NETWORK EXTERNALITIES AND THE DETERMINANTS OF NETWORK SURVIVAL

Robert J. Kauffman

Associate Professor of Information Systems and Decision Science Department of Information and Decision Science

Carlson School of Management University of Minnesota

Phone: 612-624-8562 Email: rkauffman@csom.umn.edu

Yu-Ming Wang

Associate Professor of Information Systems College of Business Administration

California State University, Long Beach Long Beach, CA

Phone: 582-985-7861 Email: ymwang@csulb.edu

Last Revision: December 8, 1999

_____________________________________________________________________________________

ABSTRACT Although much theoretical work has been done on the adoption of technologies and networks in the presence of externalities, there is scant empirical evidence of the existence of network externalities and how they influence why some networks fail and exit the market, while others remain competitive and evolve to become dominant. Using economic analysis of the sources of network externalities, this paper examines variables influencing the likelihood of a network’s survival. We use growth data for regionally shared electronic banking networks in the United States to examine the determinants of network survival in the retail electronic banking services context. We employ a duration modeling approach, which incorporates explanatory variables that enable us to discern patterns of survivability. The results show that networks with a greater installed base, higher growth rates, and higher operational efficiencies, and networks that are well established in the market are better able to survive. The results provide evidence of the importance of network externalities, which create value for consumers, adopting firms and network providers. Our results also support the idea that network goods exhibit a positive feedback effect: larger networks are often expected to prevail in the market and they indeed tend to grow larger and become dominant. Beyond our present application, the evaluative framework used in this research can be applied more broadly, to a spectrum of information technologies, electronic commerce settings and competitive interorganizational information systems that offer network externalities. KEYWORDS: ATMs, automated teller machines, duration modeling, electronic banking, electronic commerce, installed base, network externalities, network mergers, network survivability, retail banking, shared networks.

1. INTRODUCTION

Many industries whose products and services are based on information technology (IT) have been

swept by asset buyouts, mergers, consolidations, acquisitions and strategic joint ventures, a trend that

promises to bring increased competition and cooperation, and lower prices for consumers. We have seen

this happen in the traditional world of packaged software products (e.g., with Wordperfect’s purchase by

Borland and later Corel, and IBM’s acquisition of Lotus Corporation), and for CASE tools (e.g., Sterling

Commerce’s purchase of Texas Instrument’s CASE tool product development capabilities). We have

also seen this occurring for networks, tools and technologies that relate to the World Wide Web (e.g.,

Netscape’s acquisition by America Online for the former’s installed base of users, Amazon.Com’s

purchase of Junglee.Com for its XML development capabilities, etc.). In a similar vein, the cellular

communications industry has experienced a number of consolidations, with the result that already big

players have grown dramatically bigger.

On the financial services side, the focus of the research presented in this paper, the market for retail

electronic banking services has experienced a great deal of change in the last decade. And much

additional change is expected, as broader corporate and consumer acceptance of the Internet makes it a

more attractive distribution channel for various forms of e-payments. For example, we see the new

electronic bill presentment and payment network providers, Transpoint (www.transpoint.com) (a joint

venture of Microsoft and bank consortium-owned First Data Corporation) and Integrion

(www.integrion.com) (also owned by a consortium of banks and technology firms), jockeying for

position in the business-to-business financial services marketplace, with standards for e-payment

middleware and transactions exchange. Meanwhile, electronic banking networks have been merging with

and acquiring one another at a frantic pace, in their quest for scale size, market share, installed bases of

cardholders, and regional dominance. The result is that the average network has increased in size, and

although retail electronic banking systems usage by consumers has expanded even more dramatically,

today fewer and fewer electronic banking networks exist (O'Keefe, 1994).

Consider the comments offered Competition in ATM Markets: Are ATMs Money Machines? a recent

policy debate report by the Congressional Budgeting Office (1998), which has been examining the

electronic banking market to determine whether there is evidence of consolidation leading to excess

market power, and unfair consumer surcharges for ATM use:

The desire for interconnection that has played a role in decreasing the number of competing networks in the [automated teller machine] ATM market is similar to but not necessarily the same as what economists refer to as network effects [italics added for emphasis]. In that phenomenon, a connection to a network becomes more valuable as -- in this instance -- the number of ATMs connected to the network rises. Network effects have been observed for telephones and, in recent years, for the Internet. … The case of ATMs is slightly different from that of telephones and other instances of conventional

network effects. Telephone users typically wish to be able to make telephone calls to any and all other telephone users. By contrast, ATM cardholders want ATMs to afford them access to their money, but they do not require that every ATM be connected to every network. Indeed, most commercial districts in cities will have several competing ATMs in close proximity, and in that situation, a person would need only a proportion of them to accept his or her card. … Network effects at the level of the ATM owner may be stronger than such effects for cardholders. ATM owners, especially those whose income derives from surcharges, want to be "connected" to as many potential customers as possible. Yet even before surcharges, ATM owners sought the highest level of transactions possible for each machine, consistent with the lowest cost, by joining the biggest network available to them. That desire for connectivity works against a new firm in the network market, because that firm might not be able to offer as many connections as an established or consolidated firm. (Excerpted from Congressional Budgeting Office, “Network Effects and Competition in the ATM Market,” Chapter 4, in Competition in ATM Markets: Are ATMs Money Machines? U.S. Government Printing Office, Washington, DC, July 1998.)

Network Externalities. Each of the industry scenarios that we mentioned above shares an important

feature: installed base appears to give rise to network externalities that create value for users who adopt

common solutions and buy into shared technological standards. Although much has been written on the

adoption of technologies and networks in the presence of externalities by economists (e.g., Economides,

1996; Farrell and Saloner, 1985; Katz and Shapiro, 1985; Oren and Smith, 1981; Tirole, 1988) and IS

researchers (e.g., Chismar and Meier, 1992; Clemons and Kleindorfer, 1992; Riggins, Kriebel and

Mukhopadhyay, 1994; Wang and Seidmann, 1995), the literature is largely theoretical, with just a few

detailed empirical studies focusing on network externalities (e.g., Brynjolfsson and Kemerer, 1996;

Kauffman, McAndrews and Wang, forthcoming). The theoretical models of network externalities

proceed from a simple assumption: that the benefits of a network product or service depend upon the

number of participants. From there, they proceed to a variety of conclusions. However, there still is a

lack of research that illustrates the measurement and extent of network externalities. Moreover, the

examples that are frequently cited in the analytical papers to illustrate the key points often tend to be a

combination of anecdotes and speculation (Liebowitz and Margolis, 1994; Gowrisankaran and Stavins,

1999). And, almost no empirical research has been done on network evolution and the rationale for

interorganizational consolidation.

Explaining Network Survival. Network survival is of special interest to us: little is known about

how network externalities fit explain the extent to which networks survive. Why do some networks go

out of business or merge with others to maintain their viability in the marketplace, while others remain

competitive and evolve to become dominant? Do network externalities create valuable leverage or make

a network more survivable? With the economic theory of network externalities in mind, this research

examines the determinants of network survivability. We will argue that differences in network attributes

will lead to different valuations of networks by participating firms, and, as a result, that such attributes

may influence a network’s likelihood of success. We employ an econometric estimation technique called

duration modeling (Kalbfleisch and Prentice, 1980; Hosmer and Lemeshow, 1999; Josefek and

Kauffman, 1998) from healthcare and biomedical research. We use this to analyze a panel data set

involving variables that describe network growth and operation in retail electronic banking services. Our

empirical results are supportive of the theory on networks: we find that networks with a larger installed

base, higher growth rates, and greater operational efficiencies, and networks new to the market, as

expected, tend to survive. The results also suggest the importance of network externalities, which create

value for consumers, adopting firms, and network providers. Finally, our results support the idea that

network goods exhibit positive feedback effects: larger networks are often expected to prevail in the

market and they do indeed tend to grow larger and become dominant (Shapiro and Varian, 1998).

This paper is structured in the following way. We next discuss the literature on networks and identify

variables that may influence the likelihood of network success. We develop a series of hypotheses

motivated by network externality theory that relate to network evolution in retail electronic banking

services. A discussion of the data set and the econometric estimation technique used to test the

hypotheses follows. We then discuss the estimation results, and conclude with the implications of this

research and suggestions for further research.

2. LITERATURE, CONTEXT AND HYPOTHESES

Extensive efforts have been made in the field of economics to develop theory that can explain and

predict the manner and speed with which network technologies and network goods are adopted in and

diffused throughout the economy. This section introduces relevant research findings that provide a

theoretical basis for the development of a series of hypotheses that enable us to understand why networks

survive.

2.1. The Economic Theory of Network Externalities

One of the fundamental features of network goods is the consumption benefits that they confer upon

the consumer. These benefits are usually called positive consumption externalities or network

externalities (Shapiro and Varian, 1998). The externality value of a network to a potential adopter

depends upon the adoption behavior of others, and, according to the theory, it increases as the network

expands. Due in part to the presence of network externalities, a potential network subscriber will care

about the likelihood of success of a network, especially in terms of the future installed base of consumers

or members who use it. The reason is that future installed base -- and an adopter's perception of what it

will be -- are good proxies for the success of the network. As a result, the dynamics of network goods

adoption and evolution are fundamentally different from those of conventional innovations that offer few

or no externality benefits.

Critical Mass. The minimal viable size of a network is referred to as its critical mass in the

literature. Critical mass theory (e.g., Dybvig and Spatt, 1983; Oren and Smith, 1981; Rohlfs, 1974)

suggests that, in a monopolistic market with reasonable network subscription prices, three different

network sizes can be equilibrium outcomes: a zero size network, an intermediate size that is unstable and

contestible, and a large stable one. The theory predicts that the market will select the largest of the three

equilibrium network sizes, and thus networks of small sizes will not be observed in equilibrium, in the

long run. Based on this theory, one would expect that either a network is able to obtain an installed base

equal to the largest equilibrium network size, or it will have to exit from the market if it cannot surpass

the critical mass and become self-sustaining.

Networks new to the market, as a result, can be expected to experience start-up problems, and may

find it difficult to achieve sustainable competitive advantage, unless they have a significant initial

installed base to begin with, for example, through an acquisition or a strategic alliance. Another

possibility is that the market may expect the new networks to be able to generate critical mass, because

they control a sponsored technology associated with a large player in the market, or because their

technological innovation is revolutionary rather than evolutionary (Shapiro and Varian, 1998). However,

there has been little empirical validation of these proposals, though there are many anecdotes that suggest

they are true. What is most telling, however, is that critical mass theory does not shed light on the

competitive environment in which multiple networks compete against each other.

Installed Base. The installed base literature provides a sharp contrast. Farrell and Saloner (1986)

developed a theoretical model involving the concept of installed base that offers two important

implications. First, because of economies of scale and positive externalities, network goods have a

greater tendency towards monopoly and concentration than services that do not generate externalities.

Second, the strength of the network externalities that accrue from an existing installed base may lead to an

inertial effect that results in social choices of inferior network technologies. Similarly, when the market

perceives that a new network is likely to achieve installed base dominance over another, there may be a

rush to adopt the new network – the so-called bandwagon effect – even if the technology it offers is not

the technically or socially optimal. The authors conclude that network externalities have strategic

implications for technology adoption, predatory pricing, and product pre-announcements. The nature of

the evolution that casual observation reveals for industries involving network goods is generally

consistent with the prediction made by Farrell and Saloner’s model. However, there has been little work

that has yielded empirical evidence for why some networks and network goods fail, while others have

become dominant or develop into a standard within the marketplace.

Compatibility. Katz and Shapiro (1985) suggested that the value of network goods depends on the

number of consumers who adopt compatible products in the future, in addition to the installed base that

has been locked in. Their model allows for the possibility that new networks starting with very small

installed bases of participants may nevertheless grow and become the preferred choice in the market. For

example, the provider of an existing network, facing the possibility of going out of business, may wish to

connect to, or affiliate with the new network. Under such competitive conditions, a newly weakened

player, even though its place in the market may be long-standing, may prefer compatibility.

Under other conditions, compatibility may emerge as the preferred choice for multiple network

providers, for example, when both are similar in size and share other competitive characteristics. It can

also happen that networks that are new to the market may see very few firms adopt, even if the newer

networks offer some transaction processing and infrastructure cost or other technological advantages.

The perceived value of the network externalities associated with a standard in the marketplace may be

very large, outweighing the cost or technological advantages that might accrue as firms switch. Clearly,

some player will have to make the first move to adopt the new network, and whether that choice turns out

to be a wise one is largely determined by the adoption behavior of firms that follow later (or fail to adopt).

Moving to take advantage of the likely equilibrium outcome may be more prudent than trying to shift it.

Related Perspectives in the IS Literature. Following Farrell and Saloner (1985) and Katz and

Shapiro (1985), a number of theoretical models that consider network externalities have been developed

by various authors for the IS literature (e.g., Chismar and Meier, 1992; Clemons and Kleindorfer, 1992;

Riggins, Kriebel and Mukhopadhyay, 1994; Wang and Seidmann, 1995). These models generally build

on a simple installed base model, assuming that the benefits of a network product or service depend upon

the number of participants, and yield the general conclusion that network externalities should be an

important element in the firm-level and consumer-level adoption decisionmaking and valuation of

network goods. With different sizes of installed bases, we should expect to observe different levels of

willingness-to-pay on the part of potential adopters.

A number of empirical studies that focus on network externalities have begun to appear also,

providing some practical evidence of the existence of network externalities in the markets for spreadsheet

software products and electronic banking services. For example, Gandal (1994) found that consumers

were willing to pay a significant premium for spreadsheet packages compatible with the Lotus platform.

In addition, Brynjolfsson and Kemerer (1996) found a positive relationship between a spreadsheet's

market share and its price, which they interpreted as evidence for the existence of network externalities.

Other work offers evidence that is more closely related to the present research. Kauffman, McAndrews

and Wang (forthcoming) offered evidence about the role that network externalities may play in the timing

and extent of adoption of a regionally shared electronic banking network. Gowrisankan and Stavins

(1999) found evidence of network externalities in the Federal Reserve Bank’s regional Automated

Clearing House (ACH) operations, and used this evidence to make arguments that the Fed’s pricing for

this capability (which provides for automated handling of corporate payroll payment transactions to

employees’ bank accounts) should not be subsidized by taxpayers. Finally, Kauffman and Wang

(forthcoming) examined the growth patterns of CIRRUS and PLUS, the primary nationally shared retail

banking networks in the United States, and the impact of compatibility between the two networks. Their

results suggested that a “glass ceiling” on the growth of network externalities for each of the networks

individually led them to provide universal switching of one another’s card transactions across the two

networks, as a strategy to further grow network value in a network value-constrained marketplace.

Examining Assumptions for the Empirical Research Context. There is still much work needed to

provide empirical validation of the propositions suggested by the theoretical networks literature. The

theoretical models involving network externalities generally make a number of simplifying assumptions

that may not hold in common empirical settings. We caution the reader not to jump to the wrong

conclusion though: the simplifying assumptions that are used in some of these models enable specific

issues to be understood in a simpler world – and very often one that has all of the salient features of the

real world that are needed to explain some observed phenomena. Some of the more common assumptions

that we are referring to include:

• Networks are homogeneous in a range of features (cost, product and service base, geographical

coverage, etc.).

• Consumers are homogeneous in their valuation of the network externalities that are offered by

individual or competing networks.

• Price-inelastic demand exists for network-delivered goods and services.

• Consumers are able to make “sage” predictions of the future evolution of the market for network

goods and services, with reference to network sizes and installed bases, prices, and the dates of

any subsequent network introductions.

With these assumptions underlying the results of some of the theoretical models in the network

externality literature, not much is really understood in empirical terms about why networks actually

succeed. Based on prior empirical research in related financial services contexts, we know that some of

these assumptions may over-simplify the rich fabric of firm competition and the changing industry

structure. Consider some of these examples.

• Most networks in the financial services industry, where the creation of physical infrastructure in

different geographical regions is required, are actually heterogeneous in their features

(McAndrews, 1991).

• Consumers, as the assumptions in the analytical models go, are probably relatively homogeneous

in their valuation of network externalities, however, banking firm valuations of network goods

are often strategically heterogeneous – and these firms are often the key players when it comes to

adoption decisions about networks that lead to their success and failure (Kauffman, McAndrews

and Wang, forthcoming);

• Today, there are multiple distribution channels for various electronic payment services (credit

cards, ATMs, phone banking and the Internet), with the result that price-inelasticity of demand

for a specific electronic payment service may no longer be an appropriate characterization of the

competitive landscape. Instead, we expect to expect price elasticity across payment services

distribution channels.

• Finally, the marketplace for financial services networks is highly complex due to digital

convergence (Yoffie, 1998), involving firms such as Microsoft, IBM, Compaq, Intuit, Oracle and

others, as well as the major banks The technology firms challenge traditional financial industry

players with new technology-focused capabilities (e.g., Compaq’s Millicent micropayments

technology solutions, www.millicent.com, and the Internet Exporer 5.0 browser-based “electronic

wallet”), making it difficult for consumers and senior banking firm managers to predic t a future

that only the state of technology appears able to control.

Thus, it is appropriate to consider some of the following questions. Is network installed base an

important factor influencing the success of financial services and other networks? If so, to what extent?

How could that be shown? Are there other variables influencing the likelihood of network survival?

How important are they relative to the network installed base? And what about the externalities that may

accrue to the consumers of network services, and to network members who take advantage of the

participation of other firms? Can one predict network success with these things in mind?

We think that the economic theory of network externalities provides a useful starting point for our

exploration of survivability in a specific context: regionally shared electronic banking networks

(McAndrews, 1991). The theory also provides a basis for the specification of testable hypotheses that

will enable us to understand how these networks, and the structure of regional electronic payments

infrastructures, tend to evolve.

2.2. Context: Regionally Shared Electronic Banking Networks in the United States

The deployment of electronic banking systems by banking firms in the United States became popular

in 1970s, and many of the investments in this form of customer access information technology were made

on a proprietary basis: no sharing of network infrastructure was planned or envisioned at the time.

However, as competitive pressures developed in the early to mid-1980s, banking firms that were late to

adopt or found themselves to be relatively high cost providers of network services to their banking

customers came to recognize the business benefits of sharing infrastructure and offering customers retail

banking services that were made available on an anytime, anywhere basis. As a result, by 1987, the only

significant installed bases of automated teller machines (ATMs) that were not shared were the proprietary

systems of five relatively large banks: Bank of America, Citibank, First Interstate, Security Pacific and

Wells Fargo. In addition, many banking firms came to view electronic banking service infrastructures as

being utility-like: a serious competitor had to offer its customers access to electronic banking services,

either by owning network assets itself or by outsourcing to obtain them.

The Antecedents of Network Externality Value. Shared electronic banking networks are known to

be subject to strong scale economies (Saloner and Shephard, 1995; Kauffman, McAndrews and Wang,

forthcoming). By setting up a shared computing facility, a large number of network transactions can be

processed and switched more economically. On the demand side, shared electronic banking networks

exhibit two kinds of externalities:

• Locational convenience is enhanced as the size of the network expands, and more ATM locations

provide broader service coverage.

• Direct revenue opportunities are also provided in the form of switching and network interchange

fees that accrue to network members who take advantage of the opportunity to provide retail

banking transaction services to an expanded card holder base.

Shared electronic banking networks generate significant network externalities because existing

network member banks value the network more highly as new members join, bringing more ATMs and

card holders to the network. A shared ATM network can be thought of as a collection of compatible and

complementary goods. Compatibility is achieved through the use of standard ATMs, with standard size

bankcards that share a similar magnetic stripe technology, and by bringing together a group of banking

firms which agree to employ technologies that can process transactions in a similar manner.

Complementarities are created when a network provider offers a customer the means to consume multiple

banking services via the shared network (e.g., checking and saving account withdrawals, balance

payments, credit card bill transfers, and so on), enables connections to multiple banks, and allows for the

use of different kinds of plastic cards (e.g., ATM cards, debit cards or credit cards).

Network Growth, Network Consolidation. Sharing ATMs among banking firms became a widely

accepted practice, leading to increasing scale economies on an industry-wide basis. By late 1989, 175

shared networks were operating in various regions of the U.S. Since that time, however, we have seen

extensive network consolidation, as networks have been merging with, and acquiring one another in an

effort to increase network size and geographical coverage to maintain viability in the marketplace. (See

Figure 1.)

Figure 1. Consolidation of the Market for Shared Electronic Banking Services

020406080

100120140160

1970

1972

1974

1976

1978

1980

1982

1984

1986

1988

1990

1992

1994

1996

Number of Shared Networks

Year

Source: Compiled from Bank Network News, various issues.

As a result, by the end of 1996 there were fewer than sixty shared networks still operating in the

United States. As the industry continues its push to serve more customers with network operations that

offer broader geographic coverage, larger scale size, greater product availability, more value-compatible

ownership and governance arrangements, and lower operating costs, different observers have speculated

over the years that industry-wide consolidation will cause the number of shared networks to shrink to as

few as 10 to 15 networks (Bank Network News, various issues).

2.3. Hypotheses

Consider the issues that a firm will have to deal with when it faces a decision about electronic

network adoption. Questions involving “When is the right time to adopt?” and “Which network should I

go with?” are among the very first that come to mind. Typically, senior management decision makers

recognize that evaluating when the timing is right to adopt can be boiled down to a series of deterministic

costs, and most of these must be expended up front. But, in addition, there will be stochastic future

benefits, few of which can be estimated without some lingering uncertainty (Benaroch and Kauffman,

1999 and forthcoming). As a result, senior managers generally tend to express a preference for networks

that have, or are expected to achieve, a larger installed base. In a word, senior managers and electronic

banking sstrategists view value as an increasing function of installed base.

Network Externality Drivers of Network Survival. McAndrews (1992) suggested that a large and

growing installed base of member firms enables electronic banking strategists and Federal Reserve Bank

regulators to make relatively good predictions about whether an electronic banking network will become

dominant. He argued that a network will be perceived to be more valuable if it has a larger installed base

of network ATMs and bankcards, and broader geographical coverage. ATMs and bankcards are

complementary goods, as we discussed earlier, and one expects that the complementary goods should

influence market perceptions of network value, just as a pure count of installed base of member firms

would. Comments along these lines are offered by the recent Congressional Budgeting Office (1998)

study:

Networks … are merging not to get each other's hardware but each other's clients. A corollary to that proposition is that it is not the need to install a large infrastructure that is hindering new networks from forming but rather banks' contractual ties to incumbent networks. (Excerpted from Chapter 4)

Moreover, the more ATM locations and broader service coverage that are provided due to the

expansion of a network’s installed base, the greater will be the value of the externalities which are

conferred upon consumers who use the network. This leads to a greater willingness-to-pay on the part of

member firms and their customers for the network services, so that larger networks are better able to

survive and prosper. Our first three hypotheses (H) follow the logic of this kind of network externality

theory directly:

• PHYSICAL INFRASTRUCTURE HYPOTHESIS (H1). Larger networks – in terms of the number of

ATMs deployed – tend to survive.

• COMPLEMENTARY ASSETS HYPOTHESIS (H2). Networks with larger complementary assets – in

terms of the number of bankcard holders – tend to survive.

• GEOGRAPHICAL COVERAGE HYPOTHESIS (H3). Networks that offer broader geographical

coverage tend to survive.

All firms in the marketplace can observe electronic banking network adoption and growth. As Katz

and Shapiro (1992) have suggested, consumer expectations in markets with network externalities are

important in that they influence other consumers’ adoption behavior. When the observed level of

adoption of an electronic banking network in a market is high, this is also likely to increase banking

firms’ willingness-to-pay for network membership, as Shapiro and Varian (1999) would suggest, since all

the signs increasingly point to the emergence of the most rapidly growing network as the dominant

network in the future. Thus, modeling firms’ expectations regarding future network size has been a

critical, if difficult aspect of the development of the theory of network externalities. By assuming

homogeneity in potential adopters’ expectations about perceived network size in empirical research on

this problem, we may fail to capture the basis upon which their decisions are actually made. The

implication is that if we were able to characterize the expectations of individual firms that may adopt, we

would find that their different propensities to adopt are equivalent to greater (or at least heterogeneous)

levels of willingness-to-pay, which means they ascribe a higher market value to the network provider.1

1 Again, we stress that in the present research we do not assume that firm expectations with respect to network value are similar, or that firms would similarly value a given level of installed base or growth, even if they did agree upon the relevant precursors to

Thus, the network provider will have clear monetary incentives to announce rapid network growth to

condition potential adopters’ expectations about the future size of the network, thereby “adding fuel to the

fire” in Shapiro and Varian’s positive feedback loop terms, and pushing the market towards bandwagon

adoption.

Firms may use several related indicators to evaluate if a network will become dominant. For

example, network pricing is a useful signaling mechanism that a network provider can employ to

influence firms’ perceptions and expectations. Networks that are able (or expected) to obtain a large

installed base in the future may wish to implement an aggressive pricing strategy early on, conditioning

the market’s expectations that the network will soon be able to wield the market pricing power of a

dominant player. Xie and Sirbu (1995), in contrast, emphasized the importance of current network

installed base in their value-based pricing model for network services. Networks that achieve such

market power may wish to maximize profitability by setting a higher subscription price. Unfortunately,

we were unable to acquire any useful sources of pricing data for use in this study.

Another proxy variable for the potential future dominance of a network is measured by its rate of

growth. Senior managers at electronic banking industry firms whom we interviewed told us that firms

that are afraid of jumping onto the wrong bandwagon by adopting the wrong network use the growth

trend of a network to evaluate its potential. Thus, firms may wish to examine the observed growth paths

of the competing networks with respect to which an adoption decision is to be made, and use this

empirical data (rough and imperfect though it may be) as a basis for predicting which network will

become dominant. They may also decide that none of the networks has the “right stuff,” in spite of the

other indicators (for example, in terms of operational infrastructure, innovative products, attractive

network governance and organizational structure, or geographical base) to emerge successful from the

competitive fray in the near term. In this research, we will use the network growth rate in ATMs as an

indicator of the future potential success of a network.2 Network externality theory predicts that:

• GROWTH RATE HYPOTHESIS (H4). The more rapid the growth rate of a network, the more likely

it will be to survive.

Non-Network Externality Drivers of Network Survival. In addition to variables influencing the

perceived network externalities, several other factors may influence the likelihood of network survival.

For example, a solid reputation for quality service, product innovation and a strong brand or make identity

(all of which ought to increase willingness-to-pay) can be expected to increase survivability. Katz and

such value. In the present research, we employ the somewhat weaker notion of heterogeneous beliefs about network externalities: namely, that they will probably differ by firm and by network, but that they will be similar across the entire geography that a regionally shared electronic banking network serves. We permit no spatial variation in perceptions of externality value in this research. 2 This allows for the possibility that networks with a small installed base could still grow and succeed in a competitive market.

Shapiro (1985) suggested that firm reputation plays a major role in network evolution. Thus, we expect

that the make effect should be an important factor influencing network adoption. Brynjolfsson and

Kemerer (1996) found that the make effect contributed to the perceived value of Lotus spreadsheet. They

argued that because of an early and dominant market position, Lotus was perceived as a market leader.

Thus, the long-term viability of Lotus was unlikely to be perceived as a problem (in spite of what would

actually happen later as Microsoft built muscle with its Office Suite), and consumers making purchase

decisions felt they could assume that they would get continued support and further enhancements.

In our context, it is difficult to capture this kind of historical data for electronic banking member

firms -- in some cases for networks that no longer exist. Instead, we must use a proxy variable. We have

selected the number of years a network has been in business as a proxy for reputation. Networks that were

early innovators and have been long-standing participants in the market have positioned themselves to be

recognized as market leaders; they probably will be perceived as having the kinds of business capabilities

that will enable the network to be viable in the long run. Thus, we expect that:

• MAKE EFFECT HYPOTHESIS (H5). The stronger the network reputation, the more likely the

network will be to survive.

Farrell and Saloner (1986) suggested that markets sometimes are biased either for or against new

networks, depending on the relative costs of network adoption. In a later paper, Katz and Shapiro (1992)

also suggested that network operating costs influence the likelihood of network success. Regardless of its

size and growth record, the operating efficiency of a network (in terms of revenue generation and cost

structure) is critical in determining whether it can make profits and continue to operate. We expect that

networks with greater operating efficiencies will be more likely to prevail in the marketplace. What are

some of the appropriate operational indicators?

Electronic banking network service providers can be viewed as profit maximizing firms that exhibit

economic rationality in a range of operational decisions that are undertaken in transaction switching and

the provision of network services. One of the mandates of this business in the last twenty years has been

to grow transaction processing scale at the switch level, enabling the network to take advantage of

operational efficiencies derived from larger scale size. Similar to other computer service businesses, such

as Internet service providers (ISPs) and outsourcing transaction processing providers, shared electronic

banking networks require investments in telecommunications, hardware and software infrastructure that

must be made in discrete and relatively large increments. Often the networks that process the largest

number of transactions are best able to drive down their per transaction operating costs, by spreading their

cost overhead. In this context, the number of transactions processed by a network ATM is a reasonable

proxy for the cost structure of a network, since it provides a means to gauge operational efficiency. This

measure can also be used to gauge the average operating costs of member firms’ ATMs: the larger the

number of network transactions per ATM, the more likely member firms are able to profit from network

adoption, and the more member firms will value the shared network. The operational cost savings can be

passed on to adopting firms (and still margined by the network provider), increasing their willingness-to-

pay for network services and adding to perceived network value. This leads us to state the next

hypothesis:

• OPERATIONAL SCALE SIZE HYPOTHESIS (H6). The larger the number of transactions processed

per network ATM, the more likely will be the network to survive.

Shared electronic banking networks charge member firms fees for ATM transactions that are

switched between banks, and, to the extent that such charges are passed on, member firms can earn

interchange revenues when their ATMs are used by other member firms’ customers. (Whether a firm

wishes to pass those charges on to its own customers is a separate matter and we do not consider it here.)

Thus, networks with a higher percentage of transactions that require interchange, a trend that has become

increasingly evident over time in the U.S., are more likely to be perceived by adopting firms as the more

attractive and valuable choice. (See Figure 2.)

Figure 2. Interchange Pe rcentage for 60 Top Shared Electronic Banking Networks

0

10

20

30

40

50

60

1980 1982 1984 1986 1988 1990 1992 1994

Interchange Transactions as a Percentage of Total Network Transactions

Year

Source: Bank Network News, various issues.

On the other hand, networks that switch relatively fewer transactions may not be properly positioned

in their markets to capture interchange revenues – either due to the geographic scope of the network they

operate or to the disposition of their users. In both cases, the externality value for the network ought to

be somewhat lower. Thus, we predict that:

• TRANSACTION SWITCHING HYPOTHESIS (H7). Networks with a higher percentage of switched

ATM transactions are more likely to survive.

The final aspect that we consider which may have an impact on the likelihood of success of a network

is its ownership structure. Nault and Dexter (1994) examined transfers and incentives in a franchise

network and suggested that under certain conditions there may be an underinvestment problem: not all

network members will have the incentive to fully invest in marketing efforts, leading to outcomes that are

sub-optimal for the network originator. Bakos and Nault (1997) modeled the impact of alternative

ownership structures on the investment incentives of member firms in interorganizational information

systems. They suggested that network ownership structures have important implications for economic

efficiency; in many cases, joint ownership produces greater network value than sole ownership due to the

increased investment by all network participants. According to their game theoretic interpretations, joint

ownership and governance is more economically efficient for networks in the absence of an indispensable

network participant, with mutually important participants, or with assets that are essential to all network

participants.

In the context of shared electronic banking networks, network ownership and governance structure is

also viewed as a predictor of network success or failure. As in the general case that Bakos and Nault

discuss, joint ownership often leads to market-wide perceptions of greater network value than sole

ownership. Regionally shared electronic banking networks have been owned by a single firm (as in the

case of Philadelphia National Bank’s MAC network (later owned by Core States Financial, and then

parceled out in the marketplace to become EPS), or jointly owned by all or some of its members (such as

the New York Case Exchange, NYCE, and Yankee 24 of New England), and managed as a for-profit or a

not-for-profit organization (Bank Network News, various issues). In addition to lack of incentive to invest

by their network members (the agency theoretic problem of imperfect monitoring), networks with sole

ownership may not be as well accepted as those with joint ownership, since adopting firms may be afraid

of being exploited by a single network owner once they participate and become dependent. Thus, we

expect that:

• GOVERNANCE HYPOTHESIS (H8). Networks with sole ownership are less likely to survive.

We next discuss the method and data that are used to test these hypotheses.

3. METHOD AND DATA

To test our hypotheses, we employ a duration model (also known as a failure-time model or a hazard

model) developed by healthcare practitioners, biomedical statistics researchers and other social scientists

who were interested in predicting such phenomena as the efficacy of alternative cancer treatments, the

lifespans of cigarette smokers, the duration of unemployment for skilled workers, and the likelihood of

financial distress and corporate bankruptcy (Hosmer and Lemeshow, 1999; Kalbfleisch and Prentice,

1980; Kiefer, 1988; Peterson, 1991). This approach enables the researcher to model the effects of

explanatory variables on the occurrence of an event (e.g., smoking related death or corporate bankruptcy),

or, alternatively, the duration of time that passes prior to an event (e.g., months of life prior to death from

cancer, or time to bankruptcy after cash flow inadequacies have been recognized by a firm’s bankers).

Duration models have also been used to study the diffusion of a variety of technological innovations

outside the IS literature (e.g., Hannan and McDowell, 1984; Levin, Levin, and Meisel, 1987; Rose and

Joskow, 1990). Mahajan, Muller and Bass (1990) suggested that duration models overcome the

limitations of prior work on technology adoption in an important way: they eliminate the need to rely on

cross-sectional models only. They have also been used elsewhere in the IS literature, for example,

(Kauffman, McAndrews and Wang, forthcoming) to gauge the power of network externalities in

explaining corporate network adoption decisions, and to study the behavior of IS professionals in terms of

their separation from or retention by the firm over time (Josefek and Kauffman, 1998).

3.1. The Duration Modeling Approach

Through the lens of duration modeling, we can view the survivability of a network in terms of

whether network default occurs. Network default need not mean bankruptcy. For example, network

default can result from outright financial failure, when a network ceases to operate, and from other

circumstances, for example, when it is value-maximizing for some reason for a network to become

involved in a merger or an acquisition, and undergo a name change or an ownership change. In the latter

case, a network’s physical and financial assets, as well as its ATMs and bankcard holders, are blended

with those of another or several other firms; its independence disappears.

Duration and the Concept of Failure Time. In the econometrics literature, the terms duration and

failure time are often used interchangeably. We will use the term failure time here because the term is

convenient to convey the idea that the dependent variable on which our interest centers is the period of

time, or duration, it take for a shared electronic banking network to “fail” to operate independently, or

default.

To study network default, we define the failure time, T, for network default as the elapsed time from

the start of observation of the network until the network ceases independent operation, or the observation

period ends, whichever is earlier. T is conditional on a set of explanatory variables, and it follows a

distribution, which can be described in one of several ways:

• the cumulative failure time probability distribution function F(t) = Pr(T<t), specifying the

probability that the random variable, T, is less than some value t;

• the corresponding probability density function, f(t) = dF(t)/dt;

• the survivor function, S(t) = 1 - F(t) = Pr(T≥ t), defined as the probability that a network will not

have defaulted prior to the end of period t; and,

• the hazard function, h(t) = f(t)/S(t).

These four specifications are mathematically equivalent to one another.

As a convenience, we employ the hazard function, h(t) = f(t)/S(t), which enables us to take advantage

of state-of-the-art econometric techniques (Hosmer and Lemeshow, 1999). We assume that T is a non-

negative random variable representing the failure time of a shared electronic banking network from a

population of n independent networks. The hazard function, hi(t), specifies the conditional probability

that network i will default at time t, given that it has not done so in the prior period, t-1. We further

assume that failure time is conditional on observable characteristics of a network in the context of its

marketplace, such as we discussed in the comments leading up to this study’s hypotheses. Based on these

assumptions, we are able to limit our consideration to a subset of parametric failure time models; they

require the analyst to make additional assumptions about the distribution of failure times.

Censoring and Time -Varying Covariates. There are two other compelling reasons for using a

duration model in this research. First, duration models accommodate censoring. The reader should note

that our observations of network default times are right-censored: not all networks defaulted by the end of

the study period and some networks continued to operate unchanged. Second, duration models are able to

handle time-constant and time-varying covariates of network survival. Some of the explanatory variables

for network default that we propose (e.g., network installed base in terms of ATMs, ATM cards and

market coverage, interchange ratio, number of transactions per ATM, and network growth rate) can vary

over time, while others remain fixed (e.g., network governance structure and the number of years a firm

has been in business at the beginning of the period that is being analyzed). By estimating the duration

model using network growth data, we will be able to empirically assess the impact of the explanatory

variables on the probability of electronic banking network default (in terms of the duration of network’s

survival in the market).

Gauging Model Fit. The reader should note, however, that econometric analysis of a duration model

involves computing maximum likelihood estimates (MLE) of the probability of the occurrence of an

event, conditional on a unit change in an explanatory variable. Although this method produces t-statistics

indicating the relative significance of the explanatory variable on the time to network default, it does not

produce R-squares indicating the explanatory power of the model, as do ordinary least squares (OLS) or

generalized least squares regression (GLS) (Hosmer and Lemeshow, 1999; Kalbfleisch and Prentice,

1980). Instead, log-likelihood values for the estimations are used to indicate the extent to which the

model as a whole applies. Interpretation of the model is subject to these considerations, as a result.

Distributional Assumptions. Several probability distributions for the duration to network default

are possible. The choice of a particular parameterization (e.g., Weibull, exponential, normal or logistic)

depends on empirical considerations as discussed by Sinha and Chandrashekaran (1992). In this case,

our interest centers on estimating a time-varying hazard rate , as we will shortly make clear.

Among them, the exponential distribution offers the simplest model for our network default data.

When an exponential distribution for T is assumed, h(t) = γ and S(t) = exp(-γt), with γ > 0. The downside

of this model is that it assumes a time-invariant hazard, disallowing time-varying estimates of the

likelihood of network default, a key consideration for our empirical analysis. It doesn’t match our

modeling needs very well. By using time-varying explanatory variables in the model, one expects that the

hazard rate will change as the values of the explanatory variables change over time. Hannan and

McDowell (1984) incorporated the effect of time-varying covariates, Xit, on the hazard function in their

research on the adoption of proprietary electronic banking networks with hi(t) = γ = exp(-ßXit). In their

expression, Xit, is a vector of observed explanatory variables relevant to period t adoption for firm i, and ß

is a column vector of unknown parameters that may be interpreted in terms of the relationship between

explanatory variables and the conditional probability of adoption. As a result, the relative probabilities of

adoption across firms change through time.

Using the survivor function described above, the probability that network i will default during a

period T prior to the end of the study period can be shown to be Si(T-1) - Si(T). The likelihood of

network default is also a function of the values of the explanatory variables that obtain in each period up

to the end of the study period or the time of network default, depending on which occurs first. The

likelihood function is expressed as follows:

L S T S T S Tii

n

i jj

n

= − −= =

∏ ∏[ ( ) ( )] [ ( )]*

1 1

1 2

1

In this expression, T* is the end of observation period, and n1 and n2 denote the number of electronic

banking networks that defaulted by T*, and did not default by T*, respectively. The estimation procedure

we use will yield parameter estimates for ß in the hazard function that maximize a likelihood function

composed of the above probabilities (Greene, 1999). Following Hannan and McDowell, we consider a

hazard function, hi(t), which defines the conditional probability that network i will default at time t, given

that it has not done so by t-1, to stress the time-varying nature of the phenomenon that we are studying.

In contrast to the work by Hannan and McDowell, however, we employ the Weibull distribution and a

functional form for the hazard function that allows the hazard rate to be a function of time directly: hi(t) =

γαtα-1 with γ = exp(-ßXit) > 0 and α > 0. The Weibull distribution is a generalization of the exponential

hazard function, for which α = 1. (Recall that hi(t) = γ = exp(-ßXit) for the exponential distribution.) It

has been widely used in modeling technology diffusion (e.g., Rose and Joskow, 1990 and Saloner and

Shepard, 1995). In the expression for hi(t) above, Xit is a vector of explanatory variables for shared

network i. β i s a column vector of unknown coefficients to be estimated for the explanatory variables. t

is incorporated, as before, to allow the hazard rate to be a function of time.

Similar to Saloner and Shepard (1995), with γ = exp(-ßXit), this functional form enables us to capture

the effects of the explanatory variables on the hazard function. Maximum likelihood estimates for the

values of the ß’s can be used to measure the impact of the various explanatory variables on hi(t), as the

constant proportional effect on network i's conditional probability of default, by applying logarithms on

the both sides of the equation. And, with the appropriate transformation of ß, 100*(exp(-ß)-1) can be

given a quasi-elasticity interpretation: it indicates the percentage change in the hazard rate for a unit

change in the corresponding explanatory variable (Helsen and Schmittlein, 1993).

3.2. Explanatory Variables Definitions, Data Sources and Other Estimation Considerations

In the estimation that we will carry out, the duration to network default is operationalized as the time

that elapsed from 1983 until the network defaulted (i.e., was merged into another network or failed), or up

until the end of 1990, whichever is shorter. By 1990, significant change had occurred in shared electronic

banking leading to a far more consolidated industrial structure, and, despite the rapid changes, we still

believe that this period allows us to model network default with the fewest concerns about exogenous

shocks that would occur later. For example, around 1990 and 1991, significant changes were also

underway at the national retail payments infrastructure level, as the CIRRUS and PLUS networks

conducted discussions that led to universality, enabling regional shared electronic banking networks to

become members of CIRRUS and PLUS.3 In related research, we discuss empirical reasons for why

there may have been a regime change in the market for shared electronic banking services, requiring us to

specify a very different model (Kauffman and Wang, forthcoming). In this, and any other research design

that involves the study of a phenomenon that continues to exist, we expect the data to be right-censored.

In our case, right-censoring occurs for networks that continued to operate after 1990. Thus, in estimating

the model, we need to include one additional categorical variable indicating observations that are right-

censored, for which no default occurred during our study period.

The explanatory variables that are used in our model are defined and operationalized as follows:

(1) ATM_BASE: The number of a network’s ATMs, a measure of the size of the physical infrastructure, in terms of network installed base that is observable by adopting firms.

3 Up until that time, CIRRUS and PLUS operated as a duopoly of national networks composed of competing regional shared networks. This operating arrangement is often referred to in the industry as duality. In a maneuver to restrain trade and further their own positions, CIRRUS and PLUS promoted duality, in which no regional networks were allowed to be members of both. Instead, they had to choose between CIRRUS and PLUS, or develop separate (and probably, but not always more costly) sharing arrangements with regional networks in different parts of the country. These changes were indications of the range of industry-wide consolidations that were to occur in the early 1990s, as firms repositioned themselves in the marketplace and sought out improvements in firm-wide operating efficiencies, and a move to universality in the national electronic banking infrastructure, allowing for interchange between the CIRRUS and PLUS networks. This had been prohibited throughout the 1980s due to the national networks’ strategies and policies, creating structural inefficiencies for electronic banking in the American economy.

(2) CARD_BASE: The number of network cards issued (in thousands), a measure of the complementary assets to the network that increase its value.

(3) NET_COVERAGE: A binary variable for the geographic service coverage of a network; 1 if the network covers multiple states, and 0 if the network operates within a single state.

(4) GROWTH_RATE: The rate of annual network growth in ATMs, a measure for the growth potential of a network as perceived by adopting firms.

(5) MAKE_EFFECT: The number of years a network has been in business since it started, a proxy for reputation.

(6) TRANSACTIONS_PER_ATM: The number of network transactions per network ATM,as a proxy for network operating cost.

(7) %INTERCHANGE: The percentage of network transactions that require interchange, as a measure of network indispensability and operating revenue.

(8) OWNERSHIP: A binary variable indicating network ownership and governance structure; 1 if the network is jointly owned by members, 0 otherwise.

Our data set includes time-series data for all regional shared networks operating in the U.S. in 1983.

To be classified as a shared network , no single participant’s cardholder base could account for 95% or

more of network transactions. CIRRUS and PLUS, the national shared networks, were excluded because

of their national scope of operation, and because they acted as a consolidating switch for many of the

regionally shared electronic banking networks. The starting year, 1983, was selected as a base year of

observation for tactical reasons. Almost all regionally shared electronic banking networks that ever

existed in the U.S. were operational by 1983.

The data set also includes information on network default dates, network installed base, network

growth, network ownership, and so on, from 1983 up to 1990, an eight-year time series that we can use to

estimate the duration model we discussed in the prior subsection of this paper. The data were obtained

from various issues of The EFT Data Book , published by Bank Network News, 1983-1990, now owned by

Faulkner and Gray Publishers (www.faulknergray.com/cards/banknn.htm). Bank Network News has

published data on all regionally shared electronic banking networks in the U.S. annually since 1982, and

it is viewed in the industry as one of the most reliable sources of data on electronic banking. As an annual

update of the industry, The EFT Data Book represents the most comprehensive compilation of retail

electronic funds transfer statistics available anywhere. We confirmed the quality of data published there

by using information available in other published studies and Federal Reserve Bank sources, and through

discussions with a senior economist at the Federal Reserve Bank of New York, and senior managers at

shared electronic banking network firms.4 Descriptive statistics and correlations for the variables used in

this study are shown in Tables 1 and 2.

4 Personal communication with Dr. James J. McAndrews, Senior Economist, New York Federal Reserve Bank, who specializes in research and regulatory issues for the electronic payments industry in the U.S.

Table 1: Descriptive Statistics for Explanatory Variables

VARIABLES MEAN STANDARD DEVIATION

ATM_BASE 613.53 1151.10

CARD_BASE (000s) 1337.40 2387.30

NET_COVERAGE .47 .50

GROWTH_RATE .12 .26

%INTERCHANGE .38 .28

TRANSACTIONS_PER_ATM 4856.80 2291.50

MAKE_EFFECT 4.20 2.68

OWNERSHIP .29 .46

Table 2. Correlation Matrix for the Explanatory Variables ATM_

BASE

CARD_

BASE

NET_

COVERAGE

GROWTH_

RATE

MAKE_

EFFECT

TRANSACTIONS_

PER_ATM

%INTER-

CHANGE

CARD_BASE .918

NET_COVERAGE .319 .357

GROWTH_RATE .092 .154 .097

MAKE_EFFECT -.102 -.038 .069 -.042

TRANSACTIONS_ PER_ATM

.144 .187 .021 -.016 -.031

%INTERCHANGE -.014 .030 -.089 -.011 .143 .078

OWNERSHIP .323 .252 .015 .055 .057 .057 -.016

Although the number of ATMs that have been deployed in a network and the number of the

network’s cardholder are complementary aspects that we expect will drive network value, the reader

should note the high correlation (ρ =.918), that we observed. Evidently, the same thing is being

measured. Similar to estimating an OLS regression, estimating a duration model with highly correlated

explanatory variables can be expected to lead to instability in the coefficient estimates. To avoid such

colinearity problems, we used only one of the two variables at a time in model estimation, resulting in

two separate duration model estimates. Our intent is to provide an indication of the range of variation of

the estimated coefficients for the entire model, rather than to argue in favor of the relative explanatory

power of ATM_BASE versus CARD_BASE. In fact, we believe that the lack of either a solid

ATM_BASE, or a significant customer pool reflected by the CARD_BASE, will act to decrease the

survivability of a shared electronic banking network.

4. RESULTS AND DISCUSSION

We next present and interpret the results of our duration model estimations. We also illustrate the

usefulness of the results in interpreting the contrasting evolution of several shared electronic networks,

arguing that the model findings provide significant insights into why network defaults occurred. Then,

we examine potential defects in our results by examining a set of residual plots. We also performed an

analysis of neglected heterogeneity in the data set. We also plan to evaluate the performance of one and

two-year ahead out-of-sample forecasts to gauge the robustness of the model in predicting network

default and survival.

4.1. Maximum Likelihood Estimation (MLE) Results

The estimation results for the duration model with the variable ATM_BASE (excluding

CARD_BASE) and for the model with the variable CARD_BASE (similarly excluding ATM_BASE) are

shown in Table 3.

Table 3. Results for Two Alternate Models of Network Default

EXPLANATORY VARIABLES

IN THE MODELS

PHYSICAL INFRASTRUCTURE INSTALLED BASE

MODEL (ATM_BASE)

COMPLEMENTARY ASSET INSTALLED

BASE MODEL (CARD_BASE)

Coefficient t-ratio Coefficient t-ratio

ATM_BASE .0008 1.95 ** NA NA

CARD_BASE NA NA .00021 1.85 *

NET_COVERAGE .32 1.04 .42 1.38

GROWTH_RATE .75 1.72 * .76 1.67 *

%INTERCHANGE 2.40 3.82 *** 2.25 3.72 ***

TRANSACTIONS_ PER_ATM

.00013 2.69 *** .00017 3.32 ***

MAKE_EFFECT .14 3.46 *** .15 3.29 ***

OWNERSHIP .20 .66 .43 1.33

α .89 6.51 *** .98 6.42 ***

Log-likelihood values for MLE models:

-187.2

-193.4

Significance levels: * p = .10; ** p = .05; *** p = .01

The results from both models are similar in terms of the signs and statistical significance levels for the

estimated coefficients. Using either ATM_BASE or CARD_BASE in the duration model yields

qualitatively similar results, even though our initial interpretation of these explanatory variables was

somewhat different. The log-likelihood values for the models are as follows: for the Physical

Infrastructure Installed Base model, -187.2; and for the Complementary Asset Installed Base model, -

193.4; both indicate a high level of model fit. In addition, in each case, the maximum likelihood

estimates converged rapidly.

We found that the variables ATM_BASE, CARD_BASE, GROWTH_RATE, %INTERCHANGE,

TRANSACTIONS_PER_ATM, and MAKE_EFFECT have significant but negative impacts on the

likelihood of network default. The positive signs of the estimated coefficients on these variables are

indications that the larger their values, the longer is the duration of time to network default. Thus,

networks that possess larger installed bases, grow more rapidly, capture a higher percentage of

transactions that require interchange, process more transactions per ATM, and have been in business

longer are less likely to default.

Next, let’s take a closer look at the results and how they were obtained. To understand how we

arrived at these general findings, the reader should recall that our MLE results for the values of ß can be

used to measure the impact of an explanatory variable on hi(t), as the constant proportional effect on some

network i's conditional probability of default by time t, with the appropriate logarithmic transformation.

And, because we expressed γ as exp(-βXit), to emphasize the time-varying nature of the hazard in

response to changes in the independent variables, X, a positive value for an MLE coefficient in our

results leads to decreasing likelihood of default or merger, and increasing survivability. The reader

should note the negative sign on ß in the expression exp(-βXit), for γ.

The finding that CARD_BASE (ß=.00021, t=1.85, p = .10) and ATM_BASE (ß=.0008, t=1.95, p =

.05) have negative impacts on the likelihood of network merger or default (increasing survivability)

supports the idea that the physical network installed base in ATMs and the complementary installed base

of ATM cards both give rise to externality value perceived by adopting firms. It also suggests that

tangible value has been created for network providers to margin in their business. As we pointed out

earlier in the paper, another way to think about coverage – apart from the number of ATMs on the

network and the number of its cardholders – is to consider geographical coverage. We found no

significant effects for NET_COVERAGE , a binary variable that we used to indicate whether a network

operated in multiple states. This was true in the Physical Infrastructure Installed Base model .(ß = .32, t

= 1.04, not significant at .10) and in the Complementary Asset Installed Base model (ß = .42, t = 1.38, not

significant at .10). Prior studies of electronic banking network-related issues considered single and multi-

state operations, in part because such operations were either mandated or prohibited (especially multi-

state operations) for competitive reasons by the Federal Reserve Bank (Felgran, 1985)

We also found that a higher GROWTH_RATE reduces the likelihood of network default (ß = .75, t =

1.72, p = .10 in the Physical Infrastructure Installed Base model; ß = 76, t = 1.76, p = .10 in the

Complementary Asset Installed Base model). Together with the finding that larger networks are less

likely to default, this result suggests that adopting firms and consumers attribute value to a growing

installed base. The theoretical literature we discussed earlier points to the importance of expectations that

network participants develop about equilibrium outcomes in the market, especially who will become

dominant and who will survive. Once growth becomes more constrained, market expectations are likely

to shift, forcing the network operator to look for other means to grow network value. Thus, our results

point to the conclusion that the success of a network is likely to be determined by the extent of the

network externalities it can throw off, and that the market for regionally shared electronic banking

network services clearly values network externalities.

Two additional significant results suggest the importance of network efficiency. The consistent

impact of %INTERCHANGE in our models (ß = 2.40, t = 3.82, p = .01 in the Physical Infrastructure

Installed Base model; and ß = 2.25, t = 3.72, p = .01 in the Complementary Asset Installed Base model)

shows the extent to which transaction switching services between member banks are valued, and make the

network more survivable as a result. By achieving high percentages of interchange transactions the

extensive coverage of a network’s infrastructure is brought into play, creating the opportunity for the

network to provide additional spatial externality value that may be somewhat different than the numbers

of ATMs or cardholder along might suggest. With higher interchange percentages come larger revenues

from servicing banking firms’ customers within and outside of the network. In addition, a higher volume

of TRANSACTIONS_PER_ ATM (ß = .00013, t = 2.69, p = .01 in the Physical Infrastructure Installed

Base model; and ß = .00017, t = 3.32, p = .01 in the Complementary Asset Installed Base model),

implying greater operational efficiency through more effective scale size, also increases the survivability

of a network.

Another highly significant effect in both models was attributable to the MAKE_EFFECT (ß = .14, t =

3.46, p = .01; and ß = .15, t = 3.29, p = .01). The positive coefficient tells us that networks that are new

to the market, and thereby have less brand recognition, a smaller following, and no history of reliable

service, are more likely to default. The interpretation of this result is clear: networks which positioned

themselves as relatively early entrants in the market have worked the kinks out of their operating

regimens, while the new entrants face a whole range of uncertainties, including the challenge of creating

an installed base of cardholders with other competitors already active in the market, the inevitable

operational problems that will need to be resolved to fine-tune service delivery, and actions that might be

taken as competitors respond to the threat of the new entrant. Of course, this finding does not rule out the

possibility that new networks may succeed in the market; but it serves to condition a potential member’s

expectations of what is likely to happen in the future. Several networks that became highly successful,

including Mellon Bank’s CashStream, Most, Money Station and Honor, only began operations in 1983,

and grew to occupy a position among the top 50 networks nationwide by the following year (The EFT

Data Book, Bank Network News, 1984), primarily led by the technological sponsorship of large regional

banks.

We included a variable for OWNERSHIP, which the reader should recall is a binary variable

indicating the presence of joint ownership of a network by its members. However, we failed to obtain a

significant effect (ß= .20, t = .66, insignificant in the Physical Infrastructure Installed Base model; and ß

= .43, t = 1.33, insignificant in the Complementary Asset Installed Based model). Perhaps if this variable

had been coded to identify the presence of financially strong versus financially weak banking firm

participants in the network ownership structure, we might have been able to develop additional

information on the OWNERSHIP effect. Unfortunately, such data on the owning banks were not

available to us; in addition, it would have required us to identify a link between some performance

dimension (e.g., total performing assets, return on capital, prior year profits, etc.) and new network

success, which our theory does not really support.

One additional important observation that we can make, based on our estimates for α in each of the

models is that the estimated value of α is greater than 0 (α = .89, t = 6.81, p = .01 in the Physical

Infrastructure Installed Base model; and α = .98, t = 6.42, p = 6.42 in the Complementary Asset Installed

Base model). With these estimates for α, we can make the conclusion that, after taking into account all of

the variables that were included in the estimation of each model, the hazard rate is increasing over time.

This observation is consistent with the Weibull distribution, which assumes the base hazard rate

(likelihood of network default) for all observations is increasing over time. This made our choice of the

Weibull distribution especially appropriate for the setting we model here. (The reader should recall from

earlier in the paper that that the hazard rate is given by hi(t) = γαtα-1 with γ = exp(-ßXit), α > 0, and ∂

hi(t)/∂α > 0.) A second related observation is that the value of α for the Complementary Asset Installed

Base model is close to 1 (α = .98), and given the size of our data set, would not be statistically different

from it. Thus, for that model at least, it appears that the base hazard rate is constant, and thus only the

explanatory variables will influence the hazard rate. (For additional consideration of the robustness of

these results, the interested reader should see the Appendix on Assessing Model Robustness.)

Overall, our findings match what is known about the time period under study. Electronic banking

technologies were diffusing during this time, becoming more acceptable to the public, and thus expanding

the impetus for electronic monetary exchange in the American economy. But simultaneously the industry

infrastructure was being rationalized, as a greater number of competitors looked for ways to achieve

profitable operations. With α > 0, the term γ = exp(-ßXit) in hi(t) = γαtα-1 (especially the values of Xit)

will determine the hazard rate for individual firms. Thus, it is possible (and consistent with the overall

observation of an increasing hazard rate for the entire data set) that the hazard rate for an individual firm

may not be increasing. In the long run, one would expect the hazard rate not to keep increasing.5 As the

industry infrastructure for shared electronic banking networks gets sorted out in the terms that we have

discussed in this paper, very strong players will remain (e.g., Star, Pulse, NYCE, etc., each currently

achieving viable critical mass), which define an infrastructure in equilibrium.

4.2. Interpretation

Overall, our findings provide some evidence about the role of network externalities in the regional

shared electronic banking industry of the 1980s. We summarize them in Table 4 below.

--------------------------------------------

INSERT TABLE 4 ABOUT HERE

--------------------------------------------

The first four hypotheses were motivated by network externality theory as we applied it to the

electronic banking services context. Only one of the four hypotheses (H3 – Geographic Coverage) was

not supported. The second four hypotheses were controls that were motivated by prior research on retail

electronic banking networks and banking firm performance. Three of four hypotheses were strongly

supported, while the last one (H8 – Governance by Sole Ownership) was not; we speculate whether

ownership structure is as important as the financial health of the owners of a network, however, we are

unable to acquire the necessary data to test this assertion. Overall, our results point out that the existence

and growth of these network externalities, which we measured in two ways – as the physical

infrastructure network installed base of ATMs and as the complementary asset installed base of

cardholders -- played an important role in the survivability of the network service providers. Moreover,

we learned that operational efficiencies, the extent of a network provider’s transaction switching

activities, and the make effect all determined whether a network succeeded or failed in the marketplace.

Since 1983, the retail electronic banking industry in the U.S. has experienced many consolidations

among shared regional networks that were undertaken to cut processing costs as one of the only ways to

preserve the business viability of the network (Bank Network News, 1990). Consider the actions of

several representative large shared network firms. The Pennsylvania -based Money Access Center’s

(MAC) acquisition of Buffalo, New York-based Metroteller in 1989, and the New York City-based New

York Cash Exchange’s (NYCE) subsequent acquisition of Buffalo-based Instabank are cases in point.

5 There is a contrast here between what we would observe in a competitive marketplace, and what we would observe for light bulbs in a house, or machines on a factory floor. For the latter, we would expect the hazard rate to reflect the fact that each would fail at some point, and after some point in time, the likelihood of failure in the next period would become very high.

Table 4. Summary of Main Results

HYPOTHESES FINDING COMMENTS

H1: Larger networks – in terms of the number of ATMs deployed – tend to survive.

Supported (.05 level)

An indicator that network externalities may play a role in the survivability of a shared electronic banking network.

H2: Networks with greater complementary assets – in terms of the number of bankcard holders – tend to survive.

Weak support (.10 level)

Similar to the implications of H1, however, the network installed base concept originates on the demand, instead of on the supply side of the marketplace, and points more to the role of complementary assets to the network, than to the size of its physical infrastructure.

H3: Networks that offer broader geographical coverage tend to survive.

Not supported in either model

We speculated that networks operating in multiple states would be more survivable on an independent basis. Perhaps the demographic qualities of a network’s marketplace may matter more than its geographical coverage.

H4: The more rapid the growth rate of a network, the more likely it will be to survive.

Weak support in both models (.10 level)

Provides additional support for the role of network externality theory in explaining network survival and default. Emphasizes importance of firm expectations of equilibrium outcomes in the electronic banking market.

H5: The stronger the network’s reputation, the more likely it will be to survive.

Strong support in both models (.01 level)

Similar to other corporate contexts, the make effect plays an important role in firm and consumer expectations about the likely performance of electronic banking firms.

H6: The larger the number of transactions per network ATM, the more likely the network will be to survive.

Strong support in both models (.01 level)

Making efforts to achieve operational efficiency offers a strong basis for ensuring the survival of an electronic banking firm. Many firms can only achieve such efficiency by merging with others in a similar situation, however.

H7: Networks with a higher percentage of switched transactions are more likely to survive.

Strong support in both models (.01 level)

Beginning in the early 1980s, switching interchange transactions provided a solid base of revenues for network service providers. This activity has grown in importance.

H8: Networks with sole ownership are less likely to survive.

Not supported in either model

The financial strength of the network ownership may be more important than its relative concentration.

With other larger networks operating in the Middle Atlantic region, the growth of Metroteller’s and

Instabank’s scale sizes, efficiency levels and network externalities were seriously constrained. Moreover,

extending MAC to upstate New York positioned its owner, Core States Financial of Philadelphia (now

acquired by First Union National Bank of Charlotte, North Carolina), to grow its own installed base of

bankcard holders, extend its geographic coverage and buffer its franchise in core markets. By the same

token, once MAC had taken a position in New York State, it made sense for NYCE to respond. NYCE

had earlier been built by large money center banks in an effort to compete with Citibank’s highly

successful proprietary network.

With limits to growth clearly understood by most electronic banking executives in New York City,

the focus for growth had to shift to expanding regional operations, even if the initial mergers meant

becoming involved in somewhat less attractive markets. Such moves nevertheless guaranteed that

interchange and switching revenues would grow, while transaction costs at the switch would fall. Today,

industry observers recognize these maneuvers as scale size plays that attempted to minimize direct

conflict between growing networks in their primary markets. As with many industries, growing the

business, increasing its operational efficiencies, and partnering with other competitors where necessary

offered considerable protection against industry giants who targeted networks with low operating

efficiencies for acquisition (The EFT Data Book, Bank Network News, 1990).

As a result, one expects that value-maximizing choices made in the past by large banking firm

electronic banking network owners, especially where they have some link to the growth of the network’s

installed base, will influence the success or default likelihood of an electronic banking network. The

actions of smaller network providers are also worthwhile to examine. Smaller networks often were not

able to bring as much value to their members and, therefore, were more prone to default, or to become the

weak partner in a network merger. Beginning with the early merger of the Network Exchange into Most,

a network that would later emerge as a dominant player in the Southeast and the southernmost of the

Middle Atlantic States, we have seen many such examples of smaller shared networks exiting the market

since 1983.6

Other more extensive mergers led to the creation of the regional electronic banking system

infrastructure whose elements remain with us today, even if they are not immediately recognizable. They

appear to have occurred in contexts where the market equilibrium could not support very many

competitors, or where the actions of one or two banking firms may have forced the hands of other major

competitors. For example, in 1984, the amalgam of four regional networks operating in Ohio led to the

formation of the Money Station, to become the sixth largest network in 1984. Apparently, no individual

network believed it could achieve a sustainable competitive advantage over its rivals, in the absence of

greater scale size, lower switch fees, and more comprehensive coverage for banking customers in the

6 One contrasting example of a network that has newly entered the shared electronic banking network industry is worthy of comment here, because it shows that electronic banking service markets, in the presence of surcharges, user fees that apply when banking customers use other banks’ cash machines, are contestible to some extent, even though it may be difficult for other new entrants to profit. The Congressional Budget Office (1998) study on ATM network competition comments: “Although … network effects … make it difficult to create new networks, the desire to dispense with surcharges --and in doing so attract more customers -- has spurred several such efforts as well as attempts to establish surcharge-free areas within existing networks. Small and community banks have begun to form surcharge-free alliances in several states, and a new small-bank network, Cartel, is operating in 48 states. In the Cartel network, most ATMs do not impose surcharges; however, the network is quite young and whether it can compete in the long run is open to question.” (Excerpted from Chapter 4.)

region – all of which were blocked by the presence of proprietary associated with specific banking firms.

Not long after, MAC became the seventh-ranked shared network in the U.S. when the Philadelphia

National Bank went on-line to Pittsburgh’s TriNet organization. TriNet involved three Pittsburgh

banking firms, including the regionally significant Pittsburgh National Bank, and operated as a franchisee

of MAC. It also competed with Mellon Bank’s CashStream network, which had developed significant

installed base in Western Pennsylvania. This outsourcing arrangement was preferred by TriNet’s

participants over creating their own transaction switching infrastructure, which the regional electronic

banking business would not support. Around the same time, the former George network of Philadelphia’s

Girard Bank was bought out by CashStream, the Mellon Bank-owned network that became the ninth-

ranked shared network (The EFT Data Book, Bank Network News, 1985).

Although we recognize that it is inappropriate here to over-interpret our econometrics results as they

relate to specific contexts, we should point out that the many of the reasons for these network

consolidations that were reported in the press at the time are supportive of them. In several regions of the

U.S., a move by just a couple significant banking firms toward consolidation of their proprietary networks

into larger shared networks forced other non-participating competitors to respond. The failure to do so,

thereby causing the network provider to forego increases in installed ATMs and the cardholder base, the

creation of network transaction interchange services, and growth in scale size leading to reduced per

transaction costs at the ATM, would render the firm’s electronic banking services uncompetitive. These

trends with respect to the network effects of electronic banking networks appear to remain in place, even

today, and they are far more widely recognized by their owner-operators, and the government groups that

regulate their activities (Congressional Budgeting Office, 1998; Hayes and Meltzer, 1997; New York

Cash Exchange, 1999).

5. CONCLUSIONS

This research provides an analysis of network default and network survivability in a financial services

industry setting that involves network externalities. We also view this research as another contribution to

the growing body of empirical work (e.g., Brynjolfsson and Kemerer, 1996; Kauffman, McAndrews and

Wang, forthcoming) that provides an empirical test of the theory of network externalities (Shapiro and

Varian, 1998). Using a duration modeling approach and network growth data that describe the

consolidation of the market for shared electronic banking network services, we examined the

determinants of network default. Our results provide the basis for understanding how to make an

electronic banking network survivable in a highly competitive marketplace, in terms of a set of externality

theory-motivated drivers, and other well-known controls that relate to industry competition. We found

evidence that electronic banking networks deliver network externalities, creating value for network

providers and adopting firms, and, by implication, for the consumers who use the network services.

More importantly, however, we learned that those networks that appear to throw off the greatest

externality value – whether it is due to the physical infrastructure and installed base, or to the installed

base of complementary assets that relate to the network – are the ones that appear to have had greater

opportunities for survival in a consolidating industry structure.

Our findings are broadly consistent with critical mass theory, which suggests that networks new to the

market often experience start-up pains, and that a network must obtain an installed base at least equal to

the size of the critical mass that consumers recognize is necessary for a firm to be a viable service

provider. Our results are also supportive of the installed base model of Farrell and Saloner (1986), which

suggests that networks have a tendency towards greater concentration and that the strength of the network

externalities that accrue from an existing installed base may lead to a bandwagon effect. Although the

theoretical literature on networks suggests that network externalities may result in choices of inferior

technologies, the level of analysis undertaken in our research does not allow us to discriminate among the

different de facto standards that were involved – inferior or otherwise – as banking firms joined up with

one network provider or another. Further, Liebowitz and Margolis (1994) argued that there has been very

little detailed empirical support and almost no compelling examples of markets failing in the sense that

the “wrong” choice of network was made. We see no evidence of that in the electronic banking context,

other than the failures of the market to move to universal switching prior to the early 1990s.

Our findings provide useful managerial insights for effective valuation, adoption and diffusion

strategies. To succeed in the marketplace, network providers can deliver more value to adopting firms

and their downstream consumers by operating a large and growing network that is economically efficient

and offers indispensable switching services. In addition, network providers may wish to influence

adopting firms’ perception regarding future network size and network viability in their favor. This is

particularly important for new entrants faced with starting up. Similarly, firms that contemplate adopting

a network should evaluate the extent of network externalities the network can offer, and recognize that

they may be constrained by other networks’ infrastructures that are already present in the market. This

“limits to value” perspective suggests that adopting firms should consider the adoption behavior of other

firms and anticipate which networks are more likely to become dominant, and in this way gain an

understanding of exactly what potential for value exists (Davern and Kauffman, forthcoming). To do this,

we argue from what was learned in this research, it is necessary to take into account the existing installed

base, growth path, long term viability, operational efficiency, and strategic position of networks that are

under consideration for adoption.

The results of this study also suggest managerial implications for other telecommunications network

and Internet contexts. These include wireless and cellular communications providers, Internet service and

search engine providers (e.g., Inktomi and Infoseek), and World Wide Web portal service providers. In

addition, there are many non-networked information technology products and applications that have the

characteristics of networks (e.g., the Object Management Group’s CORBA (object request broker)

middleware solutions, Microsoft’s Component Object Model (COM), Rational Software Corporation’s

Unified Modeling Language (UML), the Extensilble Markup Language (XML) and others). The latter

includes technologies with evolving standards, changing installed based of users, and heterogeneous

prospects for success in the marketplace. Although we recognize the limitations inherent in selecting one

technology and one industry setting, this research nonetheless provides a general framework for analysis

that can be applied more broadly: to a spectrum of information technologies and competitive

interorganizational information systems that offer network externalities, especially those which we are

now considering in the area of electronic commerce.

In future research, we can address some of the shortcomings of the industry model we specified for

electronic banking network competition by considering other explanatory variables for network success.

Network pricing, for example, appears to be of increasing general interest in the network service market --

and in the context of electronic banking networks especially since late 1995, when the U.S. Congress

began to consider ATM service pricing (Congressional Budgeting Office, 1998; Hayes and Meltzer,

1997). However, some observers argue that the price of network services is expected to play only a

marginal role in influencing demand (Antonelli, 1989). With network pricing data available, one might

consider perform an empirical test of the proposition that the equilibrium outcome in the market for

electronic banking services entails artificially lowered network prices initially, followed by above

marginal cost pricing later. This is akin to what Gowrisankaran and Stavins (1999) have shown for the

case of Federal Reserve Bank ACH services.

More interesting to us, however, is to pursue future research that applies the general framework we

have developed to study the growth and evolution of other IT applications in electronic commerce that

offer externalities. For example, our modeling approach is readily applicable to the on-line information

services industry where America Online, CompuServe, Microsoft Network (MSN), and Prodigy, among

other networks, once competed head-to-head with each other. The competition in this section has played

out rapidly, and hardly a quarter goes by now without some major announcement that changes the

structure of that industry. In financial services on the Internet, similar opportunities for in-depth

empirical research and analysis also exist. For example, in just the last couple years, electronic bill

presentment and payment service networks for the Internet have begun to take shape, and we expect to

see head-to-head- competition between the alternative potential standards for vendor, consumer and bank

payment clearing and settlement connectivity (O’Sullivan, 1999). Other purely Internet technology and

electronics-focused competition is interesting to us as well. For example, CIO Magazine’s WebBusiness

Section recently reported that “[i]n the Internet audio realm, a VHS-versus-Betamax-like war is currently

being waged between Microsoft’s Windows Media Player and RealNetwork’s RealPlayer G2” (Edwards,

1999, p. 94).

Indeed, there are many such heated battles involving one kind of standard or another on the Internet,

each of which offers it own promise for a technological standard in the future. But some just don’t make

it, and limits to the value of the network externalities that they can create are often at the heart of the

market’s sea change. For example, VRML (virtual reality modeling language), once envisioned as a

potential three-dimensional graphics formatting standard, has now been surpassed in interest by other

developments in the marketplace, even though its core capabilities are now incorporated into some of the

newly emerging technologies. Technologies, like the electronic banking networks we studied in this

paper, are also subject to network default.

APPENDIX. ASSESSING MODEL ROBUSTNESS

With these general results in hand, the next step is to probe how robust they are. To accomplish this, we employ two separate approaches. First, we employ a test for neglected heteroscedasticity in the data set, which might be associated with an omitted variables bias. Second, we perform out-of-sample look-ahead forecasts, to gauge how our models predict outside the competitive regime we estimated them in.

A Test for Neglected Heterogeneity. Another way that we can obtain a reading on the usefulness of our models is to examine the extent to which individual observations in the data set exhibit heterogeneity . This could occur, for example, if the banking firms that own or operate a given electronic banking network exhibit different levels of financial well-being. The implication is that the health of the owner(s) may be a predictor of the health of the network.

We can perform a test for neglected heterogeneity within our sample by employing a score statistic that is attributed to Behrman, Sickles and Taubman (1990). The authors’ technique involves the hazard function, h(t), and is assumed to include such heterogeneity for whatever reasons. The hazard function, h(t), can be shown to be approximately equal to a function of a second related hazard function, h*(t), in which there is no neglected heterogeneity, as follows:

h(t) ≅ h*(t) Z ≅ h*(t) {1 + σ2 [ε[(t)2 - 2ε(t)] / 2}

We see in this expression that h(t) and h*(t) differ by the multiplier, Z, which captures the variability in the differences among individual observations.

We next take logs of both sides, and let the parameter vector characterizing the hazard be (β, σ2), with β^ the maximum likelihood estimate of β when σ2 = 0. Differentiating the resulting expression at the

point (β^,0) yields just one non-zero term: ∂∂σ

ε εβ

ln ( ) | [ ^ ( ) ^ ( )] /( , )

h t t t2 0

2 2 2= − . In the absence of neglected

heterogeneity, E[ε^(t)2 - 2ε^(t)] / 2] = 0. Under the null hypothesis, σ2 = 0, and with N observations in

the sample, the Behrman, Sickles and Taubman score test is based on: Q Nii

N

ii

N* [ ^ ^ ] /= −= =∑ ∑ε ε2

1 1 .

The asymptotic variance of [ε^(t)2 - 2ε^(t)] / 2] can be determined from the information matrix associated with h*(t) {1 + σ2 [ε[(t)2 - 2ε(t)] / 2 } with the parameters (β, σ2). Reordering the terms (σ2,

β) for σ2 = 0, the MLE information matrix for the null hypothesis is given by: Ii

T

( , )β 02 11

= The test

statistic, W, approximates the standard normal distribution, and is given by:

W = ([N/4])1/2 [ (Σe i2 - 2Σe i) / N ] / (2-l'

i-1l )1/2 .

Preliminary results that we obtained with the computed value of the W statistic allow us to decline the null hypothesis and reject the alternative of neglected heterogeneity at the 5% level of significance.

Out-of-Sample Forecasts. In addition to the data we used to estimate our models, we also collected additional years of data from the same sources to perform out-of-sample forecasts. Performing one and two-year ahead out-of-sample forecasts, for example, would enable us to examine the extent to which our model is able to predict electronic banking network defaults on the basis of the variables that were used for prior periods. And, assuming that one believes that future years are likely to exhibit the same main effects leading to network default, we can assess the quality of the predictions about network survival and network default, as an indicator of the robustness of the models to upcoming changes in the market. The work on these results is currently underway. REFERENCES Antonelli, C. The Diffusion of Information Technology and the Demand for Telecommunication Services. Telecommunications Policy 13, 3 (September 1989), 255-264. Bakos, Y. J. A Strategic Analysis of Electronic Marketplaces. MIS Quarterly 15, 3 (September 1991), 295-310. Bakos, Y. J., and Nault, B. R. Ownership and Investment in Electronic Networks, Information Systems Research, 8, 4 (December 1997). Bank Network News, various issues (see in.faulknergray.com/pubs.htm). Behrman, J. R., Sickles, R. C., and Taubman, P. Age-Specific Death Rates with Tobacco Smoking and Occupational Activity: Sensitivity to Sample Length, Functional Form, and Unobserved Frailty. Demography, 27, 2 (May 1990), 267-284.

Benaroch, M., and Kauffman, R. J. A Case for Using Option Pricing Analysis to Evaluate Information Technology Project Investments. Information Systems Research, 10, 2 (Spring 1999).

Benaroch, M., and Kauffman, R. J. Justifying Electronic Banking Network Expansion Using Real Options Analysis. MIS Quarterly , forthcoming. Brynjolfsson, E., and Kemerer, C. F. Network Externalities in Microcomputer Software: An Econometric Analysis of the Spreadsheet Market. Management Science, 42, 12 (December 1996), 1627-1647. Chismar, W. G., and Meier, J. A Model of Competing Interorganizational Systems and Its Application to Airline Reservation Systems. Decision Support Systems 8 (September 1992), 447-458. Clemons, E. K., and Kleindorfer, P. R. An Economic Analysis of Interorganizational Information Technology. Decision Support Systems 8, 5 (September 1992), 431-446.

Congressional Budgeting Office. Competition in ATM Markets: Are ATMs Money Machines? U.S. Government Printing Office, Washington, DC, July 1998 (available on the Internet at www.cbo.gov/showdoc.cfm?index=666&sequence=0&from=0#anchor). Davern, M. A., and Kauffman, R. J. "The Value of Decision Technologies: Discovering Potential and Realizing Value," Journal of Management Information Systems , forthcoming. DeMarco, R. C. The Biggest Stories of 1995. Computer Reseller News (December 11, 1995). Dybvig, P. H., and Spatt, C. Adoption Externalities as Public Goods. Journal of Public Economics 20, 2 (1983), 231-247. Economides, N. The Economics of Networks. International Journal of Industrial Organization (December 1996). Edwards, J. Standards Carpet the Internet: But Are They Creating Their Own Form of Confusion? CCIO Magazine, WebBusiness Section, December 1, 1999. The EFT Data Book . An annual publication of Bank Network News, various issues (see also in.faulknergray.com/pubs.htm). Farrell, J., and Saloner, G. Standardization, Compatibility, and Innovation. Rand Journal of Economics 16, 1 (Spring 1985), 70-83. Farrell, J., and Saloner, G. Converters, Compatibility, and the Control of Interfaces. Journal of Industrial Economics 40, 1 (March 1992), 9-35. Felgran, S. D. From ATM to POS Networks: Branching, Access, and Pricing. New England Economic Review (May-June 1985), 44-61. Gandal, N. Hedonic Price Indexes for Spreadsheets and an Empirical Test for Network Externalities. RAND Journal of Economics, 25, 1 (Spring 1994), 160-170. Gowrisankaran, G., and Stavins, J. Network Externalities and Technology Adoption: Lessons from Electronic Payments. Working paper, Department of Economics, University of Minnesota, August 1999.

Greene, W. R. Documentation for LIMDEP 7.0. Econometric Software Inc., Bayport, NY, 1999 (available at www.limdep.com).

Hannan, T., and McDowell, J. The Determinants of Technology Adoption: The Case of the Banking Firm. RAND Journal of Economics 15, 3 (Autumn 1984), 328-335.

Hayes, T., and Meltzer, J. The Future Use of ATMs: An Examination of the Current and Future State of the ATM Business. Dove Associates, Boston, MA, June 1997. Hosmer, D. W., Jr., and S. Lemeshow, Applied Survival Analysis: Regression Modeling of Time to Event Data. John Wiley and Sons, New York, NY, 1999. Kalbfleisch, J., and Prentice, R. The Statistical Analysis of Failure Time Data. John Wiley & Sons, New York, N.Y., 1990.

Katz, M., and Shapiro, C. Network Externalities, Competition, and Compatibility. American Economic Review 75, 3 (June 1985), 424-440. Katz, M., and Shapiro, C. Product Introduction with Network Externalities. Journal of Industrial Economics 40, 1 (March 1992), 55-83. Josefek, R. A., and Kauffman, R. J. Duration of IT Human Capital Employment. Presented at the 1998 Workshop on Information Systems and Economics (WISE), Stern School of Business, New York University, New York, NY, December 1998.

Kauffman, R. J., McAndrews, J. J., and Wang, Y. Opening the ‘Black Box’ of Network Externalities in Network Adoption, Information Systems Research, forthcoming. Kauffman, R. J., and Wang, Y. The Network Externality Hypothesis and Competitive Network Growth. Journal of Organizational Computing and Electronic Commerce, forthcoming. Kauffman, R. J., and Wang, Y. An Econometric Analysis of Network Diffusion in the Presence of Spatially Varying Externalities. Working paper, MIS Research Center, Carlson School of Management, University of Minnesota, 1999. Keenan, C. "Small Banks Join No-Fee Alliances in Bid to Retain ATM Customers," American Banker, January 8, 1998. Kiefer, N. M. Economic Duration Data and Hazard Functions. Journal of Economic Literature 26 (June 1988), 646-79. Levin, S., Levin, S., and Meisel, J. A Dynamic Analysis of the Adoption of a New Technology: The Case of Optical Scanners. The Review of Economics and Statistics 69, 1 (February 1987), 12-17. Liebowitz, S. J., and Margolis, S. E. Network Externality: An Uncommon Tragedy. Journal of Economic Perspectives, 8, 2 (Spring 1994), 133-150. McAndrews, J. J. The Evolution of Shared ATM Networks. Business Review, Federal Reserve Bank of Philadelphia (May-June, 1991), 3-16. McAndrews, J. J. Network Business Value Externalities. In Perspectives on the Strategic and Economic Value of Information Technology Investments, R. D. Banker and R. J. Kauffman (eds.), Middletown, PA: Idea Group Publishers, 1992. Nault, B. R., and Dexter, A. S. Adoption, Transfers, and Incentives in a Franchise Network with Positive Externalities. Marketing Science 13, 4 (Fall 1994), 412-423. New York Cash Exchange (NYCE). NYCE News (a quarterly newsletter for network members), 14, 1 (April 1999) (available on the Internet at www.nycenet.com/nl/nyce4_99.pdf). O'Keefe, M. Survival of the Fittest. Bank Systems & Technology 31, 10 (October 1994), 64-70. Oren, S., and Smith, S. Critical Mass and Tariff Structure in Electronic Communications Markets. Bell Journal of Economics 12, 2 (Autumn 1981), 467-486.

Peterson, P. The Statistical Analysis of Event Histories. Sociological Methods & Research 19, 3 (February 1991), 270-323. Riggins, F. J., Kriebel, C. H., and Mukhopadhyay, T. The Growth of Interorganizational Systems in the Presence of Network Externalities. Management Science 40, 8 (August 1994), 984-998. Rohlfs, J. A Theory of Interdependent Demand for a Communication Service. Bell Journal of Economics and Management Science 5, 1 (Spring 1974), 16-37. Rose, N., and Joskow, P. The Diffusion of New Technologies: Evidence from the Electric Utility Industry. Rand Journal of Economics 21, 3 (Autumn 1990), 354-374. Saloner, G. and A. Shepard, "Adoption of Technologies with Network Effects: An Empirical Examination of the Adoption of Automated Teller Machines," RAND Journal of Economics, 26, 3 (Autumn 1995), 479-501. Tirole, J, The Theory of Industrial Organization. MIT Press, Cambridge, MA 1988. O’Sullivan, O. E-Billers Jockey for Position, USBanker, October 1999 (excerpts available on the Internet at www.electronicbanker.com/usbanker/current.htm). Wang, E. T. G., and Seidmann, A. Electronic Data Interchange: Competitive Externalities and Strategic Implementation Policies. Management Science 41, 3 (March 1995), 401-418. Xie, J., and Sirbu, M. Price Competition and Compatibility in the Presence of Positive Demand Externalities. Management Science 41, 5 (May 1995), 909-926.