This article was downloaded by: [2.37.204.136] On: 24 July 2023, At: 09:04 Publisher: Institute for Operations Research and the Management Sciences (INFORMS) INFORMS is located in Maryland, USA Organization Science Publication details, including instructions for authors and subscription information: http://pubsonline.informs.org Reproducing Knowledge: Replication Without Imitation at Moderate Complexity Jan W. Rivkin, To cite this article: Jan W. Rivkin, (2001) Reproducing Knowledge: Replication Without Imitation at Moderate Complexity. Organization Science 12(3):274-293. https://doi.org/10.1287/orsc.12.3.274.10106 Full terms and conditions of use: https://pubsonline.informs.org/Publications/Librarians-Portal/PubsOnLine-Terms-andConditions This article may be used only for the purposes of research, teaching, and/or private study. Commercial use or systematic downloading (by robots or other automatic processes) is prohibited without explicit Publisher approval, unless otherwise noted. For more information, contact permissions@informs.org. The Publisher does not warrant or guarantee the article’s accuracy, completeness, merchantability, fitness for a particular purpose, or non-infringement. Descriptions of, or references to, products or publications, or inclusion of an advertisement in this article, neither constitutes nor implies a guarantee, endorsement, or support of claims made of that product, publication, or service. © 2001 INFORMS Please scroll down for article—it is on subsequent pages With 12,500 members from nearly 90 countries, INFORMS is the largest international association of operations research (O.R.) and analytics professionals and students. INFORMS provides unique networking and learning opportunities for individual professionals, and organizations of all types and sizes, to better understand and use O.R. and analytics tools and methods to transform strategic visions and achieve better outcomes. For more information on INFORMS, its publications, membership, or meetings visit http://www.informs.org Reproducing Knowledge: Replication Without Imitation at Moderate Complexity Downloaded from informs.org by [2.37.204.136] on 24 July 2023, at 09:04 . For personal use only, all rights reserved. Jan W. Rivkin Harvard Graduate School of Business Administration, Morgan Hall 239, Boston, Massachusetts 02163 jrivkin@hbs.edu Abstract The complexity of a firm’s strategy affects both the ease with which the firm can replicate the strategy in a new setting and the ease with which rivals can imitate it. Simple strategies are as readily imitated as replicated, and highly intricate strategies resist imitation and replication equally. At moderate levels of complexity, however, a wedge develops between the ease of replication and the difficulty of imitation, so long as the replicator has better information than the imitator about the original success. An agent-based simulation model clarifies the structural reasons that this is so. The model also shows how the wedge-maximizing level of complexity varies with the replicator’s informational edge over the imitator. The results help to pinpoint situations in which strategies requiring replication are likely to defy imitation and generate sustained competitive advantage. More generally, the analysis sheds light on the value of superior but imperfect information about good solutions to hard problems. Finally, the results suggest that a pattern long observed by organization scholars—that “loosely coupled organizations” are especially effective competitors—may arise for a very different reason than is normally posited. (Replication; Imitation; Complexity; Interaction; Loose Coupling; Competitive Advantage) imitation is elusive. Rather, replication and imitation often go hand-in-hand. A productive system whose component resources are readily assembled and coordinated is easy to replicate. If such a system is profitable, however, imitators will rapidly construct equivalent systems and, in the process, drive down output prices or bid up input costs. Conversely, the factors that permit a system to defy imitation can forge equally strong barriers to replication. The tacit knowledge and personal relationships that make a particular factory run smoothly, for instance, are difficult to reproduce in a second plant, regardless of who owns that plant. Productive routines which are challenging targets for rivals are often difficult targets for internal agents as well (Nelson and Winter 1982). What characterizes a productive system which can be replicated within a firm more easily than it can be imitated by rivals? In this paper, I propose that the complexity of a system can drive a wedge between the ease of replication and the ease of imitation. Following Simon (1962), I define complexity as having two aspects; a productive system is complex if (a) it consists of numerous elements and (b) those elements interact with one another richly. I then use NK simulation techniques (Kauffman 1993) to establish the main hypothesis of the paper: HYPOTHESIS. In the settings simulated here, the gap between replicability and imitability tends to be greatest at moderate levels of complexity. 1. Introduction Entrepreneurs crave a superior system of production that they themselves can replicate but others cannot imitate. By replicating such a system within the boundaries of the firm, the entrepreneur can exploit the system’s advantages in a wide range of markets. Yet because the system cannot be imitated by rivals, the entrepreneur does not face competitors of equal strength in those markets. In such a situation, replication may be the centerpiece of a firm’s competitive efforts (Winter and Szulanski 2000). Unfortunately for entrepreneurs, replication without ORGANIZATION SCIENCE, 䉷 2001 INFORMS Vol. 12, No. 3, May–June 2001, pp. 274–293 Simple systems are as easily imitated as replicated, and highly intricate systems resist imitation and replication equally. Systems of moderate complexity, however—for structural reasons described below—are more easily replicated than imitated. The paper is organized as follows. Section 2 argues that replication and imitation are connected, but the link can be severed. The section reviews game-theoretic, resourcebased, and organizational explanations for replicationwithout-imitation, but finds the most promising explanation in an evolutionary perspective that views replication and 1047-7039/01/1203/0274/$05.00 1526-5455 electronic ISSN Downloaded from informs.org by [2.37.204.136] on 24 July 2023, at 09:04 . For personal use only, all rights reserved. JAN W. RIVKIN Reproducing Knowledge imitation as competing search efforts. Accordingly, §3 formulates an agent-based simulation model of competitive search. In the model, imitators and replicators compete to discover, or rediscover, the knowledge embodied in a successful productive system. The model purposely suppresses game-theoretic, resource-based, and organizational factors that may impede both imitation and replication but are unlikely to account for a wedge between the two. As a result, the model focuses exclusively on how the character of knowledge and search can generate replication-without-imitation. The results of the model, presented in §4, show how moderate complexity can make imitation more difficult than replication. A final section summarizes the empirical implications of the paper. It also offers a new interpretation of a longstanding observation in organizational theory, that “loosely coupled organizations” are especially effective competitors. Under this interpretation, the correlation between loose coupling and superior performance arises spuriously, not because loose coupling is advantageous per se. 2. The Link Between Replication and Imitation Pioneering efforts in evolutionary economics identified replication and imitation as brethren phenomena. Nelson and Winter (1982) cast both as efforts to reproduce successful routines. Later research pointed out that this similarity poses a fundamental dilemma to any firm which hopes to transfer its successes to new sites: Not only do the “[f]eatures that restrain involuntary transfer tend to inhibit voluntary transfer” (Winter 1987, p. 174), but any effort undertaken to ease internal transfer can make imitation more likely. An attempt to codify knowledge for the purpose of replication, for instance, might make it easier to communicate that knowledge to outsiders or make it easier for outsiders to interpret the knowledge (Kogut and Zander 1992, Zander and Kogut 1995). 2.1. A Breakable Link (Some Examples) Case studies bear out the proposition that replication and imitation are closely related. Examples also show, however, that the link between the two is not impregnable. In the examples which follow and the subsequent discussion, I focus on the replication and imitation of entire business strategies—that is, the set of customer needs targeted by a business and the full array of functional policies deployed to satisfy those needs in a superior manner. This is a departure from most of the earlier literature on replication and imitation, which has tended to examine the copying of smaller units of analysis: isolated processes (e.g., a new distillation technique), products (a ORGANIZATION SCIENCE /Vol. 12, No. 3, May–June 2001 semiconductor), installations (a factory), or practices (total quality management). I depart in this way in order to adopt the broadest framing for the analysis. The conceptual arguments apply to productive systems ranging in scale from entire strategies to individual processes. I return to the topic of smaller-scale systems in the final section. The early history of White Castle (Hogan 1997, especially Chapter 2) illustrates the hazards of an easily replicated, but quickly imitated, business strategy. White Castle’s founders, Walt Anderson and Billy Ingram, are widely credited with inventing both the hamburger and the fast-food business. In Anderson’s four ramshackle food stands in Wichita, Kansas, Ingram saw a success formula that could be reproduced on a national scale. Partnering with Anderson in 1921, Ingram ruthlessly standardized and codified everything about a White Castle hamburger stand: the white steel exterior with crenellated walls and turrets; the interior layout with five stools facing a grill; the staffing with a single cook and “clean-up man;” the simple menu of hamburgers, CocaCola, coffee, and pie; the location in working-class neighborhoods and near factories; the emphasis on hygiene; and so forth. He then applied the formula rapidly in other cities: Omaha, Kansas City, St. Louis, Minneapolis, Louisville, Cincinnati, Columbus, Chicago, and New York, all by 1930. As the chain expanded, Ingram insisted on absolute standardization. To guarantee uniformity, the company founded subsidiaries to grind beef, manufacture steel buildings, and make paper caps and aprons. A monthly newsletter, “Hot Hamburger,” publicizing corporate successes and standards, was distributed to employees and customers (and presumably to would-be rivals). Posters entitled “Look Yourself Over” reminded employees of how they should appear and behave. Experienced employees were assigned to new territories, and Ingram and Anderson learned to fly what must have been one of the first corporate airplanes so that they could inspect far-flung sites. White Castle’s highly visible success, however, attracted a flood of copycats. Soon, lookalike white buildings with towers, turrets, or lighthouses spotted the corners of working-class neighborhoods throughout the nation. Most adopted names that were clear takeoffs from “White Castle”: among the contenders were White Palace, White Log, White House, White Tavern, White Hut, White Cap, White Shop, White Grille, White Cabin, White Plaza, White Wonder, White Knight, White Turret, White Tower, White Diamond, White Fortress, White Kitchen, White Crescent, White Spot, White Manna, White Mill, Blue Bell, Blue Beacon, Blue Tower, Blue Castle, Red Lantern, Red Barn, Red Castle, Silver Castle, Magic 275 Downloaded from informs.org by [2.37.204.136] on 24 July 2023, at 09:04 . For personal use only, all rights reserved. JAN W. RIVKIN Reproducing Knowledge Castle, Prince’s Castle, Klover Kastle, Little Kastle, Little Crown, Modern Castle, Royal Castle, and Castle Blanca. Many imitators used White Castle as a direct template. The founders of White Tower, for instance, visited the new White Castles in Minneapolis, observed operations closely, photographed the sites, took interior and exterior measurements, copied products, and “lured away a White Castle operator at princely wages” (Hogan 1997, p. 54). By the early 1930s, White Tower operated nearly as many outlets as did White Castle. Imitation of White Castle’s strategy had proven almost as easy as replication. Just as one can find replicable strategies that were quickly imitated, so it is easy to find examples of the converse: hard-to-imitate strategies that also defied replication. Marks and Spencer, for instance, succeeded for generations as Britain’s leading apparel and food retailer and one of its most admired companies (Montgomery 1994, Collis 1996). Its products, sold only under its St. Michael’s label, were manufactured by an intricate network of suppliers. Many suppliers produced almost exclusively for Marks and Spencer and did so for decades. Through product-development and manufacturing partnerships with these suppliers, the company produced high-quality items very efficiently. The quality of its products, its good prices, and its reliable service generated enormous customer loyalty. It also enjoyed loyalty from employees, to whom it had offered above-average wages, benefits, and amenities since the 1920s. Imitators failed to match Marks and Spencer’s faithful following among suppliers, customers, and employees, or to equal its prime locations on Britain’s High Streets. Yet Marks and Spencer was unable to duplicate its British success in Canada and Europe in the 1970s and 1980s. In Canada, the company tried to reconstruct its intricate supply network in order to avoid high tariffs. The new suppliers failed to meet Marks and Spencer’s quality standards, however, and the company resorted to importing goods from its traditional suppliers. Canadian and European consumers were unfamiliar with the St. Michael’s brand, and employee enthusiasm was difficult to recreate abroad. Overall, “the same difficulties that inhibit[ed] competitors from copying M & S’s position in the United Kingdom also limit[ed] M & S’s ability to replicate its U.K. position overseas” (Collis 1996, p. 1). The connection between imitation and replication is tight, but it is not ironclad. The pantheon of management includes many entrepreneurs who reproduced successful strategies within a firm but resisted imitation. Herb Kelleher replicated Southwest’s no-frills airline success in markets from California to New England even as other airlines such as Continental failed in their no-frills efforts (Hallowell 1997, Porter 1996). Under Keith Iverson, Nucor set up 276 profitable steel mini-mills in numerous independent locations while rivals such as North Star and Northwestern Steel and Wire struggled in the same business (Ghemawat and Stander 1996). Ray Kroc installed his system for producing fast food in thousands of McDonald’s locations, but kept would-be imitators such as Burger King and Hardee’s at bay. Others who have coupled replicability with inimitability, at least for some time, include John McCoy of Banc One (Uyterhoeven and Hart 1996, Winter and Szulanski 2000), Charles Lazarus of Toys “R” Us (Rivkin 1995), Sam Walton of Wal-Mart (Foley et al. 1996), and Robert Cizik of Cooper Industries (Collis and Stuart 1995). 2.2. Barriers to Replication and Imitation Under what conditions are the bonds between replication and imitation broken? To address this question, I consider four perspectives on barriers to imitation and replication, and I identify obstacles that apply unequally to the two types of effort (Table 1). The analysis makes the case that none of the candidate explanations for replicationwithout-imitation is wholly satisfying, but explanations concerning managerial search deserve closer examination. Game-Theoretic Strategic Maneuvers. Industrial economists, strategy theorists, organization theorists, and evolutionary economists have examined imitation and replication from very different angles. In the models of mainstream industrial economics, firms typically deter imitation by executing some game-theoretic maneuver: By incurring a sunk cost, the firm alters its own marginal incentives in a way that fends off rivals. By installing a great deal of productive capacity, for instance, a firm might credibly commit to resist any entrant vigorously, and this commitment may cause would-be imitators to avoid the market (Dixit 1980). Alternatively, the first firm in a market might fill the “space” of possible products (Schmalensee 1978) or write costly-to-break contracts with customers (Aghion and Bolton 1987) that make it difficult for copycats to secure a beachhead. In line with economic orthodoxy, such models assume (and the assumption is so common that it usually goes unstated) that replication is straightforward (Nelson and Winter 1982). Productive knowledge can be duplicated within a firm at no cost. Thus, for most industrial economists, the problem of replication is foreign, and imitation and replication are unconnected. Perhaps, then, cases of replication-without-imitation arise when the replicator makes some game-theoretic move that repels copycats. For two reasons, however, this explanation appears incomplete at best. First, there is little empirical evidence that such maneuvers are prevalent. It ORGANIZATION SCIENCE /Vol. 12, No. 3, May–June 2001 Downloaded from informs.org by [2.37.204.136] on 24 July 2023, at 09:04 . For personal use only, all rights reserved. JAN W. RIVKIN Reproducing Knowledge has proven difficult to detect preemptive, game-theoretic behavior in large-scale statistical tests (Lieberman 1987a, 1987b; Gilbert and Lieberman 1987; Ghemawat and Caves 1986). Indeed, the most prominent and direct statistical analyses of imitation do not even examine preemption as a candidate barrier to imitation (Levin et al. 1987, Mansfield 1985, Zander and Kogut 1995).1 The second, more fundamental, reason that the gametheoretic explanation is incomplete is that it discounts the difficulty of replication. A number of empirical studies show that replication of success is costly and far from automatic. Teece (1977), for instance, examines 26 international technology transfer projects. He finds that the transmission and assimilation of unembedded technical know-how accounts, on average, for 19% of project costs. For four of the projects, the figure exceeds 40%.2 In a study of multiplant commercial food operations, Chew et al. (1990) find transfer of best practice so difficult that the best plants within a company remain twice as productive as the worst, even after one controls for differences in technology, location, and plant size. Szulanski (1996, fn. 1) reports that internal benchmarking efforts commonly uncover such wide gaps. In sum, then, the game-theoretic perspective does not provide a definitive explanation for replication-without-imitation. Access to Key Resources. Unlike industrial economists, strategy theorists have typically acknowledged that replication of a successful strategy can be difficult (e.g., Collis and Montgomery 1997, pp. 67, 154–159). While economists have focused on preemptive moves in product markets, proponents of the resource-based view of strategy turn to factor markets in order to explore imitation and replication. In this view, an imitator fails if it cannot muster resources comparable to the incumbent rival (Dierickx and Cool 1989, Barney 1991). For instance, the imitator may have to make do with an inferior input (e.g., the second-best retail location in town), may find distribution channels already congested, may lack a brand that customers value or trust, or may fail to have the complementary assets (Teece 1986) necessary to succeed. Such factors, however, do not automatically make imitation harder than replication. In a new market, a replicator may have just as hard a time garnering inputs, breaking into distribution channels, establishing a brand, or assembling complementary assets as would an imitator. Indeed, difficulty in acquiring resources may be the key constraint on internal firm growth (Penrose 1959). With respect to some resources, the replicator and the imitator start on the same footing when they enter a new setting. There are at least two resources to which the replicator ORGANIZATION SCIENCE /Vol. 12, No. 3, May–June 2001 Table 1 Potential Barriers to Imitation and Replication Potential Barrier Barrier to Barrier to Imitation? Replication? Game-theoretic strategic maneuvers by incumbents • Capacity preemption (Dixit 1980) ✓ Implicit • Spatial preemption (Schmalensee 1978) ✓ assumption: • Contractual preemption of customers/ ✓ no barriers switching costs (Aghion and Bolton to replication 1987) Unfavorable access to key resources • Poor access to inputs • Poor access to distribution • Lack of brand • Lack of complementary assets (Teece 1986) • Inexperience with similar technology • Patent protection Organizational impediments or weaknesses • Lack of absorptive or retentive capacity (Teece 1977, Cohen and Levinthal 1990, Szulanski 1996) • Poor motivation of imitator (Porter 1996) or replicator personnel (Szulanski 1996) • Not-invented-here syndrome (Katz and Allen 1982) • Internal jealousies; arduous relation between source and recipient (Szulanski 1996) • Desire to preserve scarcity and tap monopoly rents (Winter 1995) Obstacles to effective search • Causal ambiguity (Lippman and Rumelt 1982) • Tacit nature of required knowledge (Polanyi 1962, Winter 1987) • Poor codifiability (Zander and Kogut 1995) • Poor teachability (Zander and Kogut 1995) • Complexity (Winter 1987, Rivkin 2000) • Lack of template (Nelson and Winter 1982) • Inability to observe in use (Winter 1987) ⻫ ⻫ ⻫ ⻫ ⻫ ⻫ ⻫ ⻫ ✓ ✓ ⻫ ⻫ ⻫ ⻫ ⻫ ⻫ ✓ ✓ ⻫ ⻫ ⻫ ⻫ ⻫ ⻫ ⻫ ⻫ ⻫ ✓ ⻫ ✓ Bold check marks indicate barriers that may apply unequally to imitation and replication. would have preferential access: prior production experience and patented knowledge. The “experience curve” is widely touted to give established incumbents an advantage which is hard to imitate (Boston Consulting Group 1972, Porter 1980). Such experience, however, is often 277 Downloaded from informs.org by [2.37.204.136] on 24 July 2023, at 09:04 . For personal use only, all rights reserved. JAN W. RIVKIN Reproducing Knowledge described as localized within a single installation, so it is unclear just how experience aids replication elsewhere. Moreover, empirical evidence concerning experience seems equivocal. In his study of 39 chemical product industries, Lieberman (1989) finds no evidence that incumbent experience deters entry and only weak indications that it increases entrant mortality.3 He concludes that the benefits of experience may diffuse rapidly throughout such industries, a conclusion echoed by others who examine different contexts (Mansfield 1985, Jaffe 1986, Irwin and Klenow 1994). Likewise, patents guard against imitation only in limited circumstances (Winter 1987). Mansfield et al. (1981) find that 60% of patented innovations are imitated within four years and that patenting raises the cost of imitation by a mere 6–11%.4 Reflecting this, Levin et al. (1987) report that few R&D executives consider patents an especially effective means to protect the fruits of innovation. Lead time, learning curves, sales or service efforts, and (in the case of process innovations) secrecy are judged to be more effective. Patents may be crucial in some industry contexts, but they tend to be “invented around,” are unavailable for some technologies, and require disclosure of sensitive information. The more recent survey results of Cohen et al. (2000) confirm that executives deem patents crucial in only a handful of industries—and often as bargaining chips in negotiations or suits, not only as a means to prevent copying. In sum, then, key resources do not always drive a decisive wedge between replicability and imitability. It may be hard to gain access to resources in new settings, but this is true for both the replicator and the imitator. And two resources that offer the replicator preferential access—experience and patents—do not seem to pose uniformly high barriers to imitation. While resources may account for some instances of replication-withoutimitation, it seems worthwhile to search for other explanations. Organizational Impediments or Weaknesses. A number of researchers have highlighted the organizational characteristics that induce or impede the transfer of good ideas. An interlocking set of beliefs, values, measurement systems, incentives, and staff functions must be in place to enable smooth transfers (Chew et al. 1990; Szulanski 1996). The recipient of the transfer must have the ability to absorb and retain what is conveyed (Teece 1977, Cohen and Levinthal 1990, Szulanski 1996). In an internal transfer, the relationship between the source and the recipient must be intimate and productive, and both must be sufficiently motivated to complete the handoff (Szulanski 1996); similarly, in an imitation effort, the external copycat must not have made prior choices or commitments that dampen its motivation to complete the 278 transfer successfully (Porter 1996). A replicator or imitator must overcome the tendency of an organization to reject ideas that were “not invented here” (Katz and Allen 1982). Though organizational considerations and motivation may deter both replication and imitation, there are two reasons to doubt that they drive a wedge between the ease of replication and the difficulty of imitation. First, existing empirical work suggests that organizational impediments do not pose the most important barriers to either replication or imitation. Szulanski (1996), for instance, finds that knowledge-related factors, not motivational considerations, form the major barriers to internal replication of best practices. In the quick print industry, Knott and McKelvey (1999) find that efficiency depends far more on access to successful routines than on the presence of high-powered incentives (though the role of incentives is sometimes statistically significant). Second, along organizational and motivational dimensions, a replicator may actually face higher barriers to transfer than an imitator does. Internal jealousies may impede replication while factions within an imitator may unite to face a common, external enemy. Moreover, a replicator’s incentives to expand output are dulled: The replicator knows that expansion will depress prices, including prices on the volume it was selling before expansion. If imitation were not a threat, it might prefer to restrict output and maintain high prices rather than expand and “compete against itself” (Winter 1987, 1995). The imitator, on the other hand, was not selling any units before and therefore does not worry about depressing prices on preexpansion output. Its motives to expand production are unmixed. Organizational impediments, then, do not offer a compelling explanation for replication-withoutimitation. Obstacles to Effective Search. An alternative perspective on replication and imitation views the transfer of productive knowledge—by either replication or imitation— as a search process (Nelson and Winter 1982). The imitator must discover, and the replicator must rediscover, the combination of routines, activities, and resources which made the original installation so successful. Certain types of knowledge are inherently hard to (re)discover because it is difficult to specify and communicate precisely where the original combination resides in the space of all possible routines, activities, and resources. On the figurative treasure map, it is hard to place the “X” that “marks the spot.” This difficulty may arise because no one in an organization fully understands the connection between firm actions and outcomes, a con- ORGANIZATION SCIENCE /Vol. 12, No. 3, May–June 2001 Downloaded from informs.org by [2.37.204.136] on 24 July 2023, at 09:04 . For personal use only, all rights reserved. JAN W. RIVKIN Reproducing Knowledge dition known as causal ambiguity (Lippman and Rumelt 1982, Reed and DeFillippi 1990). It may occur because production calls on tacit individual actions or connections among individuals that the involved individuals themselves do not consciously understand (Polanyi 1962) or that cannot be codified or taught inexpensively (Zander and Kogut 1995).5 Or, difficulty may arise because the productive knowledge is complex. “Complexity” demands further explanation. The definition of the term is the subject of vigorous, sometimes strident debate among advocates and skeptics of “complexity theory” (Horgan 1995). By one accounting, researchers examining complexity employ 41 different definitions of the word.6 Within the narrow domain of evolutionary and behavioral economics, however, the term is used fairly consistently. Simon (1962) defines complexity as having two aspects: An item is complex if it consists of many elements and those elements interact richly. Reflecting Simon’s first aspect, Winter (1987) follows Kolmogorov (1965) and defines the complexity of an item of knowledge as the amount of information required to characterize it. Reflecting the second aspect, Zander and Kogut (1995) argue that a piece of knowledge is complex if it integrates many distinct competencies within a firm. In this paper, I adopt Simon’s definition, which encompasses Winter’s and Zander and Kogut’s. I pay particular attention to Simon’s second aspect, the intensity of interactions among elements. Of course, the fact that certain types of productive knowledge require and resist search does not by itself imply any wedge between replicability and imitability. Knowledge that is causally ambiguous, tacit, uncodifiable, unteachable, or complex poses a challenge to both voluntary and involuntary transfer. A number of researchers have argued that valuable pieces of knowledge or winning business strategies that consist of many interwoven elements are difficult for outsiders to copy (Porter 1996, Rivkin 2000 and references therein). But why should insiders be any more able to reproduce such successes? A “basic proposition of evolutionary economics” is that “replication of an established routine is much easier than imitation . . .” (Winter 1995, p. 166). The crux of the evolutionary argument is that, in searching to (re)discover the successful combination, the replicator enjoys a key advantage: It has full access to the original success which serves as a template (Nelson and Winter 1982, Winter 1995). The template may be useful in a number of ways. It may allow the replicator to start with a better facsimile, to initiate its search in closer proximity to its ultimate target. It may give the replicating managers a place to turn when problems arise with the new installation. Or, it ORGANIZATION SCIENCE /Vol. 12, No. 3, May–June 2001 may generate experienced personnel who can staff the new operation or train their counterparts. As the replicator makes a number of attempts to reproduce its original success, it gains access to multiple templates—some more successful than others, in all likelihood. Through multiple experiments, the replicator gradually comes to understand what truly matters in the original success, what attributes deserve to be replicated, and how these attributes can be attained—an understanding that Winter and Szulanski (2000) label the “Arrow core.” A template may be an especially potent source of advantage when the roots of a replicator’s success cannot be observed in use (Winter 1987), that is, when the external output of a productive system gives few clues about the system’s internal composition. The concept of observability in use has traditionally been applied to discrete, technological innovations. Product advances, it is reasoned for instance, are more readily observed and reverse-engineered than process innovations. Accordingly, new products may be imitated more rapidly than new processes.7 The idea of observability in use also applies to innovative strategies, however. The strategy of a retailer such as White Castle is more visible to the public than, say, that of the steel maker Nucor. In sum, replication-without-imitation may arise when (a) the nature of the productive knowledge necessitates search but makes it challenging, and (b) the replicator’s access to a template gives it an advantage in the search process. Efforts to confirm this view empirically, though rare, have been supportive.8 Alternative perspectives concerning strategic maneuvers, resource impediments, and organizational considerations—while important—fall short of a thorough explanation. The remainder of the paper uses a simple simulation model of managerial search to make the evolutionary argument more precise. Specifically, it explores a key question concerning the argument: Under what conditions of search complexity and template validity is the gap between the ease of replication and the difficulty of imitation the widest? Put differently, can complexity drive a wedge between replicability and imitability? An important feature of the model is that resource impediments and organizational considerations—which can suppress both imitation and replication even if they do not make imitation harder than replication—are suppressed. By assumption there is not, for instance, some truly unique resource that makes replication and imitation downright impossible. Rather, the model creates a world—admittedly a stylized world—in which the success of replicators and imitators depends entirely on the character of the knowledge they have and seek. 279 JAN W. RIVKIN Reproducing Knowledge Downloaded from informs.org by [2.37.204.136] on 24 July 2023, at 09:04 . For personal use only, all rights reserved. 3. Simulation of Replication and Imitation My simulation approach is an extension of the NK simulation model developed by Stuart Kauffman and his colleagues in the context of evolutionary biology (Kauffman and Levin 1987, Kauffman et al. 1988, Kauffman and Weinberger 1989, Kauffman 1993). Although devised to explore how organisms and proteins evolve, the NK technique can be adapted to examine managerial search. A number of researchers have used NK models to analyze human organizations.9 Indeed, tools from complexity research are increasingly being applied to organization science (Anderson 1999). None of the research efforts to date have examined the tension between replication and imitation. A significant virtue of the NK framework, for the present paper, is that it ties the hands of the modeler fairly tightly. The modeler can manipulate the basic setup in only a handful of ways, most of which have been vetted by other researchers. Consequently, an NK simulation can be viewed as an experiment in a “world” that others have found reasonable—and importantly, not an idiosyncratic world tailormade by the modeler. This reduces the fear that the model is rigged to produce the desired result. The fear may linger, however. To address such concerns below, I emphasize the intuition behind the results and discuss the robustness of the results to alternative specifications. The simulation unfolds as follows. First, I fix the complexity of the decision problem which modeled managers will face; that is, I choose (a) the number of decisions which together constitute a strategy and (b) the degree to which those decisions interact with one another in determining firm performance. Using NK techniques described below, a computer then generates—in a stochastic but well-controlled manner—a decision problem with the specified level of complexity. The decision problem is a particular mapping from choices to payoffs. Graphically, the decision problem can be conceived of as a landscape in a high-dimensional space. Each discrete decision constitutes a “horizontal” axis, and each possible choice concerning a decision is a point along that axis. The vertical axis records the payoff from each combination of decisions, and the mapping from combinations to payoffs defines the landscape.10 Once a decision problem is generated, I assume that one firm, Firm R (for “replicator”), has happened upon the best solution to the problem, the highest point on the landscape. The firm faces the challenge of replicating the solution in a new installation. The managers of the new installation are “released” on the landscape at a point near 280 the highest peak but, reflecting the previous section, not precisely atop the summit. They then face the problem of rediscovering the solution. At the same time, a second organization, Firm I (for “imitator”), tries to imitate Firm R’s original success. That is, it tries to reproduce the set of choices that led Firm R to be so successful initially. Because the imitator lacks a template, however, its search begins from a more distant point and proceeds with less accuracy. Though the landscape was generated by a random process, it is fixed and determined from the perspective of Firms R and I. The computer records the relative success of the replicator and the imitator in reproducing the original success and solving the decision problem. It then generates another decision problem which, though it differs in its particulars, has the same degree of complexity. A second pair of firms—replicator and imitator—tackle the second problem. This process is repeated hundreds of times. From the repetition emerges a profile of how replicators and imitators fare relative to one another for a given degree of complexity. I then adjust the parameters that govern complexity and repeat the process. By doing so, I gradually build up an understanding of how complexity affects the gap between replicability and imitability. To describe the simulation thoroughly, I next explain how decision problems are generated and how modeled managers search for better choices. Table 2 summarizes the simulation parameters and symbols. 3.1. Generation of Decision Problems Two parameters, N and K, govern the complexity of a modeled firm’s decision problem. N is the first aspect of complexity, the number of decisions that a firm must make. Each decision j, j 僆 {1, 2, . . ., N}, can be resolved in two ways. Hence a particular strategy s, a configuration of decisions, is an N-vector {s1, s2, . . ., sN}, with sj 僆 {0, 1}, from the space of all possible configurations S. Note that |S| ⳱ 2N. K controls the second aspect of complexity, the degree to which the decisions interact. The efficacy of each decision is affected not only by the choice (0 or 1) made concerning that decision, but also by the choices regarding K other decisions. In the model, each decision j makes a contribution Cj to overall firm value. Cj, a number between 0 and 1, depends not only on the resolution of decision j itself (sj), but also on how K other randomly assigned decisions are resolved: Cj ⳱ Cj(sj; sj1, . . ., sjK). K ranges from 0 to N ⳮ 1. N reflects the reality that managers must make numerous decisions. They must decide, for instance, whether to make an input internally or purchase it in the marketplace; whether to install flexible production equipment or commit to specialized machinery; whether to manufacture a ORGANIZATION SCIENCE /Vol. 12, No. 3, May–June 2001 JAN W. RIVKIN Reproducing Knowledge Table 2 Simulation Parameters and Symbols Downloaded from informs.org by [2.37.204.136] on 24 July 2023, at 09:04 . For personal use only, all rights reserved. Parameters N K r i hf Symbols s S Cj P(s) s* s† The number of decisions which the modeled firm must make. All decisions are binary (0/1) The number of decisions, other than the focal decision, whose resolution (0 or 1) affects the contribution of the focal decision to firm payoff. Ranges from 0 to N ⳮ 1 The number of decisions in common between the replicator’s initial s and the benchmark s* The number of decisions in common between the imitator’s initial s and the benchmark s* The probability that, in a long jump, firm f successfully matches s*j , an individual element of s*. f ⳱ R for the replicator and I for the imitator An N-digit string of zeroes and ones, representing a full set of choices The space of all possible choices The contribution of decision j to firm payoff. For each resolution of decision j and the K randomly-assigneddecisions affecting decision j, Cj is drawn at random from a uniform distribution between 0 and 1 The payoff of a firm which adopts decisions s. An average over the contributions of the N decisions The combination of choices which yields the highest possible payoff The combination of choices which yields the second-highest possible payoff diverse product line or focus on a handful of products; whether to organize by function or by product, etc. The firm’s strategy is embodied in this nexus of choices. It may seem odd to represent a firm’s strategy as a string of N zeroes and ones; such a string feels linear while strategies are holistic, with component decisions intertwined. Adding connections across decisions is precisely the role of K, the second parameter in the model. K captures the fact that the choice made concerning one decision may affect the marginal benefit or cost associated with another decision.11 For instance, investments in machines that permit production of a complex product line are made more valuable on the margin if a firm also employs a highly trained salesforce that can promote the product line effectively. The marginal value of product inspection activities is reduced by manufacturing activities that eliminate defects in the production process. The marginal cost of a no-questions-asked merchandise return policy is lowered for a retailer whose logistics system is configured to handle returns easily. A number of recent studies document such interactions.12 Choices often complement one another or substitute for each other. As described below, these interactions constrain search processes in important ways. A payoff is assigned to each of the 2N combinations of decisions as follows. Recall that the contribution of each decision to overall firm value, Cj, is affected by K other randomly assigned decisions. For each possible realization of (sj; sj1, . . ., sjK), a contribution Cj is drawn at random from a uniform distribution between 0 and 1. The overall payoff associated with a configuration is the average over the N contributions: N P(s) ⳱ 冤 兺 C (s ; s , . . . , s )冥冫N. j j j1 jK j⳱1 ORGANIZATION SCIENCE /Vol. 12, No. 3, May–June 2001 Note that when K ⳱ 0, P(s) is the average over N contributions, each of which depends only on a single choice. At the other extreme, when K ⳱ N ⳮ 1, P(s) is the average over N contributions, each of which is affected by all N choices. 3.2. Managerial Search Once the modeler fixes N and K and the computer generates a particular decision problem (that is, a payoff P(s) for each of the 2N possible decision configurations), a modeled replicator and imitator search for good solutions to the problem. Firm R, the replicator, is assumed to have discovered a very effective solution, a high point on the landscape, in a previous period. This solution serves as a target for the replicator and the imitator. In particular, I assume that the replicator managed previously to attain the very best solution to the decision problem, the global peak, s*. (The computer finds this optimum by exhaustively searching the 2N possible configurations.) For all the reasons discussed above, the replicator cannot simply write down the combination of decisions that constitutes s*. Rather, the firm has only imperfect knowledge of the decisions that underpin the original success. Starting with this knowledge, it must search to rediscover the winning formula. At the same time, an imitating Firm I begins its search to match the success. An important aspect of the decision problems generated by this version of the NK approach is that they are difficult to solve using any optimization algorithm. Formally, for K ⬎ 2, these NK problems are NP-complete. That is, there exists no algorithm that can find the global optimum for such decision problems in polynomial time. (See Rivkin 2000 for further explanation.) For this reason, it is sensible to assume that replicating or imitating managers use search heuristics—not some algorithm—to 281 Downloaded from informs.org by [2.37.204.136] on 24 July 2023, at 09:04 . For personal use only, all rights reserved. JAN W. RIVKIN Reproducing Knowledge (re)discover successful combinations of choices. I explore two polar types of heuristic: incremental improvement and follow-the-leader long jumps. When engaging in incremental improvement, a firm considers changing only one decision at a time. Starting at configuration s, the firm examines all N of the alternatives which involve changing just one decision. It accepts a considered alternative sⴕ if P(sⴕ) ⬎ P(s). Having adopted sⴕ, it then considers another incremental change, and so forth until the firm reaches a set of decisions which offers no more incremental improvement. Note that such a set may very well constitute a local peak, not the global optimum. When taking a follow-the-leader long jump, the firm at s attempts a major reconfiguration toward the target s*. That is, it tries to adjust each of its N decisions to match the original success. No longer does it change only one decision at a time. Because the firm (either replicator or imitator) cannot discern the precise configuration of the success, the probability that each decision is matched correctly is h ⱕ 1. hN, then, is the probability that a firm leaps to precisely s*. In the simulations below, I sometimes couple incremental improvement and long jumps: A firm moves incrementally uphill, then leaps toward s*, creeps uphill again, and so forth. Because it has access to a template, the replicator has an advantage over the imitator during the search process. I consider three specific benefits. First, the replicator may initiate its search in closer proximity to s*. It resolves r of its decisions in the same way as s* at the outset of its search, while the imitator shares only i with s*. N ⱖ r ⬎ i ⱖ 0. This advantage reflects the fact that the replicator’s initial attempts at transfer will be more accurate than the imitator’s. Second, h may be higher for the replicator than it is for the imitator. This captures the ability of the replicator to examine the template closely and turn to experienced personnel as it tries to improve its operations. Third, when taking a long jump, the replicator may know which of its decisions are already accurately replicated; the imitator, in contrast, might not know what it is doing “right” and what it is doing “wrong.” The template guides the replicator to focus on the right issues. In the simulation, then, the replicator is simply endowed with a search advantage over the imitator. Clearly, this simplifies a much more complicated process in which both replicator and imitator decide how much to “invest” in understanding the original success. My presumption is that, on the margin, information about the template is less costly to the replicator and, accordingly, the replicating firm emerges with an information advantage. I return to this issue—that r, i, and h may be endogenous to some degree—in the final section. After the replicator and imitator complete their search 282 efforts, the computer records the success of each—both the performance level P(s) that each attains and the similarity between the final s and the target s*. The computer then generates a second, altogether new decision problem. This new problem reflects the same levels of N and K, but is based on a new set of random draws for the Cjs and a new set of particular interactions. A second replicator and second imitator tackle the problem, and their relative success is again recorded. This process continues until 100 to 200 decision problems, all with identical levels of complexity, have been explored. The modeler then adjusts N and K to alter the level of complexity; examines replicator and imitator success on 100–200 decision problems with the new level of complexity; alters N and K again; and so forth. In this way, one gradually isolates the effect of complexity on the gap between replication and imitation. Note that a firm’s payoff is determined solely by its own decisions. I permit no interplay at all between the payoff of the replicator and that of the imitator. For instance, I do not penalize the firms for adopting similar configurations of decisions. Thus, any benefit that the replicator enjoys comes from its advantages in the search process, not from some preemptive maneuver. In addition, if the replicator and the imitator happen to make the same configuration of decisions, they attain identical payoffs. Thus the replicator has no resource advantage that allows it to make more profit than the imitator can from a particular set of choices. It may seem odd that the simulation uses the global peak as the target configuration of decisions. Given the complexity of the decision problem, how could the replicator or imitator know for sure that the old installation constitutes the best possible outcome and therefore warrants duplication? In fact, the global peak is used purely for convenience. All that is necessary is that the replicator has superior information about some good (not specifically the best) resolution to the decision problem. For instance, one can alter the simulation so that the target s* is not the global peak, but the best peak discovered by some number of prior random searches. As §4.3 mentions, this alternative arrangement does not change the results qualitatively. My tack is to examine how the benefits of search advantages shift as K changes, that is, as the underlying decisions become more or less intertwined. To grasp the intuition behind the results, however, one must first understand how K affects the topography of the underlying landscape. 3.3. Interactions and the Landscape A key insight is that changes in interactivity (K) fundamentally alter the character of the typical landscape. ORGANIZATION SCIENCE /Vol. 12, No. 3, May–June 2001 Downloaded from informs.org by [2.37.204.136] on 24 July 2023, at 09:04 . For personal use only, all rights reserved. JAN W. RIVKIN Reproducing Knowledge Kauffman (1993) explores the effects of interactions in depth, and Levinthal (1997) and Rivkin (2000) do the same in the context of business strategy. Here, I summarize three germane effects. First, as K increases, the landscape shifts from being smooth and single-peaked to being rugged with numerous local peaks. The extreme cases of K ⳱ 0 and K ⳱ N ⳮ 1 illustrate this point. When K ⳱ 0, the N decisions make independent contributions to firm payoff. In that situation, alteration of a single decision changes the contribution of that decision alone. From any initial location on a landscape, then, a replicating or imitating firm can rise to the global peak via a series of value-improving, singledecision “tweaks” to its strategy. Management simply adjusts each choice, one by one, to the option that allows that decision to make the greater contribution in isolation. In contrast, when K ⳱ N ⳮ 1, every decision influences the contribution of every other choice. Then a small step on the landscape—a change in a single decision—alters the contributions of all N elements. Consequently, adjacent decision configurations have elevations that are altogether uncorrelated with one another. A particular configuration is a local peak if all N of its neighbors are lower than itself, and because elevations are random when K ⳱ N ⳮ 1, this occurs with probability 1/(N Ⳮ 1). Of the 2N possible strategies, 2N/(N Ⳮ 1) are expected to be local peaks. This number is large for even modest N, surpassing 100 for N ⳱ 11 and 10,000 for N ⳱ 18. More generally, as K increases, the expected correlation between P(s) and P(sⴕ)—where s and sⴕ differ in how one decision is resolved—falls toward zero. As a consequence, the number of local peaks rises. Each of these local peaks is internally consistent, in the sense that, starting from such a locale, no tweak to a single decision can boost performance. “Internally consistent,” however, is not the same as profitable, and this highlights the second effect of interactions. As K rises, not only do local peaks become more numerous, but the height of the average peak declines. As the web of connections among decisions thickens, it becomes possible for a firm to run out of opportunities for incremental improvement even at low levels of performance. When K is high, a firm can easily become trapped in an internally consistent but unfavorable set of choices. A manager in charge of, say, one of the choices has no incentive to break the organizational truce (Nelson and Winter 1982) even though an altogether different arrangement might lead to better performance. Interactions, then, constrain the search for incremental improvement. Finally, as interactions proliferate, the high peaks on the typical landscape spread apart from one another “horizontally.” On a landscape with a low level of K, high ORGANIZATION SCIENCE /Vol. 12, No. 3, May–June 2001 peaks usually arise near each other, in “mountain ranges;” for instance, if one examines the decision configuration s† that constitutes the second-highest peak, the vector {s†1, s†2, . . ., s†N} will closely resemble the global peak s* ⳱ {s*, On a landscape with high K, in 1 s*, 2 . . ., s*}. N contrast, favorable sets of decisions are often far apart; s† may look little like s*. The intuition behind this is straightforward. When choices are loosely connected, the strategy of a successful firm can typically be decomposed into separable subsystems of well-made choices (Simon 1962). Reconfiguring one of the subsystems has limited effect on the others and generates a new local peak that is quite close in decision space to the original one. Starting with a very good set of choices, one might, for example, rearrange just the marketing choices while leaving production, R&D, and logistics decisions unchanged, and this could lead to a second, internally consistent, effective configuration. In contrast, when choices are tightly interlocked, the strategy of a successful firm cannot be broken into subsystems. High peaks are then likely to be altogether different ways of doing business rather than slight variations on a theme. In such a setting, a change in the marketing function, for instance, will necessitate adjustments throughout the entire value chain. In sum, then, interactions reshape the landscapes on which replicators and imitators search. Simulations show how these effects alter the value of the replicator’s template, and hence the advantage which the replicator enjoys. 4. Simulation Results In this section, I examine the replicator’s advantage when searches proceed by incremental improvement and by follow-the-leader long jumps. The pattern that emerges is this: Under a wide range of assumptions, the wedge between replicability and imitability is widest at moderate levels of K, i.e., at intermediate degrees of complexity. 4.1. Incremental Improvement First, suppose that the replicator and imitator can (re)discover the best solution only by means of incremental search. The replicator is at an advantage because it begins its search closer to the benchmark configuration of decisions; r ⬎ i. The size of the benefit that derives from this advantage, however, depends on K. The four panels of Table 3 show this dependence on K for various combinations of r and i. (N ⳱ 12 throughout this section, and reported results are averages over 200 landscapes. Results focus on the final performance attained by each firm, but results for interim performance are qualitatively similar; see §4.3 below.) When K ⳱ 0, 283 Downloaded from informs.org by [2.37.204.136] on 24 July 2023, at 09:04 . For personal use only, all rights reserved. JAN W. RIVKIN Reproducing Knowledge the replicator has no advantage at all. The landscape is smooth, and both firms discover the global peak eventually. As K rises a little, a gap emerges between the replicator’s performance and the imitator’s. The imitator gets stranded on local peaks far from the best strategy while Firm R comes close to complete replication. As K approaches N, however, the replicator starts to founder as well. Like the imitator, it gets trapped on local peaks away from the benchmark. Its local peak is closer to the benchmark than the imitator’s peak (Column 3 vs. Column 5), but because the high peaks are spread apart, it does little or no better than the imitator. As Table 4 shows, the replicator’s advantage is often largest at levels of K well below the maximum value of N ⳮ 1. Simple decision problems stymie neither firm, and highly complex problems resist both imitation and replication. Decision problems with moderately intertwined choices, however, make the replicator’s template especially valuable. Only when the replicator can reproduce the original success with very high fidelity (r close to N) is the replicator best off with a maximally intertwined set of choices. For a fixed level of i, the following pattern emerges: The greater is the replicator’s advantage with respect to the template (r ⳮ i), the more complex the replicator would like the underlying set of decisions to be. 4.2. Follow-the-Leader Long Jumps Similar results emerge when the two firms start with the same initial transfer (r ⳱ i), but leap toward the benchmark with different degrees of accuracy (hR ⬎ hI). The intuition behind the similarity is straightforward. Consider a model in which the firms start equally far from the target (r ⳱ i ⳱ 0), leap toward the original success with differential accuracy (hR ⬎ hI), and then inch uphill to a local peak. Results for such a model are identical to one with r ⳱ hR N ⬎ i ⳱ hI N. Table 5 illustrates the similarity for two combinations of hR and hI. (Results here and below are averages over 100 landscapes.) In leaping toward the benchmark, Firm R may enjoy an advantage beyond greater accuracy. Because it has access to a template, the replicator may be better able to identify which of its decisions differ from the benchmark and which are already accurately duplicated. When decisions are tightly connected to one another, distinguishing what one is doing “right” from what one is doing “wrong” is no trivial matter. The ability to tell right from wrong may give the replicator a substantial edge. Table 6 illustrates this effect. In the simulation shown there, the replicator and imitator start out with equally poor replicas (r ⳱ i ⳱ 0) and the two are equally inac284 curate in their leaps (hR ⳱ hI ⳱ 0.5). Each firm tweaks uphill, takes a leap, then moves uphill again to a (possibly local) peak. In making its leap, however, only the replicator knows which choices are identical to the benchmark. The replicator tries to alter only the decisions that are misaligned with s*, while the imitator experiments with all N choices. The results show a familiar pattern. When K ⳱ 0, both replication and imitation are thoroughly successful. When K ⳱ N ⳮ 1, both efforts are stymied by the complexity of the task. The replicator’s advantage is greatest at an intermediate value of K well below the maximum of N ⳮ 1. 4.3. Robustness of the Result The main result, that the replicator’s advantage over the imitator is greatest at an intermediate level of interaction, holds up over a broad range of assumptions. Results available from the author show that this pattern remains when N is higher or lower than the value explored here; when the benchmark s* is set by taking the best of numerous search results rather than placing it atop the global peak; when each choice offers more than two options; when contributions Cj are drawn from distributions other than the uniform distribution assumed above; and when the assumption that all decisions are ex ante equally influential is altered. The findings reported so far focus on the final performance of each firm, after it has completed its search. If search efforts stretch over a long time, however, it may be important to consider performance during the entire process, not just in the final period. Imitators initiate their search efforts farther from the global peak than replicators do. Consequently, on smooth landscapes (low K), imitators wander in the “lowlands” longer than replicators do. This effect will widen the gap between replicator performance and imitator performance at low K, but will have a less pronounced effect on the gap at high K, when local peaks are numerous and searches brief. At least conceptually, this could create a situation in which the performance gap is widest at K ⳱ 0—contrary to the paper’s main hypothesis. To test this possibility, I examined performance over the entire search process and discounted the payoffs back to Period 0 using a discount rate as high as 100%. This reduced the degree of complexity at which the performance differential is greatest, as expected, but not to K ⳱ 0. The results of virtually any simulation model can be overturned if one changes one’s assumptions radically enough. The main result fails to hold under two alternative sets of assumptions. First, if Firm R is able to replicate its initial success perfectly (r ⳱ N) but can keep its ORGANIZATION SCIENCE /Vol. 12, No. 3, May–June 2001 Downloaded from informs.org by [2.37.204.136] on 24 July 2023, at 09:04 . For personal use only, all rights reserved. JAN W. RIVKIN Reproducing Knowledge Table 3 Results for Incremental Improvement Description: In Panel A, the replicator initially shares eight decisions with the benchmark set of decisions (r ⳱ 8) and the imitator shares two (i ⳱ 2). Both firms are released on a landscape, and each is allowed to make incremental improvements until arriving at a (possibly local) peak. N ⳱ 12 and K varies. For each level of K, the panel reports the portion of the benchmark performance eventually attained by each firm and the “distance” of the final location from the benchmark (i.e., the number of decisions not resolved in identical fashion). The reported figures are the average (and the standard deviation) over 200 landscapes. The significance level of a test that the replicator’s average and the imitator’s average do not differ is also reported. The level of K that leads to the greatest performance differential is shown in bold. Panels B, C, and D report results for other values of r and i. Panel A: r ⳱ 8, i ⳱ 2 Replicator performance Imitator performance K Percent of P(s*) attained Final distance from s* Percent of P(s*) attained Final distance from s* Performance differential 0 1 2 3 4 5 6 7 8 9 10 11 100.0 (0.0) 99.3 (1.8) 97.8 (3.5) 96.3 (4.6) 93.8 (5.3) 92.4 (6.3) 90.0 (5.9) 88.0 (5.7) 87.6 (5.8) 86.0 (5.7) 85.3 (5.5) 81.1 (4.6) 0.0 (0.0) 0.7 (1.5) 1.8 (2.2) 2.4 (2.4) 3.6 (2.3) 3.8 (2.3) 4.3 (1.9) 4.6 (1.6) 4.6 (1.6) 4.5 (1.3) 4.5 (1.2) 4.4 (1.2) 100.0 (0.0) 95.2 (4.3) 91.8 (5.6) 90.6 (5.7) 89.8 (6.0) 89.2 (5.5) 88.5 (6.0) 87.4 (5.5) 87.0 (5.7) 85.9 (5.7) 84.9 (5.4) 81.0 (4.6) 0.0 (0.0) 4.0 (2.6) 6.9 (2.2) 7.6 (1.6) 8.1 (1.4) 8.1 (1.3) 8.4 (1.3) 8.6 (1.3) 8.9 (1.3) 8.9 (1.1) 8.8 (1.1) 9.1 (1.0) 0.0 4.1 *** 6.1 *** 5.7 *** 3.9 *** 3.2 *** 1.4 *** 0.6 0.6 0.1 0.5 0.2 Panel B: r ⳱ 10, i ⳱ 6 Replicator performance Imitator performance K Percent of P(s*) attained Final distance from s* Percent of P(s*) attained Final distance from s* Performance differential 0 1 2 3 4 5 6 7 8 9 10 11 100.0 (0.0) 99.9 (0.7) 99.8 (1.0) 99.3 (2.4) 99.0 (2.7) 98.4 (4.0) 96.8 (5.2) 94.9 (6.6) 93.7 (7.5) 90.5 (8.0) 89.2 (8.3) 84.9 (7.5) 0.0 (0.0) 0.1 (0.5) 0.1 (0.6) 0.3 (1.0) 0.6 (1.4) 0.6 (1.4) 1.2 (1.8) 1.5 (1.8) 1.9 (1.9) 2.3 (1.8) 2.4 (1.7) 2.8 (1.4) 100.0 (0.0) 98.1 (2.9) 95.1 (4.9) 92.6 (5.4) 90.4 (6.5) 89.5 (5.8) 88.8 (5.7) 87.0 (5.7) 86.4 (5.7) 85.7 (6.0) 84.9 (5.8) 81.6 (4.4) 0.0 (0.0) 1.5 (1.7) 3.5 (2.5) 4.8 (2.5) 5.6 (1.9) 5.9 (1.8) 5.9 (1.6) 6.2 (1.4) 6.0 (1.4) 5.9 (1.3) 5.9 (1.2) 6.0 (1.2) 0.0 1.8 *** 4.7 *** 6.8 *** 8.6 *** 8.9 *** 8.0 *** 8.0 *** 7.3 *** 4.8 *** 4.3 *** 3.3 *** *** Performance differential significant at the 1% level. perfect knowledge to itself (r ⬎ i), then its advantage simply grows larger and larger as the complexity of the underlying problem increases. Greater complexity then undermines the efforts of the imitator, but does no harm to the replicator. Second, if the replicator and imitator ORGANIZATION SCIENCE /Vol. 12, No. 3, May–June 2001 have equal access to and knowledge of the template (i.e., r ⳱ i, hR ⳱ hI, and the firms have equal knowledge of what they are doing right), then the replicator enjoys no advantage regardless of the degree of complexity. Neither of these alternatives seems likely to arise often in reality. 285 JAN W. RIVKIN Reproducing Knowledge Table 3 (cont.) Results for Incremental Improvement Panel C: r ⳱ 6, i ⳱ 0 Downloaded from informs.org by [2.37.204.136] on 24 July 2023, at 09:04 . For personal use only, all rights reserved. Replicator performance Imitator performance K Percent of P(s*) attained Final distance from s* Percent of P(s*) attained Final distance from s* Performance differential 0 1 2 3 4 5 6 7 8 9 10 11 100.0 (0.0) 98.1 (2.9) 95.1 (4.9) 92.6 (5.4) 90.4 (6.5) 89.5 (5.8) 88.8 (5.7) 87.0 (5.7) 86.4 (5.7) 85.7 (6.0) 84.9 (5.8) 81.6 (4.4) 0.0 (0.0) 1.5 (1.7) 3.5 (2.5) 4.8 (2.5) 5.6 (1.9) 5.9 (1.8) 5.9 (1.6) 6.2 (1.4) 6.0 (1.4) 5.9 (1.3) 5.9 (1.2) 6.0 (1.2) 100.0 (0.0) 93.4 (5.8) 91.3 (5.6) 90.3 (5.6) 90.4 (5.3) 89.1 (6.0) 88.5 (5.9) 87.4 (6.2) 86.4 (5.6) 85.9 (5.7) 85.2 (5.9) 81.5 (4.1) 0.0 (0.0) 4.6 (2.8) 7.6 (2.0) 8.5 (1.4) 8.9 (1.3) 9.3 (1.2) 9.6 (1.1) 9.7 (1.1) 10.1 (1.1) 10.2 (1.0) 10.3 (1.0) 10.5 (0.9) 0.0 4.7 *** 3.8 *** 2.2 *** ⳮ0.1 0.4 0.3 ⳮ0.4 0.0 ⳮ0.2 ⳮ0.3 0.0 Panel D: r ⳱ 10, i ⳱ 2 Replicator performance Imitator performance K Percent of P(s*) attained Final distance from s* Percent of P(s*) attained Final distance from s* 0 1 2 3 4 5 6 7 8 9 10 11 100.0 (0.0) 99.9 (0.7) 99.8 (1.0) 99.3 (2.4) 99.0 (2.7) 98.4 (4.0) 96.8 (5.2) 94.9 (6.6) 93.7 (7.5) 90.5 (8.0) 89.2 (8.3) 84.9 (7.5) 0.0 (0.0) 0.1 (0.5) 0.1 (0.6) 0.3 (1.0) 0.6 (1.4) 0.6 (1.4) 1.2 (1.8) 1.5 (1.8) 1.9 (1.9) 2.3 (1.8) 2.4 (1.7) 2.8 (1.4) 100.0 (0.0) 95.2 (4.3) 91.8 (5.6) 90.6 (5.7) 89.8 (6.0) 89.2 (5.5) 88.5 (6.0) 87.4 (5.5) 87.0 (5.7) 85.9 (5.7) 84.9 (5.4) 81.0 (4.6) 0.0 (0.0) 4.0 (2.6) 6.9 (2.2) 7.6 (1.6) 8.1 (1.4) 8.1 (1.3) 8.4 (1.3) 8.6 (1.3) 8.9 (1.3) 8.9 (1.1) 8.8 (1.1) 9.1 (1.0) Performance differential 0.0 4.7 *** 8.1 *** 8.7 *** 9.1 *** 9.2 *** 8.3 *** 7.6 *** 6.7 *** 4.6 *** 4.3 *** 3.9 *** ***Performance differential significant at the 1% level 5. Conclusions This paper takes up a deceptively simple question: What distinguishes situations in which replication and imitation are equally easy, those in which the two are equally hard, and those in which the replicator fares better than the imitator? Why do the situations of, say, White Castle, Marks & Spencer, and Southwest Airlines differ? I argue that strategic, resource-based, and organizational perspectives—as important as they are—do not shed bright light on this particular question,13 and I focus instead on managerial search for knowledge that is difficult to trans- 286 fer. The particular impediment to search which I examine in depth is the complexity of the decision problem facing both replicator and imitator. One might initially suspect “the more complex the better” from the point of view of the firm which has the original success. And indeed, decision problems which are more complex do resist imitation more staunchly (Rivkin 2000). Unfortunately, they also pose a stiffer challenge to replication. The replicator’s relative advantage is greatest with respect to decision problems of intermediate complexity. It is here that the replicator’s ORGANIZATION SCIENCE /Vol. 12, No. 3, May–June 2001 JAN W. RIVKIN Reproducing Knowledge Interaction Level that Maximizes Replicator’s Relative Advantage Description: For each level of r and i, the table reports the level of K that generates the widest gap between the replicator’s performance and the imitator’s performance. Performance is indicated by the portion of P(s*) attained in simulations similar to the ones described in Table 3. Downloaded from informs.org by [2.37.204.136] on 24 July 2023, at 09:04 . For personal use only, all rights reserved. Table 4 i⳱ i⳱ i⳱ i⳱ i⳱ i⳱ 0 2 4 6 8 10 r⳱2 r⳱4 r⳱6 r⳱8 r ⳱ 10 r ⳱ 12 1 1 1 1 2 2 2 2 2 3 5 5 4 5 7 11 11 11 11 11 11 informational edge, which comes from its preferential access to the original success, confers the greatest benefit. In such a regime, landscapes are rugged enough that imitating firms with poor information get trapped on low and distant local peaks. Landscapes are smooth enough that perfect information about the benchmark is not necessary for acceptable replication. And high peaks cluster together, implying that even imperfect information about the location of the benchmark gives a valuable clue as to where other high peaks may lie. In essence, this paper examines the value of superior yet imperfect information about good solutions to complex problems. Facing simple problems, both replicator and imitator generate good information through their incremental search efforts. The replicator’s initial informational advantage does not last for long or do it much good. At high levels of complexity, when decisions depend on one another in an extremely delicate way, small errors in information spoil replication attempts altogether. The replicator’s slightly imperfect knowledge then has little more value than the imitator’s highly imperfect knowledge. At moderate levels of complexity, the replicator’s initial information guides it to a good set of decisions, and the imitator does not recreate the information through its ordinary search efforts. Viewed in this way, the results of the paper have general implications for scholars who focus on the value and management of knowledge. A growing number of management scholars contend that knowledge has become a decisive competitive resource (e.g., Grant and Spender 1996, Grant 1996, Nonaka and Takeuchi 1995). Indeed, knowledge is sometimes held up as the ultimate leveragable resource: Knowledge is often difficult for outsiders to observe and copy, and it is enhanced, not con- ORGANIZATION SCIENCE /Vol. 12, No. 3, May–June 2001 sumed, in use. The results developed here suggest that some kinds of knowledge are more likely than others to serve as a source of sustained advantage. Specifically, knowledge that is moderately complex lends itself to being replicated without being imitated. There are doubtlessly other dimensions and characteristics that delineate competitively useful knowledge from less valuable knowledge. This paper focuses on just one aspect of what makes knowledge a potential source of advantage. Though the discussion has focused on the replication and imitation of entire business strategies, the core argument concerning the value of superior yet imperfect information also applies to productive systems of smaller scale. Consider, for instance, a firm with a single, highly productive factory. If economies of scale are modest, transportation costs high, demand geographically dispersed, and raw inputs available at multiple locales, then the firm may wish to build replicas of the factory in numerous locations. In such a setting, the arguments of this paper should apply: successful replication without imitation would be most likely when choices concerning the design and operation of the template factory are modestly intertwined. Since the inner workings of a factory are often difficult for an outsider to deduce, the advantage that the replicator gets from superior access to the template might be especially great in such a context. The analysis in the paper suggests three hypotheses that could be tested in real settings. First, replicators will have the greatest advantage over imitators when underlying decisions are neither tightly intertwined nor unrelated. A business strategy that requires replication (or, for that matter, any productive item of knowledge that requires replication) is most likely to generate sustained competitive advantage if it is of moderate complexity. The second hypothesis is that replication and imitation will both ensue quickly when decisions are largely unrelated and that neither will succeed (or perhaps even be attempted) when decisions are very intensely connected. Under this interpretation, one of White Castle’s problems was that the elements of its strategy were only loosely connected to one another; the turrets had little to do with the limited menu, and both were unconnected to the emphasis on hygiene. (The fact that its strategy was easily observed, i.e., i ⬇ r ⬇ N, also played a role in its difficulties, of course.) Similarly, the intricacy of Marks and Spencer’s original British strategy made it difficult to reproduce abroad. The third and final hypothesis comes from the analysis of Table 4. For a given level of i, the degree of complexity that maximizes the replicator’s advantage rises in line with the replicator’s informational edge (r vs. i). Put differently, the greater is a firm’s informational superiority, the more complex it would like its strategy 287 Downloaded from informs.org by [2.37.204.136] on 24 July 2023, at 09:04 . For personal use only, all rights reserved. JAN W. RIVKIN Reproducing Knowledge Table 5 Results for Follow-the-Leader Long Jumps with Different Degrees of Accuracy Description: A replicator and an imitator are released on a landscape with equally poor initial positions (r ⳱ i ⳱ 0). Each firm leaps toward the benchmark set of decisions and then makes incremental improvements until arriving at a (possibly local) peak. The replicator leaps toward the benchmark with greater accuracy than does the imitator; hR ⳱ 0.8 and hI ⳱ 0.2 in Panel A, and hR ⳱ 0.6 and hI ⳱ 0.4 in Panel B. N ⳱ 12 and K varies. For each level of K, the panel reports the portion of the benchmark performance eventually attained by each firm and the “distance” of the final location from the benchmark (i.e., the number of decisions not resolved in identical fashion). The reported figures are the average (and the standard deviation) over 100 landscapes. The significance level of a test that the replicator’s average and the imitator’s average do not differ is also reported. The level of K that leads to the greatest performance differential is shown in bold. Panel A: hR ⳱ 0.8, hI ⳱ 0.2 Replicator performance Imitator performance K Percent of P(s*) attained Final distance from s* Percent of P(s*) attained Final distance from s* Performance differential 0 1 2 3 4 5 6 7 8 9 10 11 100.0 (0.0) 99.6 (1.5) 99.4 (2.1) 98.5 (3.5) 97.7 (4.0) 97.1 (5.1) 94.3 (7.3) 93.3 (7.4) 92.8 (7.6) 91.6 (8.4) 90.4 (8.6) 86.6 (8.7) 0.0 (0.0) 0.3 (1.0) 0.5 (1.6) 0.9 (1.8) 1.3 (2.0) 1.6 (2.5) 2.0 (2.4) 2.6 (2.6) 2.5 (2.3) 2.5 (2.4) 2.7 (2.4) 3.0 (2.0) 100.0 (0.0) 95.3 (5.0) 92.7 (5.7) 91.3 (5.4) 90.2 (6.8) 89.0 (5.5) 88.0 (5.2) 87.0 (5.6) 87.0 (5.9) 86.1 (5.9) 84.4 (4.7) 80.8 (4.1) 0.0 (0.0) 3.6 (2.5) 6.6 (2.6) 7.5 (1.8) 7.6 (1.8) 8.1 (1.5) 8.3 (1.5) 8.3 (1.5) 8.8 (1.5) 8.6 (1.6) 8.4 (1.2) 8.5 (1.5) 0.0 4.4 *** 7.1 *** 7.2 *** 7.5 *** 8.1 *** 6.3 *** 6.3 *** 5.8 *** 5.5 *** 6.0 *** 5.8 *** Panel B: hR ⳱ 0.6, hI ⳱ 0.4 Replicator performance Imitator performance K Percent of P(s*) attained Final distance from s* Percent of P(s*) attained Final distance from s* Performance differential 0 1 2 3 4 5 6 7 8 9 10 11 100.0 (0.0) 98.5 (2.4) 97.2 (4.4) 94.7 (5.5) 92.6 (6.9) 91.0 (7.0) 91.3 (6.0) 88.4 (7.0) 88.3 (6.2) 86.6 (6.3) 86.5 (6.4) 83.1 (7.3) 0.0 (0.0) 1.4 (2.0) 2.3 (2.8) 3.6 (2.9) 4.0 (2.7) 4.3 (2.7) 4.7 (2.6) 4.8 (2.3) 4.9 (2.0) 5.1 (2.1) 5.1 (2.0) 4.9 (2.0) 100.0 (0.0) 96.7 (4.6) 93.4 (6.4) 91.7 (6.1) 90.7 (6.0) 89.3 (5.5) 89.1 (5.7) 87.1 (5.4) 87.6 (5.3) 86.0 (5.3) 85.7 (5.2) 81.6 (4.6) 0.0 (0.0) 2.8 (2.7) 5.0 (2.9) 5.4 (2.5) 6.3 (2.3) 6.7 (1.7) 6.7 (1.9) 6.7 (1.7) 7.0 (1.8) 6.8 (1.6) 6.7 (1.7) 7.0 (1.9) 0.0 1.8 *** 3.8 *** 3.1 *** 1.9 ** 1.7 ** 2.2 *** 1.3 * 0.7 0.6 0.8 1.4 * ***Performance differential significant at the 1% level **Performance differential significant at the 5% level *Performance differential significant at the 10% level to be (assuming that its goal is to perform better than its competitors). All three hypotheses can be tested in principle, though the practical challenges of performing the tests should not be discounted.14 The analysis offers an alternative interpretation for a 288 prevalent observation in organization theory. So-called “loosely coupled” organizations, with only modest connections among decision-making elements, are widely said to be especially effective competitors (Weick 1976, Page-Jones 1980, pp. 101–103, Peters and Waterman ORGANIZATION SCIENCE /Vol. 12, No. 3, May–June 2001 Downloaded from informs.org by [2.37.204.136] on 24 July 2023, at 09:04 . For personal use only, all rights reserved. JAN W. RIVKIN Reproducing Knowledge Table 6 Results for Follow-the-Leader Long Jumps with Different Knowledge of What Is “Right” Description: A replicator and an imitator are released on a landscape with equally poor initial positions (r ⳱ i ⳱ 0). Each firm tweaks uphill to a peak, leaps toward the benchmark set of decisions, and then makes incremental improvements until arriving at a peak. The two firms are equally accurate in their leaps (hR ⳱ hI ⳱ 0.5). The replicator knows which of its decisions are different from the benchmark and, in its leap, tries to modify only those decisions. In contrast, the imitator tries to modify all of its decisions in its long jump. N ⳱ 12 and K varies. For each level of K, the panel reports the portion of the benchmark performance eventually attained by each firm and the “distance” of the final location from the benchmark (i.e., the number of decisions not resolved in identical fashion). The reported figures are the average (and the standard deviation) over 100 landscapes. The significance level of a test that the replicator’s average and the imitator’s average do not differ is also reported. The level of K that leads to the greatest performance differential is shown in bold. Replicator performance Imitator performance K Percent of P(s*) attained Final distance from s* Percent of P(s*) attained Final distance from s* Performance differential 0 1 2 3 4 5 6 7 8 9 10 11 100.0 (0.0) 98.4 (2.4) 96.8 (3.7) 94.9 (5.3) 93.0 (5.5) 92.0 (6.2) 89.6 (6.3) 88.4 (5.6) 86.9 (5.9) 85.2 (6.4) 85.6 (5.6) 82.7 (5.9) 0.0 (0.0) 1.7 (1.9) 3.0 (2.8) 3.9 (3.2) 4.4 (2.9) 4.4 (2.9) 5.3 (2.3) 5.3 (2.3) 5.3 (1.8) 5.5 (1.9) 5.3 (1.5) 5.3 (1.6) 100.0 (0.0) 97.9 (3.5) 95.9 (4.5) 93.9 (5.3) 91.0 (6.5) 90.2 (6.4) 89.6 (6.1) 88.7 (5.5) 86.8 (5.9) 84.5 (6.2) 85.4 (5.9) 81.9 (6.0) 0.0 (0.0) 1.9 (2.1) 3.6 (3.0) 4.2 (2.7) 5.1 (2.6) 5.6 (2.3) 5.8 (2.0) 5.7 (2.1) 6.0 (1.9) 6.2 (1.7) 6.2 (1.7) 5.9 (1.7) 0.0 0.6 * 1.0 ** 1.0 * 2.0 *** 1.7 ** ⳮ0.1 ⳮ0.3 0.1 0.7 0.2 0.7 ***Performance differential significant at the 1% level **Performance differential significant at the 5% level *Performance differential significant at the 10% level 1982, pp. 107–109, 306–325, Perrow 1984, pp. 62–100, Sanchez and Mahoney 1996, Brown and Eisenhardt 1998). The traditional interpretation is that such organizations balance the benefits of centralization and decentralization, adapt to changes in local conditions, and seal off internal problems before they spread (Orton and Weick 1990). The normative implication is that firms should configure themselves to be loosely coupled.15 An alternative rendering, however, begins with the observation that environments differ in the degree to which they demand coupling. Some settings present highly intertwined decisions while others pose uncoupled choices. It happens—for the reasons explored in this paper—that the widest gaps between firms, between replicators and imitators, arise in settings where decisions are inherently loosely coupled (intermediate K). For this reason, firms with particularly large performance advantages over their industry peers might exhibit loose coupling. This association arises not because loose coupling leads to superior profitability, but because the settings that encourage loose ORGANIZATION SCIENCE /Vol. 12, No. 3, May–June 2001 coupling also generate wide gaps between the best and worst performers. In such a scenario, the best (relative) performers are not those that choose to be loosely coupled, but those who face moderately interactive decision problems and happen upon good sets of decisions. If this is the case, universal prescriptions in favor of loose coupling seem misguided. This alternative interpretation is simply speculative at this point, but it does suggest empirical patterns that are distinct from the traditional interpretation. Under the traditional interpretation, different firms within an industry would exhibit different degrees of coupling, with the superior performers within each industry displaying loose coupling. Under the alternative rendering, firms within an industry would exhibit similar degrees of coupling, industries would differ in the degree of coupling of the typical firm, and the broadest dispersion of profitability would arise within those industries that display loose coupling. I am aware of no existing empirical work that distinguishes between the interpretations. 289 Downloaded from informs.org by [2.37.204.136] on 24 July 2023, at 09:04 . For personal use only, all rights reserved. JAN W. RIVKIN Reproducing Knowledge The discussion of loose coupling raises two issues which are significant gaps in the current paper and major opportunities for future research: environmental change, and managerial influence on the landscape and the search process. I consider each in turn. For simplicity, my analyses have assumed a fixed environment. Replication and imitation attempts proceed on the very same landscape as the one on which the original success arose. In reality, of course, external conditions vary from one setting to the next; in landscape parlance, each time a firm attempts replication, it faces terrain that is at least slightly altered. One of the replicator’s gravest challenges is to detect the changes and respond to them appropriately. Marks & Spencer’s woes in Canada and Europe, for instance, arose in part because it reproduced its British formula too faithfully in some respects. In Canada, it located in downtown shopping areas that resembled British High Streets rather than in regional malls which were gaining popularity. It also retained traditional clothing lines and a dark decor that were popular in Britain but not in Canada or Europe (Montgomery 1994). A natural extension of the research is to explore the conditions of environmental change under which the replicator’s access to a template remains an advantage and the circumstances under which it can become a liability. How does environmental change influence the wedge between replicability and imitability? The analyses have also simplified the role of managers in ways that clarify the present paper, but depart from reality in important respects. Managers in my simple models, for instance, cannot affect the degree to which decisions are connected (K). In real situations, of course, managers spend a great deal of time, energy, and money to make or break connections (Anderson 1999). They can, for instance, make investments that sever links among decisions, thereby making a productive system more modular (Baldwin and Clark 2000, Levinthal and Warglien 1999, Schilling 2000). Similarly, managers here do not influence the information they have about past successes (r or i). In reality, managers undertake crucial efforts to preserve secrecy or to penetrate the secrets of rivals (Winter 1987). They also make trade-offs among different approaches to information gathering. Replicating managers, for instance, may opt to reproduce a success with less initial fidelity than an imitator (r ⬍ i) because they know they can turn to the template later in order to refine their efforts. Finally, the paper treats replication and imitation largely as a one-shot effort to reproduce a success, not as an ongoing effort to refine one’s knowledge of the deeper roots of success (cf., Winter and Szulanski 2000). All three of these lacunae offer promising avenues for further modeling and future research. 290 Acknowledgments For insightful comments, the author is grateful to Jeff Dyer, Nicolaj Siggelkow, Gabriel Szulanski, Sidney Winter, three anonymous referees, and seminar participants at conferences organized by INFORMS and the Academy of Management. The author also thanks Howard Brenner for exceptional computer programming assistance and the Division of Research of Harvard Business School for generous funding. Errors remain the author’s. Endnotes 1 Because the predictions of game-theoretic models are highly sensitive to assumptions that depend on the details of a competitive context, such models might be better tested through analytical case studies than through large-scale statistical work. Case studies do validate the predictions of game theory in particular settings (Ghemawat 1997). In the handful of examples of replication-without-imitation cited above, a preemptive move seems crucial in only one case: Wal-Mart apparently deterred competitors from entering its markets by committing to build “good-sized stores in little one-horse towns” (Walton and Huey 1992, Foley et al. 1996). This helps to explain why rivals did not enter towns already served by Wal-Mart. It leaves open the question of why rival discount stores in rural areas, such as Ames, did not build aggressively in other towns. 2 All of the transfers in Teece’s sample are voluntary, but roughly half are to joint venture partners or to enterprises not owned by the transferring party. 3 Moreover, note that the chemical industries which Lieberman studies are the markets in which Mansfield (1985) finds information leakage to be the slowest. 4 Mansfield et al. (1981) report that patents boost imitation costs by much more than 11% in the ethical drug market, a finding confirmed by Levin et al. (1987). 5 Codification and teaching of knowledge may be especially expensive because, in order to codify or teach, the managers who possess the knowledge must forgo other valuable activities. 6 Personal communication from Seth Lloyd, d’Arbeloff Laboratory, Department of Mechanical Engineering, Massachusetts Institute of Technology, September 11, 1998. 7 Mansfield (1985) and Levin et al. (1987) find evidence, albeit weak evidence, that this is true. 8 Zander and Kogut (1995) provide the most direct evidence. Using a survey of managers familiar with 35 Swedish innovations, they examine the factors which influence the speed of replication and the speed of imitation. Consistent with the search point of view, technologies that can be codified and communicated more easily are more rapidly replicated. The presence of rivals working on similar technology also pressures firms to replicate their successes quickly. Contrary to the search perspective, neither the characteristics of the knowledge being transferred nor the observability of the innovation influences the pace of imitation. Imitation is faster when the innovator has lost key employees and slower when the innovator pushes aggressively to improve on its own success. 9 See, for instance, Kauffman (1995), especially Chapters 9–11, Kauffman and Macready (1995), Westhoff et al. (1996), Levinthal (1997), Sorenson (1997), McKelvey (1999), Levinthal and Warglien (1999), Rivkin ORGANIZATION SCIENCE /Vol. 12, No. 3, May–June 2001 Downloaded from informs.org by [2.37.204.136] on 24 July 2023, at 09:04 . For personal use only, all rights reserved. JAN W. RIVKIN Reproducing Knowledge (2000), Gavetti and Levinthal (2000), Kauffman et al. (2000), and Fleming and Sorenson (2001). 10 The NK framework assumes that there exists some single measure of fitness which managers maximize. Anderson (1999) points out that real managers must attend to multiple goals, which makes maximization untenable. An interesting extension to this paper would modify the present model so that each combination of decisions maps to more than one vertical “payoff” axis. Managers would then cycle through different goals and optimize with respect to each in turn. Local search with respect to one goal would often draw managers away from peaks with respect to other metrics. In such a model, I speculate that firms would ordinarily not get stuck on a local peak in a static fashion, but would reach some limiting cycle of steps. 11 This is an adaptation of Milgrom and Roberts’ (1990) definition of “complementarity.” Under their definition, complementarity arises when performing more of Activity A increases the marginal benefit of doing more of Activity B. I use the term “interaction” rather than “complementarity” to encompass both complementarity and substitution between decisions. 12 See Abernathy et al. (1995) on “lean retailing” of apparel, Athey and Schmutzler (1995) on process and product flexibility and innovation, Cockburn et al. (1999) on pharmaceutical research, Hitt and Brynjolfsson (1997) on information technology investments and organizational architecture, Hwang and Weil (1996) on apparel manufacturing, Ichniowski et al. (1997) on human resource practices in the steel finishing process, Milgrom and Roberts (1990) on “modern manufacturing,” Milgrom and Roberts (1995) on Lincoln Electric, Porter (1996) on Ikea, Vanguard, and other firms, Rivkin (1995) on Toys “R” Us and Circuit City, Siggelkow (1998) on Vanguard and Liz Claiborne, and Upton (1995) on flexible manufacturing. 13 The claim is not that strategic, resource, and organizational considerations have no effect on the success of imitation and replication. Rather, the argument is that these considerations do not explain a difference between replication and imitation thoroughly. If an original success is based on some truly one-of-a-kind resource, for example, the resource constraint could bar both replication and imitation. But the constraint would not account for a gap between the ease of replication and the ease of imitation. 14 A test would have to measure the degree of interaction, detect differences in interactivity, and size up differences in the ease of replication and imitation. The work of Athey and Stern (1996) highlights the difficulty involved in measuring interactions. See Zander and Kogut (1995) for a rare successful effort to compare the speeds of replication and imitation. 15 The diverse literature on loosely coupled systems is far from unanimous in its endorsement of loose coupling. Some studies see loose coupling as a problem to be fixed. Others recommend not loose coupling per se, but a degree of coupling commensurate with external and internal demands. See Orton and Weick (1990) for an excellent survey. References Abernathy, F., J. T. Dunlop, J. H. Hammond, D. Weil. 1995. The information-integrated channel: A study of the U.S. apparel industry in transition. Brookings Papers on Economic Activity, pp. 175–246. Aghion, P., P. Bolton. 1987. Contracts as a barrier to entry. Amer. Econom. Rev. 77 388–401. ORGANIZATION SCIENCE /Vol. 12, No. 3, May–June 2001 Anderson, P. 1999. Complexity theory and organization science. Organ. Sci. 10 216–232. Athey, S., A. Schmutzler. 1995. Product and process flexibility in an innovative environment. RAND J. Econom. 26 557–574. ——, S. Stern. 1996. An empirical framework for testing theories about complementarities. MIT Mimeo. Baldwin, C. Y., K. B. Clark. 2000. Design Rules: The Power of Modularity. MIT Press, Cambridge, MA. Barney, J. 1991. Firm resources and sustained competitive advantage. J. Management 17 99–120. Boston Consulting Group. 1972. Perspectives on experience. Technical report. Brown, S. L., K. M. Eisenhardt. 1998. Competing on the Edge: Strategy as Structured Chaos. Harvard Business School Press, Boston, MA. Chew, W. B., T. F. Bresnahan, K. B. Clark. 1990. Measurement, coordination, and learning in a multiplant network. R. S. Kaplan, ed. Measures for Manufacturing Excellence. Harvard Business School Press, Boston, MA. Cockburn, I., R. M. Henderson, S. Stern. 1999. The diffusion of science-driven drug discovery: Organizational change in pharmaceutical research. Working paper 7359, NBER. Cohen, W. M., D. Levinthal. 1990. Absorptive capacity: A new perspective on learning and innovation. Admin. Sci. Quart. 35 128– 152. ——, R. R. Nelson, J. P. Walsh. 2000. Protecting their intellectual assets: Appropriability conditions and why U.S. manufacturing firms patent (or not). Mimeo, Carnegie Mellon, Pittsburgh, MA. Collis, D. 1996. Marks and Spencer, Ltd. (A), (B), and (C): Teaching note. Harvard Business School Teaching Note 391–271, Harvard University, Cambridge, MA. ——, C. A. Montgomery. 1997. Corporate Strategy: Resources and the Scope of the Firm. Irwin, Chicago, IL. ——, T. Stuart. 1995. Cooper Industries’ corporate strategy (A). Harvard Business School Case 391–095, Harvard University, Cambridge, MA. Dierickx, I., K. Cool. 1989. Asset stock accumulation and sustainability of competitive advantage. Management Sci. 35 1504–1511. Dixit, A. 1980. The role of investment in entry-deterrence. Econom. J. 90 95–106. Fleming, L., O. Sorenson. 2001. Technology as a complex adaptive system: Evidence from patent data. Res. Policy. Forthcoming. Foley, S., T. Mahmood, S. P. Bradley, P. Ghemawat. 1996. Wal-Mart Stores, Inc. Harvard Business School Case 794–024, Harvard University, Cambridge, MA. Gavetti, G., D. Levinthal. 2000. Looking forward and looking backward: Cognitive and experiential search. Admin. Sci. Quart. 45 113–137. Ghemawat, P. 1997. Games Businesses Play: Cases and Models. MIT Press, Cambridge, MA. ——, R. Caves. 1986. Capital commitment and profitability: An empirical investigation. Oxford Econom. Papers 38 94–110. ——, H. J. Stander III. 1996. Nucor at a crossroads. Harvard Business School Case 793–039, Harvard University, Cambridge, MA. Gilbert, R. J., M. B. Lieberman. 1987. Investment and coordination in oligopolistic industries. RAND J. of Econom. 18 17–33. 291 Downloaded from informs.org by [2.37.204.136] on 24 July 2023, at 09:04 . For personal use only, all rights reserved. JAN W. RIVKIN Reproducing Knowledge Grant, R. M. 1996. Prospering in dynamically-competitive environments: Organizational capability as knowledge integration. Organ. Sci. 7 375–387. ——, J.-C. Spender. 1996. Knowledge and the firm. Special Issue of the Strategic Management J. Hallowell, R. H. 1997. Southwest Airlines: 1993 (A). Harvard Business School Case 694–023, Harvard University, Cambridge, MA. Hitt, L., E. Brynjolfsson. 1997. Information technology and internal firm organization: An exploratory analysis. J. Management Inform. Systems 14 81–101. Hogan, D. G. 1997. Selling ‘Em by the Sack: White Castle and the Creation of American Food. New York University Press, New York. Horgan, J. 1995. From complexity to perplexity. Scientific American 272 (6) 104–109. Hwang, M., D. Weil. 1996. Production complementarities and the diffusion of modern manufacturing practices: Evidence from the U.S. apparel industry. Boston University Department of Economics Mimeo. Ichniowski, C., K. Shaw, G. Prennushi. 1997. The effects of human resource management practices on productivity: A study of steel finishing lines. Amer. Econom. Rev. 87 291–313. Irwin, D. A., P. J. Klenow. 1994. Learning-by-doing spillovers in the semiconductor industry. J. Political Economy 102 1200–1227. Jaffe, A. B. 1986. Technological opportunity and spillovers of R&D: Evidence from firms’ patents, profits, and market value. Amer. Econom. Rev. 76 984–1001. Katz, R., T. J. Allen. 1982. Investigating the not invented here (NIH) syndrome: A look at the performance, tenure, and communication patterns of 50 R&D project groups. R&D Management 12 7–19. Kauffman, S. A. 1993. The Origins of Order: Self-Organization and Selection in Evolution. Oxford University Press, Oxford, U.K. ——. 1995. At Home in the Universe. Oxford University Press, Oxford, U.K. ——, S. Levin. 1987. Toward a general theory of adaptive walks on rugged landscapes. J. Theoret. Biology 128 11–45. ——, J. Lobo, W. Macready. 2000. Optimal search on a technology landscape. J. Econom. Behavior and Organ. 43 141–166. ——, W. Macready. 1995. Technological evolution and adaptive organizations. Complexity 1 26–43. ——, E. D. Weinberger. 1989. The N-k model of rugged fitness landscapes and its application to maturation of the immune response. J. Theoret. Biology 141 211–245. ——, ——, A. S. Perelson. 1988. Maturation of the immune response via adaptive walks on affinity landscapes. A. S. Perelson, ed. Theoretical Immunology, Part I, Santa Fe Institute Studies in the Sciences of Complexity. Addison-Wesley, New York. Knott, A. M., B. McKelvey. 1999. Nirvana efficiency: A comparative test of residual claims and routines. J. Econom. Behavior and Organ. 38 365–383. Kogut, B., U. Zander. 1992. Knowledge of the firm, combinative capabilities, and the replication of technology. Organ. Sci. 3 383– 397. Kolmogorov, A. N. 1965. Three approaches to the qualitative definition of information. Problems of Information Transmission, pp. 1, 4–7. 292 Levin, R. C., A. K. Klevorick, R. R. Nelson, S. G. Winter. 1987. Appropriating the returns from industrial research and development. Brookings Papers on Econom. Activity 3 783–820. Levinthal, D. 1997. Adaptation on rugged landscapes. Management Sci. 43 934–950. ——, M. Warglien. 1999. Landscape design: Designing for local action in complex worlds. Organ. Sci. 10 342–357. Lieberman, M. B. 1987a. Excess capacity as a barrier to entry: An empirical appraisal. J. Indust. Econom. 35 607–627. ——. 1987b. Postentry investment and market structure in the chemical processing industries. RAND J. Econom. 18 533–549. ——. 1989. The learning curve, technology barriers to entry, and competitive survival in the chemical processing industries. Strategic Management J. 10 431–447. Lippman, S., R. Rumelt. 1982. Uncertain imitability: An analysis of interfirm differences in efficiency under competition. Bell J. Econom. 12 418–438. Mansfield, E. 1985. How rapidly does new industrial technology leak out? J. Indust. Econom. 34 217–223. ——, M. Schwartz, S. Wagner. 1981. Imitation costs and patents: An empirical study. Econom. J. 91 907–918. McKelvey, B. 1999. Avoiding complexity catastrophe in coevolutionary pockets: Strategies for rugged landscapes. Organ. Sci. 10 294–321. Milgrom, P., J. Roberts. 1990. The economics of modern manufacturing: Technology, strategy, and organization. Amer. Econom. Rev. 80 511–528. ——, ——. 1995. Complementarities and fit: Strategy, structure, and organizational change in manufacturing. J. Accounting Econom. 19 179–208. Montgomery, C. A. 1994. Marks and Spencer, Ltd. (A). Harvard Business School Case 391–089, Harvard University, Cambridge, MA. Nelson, R., S. Winter. 1982. An Evolutionary Theory of Economic Change. Belknap, Cambridge, U.K. Nonaka, I., H. Takeuchi. 1995. The Knowledge-Creating Company. Oxford University Press, Oxford, U.K. Orton, J. D., K. E. Weick. 1990. Loosely coupled systems: A reconceptualization. Acad. Management Rev. 15 203–223. Page-Jones, M. 1980. The Practical Guide to Structured Systems Design. Yourdon Press, New York. Penrose, E. 1959 [1995]. The Theory of the Growth of the Firm. Oxford University Press, Oxford, U.K. Perrow, C. 1984. Normal Accidents: Living With High-risk Technologies. Basic Books, New York. Peters, T. J., R. H. Waterman, Jr. 1982. In Search of Excellence: Lessons from America’s Best-run Companies. Warner Books, New York. Polanyi, M. 1962. Personal Knowledge: Towards a Post-critical Philosophy. University of Chicago Press, Chicago, IL. Porter, M. E. 1980. Competitive Strategy: Techniques for Analyzing Industries and Competitors. Free Press, New York. ——. 1996. What is strategy? Harvard Bus. Rev. 74 (6) 61–78. Reed, R., R. J. DeFillippi. 1990. Causal ambiguity, barriers to imitation, and sustainable competitive advantage. Acad. Management Rev. 15 88–102. ORGANIZATION SCIENCE /Vol. 12, No. 3, May–June 2001 Downloaded from informs.org by [2.37.204.136] on 24 July 2023, at 09:04 . For personal use only, all rights reserved. JAN W. RIVKIN Reproducing Knowledge Rivkin, J. W. 1995. The rise of retail category killers. Harvard Business School Mimeo. ——. 2000. Imitation of complex strategies. Management Sci. 46 824– 844. Sanchez, R., J. T. Mahoney. 1996. Modularity, flexibility, and knowledge management in product and organization design. Strategic Management J. 17 (Special Issue) 63–76. Schilling, M. A. 2000. Toward a general modular systems theory and its application to interfirm product modularity. Acad. Management Rev. 25 312–334. Schmalensee, R. 1978. Entry deterrence in the ready-to-eat breakfast cereal industry. Bell J. Econom. 9 305–327. Siggelkow, N. 1998. Benefits of Focus, Evolution of Fit, and Agency Issues in the Mutual Fund Industry. Ph.D. Dissertation, Harvard University, Cambridge, MA. Simon, H. A. 1962. The architecture of complexity. Proc. Amer. Philosophical Society 106 467–482. Sorenson, O. J. 1997. The Complexity Catastrophe in the Computer Industry: Interdependence and Adaptability in Organizational Evolution. Ph.D. Dissertation, Stanford University, Stanford, CA. Szulanski, G. 1996. Exploring internal stickiness: Impediments to the transfer of best practice within the firm. Strategic Management J. 17 (Special Issue) 27–43. Teece, D. J. 1977. Technology transfer by multinational firms: The resource cost of transferring technological know-how. Econom. J. 87 242–261. ——. 1986. Profiting from technological innovation. Res. Policy 15 285–306. Upton, D. M. 1995. What really makes factories flexible? Harvard Bus. Rev. 73 74–84. Uyterhoeven, H. E. R., M. Hart. 1996. Banc One—1993. Harvard Business School Case 394–043, Harvard University, Cambridge, MA. Walton, S., J. Huey. 1992. Sam Walton, Made in America. Bantam Books, New York. Weick, K. E. 1976. Educational organizations as loosely coupled systems. Admin. Sci. Quart. 21 1–19. Westhoff, F. H., B. V. Yarbrough, R. M. Yarbrough. 1996. Complexity, organization, and Stuart Kauffman’s The Origins of Order. J. Econom. Behavior and Organ. 29 1–25. Winter, S. G. 1987. Knowledge and competence as strategic assets. D. Teece, ed. The Competitive Challenge: Strategies for Industrial Innovation and Renewal. Ballinger, Cambridge. ——. 1995. Four Rs of profitability: Rents, resources, routines, and replication. C. Montgomery, ed. Resource-based and Evolutionary Theories of the Firm: Towards a Synthesis. Kluwer, Boston, MA. ——, G. Szulanski. 2000. Replication as strategy. Working Paper 98– 10, Reginald H. Jones Center, Wharton School, University of Pennsylvania, Philadelphia, PA. Zander, U., B. Kogut. 1995. Knowledge and the speed of the transfer and imitation of organizational capabilities: An empirical test. Organ. Sci. 6 76–92. Accepted by Benoit Morel. ORGANIZATION SCIENCE /Vol. 12, No. 3, May–June 2001 293