Industrial Marketing Management 39 (2010) 170–177

Contents lists available at ScienceDirect

Industrial Marketing Management

Role of network structure and network effects in diffusion of innovations

Hanool Choi a,1, Sang-Hoon Kim b,⁎, Jeho Lee c,2
a
Keimyung University, Department of International Commerce, 2800 Dalgubeoldaero, Dalseo-gu, Daegu 704-701, Republic of Korea
b
Seoul National University, Graduate School of Business, 599 Gwanangno, Gwanak-gu, Seoul 151-916, Republic of Korea
c
Korea Advanced Institute of Science and Technology, KAIST Business School, 207-43 Cheongryangri 2-dong, Dongdaemun-gu, Seoul 130-722, Republic of Korea

a r t i c l e i n f o a b s t r a c t

Article history: Why does diffusion of innovation sometimes propagate throughout the whole population, and why at other
Received 4 October 2005 times does it halt in its interim process? The current paper provides a potential answer to this question by
Received in revised form 30 March 2008 developing a simple computational model of social networks. The proposed computational approach
Accepted 31 August 2008
incorporating small-world graphs enables the authors to find that diffusion of innovation is more likely to fail
Available online 19 December 2008
in a random network than in a highly clustered network of consumers. A marketing implication is that the
choice of initial target groups and their network structures matter in influencing whether an innovation
Keywords:
Diffusion of innovation
makes full or partial penetration, in markets where network effects plays a role.
New product © 2008 Elsevier Inc. All rights reserved.
Network effects
Small world
Network structure

1. Introduction To address this question, we develop a simple numerical model of

adoption dynamics with two key components: network effects and
Innovations happen everywhere. However, some innovative network structure. First, we consider a market with network effects,
products take off instantly, and others take a long time to penetrate where the benefits of adopting the innovation grow as the number of
the market—for example, facsimile (fax) machines needed a century adopters increases (e.g., Katz & Shapiro, 1985). Adoption dynamics of
to accomplish substantial diffusion. In addition, there are many new such network products or services are quite distinct from those of
products that succeed in the early market but ultimately fail to diffuse conventional ones. Network products and services are quite difficult to
throughout the whole customer base.3 The cases in point are get started and often end up being under-adopted (Rohlfs, 1978, 2001).
Momenta Corporation and Palm Computing. Both companies sought For example, consider an adopter of a communication service (e.g. email
to exploit untapped opportunities in the emerging pen-based or instant messaging). The first adopter sees no benefit in adopting this
computing market in the early 1990s. From the outset, Momenta service because there is no one to communicate with. Customer benefits
tried to target the mass market with glitzy advertising (cf. Harvard will be realized only when other people begin to adopt an interoperable
Business School Case 9-392-013), whereas Palm initially focused on a service. Because of such a lack of customer benefits at the early stage,
specific target group. Palm successfully penetrated into the main- under-adoption is natural for network products and services.
stream market, but Momenta didn't. The second essential component of our model is the direct
Why does diffusion of innovation sometimes propagate through- specification of the structure of consumer networks. As cited in
out the whole population, and why at other times does it stop in its Rogers (1995, p.25), Katz (1961) remarked: “[i]t is as unthinkable to
interim process? Although practitioners can provide a wide array of study diffusion without some knowledge of the social structures in
anecdotes, systematic research on this issue has been rather sparse. which potential adopters are located as it is to study blood circulation
The purpose of the present paper is to provide a potential answer to without adequate knowledge of the veins and arteries.” But, much of
the question, focusing on the role of social networks, an under- prior work on network effects (with only a few exceptions, which are
explored variable in marketing. Midgley, Morrison, Roberts, & Midgley, 2000; Abrahamson &
Rosenkopf, 1997) has so far largely sidestepped the role of social
⁎ Corresponding author. Tel.: +82 10 2070 6934. networks, while focusing on the installed base or a total number of
E-mail addresses: hnchoi@kmu.ac.kr (H. Choi), profkim@snu.ac.kr (S.-H. Kim), adopters. The paucity of research on social networks stems from their
jlee@business.kaist.ac.kr (J. Lee). bewildering complexity, which defies analytical tractability (Strogatz,
1
Tel.: +82 53 580 6958. 2001; Lee, Lee, & Lee, 2006). Researchers have had difficulties even
2
Tel.: +82 2 958 3678.
3
These products are different from the so-called “flops” whose failures are mainly
describing or representing social networks mathematically.
attributed to unrealistic ideas, defects, or mismarketing (Refer to Business Week (8/16/93) Recently, however, advances in complexity theory have offered
article for examples of new product flops.). tools to represent social networks systematically. In particular, Watts

0019-8501/$ – see front matter © 2008 Elsevier Inc. All rights reserved.
doi:10.1016/j.indmarman.2008.08.006
 H. Choi et al. / Industrial Marketing Management 39 (2010) 170–177 171

and Strogatz (1998) developed the small-world graph model, which is Church & Gandal. 1993; Farrell & Saloner, 1985, 1986; Katz & Shapiro,
a formal representation of Granovetter's (1973) conceptualization of 1985, 1986; Schoder, 2000; Shapiro & Varian, 1999). Network effects
the architecture of social networks. In his view, a social network play a key role in the adoption of certain types of products, especially
consists of two essential elements: (1) cliquish sub-networks and interactive communication-type innovations such as telecommunica-
(2) bridges. A cliquish sub-network consists of individuals who are tion services (Mahler and Rogers, 1999).4
interacting extensively with one another. Consider, for example, a Marketing scholars paid attention to this phenomenon from early
friends and family network. Granovetter (1973, 1974) found that this on. For example, as early as 1985, Dhebar and Oren (1985)
sort of sub-network is not very helpful when people look for jobs. The incorporated network effects into their demand function. Xie and
main reason is that information traversing through cliquish sub- Sirbu (1995) extended this work toward duopolistic competition in
networks is more likely to be limited to a few cliques, which tend to the context of dynamic pricing. Since then, other researchers have
share redundant ties. Instead, people get more useful job information examined network effects in relation to various aspects of marketing
from random contacts, or people who are not in extensive relation- practices, such as pricing, promotion policy, distribution, and first-
ships. Such connections are called bridges, which serve to connect mover strategies (e.g., Sun, Xie, & Cao, 2004; Srinivasan, Lilien, &
diverse members from different, or often socially distant, sub- Rangaswamy, 2004; Li, 2005).
networks. In social worlds, people often create bridge-building Recently, some research has begun to examine network effects in
mechanisms, such as conferences, parties, or Internet chat rooms, to the context of new product adoption and diffusion, which is more
facilitate interactions among random contacts or strangers. Onnela related to our work. Based on a survey of consumers' patronage of
et al. (2007) empirically confirmed that the above architecture indeed Microsoft's Windows, Pae and Hyun (2002) showed that network
represents the structure of real-world communication networks for effects play a key role in buyers' adoption and repurchase decisions.
mobile phone users. They also showed that compatibility, upgradability, and preannounce-
Key questions are: should marketers focus on a few cliquish sub- ments are the key sources of network effects.
networks for launching network products? Or should they exploit Lee and O'Connor (2003), on the other hand, set up an empirical
random bridges from the outset? Which network strategy is more model that provides a link between launch strategies and the success
likely to lead to full diffusion? To explore this question, we use Watts of network products. One of the key conclusions of their study is that
and Strogatz's (1998) small-world graph to represent consumer the size and development speed of the installed base plays a critical
networks. The main benefit of using this model is that we can address role in a product's long-term performance, which is consistent with
the above questions by generating diverse kinds of networks that lie the current study's finding that building up network effects at the
between a network of purely cliquish sub-networks and that of purely early stage is the key to the success of innovation diffusion.
random bridges. Specifically, they tested the impacts of various launch strategies
Our numerical analysis shows that the network structure does play such as order of entry, product advantage, penetration pricing,
a moderator role for the link between network effects and innovation bundling, mass targeting, and pre-announcing strategies upon the
diffusion. More specifically, we found that a new product is less likely development of an installed base. Although their model includes
to reach full diffusion in random networks than in cliquish networks. many factors that affect installed base development, they did not
Unlike information diffusion, the spread of network effects reveals take into account the structure of social networks, which we believe
that randomness in connection topology makes it harder for an is another crucial variable for the long-term performance of an
innovation to build up network effects or network benefits at the innovation.
initial stage, and that insufficient network effects in turn result in a From a slightly different angle, Padmanabhan, Rajiv, and Srinivasan
lack of momentum for the diffusion to reach the whole customer base. (1997) studied how network effects influence the success or failure of
However, once the diffusion process reaches a critical mass, abundant a new product. This work deserves some attention since its analytical
bridges in a random network accelerate the process. structure looks similar to that of our model. They compared one-time
A marketing implication is that the network structure of the target release vs. sequential release (or upgrades) strategies in the context of
consumer groups at the early market stage may be a critical factor new product introduction. They proved that a company can use a
affecting whether a new product makes full or partial penetration. sequential release strategy to convey a credible signal of a future
Indeed, Rosen's (2000) argument for Palm's success is consistent with installed base, and that consumer expectations about a future
our numerical result. He argued that Palm was successful because it installed base positively affect their product adoption decisions. In
effectively focused on cliquish social networks in Silicon Valley, their study, Padmanabhan et al. (1997) specified consumer utility as a
whereas Momenta's mass-market approach did not exploit such function of not only the number of adopters, but also of product
favorable properties of social networks. In addition, many shrewd quality, which is quite similar to our model (in Section 3.1.). At first
marketers have been already exploiting these properties. Tupper- they assumed that consumers are homogeneous in their product
ware's party plan, Amway's network marketing, and MCI's calling quality evaluation, and then relaxed the assumption by specifying two
circle are examples of utilizing cliquish sub-networks. levels (novice vs. expert) of marginal willingness to pay for quality. In
The present paper proceeds as follows. Section 2 reviews the our model, we assume a normal distribution for this quantity, which is
related literature, and justifies our research method that utilizes the more consistent with empirical findings (Rogers, 1995).
small-world graph. In the third section, we develop a model and a Despite vast research in this area, most diffusion models as well as
simulation procedure. Section 4 discusses simulation results, and we network effects models, including the Bass model (Mahajan, Muller &
conclude by discussing implications and future directions of the Bass, 1990), have so far described social interactions at the “aggregate”
current study. level by treating the cumulative adoption or installed base as a
surrogate for network effects. Therefore, there is still little known
2. Theoretical background about how the structure of interactions among individuals at the
micro level affects social contagion processes at the macro level.
2.1. Innovation diffusion and network effects

This section reviews prior work on network effects and highlights 4

Although network effects are frequently observed in the information technology
how our work builds on and departs from it. As mentioned before, in a market, they also exist in other industries, such as the railroad industry, the credit card
market with network effects, the benefits of adopting a product or a industry, and so on. So our key findings may not be limited only to high technology
service grow as the number of adopters increases (Arthur, 1989; innovations.
172 H. Choi et al. / Industrial Marketing Management 39 (2010) 170–177

scale-free networks—in classic epidemiology, where the role of

networks had been sidestepped, it had been thought that some
positive value for epidemic threshold would exist. In reexamining the
relationship between network effects and the winner-take-all out-
come, Lee et al. (2006) also saw consumer networks as neglected
moderators that affect the strength of the relation between the two
variables. Lee et al. showed the danger of the popular winner-take-all
hypothesis—this says that network effects always lead the market
dynamics to the winner-take-all outcome. They numerically demon-
strated that the emergence of such an outcome is moderated by the
topology of how consumers interact with one another. We follow this
tradition in analyzing the relationship between network effects and
Fig. 1. Small-world graphs.
the extent of diffusion, hoping that the addition of the under-explored
social networks to a diffusion model will shed new light on our
understanding of innovation diffusion.
In the marketing field, Goldenberg, Libai, and Muller (2001) took
Such structural effects, however, have not been entirely neglected an early step to examine the underlying process of word-of-mouth
in the literature. The study by Midgley et al. (1992) is among the few from the complex systems view. The spirit of our work is not much
that investigated the impact of different network topologies (e.g., different from theirs in that we also want to understand how macro-
completely connected networks, intra-clique networks, disconnected level patterns, such as diffusion of innovation, emerge through
networks, etc.) on adoption dynamics. They developed a simulation interactions among individuals at the micro level. But, the main
model based on the Bass model, showing that network topology distinction lies in the choice of tools to represent customer networks.
affects how fast an innovation spreads throughout the entire system. Goldenberg et al. (2001) used a two-dimensional cellular automaton,
Abrahamson and Rosenkopf (1997) developed another numerical which has been popular for describing spatial phenomena, where
model on the structural effects. Like our work, they challenged the transitivity of distance matters. In the complexity literature, it has
widely received assumption that innovations diffuse fully, or the been noted that this property is frequently violated in social
diffusion process proceeds until every potential adopter adopts them. networks.5 To relax this problem, we use the small-world graph
Then, they numerically demonstrated the possibility of under- developed by Watts and Strogatz (1998).
adoption when adoption dynamics are driven by bandwagon This graph model was a formal representation of Granovetter's
pressures—e.g., increases in the number of adopters of a fad press (1973) original conceptualization of the architecture of complex social
others to adopt it. In particular, they showed that the structure of networks. In his view, there are two essential components for
social networks (the density of network ties, network pressure points, describing social networks. The first component is a clique or a
and the weakness of internal boundaries of a network) affects cliquish sub-network, which consists of individuals who are interact-
whether an innovation diffuses fully or not. ing extensively with one another. Another key component is bridges,
We build on these prior studies. Our work, however, differs from which connect and weave sub-networks, making the whole system
them in that we adopt the complex systems approach or complexity integrated. Recently, Onnela et al. (2007) provided empirical evidence
theory, which has been developed to deal with the structure and for Granovetter's conceptualization of social networks by analyzing
dynamics of complex networks more systematically. Until recently, large-scale mobile phone call records, including about 4.6 million
researchers had difficulty analyzing them because of the lack of mobile phone users.
appropriate tools to represent complex networks like social networks Watts and Strogatz's (1998) small-world graph model uses the
(Strogatz, 2001; Lee et al., 2006). They had little choice but to ignore same two basic building blocks, cliquish sub-networks and bridges.
them or to rely on ad hoc approaches. In the next section, we review An example of the idealization of cliquish sub-networks is shown on
the recent progress in complexity theory and its tools, highlighting the the left in Fig. 1. This graph is a collection of sub-networks that are
benefits of using them to represent social networks. inter-woven on a ring substrate with n nodes and k edges per node,
representing that each of n persons maintains k relationships with
2.2. Complex systems and the small-world graph others. Fig. 1 provides an example when n = 20 and k = 4. Bridges,
then, are added to this default architecture through a random
Through an extensive review of network effects research, Farrell rewiring algorithm with parameter β, which represents a probability
and Klemperer (2001) recently noted that the most interesting aspect of rewiring a link between two nodes in a cliquish sub-network to a
of this stream of research lies in its tie to the burgeoning field of randomly selected node from the entire network. By varying this
complexity theory. However, what is missing in their extensive review single, tunable parameter, a researcher can model numerous
is the discussion of the role of complex networks, which has recently different kinds of networks lying between the two extreme types
attracted a great deal of attention in complexity theory. The dearth of of networks. First, if β = 0, a network consists of a series of cliquish
discussion on networks may be no surprise. Thus far, economists have sub-networks with no random bridges. Second, if β = 1, a network is
largely sidestepped their role primarily because their bewildering completely “random” with no cliquish sub-networks. A key benefit of
complexity defies analytical tractability. using this graph model is that it enables the researcher to observe
Fortunately, the past ten years have seen an explosion of research how macro-level patterns emerge through interactions among
on complex networks, offering promising ways to tame the complex- individuals at the micro level, while not being stuck in the complex
ity inherent in social networks (e.g., Barabási, 2002; Buchanan, 2002; details of reality.6
Redner, 2002). In this literature, complex networks have been seen as
moderators that influence the relationship between known theore-
5
tical variables. For example, Paster-Satorras and Vespignani (2001) For instance, Watts (1999, p.12) noted: “It is possible for person A to be well
considered networks as under-explored media that affect the acquainted with both person B and person C, yet for B and C to be not even remotely
familiar with each other.”
relationship between infection rate and spread of diseases, the one 6
A study by Garber et al. (2004) is the first attempt in marketing to utilize the
that had been widely studied in epidemiology. To the surprise of many small-world graph. They developed a model for predicting success of a product at the
epidemiologists, they showed that no epidemic threshold exists in early stage of the adoption process.
 H. Choi et al. / Industrial Marketing Management 39 (2010) 170–177 173

3. Method

3.1. Model

The present paper develops a computational model to address the

question of whether and how structural characteristics of commu-
nication channels affect the diffusion of innovation. Each individual's
willingness to adopt an innovative product or service is modeled by
three main elements: the product's (or service's) stand-alone benefit,
network effects, and the idiosyncratic reservation utility. Formally,
consumer i's willingness to adopt the product or service at time t is:

Uit = Qi + aNiðt − 1Þ − Ri ð1Þ

Qi represents the product's (or service's) intrinsic value (repre- Fig. 2. Effects of number of initial adopters on adoption dynamics.
sented by its quality or technical performance) perceived by consumer
i, which is not affected by whether other people adopt it or not. It is Period 0: To initiate the adoption process, there should be some
assumed to be constant with time for exposition. Qi is assumed to technology enthusiasts (say, 2.5% of the total population9) who adopt
follow a normal distribution with mean Q and variance s2. In order to the new product without any network benefits. We set this up by tuning
tease out the exact nature of the social contagion, marketing efforts the values of Q, s2, μ, and σ2. Technically, the consumers under the
are not incorporated into the model.7 condition Qi N Ri become initial adopters. If Qi b Ri or Qi =Ri, then the
The second term aNi(t − 1) represents the benefit due to network consumer i will not adopt the innovation in the initial period, and stays
effects. Unlike much of prior work, where a consumer's choice globally “inactive” until his or her first-order acquaintances influence consumer
affects all other consumers' adoption decision-making in the market, i’s the willingness to adopt by increasing the network benefit.
our model assumes network effects in a local sense.8 That is, Ni(t − 1) Period 1: The simulation procedure will calculate the network benefit
represents the proportion of the adopters in consumer i's local aNi0 for all of those who are connected to the initial adopters. Each of
network or “neighbors” (e.g., friends and family network) at time these first-order acquaintances makes the adoption decision. If the sum
t − 1. The scaling constant a represents the relative importance of of the stand-alone benefit Q i and the network benefit aN i(t − 1) is
the network benefit against stand-alone benefits. greater than the reservation utility, she adopts the product or
Lastly, Ri indicates consumer i's inherent reluctance or reservation service. Otherwise, she waits until this condition is satisfied.
about adopting the innovative product or service. This reservation Period t: All of the non-adopters at time t − 1 who are connected to
utility can be regarded as a threshold in the sense that if the benefits any of the previous adopters make adoption decisions. Again, the
from adopting the product (the sum of both the stand-alone benefit same decision rule described above applies to the adoption process at
and the network benefit) exceed the threshold, the customer will buy t. The process will be repeated until no further adoption arises or until
the product or service. In other words, consumer i will adopt the all the consumers have adopted the product or service.
innovation for which Eq. (1) is positive at a given period. We assume
that initially there is some fraction of people who are committed to 4. Results
the new technology, thereby purchasing the innovation even when
there is no network benefit (i.e., N i0 = 0). They are so-called 4.1. Basic properties
“technology enthusiasts” who satisfy the condition: Ri b Qi. As the
product diffuses throughout the market, more people may become To check the realism of the proposed simulation model, we
adopters, influenced by the early adopters connected to them. Note examined the basic properties of the model with respect to three
that R i is time-invariant because the reservation utility level exogenous factors. First, Fig. 2 shows how diffusion patterns change
represents a consumer's innovativeness or inherent disposition to with respect to the number of initial adopters. The larger the number of
the innovation, as was the case in Katz and Shapiro (1985). Ri is initial adopters, the faster the diffusion process. Secondly, analysis on
assumed to follow a normal distribution with mean μ and variance σ2. stand-alone (perceived) benefit Qi shows that, when its mean Q is
A normal distribution has been popularly assumed to represent small, the diffusion process stops before a substantial proportion of
consumer heterogeneity (e.g., Rogers, 1995; Geroski, 2000). the population adopts the product.10 This property is not inconsistent
with the intuition that without sufficient quality or relative advantage,
3.2. Simulation procedure a product cannot penetrate into the mainstream market (e.g., Godin,
2000; Rosen, 2000). Finally, the scaling constant a, which represents
Preparation Step: All parameter values for the simulation (see the relative importance of network effects, also affects the diffusion
Appendix A) are set up at this stage. Each consumer is assigned values process. We find that when a is below a certain level, the diffusion
for her quality perception and reservation utility, each drawn from a process halts halfway. And, as in the case of Q, a larger magnitude of a
normal distribution. Also, a consumer network is constructed accelerates the diffusion process.
according to the algorithm for small-world graphs as in Appendix B. We also examined the relationships between β and other network
As discussed at the end of Section 2, the value of β will determine the measures such as group centrality and density.11 Group centrality
type of network structure or connection topology for our diffusion measures the difference in centrality between the node with the
model.
9
Rogers (1983) partitioned a normal distribution into five groups and labeled them
adopter categories. The first group, called ‘innovators,’ comprises 2.5% of the whole
7
Van den Bulte and Lilien (2001) empirically showed how social contagion could be population. For a deeper understanding of the role of ‘leading edge users’ in early
confounded by marketing efforts. adoption, the authors recommend reading Morrison et al. (2000).
8 10
Network structure might become trivial under global network effects, for which a In the interest of brevity, we do not provide the figures on the influence of stand-
complete network (i.e., everyone is connected to each other) is usually assumed. alone benefit Qi and the scaling constant a. Both of them resemble Fig. 2.
11
Cowan and Miller (1998) and Lee, Lee and Lee (2006) also used the assumption of We thank an anonymous reviewer for bringing our attention to the importance of
local network effects. these relationships.
174 H. Choi et al. / Industrial Marketing Management 39 (2010) 170–177

Fig. 3. The relationship between beta and centrality.

highest centrality and all the other nodes in a network. As shown in

Fig. 3, group centrality is an increasing function of β. All the data
points here are obtained by taking averages of 10,000 simulations. It
shows that as β increases, group centrality gets larger, but the value of
group centrality levels off with a further increase in β. The overall
result implies that as the number of bridges increases (with a higher
value of β), group centrality increases, reflecting the increase in
individual heterogeneity in terms of connectivity. But the variation in
centrality is rather small in the Watts Strogatz model, primarily
because connectivity distributions with large values of β could be
approximated by Poisson distributions (Barabási et al., 1999). This
property is quite in contrast to those in scale-free networks (Barabási
& Albert 1999) and broad-scale networks (Amaral, Scala, Barthelemy,
& Stanley, 2000), where a few nodes, or what are called “hubs,”
dominate relationships in a network, while many nodes are connected
to only a small number of other nodes. In this kind of fat-tailed
network, group centrality will be much larger than that in the Watts Fig. 5. Cumulative distributions of diffusion by network topology.

Strogatz model.
We also checked the relationship between β and density. Our
analysis shows that for all values of β, density remains constant. A “failed” diffusions over 10,000 simulation runs by network structure.
change in the number of bridges does not affect density in the Watts In our study, a failed diffusion is characterized by diffusion stopping
Strogatz model because the rewiring algorithm keeps the same before it reaches 16% of the consumer population, which accounts for
number of links for the whole network, although the connectivity for innovators and early adopters according to Rogers' (1995) adopter
each individual may change. categorization.12 This operationalization is also consistent with
Moore's (1991) notion of “chasm,” which is defined as the gap
4.2. Simulation experiments between the early market (innovators and early adopters) and the
mainstream market. Fig. 4 indicates that given the parameter setting
4.2.1. Failed diffusion (see Appendix A), failed diffusion is more likely to happen when the
Now, let us turn to the key results of our analysis. First, we value of β is high. Note that all the parameter conditions except for
conducted simulation experiments by varying the structure of a network structure are the same for all experiments. This result
consumer network. As discussed before, the increase in the value of β suggests that poor cliquishness with too many bridges hinders an
is positively associated with greater availability of bridges, but with a innovation from diffusing throughout the whole population.
lower degree of cliquishness. Fig. 4 shows the proportion of partial or We can get a more comprehensive view of the diffusion outcome
by plotting the distributions of the number of adopters at the steady
state (out of 10,000 simulation runs) with varying values of β. As
shown in Fig. 5(a), when β is close to 0, the cumulative frequency
looks like an S-curve, implying that the probability distribution for
this case is a bell-curve. A majority of observations fall between 40%
and 60% penetration. The probability of having extreme results—either
100% penetration or failed diffusion—is small. Fig. 5(a) gives a detailed
view on a regime change between β = 0 and β = 0.1. As the value of β
increases, the cumulative frequency curve shifts down to the right.
When β ≥ 0.1, as shown in Fig. 5(b), the cumulative frequencies
consistently show bi-modal patterns. Their striking characteristic is that

12
The critical mass point in the diffusion process is generally expected to occur
between 10 and 20% of potential consumers. According to Rogers’(1995) adopter
categorization, innovators and early adopters account for 2.5% and 13.5%, summing up
Fig. 4. Proportion of failed diffusions by network structure. to 16%.
 H. Choi et al. / Industrial Marketing Management 39 (2010) 170–177 175

most observations fall into either partial adoption or 100% penetration.

Such an extreme bi-modal structure, with few observations in the
middle ground, is frequently detected in a non-linear regime (Lee et al.,
2006). Furthermore, an increase in β augments the frequency of failed
diffusion (below 16% penetration)—or reduces the frequencies of full
penetration. That is, the greater the number of bridges, the more likely
that partial adoption will result.

4.2.2. Diffusion speed

We also examined the relationship between network structure and
the speed of diffusion. Fig. 6 shows the average time (i.e., the length of
simulation period) to reach a steady state. Here we considered only
the cases in which the diffusion reaches beyond 50% of the total
population. The result demonstrates that the average time taken to
reach a steady state is the longest when the value of β is zero. The
speed of diffusion picks up as the value of β increases, rather fast in Fig. 7. Realized network effects over time.
the range between 0 and 0.2, slowly thereafter. This result confirms
the idea that random links facilitate rapid diffusion.

4.2.3. Network topology and propagation of network effects beginning stage typically results in a lack of momentum for further
So far, we have established some regularity such that an under- adoption, leading to failed diffusion. In sum, our analysis suggests that
adoption is more likely to happen in a network with more bridges under-adoption is more likely to happen in networks with random
or random links. We now explore why this regularity may emerge. bridges due to the insufficient buildup of network benefits.
In a market exhibiting network effects, a diffusion failure would
arise when network effects fail to grow beyond some critical point. 5. Discussion
Mahler and Rogers (1999) called it “critical mass effect.” With data
from a sample of 392 German banks, they found that reaching a 5.1. Summary of key findings
minimal number of adopters is critical in the diffusion of
interactive innovations such as telecommunication services. This In this paper, we considered the diffusion of network products and
is more likely to happen when there are relatively few actual services, whose value increases as more and more customers adopt
adopters at the early stage of the market evolution. The question is interoperable products and services. It has been shown that they are
whether network structure influences the early buildup of network prone to under-adoption unless marketers carefully manage their
benefits. diffusion processes (Rohlfs, 1978, 2001; Moore, 1991). The key
Fig. 7 is one way of visualizing the growth of realized network question we explored is whether the structure of a consumer network
effects. The horizontal axis represents time. The vertical axis affects the possibility of under-adoption.
represents a proportion of actual adopters out of the potential adopter To represent consumer networks, we used the small-world
pool. A consumer becomes a “potential” adopter when any of her network model (Watts & Strogatz, 1998), which allows us to tune
acquaintances adopts a new product. When a potential adopter adopts networks between two extreme possibilities. One extreme case is a
the product at time t, she will be eliminated from the potential highly cliquish network, where any pair of well-known customers are
adopter pool at time t + 1. likely to share acquaintances, as easily seen in friends and family
The result in Fig. 7 provides a clue as to why a failed diffusion is less networks, and where network ties are rather redundant and there are
likely to happen in a network with a smaller β. Fig. 7 demonstrates no bridges that connect individuals from remote social distances. The
that the proportion of actual adopters at early periods is relatively other extreme case is a random network, whose approximations can
high when the value of β is close to 0. On the other hand, when the often be seen in bridge-building activities, such as conferences,
value of β is 0.55 or 1, the proportion of actual adopters at early parties, or Internet chat rooms. Our numerical analysis shows that a
periods is very low. Although the figure shows the cases where the new product is less likely to fall into a trap of under-adoption in
adoption explodes at later periods, such an explosion may not happen cliquish networks than in random networks. Our model shows that
in other trials. Indeed, the tiny proportion of actual adopters at the the presence of random bridges makes it harder for an innovation to
build up network effects, or momentum to move ahead, at the initial
stage. But, once the diffusion goes beyond critical mass, the diffusion
tends to be precipitously accelerated in a network with more random
bridges. On the other hand, such an acceleration is almost absent in a
highly cliquish network, thereby causing slower diffusion after the
early stage.

5.2. Managerial implications

Our study implies that there are two network strategies for
launching network products and services: (1) exploiting bridges or
(2) focusing on highly cliquish networks. In the case of information
diffusion or disease spread, it is well known that the presence of many
bridges reduces average social distance in a network, thereby
increasing the speed of diffusion (Watts & Strogatz, 1998). The first
option may work in targeting mass markets with conventional
products and services, where customers do not have to change their
Fig. 6. Average time taken to reach a steady state. consumption habits substantially (Moore, 1991).
176 H. Choi et al. / Industrial Marketing Management 39 (2010) 170–177

However, it turns out that fast penetration comes at a price, i.e., a Appendix A. Parameter values for simulation runs13
greater risk of falling into under-adoption in the case of network
products and services. Firms can reduce this risk by choosing the
second option: cultivating cliquish networks. Indeed, Palm's success in μ σ Q s a β Other characteristics
launching its PDA has been attributed to its focus on cliquish social Fig. 2 200 50 100 10 300 0.2 Single realization
networks in Silicon Valley (Rosen, 2000). In addition, many shrewd Fig. 3 200 50 100 10 300 Variation 10000 simulation runs
marketers have long been exploiting cliquish networks of some sort. Fig. 4 200 50 100 10 300 Variation 10000 simulation runs
Fig. 5a 200 50 100 10 300 Variation 10000 simulation runs
Some examples are Tupperware's party plan, Amway's network
Fig. 5b 200 50 100 10 300 Variation 10000 simulation runs
marketing, and MCI's calling circle. Fig. 6 200 50 100 10 300 Variation 10000 simulation runs
Though a cliquish network facilitates building up an early Fig. 7 50 5 37.5 10 47.5 Variation Single realization
customer base, it tends to inhibit rapid diffusion. To improve the
balance between the two extreme types, marketers can work with
Appendix B. Small-world graph construction algorithm (Watts,
mixed types. That is, they can choose a cliquish network with a few
1999, p.67)
bridges (i.e., a small-world network) to achieve both rapid and full
diffusion. We've seen from the simulation (See Fig. 6) that a consumer
Starting with a perfect 1-lattice, in which each vertex has precisely
network with a few random bridges greatly enhances the diffusion
k neighbors (k/2 on either side), randomly rewire the edges of the
speed. This may explain why viral marketing, which is based on email
lattice, with probability β, using the following algorithm:
or other communication technologies, has been so effective. Viral
marketing appears to exploit the topological properties of both 1. Each vertex i is chosen in turn, along with the edge that connects it
cliquish ties and random bridges. It aims not only to build up w-o-m to its nearest neighbor in a clockwise sense (i, i + 1).
power among close acquaintances, but also to spread information 2. A uniform random deviate r is generated. If r ≥ β, then the edge
rapidly via bridges. Social network services such as MySpace, Face (i, i + 1) is unaltered. If r b β, then (i, i + 1) is deleted and re-
Book, and Second Life have become very popular as new marketing wired such that i is connected to another vertex j, which is
platforms, perhaps for similar reasons. Promising opportunities seem chosen uniformly at random from the entire graph (excluding
to lie ahead for empiricists to confirm this possibility. In fact, a recent self-connections and repeated connections).
study by Vilpponen, Winter, and Sundqvist (2006) verified the role of 3. When all vertices have been considered once, the procedure is
network structure in the context of electronic word-of-mouth repeated for edges that connect each vertex to its next-nearest
communication. neighbor (that is, i + 2), and so on. In total, k/2 such rounds are
In addition, our study confirms what marketers have been completed until all edges in the graph have been considered for
practicing for a while. We found that insufficient buildup of network rewiring exactly once.
effects at the early stage of the product life cycle may be the main
cause of failed diffusion. To avoid this possibility, John McAfee created References
a new launch strategy. He gave away anti-virus software products to
Abrahamson, Eric, & Rosenkopf, Lori (1997). Social network effects on the extent of
individual customers for free. Other software companies have
innovation diffusion: A computer simulation. Organizational Science, 8(3), 289–309.
subsequently followed this strategy. This strategy makes sense for Amaral, L. A. N., Scala, A., Barthelemy, M., & Stanley, H. E. (2000). Classes of small world
software products because a major part of their costs is the networks. Proceedings of the National Academy of Science of America, 97(21),
11149–11152.
development cost (fixed cost). Once a software product is developed,
Arthur, W. Brian (1989). Competing technologies, increasing returns, and lock-in by
and, as long as this cost is somehow recouped, marketers can give it historical events. Economic Journal, 99, 116–131.
away for free to avoid the possibility of under-adoption. When the Barabasi, Albert-Laszlo (2002). Linked: The new science of networks. Cambridge, MA:
diffusion of the product moves beyond the point of critical mass, they Perseus.
Barabasi, Albert-Laszlo, & Albert, Reka (1999). Emergence of scaling in random
can start charging the product, especially an updated version. networks. Science, 286, 509–512.
Barabasi, Albert-Laszlo, Albert, Reka, & Jeong, H. (1999). Mean-field theory for scale-
5.3. Limitations and future research free random networks. Physica A, 272, 173–187.
Buchanan, Mark (2002). Nexus: Small worlds and the groundbreaking science of networks.
New York: Norton.
We now discuss the limitations of our work, which provide Business Week. Flops (1993. 8. 16).
directions for future research. First, due to the bounded focus of the Church, Jeffrey, & Gandal, Neil (1993). Complementary network effects and technolo-
gical adoption. International Journal of Industrial Organization, 11(2), 239–260.
current research, our model simplifies the consumer's utility function. Cowan, Robin, & Miller, John H. (1998). Technological standards with local externalities
Specifically, the product's intrinsic value (Qi in our model) encapsu- and decentralized behavior. Journal of Evolutionary Economics, 8, 285–296.
lates the attribute-based (such as price and quality) as well as non- Dhebar, Anirudh, & Oren, Shmuel S. (1985). Optimal dynamic pricing for expanding
networks. Marketing Science, 4(4), 336–351.
attribute-based (such as size of perceived market/technological risk)
Farrell, Joseph, & Klemperer, Paul (2001). Coordination and lock-in: competition with
values of the product, which are not explicitly expressed. One switching costs and network effects. Working paper Berkeley, California: University of
possibility is that the brand name may enhance the consumers' utility California at Berkeley.
Farrell, Joseph, & Saloner, Garth (1985). Standardization, compatibility, and innovation.
via a quality evaluation (Qi) term. Incorporating such elements and
Rand Journal of Economics, 16(1), 70–83.
seeing how they affect the diffusion of innovative products is Farrell, Joseph, & Saloner, Garth (1986). Installed base and compatibility: Innovation,
obviously a meaningful venue for future research. Our model also product preannouncements, and predation. American Economic Review, 76(5),
has no marketing component. 940–955.
Garber, Tal, Goldenberg, Jacob, Libai, Barak, & Muller, Eitan (2004). From density to
Another direction that future research may pursue is a more destiny: Using spatial dimension of sales data for early prediction of new product
comprehensive view of the market. That is, endogenous determina- success. Marketing Science, 23(3), 419–428.
tion of product quality and price can be explored by theoretical Geroski, P. A. (2000). Models of technology diffusion. Research Policy, 29, 603–625.
Godin, Seth. Unleashing the Ideavirus. www.ideavirus.com. 2000.
analyses of multi-product competition. However, in the context of
research on diffusion failure, excluding competition may still be a
13
useful modeling option because a firm or a new product may fail to We tuned the parameters to exclude extreme possibilities. That is, when either Q
or a is too small or too large, the diffusion of a new product never or always propagates
diffuse fully even in a monopoly situation. Finally, although the
beyond 100 customers for all values of β. In these extreme cases, the varying structure
current model is not inconsistent with the intuitions that practitioners of the consumer network has no impact on innovation adoption. Our analysis focuses
have had, empirical studies based on field data are in immediate on the area between the two extreme cases, where a choice of initial target groups and
demand. their network structure can make a big difference in adoption dynamics.
 H. Choi et al. / Industrial Marketing Management 39 (2010) 170–177 177

Goldenberg, Jacob, Libai, Barak, & Muller, Eitan (2001). Talk of the network: A complex Schoder, Detlef (2000). Forecasting the success of telecommunication services in the
systems look at the underlying process of word-of-mouth. Marketing Letters, 12(3), presence of network effects. Information Economics and Policy, 12, 181–200.
211–223. Srinivasan, Raji, Lilien, Gary L., & Rangaswamy, Arvind (2004). First in, first out? The
Granovetter, Mark S. (1973). The strength of weak ties. American J. Soc., 78, 1360–1380. effects of network externalities on pioneer survival. Journal of Marketing, 68(Jan),
Granovetter, Mark S. (1974). Getting a job: A study of contacts and careers. Chicago: 41–58.
University of Chicago Press. Strogatz, Steve H. (2001). Exploring complex networks. Nature, 401, 268–276.
Harvard Business School Case 9-392-013. Momenta Corporation (A). Sun, Baohong, Xie, Jinhong, & Cao, H. Henry (2004). Product strategy for innovators in
Katz, Elihu (1961). The social itinerary of technical change: Two studies on the diffusion markets with network effects. Marketing Science, 23(2), 243–254.
of innovation. Studies of innovation and communication to the public. Stanford, CA: Van den Bulte, Christophe, & Lilien, Gary L. (2001). Medical innovation revisited: Social
Stanford University, Institute for Communication Research. contagion versus marketing effort. American J. of Sociology, 106(March), 1409–1435.
Katz, Michael L., & Shapiro, Carl (1985). Network externalities, competition and Vilpponen, Antti, Winter, Susanna, & Sundqvist, Sanna (2006). Electronic word-of-
network effects. American Econ. Rev., 75(3), 424–440. mouth in online environments: Exploring referral network structure and adoption
Katz, Michael L., & Shapiro, Carl (1986). Technology adoption in the presence of behavior. Journal of Interactive Advertising, 6(2), 71–86.
network externalities. Journal of Political Economy, 94, 822–841. Watts, Duncan J. (1999). Small worlds. Princeton, NJ: Princeton University Press.
Lee, E., Lee, J., & Lee, J. (2006). Reconsideration of the winner-take-all hypothesis: Watts, Duncan J., & Strogatz, Steven H. (1998). Collective dynamics of ‘small-world’
Complex networks and local bias. Management Science, 52(12), 1838–1848. networks. Nature, 393(June), 440–442.
Lee, Yukuan, O, & Connor , Gina Colarelli (2003). New product launch strategy for network Xie, Jinhong, & Sirbu, Marvin (1995). Price competition and compatibility in the
effects products. Journal of the Academy of Marketing Science, 31(3), 241–255. presence of positive demand externalities. Management Science, 41(5), 909–926.
Li, Xiaotong (2005). Cheap talk and bogus network externalities in the emerging
technology market. Marketing Science, 24(4), 531–543.
Mahajan, Vijay, Muller, Eitan, & Bass, Frank M. (1990). New product diffusion models in
Hanool Choi is an Instructor at the Department of Economic and Commerce,
marketing: A review and directions for research. J. Marketing, 54, 1–26 (January).
Mahler, Alwin, & Rogers, Everett M. (1999). The diffusion of interactive communication Kyemyung University. He teaches Internet Marketing, e-Business and Customer
innovations and the critical mass: The adoption of telecommunications services by Relationship Management. He received his Ph.D. degree in management information
German banks. Telecommunications Policy, 23(10–11), 719–740. systems from the Korea Advanced Institute of Science and Technology (KAIST). His
Midgley, David F., Morrison, Pamela D., & Roberts, John H. (1992). The effect of network research interests include organizational learning, evolutionary economics, knowledge
structure in the industrial diffusion process. Research Policy, 21(6), 533–552. management, and customer relationship management.
Moore, Geoffrey A. (1991). Crossing the chasm. New York, NY: Harper Business.
Morrison, Pamela D., Roberts, John H., & Midgley, David F. (2000). Opinion leadership
amongst leading edge users. Australasian Marketing Journal, 8(1), 5–14.
Sang-Hoon Kim is Associate Professor of Marketing at the Graduate School of
Onnela, J. P., Samaraki, J., Hyvönen, J., Szabó, G., Lazer, D., Kaski, K., Kertész, J., & Barabási,
A. -L. (2007). Structure and tie strengths in mobile communication networks. Business, Seoul National University, where he teaches marketing courses including
Proceedings of the National Academy of Sciences of the USA, 104(18), 7332–7336. High-Tech Marketing and New Product Development. He received his Ph.D. degree
Padmanabhan, V, Rajiv, Surendra, & Srinivasan, Kannan (1997). New products, upgrades, from Stanford University. He also holds an MBA degree from the University of Chicago.
and new releases: A rationale for sequential product innovation. Journal of Marketing His research interests include new product design, innovation diffusion, and other
Research, 34, 456–472. strategic issues related to high-technology products. He has published papers in
Pae, Jae H., & Hyun, Jung Suk (2002). The impact of technology advancement strategies marketing journals such as Industrial Marketing Management, Journal of Retailing,
on consumers' patronage decisions. Journal of Product Innovation Management, 19, Journal of Business Research, and Technovation. Many of his research papers are
375–383. presently under review at major marketing journals.
Paster-Satorras, Romualdo, & Vespignani, Alessandro (2001). Epidemic spreading in
scale-free networks. Physical Review Letters, 86, 3200–3203.
Redner, Sidney (2002). Networking comes of age: How an old concept has grown new
wings. Nature, 418, 127–128. Jeho Lee is an Associate Professor of Strategy at the KAIST Business School, where he
Rogers, Everett M. (1983). Diffusion of innovations. New York, NY: Free Press. teaches strategic management courses including Strategic Management and High-Tech
Rohlfs, Jeffrey H. (1978). A theory of interdependent demand for a communications Strategy. He received his Ph.D. from the Wharton School of the University of
service. Bell Journal of Economics and Management Science, 5(1), 16–37. Pennsylvania. His research interests include innovation-based competition, network
Rohlfs, Jeffrey H. (2001). Bandwagon effects in high-technology industries. Cambridge, effects, and other strategic issues related to high-tech industries. He has published
MA: MIT Press. papers in journals such as Management Science, Organization Science, and Strategic
Rosen, Emmanuel (2000). The anatomy of buzz: How to create word-of-mouth marketing. Management Journal.
New York, NY: Doubleday.
Shapiro, Carl, & Varian, Hal. R. (1999). Information rules. Boston, MA: Harvard Business
School Press.