuenced by all others. The parameter Kallows us to tune the degree of interaction from fullindependence (see Figure 1A) to full interaction (seeFigure 1B) without specifying particular patterns ofinuence narrowly.

3.2. Generating Decision ProblemsOnce the pattern of interaction is set, the computergenerates a decision problem. That is, it assigns a pay-off to each of the 2N possible congurations of choices.Recall that the contribution Ci of each decision tooverall rm value is affected by other decisions: Ci =Cidi; other dj s). For each possible realization of diand the relevant other dj s, a contribution is drawnat random from a uniform U01 distribution. Theoverall payoff associated with a conguration is theaverage over the N contributions

Pd=N

i=1Cidiother dj s

N

This procedure for generating payoff functionsstochastically, but with well-controlled patterns ofinteractionis adapted from Kauffmans (1993) NKmodel, a model originally developed in the context of

evolutionary biology. Numerous management schol-ars have used the procedure in recent years to gen-erate payoff functions that can be employed to exam-ine organizational search (Levinthal 1997, Gavetti andLevinthal 2000, Ghemawat and Levinthal 2000, Rivkin2000, Marengo et al. 2000, McKelvey 1999 and refer-ences therein). It is common to interpret such payofffunctions in terms of high-dimensional landscapes.Each of the N decisions constitutes a horizontalaxis in a high-dimensional space, and each decisionoffers different options. Resulting from each combi-nation of choices is a payoff for the rm, which isplotted on the vertical axis. The goal of organiza-tional search is to nd and occupy a high spot onthis landscape, i.e., to select a combination of choicesthat, together, are highly successful (Siggelkow 2001).Interactions among decisions cause the landscape tobecome rugged and multipeaked, making the searchfor a high peak profoundly more difcult (Kauffman1993, Rivkin 2000).

3.3. Searching the LandscapesHaving xed a pattern of interaction, we use the pro-cedure described above to generate manytypically10,000landscapes with the same underlying patternof interaction. Onto each landscape (or equivalently,decision problem), we send a set of rms. Each rm ina set searches for a good conguration of choices. Allrms in a particular set start with the same initial con-guration of choices. Firms in a set differ, however,in their organizational designs and the capabilities oftheir management teams. For instance, Firm 1 in a setmight have highly capable subordinates while Firm 2has subordinates of limited ability.

Decomposition: Allocation of Decisions. Eachrm has a management team consisting of a CEO,subordinate Manager A, and subordinate Manager B.Manager A has primary responsibility for a subset ofthe N decisions, and Manager B has responsibility forthe complementary subset. We use a string of as andbs to designate a particular allocation of decisions. Ina simulation with N = 6, for instance, the allocationabbbba would indicate that Manager A has responsi-bility for decisions 1 and 6 while Manager B controlsdecisions 25. We think of decisions 25 as Manager



Bs department. The more that related decisions areassigned to a single managerthat is, the more oftenthat xs in the inuence matrix correspond to pairsof decisions under one managerthe better an allo-cation decomposes the decisions.

Subordinate Capability. Search proceeds in aseries of periods. In each period, each subordinatemanager reconsiders the conguration of choices inhis department. Specically, he compares the statusquo to some number, AltSub, alternatives. Continu-ing with the N = 6 example mentioned above, sup-pose that AltSub = 5 and the current congurationof rm choices is 100111. This means that the currentconguration of choices in Manager Bs departmentis 0011. He considers ve local alternatives to 0011.These include all four of the adjacent alternatives(1011, 0111, 0001, and 0010) and one of the six alterna-tives that involve changing two decisions. (One of thesix is chosen at random.) AltSub reects the cogni-tive abilities of the subordinate manager. A managerwith a higher level of AltSub is smarterable toconsider more alternatives and able to assess the ram-ications of changing more choices at once within hisdepartment.

Incentives: Assessment of Alternatives. Each man-ager ranks the AltSub alternatives from most pre-ferred to least. In assessing alternatives, he puts pri-mary weight on the performance of his department,but he may also consider the effects of his changesbeyond his domain. Incent, a parameter that rangesfrom 0 to 1, captures the degree to which the subordi-nate cares about the ramications of his actions on theother department. Incent= 0 implies that each man-ager considers only effects within his department; thismay reect, for instance, a rm in which managersare paid strictly on the basis of local business unitprotability. Incent= 1 implies that each manager isequally concerned with effects outside his departmentand genuinely wants to maximize rm-wide payoff;this may reect a rm in which divisional ofcers arerewarded for overall corporate performance. Contin-uing with the N = 6 example: In assessing any alter-native d, subordinate Manager B will consider

P d = {C2d+C3d+C4d+C5d+ Incent C1d+C6d

}/6

In evaluating alternatives, each subordinate assumesthat choices in the other department will not change.4

Vertical Hierarchy and the Ability of the CEO.Each subordinate considers the AltSub alternativesand the status quo in his department, and sends upto the CEO the P proposals that he most prefers. Alow level of P reects a rm in which managers areexpected to, or permitted to, narrow down options agreat deal before turning to superiors. A high levelof P reects a rm in which senior managers wantto review many alternatives themselves. Note thatthe term CEO need not be taken literally. We usethe term as a shorthand for any vertical coordinat-ing mechanism, such as an executive committee, thatfullls functions similar to those outlined below.We consider two types of CEOs: rubberstamping

and active. The rst type simply approves ManagerAs favorite proposal and Manager Bs favorite with-out review. A rm with a rubberstamping CEO isequivalent to one with no CEO at all. In such rms,decision making has been completely decentralizedand subordinate managers have full autonomy overdecisions in their departments. In contrast, an activeCEO exercises discretion. From roughly P 2 combina-tions of proposals [(P from Manager A P fromManager B)], she selects AltCEO at random, assessesthem in light of the interests of the rm as a whole,compares them to the status quo, and selects theoption that yields the best payoff for the rm.5

AltCEO reects the cognitive power of the CEO, or

4 Our formulation requires that managers know the total contri-bution of each department or, equivalent, the performance of theentire rm and the total contribution of one department. Experts inaccounting and performance measurement have developed sophis-ticated techniques to isolate the contributions of individual divi-sions, product lines, and functional departments (e.g., Kaplan andAtkinson 1998). These techniques may, of course, err in their mea-surements of contributions. In related work, we are exploringwhat happens when managers in rms with various organizationaldesigns misperceive performance.5 Each subordinate sends up P options, which may or may notinclude the departmental status quo. The CEO always considers thedepartmental status quo as an option, even if the subordinate doesnot submit it. Hence, the CEO may have as many as P + 12 1new combinations at her disposal (if neither manager submits thestatus quo as one of his proposals) or as few as P 2 1 (if bothmanagers submit the status quo as one of their proposals).



more generally, the processing capacity of the coordi-nating unit.In the N = 6 example with decision allocation

abbbba and current choice conguration 100111, sup-pose that P = 2 and AltCEO = 2. Manager B mightsend up for review the alternatives 1011 and 0010 forchoices 25 while Manager A might propose 11 and10 for decisions 1 and 6. From the possibilities, theCEO might select congurations 110110 and 100110for comparison to the status quo 100111. Whicheverof these three yields the highest payoff for the rmis selected and implemented. The new congurationbecomes the launching point for further search inthe next period. Subordinates and CEOs, thus, differin the type of knowledge they possess. Subordinatemanagers have local knowledge that allows them togenerate proposals for their departments. CEOs pos-sess global knowledge that enables them to assess thefull ramications of departmental choices on overallrm performance.6

In sum, rms differ in their organizational arrange-ments: the grouping of decisions into departments,the amount of information conveyed to senior man-agement P, the degree to which the CEO acts uponthat information (rubberstamping vs. active), and theincentives that managers have to consider effectsbeyond their domains (Incent). Firms also differ inthe abilities of their subordinates (AltSub) and theirCEOs (AltCEO). Overall, the organization we envi-sion resembles the one examined by Bowers (1970)classic study of the resource allocation process. Seniormanagement lays out some basic structural elementsof the rm: the allocation of decision rights and theincentive system, for instance. Subject to those rulesof the game (Jensen et al. 1999), lower level managers

6 For parsimony, our formulation suppresses a number of consider-ations worthy of future research. For example, our managers haveno cognitive representations (Gavetti and Levinthal 2000), our CEOhas no agenda of her own, our rms implement plans without error(Siggelkow 2002), and our subordinate managers do not laterallycommunicate. Note, however, that our CEO is functionally equiv-alent to a form of lateral communication in which subordinatesrank departmental options, convene in a conference room, considercomposite alternatives, and pick the composite that is best for therm. P then reects the number of options that subordinates bringto the conference room, and AltCEO reects the limited time thatmanagers can afford to spend there.

select and promote proposals that they nd attractive.Senior management then exercises some discretion inselecting among, and integrating across, the proposalsthat bubble up.

3.4. Sticking PointsFirms continue to search for many periods. In many(but not all) cases, rms reach sticking points aftera number of periods. That is, they reach congura-tions of choices from which they do not move. From asticking point, there is no alternative conguration ofthe N choices within the search radius of the rm thatmeets the approval of enough actors within the rmthat the alternative can be adopted. AltSub inu-ences how broad the search radius is. Organizationalarrangements dictate the standards that an alterna-tive set of N choices must meet to be accepted. Forinstance, when the CEO exercises discretion, one suchstandard is that the alternative must yield a higherpayoff for the rm as a whole than the status quoachieves. The same standard does not apply when theCEO simply rubberstamps proposals and Incent islow. Then an alternative that is in the interest of justone department may be implemented.In conceptions of organizational search, it has been

common to think of rms as getting stuck on localpeaks (e.g., Alchian 1950, Levinthal 1997). A localpeak is a conguration of choices for which a changein any single choice leads to worse rm performance,even though simultaneous changes in several choicesmay improve rm performance. In contrast, our morerichly modeled rms may well get stuck at pointsother than local peaks (Rivkin and Siggelkow 2002).The set of sticking points is neither a sub- nor a super-set of the set of local peaks. Suppose, for instance, thata rm with low Incent and a rubberstamping CEOsits atop a local peak. It is quite possible that a sub-ordinate will discover and implement a move that isbenecial for his department but detrimental for therm, causing the rm to descend from the peak. Thus,a local peak might not be a sticking point. Similarly,a sticking point need not be a local peak. Consider asituation in which a rm is one decision away from alocal peak but the change required to attain the peakis not in the interest of the manager who controls therelevant decision. In such an instance, the manager



may never propose the needed change, and a rmmay get stuck on the hillside below a peak. At sucha sticking point, it is also possible that a subordinatewill want to make some incremental change, but theCEO will veto it. In that case, the rm is stuck eventhough it is not on a local peak of the overall rmslandscape nor is it on a local peak of the lower dimen-sional subscape dened by departmental choicesand payoffs.Likewise, sticking points are related to, but differ-

ent from, the concept of Nash equilibrium outcomes.With a rubberstamping CEO, all sticking points areindeed Nash equilibrium outcomes in a game playedbetween the two subordinates: At a sticking point,each subordinate is picking the best alternative (forhim) given the other subordinates current decisions.However, once the CEO becomes active, the setsof sticking points and Nash equilibrium outcomesdiverge and become neither sub- nor supersets of eachother. At a Nash equilibrium outcome, each playerin a game must be taking the best possible actiongiven the strategies of the other players. Unlike play-ers at a Nash outcome, our subordinate managersdo not anticipate the CEOs reaction when they pro-pose alternatives. As a result, they may forego self-benecial opportunities. For instance, they alwayssend up their most preferred proposals, even whenthose proposals are subsequently rejected by the CEO.A forward-looking subordinate might opt to makea proposal that he prefers less but the CEO willaccept. This failure to consider the CEOs strategydrives a wedge between the sets of sticking pointsand Nash equilibria when an active hierarchy is inplace. In an appendix available on this journals web-site mansci.pubs.informs.org, we discuss in detailthe subtle relationship between sticking points andNash equilibrium outcomes.

4. ResultsWe conducted a comprehensive set of analysesinvolving each of the ve featuresvertical hierarchy,incentives, decomposition, degree of decision inter-action, and managerial abilityand all the relationsamong them. A few themes recur in the results.First, certain sets of design elements encourage rms

to search and evaluate a broad array of optionswhile others lead rms to stabilize and cease theirsearch. Second, rms that perform well typically bal-ance search and stability. Third, an organizationaldesign that promotes search is especially effectivewhen underlying decisions intensely interact with oneanother.To illustrate these themes, we focus on a series of

interdependencies that involve the vertical hierarchy7

(see Table 1 for a summary). Mirroring the presenta-tion of conventional wisdom in 2, we rst examinethe effects of an active CEO in isolation. We conrmthat an active CEO can slow down decision making,but we also identify another hazard of an active CEO:She acts as a strong force for stability and can prema-turely channel her rm toward a low sticking pointbefore adequate search is undertaken. This hazard isalleviated, however, when the active CEO is coupledwith features that encourage broad searchrich infor-mation ow, capable subordinates, rm-wide incen-tives, and incomplete decision decomposition. Thestabilizing inuence of the active CEO and the searchinduced by other factors create the interdependen-cies we report here. Some of the interdependenciesconrm conventional wisdom, but others bound orchallenge it.Often in this section, we report that one type of

rm achieves a higher level of performance on aver-age than another. In each instance, the differencein mean performance is statistically signicant withp < 0001, assuring that reported differences are notsimply chance occurrences caused by the stochasticnature of the landscape generation. We report rmperformance as a portion of the highest performanceattainable on each landscape that was explored.

Robustness. We have observed the qualitative pat-terns described in this section under a far broaderrange of parameter values than reported here, and weare happy to share our simulation software with anyresearcher who wants to probe the robustness of par-ticular results in depth. That said, our primary goal in

7 Interdependencies also exist that do not involve the vertical hier-archy, but the interdependencies we examine here are sufcient toillustrate our general themes. Full results that show other interde-pendencies are available from the authors.



Table 1 Summary of Results

Conventional wisdom Results

Active vs. passivevertical hierarchyand degree ofinteraction

An active vertical hierarchy can beoverloaded, which slows down decisionmaking

An active vertical hierarchy is more valuablewhen decisions richly interact

Conrmed when interactions are sparse. When interactions are dense,however, the coordinating benet of the hierarchy outweighs the cost ofoverloading.

Conrmed for short-run performance. However, an active hierarchy canundermine long-term performance when interactions are dense andinformation ow is limited, by rapidly locking the rm into a poor set ofdecisions.

Active verticalhierarchy andmanagerial ability

An active hierarchy is less valuable whensubordinates are highly capable

An active hierarchy is more valuable when subordinates are highly capableand interactions are dense. An active hierarchy is required to stabilizeexcessive search of smart managers.

Active verticalhierarchy andrm-wideincentives

An active hierarchy is less valuable whensubordinates are rewarded for rm-wideperformance

An active hierarchy is more valuable when subordinates are rewarded forrm-level performance, especially when subordinates are highly capable.Firm-wide incentives alone may not achieve stability, because they coordinateonly intentions, not actions. The active hierarchy provides necessary stability.

Active verticalhierarchy anddecisiondecomposition

An active hierarchy is less valuable whendecisions can be decomposed so that fewcross-departmental interactions remain

Decisions should be allotted to subordinatesso that cross-departmental interactions areminimized

Conrmed. When decisions are completely decomposed, there is no benetof coordination across departments, so the value of hierarchy is zero.

Conrmed for a rubberstamping vertical hierarchy. With an active hierarchy,however, leaving unnecessary interactions between departments can yieldhigher performance. Incomplete decomposition creates additional searchthat can be exploited by an active, stabilizing hierarchy.

this paper is not to prove the generality of any single,ne-grained result. Rather, we aim to illustrate par-ticular ways in which the elements of organizationaldesign can relate to one another, and to identify broaddrivers of those relationships across a wide range ofresults.

4.1. Active vs. Passive Vertical Hierarchy andDegree of Interaction

Before considering the interdependency between anactive CEO and other design elements, one must care-fully examine the effect of an active CEO in isola-tion. Recall that the active CEO vets the proposalsof her subordinates, weighing interactions that localmanagers ignore and accepting only those changesthat serve the rm as a whole. Viewed in this light,the CEO seems to be an unalloyed source of bene-t. The coordinative value of an active CEO shouldbe particularly high when interactions are pervasive,received wisdom tells us. Conventional wisdom, how-ever, also points to a potential downside of active

vertical hierarchies: Active CEOs may slow downdecision making. In this subsection, we conrm thesetwo pieces of conventional wisdom, but put boundaryconditions on both.We analyze rms that face decision problems with

N = 6. In each rm, Manager A controls the rst threedecisions and Manager B the last three; i.e., the deci-sion allocation is aaabbb. Each manager considers onlyone alternative per period (AltSub = 1). We examineall possible degrees of interaction among the decisions(K = 01234, or 5). The rms we analyze differ inthe degree of CEO activity (i.e., whether the CEO doesanything more than rubberstamp proposals), the abil-ity of the CEO to consider multiple options (AltCEO),and the number of proposals that subordinatessubmit (P ).Specically, we consider the four rms described

on the top panel of Table 2. In Firm 2A, the CEOrubberstamps proposals. In Firm 2B, each managersends up his preferred option to a CEO who considersone composite alternative per period. Firm 2C differs



Table 2 Effect of Active vs. Rubberstamping CEO on Performance

Firm 2A 2B 2C 2DAltCEO R 1 1 3P 1 2 1

Performance in period 4

K = 0 0951 0939 0886 0950K = 1 0884 0893 0857 0905K = 2 0842 0862 0838 0877K = 3 0814 0846 0827 0861K = 4 0797 0835 0825 0850K = 5 0780 0828 0822 0842

Performance in period 100

K = 0 1000 1000 1000 1000K = 1 0943 0952 0987 0954K = 2 0910 0918 0970 0921K = 3 0894 0897 0956 0900K = 4 0886 0880 0945 0884K = 5 0882 0865 0938 0869For K = 5:Benet of active CEO in period 100 0017 0056 0013Portion of landscapes on which a rm

still wanders after period 80 39% 00% 21% 00%Number of sticking points 39 130 41 130Average height of sticking points 0869 0827 0918 0827

Note. N = 6, aaabbb, Incent = 0, AltSub = 1. R indicates rubberstampingCEO. Performance is an average across 10,000 landscapes. Sticking pointresults are an average across 500 landscapes.

Differences denoted by are signicant with p < 0001.

from Firm 2B in that each manager sends up two pro-posals. In Firm 2D, each manager sends up one pro-posal, but the CEO can consider all new compositealternatives (i.e., as many as three) in each period. Weexamine the performance of each rm in period 4 andperiod 100, representing the short run and long run.(We choose periods 4 and 100 simply because mostrms take considerably more than 4 and less than 100periods to reach their long-run levels of performance.Qualitatively, the results are not sensitive to the choiceof periods 4 and 100.)

Short-Run Performance. The rst six rows ofTable 2 compare the ability of different rms toquickly scramble uphill, for varying levels of K. Theresults for K = 0 conrm the conventional wisdomthat an active but overloaded CEO can be a lia-bility. Firm 2A, with its rubberstamping CEO and

decentralized decision making, performs as well asor better than any of the rms with active CEOs.An active CEO undermines performance because sheis overwhelmed with proposals and becomes a bot-tleneck, standing between good proposals from thedepartments and implementation of those proposals.In Firm 2B, for instance, the CEO has up to three com-posite alternatives to evaluate each period, but canassess only one. Overloaded, she may ignore goodproposals that are sent to her by the subordinates. Asone would expect, the effect is exacerbated when thenumber of proposals rises (Firm 2C) and mitigatedwhen the CEO can process more options per period(Firm 2D).As interactions across departmental boundaries

proliferate, the active CEO changes from liability toasset. At high K, Firm 2B fares better than Firm 2A inperiod 4. On rugged landscapes, the benets of a CEOwho takes account of interactions across departmentsoutweigh the danger of CEO overload. Put differently,when K is high, local managers do not have a goodidea of what proposals are valuable for the rm as awhole. Hence, even if the CEO does not get around tolooking at the proposal most preferred by a manager,not much harm is done. This nding puts a boundarycondition on the conventional wisdom about over-loaded CEOs: If interactions across departments aredense enough, an active CEO is benecial in the shortrun despite the hazard that she may delay acceptanceof good proposals.

Long-Run Performance. Intuition suggests that theCEO-overload problem should fade in the long run.If a good proposal is sent up enough times, at somepoint the CEO will consider it, accept it, and ceaseto be a bottleneck. This is particularly salient on asmooth landscape (i.e., K = 0). Given enough time,every rmeven one with a very overloaded CEOshould reach the global peak of such a landscape.In the following results, we conrm this intuition,but identify another persistent drawback of an activeCEO.The middle panel of Table 2 reports the perfor-

mance of each rm in period 100. When K = 0, weexpect and nd that all rms reach the global peak inthe long run. For high values of K, however, Firm 2A,which has a rubberstamping CEO, performs better



than Firm 2B, which has an active CEO. Why is theactive CEO a detriment even in the long run forFirm 2B? Firms with active CEOs never move down-hill on a landscape; the CEO vetoes such a maneuver.As a result, these rms quickly reach sticking pointsbefore considering a wide range of alternatives. Theyrun the risk of excessive stability. In contrast, rmswith rubberstamping CEOs will sometimes imple-ment alternatives that cause overall performance todecline temporarily. This promotes a wider search ofpossibilities and can lead, in the long run, to higherperformance. The nal two rows of Table 2 supportthis interpretation for the case of K = 5. Firm 2B per-ceives far more sticking points than Firm 2A and ismuch more likely to get stranded on a low stickingpoint, i.e., a poor compromise among the subordi-nates and the CEO. In sum, Firm 2B has more stabil-ity and undertakes less search than is optimal wheninteractions are pervasive.Employing a smarter CEO, as Firm 2D does, does

not alleviate the problem that an active CEO causeswhen K is high; an ability at the top of the orga-nization to assess more alternatives does not help iftoo few alternatives are being proposed. On the otherhand, it is helpful to insist on a greater ow of infor-mation from subordinates to the CEO as Firm 2Cdoes. With more proposals (P ) coming in, the dan-ger of premature lock-in diminishes and the benetsof an active CEOespecially her ability to keep sub-ordinates from acting in ways that undermine over-all performancereassert themselves. The search pro-moted by higher P productively balances the stabilityprovided by the CEO.The active CEO, thus, has quite different draw-

backs in the short run and long run. In the shortrun, especially if interactions are sparse, the CEO canpose a bottleneck, blocking the rapid progress that adecentralized rm would enjoy. The problem is exac-erbated by a greater ow of information and allevi-ated by the hiring of a smarter CEO. In the long run,an active CEO can funnel a rm prematurely to amediocre sticking point. The problem is alleviated bya greater ow of information, but not by an increasein the CEOs processing power. The risk of prematurelock-in is especially acute when interactions are per-vasive. Interactions make the underlying landscape

rugged and multipeaked, which provides a naturalsource of stability. It is in this setting that the addi-tional stability provided by an active, low-P CEO canharm rm performance. Thus, we pinpoint a bound-ary condition on the conventional wisdom that activehierarchies are more valuable when interactions arepervasive: This appears to be true only if the CEOreceives a rich ow of information. If interactions arepervasive and the CEO gets little information, sheserves the rm better by rubberstamping subordi-nates proposals than by actively exercising oversight.The long-run results in Table 2 illustrate a gen-

eral pattern: As interactions among decisions becomemore pervasive, design elements that encouragebroad search grow more important. Comparing Firms2B and 2C, for instance, we see that the incrementalbenet of a richer ow of information, which pro-motes more search, rises steadily from 0.000 when K=0 to +0073 when K = 5.The ndings in this section, taken as a whole,

also suggest that very different vertical hierarchiesmay be suitable for volatile and stable environments.In volatile settings, rms essentially face a series ofshort-run problems. Attaining decent results quicklyrequires either a passive CEO who lets subordinatesmake nal choices (when K is low) or a very smartCEO who acts on the basis of limited information(when K is high). In stable settings, rms can focuson long-run performance, which is best delivered byan active, not-necessarily-brilliant CEO who receivesand reviews numerous proposals. This conclusion isreminiscent of Burns and Stalkers (1961) ndingson organic and mechanistic organizations and Brownand Eisenhardts (1997) observations of successfulorganizational structures in highly dynamic environ-ments. This speculative interpretation deserves fur-ther research.

4.2. Active Vertical Hierarchy andManagerial Ability

In the last subsection, we identied the stabilizingeffect of the active CEO in the long run, noted thatthe effect may be harmful when interactions pro-vide natural stability, and illustrated how the searchinduced by richer information ow may counterbal-ance the CEOs stability. In the next three subsections,we examine how other design elements and contex-



Table 3 Interdependency Between Active CEO and Subordinate Capability with Incent= 0Firm 3A 3B 3C 3D 3E 3F 3G 3H 3I 3J 3K 3LAltCEO R 1 1 3 R 1 1 3 R 1 1 3P 1 2 1 1 2 1 1 2 1AltSub 1 1 1 1 4 4 4 4 7 7 7 7

Performance in period 100 for:K = 0 1000 1000 1000 1000 1000 1000 1000 1000 1000 1000 1000 1000K = 1 0943 0952 0987 0954 0955 0964 0981 0967 0951 0960 0979 0964K = 2 0910 0918 0970 0921 0921 0938 0965 0943 0897 0929 0962 0938K = 3 0894 0897 0956 0900 0900 0920 0953 0926 0854 0912 0948 0922K = 4 0886 0880 0945 0884 0884 0909 0943 0917 0825 0900 0937 0915K = 5 0882 0865 0938 0869 0870 0900 0936 0910 0810 0896 0931 0912For K = 5:Benet of active CEO 0017 0056 0013 0030 0066 0040 0086 0121 0102Portion of landscapes on which a rm

still wanders after period 80 39% 00% 21% 00% 367% 00% 03% 00% 608% 00% 00% 00%Number of sticking points 39 130 41 130 12 74 29 74 10 92 51 92Average height of sticking points 0869 0827 0918 0827 0928 0873 0939 0873 0936 0852 0898 0852

Note. N = 6, aaabbb, Incent = 0. R indicates rubberstamping CEO. Performance for each level of K is an average across 10,000 landscapes. Sticking pointresults are an average across 500 landscapes. Differences denoted by are signicant with p < 0001.

tual features may also provide balance. We start byanalyzing the effects of smart subordinates who canconsider a wide array of options within their depart-ments. We show that the search undertaken by suchsubordinates can balance the stability of an activeCEO, making capable subordinates and active CEOscomplementary in our model. This runs contrary toconventional wisdom for reasons we explain below.To analyze the interdependency between hierar-

chy and subordinate ability, we engage in simula-tions with the same parameter settings as in theprevious subsection, but vary the number of alterna-tives that each manager is able to evaluate in eachperiod (AltSub). The left, middle, and right panelsof Table 3 show performance in period 100 for rmswith subordinates who are able to evaluate one, four,and seven alternatives per period, respectively. Asbefore, all subordinates pursue departmental perfor-mance (Incent= 0).8The rst striking result shown in Table 3 is that

a rm can undermine its long-run performance by

8 This subsection and the following two focus on interdependenciesas exhibited in long-run performance. Short-run results, availablefrom the authors, display interdependencies that are qualitativelysimilar.

hiring smarter managers, especially when decisionsrichly interact and the CEO rubber-stamps decisions.A comparison of the three rms with rubberstamp-ing CEOsFirms 3A, 3E, and 3Imakes this clear.For K = 5, for instance, performance declines steadilyas AltSub rises from 1 to 4 to 7. For all valuesof K > 0, an increase in AltSub from 4 to 7 leadsto worse performance. Why can smarter managersundermine long-run performance? The answer lies inthe problem that smart subordinates create for eachother when their domains inuence one another. Asmart subordinate searches broadly and undertakesfar-reaching changes to improve the performance ofhis department. In doing so, however, he under-mines the improvement efforts that the other equallysmart subordinate is making. Hence, a pair of smartsubordinates can dance abouteach making radicalchanges that seem like uphill movements from hisperspective but that deform the landscape as the othersees it. The higher is AltSub, the graver is the dangerof this ongoing dance. Reecting this, the portion ofrms that still wander after period 80 rises from 3.9%for Firm 3A to 60.8% for Firm 3I, and the number ofsticking points falls from 3.9 to 1.0.



The broad search pursued by smart subordinatesmay be harnessed to good effect if an active, stabiliz-ing CEO is present. Accordingly, we see performancerise with AltSub, or at least fail to fall dramaticallyas AltSub increases, when a CEO reviews and ratiesproposals. For any given level of AltSub, the portionof rms wandering after period 80 falls dramaticallyand the number of sticking points rises once an activeCEO is in place. Moreover, the performance benetof an active CEO relative to one who rubberstampsproposals, a gure reported in the bottom portion ofTable 3 grows with AltSub. For instance, for a CEOwith P = 1 and AltCEO= 1, the benet increases from0017 to +0086 as AltSub rises from 1 to 7.An active hierarchy and capable subordinates are

complements in our model; that is, the benet of hav-ing an active CEO rises with AltSub. In contrast,received wisdom holds that the two are substitutes:Senior management oversight is less necessary whensubordinates are highly capable. The difference stemsfrom distinct perspectives on the role of senior man-agement. To simplify, the conventional view holdsthat senior managers are responsible for handlingexceptions thrown up by the outside world: Thehierarchy of authority is employed on an exceptionbasis. That is, the new situation, for which there isno preplanned response, is referred upward in thehierarchy to permit the creation of a new response(Galbraith 1973, p. 11). Senior managers are presumedto have the expertise and experience to create newresponses and to solve unusually difcult problems(Garicano 2000). The more capable are subordinates,the rarer are such exceptions and the less necessaryis an active hierarchy. In contrast, we envision seniormanagers as the integrators and ratiers of subor-dinates proposals for change. An increase in thepotential scope of subordinate proposals makes seniormanagement all the more necessary, especially whendecisions interact with one another. We interpret ourndings as a boundary condition on conventionalwisdom. Where senior managers are repositories ofexpertise that enable a rm to handle surprising cir-cumstances, an active hierarchy and capable subor-dinates are likely to be substitutes. But where seniormanagers serve to check and integrate internal pro-posals, the two may be complements. Smart managers

may offset any excess stability of the active hierar-chy while the hierarchy stabilizes the radical searchundertaken by smart managers.9

So far, we have presented smarter managers as asource of broader search. There is, however, a wayin which smarter managers can prevent wide searchand exacerbate the potential for excessive stability.Note in Table 3 that performance sometimes declineswith higher AltSub even for a rm with an activeCEO; for instance, the performance of Firm 3K withAltSub = 7, is modestly lower than that of Firm 3Cwith AltSub = 1 for all K > 0. Because active CEOsprevent excessive wandering, the reason for the per-formance difference in such cases cannot be a failureto achieve stability. Rather, the difference is drivenby restricted search. In Firm 3C, each subordinate isessentially forced to send up two random propos-als every period. In contrast, in Firm 3K, each sub-ordinate ranks seven alternatives and sends up histwo most preferred. As a result, the heterogeneity ofproposals received by the CEO is larger in Firm 3Cthan in Firm 3K, leading Firm 3C to experience widersearch. When a subordinate is so smart that the num-ber of alternatives he considers is much greater thanthe number of proposals he must reveal, the pre-screening he performs can restrict search. In essence,the subordinate is able to hide options he dislikesbehind the tall stack of alternatives he has considered.This dynamic plays a part in the next subsection.

4.3. Active Vertical Hierarchy andFirm-Wide Incentives

Conventional wisdom suggests an interdependencybetween incentive systems that stress rm-wide out-comes and the presence of an active CEO: Firm-wideincentives may reduce or even eliminate the need foran active CEO. Bounding this wisdom, we nd situa-tions in which rm-wide incentives can induce broad

9 Smart managers and an active hierarchy are complements in thesense of Milgrom and Roberts (1990): The adoption of one increasesthe benet of adopting the other. To an actual manager in such arm, however, the two countervailing items may not feel particu-larly consistent. A smart manager may be frustrated, for instance,by what he perceives as the meddling of an active CEO. This mayeven undermine the subordinates motivation to search for betterchoices, a consideration we do not model here.



search, thereby making the stability of an active CEOmore valuable.In all previous simulations, each subordinate man-

ager was under a parochial incentive system, eval-uating alternatives only from the perspective of hisindividual department. Here, we conduct simulationsthat have the same parameter settings as those in theprevious subsection except that the value of Incentis set to 1. Thus, in evaluating alternatives, each sub-ordinate manager takes into account the full effect ofhis actions on the rm as a whole. Table 4, whichreports the results, has precisely the same structure asTable 3.The key results are most easily seen by comparing

the lines labeled Benet of active CEO in Tables 3and 4. Especially for rms with AltSub = 4 or 7, thebenet of having an active CEO increases as Incentrises from 0 to 1. For example, for Firm 3J withIncent = 0, the benet of having an active CEO is+0086, while for Firm 4J with Incent = 1, the ben-et is +0211. Accordingly, the active CEO and rm-wide incentives complement one another. This resultarises because rm-wide incentives within an activevertical hierarchy promote search that balances thestability provided by the active CEO. Why do rm-

Table 4 Interdependency Between Active CEO and Subordinate Capability with Incent= 1Firm 4A 4B 4C 4D 4E 4F 4G 4H 4I 4J 4K 4LAltCEO R 1 1 3 R 1 1 3 R 1 1 3P 1 2 1 1 2 1 1 2 1AltSub 1 1 1 1 4 4 4 4 7 7 7 7

Performance in period 100 for:K = 0 1000 1000 1000 1000 1000 1000 1000 1000 1000 1000 1000 1000K = 1 0971 0970 0987 0972 0973 0982 0987 0986 0941 0981 0987 0987K = 2 0946 0945 0970 0946 0953 0964 0972 0968 0864 0964 0974 0972K = 3 0922 0921 0956 0923 0926 0946 0957 0952 0796 0948 0960 0959K = 4 0904 0902 0945 0905 0894 0935 0946 0942 0755 0938 0948 0951K = 5 0884 0883 0937 0885 0845 0925 0937 0934 0718 0929 0937 0945For K = 5:Benet of active CEO 0001 0053 0001 0080 0092 0089 0211 0219 0227Portion of landscapes on which a rm

still wanders after period 80 00% 00% 23% 00% 303% 00% 01% 00% 716% 00% 00% 00%Number of sticking points 92 92 40 92 49 49 29 49 43 43 40 43Average height of sticking points 0857 0857 0921 0857 0903 0903 0939 0903 0913 0913 0917 0913

Note. N = 6, aaabbb, Incent = 0. R indicates rubberstamping CEO. Performance for each level of K is an average across 10,000 landscapes. Sticking pointresults are an average across 500 landscapes. Differences denoted by are signicant with p < 0001.

wide incentives promote search in the presence of anactive CEO? Recall that the active CEO rejects alter-natives that are detrimental to the rm as a whole.When incentives are parochial, the CEO often receivesproposals that do not benet the rm. She rejectsthese proposals out of hand so the proposals do notlead to effective exploration of the landscape. In con-trast, when subordinates face rm-wide incentives,the CEO receives far more proposals that are accept-able to her, and much more movement ensues. In linewith this intuition, we see that rms with an activeCEO experience fewer sticking points when Incent=1 than when Incent= 0.In the absence of an active CEO, however, rm-

wide incentives can lead to less search than parochialincentives, because subordinates have to nd alterna-tives that are benecial for the rm as a whole, ratherthan only for their departments. This effect is sug-gested by the larger number of the sticking points per-ceived by the rubberstamping rms in Table 4 versusthose in Table 3.Our nding that an active hierarchy is a comple-

ment to rm-wide incentives runs contrary to conven-tional wisdom. The intuition behind the conventional



wisdom is straightforward: If subordinates have theinterests of the overall rm at heart, why does oneneed oversight? Our answer is that rm-wide incen-tives can coordinate the intentions of subordinates,but they do not necessarily coordinate the actions ofsubordinates when decisions interact. Capable sub-ordinates can engage in aggressive, well-intentionedsearch that results in mutually destructive improve-ment. This possibility is most vividly illustrated byFirms 3I and 4I, which have highly capable man-agers and rubberstamping CEOs. Firm-wide incen-tives cause a precipitous drop in performance forK > 0 as subordinates engage in search uncheckedby an active CEO. In contrast, rm-wide incentivesimprove performance when an active CEO is present.The active hierarchy provides a device to coordinateactual moves, not just motives.If rm-wide incentives encourage broad search

in rms with active CEOs, one might expect themarginal benet of such incentives to rise as interac-tions among decisions become pervasive, which intro-duces natural stability. A comparison of Tables 3and 4 conrms this: For rms with active CEOs, themarginal benet of rm-wide incentives increases asK rises from 0 to 2, then levels off. Comparing Firms3J and 4J, for instance, we see that the marginal ben-et of rm-wide incentives rises from 0.000 for K = 0to +0021 for K = 1 to +0035 for K = 2, then stabilizesfor higher K.

4.4. Active Vertical Hierarchy andDecision Decomposition

Lastly, we turn to the interdependency between anactive CEO and the decomposition of decisions intodepartments. As discussed in 2, a consensus existsin the qualitative literature on organizational designthat rms should, as much as possible, assign deci-sions that inuence one another to the same manager.This manager is able to internalize the interactionsamong decisions and nd the departments choiceconguration that is best for the rm. Moreover, con-ventional wisdom points out that when decisions arefully decomposedgrouped such that all interactionsare internalizedan active vertical hierarchy mightbe unnecessary.

To study the interdependency between decisiondecomposition and the active CEO, we examine aseries of simulations in which the inuence matrix isblock-diagonal as shown in Figure 1C. That is, rmsface decision problems in which N = 6, and all inter-actions are among decisions 13 and decisions 46.While the particular prot contributions change fromrun to run, this pattern of interaction stays the same.Incentives are parochial (Incent= 0), and each man-ager considers one local alternative to his current con-guration of choices (AltSub= 1). Firms differ in themanner in which decisions are allocated to subordi-nates and whether their CEOs exercise active discre-tion (see Table 5). Firms 5A and 5B have the decisionallocation aaabbb, which completely decomposes therm into two independent parts, while Firms 5C and5D have decision allocation aabbba, which leaves inter-actions between the departments. Firms 5A and 5Chave rubberstamping CEOs, whereas Firms 5B and5D have active CEOs.The simulation results are in line with the sec-

ond aspect of conventional wisdom noted above:when decisions are completely decomposed, the per-formance benet of an active CEO is nil (per-formance of Firm 5A = performance of Firm 5B).Suggested improvements that are sent up benet theproposing department andbecause of the completedecompositionhave no effect on the other depart-ment. Moreover, the most preferred departmentalimprovements are also the most benecial for therm as a whole. As a result, the CEO always accepts

Table 5 Interdependency Between Active CEO and Decomposition

Firm 5A 5B 5C 5DDecision allocation aaabbb aaabbb aabbba aabbbaAltCEO R 3 R 3P 2 2

Performance in period 100 0937 0936 0898 0979Benet of active CEO 0001 0081Portion of landscapes on which a

rm still wanders after period 80 00% 00% 221% 00%Number of sticking points 41 41 18 18Height of average sticking point 0916 0916 0897 0968

Note. N = 6, block-diagonal I, Incent = 0, AltSub = 1. R indicates rubber-stamping CEO. Performance is an average across 10,000 landscapes. Stick-ing point results are an average across 500 landscapes. Differences denotedby are signicant with p < 0001.



preferred proposals of each subordinate and, conse-quently, acts much like a rubberstamper. Accordingly,the performance of Firm 5B is indistinguishable fromthat of Firm 5A. In contrast when interactions remainacross the departments, the CEO provides valuablestability (performance of Firm 5D performance ofFirm 5C). Firm 5C, with a rubberstamping CEO, failsto come to a sticking point on 22.1% of the landscapes.In contrast, Firm 5D with an active, stabilizing CEO,always comes to a steady conguration of decisions.Intriguingly, the best performance in Table 5 is

attained by Firm 5D, which has an active, well-informed CEO and unnecessary overlap betweendepartments. This result contradicts one of the mostcommon pieces of received wisdom, that decisionsshould be allocated to minimize cross-departmentalinteractions. The cause of Firm 5Ds superior per-formance is again a helpful balance of search andstability: The overlap across departments generateswide search, as each subordinate proposesand issometimes allowed to enactoptions that changeconditions in the other department, prompting newsearch in that department. The active CEO ensuresthat the rm eventually stabilizes around any greatoption that the wide search produces. Put differ-ently, the incomplete decomposition generates searchthat the active CEO can take advantage of. Thus,we see imperfect decomposition and an active CEOas complements, the search generated by one work-ing well in concert with the stability generated bythe other.10 In sum, we place an important bound-ary condition on the conventional wisdom that rmsshould strive for complete decomposition. Unneces-sary overlap between departments can induce subor-dinates to make creative proposals that pry rms offof low sticking points. The ensuing search, coupled

10 The bottom of Table 5 supports this interpretation. Compared tothe completely decomposed Firm 5B, Firm 5D gets stuck on fewerand higher points; departmental overlap shakes Firm 5D off oflow sticking points and encourages it to explore possibilities morewidely. Compared to Firm 5C, which has a rubberstamping CEO,Firm 5D is unlikely to wander forever; it stabilizes around its highsticking points. While Firm 5B has only the advantage of stabil-ity and Firm 5C has only the advantage of search, Firm 5D enjoysboth.

with an active, stabilizing CEO, can produce superiorperformance.11

5. DiscussionThe existence of interdependencies among elementsof organizational design has become a bedrock propo-sition in the literature on organizations, yet rela-tively little is known about the underlying forcesthat create these interdependencies. Using a simu-lation model of organizational design and search,we identify one such force: Interdependencies arisebecause design elements inuence both how broadlya rm searches its environment to discover goodsets of coordinated choices and whether the rm isable to stabilize around those sets once they are dis-covered. The adoption of an element that encour-ages broad search typically raises the marginal benetof other elements that provide offsetting stability.This duality between search and stability has playeda central role in substantial prior researchon theproductivity dilemma (Abernathy 1978), static anddynamic efciency (Ghemawat and Ricart i Costa1993), exploration and exploitation (March 1991),and the ambidextrous organization (Tushman andOReilly 1996), for instance. We add to these prece-dents in three ways.First, in the context of a formal model of search, we

associate search and stability with specic combina-tions of design elements (see Figure 2). The designersof an organization can promote stability by employ-ing an active CEO, who rejects proposals that makethe rm as a whole worse off, or by decomposingdecisions such that no cross-departmental interactionsremain. Designers can broaden search by increasingthe number of proposals sent to the CEO. The effectsof rm-wide incentive systems and subordinates whoare able to evaluate many alternatives are more sub-tle; they depend on the degree of discretion the CEOexercises. When the CEO is active and subordinatesare allowed to send only a few proposals, smartersubordinates can curtail search because they are more

11 As Siggelkow and Levinthal (2003) show, such unnecessary over-lap can also be helpful for a rubberstamping rm if the rm sub-sequently centralizes decision making by assigning all decisions toone department.



Figure 2 Organizational Design Elements and Their Effects on Searchand Stability

Rubberstamping CEO Active CEO

Active CEO stability

Rich informationow search

Smart managers search stability*

Firm-wideincentives stability search

Decomposition stability

*if information ow is limited

able to hide parochially distasteful alternatives fromthe CEO. In contrast, when the CEO rubberstampsproposals, rms with smarter subordinates engage inbroader search because managers are given free rein.Firm-wide incentives can broaden search when theCEO is active by reducing the proportion of proposalsthat are rejected outright by the CEO, but they tend torestrict search when the CEO is passive because eachsubordinate then has to nd solutions that benet therm as a whole, not his department alone.Our second contribution is to emphasize the need

for an organization to strike a balance between searchand stability. While much of the prior literature high-lights the tension between the two, we focus onways in which they can work together. We nd, forinstance, that it can be useful to couple an active,stabilizing CEO with a rich vertical ow of informa-tion that promotes search (4.1). Similarly, the broadsearch generated by smarter managers (4.2), by rm-wide incentives (4.3), or by an incomplete decisiondecomposition (4.4) can be harnessed if it is balancedby the stability of an active CEO. Our results thus pin-point the conditions that make an active vertical hier-archy especially valuable: very capable subordinates,incentives that stress rm-wide outcomes, and deci-sion interactions across departments. Our ndingsthat capable subordinates and rm-wide incentivescomplement an active hierarchy depart from conven-tional wisdom because we view top-level managersas integrators, not as exception handlers. Smart man-agers, even if they have the rms interest at heart,may still require coordination. Similarly, rm-wide

incentives, while aligning intentions, may still fail toalign actions.Third, we show how the underlying pattern of

interaction among decisions affects the appropriatebalance between search and stability. The greateris the degree of interaction among decisions, themore rugged are the landscapes that rms face. Thisruggedness provides built-in stability. A rm can pro-ductively counter this stability by shifting its organi-zational arrangements in favor of search. Hence, wesee a need for a rubberstamping CEO or a rich verticalow of information when interactions are pervasiveand subordinates do not evaluate many alternatives(4.1). Similarly, we nd that the marginal benets ofsearch-promoting, rm-wide incentives increase withthe density of interactions (4.3).The three contributions we just identied suggest a

set of empirical propositions.

Hypothesis 1. Both an active CEO and decisiondecomposition encourage stability in rm choices, ceterisparibus. A rich vertical ow of information, rm-wideincentives in the presence of an active CEO, and more capa-ble subordinates in the absence of an active CEO promotewide search.

Hypothesis 2. Organizations that couple design ele-ments that foster search with elements that promote stabil-ity will be more successful than those that rely exclusivelyon one set of elements or the other.

Hypothesis 3. Successful organizations in environ-ments with pervasive interactions among decisions will relymore heavily on design elements that promote search thando successful organizations in environments with less per-vasive interactions.

Our ndings both support and extend the extensiveliterature on organizational congurations (e.g., Milesand Snow 1978, Mintzberg 1979, Doty et al. 1993).This line of work contends that only a few internallyconsistent congurations of organizational design ele-ments exist. This may be true in part because ofthe interdependencies generated by the need to bal-ance search and stability. In this paper, we have envi-sioned rms with xed organizational designs thatstruggle to nd effective combinations of operationalchoices. One can imagine, however, that in the long



term, rms are also engaged in a search for goodorganizational designs. That is, they struggle on ahigher order landscape whose horizontal dimensionsare elements of organizational design. Interdependen-cies among organizational design elements may giverise to multiple local peaks on this landscape, witheach peak corresponding to an organizational cong-uration. A natural extension of this paper is to allowmodeled rms to tweak their internal structuresthatis, explore the landscape of organizational elementsand examine whether rms gravitate toward effectiveorganizational designs.Though the interdependencies we identify may

underpin congurations, our results call into ques-tion the recommendation that rms pursue pureconsistent congurations. Some of the congura-tion literature suggests that a rm should be fullygeared toward search or toward stability (e.g., Milesand Snows (1978) prospectors and defenders). Incontrast, our results highlight the need to balanceboth attributes in each organization. The differencebetween our ndings and the prior emphasis on purecongurations may arise for at least two reasons.First, scholars of conguration typically acknowledgethat pure congurations are simply ideal types, use-ful for exposition, yet hybrids may arise in realityand have higher performance because of the needto respond to more than one valid force at the sametime (Mintzberg 1979, p. 474).Second, while we emphasize the need to balance

search and stability, we also stress that contextualvariables affect the nature of that balance. Hypothesis3, for instance, contends that an increase in underly-ing interactions should tilt the balance toward search.Another critical contextual feature that we have pur-posely suppressed in this paper is environmentalchange. A distinguished line of research emphasizesthat organizations design themselves in part to copewith environmental change (Burns and Stalker 1961,Chandler 1962, Lawrence and Lorsch 1967). A valu-able extension of this paper would place our sim-ulated rms on uctuating landscapes and exam-ine what kinds of organizational designs deal wellwith external turbulence. Different types of environ-mental change might necessitate different combina-tions of search and stability. Looking at a cross-section

of contexts, one might very well see a pattern thatconguration scholars would predictthat numer-ous organizational elements associated with searchare adopted in some settings and many elementsassociated with stability are adopted in otherseventhough any single organization requires a balancebetween search and stability, as we contend.More broadly, this paper illustrates a general theme

that has emerged from agent-based simulations oforganizations: connections at one level of analy-sis drive connections at other levels. In this paper,the underlying interactions among a rms deci-sions shape interdependencies among organizationaldesign elements. Interdependencies among designelements set the stage for organizational congura-tions, which, in turn, might mold the interplay amongcompeting rms. Such layered connectivity is one ofthe features that make organizations fascinating yetchallenging to study.

AcknowledgmentsAn earlier version of this paper circulated under the title ChoiceInteraction and Organizational Structure. Special thank to HowardBrenner for computer programming heroics and to Atul Tantiaand Kevin Wang for excellent research assistance. For helpful com-ments, the authors thanks George Baker, Carliss Baldwin, Kath-leen Eisenhardt, Lee Fleming, Daniel Levinthal, Will Mitchell, OlavSorenson, Michael Tushman, Sid Winter, two anonymous referees,an anonymous associate editor, and seminar participants at the Uni-versity of Michigan, Harvard Business School, University of Penn-sylvania, the 2000 Academy of Management meetings, the Novem-ber, 2000, INFORMS conference, and the 2001 EGOS conference. Weare grateful to the Mack Center for Technological Innovation andthe Division of Research of Harvard Business School for generousfunding. Errors remain our own.

ReferencesAbernathy, W. J. 1978. The Productivity Dilemma: Roadblock to Innova-

tion in the Automobile Industry. John Hopkins University Press,Baltimore, MD.

Aghion, P., J. Tirole. 1997. Formal and real authority in organiza-tions. J. Political Econom. 105 129.

Alchian, A. A. 1950. Uncertainty, evolution, and economic theory.J. Political Econom. 58 211221.

Anderson, P., A. Meyer, K. M. Eisenhardt, K. Carley, A. Pettigrew.1999. Applications of complexity theory to organization sci-ence. Organ. Sci. (Special Issue) 10 233236.

Axelrod, R. M., M. D. Cohen. 1999. Harnessing Complexity: Organi-zational Implications of a Scientic Frontier. Free Press, New York.



Baldwin, C. Y., K. B. Clark. 2000. Design Rules: The Power of Modu-larity. MIT Press, Cambridge, MA.

Bower, J. L. 1970. Managing the Resource Allocation Process. HarvardBusiness School Press, Boston, MA.

Brown, S. L., K. M. Eisenhardt. 1997. The art of continuous change:Linking complexity theory and time-paced evolution in relent-lessly shifting organizations. Admin. Sci. Quart. 42 134.

Burns, T., G. M. Stalker. 1961. The Management of Innovation.Tavistock, London, U.K.

Camerer, C., A. Vepsalainen. 1988. The economic efciency of cor-porate culture. Strategic Management J. 9 115126.

Carley, K. M., Z. Lin. 1997. A theoretical study of organizationalperformance under information distortion. Management Sci. 43976997., D. M. Svoboda. 1996. Modeling organizational adaptation asa simulated annealing process. Sociological Methods and Res. 25138168.

Chandler, A. D., Jr. 1962. Strategy and Structure: Chapters in the His-tory of Industrial Enterprise. MIT Press, Cambridge, MA.

Chang, M.-H., J. E. Harrington, Jr. 2000. Centralization vs. decen-tralization in a multi-unit organization: A computationalmodel of a retail chain as a multi-agent adaptive system. Man-agement Sci. 46 14271440.

Child, J. 1984. Organization: A Guide to Problems and Practice, 2nd ed.Harper & Row, New York.

Doty, D. H., W. H. Glick, G. P. Huber. 1993. Fit, equinality, andorganizational effectiveness: A test of two congurational the-ories. Acad. Management J. 36 11961250.

Eppinger, S. D., D. E. Whitney, R. P. Smith, D. A. Gebala. 1994. Amodel-based method for organizing tasks in product develop-ment. Res. Engrg. Design 6 113.

Date post:	08-Jan-2016
Category:	Documents
Upload:	ysumaryan-doni
View:	216 times
Download:	0 times

Balance Ms

Documents