sendil k. ethiraj daniel levinthal · tion or departmentalization of activity and the allocation of...

We employ a computational model of organizationaladaptation to answer three research questions: (1) Howdoes the architecture or structure of complexity affect thefeasibility and usefulness of boundedly rational designefforts? (2) Do efforts to adapt organizational forms com-plicate or complement the effectiveness of first-orderchange efforts? and (3) To what extent does the rate ofenvironmental change nullify the usefulness of designefforts? We employ a computational model of organiza-tional adaptation to examine these questions. Our results,in identifying the boundary conditions around successfuldesign efforts, suggest that the underlying architecture ofcomplexity of organizations, particularly the presence ofhierarchy, is a critical determinant of the feasibility andeffectiveness of design efforts. We also find that designefforts are generally complementary to efforts at localperformance improvement and identify specific contin-gencies that determine the extent of complementarity. Wediscuss the implications of our findings for organizationtheory and design and the literature on modularity inproducts and organizations.•Simon’s (1962) work on the architecture of complexity pro-vides the foundation for viewing organizations, as well asother systems such as products or technologies, as complexadaptive systems (Cohen and Axelrod, 1999). A central ele-ment of Simon’s argument is that the fundamental featuresof complex systems are hierarchy, the fact that some deci-sions or structures provide constraints on lower-level deci-sions or structures, and near-decomposability, the fact thatpatterns of interactions among elements of a system are notdiffuse but will tend to be tightly clustered into nearly isolat-ed subsets of interactions. These dual properties of hierarchyand near-decomposability are argued to enhance the evolv-ability of such systems (Simon, 1962). Although hierarchyand near-decomposability may be desirable attributes of anadaptive entity, how can we presume that boundedly rationalmanagers will be able to identify and uncover some true,latent structure of hierarchy and near-decomposability? Thisis a particularly important question for the growing literatureon the power of modular organizational and product architec-tures (Sanchez and Mahoney, 1996; Baldwin and Clark,2000), which has offered considerable insight into the powerof modular designs but has left largely unaddressed thequestion of the feasibility of boundedly rational actors identi-fying more or less appropriate modular architectures.

The essential tension between designing complex systemsthat are hierarchical and nearly decomposable and the limitsto rationality of human agents is captured in the followingpassage from Simon (1962: 477–478, emphasis added):

The fact, then, that many complex systems have a nearly decom-posable, hierarchic structure is a major facilitating factor enabling usto understand, to describe, and even to see such systems and theirparts. Or perhaps the proposition should be put the other wayround. If there are important systems in the world that are complexwithout being hierarchic, they may to a considerable extent escapeour observation and our understanding. Analysis of their behaviorwould involve such detailed knowledge and calculation of the inter-actions of their elementary parts that it would be beyond our capaci-

© 2004 by Johnson Graduate School,Cornell University.0001-8392/04/4903-0404/$3.00.

•We thank the associate editor, Reed Nel-son, three anonymous referees, and sem-inar participants at Stanford University,Northwestern University, New York Uni-versity, University of Pennsylvania, Uni-versity of Michigan, the University ofMichigan–Santa Fe Institute Workshop,and the Lake Arrowhead Conference onComputational Modeling in the Social Sci-ences for useful comments and sugges-tions. We also thank Linda Johanson forexcellent editorial guidance. Errors andomissions remain our own.

Bounded Rationality andthe Search forOrganizationalArchitecture: An EvolutionaryPerspective on theDesign of Organizationsand Their Evolvability

Sendil K. EthirajUniversity of MichiganDaniel LevinthalUniversity of Pennsylvania

404/Administrative Science Quarterly, 49 (2004): 404–437

#2038-ASQ V49 N3-September 2004—file: 49303-ethiraj

ties of memory or computation. I shall not try to settle which ischicken and which is egg: whether we are able to understand theworld because it is hierarchic, or whether it appears hierarchicbecause those aspects of it which are not elude our understandingand observation. I have already given some reasons for supposingthat the former is at least half the truth—that evolving complexitywould tend to be hierarchic—but it may not be the whole truth.

In the quote above, Simon raises a puzzle about the directionof causality between bounded rationality and complexity.Does bounded rationality allow us to perceive and analyzeonly hierarchical systems, or is it that because complex sys-tems are hierarchical, we are able to observe and analyzethem? The essential question remains: to what extent is thearchitecture of complexity (i.e., hierarchy and decomposabili-ty) the ultimate arbiter of feasible design efforts? This ques-tion is intimately related to a central research agenda in orga-nization theory concerning the design of organization forms.Two contrasting themes in organization theory provide thestarting point for thinking about the feasibility and value ofadaptation in organizational forms themselves. Contingencyarguments implicitly assume that high-level actors in an orga-nization are able to identify and comprehend the demandsimposed by their current environment and are able to designthe appropriate organizational architecture to respond tothose demands. The role of the manager, therefore, is torespond to the changing environment by continuously adapt-ing to the contingencies that confront the organization (Astleyand Van de Ven, 1983). The research in this tradition, howev-er, has devoted little attention to the feasibility and efficacy ofadaptation in organizational forms, focusing instead on the fitbetween environmental contingencies and organizationalforms and the nature of the lower-order adaptation or flexibili-ty that different organizational forms make possible.

Population ecologists (Hannan and Freeman, 1977), in con-trast, argue quite explicitly about the difficulty of identifyingwhat form may seem most apt for a particular environmentor niche and, furthermore, make salient the challenges ofshifting from one form to another. This perspective highlightsthe limits on the degree of strategic choice and the capacityof organization structures to adapt to different niches(Aldrich, 1979). Moreover, even if organizations are able toengage in adaptation of organizational forms, if the rate ofenvironmental change is faster than the rate at which organi-zations can adapt their organizational forms, then such adap-tation efforts are likely to be fruitless (Hannan and Freeman,1984).

The roots of the contrasting positions of the contingency andthe ecological views are embedded in both implicit andexplicit assumptions about individual behavior that underpinsefforts to adapt organization forms and the complexity of thechallenge represented by such efforts. Because contingencytheories of fit between environmental contingencies andorganization design features depend crucially on the ability ofmanagers to discover and achieve such a fit, it is important toexamine whether managers, viewed more realistically as“boundedly rational,” can engage in effective adaptation oforganization forms to achieve such a fit with the environ-ment. At the same time, if the relative futility of managerial

405/ASQ, September 2004

Organizational Architecture


action or the benefits of inertia really depend on the rate ofchange in the environment, then it is also useful to examinethe relationship between the rate of environmental changeand the effectiveness of efforts at adaptation.

A related theme, but treated largely as a distinct line of work,is the examination of first- and second-order change process-es (Argyris and Schön, 1978; Bartunek, 1984; Tushman andRomanelli, 1985; Miner and Mezias, 1996). First-order changeis viewed as incremental, local adaptation within a givenstructure (e.g., changes in pricing policies, product launchesor withdrawals, changes in investments in research anddevelopment or advertising) involving the working out of spe-cific choices within a given organizational structure. In con-trast, second-order change represents change in the underly-ing structure itself (e.g., change from a unitary to M-formstructure). Related to the prior questions of whether changein organizational forms is feasible and useful and how it isinfluenced by environmental change are questions related tothe interrelationship between first- and second-order changeprocesses.

The interrelationship between first-order and second-orderadaptation is far from straightforward. On the one hand,because first-order adaptation is likely to yield diminishingreturns as the space of possibilities within an existing organi-zational architecture is exhausted, a major shift in the organi-zational form (via second-order adaptation) may enhance theeffectiveness of first-order adaptation by creating new config-urations for experimentation (Levinthal, 1997), much asbreakthrough innovations set the stage for subsequentrefinements (Nelson and Winter, 1982). Empirical research onlearning curves (Argote, Beckman, and Epple, 1990) and qual-ity-improvement efforts support this possibility. On the otherhand, aggressive second-order adaptation can undo the learn-ing and first-order adaptations of the past, creating conditionsakin to the liability of newness (Amburgey, Kelly, and Barnett,1993). The ambiguity of predictions suggests that it is usefulto examine the impact of second-order adaptation on the effi-cacy of first-order adaptation and the circumstances underwhich they are complementary or conflicting.

In this paper, we connect Simon the “system designer,” whoexposes the power of hierarchical and nearly decomposablesystems, with Simon the forceful proponent for the notion ofbounded rationality. Uniting Simon’s contrasting ideas aboutbounded rationality and complex system design, we addressthree questions that speak to the literature on organizationdesign and change processes: (1) How does the architectureor structure of complexity affect the feasibility and useful-ness of boundedly rational design efforts? (2) Do efforts toadapt organizational forms complicate or complement theeffectiveness of first-order change efforts? and (3) To whatextent does the rate of environmental change nullify the use-fulness of design efforts? In the process of addressing thesequestions, we hope to delineate at least a skeleton of amicro-foundation for some of the important “macro” ques-tions about the design and evolution of organizational forms.



We examine these questions in the context of a computa-tional model of organizational adaptation. Such a methodolog-ical approach allows us to examine the complex interactionamong search processes at different levels of analysis (thespace of alternative structures and the set of alternativeactions) and, in a controlled manner, to consider how theseprocesses are affected by different environmental settingsand varying degrees of environmental change. As a model,by necessity, it is a stylized representation of actual process-es of organizational adaptation. We build closely on priorwork on models of organizational adaptation (Lant andMezias, 1990; Levinthal, 1997) and specify a process of adap-tive search for organizational form that roughly parallels theseexisting specifications of local search processes. Our charac-terization of an organizational form focuses on the segmenta-tion or departmentalization of activity and the allocation ofspecific functions to a particular organizational subunit(Marengo et al., 2000; Rivkin and Siggelkow, 2003). This isnot to suggest that there are not other important facets of anorganization’s form one might want to consider, including theinformal pattern of interaction among actors or some notionof values or organizational culture. But, as Scott (1998: 26)suggested, a distinctive feature of organizations is their“relatively formalized social structures,” and these featuresof structure on which we do focus constitute the central ele-ments that are manipulated in the process of restructuring(Kelly and Amburgey, 1991).

COMPLEX ORGANIZATIONS AND THEIRENVIRONMENTS

Complex Organizations and the Design ProblemBuilding on the work of Simon (1962: 468) and Perrow(1972), complex organizational systems can be characterizedas consisting of a large number of elements that interact in anon-simple way.1 The complexity that stems from a largenumber of elements interacting in non-trivial ways is twofold.First, the large number of elements creates difficulty in com-prehending the structure that binds them. Second, even ifone is able to uncover and comprehend the structure thatbinds the elements, anticipating the effects of the interac-tions on system behavior and performance is non-trivial(Kauffman, 1993; Cohen and Axelrod, 1999). The complexityincreases as the system gets larger because designers needto first discover which elements interact with which othersand then discover the nature of the interaction relationship(see Perrow, 1999, chap. 3 for a discussion). The followingquote describing the complexity of Xerox’s photocopiers(Adler and Borys, 1996: 68) vividly portrays the difficultyinherent in comprehending and intervening in complexstructures:

During the 1970s, Xerox photocopiers grew vastly more sophisticat-ed in their functionality. As a result, even simple tasks such as copy-ing, loading paper, and resupplying ink became more complex, andrecovery from routine problems such as paper jams became moredifficult. It became increasingly common for users to walk awayfrom the machine rather than waste time trying to work out how toclear a paper jam or replace the ink supply. This resulted in unneces-sary downtime and expensive service calls.

1For instance, predicting system behavioris relatively easy if the interactionsbetween the elements are linear. But ifelements interact such that the relation-ship within and between them is non-lin-ear, i.e., positive over some range andnegative or unrelated over other ranges,then predicting system performancebecomes a difficult problem.




Similarly, Chandler’s (1962: 91) description of Du Pont’s grow-ing pains highlights the challenges of complexity in an organi-zational setting:

The essential difficulty was that diversification greatly increased thedemands on the company’s administrative offices. Now the differ-ent departmental headquarters had to coordinate, appraise, and planpolicies and procedures for plants, or sales offices, or purchasingagents, or technical laboratories in a number of quite different indus-tries. .|.|. Coordination became more complicated because differentproducts called for different types of standards, procedures, andpolicies. .|.|. Appraisal of departments performing in diverse fieldsbecame exceedingly complex. Interdepartmental coordination grewcomparably more troublesome. The manufacturing personnel andthe marketers tended to lose contact with each other and so failedto work out product improvements and modifications to meetchanging demands and competitive developments. .|.|.

The problem of adaptive change in such settings is madeeven more salient, particularly if we assume that managersare boundedly rational, because any adaptive attempt isbased on guesses about the nature of interactions and inter-action relationships among organizational choices and deci-sions.

The design of complex organizations, in its simplest form,invokes two important principles: reductionism, breaking up acomplex whole into simpler units, and the division of labor,grouping tasks or units based on similarity in function. Thetwo principles serve to economize on the cognitive demandsplaced on the designer (Miller, 1956) and also minimizeredundancies in task performance. This idea has persisted inand remained central to the ideas of more recent organizationtheorists, albeit under various labels, such as nearly decom-posable systems (Simon, 1962), loosely coupled systems(Weick, 1976), or pooled, sequential, and reciprocal interde-pendence (Thompson, 1967).

Among the many coordination benefits of specialization andthe division of labor, one of the most important ones from anadaptive standpoint is the potential for relatively autonomousadaptation within the specialized units or departments. Thisallows for localized adaptation within problematic parts of theorganization, while simultaneously buffering the unaffectedparts (Thompson, 1967), or engaging in parallel and simulta-neous adaptive attempts in different departmental units (seeWeick, 1976, for seven adaptive benefits of loose coupling).

A second basic principle for dealing with complexity is hierar-chy. Specialization or the division of labor can help eliminateredundancies and duplication of effort, though the decisionsor activities that are compartmentalized still may need to becoordinated. Among other roles, hierarchy serves the impor-tant function of temporally ordering activities and helps elimi-nate endless cycling in their performance. Simon (1962) sug-gested that complex systems that resemble hierarchies tendto evolve faster and toward a stable, self-reproducing form ascompared with non-hierarchic systems. The property of hier-archy is argued to be found not by chance, but favored byevolutionary selection processes (Simon, 1962).



The notion of hierarchy is often used interchangeably withorganizational fiat (Williamson, 1975) or authority (see Rivkinand Siggelkow, 2003, for an implementation of hierarchy asauthority using a modeling structure similar to ours). Rather,in this paper, the notion of hierarchy is used in the sense ofnested hierarchies (Baum and Singh, 1994) in which there isa precedence ordering of tasks or activities in the organiza-tion. In the context of organization design, the line of hierar-chy denotes the flow of information, or constraints, from theimmediately higher department or unit. Similarly, in an organi-zational context, Simon’s (1957) notion of decision premiseshas the property of a nested set of constraints on organiza-tional decision making. The important function of hierarchy isnot only to resolve conflicts between subsystems (Thomp-son, 1967: 60) but also to facilitate learning through trial anderror by allowing systematic and orderly local search andexploration. Hierarchy enables the recognition of progresstoward one’s goals and the evolution of a system throughseveral intermediate stable configurations (Perrow, 1972:44–52). In contrast, non-hierarchic systems, which display noparticular order in their configuration, make local search andtrial-and-error learning much more difficult to accomplish.

Interestingly, Xerox’s response (Adler and Borys, 1996) to thegrowing complexity of its copiers and that of Du Pont to itsdiversification (Chandler, 1962) was to redesign their productand organization, respectively, conforming to the twin princi-ples of decomposability and hierarchy:

Through its physical structure and the displays it offered, themachine provided a succession of informative views of the copier’sfunctioning and of the user’s interaction with it at various stages ofthe copying experience. As the views unfolded, they helped usersform mental models of the machine’s subsystems and of the experi-ence of interacting with those subsystems. The views includedstep-by-step presentations of machine subsystems, their functions,and the corresponding task sequences. The views supported copy-ing tasks by talking the users through them—neither concealinginformation nor overloading users with incomprehensible or unrelat-ed information. The interiors, for example, were designed toexpress various layers and degrees of interaction to users and ser-vice people. The user-accessible components of the interiors (suchas paper loading, jam clearing, and simple maintenance) wereplaced in the foreground of the visual field, and the technician-accessible components of the interior (for more complex mainte-nance and repairs) were placed in receding layers in the back-ground. (Adler and Borys, 1996: 68–69)

No member of the Executive Committee should have direct individ-ual authority or responsibility which he would have if he was incharge of one or more functional activities of the Company. .|.|. Forexample, our plan provides that one member of the Executive Com-mittee, who may be best fitted by experience for his duty, will coor-dinate the sales function by holding regular meetings with appropri-ate representatives of the five Industrial Departments. .|.|. Accordingto this plan, the head of each Industrial Department will have fullauthority and responsibility for the operation of his industry, subjectonly to the authority of the Executive Committee as a whole. He willhave under him men who will exercise all the line functions neces-sary for a complete industry, including routine and special purchas-ing, manufacture, sales, minor construction. .|.|. (Chandler, 1962:107)




The description of the copier redesign effort and the restruc-turing of Du Pont invoke the twin principles of complex sys-tem architecture advocated by Simon (1962). The design ofthe copier into subsystems and concealing irrelevant informa-tion within modules or subsystems (Baldwin and Clark, 2000)corresponds to near decomposability or the loose coupling ofstructural elements. Similarly, identifying task sequences andseparating the various layers and degrees of interaction isconsistent with the principle of hierarchy. By the same token,the creation of multiple, autonomous divisions in Du Pontcorresponds to the principle of decomposability, i.e., groupingof activities that are strongly interdependent, whereas themembers of the Executive Committee approximate the prin-ciple of hierarchy in achieving coordination when interdepart-mental interdependence was involved. While the flexibilityand coordination benefits of design architectures that arehierarchical and nearly decomposable are widely reported inthe literature (Lawrence and Lorsch, 1967; Adler, Goldoftas,and Levine, 1999; Baldwin and Clark, 2000), how one discov-ers or designs such an architecture for a complex product ororganization is relatively underexplored. The design challengeis compounded when we consider not omniscient designersbut boundedly rational designers engaging in local and imper-fect search in the space of design possibilities. Does thetractability of the design challenge for boundedly rationalagents vary with the underlying architecture of the complexproblem? We explore this complex design challenge in a set-ting in which there is an explicit recognition of the limits tothe adaptive and inferential capacity of the design effort.

Contingent Logic of Alternative Environments

As suggested above, we focus on two aspects of organiza-tion design: (1) how to group organizational functions intotwo or more departments or units and (2) specify the hierar-chy of information flow or task ordering between the depart-ments. As a result, the critical feature of the environmentfrom the perspective of our analysis is the net effect of task,technology, and institutions on the choices of the numberand nature of departments and the information flow betweenthem. Consistent with contingency logic, there should besome ordering among the set of possible organizationalforms based on their fit with a given set of environmentalconditions—different groupings of organizational choices andthe direction of information flow among them will vary intheir efficacy, depending on the myriad environmental influ-ences at work.

Accepting the existence of some inherent contingency logicabout the desirability of alternative forms, if managers makeguesses about structures and observe the consequences oftheir choices, it is reasonable to expect to observe (bounded-ly rational) efforts at adaptive organizational change. Forinstance, if a multinational corporation (MNC) operating in dif-ferent countries all over the world employs a functional struc-ture to coordinate its activities and observes the dysfunction-al effects of the uniformity of policies in different countries,such information constitutes feedback about the inappropri-ate grouping of organizational choices and functions. It is rea-sonable to expect the managers of the MNC to engage in



efforts to adapt the organization structure in the light of suchfeedback. While bounded rationality suggests that they areunlikely to discover the appropriate structure in the firstattempt, it is certainly possible that repeated, small adaptiveattempts will generate progress toward the appropriate struc-ture.

There is, however, a second complication. The adaptive walktoward discovering the appropriate structure of a complexorganization, given its environment, is likely to be relativelyeffective only if the environment is itself stationary. Overtime, the MNC might enter or exit from different technolo-gies, countries, and businesses, making the appropriatestructure defined by the environment a moving target. Whatconstituted an appropriate grouping of organizational func-tions at one time might be inappropriate at another time. Thisbrings us back to the central questions posed in this paper.Can boundedly rational managers of complex organizationsengage in effective design of organizational forms given thatthe environment that defines the appropriate organizationalform is itself changing?

First-order and Second-order Adaptation

The roots of the distinction between first-order and second-order change processes can be traced to Argyris and Schön(1978). First-order adaptation occurs within the parameters ofan existing architecture or design and is geared to improvingperformance and finding durable fixes and solutions to identi-fied problems (Bartunek, 1984). For instance, in the photo-copier example, first-order adaptation would involve incre-mental tweaks to the design, such as making it easier toopen a door to remove paper jams or load paper. Such incre-mental changes are a result of cumulative learning by doingover long periods of time (Nelson and Winter, 1982; Adlerand Clark, 1991). In contrast, second-order adaptationinvolves a change in the existing architecture itself. Thus,fundamental shifts in strategy or structure would fall withinthe ambit of second-order change (Bartunek, 1984). Analo-gously, adaptation of organizational forms or designs is akinto second-order adaptation, whereas incremental adaptationwithin a given organizational form is akin to first-order adapta-tion (Fiol and Lyles, 1985). The obvious downside risk of sec-ond-order adaptation is the obliteration of prior first-orderadaptation efforts (Tushman and Romanelli, 1985). Forinstance, with a complete redesign of the architecture of thecopier, the difficulty of retaining all the incremental first-orderadaptive efforts of the past is quite apparent. Often, realizingthe full benefits of second-order adaptation requires the elim-ination of prior first-order adaptations.

On the upside, second-order adaptation can open up newopportunities for successful first-order adaptation effortsbecause opportunities for first-order adaptation often exhibitdiminishing returns over time. In complex interdependentsystems, first-order adaptation efforts are rarely costless.Improvement in one dimension often comes at the cost ofdeterioration in one or more other dimensions. As more andmore such adaptations are carried out, the available opportu-nities for productive adaptation decline over time (Tushman




and Anderson, 1986). In such cases, a change in the architec-ture or second-order adaptation can create new opportunitiesfor first-order adaptation (Henderson and Clark, 1990).

The preceding discussion hints at a trade-off between first-order and second-order adaptation. But what is unclear is theprecise nature of the trade-off. The empirical literature inorganization theory remains inconclusive. Carroll and Hannan(2000: 371) presented a list of 15 papers in organization theo-ry, with eleven studies finding a positive relationshipbetween organizational change and mortality and eight (somestudies show both findings) studies documenting a negativerelationship. Carroll and Hannan (2000) suggested thatchanges in core aspects of organizations are likely to threatensurvival, while incremental and peripheral changes are likelyto be less disruptive. As a field, however, we need greaterconceptual clarity as to what constitutes core versus periph-eral changes. As an initial step in this direction, we examinethe boundary conditions under which first-order and second-order adaptations are complements and the conditions underwhich they counteract each other.

We seek therefore to investigate three distinct but interrelat-ed questions that constitute the foundational structure oforganization theory: (1) How does the architecture or struc-ture of complexity affect the feasibility and usefulness ofboundedly rational design efforts? (2) To what extent doesthe rate of environmental change nullify the usefulness ofdesign efforts? and (3) How does second-order adaptation—design efforts—affect first-order adaptation—incremental per-formance improvement efforts? The first question speaks tothe fundamental dilemma of causality between boundedrationality and complexity posed by Simon (1962). The nexttwo questions examine how our understanding of environ-mental change and organizational adaptation processes tem-pers the practical usefulness of design efforts. In otherwords, does the primacy of environmental and organizationalchange processes overwhelm the feasibility and/or function-ality of organization design efforts? We set up a formal mod-eling structure to explore the research questions outlinedabove.2

MODEL

The first question was how the architecture of complexityaffects the feasibility and usefulness of boundedly rationaldesign efforts. In addressing this question, we sought to holdconstant the boundedly rational nature of the search process-es and show the relative effectiveness of this search processas the architecture of complexity varies. Toward this end, weset up four alternative states of the world (what we termgenerative structures from here on) that vary in the nature ofthe underlying complexity discussed above—decomposabilityand hierarchy: (1) hierarchical and loosely coupled (figure 1a);(2) non-hierarchical and loosely coupled (figure 1b); (3) hierar-chical and tightly coupled (figure 1c); and (4) non-hierarchicaland tightly coupled (figure 1d).3 In the figures, an alphanu-meric notation represents each decision variable. The alpha-betic portion denotes the department, while the numeric por-tion denotes the respective decision choice. The x’s in each

2In the interest of readability, details of theformal implementation are provided in theAppendix.

3Note that figures 1a–1d bear a resem-blance to two recent published papers(Rivkin and Siggelkow, 2003; Ethiraj andLevinthal, 2004). The similarity is a func-tion of the common modeling apparatusemployed. The research questions exam-ined here, however, share no overlap withthe research questions of either of thosetwo papers.



row-column intersection identify interdependence betweendecision choices. Reading across a row, an x indicates thatthe row variable is affected by the column variable. Con-versely, reading down a column, an x indicates that the col-umn variable affects the row variable. Therefore, x’s posi-tioned symmetrically above and below the principal diagonalrepresent reciprocal interdependence between decisionchoices.

Within each department, each decision choice is tightly cou-pled with other decision choices in the same department,what Thompson (1967) termed reciprocal dependence. Fig-ure 1a depicts a structure that is hierarchical and loosely cou-pled, i.e., departments 2 and 3 have a weakly coupled rela-tionship with the next higher department, denoted by asingle x below the principal diagonal. The presence (or




Figure 1a. Hierarchical and loosely coupled.

a1 a2 a3 a4 b1 b2 b3 b4 c1 c2 c3 c4

a1 x x x

a2 x x x

a3 x x x

a4 x x x

b1 x x x

b2 x x x x

b3 x x x x

b4 x x x

c1 x x x x

c2 x x x

c3 x x x

c4 x x x x

Figure 1b. Non-hierarchical and loosely coupled.

a1 a2 a3 a4 b1 b2 b3 b4 c1 c2 c3 c4

a1 x x x

a2 x x x x

a3 x x x

a4 x x x

b1 x x x

b2 x x x

b3 x x x x x

b4 x x x

c1 x x x

c2 x x x

c3 x x x

c4 x x x x

absence) of hierarchy is identified by the asymmetry (or sym-metry) in between-department interaction. The degree ofloose coupling is a function of the strength (i.e., magnitude)of between-department interactions. If there are no interac-tions between departments, then the organization is fullydecoupled. In figure 1a, decision a2 influences the payoffassociated with decision b3. This also corresponds to sequen-tial or hierarchical interdependence between departments(Thompson, 1967).

Figure 1b denotes a loosely coupled but non-hierarchicalstructure. The structure is still loosely coupled, because theinteraction within departments is stronger than the interac-tion between departments, but the structure is not hierarchi-cal, because there is no precedence ordering of activitiesbetween the departments. Departments a and b are charac-



Figure 1c. Hierarchical and tightly coupled.

a1 a2 a3 a4 b1 b2 b3 b4 c1 c2 c3 c4

a1 x x x

a2 x x x

a3 x x x

a4 x x x

b1 x x x x x x x

b2 x x x x x x x

b3 x x x x x x x

b4 x x x x x x x

c1 x x x x x x x

c2 x x x x x x x

c3 x x x x x x x

c4 x x x x x x x

Figure 1d. Non-hierarchical and tightly coupled.

a1 a2 a3 a4 b1 b2 b3 b4 c1 c2 c3 c4

a1 x x x x x

a2 x x x x x x

a3 x x x x x x

a4 x x x x x

b1 x x x x x x x x x

b2 x x x x x x x

b3 x x x x x x x x x

b4 x x x x x x x

c1 x x x x x x

c2 x x x x x x

c3 x x x x x

c4 x x x x x

terized by reciprocal interdependence (Thompson, 1967) andsymmetry in between-department interactions. Figure 1cdescribes a hierarchical but tightly coupled structure. Theinteraction between departments a and b and departments band c is as strong as the interaction within the respectivedepartments. In this organization, however, hierarchy is pre-served because there is only sequential interdependencebetween departments, i.e., department a affects departmentb, but not vice versa. Finally, figure 1d represents a non-hier-archical and tightly coupled structure. In the four settings, thedegree of coupling in figure 1a and figure 1b (and figures 1cand 1d, respectively) is held constant, as the total number ofinteractions off the principal diagonal is equal.

Each of the four structures, in a stylized manner, representsdifferent contexts that managers encounter. For instance, thehierarchical and loosely coupled structure might characterizethe relationship between the research and development(R&D) and manufacturing departments of a pharmaceuticalcompany. The process of drug discovery, including drugdevelopment and clinical trials usually spans between 3.5 and13 years (Dranove and Meltzer, 1994). Moreover, becausethe Food and Drug Administration (FDA) approval rate fornew drugs is only about 17–20 percent (DiMasi, 2001), it islikely that the R&D process is highly decoupled from themanufacturing process, with the latter emerging as a signifi-cant issue only after the FDA’s approval of the drug. This sug-gests that figure 1a might be representative of the hierarchi-cal and loosely coupled relationship between R&D andmanufacturing in pharmaceutical firms.4 In more process-intensive industries such as chemicals and semiconductors,however, it is likely that manufacturing considerations will betightly coupled with R&D decisions (Ulrich and Eppinger,1999: 20). This leads us to expect that the R&D-manufactur-ing relationship in the semiconductor industry is likely to benon-hierarchical and tightly coupled, as in figure 1d.

Both situations are in contrast to a case in which the relation-ship between R&D and manufacturing is loosely coupled butmutually consultative in nature (i.e., figure 1b), i.e., R&D iter-ates its designs based on input from manufacturing. This islikely to be the case in the automotive industry, though thestrength of the relationship likely depends on the extent towhich manufacturing-cost considerations dominate (Clark andFujimoto, 1991). Lastly, the relationship between R&D andmanufacturing in the biotechnology industry seems represen-tative of the hierarchical and tightly coupled structure. Abiotechnology product is a protein-based drug, in contrast toa pharmaceutical product, which is chemical-based. Protein-based drugs are derived from living organisms, human bloodand plasma, and proteins. As a result, the manufacturingprocess cannot be divorced from product-development R&D.Thus the separation between R&D and manufacturing thatexists in chemical-based drugs is not possible in protein-based drugs (Dove, 2001) and leads us to expect that theR&D-manufacturing relationship in the latter will be hierarchi-cal but tightly coupled, as in figure 1c. The descriptions of thefour contrasting contexts of the R&D-manufacturing relation-ship suggest that the four structures might depict alternative

4In the pharmaceutical example there is atemporal separation between R&D andmanufacturing. The meaning of hierarchyin our models is simply the unidirectionalflow of decision constraints. Such a unidi-rectional flow may be a result of the tem-poral separation of decisions (as in thepharmaceutical example) or simply theordering of decision constraints betweensets of activities at a point in time (seeCusumano and Selby, 1998, for an exam-ple of how Microsoft partitions its deci-sion constraints in organizing its softwaredevelopment activity).




states of the world, each of which might pose differentdesign and coordination challenges.

Addressing the three research questions posed earlierrequires specifying the following: (1) the four generativestructures, (2) boundedly rational second-order adaptation, (3)boundedly rational first-order adaptation, (4) environmentalchange, and (5) selection. The following subsections providean intuitive explanation for the modeled processes. TheAppendix provides a formal description.

Modeling the Generative Structures

We represent an organization as a set of N decision variables,some subset of which is interdependent. For simplicity andwithout loss of generality, each decision variable in our modelis assumed to take on two possible values (0,1). Thus, thespace of possible organizational action consists of 2N possiblesets of behaviors. For instance, if we consider the manufac-turing strategy of a business firm, then a setting of 1 mightrepresent a policy of outsourcing production activity, and asetting of 0 might connote engaging in in-house manufactur-ing. It follows that different settings for the decision variablesof the organization have different performance implications.Choices about production are likely to have interactions withthe investments in information technology, rate of productintroductions, and so on. For instance, it is conceivable thatrapid and highly variable patterns of product introductionsmay enhance the value of outsourcing and of a sophisticatedinformation-technology system that can link retail activity toexternal suppliers. In other words, some combinations ofdecision choices may yield performance improvements whileothers may undermine it (Macduffie, 1995).

The performance of the organization ultimately depends onthe settings (1s or 0s) of the decision variables, but the abilityof the organization to engage in effective first-order adapta-tion and identify a more or less desirable set of choices is, inturn, a function of the organization’s structure, in particular,the set of interactions among the decision choices, asdescribed by figures 1a–1d. When there are no interactionsbetween decision variables, each decision makes an indepen-dent contribution to organizational performance. As the inter-actions between decision variables increase, the contributionof each decision variable to organizational performancebecomes increasingly interdependent. Overall performance isan average of the performance contribution of individual deci-sion variables.

In modeling the four generative structures, there are bothsystematic and stochastic manifestations of each structure.First, the number of departments, D, was specified subjectto the constraint that each department contained an equalnumber of decision choices. This was done so as to reducethe combinatorics of the possible structures we need to con-sider as we vary N and D. In specifying the interaction struc-ture across decision variables, we assume that, as depictedin figures 1a–1d, all decisions within a department interactwith one another. Further, for the loosely coupled structures,we assume that the number of cross-departmental interac-tions is 2(D-1), or, on average, two interactions between each



pair of departments. For the tightly coupled structures, weassume that the number of interactions across departmentsis 2(N/D * N/D), or half the degree of within-department inter-actions. Although the magnitude of the degree of interactionsis specified in this manner, the particular variables that inter-act with one another are chosen randomly.5 Thus, overallinterdependence in figures 1a and 1b (the loosely coupledstructures) is always equal, with hierarchical and non-hierar-chical structures having the same total number of between-department interactions. Similarly, for the two tightly coupledstructures (figures 1c and 1d), the total magnitude ofbetween-department interactions is the same.

Modeling Boundedly Rational Second-order Adaptation

We assume that managers make two organization designchoices: (1) how many departments or units to create and (2)the assignment of functions to departments. We model anevolutionary search process wherein managers engage inboundedly rational adaptive attempts to discover superiorstructures, as defined by the appropriate number of depart-ments and the appropriate mapping of decision variables todepartments in the context of one of the four underlying gen-erative structures. We implement three search operators thatcollectively represent second-order adaptation: (1) splitting,(2) combining, and (3) reallocation.

Splitting may be seen as the breaking up of existing depart-ments into two or more new departments. As departmentsin organizations grow larger, the same piece of information islikely to have to pass through a larger number of potentiallyredundant individuals before resulting in an action or deci-sion. In such cases, splitting an existing department can helpeconomize on information flow (Arrow, 1974). Splitting is alsonecessitated when a department is engaged in a number ofunrelated activities, each of which requires different skills,people, and/or resources. In such settings, splitting facilitatesdifferentiation (Lawrence and Lorsch, 1967) among organiza-tional units to allow specialization to their specific contexts.

Combining is the opposite of splitting, akin to combining orintegrating two or more departments. From the seminal workof Lawrence and Lorsch (1967), we know that organizationalstructures are constantly balancing the contrasting forces ofdifferentiation and integration. While pressures for local adap-tation dictate greater differentiation, the pressures for broad-er efficiencies in organizational performance demand greaterintegration. Empirical evidence shows that organizationsoften cycle between extended periods of increasing decen-tralization (differentiation) and increasing centralization (inte-gration) (Mintzberg, 1979; Cummings, 1995; Nickerson andZenger, 2002), suggesting that the combining operator mightbe an important counterpart to splitting.

In reallocation, organizations transfer or reassign functionsfrom one subunit to another. Such reorganization does notlead to more (as in splitting) or less (as in combining) parti-tioning of tasks but simply involves the reallocation or reas-signment of functions between departments. For instance,the shift from an organization structured along geographiclines to a product structure (Bartlett and Ghoshal, 1989) or

5We did robustness checks (available onrequest) that vary the degree of within-department interactions and the treat-ment of loosely versus tightly coupledstructures. In the loosely coupled struc-ture, the qualitative results hold as longas the intensity of interactions withindepartments is greater than the intensityof interactions between departments.Correspondingly, in tightly coupled struc-tures, the qualitative results hold as longas the intensity of interactions acrossdepartments is greater than or equal tothe intensity of interactions within depart-ments.




from a functional structure to a product structure (Chandler,1962) has no clear implication for the number of subunits butobviously would involve the wholesale reconfiguration oforganizational activity.6 Collectively, these operations of com-bining, splitting, and transfer are the mechanisms that gener-ate change in both the number of departments within theorganization and the assignment of decision variables todepartments, what we consider to be an organization’s archi-tecture.

The inference involved in the three operations is boundedlyrational. Managers employ the operations of splitting, com-bining, or transfer based on their understanding of whetherone or more decision choices belong to their respectivedepartment. The inferential process is imperfect in that onlypairs of departments are compared at a time, and the exami-nation of each department is only local. As a result, the even-tual outcome of the attempts at redesign is not always func-tional from the organizational standpoint, thus renderingsecond-order adaptation imperfect. In addition, observingsuch patterns of influence does not assume that managersunderstand cause-and-effect processes at the departmentlevel. The only behavioral assumption made is that thedesigners are able to observe the effect of their actionsthrough a crude form of root-cause analysis (Macduffie,1997). The formal specification of the splitting, combining,and transfer operators is provided in the Appendix.

Modeling Boundedly Rational First-order Adaptation

First-order adaptation is implemented as follows. In eachperiod of the experiment, the actors in each departmentattempt to enhance the performance of their particulardepartment. Actors are assumed to “see” the performanceof their given department and can anticipate what incremen-tal changes from the existing decision string would imply fordepartment performance. Thus adaptation occurs through aprocess of off-line, local search implemented simultaneouslyin each of the departments (cf. Marengo et al., 2000; Rivkinand Siggelkow, 2003).7 Within each department, a decisionchoice is selected at random, and actors within each depart-ment evaluate the efficacy of flipping the decision choice(0,1) by the criterion of improvement in department perfor-mance. The change is implemented if there is a perceivedincrease in department performance. Because these changeattempts occur in parallel in each of the departments, howev-er, and the departments may have some degree of interde-pendence, there is no presumption that these change effortswill in fact improve organizational performance, or evendepartment performance.

Modeling Environmental Change

Environmental change, according to the contingency view,causes misalignment of organizational structures, choices,and environmental demands. As described above, managersengage in second-order adaptation to align the organizationstructure with the unknown generative structure. A change inthe environment will have at least two effects on organiza-tions. First, environmental change can obviate prior first-order

6In the context of the splitting and combin-ing operators, we considered substantial,discrete points of restructuring. Withrespect to the transfer of activities, weexamined reallocation of individual ele-ments from one subunit to another, orwhat Eisenhardt and Brown (1999)referred to as a form of “patching.” Overtime, such incremental efforts at reorgani-zation can cumulate in broad changes.

7This capability, as suggested by Rivkinand Siggelkow (2003), may be a functionof effective accounting systems that canfacilitate the evaluation of department-level performance. Indeed, activity-basedcosting is widely deployed to track theperformance of organizational subunits(Cooper and Kaplan, 1992).



Figure 1e. Hierarchical and loosely coupled structureafter environmental change.

a1 a2 b4 c2 a3 b1 b3 c1 a4 b2 c3 c4

a1 x x x

a2 x x x

b4 x x x

c2 x x x

a3 x x x

b1 x x x x

b3 x x x x

c1 x x x

a4 x x x x

b2 x x x

c3 x x x

c4 x x x x

adaptations, and in that sense, environmental change can becompetence destroying (Tushman and Anderson, 1986). Sec-ond, environmental change may render less appropriate agiven organizational form (Stinchcombe, 1965). For instance,in a multinational firm, if political change in the host countryin which the firm operates increases the asset expropriationthreat, then the old policies that guide investment andgrowth are unlikely to remain relevant. The organization’smanagers now need to balance the pursuit of growth againstthe threat of expropriation.

To capture the effects of the possible obsolescence of orga-nizational competence, as expressed in the reduced perfor-mance associated with the current set of policy choices, andthe possible misalignment between environmental demandsand the organization structure, we allow the environment tochange every period with some probability, ∆. A stable envi-ronment is specified as ∆ = 0, and the environment changesevery period with certainty when ∆ = 1. For all intermediatevalues of ∆, the environment changes probabilistically.

Specifically, following each environmental change, werespecify the coupling of decision variables within depart-ments and the nature of between-department interactionsthat make up the generative structure. Contrasting figure 1awith figure 1e illustrates the effect of environmental changefor the hierarchical and loosely coupled structure. Twochanges are visible in a comparison of the two figures. First,the composition of the departments is altered in the sensethat decision variables subscripted a, b, and c are no longerclustered together in the same department. Second, thebetween-department interactions are altered as well. Weimplemented environmental change in the three other struc-tures along similar lines.8 Given the newly specified struc-ture, the performance landscape is re-seeded to generate anew mapping between decision variables and performanceoutcomes. The form of environmental change we implement

8We also implemented environmentalchange as also triggering a change in thenumber of departments that representthe generative structures. The resultswere identical and are available from theauthors on request.




is best described as radical in the sense that all prior adapta-tions and learning are destroyed after the environmentalchange, which we recognize may be quite rare. Change ismore often incremental and tends to devalue some prioradaptations while preserving others. We are interested, how-ever, in assessing the usefulness of second-order adaptationin the context in which it is most favorable. If inert structuresare shown to outperform those that undergo second-orderchange in such settings, then clearly inertia with respect toform will prevail in settings with less radical change. We rec-ognize, however, that as one moves along the continuumfrom a stable to a radically changing environment, the valueof second-order change efforts changes commensurately.

Modeling Selection

Organizational selection processes are modeled as being pro-portionate to fitness. The probability that an organization willbe selected equals its performance level divided by the sumof the performance of all organizations in the population atthat time. This is a standard assumption in modeling biologi-cal processes (Wilson and Bossert, 1971) and has been usedin a number of models of organizational selection (see Lantand Mezias, 1990, 1992; Levinthal, 1997).

ANALYSIS

For each run of the model, a generative structure is specifiedas described above. In addition to initializing the performancelandscape, we drew the states (0,1) of the vector of decisionchoices at random at the start of an experiment. Becauseany single run is sensitive to the inherent randomness inboth the initial states of the decision choices and the perfor-mance landscape, we replicated each experiment 100 timeswith different starting seeds for both the specification of theperformance landscape and the starting state of the system.The reported results, unless mentioned otherwise, are aver-aged over the 100 runs to remove the stochastic componentendemic to any single run. Figure 1f illustrates an interactionmatrix for a single organization at the start of a typical experi-ment. As can be seen from the figure, the starting organiza-tion design contains unequal-sized departments with no sys-tematic pattern of interactions among decision choices. Ineach run, we generate 100 organizations in which for eachorganization, the number of departments and the allocationof activities to departments is randomly specified, as are thesettings for the decision choices. Thus at the start of eachrun, there are typically 100 different organizational forms,each of which independently engages in second-orderadaptation.

The Architecture of Complexity and the Effectiveness ofDesign Efforts

The first experiment was designed to answer the question ofhow the architecture of complexity affects the feasibility andusefulness of boundedly rational design efforts (i.e., second-order adaptation). The second-order adaptation challengeentails discovering the set of interactions among the N deci-sion variables and clustering those decision variables thatseem to have strong interactions with each other. We exam-



ined whether the problem of second-order adaptation istractable and whether and how it varies systematically withthe architecture of complexity.

A key performance variable is the number of organizationalforms as the simulation progresses. If second-order adapta-tion is fully successful, then the number of organizationalforms would converge to one, i.e., the generative structurethat represents the correct number of departments and thecorrect mapping of decision choices to the departments. Thisis an important performance metric because, as Simon(1962) pointed out, an important goal of organization designis to reach evolutionary stability. In the absence of anydegree of stability, the efficacy of adaptation efforts isimpaired.

Figure 2 plots the number of organizational forms (averagedover 100 runs) as the simulation progressed in each of thefour generative structures. The figure shows that as long asthe structure is hierarchical, even at extreme levels of tightcoupling (figure 1c), when all decision choices of one depart-ment affect all decision choices of the immediately succeed-ing department (i.e., all decision choices of department 1affect all decision choices of department 2 and so on), theprocess of second-order adaptation is always able to arrive atand stabilize on the corresponding generative structure usedto generate the performance landscape. In contrast, whenwe introduced reciprocal interaction between departments,i.e., the generative structure is non-hierarchical (figures 1band 1d), organizations never manage to reach a stable state.In the non-hierarchical and loosely coupled structure (figure1b), organizations converge on the generative structure mostof the time, but the violation of hierarchy triggers instabilitywith about six different organizational forms continuing tosurvive even at the end of the experiment. In the case of thenon-hierarchical and tightly coupled structure (figure 1d), the




Figure 1f. Typical perceived interaction matrix of decision choices at thestart of the experiment.

a1 a2 a3 a4 b1 b2 b3 b4 c1 c2 c3 c4

a1 x

a2 x x

a3 x x x x

a4 x

b1 x x x x

b2 x

b3 x x

b4 x

c1 x x x x

c2 x x x x

c3 x x x x

c4 x x

violation of both principles (hierarchy and loose coupling)results in the generative structure never being identified. Theinitial diversity of organizational forms continues to be pre-served, suggesting that second-order adaptation is relativelyineffective.

We examined the sensitivity of the results in figure 2 tochanges in the size of the organization (N) and the number ofunderlying departments (D). We found that the results arerobust to changes in both the size of the organization and thenumber of underlying departments.9 We observed an approx-imately linear positive relationship between organization sizeand the time periods to converge on the generativestructure.10

The Impact of Environmental Change on Design Efforts

The previous section showed that managers were able tosuccessfully converge on the underlying generative structurewhen such structures were hierarchical. These results wereobserved in a stable environment, but environments arerarely stable over long periods of time and, as a result, thequestion arises as to whether second-order adaptation effortscontinue to be effective when the environment changes peri-odically. We reran the models presented above, introducingenvironmental change every period with a probability of ∆ =0.05, (implementing environmental change as describedabove).

The results with environmental change are presented in fig-ure 3. Whenever environmental change occurs, the cumula-

9These results are not included due tospace constraints but are available fromthe authors.

10In sharp contrast, Schaefer (1999) foundthat, in general, with a randomly specifiedstructure, the problem of identifying thecorrect modularization of an organizational(or product) design is NP-complete (Gareyand Johnson, 1990).



Figure 2. Second-order adaptation in a stable environment (D = 5 and N = 30; 100 firms, 100 runs).

100

90

80

70

60

50

40

30

20

10

0

No

. of

Org

aniz

atio

nal

Fo

rms

1 6 11 16 21 26 31 36 41 46 51 56 61 66 71 76 81 86 91 96 101

Periods

Hierarchical & loosely coupled

Non-hierarchical & loosely coupled

Hierarchical & tightly coupled

Non-hierarchical & tightly coupled

Figure 3. Second-order adaptation in a changing environment (D = 5, N = 30, = .05; 100 firms, 100 runs).

100

90

80

70

60

50

40

30

20

10

0

No

. of

Org

aniz

atio

nal

Fo

rms

1 6 11 16 21 26 31 36 41 46 51 56 61 66 71 76 81 86 91 96 101

Periods





tive adaptations of the previous periods are rendered ineffec-tive, and the organization designs are no longer aligned withthe changed generative structure. The second-order adapta-tion efforts of the past are effectively reset and the processbegins again. The figure shows that the effectiveness of sec-ond-order adaptation in the presence of environmentalchange is considerably reduced. About 25–35 organizationalforms continue to survive in both the loosely coupled struc-tures (both with and without hierarchy) and the hierarchicaland tightly coupled structure. The non-hierarchical and tightlycoupled structure, as in the stable environment, makes theleast progress in the process of second-order adaptation and,in this case, the initial diversity of organizational forms contin-ues to persist.

The pattern of results suggests that the effectiveness of sec-ond-order adaptation in the presence of environmentalchange is likely to be sensitive to two parameter settings ofthe model. First, the frequency of environmental change iscritical to the effectiveness of second-order adaptation. Asthe rate of environmental change increases, the process ofsecond-order adaptation becomes relatively ineffective. Moresubtly, the size of the organization is also critical to the effec-tiveness of second-order adaptation. As observed in the pre-vious section, there is a linear positive relationship betweenthe size of the organization and the time periods for the sec-ond-order adaptation process to be effective. Thus, if theenvironment changes faster than the time it takes for second-order adaptation to work, we can expect such efforts to be




futile. For instance, we ran a set of models setting N = 60and the probability of environmental change at 0.20 andfound that, on average, the initial population of 100 organiza-tional forms reduced to only about 50, even for hierarchicalgenerative structures. Similarly, as the radicality of environ-mental change declines, the efficacy of second-order adapta-tion efforts correspondingly improves.

The experiment shows that the process of second-orderadaptation is effective when there is a hierarchical prece-dence structure underlying between-department interactionsand the pace of environmental change is moderate. Devia-tions from hierarchy lead to a complete collapse of the effec-tiveness of second-order adaptation for tightly coupled struc-tures. This finding regarding hierarchy formalizes Simon’s(1962) intuition that complex systems that are hierarchicaltend to evolve faster and toward more stable structures.Hierarchy appears to be a necessary and sufficient conditionfor successful second-order adaptation. Nevertheless, as apractical matter, second-order adaptation is reasonably suc-cessful even when the generative structure is not hierarchicalbut is loosely coupled. With a loosely coupled structure, evenwhen hierarchy is violated, the set of organizational formsreduces to a modest number (about 6 in the stable environ-ment and about 30 in a changing environment). In contrast,the search process is less effective when the generativestructure violates both hierarchy and loose coupling.

Apart from the diversity of organizational forms that are pre-sent, there is the important question of the degree of equifi-nality among these forms. Even if efforts at adaptive second-order change prove unsuccessful or lead to fixation on astructure inconsistent with the generative structure, thisneed not imply that such organizations will fail to engage ineffective first-order change.

Interaction between First-order and Second-orderAdaptation

As highlighted earlier, the extant literature on organizationdesign offers ambiguous predictions about the relationshipbetween first-order and second-order adaptation in complexorganizations. The conditions under which they are substi-tutes and/or complements are unclear. The set of experi-ments reported in this section is designed to clarify the inter-action relationship between first-order and second-orderadaptation efforts.

Adaptation in loosely coupled structures. To explore theinteraction between first-order and second-order adaptation,we preserved the process of second-order adaptation asimplemented in experiment 1. In addition, we included theprocess of first-order adaptation. Figure 4 plots the averageperformance results from 100 runs of four models in whichthe generative structure was loosely coupled, both with andwithout hierarchy (i.e., figures 1a–1b) with an underlyingstructure of five departments (N = 30). In two of the settings,there is a simultaneous process of first-order and second-order adaptation. In the other two settings, only first-orderadaptation occurs, and the organization persists in the initialrandom initialization of the department structure.



Comparing the four models in figure 4 suggests that first-order adaptation shares strong complementarities with sec-ond-order adaptation. In settings with both first-order andsecond-order adaptation, performance not only increasesfaster but also asymptotes at a higher level than when thereis first-order adaptation alone. But a surprising and, in somesense, reassuring finding is that the process of first-orderadaptation continues to be quite useful even when the orga-nization design is misaligned with the corresponding genera-tive structure. This property manifests itself in two ways.First, even in the absence of second-order adaptation, inwhich, in almost all circumstances, the organizational struc-tures that provide the context for first-order adaptationefforts are misspecified, the process of local adaptation iseffective, reaching a performance level of 0.614 by the 100thperiod.11 Second, when the generative structure is non-hier-archical, we know from the prior analysis that the process ofsecond-order adaptation results in modest diversity in the setof organizational forms. Despite this, we find no significantperformance differences between the hierarchical and non-hierarchical structures. Thus the process of second-orderadaptation, even in non-hierarchical structures, is generallyeffective in arriving at a small subset of functionally equiva-lent designs in that each of these designs tends to yieldequivalent first-order adaptation benefits, lending some sup-port to the principle of equifinality (Gresov and Drazin, 1997).At least in loosely coupled systems, it appears possible to

11In this setting, the performance asymp-tote is not reached by the 100th period.The process is still moving, though admit-tedly quite slowly, uphill.




Figure 4. Adaptation in loosely coupled structures in stable environments (D = 5, N = 30; 100 runs).

0.68

0.66

0.64

0.62

0.6

0.58

0.56

0.54

0.52

.05

0.48

Per

form

ance

1 6 11 16 21 26 31 36 41 46 51 56 61 66 71 76 81 86 91 96 101

Periods

First-order adaptation in hierarchical structures

First- & second-order adaptation in hierarchicalstructures

First-order adaptation in non-hierarchicalstructuresFirst- & second-order adaptation in non-hierar-chical structures

realize the benefits of parallelism and localized adaptationeven when the organization design is misaligned with envi-ronmental demands. Departures from hierarchy do cause theprocess of first-order adaptation to be less monotonic, how-ever, particularly in the absence of second-order adaptation.12

In the absence of hierarchy, actors engage in local adaptationefforts that turn out to be damaging to organizational perfor-mance as a whole. This is a consequence of the unanticipat-ed and unknown reciprocal interactions between depart-ments.

Adaptation in tightly coupled settings. We engaged in aparallel analysis in which the generative structure was tightlycoupled (figures 1c–1d). Figure 5 indicates the results of first-order adaptation in four tightly coupled structures: hierarchi-cal, with and without second-order adaptation, and non-hier-archical, with and without second-order adaptation. Theresults here partially diverge from those observed in looselycoupled structures. First, the non-monotonicity in organiza-tion performance over time is much greater, because the like-lihood of incorrect first-order adaptation efforts is increasingwith the degree of coupling between departments. Second,the complementarity between first-order and second-orderadaptation is robust even in tightly coupled structures. First-order adaptation in conjunction with second-order adaptationtends to outperform the former alone.

12The degree of non-monotonicity is some-what masked by the fact that the report-ed results are averages over 100 runs.Single runs of the model exhibit a greaterdegree of non-monotonicity.



Figure 5. Adaptation in tightly coupled structures in stable environments (D = 5, N = 0; 100 runs).

0.68

0.66

0.64

0.62

0.6

0.58

0.56

0.54

0.52

.05

0.48

Per

form

ance

1 6 11 16 21 26 31 36 41 46 51 56 61 66 71 76 81 86 91 96 101

Periods

First-order adaptation in hierarchical structures

First- & second-order adaptation in hierarchicalstructures

First-order adaptation in non-hierarchicalstructures

First- & second-order adaptation in non-hierar-chical structures

These results lend robustness to the findings of the first setof experiments. If the condition of hierarchy is met, second-order adaptation is quite effective, and first-order adaptationis then useful in generating performance improvements atthe department level. Interestingly, the performance asymp-tote (0.676) here is not statistically significantly different fromthe performance asymptote (0.684) for loosely coupled struc-tures (see figure 4). This suggests that if the underlyingstructure is hierarchical and designers engage in second-order adaptation, the success of first-order adaptation effortsis not sensitive to the degree of coupling.

The violation of hierarchy, however, tends to be damaging tofirst-order adaptation efforts. The disruptive consequences offirst-order adaptation when hierarchy is violated are some-what mitigated by second-order adaptation. Even though wefound that the initial diversity of organizational forms persistsin the face of second-order adaptation in non-hierarchical andtightly coupled structures, it seems that identifying this sub-set of designs enhances the effectiveness of first-order adap-tation as compared with the random grouping of functionswhen organizations engage in first-order adaptation alone.Thus even when hierarchy is violated, the process of second-order adaptation is an extremely useful design activitybecause it results in more orderly structures that facilitate theprocess of first-order adaptation. This strengthens our initialfindings on equifinality in loosely coupled structures.

The results in the stable environment therefore indicate thatsecond-order adaptation shares strong complementaritieswith first-order adaptation efforts. In the extreme case, whenthe generative structures are neither hierarchical nor looselycoupled, first-order adaptation efforts without second-orderadaptation are largely futile. In general, the violation of hierar-chy is less critical to first-order adaptation efforts than theviolation of loose coupling.

The impact of environmental change on the complemen-tarity of first- and second-order adaptation. We found thatfirst-order and second-order adaptation share strong comple-mentarities in stable environments. Because environmentalchange renders second-order adaptation relatively less effec-tive, we sought to examine whether the observed comple-mentarity of first-order and second-order adaptation is robustin the presence of environmental change by replicating theexperiment including first-order and second-order adaptationwith the addition of environmental change set at ∆ = 0.05. Asexpected, the performance of organizations declined some-what in a regime of modest environmental change. Other-wise the pattern of results for both loosely coupled struc-tures and tightly coupled structures were largely identical tothat in figure 4 and figure 5, respectively, with the comple-mentarities between first-order and second-order adaptationlargely robust to modest levels of environmental change.13

In the first set of experiments that modeled the process ofsecond-order adaptation alone, as the size of the organizationand the probability of environmental change increased,respectively, the marginal value of second-order adaptationdeclined. The simple intuition here was that when the envi-

13These results are not included due tospace constraints, but they are availablefrom the authors on request.




ronment changes faster than the rate at which organizationscan effectively adapt their organization designs, then suchadaptations are likely to be futile (Hannan and Freeman,1984). We examine this possibility here as well.

From an evolutionary standpoint, because selection will tendto favor higher-performing organizations, if second-orderadaptation generates performance gains, then organizationsthat engage in second-order adaptation will have an evolu-tionary edge over organizations that do not engage in second-order adaptation. Conversely, if increases in organization sizeand the frequency of environmental change, respectively,obliterate the value of second-order adaptation, then weshould find that such efforts at second-order change do notprovide any evolutionary edge. In this case, we should findthat organizations that are inertial, i.e., do not engage in sec-ond-order adaptation, should face no differential selectionpressures as compared with organizations that do engage insecond-order adaptation. We investigate this possibility bypopulating the landscape with 100 organizations with a ran-domly specified number and composition of departments andsettings for the decision variables. We set N = 60, D = 5, and∆ = 0.25. All 100 organizations engage in first-order adapta-tion, but 50 randomly selected organizations also engage insecond-order adaptation every period, while the remaining 50are inert. In each period, 100 organizations are selected withreplacement, and we track the number of surviving organiza-tions that are inert with respect to their structure and thenumber of survivors that engage in second-order adaptation.An increase in the population of inert firms would suggestthat inert firms enjoy an evolutionary advantage, whereas asignificant decrease would indicate an evolutionary disadvan-tage.

Figure 6 graphs the number of inert organizations that survivein each period of the simulation in each of the four generativestructures. The results confirm the relative futility of second-order adaptation.14 The population of inert firms continues todrift randomly around 50, suggesting that they face no signifi-cant advantage or disadvantage compared with firms thatengage in second-order adaptation.

At the end of the experiment, there was no statistically sig-nificant difference in the average performance levels of theinert and non-inert organizations in all four structures. Exam-ining the variance in performance over the 50 simulation peri-ods, however, we find that populations of organizations thatinclude second-order adaptation exhibit twice the variance inperformance as compared with organizational populationsthat remain inert with respect to their structure. This resultconfirms the intuition behind Hannan and Freeman’s (1984)argument that adaptation of organization forms reduces relia-bility rather than enhancing performance when the environ-ment changes faster than the pace at which organizationsadapt.

The results including environmental change suggest thatwhen change is episodic and infrequent, then the comple-mentary effects of first-order and second-order adaptation arerobust. But as organizations grow larger, slowing the effec-

14We ran these models for only 50 periodsrather than 100 because there is no addi-tional information conveyed in modelingthe second 50 periods. Also, this analysistook about 200 hours to run on a state-of-the-art PC. Thus, reducing the simulatedperiods was a pragmatic concern as well.



DISCUSSION

The primary objective of this paper was to begin to explore afundamental set of questions about organization design,including the feasibility and value of design efforts (i.e., sec-ond-order adaptation) and the interrelationship betweendesign efforts and incremental adaptation efforts (i.e., first-order adaptation). Simon (1962) argued strongly both thatcomplex systems tend to be hierarchical and loosely coupledand that individual cognition is substantially bounded relativeto the complexity of the task environments that people face(Simon, 1955, 1957). Again, for Simon, these two properties,




tiveness of second-order adaptation, and the rate of environ-mental change increases, then second-order adaptationbecomes largely ineffective. In such settings, the processesof first-order and second-order adaptation cease to be com-plements. Table 1 summarizes these results, the implicationsof which we discuss in the following section.

Figure 6. Adaptation and selection in changing environments (D = 5, N = 60, = 0.25; 100 firms, 100 runs).

60

58

56

54

52

50

48

46

44

42

40

No

. of

Iner

tial

Fir

ms

1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51

Periods





Table 1

Summary of the Results of the Analyses

Second-order First-order Complementarity of first-orderGenerative structure adaptation adaptation and second-order adaptation

Hierarchical and loosely coupled Effective Effective StrongNon-hierarchical and loosely coupled Moderately effective Effective StrongHierarchical and tightly coupled Effective Ineffective StrongNon-hierarchical and tightly coupled Ineffective Ineffective Moderately strong

one of systems and the other of individuals, raise the ques-tion of to what extent the architecture of complexity (i.e.,hierarchy and decomposability) is the ultimate arbiter of feasi-ble and effective boundedly rational design efforts. This con-stitutes a rather fundamental puzzle for modern organizationtheory as well, one closely related to the long-standingdebate in the literature between contingency perspectives onorganizational change that suggest both a high level of plas-ticity in organizations and, at least implicitly, a high level ofcognitive reasoning on the part of managers and ecologicalperspectives that treat organizations as being relatively inertor subject to dysfunctional consequences as a result ofchange efforts. Our work sought to engage these importantquestions in an even-handed manner that recognizes boththe constraints on adaptation efforts and the importance ofthe structure of the firm’s task environment.

In general, we found that the underlying structure of com-plexity is an important determinant of the success of designefforts. In particular, hierarchy was shown to be a necessaryand sufficient condition for the success of design efforts, incontrast to the relatively greater saliency given to the proper-ty of loose coupling in this literature (Simon and Ando, 1961;Simon, 2002). But the finding shows that loose couplingmoderates the violation of hierarchy in terms of facilitatingdesign efforts. This finding on the role of hierarchy and loosecoupling is generally robust to both variations in the size ofthe organization and its complexity (i.e., number of depart-ments and the nature of interactions between them). Theprinciple of hierarchy (i.e., asymmetry in between-departmentinterdependence) facilitates the process by which boundedlyrational search for organizational forms helps designersevolve toward and stabilize on appropriate forms. In contrast,in the absence of hierarchy (i.e., between-department interac-tions are reciprocal), the local search process never ceasessearching for the appropriate assignment of decision choicesto departments. Reciprocal interdependence triggers acycling behavior wherein the reciprocally interdependentdecisions are continually reassigned from one department toanother. The locally rational designers have no way of stop-ping this incessant searching. Thus the appropriate design ofnon-hierarchic structures requires actors to have a sophisti-cated, perhaps implausibly so, global sense of the interde-pendencies. An empirical implication of this finding is thatinstances in which one observes an on-going pattern of orga-nizational restructuring (see Eccles and Nohria, 1992) maystem from reciprocal interdependence among activities. Insuch settings, stability in the organizational form can onlyresult if the organization is willing to accept some degree ofapparent misspecification of the organizational structure.

In terms of the chicken-and-egg dilemma that Simon posed,our analysis suggests that the underlying structure of com-plexity is an important arbiter of the success of humandesign efforts. Structures that are non-hierarchical and tightlycoupled do not easily lend themselves to effective analysisand design efforts, while structures that are hierarchical (or ifnot hierarchical, loosely coupled) are amenable to boundedlyrational design efforts. Even for non-hierarchical and tightly



coupled structures, the product of design efforts is still betterthan random designs. Though this is not observed in thereduction in diversity of organizational forms, the usefulnessof design efforts to first-order adaptation is clear from theresults of experiment 3. Thus we infer that the usefulness ofboundedly rational design efforts is not limited to just hierar-chical and loosely coupled structures. It broadly extends toother structures that violate the two properties. Neverthe-less, the benefits of design efforts differ both qualitativelyand quantitatively across the four different structures. In thisregard, our analysis helps formalize and extend Simon’s intu-itions and provide some boundary conditions around thedirection of causality between the architecture of complexsystems and bounded rationality.15

The second research question examined how environmentalchange hampers the usefulness of design efforts. We foundthat the relative efficacy of adaptive efforts in hierarchicalstructures persists with moderate levels of environmentalchange, but as the rate of environmental change increases ororganizations get larger, the capacity to adapt effectivelyrecedes. Our results thus provide a potential resolution to theambiguous empirical findings in the literature on the effectsof organizational change (see Carroll and Hannan, 2000: 371).When the rate of environmental change is modest, adapta-tion yields survival benefits. In contrast, when the rate ofenvironmental change is high, adaptation does not yield sur-vival benefits. Similarly, larger organizations are slower toadapt, suggesting that they are likely to be more vulnerablewhen environmental change is rapid. Perhaps the divergencein empirical findings is a function of examining different ratesof environmental change or is due to variations in organiza-tion size across studies.

We also examined the conditions under which design efforts(i.e., second-order adaptation) inhibit or facilitate the processof first-order adaptation. The main contingencies we exam-ined were (1) the four states of the world circumscribed bythe two dimensions of complexity, i.e., hierarchy and degreeof coupling, and (2) the frequency of environmental change.We found that first-order and second-order adaptation weregenerally complements, though second-order adaptation byitself was completely ineffective in non-hierarchical and tight-ly coupled structures. The degree of complementarity,though, depends on the nature of the underlying interactionstructure. Whereas the complementarity is non-zero andhighly positive in loosely coupled structures, it is significantlylower when the underlying structure is non-hierarchical andtightly coupled. The reason we observed the complementari-ty is that second-order adaptation, even if unsuccessful, isreasonably effective in identifying the neighborhood of high-performing organizational forms. From the standpoint of first-order adaptation, specifying a design that is in the vicinity ofthe correct design is still significantly better than a randomconfiguration.

An examination of the summary results in table 1 reveals aninteresting paradox. On the one hand, design efforts (second-order adaptation) are highly effective in hierarchical structuresand less useful in non-hierarchical structures. This finding

15In examining Simon’s chicken-and-eggdilemma, we did not vary the extent ofpresumed rationality in search processes.Holding the rationality underlying thesearch processes constant but varyingthe underlying architecture of complexity,however, allowed us to address whetherboundedly rational search processes arelargely ineffective in settings that are non-hierarchical and non-decomposable. Thuswe were able to address the causalityquestion without varying the rationalityunderlying the search processes.




encourages design efforts in the former and cautions againstit in the latter. On the other hand, first-order adaptation byitself is reasonably effective in loosely coupled structures butcompletely ineffective in tightly coupled structures in theabsence of second-order adaptation. From the standpoint ofperformance improvement through first-order adaptation, thesuccess of design efforts in loosely coupled structures is rel-atively less valuable than the imperfect design efforts in non-hierarchical and tightly coupled structures. Thus the results offirst-order and second-order adaptation taken together sug-gest that incremental performance improvement efforts willbenefit from organization design efforts even in non-hierarchi-cal and tightly coupled structures in spite of the relative inef-fectiveness of design efforts in such settings. The contingentquality of the architecture of complexity is starkly visible inthis result.

In a related vein, our results also reaffirm and formalize theintuition behind Simon’s architecture of complexity: the dualproperties of hierarchy and near-decomposability. Whereashierarchy is a necessary and sufficient condition for the suc-cess of design efforts (second-order adaptation), near-decom-posability is a necessary and sufficient condition for the suc-cess of incremental performance improvement efforts(first-order adaptation). If first-order and second-order adapta-tion are both crucial activities of complex organizations, thedual properties of hierarchy and near-decomposability areundeniably central to their architectures, though each plays adistinct role, a distinction that the prior literature had notclearly identified.

Finally, the results of our analysis provide a useful micro-foun-dation for the burgeoning research on modularity (Baldwinand Clark, 2000; Garud, Kumaraswamy, and Langlois, 2001).Modularity is a design principle that advocates designingstructures based on minimizing interdependence betweenmodules and maximizing interdependence within modules.Much of the extant research on modularity has sought tocontrast modular architectures with integrated architecturesand document the benefits of modular designs. Littleresearch has thus far grappled with the issue of whether andhow good modular designs may be achieved in the face ofcomplexity (Ethiraj and Levinthal, 2004). This is an importantquestion, particularly if the benefits of modularity are contin-gent on achieving good modular designs. If such modulararchitectures are unrealizable through boundedly rationaldesign efforts, then the benefits of modularity are moot. Inthis respect, the results reported here are encouraging. Wefound that relatively local and incremental processes can beused to identify useful, if not optimal modules in structuresthat have some inherent hierarchy and decomposability.

Collectively, this is an important set of findings that brings tothe surface the question of how boundedly rational actors areto design organizations that are intended to economize onthe coordination capabilities of these actors. We showed thatorganization design need not be the product of divine designbut may derive from an evolutionary process. At the sametime, the work leaves unanswered many other importantquestions. We have treated the design problem as one of



Adler, P. S., and B. Borys1996 “Two types of bureaucracy:

Enabling and coercive.”Administrative Science Quar-terly, 41: 61–89.

Adler, P. S., and K. B. Clark1991 “Behind the learning curve: A

sketch of the learningprocess.” Management Sci-ence, 37: 251–267.

Adler, P. S., B. Goldoftas, and D. I.Levine1999 “Flexibility versus efficiency?

A case study of modelchangeovers in the Toyotaproduction system.” Organi-zation Science, 10: 43–69.

Aldrich, H.1979 Organizations and Environ-

ments. Englewood Cliffs, NJ:Prentice-Hall.

Amburgey, T. L., D. Kelly, andW. P. Barnett1993 “Resetting the clock: The

dynamics of organizationalchange and failure.” Adminis-trative Science Quarterly, 38:51–73.

Argote, L., S. L. Beckman, and D.Epple1990 “The persistence and transfer

of learning in industrial set-tings.” Management Science,36: 140–154.

Argyris, C., and D. A. Schön1978 Organizational Learning: A

Theory of Action Perspective.Reading, MA: Addison-Wes-ley.

Arrow, K. J.1974 The Limits of Organization.

New York: W. W. Norton.

Astley, W. G., and A. H. Van deVen1983 “Central perspectives and

debates in organization theo-ry.” Administrative ScienceQuarterly, 28: 245–274.

Baldwin, C. Y., and K. B. Clark2000 Design Rules: The Power of

Modularity. Cambridge, MA:MIT Press.

Bartlett, C. A., and S. Ghoshal1989 Managing Across Borders:

The Transnational Solution.New York: McGraw-Hill.

Bartunek, J. M.1984 “Changing interpretive

schemes and organizationalrestructuring: The example ofa religious order.” Administra-tive Science Quarterly, 29:355–372.

Baum, J. A. C., and J. V. Singh1994 “Organizational hierarchies

and evolutionary processes:Some reflections on a theoryof organizational evolution.”In J. A. C. Baum and J. V.Singh (eds.), EvolutionaryDynamics of Organizations,3–20. New York: Oxford Uni-versity Press.

Carroll, G. R., and M. T. Hannan2000 The Demography of Corpora-

tions and Industries. Prince-ton, NJ: Princeton UniversityPress.

Chandler, A. D.1962 Strategy and Structure: Chap-

ters in the History of theAmerican Industrial Enter-prise. Cambridge, MA: MITPress.

Clark, K. B., and T. Fujimoto1991 Product Development Perfor-

mance. Boston: Harvard Busi-ness School Press.

Cohen, M., and R. Axelrod1999 Harnessing Complexity: Orga-

nizational Implications of aScientific Frontier. New York:Free Press.

Cooper, R., and R. S. Kaplan1992 “Activity-based systems:

Measuring the costs ofresource usage.” AccountingHorizons, 6: 1–13.

Cummings, S.1995 “Centralization and decentral-

ization: The never-endingstory of separation andbetrayal.” Scandinavian Jour-nal of Management, 11:103–117.

Cusumano, M. A., and R. W.Selby1998 Microsoft Secrets: How the

World’s Most Powerful Soft-ware Company Creates Tech-nology, Shapes Markets, andManages People. New York:Simon & Schuster.

DiMasi, J. A.2001 “Risks in new drug develop-

ment: Approval success ratesfor investigational drugs.”Clinical Pharmacology andTherapeutics, 69: 297–307.

Dove, A.2001 “Betting on biogenerics.”

Nature Biotechnology, 19:117–120.

Dranove, D., and D. Meltzer1994 “Do important drugs reach

the market sooner.” RANDJournal of Economics, 25:402–423.

Eccles, R. G., and N. Nohria1992 Beyond the Hype: Rediscov-

ering the Essence of Manage-ment. Boston: Harvard Busi-ness School Press.

Eisenhardt, K. M., and S. L.Brown1999 “Patching: Restitching busi-

ness portfolios in dynamicmarkets.” Harvard BusinessReview, 77(3): 72–82.

Ethiraj, S. K., and D. A. Levinthal2004 “Modularity and innovation in

complex systems.” Manage-ment Science, 50: 159–173.

Fiol, M. C., and M. A. Lyles1985 “Organizational learning.”

Academy of ManagementReview, 10: 803–813.

Garey, M. R., and D. S. Johnson1990 Computers and Intractability:

A Guide to the Theory of NP-Completeness. San Francis-co: W.H. Freeman.

Garud, R., A. Kumaraswamy, andR. Langlois2001 Managing in the Modular

Age: Architectures, Networksand Organizations. London:Blackwell.

Goldberg, D. E.1989 Genetic Algorithms in Search,

Optimization and MachineLearning. Reading, MA: Addi-son-Wesley Longman.

Gresov, C., and R. Drazin1997 “Equifinality: Functional

equivalence in organizationdesign.” Academy of Man-agement Review, 22:403–428.




discovering an unknown but latent generative structure. Wehave not addressed the possible endogeneity of the underly-ing interaction structure itself. Nevertheless, we view thecurrent work as an important step forward in considering thesearch for the architecture of complex organizations.

REFERENCES

Hannan, M. T., and J. Freeman1977 “The population ecology of

organizations.” AmericanJournal of Sociology, 82:929–964.

1984 “Structural inertia and organi-zational change.” AmericanSociological Review, 49:149–164.

Henderson, R. M., and K. B. Clark1990 “Architectural innovation: The

reconfiguration of existingproduct technologies and thefailure of established firms.”Administrative Science Quar-terly, 35: 9–30.

Kauffman, S. A.1993 The Origins of Order: Self-

organization and Selection inEvolution. New York: OxfordUniversity Press.

Kelly, D., and T. L. Amburgey1991 “Organizational inertia and

momentum: A dynamicmodel of strategic change.”Academy of ManagementReview, 34: 591–612.

Lant, T. K., and S. J. Mezias1990 “Managing discontinuous

change: A simulation study oforganizational learning andentrepreneurship.” StrategicManagement Journal, 11:147–179.

1992 “An organizational learningmodel of convergence andreorientation.” OrganizationScience, 3: 47–72.

Lawrence, P. R., and J. W. Lorsch1967 Organization and Environ-

ment: Managing Differentia-tion and Integration. Home-wood, IL: Richard D. Irwin.

Levinthal, D. A.1997 “Adaptation on rugged land-

scapes.” Management Sci-ence, 43: 934–950.

Macduffie, J. P.1995 “Human resource bundles

and manufacturing perfor-mance: Organizational logicand flexible production sys-tems in the world auto indus-try.” Industrial and LaborRelations Review, 48:197–221.

1997 “The road to ‘root cause’:Shop floor problem-solving atthree auto assembly plants.”Management Science, 43:479–502.

Marengo, L., G. Dosi, P. Legrenzi,and C. Pasquali2000 “The structure of problem-

solving knowledge and thestructure of organizations.”Industrial and CorporateChange, 9: 757–788.

Miller, G. A.1956 “The magical number seven,

plus or minus two: Some lim-its on our capacity for infor-mation processing.” Psycho-logical Review, 63: 81–97.

Miner, A. S., and S. J. Mezias1996 “Ugly duckling no more:

Pasts and futures of organiza-tional learning research.”Organization Science, 7:88–99.

Mintzberg, H.1979 The Structuring of Organiza-

tions. Englewood Cliffs, NJ:Prentice-Hall.

Nelson, R. R., and S. Winter1982 An Evolutionary Theory of

Economic Change. Cam-bridge, MA: Belknap Press.

Nickerson, J., and T. Zenger2002 “Being efficiently fickle: A

dynamic theory of organiza-tional choice.” OrganizationScience, 13: 547–566.

Perrow, C.1972 Complex Organizations: A

Critical Essay. Glenview, IL:Scott, Foresman.

1999 Normal Accidents. Princeton,NJ: Princeton UniversityPress.

Rivkin, J., and N. Siggelkow2003 “Balancing search and stabili-

ty: Interdependencies amongelements of organizationaldesign.” Management Sci-ence, 49: 290–321.

Sanchez, R., and J. T. Mahoney1996 “Modularity, flexibility, and

knowledge management inproduct and organizationdesign.” Strategic Manage-ment Journal, 17: 63–76.

Schaefer, S.1999 “Product design partitions

with complementary compo-nents.” Journal of EconomicBehavior and Organization,38: 311–330.

Scott, W. R.1998 Organizations: Rational, Natur-

al, and Open Systems, 4thed. Englewood Cliffs, NJ:Prentice-Hall.

Simon, H. A.1955 “A behavioral model of ratio-

nal choice.” Quarterly Journalof Economics, 69: 99–118.

1957 Administrative Behavior: AStudy of Decision-makingProcesses in AdministrativeOrganization, 2d ed. NewYork: Macmillan.

1962 “The architecture of complex-ity.” Proceedings of theAmerican Philosophical Soci-ety, 106: 467–482.

2002 “Near decomposability andthe speed of evolution.”Industrial and CorporateChange, 11: 587–599.

Simon, H. A., and A. Ando1961 “Aggregation of variables in

dynamic systems.” Econo-metrica, 29: 111–138.

Stinchcombe, A. L.1965 “Organizations and social

structure.” In J. G. March(ed.), Handbook of Organiza-tions, 153–193. Chicago:Rand-McNally.

Thompson, J. D.1967 Organizations in Action. New

York: McGraw-Hill.

Tushman, M. L., and P. Anderson1986 “Technological discontinuities

and organizational environ-ments.” Administrative Sci-ence Quarterly, 31: 439–465.

Tushman, M. L., and E. Romanelli1985 “Organizational evolution: A

metamorphosis model of con-vergence and reorientation.”In L. L. Cummings and B. M.Staw (eds.), Research inOrganizational Behavior, 7:171–222. Greenwich, CT: JAIPress.

Ulrich, K. T., and S. D. Eppinger1999 Product Design and Develop-

ment, 2d ed. New York:McGraw-Hill.

Weick, K. E.1976 “Educational organizations as

loosely coupled systems.”Administrative Science Quar-terly, 21: 1–19.

Williamson, O. E.1975 Markets and Hierarchies:

Analysis and Antitrust Implica-tions. New York: Macmillan.

Wilson, E., and W. Bossert1971 A Primer of Population Biolo-

gy. Sunderland, MA: SinauerAssociates Inc.



APPENDIX

Modeling the Generative Structures

Organizations are represented as making a set of N choices or decision vari-ables [a1, a2, a3, … an]. In figure 1a, the contribution of an individual decisionvariable, ai, depends on other decision variables. Thus, decision variable a1depends on decision variables a2, a3, and a4. In contrast, decision variable b3depends on four other decision variables (a2, b2, b3, b4). As a result, decisionvariable a1 can result in 16 possible levels of performance, depending on itsown value (a 0 or 1) and the value of the three other decision variables onwhich it depends, while decision variable b3 can take on 32 possible levels ofperformance, depending on its own value and the value of the four otherdecision variables on which it depends.

The performance contribution (ωi) of each decision choice (ai) is determinedboth by the state (0 or 1) of the ith decision choice and the states of the jother decision choices on which it depends. Thus,

ωi = ωi(ai; ai1, ai

2 … aij)

The value of ωi is treated as an i.i.d. random variable drawn from the uniformdistribution U [0,1] for each configuration of ai and the j other decision choic-es on which it depends. Organization performance Ω is a simple average ofthe ωi over the N decision choices.

Ω = 1N[ N

Σi=1

ωi (ai; a1i, ai

2 , ... aij)]

The results are robust to alternative distributional assumptions as well. Weran the analysis with exponential and log-normal distributions and obtainedresults that are qualitatively similar to those reported here. These results areavailable from the authors.

We also specified the number of departments and their composition (i.e.,the decision variables that would be assigned to a department). For an orga-nization with N decision variables, we created D departments, where the kthdepartment, Dk comprised (N/D) decision variables assigned at random. For agiven value of N, we varied the value of D for robustness. We specified theinterdependence between departments randomly. For instance, in experi-ments in which figure 1a characterized the organizational form, we randomlychose two decision variables from each department Di that affect two ran-domly chosen decision variables in department Dj (for all i < j). Similarly, infigure 1b, we randomly chose one policy each from departments Di and Djthat affect each other. More generally, overall interdependence in figures 1aand 1b (the loosely coupled structures) was always equal and determined bythe formula D[(N/D) * ((N/D) –1)] + 2(D-1), where the first term captures thetotal interdependence within departments and the second term captures thelevel of interdependence between departments. Similar procedures wereemployed in specifying the between-department interactions in figures 1cand 1d. Formally, the total interdependence in figures 1c and 1d was alwaysequal and represented as D[(N/D) * ((N/D) – 1)] + 2*[(N/D) * (N/D)]. The firstterm captures interdependence within departments and the second term theinterdependence between departments.

Second-order Adaptation

The inferential process for second-order adaptation can be formalized as fol-lows. Consider a set of decision variables that are perceived to belong to adepartment, (ai, a¬i) ∈ D, where ai represents a focal decision choice and a¬iare the remaining set of decision choices within the department D and theorganization is defined by a set of departments, D = Dα, D, D, … DK. Theperformance of the department is given by

nD

1 ΣnD

ωi(ai; a¬i),




where nD is the number of decision choices in department D . Now, con-sider a single decision variable, aj ∈ D, that is flipped to aj′ and the resultingperformance of each decision choice in the department is observed. Then letA be defined as a set of all decision choices such that

A = ai : ωi(ai; a¬i¬j, aj) ≠ ωi(ai; a¬i¬j, aj′) for all i ∈ D

The set A identifies the set of all decision choices in D whose performancechanges as a result of flipping decision choice aj. The designers then adopt asimple rule that all decision choices that were unaffected by the search donot belong to the focal department. All such decision choices are then eithertransferred to a randomly chosen different department or split into a sepa-rate department if they constitute a large enough set. The unchanged deci-sion choices in D are transferred into a randomly chosen other existingdepartment if they constitute less than half the total number of decisions inD; otherwise, the unaffected decisions are split into a new departmentDK+1. If A is an empty set, it means that the performances of all remainingdecision choices in D, (a¬j), were unchanged by the flip in aj suggesting thataj does not belong to the department. Thus, aj is transferred to a randomlychosen department DK. More formally, if

A = , then aj ∈ DK, otherwise,

If nA ≤nDγ

2 (¬A ∩ a¬j) ∈ DK+1 otherwise,(¬A ∩ a¬j) ∈ DK

where, nA and nD represent the number of decision choices in A and D,respectively.

In each period, we also consider combining each department with anotherrandomly chosen department. The departments combine if changes in eachdepartment affect the other and remain separate otherwise. For instance,consider two teams representing two departments, Dα with decision choices[a1, a2, a3, … ai], and D with decision choices [b1, b2, b3, … bj]. A randomlychosen decision choice from each department, ai and bj, respectively, areflipped. The departments Dα and D combine to create a new department,D if the performance of department D is affected by the flipping of ai andthe performance of department Dα is affected by the flipping of bj. More for-mally, if

nDα

1 Σn

Dα

ωi (ai; a¬i, b′j) ≠ nDα

1 Σn

Dα

ωi(ai; a¬i, bj)

nD

1 ΣnD

ωj(bj; b¬j, a′i) ≠ nD

1 ΣnD

ωj (bj; b¬j, ai)

Then, D = Dα ∪ D

First-order Adaptation

Formally, consider a decision choice aj ∈ DK is flipped to aj′. Then, if

nDK

1 Σn

DK

ωi (ai; a¬(i,j) , a′j) > nDK

1 ΣnDK

ωi (ai; a¬(i,j) , aj)

then Dk = (a¬j , a′j) otherwise

Dk = (a¬j , aj)



Modeling Environmental Change

For an organization with N decision variables, we retained the D depart-ments that we specified at the start of the simulation, and following everyenvironmental change, we again randomly reassigned the (N/D) decision vari-ables to each of the departments. We also respecified the interdependencebetween departments randomly. For instance, in experiments in which figure1a was the generative structure, in the period when environmental changeoccurred, we again randomly chose two decision variables from each depart-ment Di that affects two randomly chosen decision variables in departmentDj (for all i < j). Thus, following each environmental change, we respecifiedthe coupling of decision variables within departments and the nature ofbetween-department interactions that make up the generative structure.

Modeling Selection

We used the standard roulette wheel algorithm (Goldberg, 1989) for model-ing selection. More formally,

Ωip(si) = S

Σi=1

Ωi

where, p(si), the probability of selecting the ith organization is given by theratio of the performance, Ωi , of the ith organization to the total performanceof all S organizations in the population. The cumulative probability, P(si), isthen computed as,

P(si) =i

Σj=1

p(sj)

A total of S random numbers rs distributed i.i.d. in the interval [0, 1] aredrawn, and the organizations whose cumulative probability spans a randomdraw are selected according to the rule, P(si–1) ≤ rs ≤ P(si).




sendil k. ethiraj daniel levinthal · tion or departmentalization of activity and the allocation of...

Documents