strategy making in novel and complex worlds: the power of ... · decomposed decision problems. in...

Strategic Management JournalStrat. Mgmt. J., 26: 691–712 (2005)

Published online 7 June 2005 in Wiley InterScience (www.interscience.wiley.com). DOI: 10.1002/smj.475

STRATEGY MAKING IN NOVEL AND COMPLEXWORLDS: THE POWER OF ANALOGY

GIOVANNI GAVETTI,1 DANIEL A. LEVINTHAL2* and JAN W. RIVKIN1

1 Harvard Business School, Boston, Massachusetts, U.S.A.2 The Wharton School, Philadelphia, Pennsylvania, U.S.A.

We examine how firms discover effective competitive positions in worlds that are both novel andcomplex. In such settings, neither rational deduction nor local search is likely to lead a firm to asuccessful array of choices. Analogical reasoning, however, may be helpful, allowing managersto transfer useful wisdom from similar settings they have experienced in the past. From a longlist of observable industry characteristics, analogizing managers choose a subset they believedistinguishes similar industries from different ones. Faced with a novel industry, they seek afamiliar industry which matches the novel one along that subset of characteristics. They transferfrom the matching industry high-level policies that guide search in the novel industry.

We embody this conceptualization of analogy in an agent-based simulation model. The modelallows us to examine the impact of managerial and structural characteristics on the effectivenessof analogical reasoning. With respect to managerial characteristics, we find, not surprisingly,that analogical reasoning is especially powerful when managers pay attention to characteristicsthat truly distinguish similar industries from different ones. More surprisingly, we find that themarginal returns to depth of experience diminish rapidly while greater breadth of experiencesteadily improves performance. Both depth and breadth of experience are useful only when oneaccurately understands what distinguishes similar industries from different ones. We also discoverthat following an analogy in too orthodox a manner—strictly constraining search efforts to whatthe analogy suggests—can be dysfunctional. With regard to structural characteristics, we findthat a well-informed analogy is particularly powerful when interactions among decisions crosspolicy boundaries so that the underlying decision problem is not easily decomposed. Overall,the results shed light on a form of managerial reasoning that we believe is prevalent amongpracticing strategists yet is largely absent from scholarly analysis of strategy. Copyright 2005John Wiley & Sons, Ltd.

INTRODUCTION

Strategy making is most critical in times of changeand in unfamiliar environments. It is in such a con-text that the strategy makers of a firm must identifya viable new strategic position or face the potential

Keywords: analogy; cognition; complex systems; strate-gic decision making; modularity∗ Correspondence to: Daniel A. Levinthal, The Wharton School,University of Pennsylvania, 2028 Steinberg-Dietrich Hall,Philadelphia, PA 19104, U.S.A.E-mail: [email protected]

demise of their enterprise. Yet how are managers tomake intelligent choices in these novel contexts?Popular accounts suggest that managers in suchsettings fall back on past learning and their expe-rience in a variety of business settings. Managerswith such experience may be capable of seeingthrough surface features of the current context todeeper truths that underlie it. This paper aims toshed light on the experiential wisdom that enablesstrategy makers to cope with novel environments.We argue that the basis of this sort of experien-tial wisdom lies in analogical reasoning. For past

Copyright 2005 John Wiley & Sons, Ltd. Received 25 March 2003Final revision received 24 December 2004

692 G. Gavetti, D. A. Levinthal and J. W. Rivkin

experience to be of value in a world of novelty,actors must be able to generalize from prior set-tings to the current environment. This process ofmapping from a source context of prior experi-ence to the current, ‘target’ context is preciselywhat constitutes analogical reasoning (Gick andHolyoak, 1980). To shed light on strategy mak-ing in the face of novelty, we tackle two relatedtasks. First, we describe the anatomy of analogi-cal reasoning, discussing in detail the componentsthat make such logic work in the context of busi-ness strategy. Second, we build a simulation modelof firms that use analogies to seek good strate-gic positions in unfamiliar settings. An analysis ofthe model allows us to identify situations in whichanalogical reasoning works well or poorly.

Our focus on analogical reasoning contrasts withthe two dominant perspectives on strategic choicein prior literature. In one corner of the schol-arly ring, the positioning school within strategicmanagement has emphasized the importance ofdeductive reasoning and rational choice in thestrategy-making process. The imagery is that of ageneral and staff in a situation room, coolly survey-ing topographical maps of the business landscape(Ghemawat, 1999), picking out a peak, and direct-ing the troops to take the summit. This image ishard to sustain, however, in the world that posi-tioning scholars have increasingly painted of firms’strategies embodied in intricate arrays of inter-locking choices (Porter, 1996; Ghemawat, 1999;Rivkin, 2000; Siggelkow, 2002). With this recentemphasis on nuanced interactions among strategicchoices, the positioning school has created a bit ofa logical predicament for itself: it is precisely incomplex worlds of highly interactive systems thatdeductive reasoning and rational choice seem leastable to pinpoint effective positions. If the choicesinvolved in a strategy are numerous and eachaffects the pay-offs associated with many others,the computational load created by a deductive pro-cess can quickly outstrip the bounded processingpower (Simon, 1955) of any management team.1

We explore below whether analogical reasoningcan cope not only with novel settings, but alsowith novel settings that are complex.

In the opposite corner of the ring are thosewho argue that firms discover effective positions

1 Indeed, Rivkin (2000) shows that strategic problems canbecome NP-complete—that is, intractable to algorithmic solu-tion—as the degree of interaction among component choicespasses a critical threshold.

through local, boundedly rational search and luck.In any given industry, a broad set of firms begintheir lives with a wide variety of strategic choices.Each firm’s decisions and its routines are grad-ually perturbed in ways that enhance immediateperformance, in a process that behavioral theoristsof the firm (Cyert and March, 1963) and evolu-tionary scholars (Nelson and Winter, 1982) depictas largely automatic, experiential, and emergent(Mintzberg, 1978). A few fortunate firms happenupon highly effective sets of choices and survivean ensuing shakeout. The imagery here depictsblindfolded individuals spread out widely on arugged landscape, each engaged in local hill climb-ing (Levinthal, 1997). The water level rises overtime, and a few lucky survivors discover effec-tive positions not by virtue of deductive power butsimply because they started their search in a for-tunate place and scrambled up the right inclines.This imagery doubtlessly paints an accurate pic-ture of some successes. Pascale’s (1984) well-known description of how Honda discovered itseffective position in the U.S. motorcycle indus-try, for instance, is marked by happenstance andincremental response to unforeseen problems andopportunities.

We hesitate, however, to ascribe all effectivepositions to luck and local search. Human rational-ity may be limited, but intendedly rational actionsurely remains possible (March and Simon, 1958).By simplifying the problems they face, managerscan bring problems within the bounds of their pro-cessing power and possibly come up with effectivesolutions (Simon, 1991). This is a central role ofmanagerial cognition. Cognition can be especiallypowerful when coupled with local search (Gavettiand Levinthal, 2000); cognition can raise the oddsthat a firm begins its search near a high, promisingmountain rather than in the valley beneath a merefoothill, and local search can complete the journeyuphill. We see cognition, then, as a means to copewith complexity.

Novelty, however, remains a challenge. Whatcognitive processes are available to managerswhen they face unfamiliar problems? In novel sit-uations, where deduction is likely difficult, pastlessons from prior settings can be a tremendouslypowerful source of wisdom (Neustadt and May,1986). One process for transferring such wis-dom is analogical reasoning. As Thagard (1996:80) points out, ‘analogies can be computation-ally powerful in situations when conceptual and

Copyright 2005 John Wiley & Sons, Ltd. Strat. Mgmt. J., 26: 691–712 (2005)

The Power of Analogy 693

rule-based knowledge is not available.’ Analogiesallow actors to take the insight developed in onecontext and apply it to a new setting.

In this paper, we examine how managers can useanalogy to tackle the twin challenges of complexityand novelty. The managers we consider discovernew positions neither by reasoning from first prin-ciples of economics nor by undertaking unguidedlocal search. Rather, when faced with a new andcomplex setting, managers identify the features ofthe setting that seem most pertinent, think backthrough their experiences in other settings withsimilar features, and recall the broad policies thatworked well in those settings. These broad policiesthen form the starting point for a local search pro-cess. Analogies to other settings, drawn from director vicarious experience, guide the strategy-makingprocess. The image of the strategist here is that of awizened trail guide who might not know the detailsof the local landscape, but can recognize types ofterrain well enough to give some general guidance.Analogical reasoning gives managerial cognition asignificant hand in strategy making, and it empha-sizes aspects of strategy making like pattern recog-nition, judgment, and even wisdom—aspects that,in our view, are prominent among practicing strate-gists but are understated in the academic literatureon strategy.

To begin to address this gap, we develop ananalytical structure with which we can exploreboth the power and limits of analogical reason-ing in discovering strategic positions. We do soby first describing qualitatively what we view asthe basic features of analogical reasoning in strat-egy making. We next formalize this argument inan agent-based simulation of firms that struggle tofind superior competitive positions in unfamiliarindustries. The simulation clarifies our depictionof analogical reasoning. Results of the simulationidentify the managerial and structural characteris-tics that make analogical reasoning more or lesspowerful. Our analysis highlights the critical roleplayed by a manager’s mental representations, theconditions that make breadth and depth of man-agerial experience valuable, the dangers of holdingtoo closely to the guidance offered by an analogy,and the power of analogy in the face of poorlydecomposed decision problems.

In our concluding discussion, we suggest that theapproach taken here helps to bridge two perspec-tives on strategy making: behavioral approaches

that emphasize limits to cognition and the impor-tance of search processes, and the positioningapproach, with its emphasis on conscious, a pri-ori strategic choice. Each approach encompassesimportant elements of reality. The strategy field, atan aggregate level, has recognized the importanceof pluralism, but we often lack individual analyt-ical frameworks that embrace and meld multipleperspectives. We suggest that our analytical struc-ture has this property and therefore, beyond thevalue of the particular results we offer, we viewthe conceptual framework as a useful contributionthat may facilitate a more integrated analysis ofstrategic decision making.

MAKING STRATEGY BY ANALOGY

Reasoning by analogy is a common form of logicamong business strategists. Facing a novel oppor-tunity or predicament, strategists think back tosome similar situation they have faced or heardabout, and they apply the lessons from that previ-ous experience. Analogies—to the past, to otherfirms or industries, and to other competitive set-tings like sports or war—come up frequently instrategy discussions. Popular management writershail pattern recognition and associated analogicalthinking as the best way for managers to copewith rapid change (e.g., Slywotzky and Morrison,1999). The case method, perhaps the most popularform of management education, is designed in partto give students a rich base from which to drawanalogies. Through the case method, students ‘areled to active consideration of a tremendous num-ber of diverse and related real situations, whichit would take them at least a lifetime of experi-ence to encounter, and they are thus given a basisfor comparison and analysis when they enter upontheir careers of business action’ (Gragg, 1940).

In this section, we lay out the key elementsof analogical reasoning. A few examples illustratehow analogies are commonly used and give us con-crete instances in which we ground the followingconceptual discussion:

• The supermarket has served repeatedly as thebasis for analogical reasoning. Charlie Merrillfashioned Merrill Lynch’s distinctive approachto retail brokerage after the practices he witness-ed as a long-time manager in the supermarketindustry. ‘Although I am supposed to be an



investment banker,’ he confessed, ‘I think I amreally and truly a grocery man at heart’ (Perkins,1999). Likewise, when Charles Lazarus foundedthe toy superstore Toys ‘R’ Us in the 1950s, herelied explicitly on an analogy to supermarketretailing. Indeed, he called his retail outlet theBaby Furniture and Toy Supermarket until thesignage at a new site demanded a shorter name(Hast, 1992). And when Thomas Stemberg, aformer supermarket executive, established theoffice supply superstore Staples, he posed theinitial strategic insight as an analogical question:‘Could we create a Toys ‘R’ Us for office sup-plies?’ (Stemberg, 1996). The basic supermarketformula—exhaustive selection, low prices andmargins, and high volume—has been appliedin a wide range of retail categories.

• Starting in the 1970s, Circuit City thrived byselling consumer electronics in large super-stores. A wide selection of products, profes-sional sales help, and a policy of not hag-gling with customers distinguished the com-pany’s stores. In 1993, Circuit City surprisedinvestors by announcing that it would open Car-Max, a chain of used car outlets. The com-pany argued that the used car industry of the1990s bore a close resemblance to the elec-tronics retailing environment of the 1970s. Theindustry was dominated, for instance, by smallMom & Pop dealers with questionable reputa-tions. The company hoped that its success for-mula from electronics retailing would work wellin an apparently analogous setting.

• In 1997, the Board of the Internet portal Lycosdecided to grow rapidly and become a full-fledged new-media company, in part by meansof acquisition. After making its first acqui-sition—of homepage builder Tripod—Lycos’smanagers faced a decision that they later identi-fied as the pivotal choice in the company’s his-tory: should Lycos maintain Tripod as a separateoperation with its own brand name or integrateTripod more fully into Lycos-branded opera-tions? The management team argued the issuefor weeks, and ultimately an analogy proveddecisive in the debate. Traditional media compa-nies, the managers observed, tended to maintainmultiple divisions with separate brands but tocoordinate back-office operations. Since Lycoswanted to become a large media company,the managers reasoned, the company should dolikewise. Accordingly, the company kept alive

the brands of Tripod and subsequent acquisi-tions, but integrated the back offices (Gavettiand Rivkin, 2004).

• Many factors contributed to Enron’s startlingfailure, but headlong diversification based onloose analogies played an important role. Afterapparent success in trading natural gas and elec-tric power, Enron executives moved rapidly tocreate markets for other goods ranging fromcoal, steel, and pulp and paper to weather deriva-tives and broadband telecom capacity. Ana-logical reasoning seemed to drive the expan-sion. Executives looked for markets with cer-tain characteristics: fragmented demand, rapidchange due to deregulation or technologicalprogress, complex and capital intensive distribu-tion systems, lengthy sales cycles, opaque pric-ing, and mismatches between long-term supplycontracts and short-term fluctuations in customerdemand (Salter, Levesque, and Ciampa, 2002).On the broadband opportunity, for instance,Enron Chairman Kenneth Lay said: ‘[Broad-band]’s going to start off as a very inefficientmarket. It’s going to settle down to a businessmodel that looks very much like our businessmodel on [gas and electricity] wholesale, whichobviously has been very profitable with rapidgrowth’ (Gas Daily, 2000). However, the ana-logical reasoning failed to appreciate important,deeper differences between the market for nat-ural gas and the market for bandwidth. Thebroadband market was based on unproven tech-nology and was dominated by telecom compa-nies that largely resented Enron’s encroachment.The underlying good, bandwidth, did not lenditself to the kinds of standard contracts that madeefficient trading possible in gas and electricity.Perhaps worst, in broadband trading Enron hadto deliver capacity the ‘last mile’ to a customer’ssite, an expensive chllenge that gas wholesalersdid not face. Even according to Enron’s mis-leadingly rosy financial reports, the broadbandventure was disastrous (Salter et al., 2002).

In each of these cases, a management team borrow-ed a broad set of related choices from one industryand applied the system to a new industry that itbelieved to be similar on some crucial dimensions.In some instances (e.g., Merrill Lynch, Lycos), theteam was motivated by a problem that required asolution. In other cases (e.g., Enron), the analogytook the form of a solution seeking a problem. In



the following discussion, we focus on the problem-requiring-a-solution type of analogy, though wesuggest that the solution-seeking-a-problem formof analogical reasoning is important as well, par-ticularly as a source on entrepreneurial activity.

Well beyond the context of business, the useof analogical reasoning has been of long-standinginterest among cognitive scientists (Gick andHolyoak, 1980; Holyoak and Thagard, 1995; Tha-gard, 1996). When reasoning by analogy, an indi-vidual starts with a situation to be handled—thetarget problem. The actor develops a mentalrepresentation of the target problem, a lower-dimensional sketch that, in the actor’s view, cap-tures the salient characteristics of the situation. Shethen uses some computational procedure to scourother settings with which she is familiar, due toeither direct or vicarious experience, and identifiesa setting that displays similar salient characteris-tics. This setting serves as a source of a candidatesolution. The individual then transfers the candi-date solution and applies it to the target problem.Clearly the power of analogy depends on the valid-ity of the similarity mapping between source andtarget contexts as well as the quality of the solutionsuggested by the source context.

The danger of ‘superficial mappings’ betweenthe source and the target is a widely studied phe-nomenon in cognitive psychology. Experimentalwork and field evidence strongly suggest that pooranalogies are typically based on representationsthat capture only superficial features of the prob-lems (Holyoak and Thagard, 1995). Indeed, re-search in this field shows that superficial similaritycan readily induce experimental subjects, evenwell-educated individuals, to adopt poor analo-gies.2 We suspect the same is true of practicingmanagers.

2 In one study (Gilovich, 1981), for example, students ofinternational conflict at Stanford were told of a hypotheticalforeign policy crisis: a small democratic nation was beingthreatened by an aggressive, totalitarian neighbor. Each studentwas asked to play the role of a State Department official and torecommend a course of action. The descriptions of the situationwere manipulated slightly. Some of the students heard versionswith cues that were intended to make them think of the crisesthat preceded World War II. The President at the time, they weretold for example, was ‘from New York, like Franklin Roosevelt,’refugees were fleeing in boxcars, and the briefing was held in‘Winston Churchill Hall.’ Other students heard versions thatmight have reminded them of Vietnam. The President was‘from Texas, the same state as Lyndon Johnson,’ refugees wereescaping in small boats, and the briefing took place in ‘DeanRusk Hall.’ Clearly, there is little reason that the home stateof the President, the vehicles used by refugees, or a briefing

In the context of business strategy, theobservable characteristics of an industry mayconstitute the dimensions of a representation.Three features of any industry, for instance, arethe size of economies of scale, the size of customerswitching costs, and the heterogeneity of customertastes. Suppose a manager in a novel settingopts to represent her target problem along thesedimensions. On the basis of an initial assessmentof the target, the manager judges that the targetindustry is characterized by modest economies ofscale, large switching costs, and diverse customertastes. The manager then engages in a simplecomputational procedure: where has she seenmodest economies of scale, large switching costs,and diverse tastes before, in other industries?The manager reviews her experience and realizesthat, in a specific industry that was similaralong these three dimensions, a particular nicheprovider of high-end, premium products washighly successful. She then transfers this solutionto the target industry, adopting a small-scalemanufacturing policy, a cream-skimming pricingpolicy, a targeted sales policy, and so forth. If thedimensions she chose to focus on are the ones thatbest summarize the true drivers of performance,the firm improves its odds of success.

The difficulty that faces the analogizing manageris that there are innumerable dimensions alongwhich one can form a representation and somedimensions may be misleading. Suppose, forinstance, that a different manager facing thesame target industry ignores economies of scale,switching costs, and customer heterogeneity.Instead he pays attention only to the prevalenceof Internet technology in the industry. He notesthat the target industry relies heavily on theInternet for sales, marketing, and distribution.

room name should influence one’s recommendation. Yet thesurface features caused the two groups to reach very differentconclusions. Students in the first group were significantly morelikely to apply the lessons of World War II—that aggressionmust be met with force—than were students in the second group,which veered toward a hands-off policy inspired by Vietnam.

Not only were the subjects of the experiment lured bysuperficial likeness, but—perhaps more disturbing—they werenot even aware that they had been lured. After making arecommendation, each student was asked, ‘How similar is thesituation to World War II?’ and ‘How similar is the situationto Vietnam?’ The two groups gave identical answers. Small,superficial, irrelevant features led well-educated students to drawtheir analogies from different sources. This caused them torecommend very different candidate solutions. Yet the studentsseemed unaware that they were drawing from different sources.



In his experience with Internet-based industries,successful companies in such settings spendaggressively in order to get big fast. He deploysthis candidate solution in the target industry. Ifthe prevalence of Internet technology does nottruly shape the target landscape, he may findthat his mass-market product is far less appealingto diverse customers than the offerings of nichecompetitors and that his large-scale manufacturingoperations do not lower his costs. By focusing onan irrelevant dimension and ignoring three other,more pertinent characteristics, he has been led toa poor analogy.

Analogical reasoning and strategic positioning

The target problem is the business situation thestrategist hopes to resolve. In our particular setting,the challenge of interest is how to position a firmin an industry that is novel—novel either for themanagers of the firm itself (e.g., the used carbusiness for Circuit City’s managers) or for allmanagers (e.g., the Internet portal industry). Toposition itself, the company must make a vast arrayof detailed choices about how to develop, design,produce, sell, deliver, and service products (Porter,1985). These choices incur costs and generatebuyer value, and therefore they shape the economicsuccess of the firm.

We find it useful to conceive of the target prob-lem in terms of a high-dimensional performancelandscape (Kauffman, 1995; Levinthal, 1997;McKelvey, 1999, and references therein). Eachdetailed choice constitutes a horizontal dimen-sion on this landscape, and the vertical dimensionrecords the economic success associated with eachcombination of choices, thereby creating a surface.The task of management is to discover a com-bination of choices that, together, produce highperformance; in graphical terms, the challenge isto find a high peak on the performance land-scape. Because the decisions may interact with oneanother (that is, the choice made on each may alterthe marginal costs and benefits associated with theothers), the performance landscape may containnumerous local peaks. A local peak is a config-uration of choices from which one cannot improveperformance by altering any single choice, eventhough simultaneous change in many choices mayreposition a firm to higher ground.

In many business settings, detailed choices clus-ter together to form policy domains. Choices about

automation, lot sizes, work flow, factory scale,and so forth may aggregate into a manufacturingpolicy, for example, with combinations of thesedetailed choices taking on labels that summarizethe policies (e.g., mass production, job shop oper-ations, cell manufacturing, or batch production).Obviously, the choice of a particular approachfor higher-order policies typically has an influenceon detailed choices. Three aspects of these pol-icy domains shape the modeling choices we makebelow. First, the interactions within the decisionsthat make up a policy domain may be more intensethan the interactions across domains. This createsa type of interdependence that Simon (1962) haslabeled near-decomposability. In our subsequentanalysis, we consider the effect of decomposabilityon the efficacy of analogy.

Second, the effect of one policy domain onthe efficacy of another may depend not on thedetailed choices made in the first domain, but onthe net effect of those choices. Take, for instance,the example of Circuit City’s original successin consumer electronics. Circuit City’s successin marketing and selling consumer electronicsclearly had an effect on its policy for procuringelectronics. When the marketing and sales policyproduced strong demand among consumers inthe 1970s, the marginal value of large-scaleprocurement increased dramatically. The effect ofmarketing and sales on procurement dependednot on detailed marketing and sales choices suchas pricing, sales tactics, and sales compensation,but only on the total demand that those choicesgenerated in aggregate.

Third, the high-level policies adopted by a com-pany are often more visible to outsiders and per-haps more memorable to insiders than the detaileddecisions that underpin them. In Circuit City’scase, the consumer-friendly marketing and salespolicy that it created in the electronics retailingmarket was highly visible, but the compensationschemes and information technology that enabledthat policy remained more obscure. We assumebelow that analogies provide guidance about high-level policies, not detailed decisions per se.

Quality of analogically based solutions

On the basis of representational similarity, theanalogizing manager chooses a source industryfrom her library of experience. Beyond thechoice of representation, we see three factors that



influence the quality of the guidance providedby an analogy. First, there is the breadth ofthe manager’s experience. The larger the numberof other industries the manager has experienced,personally or vicariously, the greater is thechance that the manager will be familiar withan industry that matches the target well alongthe representational dimensions. It is this kind ofbreadth that strategy consultants often purport tobring to a client.

A second factor that affects the quality of ananalogy’s guidance is the depth of the manager’sexperience in the source industry. If the managerhas spent a great deal of time in the sourceindustry and understands deeply what distinguishesa good position from a poor one in that setting,it is more likely that the analogy will accuratelyguide the manager’s firm to a favorable set ofpolicies. On the other hand, deep experience ina source setting might be a double-edged sword(Levinthal and March, 1993). If a manager couplesdeep experience with a poorly chosen source, theresult can be disastrous: profoundly convinced thatshe knows exactly what to do, the manager maypersist in applying the policies that succeededin a prior, but very different, setting and mayignore negative feedback from the environment.In this sense, Enron’s extensive experience withgas and electricity trading may have impeded itsability to adapt in its broadband venture. Moregenerally, it is useful to consider how seriouslymanagers take their analogies and how closelythey abide by the candidate solutions. In theformal model we develop below, we consider bothorthodox analogizers, who consistently hold to theset of policies recommended by their analogy, andheterodox analogizers, who use the recommendedset only as a starting point for further exploration.

Analogies typically give high-level guidancewhose details must be worked out at the groundlevel of the target industry. This points to a thirdfactor that influences the quality of the guidanceprovided by an analogy: the degree to which ahigh-level policy choice constrains the detailedchoices that comprise it. In some settings, a policychoice may tightly constrain detailed choices,while in others a policy choice may leave widelatitude for what goes on below. Since candidatesolutions are transferred at the level of policies,not detailed choices, an analogy provides sharperguidance in the former setting than in the latter.

When an analogy is chosen well, we expect clearerguidance to produce higher performance.

The subtlety of analogizing and its implicationsfor the establishment of effective competitivepositions suggest that a formal model of theprocess may generate valuable insights. In thefollowing section, we describe such a model,built as an agent-based simulation. The modelincorporates the features we have described here:representations of variable quality; the possibilityof analogies based on superficial similarity; targetand source industries of variable complexity; adistinction between broad policies and detailedchoices; interactions across choices within policiesand interactions across policies; and managerswho differ in the breadth and depth of theirexperience and in the orthodoxy with whichthey follow their candidate solutions. The modelallows us to confirm some of the relativelystraightforward arguments we have made here(e.g., that better representations make analogicalreasoning more effective). It also reveals some lessobvious factors, such as the role of the underlyinginteraction structure among policy choices, thatmake analogical reasoning a more or less powerfultool for discovering effective competitive positionsin novel and complex settings.

A MODEL OF SEARCH IN NOVEL ANDCOMPLEX WORLDS

The simulation model undertakes two basic setsof operations. First, it generates a family ofperformance landscapes—a target landscape anda set of potential sources—in which the modelercan tune the relationship between the sources andthe target. The performance landscape itself canbe tuned with respect to the degree to which firmchoices on these landscapes are decomposable.Second, the model permits firms to search foreffective positions on the target landscape in avariety of ways. Firms may differ, for example,in terms of the mapping between targets andsources that each firm uses, the way in whichattractive candidate solutions are identified on eachsource landscape, and the manner in which eachfirm couples analogical reasoning with incrementalsearch. We discuss the generation of families oflandscapes and the search for effective positions inturn. (Table 1 summarizes the model’s parametersand symbols.)



Table 1. Parameters and symbols

Parameters related to generationof families of landscapes

P The number of high-level policy decisions that each firm faces on atarget or source landscape.

D The number of detailed decisions each firm must make within eachhigh-level policy. P × D is the total number of decisions eachfirm makes. Each decision is binary so there are 2P×D possibleconfigurations of detailed decisions.

X The total number of observable industry characteristics. Eachcharacteristic is binary, so there are 2X industries or landscapes ina family: one target and 2X − 1 potential sources.

Kw The probability that the performance contribution of a focal detaileddecision is affected by the resolution of each of the other D − 1detailed decisions within the focal decision’s policy. The ‘w’ inKw signifies ‘within’.

Kb The probability that the performance contribution of a focal detaileddecision is affected by the resolution of each of the other P − 1high-level policies. ‘b’ signifies ‘between’.

XPROB An X-digit vector of probabilities. The first element is the probabilitythat the first observable characteristic affects the performancecontribution of each detailed decision. Likewise for elements 2, 3,. . ., X.

Parameters related to firm search DEPTH The portion of each source landscape that is evaluated in order togenerate the library of most promising policy configurations. Thehigher this value is, the more likely it is that the library trulyidentifies the best policies for each source.

BREADTH The portion of the library of most promising policy configurationsthat is available to a particular analogizing firm.

〈# # #〉 Numbers within angle brackets refer to the representation of eachanalogizing firm. The first number is the observable characteristicthe firm’s managers deem the most important; the second numberis the second most important; and so forth.

Orthodoxy Each analogizing firm is set to be orthodox or heterodox. Duringincremental search, orthodox firms never violate the policyguidance provided by the analogical source, while heterodox firmsmay ignore that guidance.

Symbols used to describe themodel

d A particular configuration of detailed decisions; a P × D string ofzeros and ones.

di The resolution, zero or one, of a specific detailed decision. i isbetween 1 and P × D.

{# # #} Numbers within braces consistently refer to lists of P × D detaileddecisions. Within a list, vertical lines separate the P policydomains.

[# # #] Numbers within square brackets refer to lists of P high-level policychoices.

(# # #) Numbers within parentheses refer to lists of X observable industrycharacteristics.

Families of performance landscapes

We generate families of landscapes using anadaptation of Kauffman’s (1993) NK model.3

3 Developed in the biological sciences, Kauffman’s modelhas now been used to explore a variety of issues relatedto organizational and technological search. See, for instance,Kauffman (1995), Westhoff, Yarbrough, and Yarbrough (1996),Levinthal (1997), McKelvey (1999), Rivkin (2000), Gavetti andLevinthal (2000), Rivkin (2001), and Rivkin and Siggelkow(2002, 2003). Fleming and Sorenson (2001, 2004) have found

The NK structure highlights the interdependencyamong choices. Settings in which choices aremore interdependent (a higher K value in thestandard structure) result in more rugged, multi-peak performance surfaces because a change inone choice will have repercussions for many ofthe N choices.

that subtle predictions of the NK model concerning patterns intechnological search are borne out in patent data.



In our elaboration of Kauffman’s basicframework, we introduce a hierarchy of choices.Each firm is assumed to face P high-level policydecisions, each of which includes D detailedchoices. A competitive position, then, is defined byP × D detailed choices. For simplicity, we assumethat each choice offers two options. A strategicposition can then be represented as a P × D-digitstring of zeros and ones: d = {d1 d2 . . . dP×D} withdi = 0 or 1 for all i. Therefore, a firm has 2P×D

positions available to it in the target industry.Likewise, in each source industry, there are 2P×D

possible configurations of choices.Policy choices are related to detailed choices by

a simple majority rule. Suppose, for instance, thatD = 3. We say that a particular policy choice is1 if the detailed decisions within the policy areconfigured such that the majority of them are 1:{111}, {110}, {101}, or {011}. It is 0 if most ofthe detailed decisions are 0: {000}, {001}, {010},or {100}. Each configuration of detailed choicesgives a unique configuration of policies, but eachconfiguration of policies is consistent with manyconfigurations of detailed choices.4

In order to explore analogical reasoning andpossible similarity between source and targetcontexts, we also need to characterize and beable to tune the degree to which one businesscontext is like another. We postulate that associatedwith all industries, including the target industryand each potential source industry, is a set ofobservable characteristics. These are the possibledimensions along which representations can bedrawn. In particular, we assume that there are X

observable characteristics, each of which offerstwo options. As a result, there are 2X combinationsof observable characteristics. Given a particularspecification of the X characteristics for the targetindustry, there are 2X − 1 possible combinationsof characteristics for potential source industries.5

4 Suppose, for instance, that P = 3 and D = 3. Then the detailedconfiguration {011|111|001} implies the policy choice [1 1 0].But [1 1 0] is also consistent with {101|011|000}, as well asmany other configurations of detailed choices. For clarity, weplace policy configurations in square brackets and detailedconfigurations in braces. Within detailed configurations, we willseparate policy domains by vertical bars. We require D to be anodd number so that the majority rule provides a unique policychoice for each configuration of detailed decisions.5 There are 2X possible combinations of characteristics; however,one of them signifies the target landscape itself.

Thus each industry has an X-digit ‘tag.’ Recall,for instance, the hypothetical situation we de-scribed earlier in which industries had fourobservable characteristics: the size of economiesof scale (large or small), the size of customerswitching costs (large or small), the heterogeneityof customer tastes (high or low), and theprevalence of Internet technology (high or low).The particular industry our two managers faced,with small economies of scale, high switchingcosts, diverse customers, and prevalent Internettechnology, might be represented as (0 1 1 1). Inchoosing an analogical source, each manager paidattention to only part of this vector; we return tothis consideration below.

Our goal is to create families of landscapesin which (a) we control the pattern of interac-tion among detailed choices within and acrosspolicy domains, (b) pairs of landscapes with sim-ilar observable characteristics—similar tags—aremore alike than are pairs with very different char-acteristics, and (c) some observable characteris-tics have more influence than others on landscapetopography.

To meet this goal, we assume that thecontribution to performance of each detaileddecision on each landscape depends not only onthe resolution of that decision (0 or 1), but also,possibly, on the resolutions of other decisionswithin the focal decision’s policy, the resolutionsof other policies, and the states of observablecharacteristics. The pattern of dependence isinfluenced by three parameters. Kw, a numberbetween 0 and 1, is the probability that eachdetailed choice’s contribution depends on theresolution of each other choice within the focalchoice’s policy domain. Kb, a number between 0and 1, controls interdependence between policydomains. Specifically, it is the probability thatthe focal choice’s contribution depends on theresolution of each other policy. XPROB, a vectorof X numbers each between 0 and 1, controlsthe probability that each observable characteristicinfluences the contribution of the focal choice. Anobservable characteristic with a large component inXPROB has a large impact on landscape topography.

For a given set of values for P , D, X, Kw,Kb, and XPROB, the simulating computer generatesan influence matrix—a matrix that records whichchoices, policies, and observable characteristicsinfluence the contribution of each detailed choice.(See Figure 1 for an example.) This influence



Figure 1. Example of an influence matrix

matrix is then used to generate target and sourcelandscapes. We lay out the details of this procedurein an Appendix, and we describe the upshot of theprocedure in intuitive terms in the following twoparagraphs.

• First consider observable characteristics. Eachof the 2X combinations of characteristics definesa particular business context. In the example weused above, the target context was characterizedby the vector (0 1 1 1). There are 15 possiblesource settings with profiles of characteristicssuch as (0 0 0 1), (0 0 1 0), (0 0 1 1), and soforth. Suppose the influence matrix looks likeFigure 1. The fourth observable characteristicdoes not influence the contribution of anychoice. Therefore, on source landscape (0 1 1 0),the contribution of every decision is the sameas it is on the target landscape (0 1 1 1); thedifference in the fourth characteristic altersnothing. Source landscape (0 1 1 0) is identicalto the target landscape in every way andis therefore an excellent source for analogy.In contrast, the first observable characteristicinfluences the contribution of every choice.On source landscape (1 1 1 1), which differsfrom the target (0 1 1 1) only along the firstcharacteristic, the contribution of every decisionis different from what it is on the target.The difference in the first characteristic shiftseach contribution altogether so that (1 1 1 1)bears no resemblance whatsoever to the

target (0 1 1 1). More generally, the larger isan element of XPROB, the more profoundlydoes a difference in that characteristic alterthe landscape. In choosing a source for ananalogy, it is crucial to pay attention toobservable dimensions with high values ofXPROB. Our procedure for constructing familiesof landscapes ensures that landscapes withsimilar observable characteristics have similarshapes. Moreover, it allows us to control whichcharacteristics strongly delineate groups oflandscapes with similar topographies and whichcharacteristics are weak, or even irrelevant,delineators.

• Next consider the effects of Kw and Kb. WhenKw = Kb = 0, the contribution of each choicedepends only on the resolution of that choiceand the state of observable characteristics. Onany given landscape, i.e., for any particularset of observable characteristics, the decisionproblem facing a firm is easy to address; achange in any choice alters the contribution ofno other choice so a firm can find the optimalconfiguration of choices simply by adjustingeach choice to the resolution, 0 or 1, thatmakes the greater contribution in isolation. Thelandscape is smooth, single peaked, and easy toscale. When Kw is high and Kb remains low, thesystem is nearly decomposable (Simon, 1962).A change in any choice alters the contributionsof other choices within the same policy, but itis possible for the firm to tackle its decision



problem policy by policy. As Kb rises, however,decomposition becomes more difficult because achange that alters a policy now has ramificationsfor choices in other domains. Graphically, wefind it useful to think of Kw and Kb in terms ofplateaus. When Kw and Kb are low, landscapesare smooth. As Kb rises, they develop distinctplateaus, with the edges of plateaus defined bychanges in policies. As Kw rises, the surfaces ofindividual plateaus become internally rugged.

Having generated a family of landscapes—targetand potential sources—whose relations to oneanother are well controlled, the computer pinpointsan attractive candidate solution on each of the(2X − 1) potential sources. Specifically, for agiven source landscape, the computer considerseach of the 2P combinations of policies, surveysa fraction DEPTH of the detailed configurationsconsistent with each policy configuration, assessesthe performance of each detailed configurationit surveys, and remembers the policy thatproduces the best results on average for thedetailed configurations considered. The result isan exhaustive library of the most promising policyconfigurations on each and every potential sourcelandscape. Table 2 shows what the library might

Table 2. Example of a library of source landscapes andpromising policies

Sourcelandscape

Most promisingpolicy

configuration

Average performancewithin most promising

configuration

(0 0 0 0) [1 0 0] 0.65(0 0 0 1) [1 0 0] 0.65(0 0 1 0) [0 1 1] 0.62(0 0 1 1) [0 1 1] 0.62(0 1 0 0) [0 0 0] 0.55(0 1 0 1) [0 0 0] 0.55(0 1 1 0) [0 1 0] 0.74(1 0 0 0) [1 0 0] 0.61(1 0 0 1) [1 0 0] 0.61(1 0 1 0) [1 1 0] 0.65(1 0 1 1) [1 1 0] 0.65(1 1 0 0) [0 1 0] 0.73(1 1 0 1) [0 1 0] 0.73(1 1 1 0) [1 1 0] 0.70(1 1 1 1) [1 1 0] 0.70

Note that in this example, pairs of source landscapes thatdiffer only in the last characteristic, which has no influence onperformance levels, always have the same most promising policyconfiguration and the same average performance within the mostpromising policy configuration (assuming DEPTH = 1).

look like for the example used above. The libraryincludes the information, for instance, that thepolicy [1 0 0] produced the best results on thesource landscape with observable characteristics(0 0 0 1). The promising policy configurationswill later serve as candidate solutions. DEPTH

parameterizes the depth of experience on whichcandidate solutions are based.

Search for effective positions

The stage is now set for a discussion of howfirms search the target landscape. We considertwo basic classes of search strategies: local searchand analogical reasoning. Local searchers relyon simple hill-climbing and make no use ofanalogy. Each is given an initial configuration ofdetailed choices and a corresponding set of policychoices by chance. That is, each is released at arandom location on the target landscape. In eachsubsequent period, each local searcher considersthe P × D alternatives to its current configurationthat involve a change in a single detailed choice.If it spots opportunities for improvement, thelocal searcher pursues one of the opportunities.Otherwise, it has arrived atop a local peak andremains there for the duration of the simulation.

Analogizers, in contrast, base their initialconfigurations on the library that was generatedearlier. Each analogizer has access to a fractionBREADTH of the full library of potential sourcelandscapes. The particular source landscapes towhich an analogizer has access are chosenat random. In the example above with 15possible sources, an analogizer with BREADTH =0.33 might be familiar with landscapes whoseobservable characteristics are (1 1 1 0), (1 1 0 1),(1 0 1 0), (0 1 0 0), and (1 1 0 0). A less experiencedfirm with BREADTH = 0.2 might know only thelandscapes with the observable characteristics(0 1 1 0), (1 0 0 1), and (1 0 1 0).

Each analogizing firm also has an ordered listof the representational dimensions, or observablecharacteristics, that it considers to be important.We consider this list to be the firm’s represen-tation of its problem context. The representation〈3 2 1〉, for instance, implies that the firm consid-ers the third observable characteristic to be mostimportant, the second characteristic to be second-most important, the first characteristic to be third-most important, and the fourth characteristic alto-gether irrelevant. The representation 〈4〉 reflects



a belief that only the fourth characteristic mat-ters. The first manager we described before—whoconsidered scale economies, switching costs, andcustomer heterogeneity—might have representa-tion 〈3 2 1〉 while the second manager—who paidattention only to the prevalence of Internet tech-nology—might have representation 〈4〉.

The analogizer uses its representation to chooseamong source landscapes within its breadth ofexperience. It does so in a lexicographic manner.Take, for instance, a firm familiar with sourcelandscapes (1 1 1 0), (1 1 0 1), (1 0 1 0), (0 1 0 0),and (1 1 0 0). Assume that the firm has therepresentation 〈3 2 1〉, and the target landscape is(0 1 1 1), as in our earlier example. The firm firstlooks for a source landscape that matches thetarget perfectly on characteristics 3, 2, and 1. Itfinds none. It then disregards the characteristicit considers least important, characteristic 1, andlooks for a source landscape that matches thetarget on characteristics 3 and 2. It finds one suchlandscape, (1 1 1 0), and chooses to focus on thatsource.

Having chosen a source, the analogizer transfersthe candidate solution—in this example and inline with Table 2, [1 1 0]—back to the target.Consistent with our discussion in the previoussection, candidate solutions carry only high-levelpolicy guidance. The firm chooses, at random, aset of detailed choices that abide by the policyguidance. In the example, the firm may choose{110|111|010}. This configuration serves as thefirm’s starting point for subsequent search. Infollowing periods, the analogizing firm engages inincremental improvement just as a local searcherwould. If we force the firm to abide by itsanalogy in an orthodox manner, it never considersincremental improvements that violate the initialpolicy guidance. From {110|111|010}, for instance,an orthodox analogizer would never consider{010|111|010} because to do so would switch thefirst policy from 1 to 0. On the other hand, ifwe allow the firm to be heterodox in its use ofanalogy, its incremental improvement efforts arenot constrained in this manner.

Note the crucial role that the firm’s represen-tation plays in this process. If the representationmatches XPROB closely, in the sense that observ-able characteristics with high influence are earlyin the firm’s list, the firm is likely to focus on asource landscape that closely resembles the target.The candidate solution will then guide the firm

to a promising starting point. On the other hand,if the representation causes the firm to pay atten-tion to industry characteristics that are secondaryor irrelevant, the source will typically bear littleresemblance to the target, and the candidate solu-tion will likely provide poor guidance.

RESULTS

Our simulation results identify factors thatmake analogical reasoning especially powerfulas well as relationships among those factors.We start by looking at factors related to theanalogizing management teams: the quality oftheir representations, the breadth and depth oftheir experience, and the orthodoxy with whichthey use the candidate solution. We then turn tostructural characteristics of the industries they face,especially the decomposability of the problemspace and the breadth of each policy domain.Throughout this section, we examine families oflandscapes in which there are four observablecharacteristics (X = 4), three of which truly affectlandscape structure and one of which is irrelevant(i.e., XPROB = (0.5, 0.5, 0.5, 0.0)).

We compare the position-seeking success offive types of firms: a local searcher and fouranalogizers that differ in their representations.The first analogizer, with representation 〈1 2 3〉,correctly surmises that the first three observablecharacteristics affect the choice/payoff mapping.The second, with 〈1〉, heeds only one of the threerelevant characteristics. The third, with 〈4 3 2〉,gives primacy to the irrelevant characteristic,but also heeds some relevant factors. The finalanalogizer, with 〈4〉, heeds only the characteristicthat has no bearing on the landscapes. Amongrepresentations that include a subset of observablecharacteristics, these four cover the spectrum fromthe longest (〈1 2 3〉 and 〈4 3 2〉) to the shortest(〈1〉 and 〈4〉) as well as the range from the mostaccurate given length (〈1 2 3〉 and 〈1〉) to the leastaccurate (〈4 3 2〉 and 〈4〉).6 Results for firms withother representations fall between results for thesefour firms in an intuitive way; the performanceof a firm with representation 〈1 2〉, for instance,lies between that of the firm with 〈1〉 and that of

6 Our premise is that boundedly rational managers cannot attendto all aspects of their environment. For this reason, we do notexamine firms that consider all four observable characteristics.



Figure 2. Firm performance over time

the firm with 〈1 2 3〉. Results for firms with otherrepresentations are available from the authors onrequest.

We report firm performance as a portion ofthe highest performance attainable on the targetlandscapes that were explored. By this measure,a type of firm that always attains the globalmaximum will achieve a performance of 1.00.Unless otherwise noted, each result is an averageover 1,000 families of landscapes, each exploredby 10 firms of each type. Each of the 1,000 familiesof landscapes shares the same structural parametersbut constitutes an independent draw from the sameunderlying distribution.

Management characteristics

We first examine a situation in which P = 3and D = 3. To highlight the effect of choosingthe correct observable characteristics on whichto base one’s representation, we choose a setof baseline parameter settings in which cognitionis most powerful. DEPTH = BREADTH = 1 for allfirms, and all analogizing firms use candidatesolutions in an orthodox manner. In addition, weset the probability of interaction across policychoices, Kb, equal to one and the probability ofinteraction among decisions within a policy, Kw,to zero.7 For the sake of comparability, we model

7 The results we obtain from this first set of analyses are robust todifferent specifications for the interaction structure. We explore

a local searcher that abides by its initial, randomlyassigned policy choices in an orthodox manner.

Figure 2 shows the average performance of eachtype of firm over the course of the simulation.On average, firms with better representationsachieve higher performance in the initial periodthan do firms with worse representations becausetheir initial policy choices are guided by asource that more closely resembles the target.Within the constraints of these initial policychoices, the sets of detailed decision choicesare randomly specified in the initial period. Asa result, there is considerable room for eachfirm to adapt and improve its performance oversubsequent time periods, even while it maintainsits set of overarching policies. Higher-qualityrepresentations place firms in more opportunelocales, as measured by their initial superior fitnessas well as by their higher asymptotic value.While all firms search the landscape to identifymore favorable sets of decisions, firms with betterrepresentations carry out this search in morefavorable regions of the landscape.

An intriguing aspect of Figure 2 is that thefirm with representation 〈4〉, the firm that heedsonly an irrelevant dimension, fares substantiallybetter than the local searcher. Even a ‘false’

the impact of the interaction structure in the next set of analyses.We choose this particular combination of structural parametersbecause it yields the greatest power of cognition. Below weexplore how the power of cognition depends on the structure ofthe decision problem.



analogy has power. This result arises because, inchoosing a source landscape on the basis of anirrelevant characteristic, the firm may, by chance,focus on a source landscape that happens to sharerelevant characteristics with the target as well.Suppose, for instance, that the target landscapehas characteristics (0 1 1 1), as in the example inthe previous section. The firm with representation〈4〉 might choose as its source a landscape suchas (1 1 1 1) or (0 0 1 1) or (0 1 0 1)—a landscapethat does overlap with the target on some relevantdimensions. Based on this ‘false’ analogy, thefirm identifies a set of choices coherent with thisrepresentation which, purely by chance, may proveuseful in the target problem. In the discussionsection below, we return to the virtues of falseanalogies.

We turn next to the effects of experienceon the power of analogy. Table 3 shows thelong-run differential between the performanceof each analogizer and the performance of thelocal searcher for various levels of BREADTH, theportion of the source library to which the firmhas access. Table 4 presents similar results forDEPTH, the portion of each policy domain thatis explored to generate the library of candidate

solutions. There are at least two striking featuresof these results. First, while greater BREADTH

of experience steadily improves performance, themarginal returns to DEPTH of experience diminishrapidly; a management team with very limitedexperience on each source landscape fares nearlyas well on the target as does a team with extensiveexperience.8 This result suggests that it is far moreimportant to identify an analogical source thatshares the structure of the target problem thanto pinpoint the absolute best solution within thesource problem context. Second, greater breadthand depth of experience boost performance morewhen a firm has a high-quality representation.Put differently, broader and deeper experiencedoes not improve a firm’s lot if it draws fromthat experience on the basis of less relevantcharacteristics.

Tables 3 and 4 obscure a hazard of experiencewe discussed above: broad and especially deepexperience may motivate a management team toabide by its candidate solution even though thesubsequent local search process identifies superior

8 In each row, the results for firms with DEPTH = 0.25, DEPTH =0.50, and DEPTH = 1.00 are statistically indistinguishable.

Table 3. Effect of breadth of experience on analogizer’s long-run performance advantage

Long-run performanceadvantage over a local

BREADTH =searcher for an analogizerwith representation . . . 0.10 0.25 0.50 0.75 1.00

〈1 2 3〉 0.069 (0.002) 0.086 (0.002) 0.111 (0.001) 0.132 (0.001) 0.146 (0.001)〈1〉 0.065 (0.002) 0.074 (0.002) 0.081 (0.002) 0.086 (0.002) 0.090 (0.002)〈4 3 2〉 0.055 (0.002) 0.055 (0.002) 0.066 (0.002) 0.069 (0.002) 0.073 (0.002)〈4〉 0.054 (0.002) 0.053 (0.002) 0.056 (0.002) 0.057 (0.002) 0.059 (0.002)

Kb = 1, Kw = 0, DEPTH = 1, all firms are orthodox. Each figure is an average over 10,000 firms on 1,000 families of landscapes.Each number in parentheses is the standard deviation of the average figure to the left of it.

Table 4. Effect of depth of experience on analogizer’s long-run performance advantage

Long-run performance advantage DEPTH =over a local searcher for ananalogizer with representation . . . 0.01 0.10 0.25 0.50 1.00

〈1 2 3〉 0.119 (0.001) 0.141 (0.001) 0.146 (0.001) 0.146 (0.001) 0.146 (0.001)〈1〉 0.083 (0.001) 0.090 (0.001) 0.093 (0.001) 0.093 (0.001) 0.092 (0.001)〈4 3 2〉 0.060 (0.001) 0.072 (0.001) 0.072 (0.001) 0.071 (0.001) 0.072 (0.001)〈4〉 0.054 (0.001) 0.059 (0.001) 0.062 (0.001) 0.061 (0.001) 0.060 (0.001)

Kb = 1, Kw = 0, BREADTH = 1, all firms are orthodox. Each figure is an average over 20,000 firms on 2,000 families of landscapes.Each number in parentheses is the standard deviation of the average figure to the left of it.



Figure 3. Performance differential due to orthodoxy

solutions. To size up this danger, we repeat theexperiment shown in Figure 2, but now allowfirms to be heterodox; that is, when improvingincrementally, firms are permitted to make changesin detailed decisions that alter high-level policies.We then plot in Figure 3 the difference betweenthe performance of the orthodox firm and theperformance of the corresponding heterodox firm.One can think of this difference as the hidden costa firm bears if it takes its candidate solution tooseriously. The figure shows that this cost dependssensitively on the quality of a firm’s representation.Firms with very good representations incur nocosts of orthodoxy; they are in such promisingparts of the target landscapes that they can performwell without crossing policy borders. Heterodoxfirms have little incentive to cross policy borders,so the search behaviors of orthodox and heterodoxfirms are effectively the same. Firms with poorerrepresentations, in contrast, pay a heavy price forholding on too firmly to their analogies. Indeed,in results not reported here, we find that firmsthat take poor analogies seriously typically fareworse than heterodox local searchers; the penaltyof orthodoxy outweighs the benefit of analogicalguidance.

Structural characteristics

The power of analogy depends not only on therepresentations, experience, and orthodoxy of themanagement team, but also on the structure of

the target landscape. We explore the effect ofboth ‘vertical’ and ‘horizontal’ structural changeson our results. By vertical structure, we meanthe relationship between the number of detaileddecisions per policy (D) relative to the number ofhigher-order policies (P ). By horizontal structure,we mean the intensity of interaction within eachpolicy domain (Kw) and the intensity of interactionacross domains (Kb).

To explore vertical structure, we compare theperformance of analogizing firms to that of localsearchers in two settings: P = 5, D = 3 vs. P = 3,D = 5. The total number of detailed decisionsis 15 in both cases, but the high-level guidancegiven by analogy is more thorough when policydomains are numerous and narrow (P = 5, D =3). Analogical reasoning is more powerful inthis situation, as the results in Table 5 show.Comparing across rows in Table 5, we see thatthe differential created by more thorough guidancediminishes as the quality of the representationdeteriorates. Detailed guidance is less valuable ifit is based on misleading keys.

To explore horizontal structure, we first examinethe effects of Kb and Kw on landscape topographyand then turn to their effects on the performanceof analogizing firms. One can think of eachlandscape as consisting of 2P policy domains,with 2D configurations of detailed decisions withineach domain. Tables 6a and 6b reveal typicaltopography within and across policy domains, asa function of Kb and Kw. To construct the tables,



Table 5. Effect of policy breadth on analogizer’s long-run performance advantage

Long-run performanceadvantage over a localsearcher for an analogizerwith representation . . .

P = 5D = 3

P = 3D = 5

Differential

〈1 2 3〉 0.162 0.109 0.053∗

(0.002) (0.002)〈1〉 0.102 0.073 0.029∗

(0.002) (0.002)〈4 3 2〉 0.085 0.054 0.031∗

(0.002) (0.002)〈4〉 0.066 0.048 0.018∗

(0.002) (0.002)

Kb = 1, Kw = 0, BREADTH = DEPTH = 1; all firms are orthodox.Each figure is an average over 3000 firms on 300 families oflandscapes. Each number in parentheses is the standard deviationof the average figure above it. ∗ Indicates that a differential isstatistically significant at the 0.1% level.

Table 6a. Absolute change in performance associatedwith a within-policy change in a detailed decision

Kw = 0 Kw = 0.5 Kw = 1

Kb = 0 0.056 0.074 0.087Kb = 0.5 0.051 0.069 0.085Kb = 1 0.050 0.068 0.083

P = 3, D = 3; each figure an average over 40 firms walking atrandom for 125 periods on 100 landscapes.

Table 6b. Standard deviation across average per-formance in each policy domain for a typical landscape

Kw = 0 Kw = 0.5 Kw = 1

Kb = 0 0.069 0.060 0.055Kb = 0.5 0.092 0.081 0.061Kb = 1 0.110 0.088 0.071

P = 3, D = 3; each figure an average over 40 firms walking atrandom for 125 periods on 100 landscapes.

we release firms in each policy domain on a largenumber of landscapes, let each walk at randomwithin a domain, and record performance duringthe random walk. Table 6a shows the averageabsolute value of the performance change causedby altering one detailed decision within a policydomain; this reflects within-policy ruggedness.Table 6b reports the standard deviation across theaverage level of performance within each domain

for the typical landscape; this captures across-policy ruggedness. Within-policy ruggedness isdriven upward by increases in Kw, while across-policy ruggedness is boosted by Kb and mitigatedby Kw.

Together, Tables 6a and 6b paint a clear pictureof landscape topography. When Kb = Kw = 0,there are no interactions among decisions and thelandscape is relatively smooth, with changes inindividual decisions having relatively little impacton overall performance. As Kb rises with Kw still0, the surface becomes more rugged. In particular,plateaus defined by policy configurations arise.Within a policy configuration, the surface isrelatively smooth; a change in a detailed choicethat does not change a policy alters only thecontribution of that decision in isolation. Incontrast, a change in a detailed choice that doesalter a policy—a step over the edge of theplateau—causes many contributions to change.The result is a surface with internally smoothplateaus of quite different elevations. In contrast,as Kw rises with Kb fixed, each plateau becomesinternally rugged.

This sets the stage for understanding the effectsof Kb and Kw on analogical reasoning. The surfaceplot in Figure 4 shows, for various combinations ofKb and Kw, the long-run performance differentialbetween an analogizing firm with representation〈1 2 3〉 and a local searcher. (P = 3, D = 3, andDEPTH = BREADTH = 1.) When Kb = Kw = 0, thelandscape is smooth, both the analogizer andthe local searcher fare well, and as a result thedifferential is modest. Indeed, a differential existsonly because firms are constrained to abide by theiroriginal policies. If firms were heterodox, bothwould attain the global peak, and the differentialwould be precisely zero.9 As Kb rises with Kw

9 While it is true that analogy offers no long-run advantagewhen target landscapes are smooth (Kb = Kw = 0) and firmsare heterodox, this outcome masks important dynamics. Thoughall firms eventually scale the global peak in such a situation,analogical reasoning helps firms with good representations todo so quickly. As a result, analogizing firms enjoy a transientadvantage. The better is the firm’s representation, the biggeris this advantage. (Simulation results that illustrate this effectare available from the authors on request.) There are numerous,unmodeled reasons to believe that, in reality, such an advantagemight be of lasting importance. Discovering a strong competitiveposition first (Lieberman and Montgomery, 1988) may beespecially important if the first firm to arrive at the positioncan deter others from copying its configuration of choices; ifcustomers are loyal and reluctant to switch once others reproducethe position; if the selection environment is so vigorous that slow



Figure 4. Long-run advantage of analogizer with representation 〈1 2 3〉 as a function of Kw and Kb

low, internally smooth plateaus of different heightsemerge on the target landscape. In such a setting,it is quite valuable to start one’s local searchin a favorable policy domain—on a promisingplateau. Accordingly, analogy is very powerful forhigh Kb and low Kw. In contrast, an increase inKw undermines the power of analogy: interactionswithin policy domains make each plateau rugged,and even a firm with a good representation,which begins its local search in a favorable policydomain, is likely to get stuck on a low localpeak. Good analogies in our model give policy-level guidance, but if guidance at that level is notsufficiently specific to lead a firm to success, thenanalogical reasoning loses much of its force.

Overall, we find that analogy is least powerfulin fully decomposed settings (Kb = 0, Kw = 1) ornearly decomposed settings (Kb low, Kw high);in such settings, the primary challenge facingmanagement is to achieve success within each ofseveral, fairly independent policy domains, andanalogy—at least as we have modeled it—offerslittle toward this end. Analogical reasoning is alsoof limited use in situations of full independence(Kb = Kw = 0), where incremental logic alonewill deliver a good configuration of choices,

movers don’t survive their temporary disadvantage; or if theenvironment changes so rapidly that one must reap the benefitsof an advantage quickly. We speculate that analogical reasoningmight be especially valuable in such settings, even in the absenceof interactions among decisions.

and in situations of very high interdependence ofdetailed decisions (Kw = 1), where proliferatinglocal peaks trap analogizers and local searchersalike. Analogy is most powerful when cross-policyinteractions dominate (Kb high, Kw low)—that is,when the plateaus defined by policy configurationsare internally smooth, but heights are sufficientlyvaried across plateaus that there is a dangerof being stranded on a low plateau. A numberof scholars have argued that piece-by-piece,subsystem-by-subsystem learning is most effectivein nearly decomposable systems (Simon, 1962;Baldwin and Clark, 2000; Frenken, Marengo,and Valente, 1999). Our results suggest thatanalogical reasoning offers its greatest relativebenefit precisely where these approaches fail, insystems that are not nearly decomposable.

DISCUSSION AND CONCLUSION

Discovering an effective competitive position isa difficult endeavor. It is difficult because themapping from strategic decisions to performanceis typically complex and, especially in novelsettings, unknown to the decision-makers. Thoughcognizant of such difficulties, positioning scholarsemphasize the role of deductive reasoning andrational choice in the origin of positions (Porter,1996; Ghemawat, 1999). In contrast, evolutionarytheorists highlight the bounds of individual



rationality and posit that effective positions emergethrough a mix of luck and experiential, localsearch, thus leaving little space for the cognitionof managers (Nelson and Winter, 1982). Althoughrecognizing the merits of both such perspectives,we hesitate to ascribe all effective positionsto either the omniscience assumed in economicanalysis or the myopia of experiential learning.

We put forth a model of the origin of strategiesthat, while recognizing the limits of managerialrationality and the intelligence of local search,also respects the power of managerial cognition.Indeed, the notion of bounded rationality, whichwe take as a cornerstone of our conceptualapparatus, does not rule out the possibility ofintendedly rational choice (March and Simon,1958). Bounded rationality suggests that thinkingis typically premised on simplified cognitiverepresentations of the world (Simon, 1991).As boundedly rational actors, managers createcognitive simplifications of their decision problemsand come up with solutions on the basis ofsuch simplifications. These solutions, in turn, mayimprint subsequent efforts at local search, thusplaying a central role in the discovery of strategicpositions (Gavetti and Levinthal, 2000). Thisperspective, which represents a middle groundbetween positioning and evolutionary arguments,suggests that the roots of superior competitivepositions may lie in the cognition of managers,particularly in the way they represent the world.

The effect of cognition on managerial actionis an enormous area of inquiry. We cut into thistopic by looking at a particular, important type ofaction: the creation of a competitive position in anunfamiliar industry. This type of action confrontsmanagers with daunting complexity so it is naturalto expect cognition—a fundamental instrument tohandle complexity—to play a major role. It alsoforces decision-makers to cope with novelty. Innovel situations, wisdom from prior experience inother contexts can be particularly powerful. Forthis reason, we focus on a particular vehicle fortransporting experience across contexts: analogy.

Our conceptual model of analogical reasoningis straightforward. We surmise that, in any setof industries, a large number of underlyingcharacteristics drive the relationship between firmaction and performance. There are so manycharacteristics and their effects are so difficultto discern that boundedly rational actors focustheir reasoning efforts on a subset of the

characteristics. These subsets form representations.Some representations are more effective bases ofreasoning than are others. A good representationdistinguishes sets of similar payoff functionsfrom one another, while a poor representationleaves very different payoff functions in the samecategory. In other words, a representation is aclassification scheme. An effective scheme putssimilar objects in the same class and differentobjects in distinct classes.

Armed with a good representation and adequateexperience, a firm is well prepared to draw a can-didate solution from a germane source and applyit to a target industry. Accordingly, our simulationmodel shows the best performance among firmswith high-quality representations—good classifi-cation schemes. That said, we also find it better totake guidance from some source, even one basedon a poor representation, than to start one’s localsearch at a randomly assigned configuration. Thereis some chance that, purely by luck, the source willprove to be a good one even though it is basedon a poor representation. This power of ‘false’analogies is reminiscent of Weick’s (1990) tale ofa Hungarian military reconnaissance unit. Lost inthe snowy Alps, the troops prepare for the worst.Then one soldier discovers a map in his pocket.Once equipped with the map, the unit outlasts thestorm, finds its bearings, and returns to safety.Only later does a commanding officer realize thatthe map is of the Pyrenees, not the Alps. The taleis often interpreted in terms of presence of mindand motivation: the map calmed the troops andmoved them to take coordinated action. We wouldsuggest a second possibility: perhaps parts of thePyrenees and parts of the Alps truly do resem-ble one another. The soldiers may have receiveda good candidate solution, though only by chance,just as do some (though not all) of our firms withpoor representations. In business also, it is pos-sible to have a poor representation yet obtain agood candidate solution. For instance, Lycos’s par-tial integration of Tripod worked out well, but evensome supporters of the decision felt the analogicalreasoning itself was dubious (Gavetti and Rivkin,2004). An intriguing avenue for future research isto examine what makes some analogies useful evenif based on similarity along irrelevant dimensions.We suspect the key is to focus on dimensions that,even if irrelevant, are correlated with true driversof success.



Our results also shed light on the roles ofbroad and deep experience. We find breadthand depth of experience to be valuable onlyif a manager has a good representation—avalid system for categorizing environments andclassifying lessons learned. In addition, beyonda modest level of depth, performance is notsensitive to depth. As a result, if top managementof a diversified firm believes that business unitmanagers have a good grasp of what factorsdrive the choice–performance relationship in theirparticular business, they should be more willing torotate managers across divisions, to invest in andexploit breadth of experience. Otherwise, it maybe better to keep each business unit manager ina single business for a longer period, to developdepth.

A possibility we do not model, worthy offuture research, is that experience in a widerange of industries may help a manager to buildan understanding of what drives the relationshipbetween choice and performance. Put simply,breadth may improve representations. As businessschool instructors who teach by the case method,we are sympathetic to this perspective. Much ofwhat goes on in business strategy courses, webelieve, is representation building. We aim to focusstudent attention on dimensions of environmentsthat shape the relationship between action andoutcome, and we do so in part by exposingstudents to a broad variety of cases. A frameworksuch as the Five Forces (Porter, 1980) is helpfulbecause it synthesizes across cases and identifiesthe dimensions that are most salient across therange of particulars. In doing so, the frameworkenhances pattern recognition. Frameworks thatclassify environments and firms along differentdimensions—say, by the durability of underlyingresources (Williams, 1992) or by the character ofcapabilities (Teece, Pisano, and Shuen 1997) ratherthan by the nature of competitive forces—mightlead managers to see very different patterns.

More broadly, the pervasive impact of repre-sentations in our findings affirms the importanceof research that examines the mental models ofstrategists (e.g., Porac, Thomas, and Baden-Fuller,1989; Huff, 1990; Huff and Jenkins, 2002). Espe-cially intriguing is the question of where repre-sentations come from (other than business schoolcurricula). Future empirical efforts might exam-ine the processes by which management teamsdevelop, track, and alter their beliefs about what

characteristics distinguish similar settings from dif-ferent ones. Modeling efforts might expose simu-lated firms to a series of decision problems, not thesingle problem we model here, and allow manage-ment teams to alter their representations as theyupdate their beliefs about the true XPROB. Modelsmight also incorporate shocks to XPROB. Our spec-ulation is that established teams with solid beliefsabout XPROB will find such shocks especially haz-ardous. Such a team may very well persist in pay-ing attention to observable characteristics that usedto be relevant but are no longer. As a result, theymay draw analogies based on similarities that havebecome superficial.

Assessing the role and performance implicationsof analogies also requires analyzing how analogiesare used. In the context of our model, wedistinguish between orthodox and heterodox usesof analogy. Not surprisingly, orthodoxy is mostcostly for firms with poor representations. Moresurprisingly, orthodoxy provides no advantageover heterodoxy even when the analogy is based ona good representation. Good representations seedsubsequent search efforts on such promising areasof the landscape that heterodox analogizers arenot motivated to violate the policies identified bythe analogy. This result suggests that analogy ismore effective as an instrument to seed searchefforts than as a means to constrain them. Thepsychological literature on analogy focuses onrepresentations and experience as central elementsunderlying the quality of analogies (Holyoak andThagard, 1995). Particularly in the context oforganizational search efforts, we believe, that inaddition to these fundamental elements, managersshould pay explicit attention to the orthodoxy withwhich analogies are used.

Our final set of results shows that analogicalreasoning produces the greatest long-run advantageover incremental local search in settings thatare poorly decomposed. In these contexts,the high-level policy guidance offered byanalogy is necessary to identify consistent policyconfigurations—policy configurations that cannotbe recreated via incremental search at the level ofdetailed choices. Others have touted the virtues ofdecomposable or nearly decomposable systems ofchoices (Simon, 1962; Baldwin and Clark, 2000).In nearly decomposable or modular systems, onecan learn by means of parallel experiments atthe subsystem level (Baldwin and Clark, 2000).Moreover, boundedly rational managers might be



able to apply deductive reasoning at the subsystemlevel even if the system as a whole outstripstheir processing abilities. However, not all systemsof choices are neatly decomposable. We suggestthat, for dealing with systems of choices thatare inherently non-decomposable or have notyet been decomposed, reasoning by analogyis particularly powerful. Good representationsunderlie the transfer of powerful strategic solutionsfrom managers’ past experience. These solutions,in turn, allow a rich appreciation of the architectureof the strategic problem (Henderson and Clark,1990). It is in poorly decomposed systems thatthis kind of architectural wisdom, and therefore thequality of the cognitive representations underlyingthe analogy, play a particularly important role.These results emphasize that analogy and, moregenerally, cognition are especially powerful insettings where parallel local search fails to guideorganizational adaptation effectively.

More broadly, our analyses may help us betterunderstand observed variation in the processesthat underlie the origins of successful strategicpositions. Our findings lead us to hypothesize,for instance, that cases in which local search isthe main process guiding firms toward successfulpositions may well reflect some select structuralcharacteristics of the industries where suchcompanies succeed. In particular, such settings arelikely to have intense interactions among decisionswithin policies. Rich sets of interactions amongthe detailed decisions lying below the surface ofhigher-level policies limit the power of a prioricognition in these settings. We believe that thislinkage between the structural characteristics ofthe context and the mechanisms underlying thedevelopment of successful positions is a crucialone—one that future empirical studies shouldconsider carefully.

Cognition in complex worlds inevitably involvessimplification. The precise basis of simplification,however, is not inevitable. As academics andpractitioners, we are often apologetic aboutoperating in the space of such simplifications—thecommonly ridiculed ‘two-by-twos’ of the strategyfield. But as boundedly rational individuals, wecannot think in high-dimensional spaces. Therelevant question is not whether we conceiveof complex strategic problems in terms of afew overarching variables, but rather what thosevariables will be. The choice of variables canhave a major impact on performance, both directly

and by altering the influence of other factorssuch as managerial experience and orthodoxy.Our findings support this perspective in thecase of analogical reasoning, and we speculatethat it will also hold true for other forms ofmanagerial cognition. Our hope is that rigorousanalysis of cognition will help bridge thechasm between rational, positional perspectives onstrategy and behavioral, evolutionary approaches.Understanding how firms identify effectivecompetitive positions requires both perspectives.With the current work, we try to provide somesubstantiation of that link and a platform on whichothers can build.

ACKNOWLEDGEMENTS

For helpful comments, we are grateful to NicolajSiggelkow, anonymous referees, and seminaraudiences at Babson, Harvard Business School,MIT’s Sloan School, the University of Calgary,UCLA, the University of Toronto, and the WhartonSchool. We also thank Howard Brenner forcomputer programming efforts, Simona Giorgi andElizabeth Johnson for research assistance, and theDivision of Research of Harvard Business Schoolfor generous funding. Errors remain our own.

REFERENCES

Baldwin CY, Clark KB. 2000. Design Rules: The Powerof Modularity . MIT Press: Cambridge, MA.

Cyert R, March J. 1963. A Behavioral Theory of the Firm .Prentice-Hall: Englewood Cliffs, NJ.

Fleming L, Sorenson O. 2001. Technology as a complexadaptive system: evidence from patent data. ResearchPolicy 30: 1019–1039.

Fleming L, Sorenson O. 2004. Science as a map intechnological search. Strategic Management Journal ,Special Issue 25(8–9): 909–928.

Frenken K, Marengo L, Valente M. 1999. Interdependen-cies, near-decomposability and adaptation. In Com-putational Techniques for Modelling Learning in Eco-nomics , Brenner T (ed). Kluwer: Dordrecht; 145–165.

Gas Daily . 2000. Broadband to be as big as Enron in fiveyears. 5 April.

Gavetti G, Levinthal D. 2000. Looking forward andlooking backward: cognitive and experiential search.Administrative Science Quarterly 45: 113–137.

Gavetti G, Rivkin J. 2004. Rationality and plasticity overtime: toward a grounded theory of the origin ofstrategies. Working paper, Harvard Business School.

Ghemawat P. 1999. Strategy and the Business Landscape.Addison-Wesley: Reading, MA.



Gick ML, Holyoak KJ. 1980. Analogical problemsolving. Cognitive Psychology 12: 306–355.

Gilovich T. 1981. Seeing the past in the present: theeffect of associations to familiar events on judgmentsand decisions. Journal of Personality and SocialPsychology 40: 797–808.

Gragg CI. 1940. Because wisdom can’t be told. HarvardAlumni Bulletin 43(2): 78–84.

Hast A (ed). 1992. International Directory of CompanyHistories. St James Press: London.

Henderson RM, Clark KB. 1990. Architectural innova-tion: the reconfiguration of existing product technolo-gies and the failure of established firms. AdministrativeScience Quarterly 35: 9–30.

Holyoak K, Thagard P. 1995. Mental Leaps: Analogy inCreative Thought . MIT Press: Cambridge, MA.

Huff AS. 1990. Mapping Strategic Thought. Wiley: NewYork.

Huff AS, Jenkins M. 2002. Mapping Strategic Knowl-edge. Sage: Thousand Oaks, CA.

Kauffman SA. 1993. The Origins of Order: Self-Organization and Selection in Evolution . OxfordUniversity Press: Oxford.

Kauffman SA. 1995. At Home in the Universe. OxfordUniversity Press: Oxford.

Levinthal D, March J. 1993. The myopia of learning.Strategic Management Journal , Winter Special Issue14: 95–112.

Levinthal D. 1997. Adaptation on rugged landscapes.Management Science 43: 934–950.

Lieberman MB, Montgomery DB. 1988. First-moveradvantages. Strategic Management Journal , SummerSpecial Issue 9: 41–58.

March J, Simon HA. 1958. Organizations . Wiley: NewYork.

McKelvey B. 1999. Avoiding complexity catastrophein coevolutionary pockets: strategies for ruggedlandscapes. Organization Science 10: 294–321.

Mintzberg H. 1978. Patterns in strategy formation.Management Science 24: 934–948.

Nelson R, Winter S. 1982. An Evolutionary Theory ofEconomic Change. Belknap: Cambridge, MA.

Neustadt RE, May ER. 1986. Thinking in Time: The Usesof History for Decision-Makers . Free Press: New York.

Pascale RT. 1984. Perspectives on strategy: the realstory behind Honda’s success. California ManagementReview 26(3): 47–72.

Perkins EJ. 1999. Wall Street to Main Street . CambridgeUniversity Press: New York.

Porac JF, Thomas H, Baden-Fuller C. 1989. Competitivegroups as cognitive communities: the case of Scottishknitwear manufacturers. Journal of ManagementStudies 26: 397–416.

Porter ME. 1980. Competitive Strategy: Techniques forAnalyzing Industries and Competitors . Free Press:New York.

Porter ME. 1985. Competitive Advantage: Creating andSustaining Superior Performance. Free Press: NewYork.

Porter ME. 1996. What is strategy? Harvard BusinessReview 74(6): 61–78.

Rivkin JW. 2000. Imitation of complex strategies.Management Science 46: 824–844.

Rivkin JW. 2001. Reproducing knowledge: replicationwithout imitation at moderate complexity. Organiza-tion Science 12: 274–293.

Rivkin JW, Siggelkow N. 2002. Organizational stickingpoints on NK landscapes. Complexity 7: 31–43.

Rivkin JW, Siggelkow N. 2003. Balancing searchand stability: interdependencies among elementsof organizational design. Management Science 49:290–311.

Salter MS, Levesque LC, Ciampa M. 2002. Innovationcorrupted: the rise and fall of Enron. Harvard BusinessSchool Working Paper 02–102.

Siggelkow N. 2002. Evolution towards fit. AdministrativeScience Quarterly 47: 125–159.

Simon HA. 1955. A behavioral model of rational choice.Quarterly Journal of Economics 69: 99–118.

Simon HA. 1962. The architecture of complexity.Proceedings of the American Philosophical Society106: 467–482.

Simon HA. 1991. Bounded rationality and organizationallearning. Organizational Science 2: 125–134.

Slywotzky AJ, Morrison DJ. 1999. Profit Patterns: 30Ways to Anticipate and Profit from Strategic ForcesReshaping Your Business . Times Business: New York.

Stemberg TG. 1996. Staples for Success: From BusinessPlan to Billion-dollar Business in Just a Decade.Knowledge Exchange: Santa Monica, CA.

Teece DJ, Pisano G, Shuen A. 1997. Dynamic capabili-ties and strategic management. Strategic ManagementJournal 18(7): 509–533.

Thagard P. 1996. Mind: Introduction to CognitiveScience. MIT Press: Cambridge, MA.

Weick K. 1990. Cartographic myths in organizations.In Mapping Strategic Thought , Huff AS (ed). Wiley:New York; 1–10.

Westhoff FH, Yarbrough BV, Yarbrough RM. 1996.Complexity, organization, and Stuart Kauffman’s TheOrigins of Order. Journal of Economic Behavior andOrganization 29: 1–25.

Williams JR. 1992. How sustainable is your competitiveadvantage? California Management Review 34(3):29–51.

APPENDIX: DETAILED DESCRIPTIONOF HOW FAMILIES OF LANDSCAPESARE GENERATED

For a given set of values for P , D, X, Kw, Kb,and XPROB, the computer generates an influencematrix—a matrix that records which choices,policies, and observable characteristics influencethe contribution of each detailed choice. Figure 1shows what the matrix might look like for a casein which P = 3, D = 3, X = 4, Kw = 1, Kb =0.33, and XPROB = (1, 0.5, 0.5, 0). A bullet (•) inthe matrix indicates that the column element of



the matrix influences the contribution of the rowelement. The second row, for instance, indicatesthat the contribution of the second detailed choiceis affected by the resolutions of choices 1, 2, and 3,the resolution of policy 3, and the states of the firstand third observable characteristics. Because Kw =1, there is thorough interaction within each policydomain. Because Kb = 0.33, the typical decisionis affected by one-third of the other policies. Thefirst observable characteristic is highly influential,affecting the contributions of all detailed decisionsbecause the first element of XPROB is 1. The fourthobservable characteristic is altogether irrelevantsince the fourth element of XPROB is 0.

Once the influence matrix is set, the computergenerates target and source landscapes. Thatis, it assigns a pay-off to each of the 2P×D

possible configurations of choices on each ofthe 2X landscapes. The contribution of eachdetailed choice depends on the resolution ofthat choice, the resolution of other choices andpolicies, and the state of observable characteristics.

For each possible combination of the influentialfactors, the computer draws a contribution atrandom from a uniform U[0, 1] distribution.In Figure 1, for instance, the second choice isaffected by six factors (as indicated by thesix bullets in the second row). Each of thesefactors can be resolved in two ways, so thecomputer draws 26 possible contributions for thesecond choice: one corresponding to each of thepossible configurations of the six factors. Theoverall performance associated with a particularconfiguration of detailed choices and observablecharacteristics is then the average over the P × D

contributions:

Performance(d given a set of observable

characteristics) =[

P×D∑i=1

contribution of decision i

given d and observable characteristics

] /(P × D).


strategy making in novel and complex worlds: the power of ... · decomposed decision problems. in...

Documents