Demystifying Disruption: A New Model for Understanding and Predicting Disruptive Technologies

Published online ahead of print December 30, 2010

informs ®
Articles in Advance, pp. 1–16
issn 0732-2399 eissn 1526-548X doi 10.1287/mksc.1100.0617
© 2010 INFORMS
not be posted on any other website, including the author’s site. Please send any questions regarding this policy to permissions@informs.org.
Copyright: INFORMS holds copyright to this Articles in Advance version, which is made available to institutional subscribers. The file may

Demystifying Disruption: A New Model for
Understanding and Predicting Disruptive Technologies
Ashish Sood
Goizueta School of Business, Emory University, Atlanta, Georgia 30322, ashish_sood@bus.emory.edu

Gerard J. Tellis
Marshall School of Business, University of Southern California, Los Angeles, California 90089, tellis@usc.edu

T he failure of firms in the face of technological change has been a topic of intense research and debate,
spawning the theory (among others) of disruptive technologies. However, the theory suffers from circular
definitions, inadequate empirical evidence, and lack of a predictive model. We develop a new schema to address
these limitations. The schema generates seven hypotheses and a testable model relating to platform technologies.
We test this model and hypotheses with data on 36 technologies from seven markets. Contrary to extant theory,
technologies that adopt a lower attack (“potentially disruptive technologies”) (1) are introduced as frequently
by incumbents as by entrants, (2) are not cheaper than older technologies, and (3) rarely disrupt firms; and (4)
both entrants and lower attacks significantly reduce the hazard of disruption. Moreover, technology disruption
is not permanent because of multiple crossings in technology performance and numerous rival technologies
coexisting without one disrupting the other. The proposed predictive model of disruption shows good out-of-
sample predictive accuracy. We discuss the implications of these findings.
Key words: technology disruption; firm disruption; demand disruption; correlated hazards; prediction of
disruption
History: Received: November 19, 2008; accepted: September 10, 2010; processed by Gary Lilien. Published
online in Articles in Advance.

Introduction He posited the theory of S-curves, which suggested
Technological change is critically important to firms that technologies evolve along successive S-curves;
for several reasons. First, it has the potential to incumbents fail if they miss to switch to a new tech-
obsolete assets, labor, and intellectual capital of nology that passes the incumbent’s technology in
incumbents in the market. For example, electronic performance. Tushman and Anderson (1986) refined
commerce has obsoleted many of the old business this theory by distinguishing between competence-
processes in the banking industry. Second, it can cre- enhancing and competence-destroying technological
ate entirely new markets, with new products, new changes. They argued that failure occurred only when
customers, and exploding demand. For example, MP3 the new technology destroyed, rather than enhanced,
technology facilitated the iPod revolution, with mas- the expertise of the incumbents. Other researchers
sive demand for products, services, and accessories. built on the theory of punctuated equilibrium (Gould
Third, technological evolution enables firms to tar- and Eldredge 1977) to propose a demand-side expla-
get new segments within a market with improved nation for the phenomenon of disruption (Levinthal
products. For example, improvements in LCD mon- 1998, Adner 2002, Adner and Zemsky 2005, Mokyr
itors enabled firms to target the segment of con- 1990). They suggested that disruption occurs when
sumers with mobile computing needs. Fourth, and a new technology that starts in one domain moves
most importantly, incumbents often misinterpret the to a new domain with potentially higher demand
potential impact of the new technology, and this error and additional resources. Christensen (1997) proposed
causes their demise. For example, microcomputers the theory of disruptive innovations. It posited that
killed off manufacturers of minicomputers. disruption occurred when an initially inferior tech-
The failure of firms in the face of technological nology introduced by a new entrant improved to
change has been a topic of intense research and debate meet the needs of the mass market (Bower and
in the strategy literature (e.g., Schumpeter 1934, Christensen 1995).
Freeman 1974, Henderson and Clark 1990, Henderson Of the three theories, Christensen’s (1997) theory
1993, Cohen and Levinthal 1990). An early attempt has won the most attention and widest acclaim from
to understand this phenomenon was by Foster (1986). both managers and researchers (Henderson 2006,
1

Sood and Tellis: Demystifying Disruption: New Schema, Model, and Findings
2 Marketing Science, Articles in Advance, pp. 1–16, © 2010 INFORMS

Gilbert 2003, King and Tucci 2002, Adner 2002, can predict the hazard of disruption of a new technol-
Adner and Zemsky 2005, Grove 1998, Gilbert and ogy. The next section presents the method and results

Bower 2002). Books on disruption have sold hundreds of the study. The last section discusses the findings,
of thousands of copies, readings on disruption are limitations, and implications of the study.
among the most used in MBA classes, and a Google
search suggests that the term “disruptive innovation”
is the most popular innovation term. Theory
However, researchers have pointed to at least four This section presents a new schema and develops
weaknesses in the theory. First, researchers claim that hypotheses about the impact of technological change
the central thesis about a disruptive technology caus- on markets.
ing disruption appears to be tautological (Cohan 2000,
Danneels 2004, Markides 2006). Christensen’s writ- Schema
ings alone suggest that the term could take on differ- Our new schema defines constructs, expands the set
ent meanings (Danneels 2004, Tellis 2006). The major of drivers of disruption, and provides the foundation
issue is the use of the same term to describe both the for hypotheses and a predictive model. Our defini-
causative agent (disruptive technology) and the effect tions cover types of technologies, types of technolog-
(disruption). For example, Kostoff et al. (2004, p. 142) ical attack, dynamics of competition, and domains of
state, “disruptive technologies can be revealed as disruption. To avoid circularity, we define concepts
being disruptive only in hindsight.” in terms of technological characteristics rather than
Second, the theory is ambiguous as to which do- effects that lead to premises true by definition, e.g.,
main of disruption the theory applies (Danneels 2004, “disruptive,” “sustaining,” or “revolutionary.”
Markides 2006). We identify three domains of dis- Definition of Technologies. What is a technology?
ruption: technology domain (performance evolution), Following Sood and Tellis (2005), we define a technol-
firm domain (competitive survival), and demand ogy as a platform based on a unique scientific prin-
domain (market acceptance). ciple, on which firms manufacture products to serve
Third, many authors point to a scarcity of empir- customers’ needs in a particular market. For example,
ical evidence to validate the generalizability of the in the lighting market, incandescence, fluorescence, or
claims (Govindarajan and Kopalle 2006, Danneels light-emitting diodes (LED) are three entirely inde-
2004, Tellis 2006, Utterback and Acee 2005). Danneels pendent scientific principles, each of which provides
(2004, p. 251) calls for new research on a “comprehen- a platform on which firms produce products to
sive list of technologies” to examine “the mechanisms serve consumers’ need for light. Thus, they constitute
and effects” of disruptive technologies on firms and three independent technologies for lighting. Innova-
markets. Cohan (2000) suggests that the results on the tions within each technology (platform or scientific
effects of disruptive technologies might not hold as principle) could cause it to improve in performance
well if the sample were drawn randomly. over time. We classify these innovations as belonging
Fourth, the theory lacks predictive ability (Tellis either to component innovations (in parts or mate-
2006, Kostoff et al. 2004). Barney (1997) urges devel- rials) or design innovations (in layout or links) (see
opment of a predictive model to rule out cherry- Table 1). However, as long as the scientific principle
picking or luck as an alternative explanation of why remains the same, we assign all these innovations to
some technologies are more disruptive than others. the same technology. For example, large and compact
To summarize, we seek answers to the following fluorescent bulbs exemplify various design innova-
specific questions with the goal of infusing this the- tions within the fluorescence technology. Carbide and
ory with validity: (1) What is a disruption? (2) Who tungsten filaments exemplify component innovations
introduces a disruptive technology, and who survives within incandescence technology. The improvement
disruption? (3) What are the causes of disruption? in performance of a platform technology over time is
(4) When does disruption occur, and how can we due to these design and component innovations.
predict it? Definition of Technological Attack. How does a
We make three contributions to prior literature in new technology attack the dominant technology? To
this paper. First, we develop a new schema that iden- answer this question, we first identify an objective
tifies key variables, defines key terms, and allows us measure of the performance of a technology, which
to derive seven testable hypotheses. Second, we con- is important to the mainstream segment and forms
duct an empirical test of the hypotheses by sampling the primary dimension of competition in the market.
all platform technologies in seven markets, rather We define a market as a set of consumers whose sim-
than selectively sampling those that may or may not ilar needs are being served by a set of competing
fit the hypotheses. Third, we develop a model that technologies, firms, and brands. For example, storage

Marketing Science, Articles in Advance, pp. 1–16, © 2010 INFORMS 3

Table 1 Definition of Technologies

Our schema Christensen’s (1997) schema

Our terms Our basis Christensen’s termsa Christensen basis Examples

Platform Unique scientific principle Disruptiveb New technology inferior on Digital vs. analog cameras
primary but superior on
secondary dimension
Sustaining breakthrough New technology superior on Fiber optics vs. analog
primary communications
Design Linkages or layout within Disruptiveb (disruptive as defined above) 5.25” vs. 3.5” floppy drives
same scientific principle
Sustaining incremental Small improvement in current Compact vs. regular fluorescent
technology on primary
dimension
Component Materials or parts within same Disruptiveb (disruptive as defined above) Thin-film vs. ferrite heads
scientific principle
Sustaining incremental Small improvement in current DVD vs. CD
technology on primary
dimension
a
Christensen’s (1997) terms will never perfectly match with ours because ours are defined on characteristics of technology, whereas his seem to be defined
on effects happening to firms.
b
Christensen uses the term disruptive for all three levels: platform, design, and component.

capacity is an important primary dimension of com- Again, following Christensen (1997), we assume
petition in the market for computer storage technolo- the segments have fixed preferences but technolo-
gies. All other attributes of technologies would be gies improve over time, as shown by the arrows
secondary dimensions of competition. We then define in Figure 1. Both technologies improve on the pri-
two types of attacks: lower and upper attacks. A lower mary dimension over time. At time t2 , the dominant
attack occurs when, at the time of its entry, a new tech- technology exceeds the needs of the mainstream seg-
nology performs worse than the dominant technology ment on this dimension. However, the new technol-
on the primary dimension of performance. An upper ogy improves sufficiently on the primary dimension
attack occurs when, at the time of its entry, a new tech- so as to appeal to the mainstream segment, because
nology performs better than the dominant technology it now meets its needs on both the primary and sec-
on the primary dimension of performance. ondary dimensions. Thus, at time t2 , demand of both
Dynamics of Competition. What are the dynamics segments shifts from the dominant technology to the
of competition between the new technology and dom- new technology. Christensen refers to this event as dis-
inant technology? For simplicity of exposition, follow- ruption. The niche segment plays the role of providing
ing Christensen (1997), we assume the market has two a demand for the new technology while it improves
technologies (dominant and new), two dimensions
(primary and secondary), and two segments: a main-
Figure 1 Theory of Disruptive Innovations
stream and a niche. (The empirical analysis allows
for multiple technologies and dimensions.) Figure 1
illustrates the dynamics of competition between the Time t2
dominant technology and the new technology on the
primary and secondary dimensions in one market.
Both segments have similar needs but differ in their Mainstream
Time t1 customers
preferences: the mainstream segment favors the pri-
Tdominant
mary dimension, whereas the niche segment favors
the secondary dimension, as shown by their locations
Primary dimension

in Figure 1. However, both dimensions are both objec-
tive and vector—i.e., more is better. At time t1 , the
dominant technology is strong on the primary dimen-
sion but weak on the secondary dimension, whereas
the reverse holds for the new technology. Given this Tnew
Niche
preference distribution, at time t1 , the mainstream
segment prefers the dominant technology, whereas
the niche segment prefers the new technology. Secondary dimension


Table 2 Classification of Studies in Literature Based on Proposed Schema

Performance (entry) of new technology

Domain of disruption Lower attack Upper attack Both attacks

Firm (competitive survival) Disruptive innovations (Christensen Sustaining breakthrough innovations Low-end and high-end disruptions
1997), potentially disruptive (Christensen 2003), technological (Govindarajan and Kopalle 2006)
innovations (Raffi and Kampas discontinuity competence
2002), radical innovation enhancing/destroying innovations
(Utterback and Acee 2005), (Tushman and Anderson 1986)
low-end disruptions (Govindarajan
and Kopalle 2006)
Demand (market acceptance) Disruptive technologies (Adner Down-market progression (Utterback Attack from below and down-market
2002), new technology (Levinthal and Acee 2005), new technology progression (Utterback and
1998, Utterback and Acee 2005) (Levinthal 1998) Acee 2005)
Technology (performance New technology (Levinthal 1998), Platform technologies (Sood and
evolution) radical innovation (Utterback Tellis 2005)
1994), theory of S-curves (Foster
1986), discontinuous innovation
(Dosi 1982)
All domains This paper

in performance on the primary dimension and meets terms used in the prior literature (see Table 2). In par-
the needs of the mainstream segment. Note that for ticular, Christensen’s (1997) term disruptive technology
this analysis, it is sufficient to assume segments with would be equivalent to a new technology adopting
fixed preferences, as does Christensen (1997), so long a lower attack that is also superior to the dominant
as technologies improve over time. technology on a secondary dimension (see Tables 1
Domains of Disruption. We identify three domains and 2). The term used by Christensen et al. (2004),
of disruption, in each of which disruption could occur sustaining breakthrough, would be equivalent to a
independently: technology, firm, and demand. Tech- new technology adopting an upper attack. Note that
nology disruption occurs when the new technology Christensen’s (1997) term sustaining incremental seems
crosses the performance of the dominant technology equivalent to design and component innovations that
on the primary dimension of performance. We use improve a current technology’s performance along the
the term dominant technology to refer to the technol- primary dimension of performance. Design innova-
ogy with the best performance on the primary dimen- tions are also what Henderson and Clark (1990) call
sion at the time a new technology enters the market. architectural innovations. From Table 2, note also that
Firm disruption occurs when the market share of a relative to the literature, this is the only study that
firm whose products use a new technology exceeds covers all three domains of disruption and both types
the market share of the largest firm whose products of attacks in one empirical analysis.
use the highest-share technology. We use the term
highest-share technology to refer to the technology with Hypotheses
the highest market share at the time a new tech- With the help of the above schema, we formu-
nology enters the market. Note that by this defini- late seven testable hypotheses—three on technological
tion, either an entrant or an incumbent can disrupt entry and four on the hazard of disruption.
the largest firm whose products use the dominant
technology.1 Demand disruption occurs when the total Technological Entry. Who introduces technologies
share of products in the market based on a new tech- that use a lower attack (potentially disruptive)? Pro-
nology exceeds the share of products based on the ponents of the theory of disruptive innovations assert
dominant technology. We use the term market dis- that “the firms that led the industry in every instance
ruption to refer inclusively to all three domains of of developing and adopting disruptive technologies
disruption. were entrants to the industry, not its incumbent lead-
Summary. These constructs for technology, direc- ers” (Christensen 1997, p. 24). Why does this occur?
tion of attack, and domains of disruption constitute According to the theory, entrants are willing to exper-
our new schema. The schema allows us to organize iment with new technologies targeted toward niche
segments (Christensen 1997). These firms are also not
1
However, if the firm with the highest market share is farsighted
deterred by the lower profit margins and smaller sales
and itself builds the highest market share in this new technology, volumes from niche segments relative to the main-
then no firm disruption would occur. stream segment (Christensen and Rosenbloom 1995).


On the other hand, incumbents’ firms get most of Hazard of Disruption. Which type of firm is more
their revenues and profits from the existing technol- likely to disrupt? The theory suggests that incum-

ogy marketed to the mainstream segment (Raffi and bents are unlikely to disrupt because they focus
Kampas 2002). So they devote all their efforts and predominantly on their current customers for sev-
energies to perfect their current technology marketed eral reasons. First, incumbents get their revenues
to the mainstream segment. The established routines primarily from their current mainstream segment,
within the incumbent firms do not provide sufficient whereas entrants target the less profitable segments
incentives to develop these new skills and knowledge and the less demanding customers (Christensen et al.
associated with the new technology. These arguments 2004). Second, incumbents do not possess appropriate
suggest the following hypothesis. resources and competencies to compete with entrants,
Hypothesis 1 (H1). Technologies using a lower attack who introduce a new value proposition and serve
(potentially disruptive) come primarily from entrants. demand on a new secondary dimension (see Fig-
ure 1). For example, incumbents making CRT mon-
Who introduces technologies that use an upper
itors could not compete effectively with entrants
attack (sustaining breakthrough)? Christensen et al.
making LCD monitors on the secondary dimension
(2004) suggest that the technologies adopting an
of compactness even though they made efforts to
upper attack (sustaining breakthrough) are introduced
reduce the size of old CRT monitors by introduc-
mainly by incumbents. Incumbents focus on satisfying
ing flat-screen CRT monitors. Third, incumbents often
their current demanding customers with both simple
do not appreciate the real threat of a new technol-
incremental improvements and breakthrough jumps
ogy (Christensen and Raynor 2003, Henderson 2006,
up the current trajectory of performance improve-
Gilbert 2003). For example, incumbents making CRT
ment. Incumbents have more resources, higher profits,
monitors discounted the potential increase in resolu-
and more at stake than new entrants. Incumbents can
readily deploy sustaining breakthrough innovations tion of LCD monitors. These arguments suggest the
because they may not require substantial changes to following hypothesis.
their overall value-creating system (business model). Hypothesis 4 (H4). The hazard of disruption is higher
They can use the same manufacturing and distribution from an entrant than from an incumbent.
process if the new technology fits their R&D capabil-
ities and delivers benefits that are consistent with the What type of technological attack is more likely to
brand promise. These arguments lead to the following cause firm or demand disruption? The theory sug-
hypothesis. gests that a lower attack is deceptively more danger-
ous than an upper attack because firms that focus
Hypothesis 2 (H2). Technologies using an upper at- on the dominant technology often do not perceive
tack (sustaining breakthrough) come primarily from the new technology as a threat until it is too late.
incumbents. The lower performance lulls incumbents into think-
How do technologies using a lower attack differ ing that these new technologies will not appeal to
from the dominant technology? The theory of disrup- the mainstream segment, which values the high per-
tive innovations suggests that firms target the less- formance of the dominant technology. Over time,
demanding niche customers with lower-performing the improvement of the dominant technology on the
technologies. The technologies using a lower attack primary dimension exceeds the needs of the main-
are “typically simpler, cheaper, easier, and more con- stream segment creating conditions of “performance
venient than dominant technologies” (Christensen oversupply” (Christensen 1997, p. 211). Disruption
1997, p. 267). Even though these technologies may occurs when the improvement of the new technol-
improve over time, at entry these technologies are ogy increases its appeal to the mainstream segment.
crude but more affordable than dominant technolo- When this change occurs, Utterback (1994) asserts that
gies. Underlying all these arguments is Christensen’s incumbents lack the required set of capabilities to
(1997) assumption that performance and cost are cor- compete with entrants regardless of how well they are
related, and a lower attack also makes the technol- positioned to serve the mainstream segment. These
ogy less expensive. Moreover, new technologies are arguments suggest the following hypothesis.
initially less feature-rich and focus on primarily pro-
viding the basic consumer benefit. By targeting only Hypothesis 5 (H5). The hazard of firm or demand dis-
the small niche segments, firms also reduce costs by ruption is higher if a new technology uses a lower attack.
limiting the product range. These arguments suggest How does firm size affect disruption? Extant the-
the following hypothesis. ories relate strategies on technology to size of firms.
Hypothesis 3 (H3). Technologies using a lower attack Small firms lack the weaknesses that often beset
(potentially disruptive) are priced lower than dominant large firms like technological inertia (Ghemawat
technologies at entry. 1991), complacency (Robertson et al. 1995), arrogance


(Lieberman and Montgomery 1988), and reluctance new technology as a control variable. Second, prior
to cannibalize existing products (Chandy and Tellis literature suggests that technological change increases

1998). Small firms are more research productive than with time (Sood and Tellis 2005). New technologies
large firms, especially in highly innovative indus- may find it easier to disrupt older technologies than
tries requiring skilled labor (Acs and Audretsch 1988). other technologies. Hence, we also include the order
During the early life of technologies, these capabilities of entry of technologies in a market as an additional
are more important than the advantages of scale and control variable.
scope of large firms (Pavitt and Wald 1971, Acs and In summary, our new schema leads to seven dis-
Audretsch 1988). These arguments suggest the follow- tinct, falsifiable hypotheses about disruption and two
ing hypothesis. control variables, which we now proceed to test.
Hypothesis 6A (H6A). The hazard of disruption is
higher if a new technology is introduced by a small firm. Method
However, some recent research suggests that We test these hypotheses using data from seven mar-
incumbents may be better positioned to take advan- kets. We collected these data using the historical
tage of new technologies because of superior finan- method (Golder and Tellis 1997, Sood and Tellis 2009).
cial and managerial resources (Hill and Rothaermel Below we detail the sample selection and sources for
2003, Rothaermel 2001), R&D capability (Rothaermel collecting the data. Online Appendix A in the elec-
and Hill 2005), and complementary assets (Tripsas tronic companion describes the procedure. An elec-
1997). Chandy and Tellis (2000) find that in recent tronic companion to this paper is available as part of
decades, radical innovations come mainly from large the online version that can be found at http://mktsci
firms. These arguments lead to the following rival .pubs.informs.org/.
hypothesis.
Sample Selection
Hypothesis 6B (H6B). The hazard of disruption is We used three criteria in selecting markets. First, we
lower if a new technology is introduced by a small firm. need markets with a minimum of two technologies
Is the hazard of disruption higher if a technol- per market to observe the phenomenon of disrup-
ogy is priced lower than the dominant technology tion. Second, we need a mix of relatively young and
at entry? The theory of disruptive innovations sug- relatively old markets. Third, we need some overlap
gests that products based on technologies that adopt with past research to enable comparison. On the basis
a lower attack are initially priced lower and are of of these criteria, we chose seven markets: electrical
a cruder design than the dominant technology. Char- lighting, data transfer, computer memory, computer
acteristics of such technologies make them attractive printers, display monitors, music recording, and anal-
to niche customers but not the mainstream segment. gesics markets. Note that the first two are utilities,
For example, lower costs reduce the perceived risk, the next four are consumer electronics, and the last
whereas crude designs reduce the perceived complex- is pharmaceutical. Thus, the sample crosses a broad
ity of the new technology (Rogers 2003). Moreover, spectrum of technologies, markets, and products with
such technologies target new consumers or those in technologies that vary in age from a few years to
low-end markets (Christensen 1997) avoiding direct more than a century. A unique feature of our sam-
competition with the dominant technology. Reduced ple is that we selected all platform technologies that
competition may help firms to maintain lower costs were ever commercialized within each market. Some
by reducing marketing expenditures and to transfer of these technologies did not achieve much of a pres-
these advantages to customers via lower prices. These ence in the mainstream segment and remained limited
arguments suggest the following hypothesis. to a niche. In all, we identify 36 technologies: nine in
computer memory, six in display monitors, five each
Hypothesis 7 (H7). The hazard of disruption is higher in computer printers, electrical lighting, and music
if a new technology is lower priced than the dominant tech-
recording, and three each in analgesics and data trans-
nology at entry.
fer. Online Appendix B in the electronic compan-
Control Variables. We use two control variables: ion describes these 36 technologies briefly. In each
change in performance of new technology and order of these technologies, improvements occur because
of entry for two reasons. First, extant theory sug- of design and component innovations. Because the
gests that higher performance of the new technol- latter number in the hundreds, for ease of analysis
ogy increases its appeal to the mainstream segment. and exposition, we track disruptions only in plat-
This improved performance of the new technology form technologies and not in design and component
increases the hazard of disruption of the dominant innovations. Thus, our results apply to platform
technology. So we use change in performance of the technologies.


Sources S dummy variable for firm size, which is 1 if firm
The primary sources of our data are technical jour- introducing the new technology is small at the

nals, industry publications, press releases, time lines time of entry of new technology and 0 otherwise;
of major firms, white papers published by R&D C dummy variable for relative price, which is 1
organizations, annual reports of industry associa- if the new technology is priced lower than the
tions, and records in museums that profiled inno- dominant technology at the time of entry and 0
vations and the development of markets. We collect otherwise;
information on technologies available in each mar- O order of entry of the new technology; and
ket, the performance of these technologies at various P percentage change in performance of the new
stages of technological evolution, the supplier of these technology over the prior year.
technologies, and the market success of each technol-
The subscripts i and t refer to technology and
ogy. We also collect information on technological per-
time, respectively. T and F are error terms assumed
formance on both primary and secondary dimensions.
to be normally distributed; T ∼ N 0 T and F ∼
2
2
Model N 0 F . Further, we allow the two error terms to be
We develop a correlated hazards model based on correlated and assume joint normality such that
the method developed by Lillard (1993). The model T
0 2
i T TF T F
may be characterized as follows. A new technology ∼N (3)
F 2
is introduced in an existing market. From the point i 0 TF T F F
of introduction, the new technology threatens to dis-
rupt both old technologies and incumbent firms using Note that we do not include the direction of attack
old technologies in the market. The hazards of both in Equation (1) for the hazard of technology dis-
technology and firm disruption are influenced by ruption to avoid circularity. The terms 1 T1t and
a number of time-related factors, including perfor- 1 T2t represent the dependence of respective hazards

mance of the technology and age of the market, and on time via piecewise-linear splines, as follows. We
by a set of exogenous covariates like relative price, denote the time at which the dominant technology or
order of entry, direction of attack, and source of new firm becomes at risk of disruption by t0 and subdi-
technology. We limit the analyses to only firm and vide the duration t − t0 into Ni + 1 discrete periods
technology disruption because demand disruption is that sum to the calendar time, but which allow the
conflated with firm disruption in our sample; i.e., slope coefficients to differ within ranges of time sepa-
demand disruption generally occurs with firm disrup- rated by the Ni nodes. The spline variable for the kth
tion or always follows it within a short time. How- period between nodes k−1 and k is given by Tk t =
ever, the same model can be extended to investigate max 0 min t − k−1 k − k−1 . So the two baseline
hazard of demand disruption for other data using hazards can be written as
the same approach. We account for the correlation N1 +1 N2 +1
between the two hazards to avoid inconsistent stan- 1 T1t = 1k T1kt and 1 T2t = 1k T2kt (4)
dard errors (Lillard 1993). k=1 k=1

The model is essentially a proportional hazard, with Let T T and F F represent the conditional
covariates shifting the baseline hazard (Allison 1995). likelihood functions of the time to next technology
In particular, we model the log hazard of technology and firm disruption, respectively. Then we can write
and firm disruption, respectively, the joint marginal likelihood as
ln hT =
it 0 + 1 T1t + 2 Ei + 3 Si + 4 Ci T T F F T F T F
f d d (5)
+ 5 Oi + 6 Pit + T
it (1) T F

ln hF =
it 0 + 1 T2t + 2 Ei + 3 Li + 4 Si Here, f T F is the joint distribution of the unob-
served heterogeneity components specified in Equa-
+ 5 Ci + 6 Oi + 7 Pit + F
it (2)
tion (3). Thus, conditional on , technology disruption
and firm disruption are independent of each other
where
and the conditional joint likelihood can be obtained
E dummy variable for incumbency, which is 1 if firm by simply multiplying the individual likelihoods. The
is an entrant at the time of entry of new technol- marginal joint likelihood is obtained by integrating
ogy and 0 otherwise; out the heterogeneity term (see Online Appendix C
L dummy variable for attack, which is 1 if the new in the electronic companion for details).
technology employs a lower attack at the time of We estimate Equations (1) and (2) jointly as a
entry and 0 otherwise; system of equations with technology-specific errors


correlated across the two equations with aML, a mul- Example of Evolution of Technologies in the
tiprocess multilevel modeling software (Lillard and Lighting Market

Panis 2003). The complete model is estimated We describe the technological competition and dis-
using full-information maximum likelihood. Where ruption in the external lighting market. The mar-
a closed-form solution does not exist, numerical ket exhibits a total of five platform technologies
approximation can be used (Schweidel et al. 2008). between 1879 and 2000 (see Figure 2(a)). Only the first
This software employs the Gauss-Hermite quadrature two technologies were introduced by small firms—
to approximate the normal integrals. incandescent lighting by Edison Lamp Works in 1879
and arc-discharge lighting by Cooper Hewitt Lamp
Co. in 1908. Some years after the entry of the arc-
Results discharge lighting, General Electric acquired Cooper
We identify the primary dimension of competition Hewitt Lamp Co. Philips, an incumbent, introduced
among competing technologies for each market two of the other three technologies (gas-discharge
and an objective measure of this dimension (see lighting and microwave electrodeless discharge light-
Table 3(a)). The data on the markets range in time ing) in 1908 and 1932, respectively. RCA, an entrant
from 53 years for the computer printers market to to the market, introduced LED lighting in 1971. Two
127 years for the external lighting market. In all, we technologies, arc-discharge lighting and gas-discharge
have 1,942 technology-years of data for testing the lighting, used upper attacks at the time of entry.
seven hypotheses. Across the sample, only 55% of all The other three technologies used lower attacks. We
technologies cause disruption. Of these, 33% cause observe instances of technology disruption in the mar-
only technology disruption and 22% cause both tech- ket that occurred each time the arc-discharge lighting
nology and firm disruption. The remaining 45% of all and gas-discharge lighting crossed each other in per-
technologies cause no disruption at all. formance. Figures 2(b) to 2(d) illustrate the evolution
We first present an example of technology evolution of technologies in three other markets.
and market disruption in the lighting market. We then
present the results of descriptive analysis, estimates Analysis of Technological Entry
of the hazard model, and out-of-sample predictions of We observe only the technologies that enter and not
the hazard model. Finally, we present results on var- those that do not enter a market. Hence, we use a
ious patterns of disruption, including the emergence cross-tabular analysis (and not log-linear models) to
of new secondary dimensions and the robustness of test the first three hypotheses. Based on the extant the-
results. ory, H1 predicts that technologies entering via a lower
attack come primarily from entrants. However, con-
Table 3 Patterns of Entry trary to the theory and the hypothesis, 47% of lower
attacks are from entrants and the remaining 53% are
Market Primary dimension Measure
from incumbents (see Table 3(b)). This difference is
(a) Dimensions of competition not significantly different from 0 ( 2 = 0 1; p = 0 80).
Electrical lighting Lighting efficacy Lumens per watt Based on the extant theory, H2 predicts that tech-
Data transfer Transfer speed Bits per second nologies entering via an upper attack come primar-
Computer memory Storage capacity Megabytes per
square inch
ily from incumbents. However, contrary to the theory
Computer printers Print resolution Dots per square inch and the hypothesis, only 42% of upper attacks are
Display monitors Screen resolution Pixels per square inch from incumbents, whereas the majority (58%) are
Music recording Storage capacity Megabytes per from entrants (see Table 3(b)). The difference is not
square inch significantly different from 0 ( 2 = 0 7; p = 0 39).
Analgesics Efficacy in pain reduction Number needed to
treat (NTT) H3 predicts that technologies entering via a lower
attack are cheaper than dominant technologies at the
Lower attack (%) Upper attack (%) time of entry. However, contrary to the theory and the
(b) Frequency of new technologies by attack, source, and price hypothesis, only 12% of technologies using a lower
Source attack are cheaper than dominant technologies at
Entrant 47 58 entry (see Table 3(c)). The rest (88%) are more expen-
Incumbent 53 42 sive. The difference is significant ( 2 = 9 9; p < 0 001).
Total 100 100
Price relative to dominant Analysis of Hazard of Disruption
technology at entry The results of the hazards model are in Table 4.
High price 88 89 The coefficients of the independent variables in this
Low price 12 11
model test the hypotheses H4 to H7. We estimated the
Total 100 100
model for technology disruption and firm disruption


Figure 2 Empirical Path of Technological Evolution in Four Markets

(a) External lighting (b) Desktop printers*
200
1.E+02
Incandescent Dot matrix
180
Arc discharge Ink jet
160 Gas discharge
1.E+01 Laser
LED
140 MED Thermal
Lighting efficacy

120

Resolution
1.E+00
100
80 1.E–01
60
40 1.E–02
20
0 1.E–03
1879 1894 1909 1924 1939 1954 1969 1984 1999 1978 1981 1984 1987 1990 1993 1996 1999 2002 2005

(c) Data transfer * (d) Analgesics
1.E +13
0.800
Cu/Al wires
1.E +12
Fiber optics
0.700
1.E +11 Wireless
1.E +10 0.600
Data transfer speed

1.E +09
0.500
Efficacy
1.E +08
0.400
1.E +07
1.E +06 0.300
1.E +05
0.200 Opiods (narcotics)
1.E +04 NSAIDs
0.100 Acetaminophen
1.E +03
1.E +02 0.000
1962 1966 1970 1974 1978 1982 1986 1990 1994 1998 2002 2006 1950 1960 1970 1980 1990 2000
∗
Performance of Y axis is in log scale.

separately (Equations (1) and (2), respectively) and for technology disruption and at 8 and 28 years for
jointly under the assumption of correlation (Equa- firm disruption. The difference in the distribution of
tion (3)). Table 4 shows that ignoring unobserved nodes reflects the different distributions of disrup-
heterogeneity results in biased and inconsistent esti- tions for technologies and firms.
mates. Unobserved heterogeneity for both technol- Based on extant theory, H4 predicts that the haz-
ogy disruption ( T and firm disruption ( F2 are
2
ard of disruption is higher from an entrant than from
significant (t = 2 7 and t = 15 2, respectively). Also, an incumbent. However, contrary to the theory and
the correlation between the unobserved heterogene- H4, entrants are less likely than incumbents to dis-
ity coefficients ( T F is statistically significant (t = 5 4). rupt (i.e., the sign of entrant is negative) for both
In addition, the maximized value of log-likelihood is technology disruption (t = −3) and firm disruption
much higher for the correlated hazard model. Thus, (t = −4 6). Consistent with this result, we find that
we only discuss the results of the joint model. incumbents more often than entrants cause technol-
The baseline hazards are specified as splines. To ogy disruption (63% versus 57%) and firm disruption
identify the location of splines, we used the follow- (29% versus 16%) more frequently than entrants. Our
ing procedure. First, we estimate the hazard model result contrasts dramatically with Christensen’s claim
with only an intercept and a linear log-hazard, i.e., a (1997, p. 24) that “the firms that led the industry in
spline without nodes. This run provides us with esti- every instance of developing and adopting disruptive
mates of an intercept and a slope. We then specify two technologies were entrants to the industry, not its
or three nodes, spread out roughly evenly over the incumbent leaders.”
years, to approximate the occurrences of disruption in Based on extant theory, H5 predicts that the hazard
our sample. If the slopes of any two adjacent splines of firm disruption is higher if a new technology enters
are not significantly different, then we combine them via a lower attack. However, contrary to the theory
into one spline in the interests of parsimony. For the and H5, a lower attack significantly lowers the hazard
baseline hazard, we select nodes at 5, 15, and 25 years of firm disruption (t = −3). Because technologies


Table 4 Results of Hazard Model on Disruption

Uncorrelated hazards Correlated hazards

Technology Firm Technology Firm
disruption disruption disruption disruption

Parameter Est. (t-value) Est. (t-value) Est. (t-value) Est. (t-value)

Technology disruption spline
Spline: 0–5 years −0 77 (−5.9) −0 74 (−26.7)
Spline: 5–15 years 0 52 (8.9) 0 2 (8)
Spline: 15–25 years 0 51 (9.0) −0 34 (−12)
Spline: >25 years −0 81 (−6.1) −1 75 (−2.7)
Firm disruption spline
Spline: 0–8 years 0 38 (5.8) 0 47 (5.5)
Spline: 8–28 years −0 06 (−3.7) −1 88 (−19.3)
Spline: >28 years −0 22 (−8.2) −2 59 (−2.1)
Intercept 0 89 (3.9) 0 89 (0.6) 0 69 (0.3) 0 84 (1.2)
Entrants E −5 72 (−9.2) −3 22 (−4.4) −3 64 (−3) −3 15 (−4.6)
Lower attack L NA −2 00 (−2.2) NA −2 66 (−3)
Small firm S 1 7 (1.1) −1 56 (−1.9) 0 34 (0.7) −1 57 (−2.5)
Low priced C 1 22 (5.1) 32 (4.8) 0 11 (2.4) 0 41 (6.4)
Order of entry O 0 004 (8.3) 1 94 (2.5) 0 27 (2.4) 0 88 (1.7)
Performance improvement P 0 92 (9.4) 1 33 (5.0) 1 14 (4.7) 1 52 (6.9)
2 2 2 2
Heterogeneity T 0.9 (7.8) F 1.70 (2.5) T 2.36 (2.7) F 1.92 (15.2)
Correlation TF 0.15 (5.4)
Log-likelihood −799.9 −7,595.2 −7,954.2

entering via a lower attack are equivalent to “poten- Out-of-Sample Prediction of Disruption
tially disruptive technologies,” and only six technolo- Following Golder and Tellis (1997) and Sood et al.
gies in our sample disrupt using a lower attack, the (2009), we use a jackknife approach to ascertain the
absolute frequencies suggest that potentially disrup- out-of-sample predictive validity of the hazard model
tive technologies rarely cause firm disruption. in (Equations (1) and (2)) as follows. We reestimate the
Based on extant theory, H6A predicts that the haz- model n times, each time excluding one target tech-
ard of disruption is higher if a new technology is nology, where n is the number of technologies in our
introduced by a small firm, whereas H6B predicts the sample. We carry out this analysis by iteratively rees-
reverse. We find that firms’ size only affects the haz- timating this model in aML using a batch mode in
ard of firm disruption. Small firms do not increase the DOS. For each of these n runs, we multiply the esti-
hazard of technology disruption (t = 0 7) but lower mated parameters of the model with the values of the
variables of the excluded target technology (in Excel)
the hazard of firm disruption (t = −2 5).
to predict the hazard of disruption for the excluded
Based on extant theory, H7 predicts that the haz-
target technology. We compare the predicted value
ard of disruption is higher if a new technology is
priced lower than the dominant technology at entry.
The results support the hypothesis. Relative price of Figure 3 Baseline Hazards: Technology and Firm Disruption
the new technology at entry relative to the dominant 1.0
technology increases the hazard of both technology 0.9 Baseline technology disruption
disruption (t = 2 4) and firm disruption (t = 6 4). Baseline firm disruption
0.8
The hazard of technology disruption increases with
0.7
both an increase in performance (t = 4 7) and in the
order of entry (t = 2 4). However, only an increase 0.6

in performance affects the hazard of firm disruption 0.5
(t = 6 9), but the order of entry has no impact (t = 1 7). 0.4
Figure 3 plots the baseline hazard for both technol- 0.3
ogy and firm disruption for a new technology. Both 0.2
hazards peak early and decline subsequently but fol-
0.1
low somewhat different paths. Firm disruption lags
0.0
technology disruption by approximately 10 years in 0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30
our sample. Years after introduction


with a cutoff point (as explained under the Predic- Table 5 Out-of-Sample Predictive Accuracy
tive Statistics section) to predict a disruption. Based

Technology Firm
on this approach, we make two types of predictions, At entry disruption disruption
one at entry and the other, one year later, updated
with the most recent prior-year information. The dif- Specificity (%) 75 82
Sensitivity (%) 80 75
ference in the two approaches lies in the difference in Updated forecasts
the information used to estimate the models, either at Specificity (%) 72 82
entry or when including each additional year of the Sensitivity (%) 65 52
subsequent evolution of the target technology. In the Predictive accuracy
latter case, if the jth technology has mj years, then Mean absolute error 0.22 (0.42) 0.19 (0.40)
in prediction 1a (SEb
there will be a total of 36 mj predictions. See Online
j=1 Mean absolute error 1.9 (6.1) years 1.8 (5.1) years
Appendix D in the electronic companion for more in prediction 2c (SEb
details. a
We compute the error in prediction as difference in the ability to predict
In total, there were 72 iterations for predictions at disruption (1) or not (0).
the time of entry and 1,969 predictions for updated b
We compute standard error as SE = Y − Y 2 / N − 1 , where
forecasts. We show the predictive accuracy of the haz- Y − Y is the error in prediction and N is the number of predictions.
c
ard model in three ways: predictive statistics, graphi- We compute the error in prediction as the difference in predicted year of
disruption and actual year of disruption.
cal comparison of actual versus predicted disruptions,
and error in the prediction of disruption.
Predictive Statistics. Traditional summary statistics and firm disruptions, respectively. Note that for both
of the accuracy of the model in predicting disruption graphs, the model predicts the disruptions reasonably
are specificity and sensitivity. Specificity and Sensitivity well, even at the time of introduction.
are the power of the model to detect true negatives Error in Prediction of Disruption. We calculate this
and true positives, respectively, computed as follows: error in two ways: First, the error in correctly pre-
dicting the occurrence of disruption (1) versus no dis-
True Negatives ruption (0). Second, the difference in years between
Specificity =
Actual Negatives when the model predicts a disruption and when the
True Negatives disruption actually takes place. For the first approach,
= (6) the mean error is 0.22 for technology disruption and
True Negatives + False Positives
0.19 for firm disruption, respectively. For the second
True Positives approach, the mean errors range from 1.9 years for
Sensitivity =
Actual Positives technology disruption to 1.8 years for firm disrup-
True Positives tion, respectively (see Table 5). Although these figures
= (7)
True Positives + False Negatives may seem large, recall that these events occur rarely
in the life of a technology that spans decades and
The false-positive rate and the false-negative rate that as of now the literature has no model whatsoever
are simply (1 − Specificity) and (1 − Sensitivity), respec- that can predict disruption, especially so many years
tively. The determination of a disruption is made by ahead of the event. Also, these error rates compare
the analyst when the predicted value falls below a well with past studies using this method (Golder and
cutoff value. Choosing too low a cutoff leads to low Tellis 1997).
false positives but high false negatives. The reverse is
true for choosing too high a cutoff, so we choose a Patterns of Disruption
cutoff that balances the two error rates. We find some patterns in the two types of disruption
Table 5 presents the results for each domain of dis- that are noteworthy.
ruption. Note that for prediction of both technology First, at many points in time, competing technolo-
and firm disruption at the time of entry, the out-of- gies coexist. In some cases, disrupted technologies
sample sensitivity and specificity are both high. For continue to survive and coexist with the new technol-
predicting disruption one year ahead, specificity is ogy by finding a niche. For example, impact printers
high for both technology and firm disruption and continue to coexist with laser and inkjet printing tech-
sensitivity is high for firm disruption. The only pre- nologies. This suggests that the phenomenon is not as
diction that is not good is that of sensitivity of firm “fatal” or “final” as the term implies. It is true that
disruption for updated forecasts. some technologies do die, but many continue to sur-
Graphical Comparison of Actual vs. Predicted vive even after being disrupted.
Disruptions. Figure 4 compares the actual disrup- Second, some technologies experience disruption in
tion at entry with that predicted by the models. Fig- one domain but not in another domain. For example,
ures 4(a) and 4(b) display the results for technology in the lighting market, incandescence continues its


Figure 4 Predictive Ability of Hazard Model 20 years propelled gas discharge into a position of
superiority again.

(a) Technology disruption
Fourth, there is a fascinating dynamic of emer-
8
gence of new secondary dimensions of performance.
7 We find that a new technology almost always intro-
6 duces a new dimension of importance even while
Number of disruptions

competing with old technologies on the primary
5
dimension (see Table 6, panel a). For example, in
4 display monitors, LCDs introduced the dimension of
Actual compactness, plasma brought into focus the dimen-
3 Predicted
sion of screen size, and organic light-emitting diode
2 brought into play the dimensions of convenience and
1
low power consumption. These secondary dimen-
sions appeal to various niche segments. However, in
0 all cases, the competition for the mainstream segment
1 4 7 10 13 16 19 22 25 28 31 34
Years since entry was still on the primary dimension of performance
(e.g., resolution in desktop monitors), which contin-
(b) Firm disruption ued to improve substantially over time.
3 Finally, contrary to current belief, we observe mul-
tiple disruptions or crossings between paths of tech-
nological performance. This pattern occurs when
technology disruption by a new technology is not
Number of disruptions

2

Table 6 Patterns of Disruption

Market Secondary dimensions
1
(a) Secondary dimensions of competition
Electrical lighting Cleanliness/safety, brightness, life, size, modularity
Data transfer Mechanization, bandwidth, connectivity
Computer memory Mechanization, mutability, accessibility,
0 addressability, transfer speed, life, capacity
1 4 7 10 13 16 19 22 25 28 31 34
Years since entry Computer printers Mechanization, graphics quality, speed, simple
design
Display monitors Mechanization, compactness, screen size,
dominance in the demand domain for many decades brightness, flexibility, low power consumption
Music recording Play time, duplication, mutability, size, life
even though it was disrupted in the performance Analgesics Recovery speed, targeted action, risk-benefit balance
domain by higher-performing technologies. We also
observe that firms that introduce a new technology (b) Occurrence of disruption by time period
may not be the ones to cause disruption. In many Technology Firm
cases, other firms may subsequently promote the new disruption (%) disruption (%)
technology and cause disruption. For example, even
Time of No No
though Optel Inc. introduced the LCD technology, introduction disruption Disruption disruption Disruption
it was Samsung that disrupted the incumbents and
became the market leader. Hence, first-mover advan- Before 1960 28 22 41 9
tages are not sufficient for disruption. After 1960 17 33 36 14

Third, most technologies do not improve smoothly (c) Technology dynamics in printer marketa
over time (see Figures 2(a) to 2(d)) as the theory of
Characteristics of new technology at entry
disruptive innovations predicts; neither do most tech-
nologies improve in the shape of S-curves (Foster Printer technology Lower attack Entrant Small firm Low-priced
1986). Rather, improvement is sporadic, with many Impact Yes No No No
periods of no improvement followed by spurts of Pen plotter No Yes Yes No
big improvements. For example, gas discharge was Laser Yes No No No
stagnant for many years and lost technological supe- Inkjet Yes No No No
riority to a competing technology, arc discharge, Thermal Yes Yes No No
which improved frequently every few years after its Note. Percentage of all technologies: 36.
a
entry. However, substantial improvement after almost Using resolution per dollar.


permanent, because a technology that has been Figure 5 Technology Dynamics for Desktop Printers (When
surpassed in performance regains technological lead- Performance Is Measured as Resolution per Dollar)

ership. We find a total of four cases of multiple tech- 1.E+ 00
Impact
nology disruptions: two in computer memory and InkJet
two in electrical lighting. Thus disruption is not per- 1.E–01 Laser
manent as extant theory suggests. At the same time, Thermal
we do not find cases of multiple firm or demand dis- 1.E–02

Resolution/$
ruption so far in our sample.
1.E–03
Robustness of Results
We carried out five analyses to assess the robust- 1.E–04
ness of our results. First, one would be concerned
that the theoretical relatedness in some of our inde- 1.E–05
pendent variables may create a problem of multi-
collinearity. However, we find that the results of the 1.E–06

1978

1980

1982

1984

1986

1988

1990

1992

1994

1996

1998

2000
2002
hazard model are robust to the selection of variables.
In particular, the significance and effect of each vari-
able does not change much, whether each variable results are in Figure 5 and Table 6, panel c. The results
is included individually or combined with all oth- are consistent with the original analyses for all mar-
ers. The correlation between the two key variables of kets using absolute performance. Four technologies
interest—incumbency and type of attack—is low (0.1). use a lower attack, but only one of these (pen plot-
Estimates of variance inflation factors, using a mul- ter) is from an entrant; the rest are from incumbents.
tiple regression model with the same data and vari- The one technology using an upper attack is from an
ables, suggest variance inflation factor values of less entrant. All five new technologies are more expensive
than 2. Thus, multicollinearity is not a problem in our than the dominant technology at the time of entry.
data (Hair et al. 2006). There is only one disruption: inkjet disrupts impact
Second, one could argue that the frequency of printers in 1987. Inkjet was introduced by IBM, which
occurrence reported in our results could suffer from was an incumbent. Thus, in this market, using reso-
censoring bias, because not enough time has elapsed lution per dollar, the pattern of results is very similar
for disruption to occur. To assess the severity of to that in other markets using absolute performance.
this problem, we do a split-sample analysis, divid- Fifth, we tested many interactions in the model.
ing our sample by a median split on the year of However, the correlated hazards model fails to con-
entry. This yields two sets of technologies—one intro- verge when these interaction terms are added to the
duced before 1960 and the other after 1960, each model, probably because of few events per interaction
with 18 technologies. Note that, in general, disrup- term. So we chose to retain and test only the vari-
tion occurs more frequently in the sample after 1960 ables directly suggested by the theory of disruptive
than in that before 1960 (see Table 6, panel b). How- innovations.
ever, even in the post-1960 sample, which allows for
a time period of at least 40 years, the occurrence of Discussion
disruption is not high, contrary to the dire warnings Although making strong claims that are quite popu-
of extant theories. lar, the theory of disruptive innovations lacks precise
Third, we also tested the impact of two more vari- definitions, suffers from tautologies, lacks adequate
ables in the hazard model—change in performance of empirical testing, and has no predictive model. We
dominant technology and difference in the change in attempt to remedy these problems with a new schema,
performance of the two competing technologies. Both new empirical data, and a new predictive model.
these variables could affect the hazard of disruption The proposed schema has clear definitions and dis-
by the new technology for the following reasons. First, tinguishes between types of technologies, types of
disruption may become easier as the dominant tech- attacks, and domains of disruption. The schema allows
nology matures and improves slowly (Foster 1986). us to derive seven testable hypotheses. We test these
Second, difference in the performance of the two hypotheses with a hazard model on data from all
competing technologies may increase the hazard of 36 technologies in seven different markets. Further,
disruption of the dominant. We added these vari- we carry out an out-of-sample predictive analysis that
ables in the hazard model to test these expectations. shows good to high sensitivity and specificity. The test
The results were not materially different from those and results apply to platform technologies. This sec-
reported here. tion summarizes the findings from this test, discusses
Fourth, we redo the analysis for one industry using implications, and points out some limitations of the
resolution per dollar rather than only resolution. The research.


Summary of Findings for many disruptions, often without the expertise,
Contrary to extant theory, market knowledge, or resources of the incumbents,

1. Technologies that adopt a lower attack or is quite impressive. A key issue is why some incum-
“potentially disruptive technologies” bents fail whereas others succeed. We suspect that the
• are introduced as frequently by incumbents as internal culture of the firms is probably a key factor
by entrants, responsible for disruption, rather than any external
• are not cheaper than old technologies, and threat per se, such as a new technology or strategy
• rarely disrupt firms. (Tellis et al. 2009).
2. The hazard of disruption by incumbents is sig-
nificantly higher than that by entrants.
3. Lower attack reduces the hazard of firm Limitations
disruption. We acknowledge some limitations of the study, which
However, consistent with extant theory, could be the basis of future research. First, because
• low price of new technologies increases the haz- of the time-consuming nature of data collection, we
ard of disruption. However, most new technologies, were able to analyze only seven markets. However,
unfortunately, are not priced lower than dominant that number still yields 36 technologies, which we
technologies at entry. track for an average of 50 years. This is probably
a more comprehensive sampling than prior research
in the field. Second, we were not able to get data
Implication on the performance per dollar of all technologies for
These results suggest that many aspects of the theory all years. A number of authors emphasize the need
of disruption are exaggerated, if not inaccurate. They to incorporate such metrics for a richer analysis of
raise one big question: Is the theory totally wrong? performance. Third, our results apply to platform
Not so. The theory is right in one aspect: the hazard of technologies because we do not test the disruptive
disruption by low-priced new technologies is higher. potential of design and component innovations, prod-
Although entrants with lower attacks do cause dis- uct innovations, or business model innovations due to
ruption, this event has been exaggerated. Although both limitations of data and the large number of such
an entrant disrupting a well-funded, giant incumbent innovations. However, each of these levels of inno-
with a lower attack always makes for a good story, vations may also have disruptive potential. Fourth,
such disruptions account for only a small fraction of because of the extensive technological and historical
all cases. For example, only 8% of all technology dis- focus of this study, we did not obtain behavioral and
ruptions and 25% of all firm disruptions were caused cultural aspects of the firms involved in technology
by entrants using a lower attack.
competition. We suspect that these may be impor-
The term “disruptive technology” has been at-
tant predictors of firm disruption. Fifth, our results
tributed to technologies entering via a lower attack.
may be susceptible to censoring. However, even when
By our results, the frequency of the latter event has
given over 40 years of time, the occurrence of dis-
been exaggerated, and so-called “disruptive technolo-
ruption never came close to the values claimed by
gies” rarely disrupt. For example, although 47% of
extant theories. Sixth, we did not encounter any cases
all technologies adopt a lower attack, only 16% of
of an incumbent acquiring a potentially disruptive
all technologies cause technology disruption and only
technology before a disruption occurred. Neverthe-
14% of all technologies cause firm disruption via a
less, this could be a viable strategy and needs to be
lower attack. However, the threat of lower attacks
studied.
should not be completely discounted. Lower attacks
are important because managers of incumbent firms
may tend to ignore or belittle a new technology that Electronic Companion
initially seems inferior to the dominant technology. An electronic companion to this paper is available as
Some of these new technologies can improve enough part of the online version that can be found at http://
to disrupt the initially superior technology. mktsci.pubs.informs.org/.
Incumbents may take hope from our results in that
incumbents cause 50% of all technology disruptions
Acknowledgments
and 62% of all firm disruptions. However, in all mar- The authors are grateful to the insightful comments of the
kets, even though incumbents introduced more tech- editor, anonymous reviewers, the support of the Market-
nologies and caused more disruption than entrants, ing Science Institute, Michael Parzen, Paul Allison, and the
many incumbents lost market dominance and subse- research assistance of Esra Kent, Vivek Pundir, and K. L.
quently failed. Hence, there is no room for compla- Dang. This study benefited from a grant from Don Murray
cency. Entrants do disrupt, and for entrants to account to the USC Center for Global Innovation.


References Grove, A. 1998. Keynote. Presentation, Annual Meeting of the
Academy of Management, August 9, Academy of Manage-

Acs, Z. J., D. B. Audretsch. 1988. Innovation in large and small

firms: An empirical analysis. Amer. Econom. Rev. 78(4) 678–690. ment, Briarcliff Manor, NY.
Hair, J. F., T. R. L. Anderson, W. C. Black. 2006. Multivariate Data
Adner, R. 2002. When are technologies disruptive? A demand-
Analysis. Prentice Hall, Englewood Cliffs, NJ.
based view of the emergence of competition. Strategic Manage-
ment J. 23(8) 667–688. Henderson, R. M. 1993. Underinvestment and incompetence as
responses to radical innovation: Evidence from the photolitho-
Adner, R., P. Zemsky. 2005. Disruptive technologies and the emer- graphic alignment equipment industry. RAND J. Econom. 24(2)
gence of competition. RAND J. Econom. 36(2) 229–254. 248–270.
Allison, P. D. 1995. Survival Analysis Using SAS® : A Practical Guide. Henderson, R. M. 2006. The innovator’s dilemma as a problem of
SAS, Cary, NC. organizational competence. J. Product Innovation Management
Barney, J. B. 1997. On flipping coins and making technology 23(1) 5–11.
choices: Luck as an explanation of technological foresight and Henderson, R. M., K. B. Clark. 1990. Architectural innovation: The
oversight. R. Garud, P. R. Nayyar, Z. B. Shapira, eds. Technolog- reconfiguration of existing product technologies and the failure
ical Innovation: Oversights and Foresights. Cambridge University of established firms. Admin. Sci. Quart. 35(1) 9–30.
Press, New York, 13–19. Hill, C. W. L., F. T. Rothaermel. 2003. The performance of incum-
Bower, J. L., C. M. Christensen. 1995. Disruptive technologies: bent firms in the face of radical technological innovation. Acad.
Catching the wave. Harvard Bus. Rev. 73(1) 43–53. Management Rev. 28(2) 257–274.
Chandy, R. K., G. J. Tellis. 1998. Organizing for radical product King, A. A., C. L. Tucci. 2002. Incumbent entry into new market
innovation. J. Marketing Res. 35(4) 474–487. niches: The role of experience and managerial choice in the
Chandy, R. K., G. J. Tellis. 2000. The incumbent’s curse? Incum- creation of dynamic capabilities. Management Sci. 48(2) 171–186.
bency, size, and radical product innovation. J. Marketing 64(3) Kostoff, R. N., R. Boylan, G. R. Simons. 2004. Disruptive technology
1–17. roadmaps. Tech. Forecasting Soc. Change 71(1) 141–159.
Christensen, C. M. 1997. The Innovator’s Dilemma: When New Tech- Levinthal, D. A. 1998. The slow pace of rapid technological change:
nologies Cause Great Firms to Fail. Harvard Business School Gradualism and punctuation in technological change. Indust.
Press, Boston. Corporate Change 7(2) 217–247.
Lieberman, M. B., D. B. Montgomery. 1988. First-mover advantages.
Christensen, C. M., M. Raynor. 2003. The Innovator’s Solution: Creat-
Strategic Management J. 9(S1) 41–58.
ing and Sustaining Successful Growth. Harvard Business School
Press, Boston. Lillard, L. A. 1993. Simultaneous equations for hazards: Marriage
duration and fertility timing. J. Econometrics 56(1/2) 189–217.
Christensen, C. M., R. S. Rosenbloom. 1995. Explaining the
Lillard, L. A., C. W. A. Panis. 2003. aML Version 2 User’s Guide
attacker’s advantage: Technological paradigms, organizational
and Reference Manual. EconWare, Los Angeles. http://www
dynamics, and value network. Res. Policy 24(2) 233–257.
.applied-ml.com/download/amldoc.pdf.
Christensen, C. M., S. Anthony, E. Roth. 2004. Seeing What’s Next: Markides, C. 2006. Disruptive innovation: In need of better theory.
Using Theories of Innovation to Predict Industry Change. Harvard J. Product Innovation Management 23(1) 19–25.
Business School Press, Boston.
Mokyr, J. 1990. Punctuated equilibria and technological progress.
Cohan, P. S. 2000. The dilemma of the “innovator’s dilemma”: Amer. Econom. Rev. 80(2) 350–355.
Clayton Christensen’s management theories are suddenly all Pavitt, K., S. Wald. 1971. The Conditions for Success in Technolog-
the rage, but are they ripe for disruption? Indust. Stan- ical Innovation. Organisation for Economic Co-operation and
dard (January 10) http://www.computerworld.com.au/article/ Development, France.
63941/dilemma_innovator_dilemma_/.
Raffi, F., P. J. Kampas. 2002. How to identify your enemies before
Cohen, W. M., D. A. Levinthal. 1990. Absorptive capacity: A new they destroy you. Harvard Bus. Rev. 80(11) 115–124.
perspective on learning and innovation. Admin. Sci. Quart. Robertson, T. S., J. Eliashberg, T. Rymon. 1995. New product
35(1) 128–152. announcement signals and incumbent reactions. J. Marketing
Danneels, E. 2004. Disruptive technology reconsidered: A critique 59(3) 1–15.
and research agenda. J. Product Innovation Management 21(4) Rogers, E. M. 2003. Diffusion of Innovations, 5th ed. Free Press,
246–258. New York.
Dosi, G. 1982. Technological paradigms and technological trajecto- Rothaermel, F. T. 2001. Complementary assets, strategic alliances,
ries: A suggested interpretation of the determinants and direc- and the incumbent’s advantage: An empirical study of indus-
tions of technical change. Res. Policy 11(3) 147–162. try and firm effects in the biopharmaceutical industry. Res. Pol-
Foster, R. N. 1986. Innovation: The Attacker’s Advantage. Summit icy 30(8) 1235–1251.
Books, New York. Rothaermel, F. T., C. W. L. Hill. 2005. Technological discontinuities
Freeman, C. 1974. The Economics of Industrial Innovation. Penguin, and complementary assets: A longitudinal study of industry
Harmondsworth, UK. and firm performance. Organ. Sci. 16(1) 52–70.
Ghemawat, P. 1991. Market incumbency and technological inertia. Schumpeter, J. A. 1934. The Theory of Economic Development. Harvard
Marketing Sci. 10(2) 161–171. University Press, Cambridge, MA. [Reprint, Oxford University
Press, New York, 1961.]
Gilbert, C. 2003. The disruption opportunity. Sloan Management Rev.
Schweidel, D. A., P. S. Fader, E. T. Bradlow. 2008. A bivariate tim-
44(4) 27–32.
ing model of customer acquisition and retention. Marketing Sci.
Gilbert, C., J. L. Bower. 2002. Disruptive change: When trying 27(5) 829–843.
harder is part of the problem. Harvard Bus. Rev. 80(May) Sood, A., G. J. Tellis. 2005. Technological evolution and radical
94–101. innovation. J. Marketing 69(3) 152–168.
Golder, P. N., G. J. Tellis. 1997. Will it ever fly? Modeling the growth Sood, A., G. J. Tellis. 2009. Do innovations really pay off? Total stock
of new consumer durables. Marketing Sci. 16(3) 256–270. market returns to innovation. Marketing Sci. 28(3) 442–456.
Gould, S., N. Eldridge. 1977. Punctuated equilibria: The tempo and Sood, A., G. M. James, G. J. Tellis. 2009. Functional regression:
mode of evolution reconsidered. Paleobiology 3(2) 115–151. A new model for predicting market penetration of new prod-
Govindarajan, V., P. K. Kopalle. 2006. The usefulness of measur- ucts. Marketing Sci. 28(1) 36–51.
ing disruptiveness of innovations ex-post in making ex ante Tellis, G. J. 2006. Disruptive technology or visionary leadership?
predictions. J. Product Innovation Management 23(1) 12–18. J. Product Innovation Management 23(1) 34–38.


Tellis, G. J., J. C. Prabhu, R. K. Chandy. 2009. Radical innovation Tushman, M. L., P. Anderson. 1986. Technological discontinu-
across nations: The preeminence of corporate culture. J. Mar- ities and organizational environments. Admin. Sci. Quart. 31(3)

keting 73(1) 3–23. 604–633.
Tripsas, M. 1997. Unraveling the process of creative destruction: Utterback, J. M. 1994. Radical innovation and corporate regenera-
Complementary assets and incumbent survival in the typeset- tion. Res. Tech. Management 37(4) 10–18.
ter industry. Strategic Management J. 18(Summer special issue) Utterback, J. M., H. J. Acee. 2005. Disruptive technologies: An
119–142. expanded view. Internat. J. Innovation Management 9(1) 1–18.

Demystifying Disruption: A New Model for Understanding and Predicting Disruptive Technologies

More Related Content

Similar to Demystifying Disruption: A New Model for Understanding and Predicting Disruptive Technologies (20)

Recently uploaded (20)

Demystifying Disruption: A New Model for Understanding and Predicting Disruptive Technologies