Among the most influential statements on network externality have been those by Michael Katz and Carl Shapiro. Their 1985 paper in the American Economic Review defines the concept: "There are many products for which the utility that a user derives from consumption of the good increases with the number of other agents consuming the good." Elaborating they add, "[T]he utility that a given user derives from a good depends upon the number of other users who are in the same network." This idea of a network embraces far more than the physically connected examples of computer networks and telecommunications systems. They also mention goods, such as computer software, automobile repair, and video games as exhibiting positive consumption externalities. These are just a few examples from what we believe would be a very large set. It is easy to come up with many more examples illustrating the concept of a network as they have defined it. When gourmet cooks find it easier to find preferred ingredients because more people are taking up their avocation, this would be a gourmet-network externality. When fans of live entertainment prefer big cities because the large market for entertainment assures a full variety of acts, this would be an audience-network externality. Indeed, we can say that the urban amenities and agglomeration economies that we associate with urban life are urban precisely because some activities function best, if at all, where large numbers of participants can be involved.
Nor is there any reason that a network externality should necessarily be limited to positive effects, although positive effects have been the main focus in this literature. If, for example, a telephone or computer network becomes overloaded, the effect on an individual subscriber will be negative. When we admit the possibility of a negative network externality, the set of goods that exhibit network externalities expands strikingly. As members of a network of highway users, we suffer from a negative network externality because freeways are subject to crowding. And although a larger installed base of computer users might lower the price of computer software, there are many goods, such as housing and filet mignon, where larger networks of users appears to increase the price.
According to the received definition, then, goods exhibit network externality wherever the consumer enjoys benefits or suffers costs from changes in the size of an associated network, that is, changes in quantities demanded. Such These benefits and costs result from such considerations as compatibility, brand familiarity, product information, status, service availability or the prices of network related goods. This almost perfect generality should sound an alert: what we have here is either much bigger or much smaller than its current position in the literature would indicate. It is either the case that most goods exhibit network externalities, and the unique insights of this literature are of far broader consequence than is currently understood, or it is the case that network externalities are limited in ways not yet specified.
As the reader will likely suspect, we believe the concept of network externalities must be limited. A classic network like CompuServe does represent an interesting class of economic phenomena, but the definition of that class must be considerably refined if it is to support fruitful analysis. The purpose of this paper, then, is to elucidate, refine, and ultimately limit the implied scope of network externality.1
Our concern is not merely semantic. The application of the concept of network externality has been broad and rapid. Further, both by name and by some of the explicit results of the network literature, network externalities are asserted to constitute market failure.2 If almost every aspect of the economy exhibits network externality, and if externalities presumptively are market failures, then our most basic results about the efficiency of markets may be in error and dramatic policy changes might be warranted.
We believe such changes would be misguided. While network effects are common and important, network externalities as market failures, we will argue, are theoretically fragile and empirically undocumented.
Katz and Shapiro draw a distinction between direct and indirect network externalities. Direct network externalities are those generated "through a direct physical effect of the number of purchasers on the quality of the product," as in the attachment of homes to a telephone network. Indirect network externalities, on the other hand, involve instances that lack that direct physical effect; for example, software being more plentiful and lower in price as the number of computer users increases. Katz and Shapiro also note another instance of indirect network externality, the availability of post-purchase service for durable goods, such as automobiles.3 In general, they ascribe indirect network externalities to any situation where complementary goods become more plentiful and lower in price as the number of users of the good increases.
A related distinction has been made between networks that are (or can) be owned and networks that are not (or cannot) be owned. This distinction is partly a function of the legal structure and partly a result of physical characteristics of networks. In some networks, participants are literally connected to each other in some fashion. The telephone system is one such network, as are pipeline, telex, electrical, and cable television systems. These literal networks require an investment of capital, and there is a physical manifestation of the network in the form of pipelines, cables, transmitters, and so on. It is not only feasible but almost inevitable for property rights to be established for these types of networks. Those who attach to such networks without permission from the owner, or who attach without adhering to the rules may be disconnected, a characteristic that removes the problem of nonexclusion.4
In contrast, other networks, which we might refer to as "metaphorical networks," such as the network of speakers of English, provide interrelationships in which there are no physical connections (though there may be direct interaction). The network of Chevrolet owners, whose relationship to each other is that they draw on common repair expertise, is such a network. So is the network of recreational fishermen, who enjoy exchanging stories and who may provide assistance to each other in emergencies. While the ownership of such a metaphorical network is possible, it may be difficult. Unlike the telephone company, which can monitor uninvited users because they must all hook up to the same electronic circuits, an owner of a language would have great difficulty monitoring illicit use, even if he were legally empowered to prevent it.5 Thus metaphorical networks are less likely to be owned, and in some instances may not be ownable.
As we discuss below, the recognition that some networks can be owned solves many of the problems that might be thought to occur when network participants create what have been called direct network externalities. First, however, we examine the efficiency consequences of indirect network externalities.
Pecuniary externalities are external effects that work through the price system. When firm A produces one more unit of output, marginally lowering price, that increased production harms rival firms B and C. Indeed, the sine qua non of competition is each firm's indifference to the impact of its actions on other firms in the industry. Each firm acts in a way that harms the other firms in the industry, but that harm is offset by a transfer of wealth to consumers. Technological externalities, on the other hand, comprise the class of externalities that economists have focused on in recent decades as market failures: pollution, congestion, and so on. Although technological externalities are often thought to be market failures, some interactions that have been identified by economists as externalities have in fact been internalized by markets. For example, while economists were writing of the positive externality brought to apple growers by the pollination activities of bees, beekeepers were internalizing this activity (and consequently invalidating the arguments of economists) by contracting with owners of apple orchards.6 Economists have demonstrated great resourcefulness in reconstructing reality to fit economic theory (as documented by Coase, 1960; Cheung, 1973; Cowan, 1988).
But at least in principle, the difference between technological and pecuniary externalities is clear. With technological externalities, actual benefits or costs are imposed outside of market mechanisms. Resolution of such problems may occur through property rights, private negotiations, or government interventions that allow the externalities to be internalized. However, in the case of pecuniary externalities, those on one side of the market (say, buyers) benefit, while those on the other side of the market (say, sellers) suffer. Here, the market outcome is the one that would be reached if all these external effects were internalized; piecemeal internalization is harmful. If firms in an otherwise competitive industry were to internalize their impacts on each other, the industry would be a monopoly or a cartel and the internalization would cause inefficiency.
Today this distinction between technological and pecuniary externalities has largely been forgotten, perhaps because it is no longer needed to correct Pigou. But the trap that caught Pigou is still set, ready to spring on the unwary analyst. In fact, the pecuniary externalities that so perplexed Pigou walk and quack very much like the indirect network externalities that are waddling through the literature today. As an example, listen to Farrell and Saloner (1985, p. 70) discussing indirect network externalities: "There may be a market-mediated effect, as when a complementary good (spare parts, servicing, software) becomes cheaper and more available the greater the extent of the (compatible) market." These "market-mediated" (read price) effects are likely pecuniary.
For a negative indirect network externality, the analogy is obvious: if a group of breakfast-eaters joins the network of orange juice drinkers, their increased demand raises the price of orange juice concentrate, and thus most commonly effect a transfer of wealth from their fellow network members to the network of orange growers.7 Certainly, the breakfast eaters have affected the orange juice drinkers. Just as certainly, they have not compensated the orange juice drinkers for the harm they have caused them. And most assuredly, we would not want them to make such recompense.
This is a pecuniary externality.
The positive indirect network externality is a bit more complicated, but it amounts to largely the same thing. We consider these situations in much greater detail in Liebowitz and Margolis (1994a). The gist of our argument is that if price falls as a network gets larger, that fall in price could be due to one of three factors: a positive technological externality across producers; an input that decreases in cost; or a decrease in rents to some (perhaps specialized) agents as the industry grows. If outside observers can only see that price decreases as output increases, they will be unable to distinguish among these cases. Without some additional information to differentiate between pecuniary effects (including redistributions in rents) and genuine externalities, it is impossible to prescribe appropriate public policy.
As an example, consider the oft-cited case of the "network" of computer users. Suppose that personal computer prices fall as the number of users increases. If this is a true externality, say because there are positive technological externalities between computer producers, there would be a justification for some sort of subsidy to the industry. However, a second possibility for declining computer prices could be declining costs of an input, such as microprocessors. In this case, a subsidy to the computer industry will be a poorly directed, and perhaps counterproductive policy tool. If a subsidy is required anywhere, it would be in the microprocessor industry, since a subsidy to the computer industry will not optimally address the number of processors per computer, or the use of processors in non-computer applications. And without further investigation into the cause of the declining microprocessor price, we can not know whether a subsidy is warranted in the microprocessor market. Clearly, treating as network externalities all cases where price and quantity move inversely can easily lead both to bad economics and to bad policy.
There is yet an additional problem with indirect network effects. Again, consider the common modern experience that rapidly declining prices of many activities are associated with dramatic increases in participation in those activities. So, for example, the costs of videocassette recorders, computers, and fax machines have declined rapidly at the same time that use of these machines has increased. (The association is not confined to contemporary new technologies: automobiles, electrical power generation equipment, refrigeration, and other "old tech" items experienced the same pattern in their day.) The problem is that this simple observation of rapidly declining prices does not distinguish between economies of scale and ordinary technological progress.8 Is the equilibrium moving along a downward-sloping supply function as demand increases, or is it moving down a demand curve as changes in technology move the supply curve? This distinction is critical to whether the phenomenon exhibits a network externality. The choice among networks, which is so central to this literature, is only central if there are economies of scale in particular networks.9 Where observed cost reductions are due to general advances in technology, rather than network size per se, then there is no network effect and no necessity to a collective choice of a particular network. To the extent that the observed association of cost decreases and output increases is the result of improved technology and not economies to scale for particular networks, the implied scope for the concept of network externality is smaller than is currently understood. Again, theory, empirics, and policy require that we note a distinction that the mainstream literature ignores. The concept of indirect network externalities, therefore, suffers from two main weaknesses. First, it is not an externality in the modern sense since it often describes nothing more than welfare neutral interactions that occur in properly functioning markets. Second, the implied scope of positive network externalities is likely misunderstood where analysis fails to identify properly the cause of a declining price.
The idea of an inframarginal externality was introduced thirty years ago by Buchanan and Stubblebine. In their very general definition of externality, one party's activity level shows up as an argument in another's utility function. For inframarginal externality, the marginal utility of the external activity is zero. Very simply, the affected party is not affected by marginal changes in the externality-causing activity. To find the optimum, we can no longer look at the marginal conditions, but must examine the total benefits and costs. The concept has been neglected, perhaps because the usual interest on external harms (like those of pollution) offers no natural analog to satiation, since we are usually comfortable assuming increasing marginal damage in those cases. Consideration of network externality, however, prompts renewed attention to this idea.
Many activities require a critical mass but are not much helped by participation beyond that level. City size is limited because urban agglomeration economies are exhausted, at the margin, where crowding offsets the benefits of additional interaction.10 Similarly, the fact that other people use the same sort of VCR that we use makes a tape rental market available to us, but the marginal benefits of increasing the number of households that our kind of VCR are likely exhausted now that businesses that rent video tapes are just about as prevalent as ones that sell milk Attention to inframarginal externality alters the problem in several ways. First, where marginal benefits of network size are exhausted there are no unexploited gains from trade regarding network size. Further, where marginal gains of network size are exhaustible at network sizes that are small relative to the market, there is no impediment to the coexistence of multiple networks. The choice of the "best" network becomes one of choosing the best set of networks. Finally, the inframarginal externality that may afflict the discrete choice of a network is not different from other coordination problems that exist in many other market choices, as we discuss later in this paper.
To analyze this situation, we begin with a conventional representation of the incentives of an owned network in which ownership solves the network problem. The model is neither original nor complicated: It is the tragedy-of-the-commons problem turned on its head.
In Figure 1, B(N) is the benefit that any participant (all individuals are assumed identical) derives from participation in the network. As such, it is both the marginal private benefit and the average social benefit, and it increases with N, the number of participants in the network. To the owner of the network, it represents the average revenue (demand) curve. The marginal social benefit (MSB) lies above B(N), because MSB includes the positive impact of an additional participant on the other network participants. The MSB, being the first derivative of NB(N), is also the marginal revenue available to the network owner. (Marginal revenue here is the price paid by the marginal participant plus the effect of the marginal participant on the willingness to pay of all participants.) The marginal cost of serving another network participant is also shown (MC), and it is assumed to increase with network size. The profit-maximizing behavior for the network owner will be to charge the price P*, and accommodate N* participants on the network, since this equates marginal revenue with marginal cost. In this case, the network owner does charge a price that is below marginal cost, and there is no problem with the credibility of this action.
The model illustrated in the diagram has a number of special assumptions, but these assumptions are the exact analogs to those used in the simple models of the tragedy of the commons, frequently illustrated as the fisheries problem. And although the model certainly cannot claim to be perfectly general, it does counter the impression of the inevitability of a suboptimal network.
There is nothing in this model that denies the possibility of competition in networks, just as there is nothing in the fisheries model that dictates that a single owner of a lake becomes a monopolist in fish. In the perfectly competitive case, competition shifts the B(N) curve down until P* will just cover the costs of operating the network. For example, when America On-line enters the market with lower prices, the willingness to pay for the Prodigy network falls. Thus, while the network owner can be said to capture the participant's full value from the use of the particular network, the value captured is not the value of network participation in general.12
One assumption embodied in the model is increasing marginal cost of serving a network participant. This assumption is conventional in most of economics, but it is possible to offer specific support for its application to networks. For literal networks, where addition of customers ordinarily means connecting to subsequently more distant or otherwise more costly participants, the assumption seems particularly compelling, and there is some empirical support for it (see Kahn v. 1 p. 124, and v. 2 p. 123). Even for metaphorical networks, it is reasonable that the first participants will be those most suited to use the network, requiring the least support for their participation. As examples of each effect: Cable television reaches the boondocks only after the more urban areas are covered, and the Internet was first used by UNIX wizards, not computer neophytes.
Of course, other cost conditions are possible. Constant or decreasing marginal cost, which is the assumption that has most commonly been made in the network externality literature, converts the problem to one of natural monopoly. If MC were constant in figure 1, the optimal size of the network would be infinite (or the entire population). The existence of multiple networks competing with each other would now be inconsistent with efficiency. But although it is possible to impose on networks the natural monopoly problem (and some networks have provided some of the best examples of natural monopoly), this is not a new or different problem, but only a familiar problem with a new name. Note also that if the "externality" were internalized by network users, the would follow the MSB curve would now represent behavior of network users. Yet this would not solve the natural monopoly problem (choosing the best network), although it would lead to the proper size for any given network. Thus it is not clear that the externality framework is useful in resolving problems of network choice.
Difficulties arise when the problem is configured such that willingness to pay in an initial period is contingent on the expected network size in a second period. Katz and Shapiro (1986) present a case in which a firm with lower costs in period two has an advantage over a firm with lower costs in period one because the former can credibly commit to lowering price the first period, but the later can not commit to lowering prices in the second period. This result is fairly specialized, however, arising in a particular context that gives rise to a time-inconsistency problem. Certainly the commitment problem would not apply to those networks where participants make a payment each period for their place in the network (or rentals for durable goods), since in this case first period consumers need not fear getting 'stuck' with the wrong product in period two. By assuming that consumers' contractual obligations span multiple periods, and stipulating a difficulty of contracting to constrain actions across these periods, we can, of course, construct problems for this market.
While it may be of interest to note that contracting costs or time inconsistencies affect networks in the same ways that they affect other economic relationships, it is misleading to present these difficulties as fundamental economic characteristics of networks. Our argument here and elsewhere is that networks that are owned are no more (or less) afflicted by market pathologies than are other economic relationships.
While ownership would seem straightforward for most literal networks, it is not impossible for the kinds of networks that we have called metaphorical. Franchise systems of various sorts are mechanisms for introducing ownership to networks of restaurants, retailers, or other service firms. And less obvious networks can be owned: Bars, country clubs, and private schools can all be thought of as facilitating networks -- internalizing network effects. They are valued both for the services that they provide and also for the social networks that they support. You can drink alone for less, but you pay a barkeeper for the service of creating a network. Such network services will not come free, given that there are real costs of creating them, but they needn't carry a monopoly price. Competition among such networks is certainly possible.
Interactions of this sort involve transactions. Although we may not be able to transact with every motorist who fouls the air that we breathe, we can advise our parents to buy a VHS video cassette recorder if they want to see home movies of their grandchildren; and we can negotiate with a co-author regarding which word-processing software to use. Since the exchange of such materials is itself a transaction, it seems unreasonable to assume that in such cases transactions costs are prohibitive. But if transactions are relatively easy, then the existence of an externality is unlikely.
Economics has not generally done well in explaining transition (Fisher, 1983). Speaking metaphorically, neoclassical economics has put transition into black boxes, assuming that an unknown process is responsible for costless and timeless movement. This might be thought of as a Panglossian view of market transitions. But the analysis of discrete choice problems in the network externality literature goes to the opposite extreme, assuming that if a result has that which has not been modeled, that it does not exist. Proponents of network externalities have attempted to fill these black boxes, but do so with restrictive models in which these transitions often do not occur. A clear implication of the network externalities literature is that often we cannot move from one technology to a superior one, from one standard to a better one, from one kind of network to a better one. This view might be thought of as the Chicken-Little view of market transitions. While it is inevitable and probably desirable that we work with restricted models, we should avoid the presumption that the things that are excluded from these models are unimportant or nonexistent.
Of course, potential problems with transitions afflict all components of the economy. For example, imagine that a new automobile manufacturer, Superior Motors, devises a new automobile design that lowers the cost of producing automobiles. The common assumption in economics is that information is not prohibitively costly, that consumers will soon find out about this new option, and that the new lower-cost automobile will soon come to dominate the market. Those less sanguine about this result might point out that building a better mousetrap is not enough to ensure success, that the mousetrap needs to be properly marketed, financed, and so on. But the usual reply would be that since Superior Motors has lower costs, it can profitably invest at least as much in marketing as its rivals, and thus would come to dominate its market.
It is nevertheless possible to construct out of these circumstances a coordination problem. Although Superior Motors has a lower cost curve, it has to overcome diseconomies of small scale. Will the innovation prevail in the face of lagged consumer response and these diseconomies? One argument is that the firm (and capital markets) knows its cost curve, can anticipate the consequences of operation at a particular scale, and will proceed to maximize profits. But by installing a kind of myopia in the problem, and focusing on the time lag between first operation and attaining efficient scale, the problem of implementing the new technology can appear fraught with pitfalls.
Indeed, with this Chicken-Little perspective on the world, we will find it remarkable that even the most ordinary new technology or action is ever implemented. It will seem remarkable that gasoline stations ever became available because, after all, at one time there were no cars. It will seem just as remarkable that cars ever became available because, after all, at one time there were no gas stations. We wonder at farmers' willingness to plant seeds and wait the months necessary for the crop to mature. Life is fraught with uncertainty.
Clearly, even the simplest act of production requires the capacity to form an expectation of the outcome and to take the risk that what is produced today can be sold tomorrow. It can be argued that the whole area of entrepreneurship continues to reside in those black boxes that contain the subject of economic transition. A transition to a standard or technology that offers benefits greater than costs will constitute a profit opportunity for entrepreneurial activities that can arrange the transition and appropriate some of the benefits. Granted, modeling these activities is difficult. Many topics have resisted successful modeling. But if science is unable to explain how bees fly, it is science that needs to be amended, not our acknowledgment that bees, in fact, manage to stay aloft. Similarly, the fact that current economic models of transition indicate that worthwhile transitions may not occur is not sufficient reason to abandon the presumption that they usually do occur. Economies do, in fact, seem to move from one state to another. This is not to say that mistakes are never made, in markets or elsewhere. But we have overwhelming evidence that markets do make transitions to superior products and standards -- from horses and buggies to automobiles, from typewriters to computers, from mail to fax. Given the march of technological progress, claims that wrong choices were made, or that superior options were not implemented in a timely fashion require a fairly high standard of countervailing evidence. We turn now to some consideration of evidence.
Probably the most frequently cited example of market failure due to network externalities is the design of the typewriter keyboard (David, 1985).14 The beguiling and often told story is that the strike mechanism of the earliest mechanical typewriters was prone to jamming, so the typewriter's inventors designed the (now standard) QWERTY keyboard in order to slow down typing speed. This arrangement became the market leader, largely by accident, because it became associated with the world's only touch typist. Typists remain burdened by this speed-reducing design today, even though there exists a competing Dvorak keyboard -- scientifically designed to be easier to learn and to allow greater speed. Nevertheless, we all learn touch typing on the QWERTY design because there are so few Dvorak typewriters, and there are so few Dvorak keyboard typewriters because almost no one knows how to type on them. This vicious cycle keeps us stuck on the wrong standard. The empirical support for the story is a U.S. Navy study conducted during the World War II. Purportedly, that study conclusively demonstrates the superiority of the Dvorak design, determining that the costs of retraining QWERTY typists on the Dvorak design will be recouped within ten days from the start of training.
Alas, almost every element of this tale is false, as we show in Liebowitz and Margolis (1990). The QWERTY keyboard was not created to slow down typing speed. Early on, there were other publicized touch typists using other keyboards. The Navy study was very poorly documented and designed, and appears to have been conducted by Navy Lieutenant Commander August Dvorak, creator and patent holder on the keyboard bearing his name. A later, carefully constructed and controlled study, performed for the General Services Administration in the 1950s, demonstrated quite the opposite results from the Navy study. More recent studies indicate that there is practically no difference in typing speed between the two keyboard designs. The Dvorak typewriter keyboard, it turns out, is a rather poor empirical base upon which to support a theory.
Probably the second most popular example is the claim that the Beta videotaping format is superior to VHS. This claim may achieve much of its popularity because the competition between these formats is so widely known, because so many of us have firsthand experience with this choice, and because a significant number of the people who chose the Beta format did so because they believed it to perform better than VHS.15 Although it is a common belief that Beta was better in some way, the evidence does not support the view that Beta was better in any way that counted to the bulk of consumers. For example, Klopfenstein (1989, p.28) writes:
Although many held the perception that the Beta VCR produced a better picture than VHS, technical experts such as Weinstein (1984) and Prentis (1981) have concluded that this was, in fact, not the case; periodic reviews in Consumers Reports found VHS picture quality superior twice, found Beta superior once, and found no difference in a fourth review. In conclusion, the Beta format appeared to hold no advantages over VHS other than being the first on the market, and this may be a lesson for future marketers of new media products.
Lardner's history of the videorecorder market (1987) provides additional support for this conclusion. There were no real technical differences initially between Beta and VHS. The major differences were the size of the cassette, the threading of the tape, and the tape speed. This similarity in technical specification was due a prior patent-licensing agreement between Sony and Matsushita (creators of Beta and VHS respectively), who had previously cooperated in selling a professional videorecorder called the U-matic. Sony offered its Beta technology and design to Matsushita, but the latter decided to pursue its own machine (produced by JVC). Matsushita's decision was partly based on its different perception of consumer desires. Sony management believed the paramount concern to the consumer would be transportability of the cassette, so they produced a paperback-sized cassette even though this limited recording time to one hour. Matsushita management believed that consumers would be more concerned with the capacity of the tape, so they opted for a larger cassette that allowed a two-hour recording time, making the taping of a complete movie or sports event possible. Sony's headstart gave Beta the entire market for several years. But within two years of VHS's introduction, thanks to its lower price and longer playing time, VHS had surpassed Beta and soon after came to dominate the market.
The typewriter keyboard and VCR format are not the only claims of market failure in the network externalities literature, but these are the claims most often repeated. Many of the other claims are of a highly conjectural nature, based on hypothetical technological developments that might have borne extraordinary fruit if only they had been more thoroughly explored. An example is the claim (reported in Arthur) that the internal combustion engine might have been the wrong choice of automobile engine (vis-à-vis steam). Though our skepticism about the validity of this claim is apparently not universal, we find this particular example difficult to take seriously.
All of this points to a challenge to those claiming that network externalities are important in the economy. The theoretical literature establishes only that within models that incorporate particular abstractions, market failure of the type that causes the wrong network to be chosen, is possible. Given these abstractions, it is essential that the literature present real examples of demonstrable market failure if the concept of network externality is to have any relevance.
The debate over network externalities is a reminder of more general methodological concerns. It demonstrates that rigor comes in small and incomplete packages. The models of network externality proceed with great rigor from a simple and plausible assumption -- that the benefits of an activity depend upon the number of participants -- to a variety of conclusions. But these models can not tell whether such a problem is important. After we economists have had our fun, thinking about network effects and considering how social interactions have a similarity to networks, we need to acknowledge that the a priori case for network externalities is treacherous and the empirical case is yet to be presented. Most constructs in economics find their way only very slowly into either public policy or established theory. The construct of network externalities should be one of them.