University of Southern California Dissertations and Theses

Essays on technological evolution and financial returns to innovation

(USC Thesis Other)

Essays on technological evolution and financial returns to innovation

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content

ESSAYS ON TECHNOLOGICAL EVOLUTION AND
FINANCIAL RETURNS TO INNOV ATION

By

Ashish Sood

A Dissertation Presented to the
FACULTY OF THE GRADUATE SCHOOL
UNIVERSITY OF SOUTHERN CALIFORNIA
In Partial Fulfillment of the
Requirements for the Degree
DOCTOR OF PHILOSOPHY
(BUSINESS ADMINISTRATION)

December 2005
Copyright 2005 Ashish Sood
UMI Number: 3222046
3222046
2006
UMI Microform
Copyright
All rights reserved. This microform edition is protected against
unauthorized copying under Title 17, United States Code.
ProQuest Information and Learning Company
300 North Zeeb Road
P.O. Box 1346
Ann Arbor, MI 48106-1346
by ProQuest Information and Learning Company.

ii
ACKNOWLEDGEMENTS
This is my opportunity to express my heartfelt gratitude to all the people who
helped me to reach this stage.
My friend and wife Amita Sood was the one who believed that I could do this
even before I did. She supported me all these years by providing encouragement and
love and by pitching in with much more than a fair share of the load of household
responsibilities while I sat staring into the computer screen for long hours. Thank
you very much, my love – I remember I said we will have a lot of time after the PhD
– but there is this small issue of tenure…
My advisor and guide Gerry Tellis was the constant source of inspiration that
I was fortunate enough to latch on early in the PhD journey. We have developed a
wonderful relationship with him in the last four years. From teaching me how to
conduct meaningful research to how to play the guitar, he was a mentor to me in
more way than one. I would like to emulate the fine balance of work and play he
fills his days with.
My dissertation committee members – Gareth James, Siddarth, Micah Officer
and Geert Ridder were always available with valuable support and insights. I am
deeply grateful to all of them for their help and guidance in raising the level of my
research and providing me with new ideas during the discussions.
The research support of the Marketing Science Institute and the Center for
Research in Technology and Innovation injected critical support. The financial
assistance enabled the collection of data for the research projects.

iii
The folks of the Marketing Department – Valerie Folkes, Shantanu Dutta,
Allen Weiss, Debbie MacInnis, Gary Frazier, C. Whan Park, David Stewart and
Joseph Nunes were always forthcoming with ready support whenever I approached
them in the last 4 years. Moreover, Carolyn, Elizabeth and Sandra always helped
smooth out any last minute problems I had.
I owe a special thanks to Rakesh Niraj for all the moral and financial support
he gave at critical times. The long hours he spent in pushing me to learn
programming skills made the projects move much faster.
My parents Rita Sood and Raman Sood played a crucial role in instilling in
me the qualities of perseverance and risk taking by setting an example. My late aunt
Punita Sood and me late grandparents prepared me for this monumental task of
finishing a doctorate by teaching me the value of education as a child. I am grateful
to my parents for believing in me and their understanding why I could not be there
with them the last four years during many times when they needed me. I am grateful
to all of them.

iv
TABLE OF CONTENTS

ACKNOWLEDGEMENTS .........................................................................................ii
LIST OF TABLES ......................................................................................................ix
LIST OF FIGURES .....................................................................................................x
LIST OF APPENDICES.............................................................................................xi
ABSTRACT...............................................................................................................xii

I CHAPTER 1: INTRODUCTION AND OUTLINE ............................................1
I.1 MOTIVATION............................................................................................1
I.2 OUTLINE....................................................................................................2
I.3 OBJECTIVE................................................................................................2

II CHAPTER 2: TECHNOLOGICAL EVOLUTION AND RADICAL
INNOVATION ............................................................................................................5

II.1 ABSTRACT.................................................................................................5

II.2 INTRODUCTION.......................................................................................6

II.3 PREVAILING THEORY.............................................................................8

II.4 DEFINITIONS.............................................................................................8

II.5 THEORETICAL BACKGROUND.............................................................9
II.5.1 Introduction Stage..............................................................................10
II.5.2 Growth Stage......................................................................................11
II.5.3 Maturity Stage....................................................................................11

II.6 HYPOTHESES..........................................................................................13
II.6.1 Shape of technological progress.........................................................13
II.6.2 Technological Transition and Performance of Competing
Technologies ......................................................................................................14
II.6.3 Dimensions of Technological Competition .......................................16
II.6.4 Pace of technological transition .........................................................17
II.6.5 Source of new technologies ...............................................................18

II.7 METHOD...................................................................................................19
II.7.1 Sample Selection................................................................................19
II.7.2 Sources...............................................................................................20
II.7.3 Procedure...........................................................................................21

v
II.8 RESULTS..................................................................................................22
II.8.1 Identification of platform innovations and performance attributes....23
II.8.2 Shape of technological evolution.......................................................24
II.8.3 Performance of competing technologies............................................30
II.8.4 Progress on secondary dimensions ....................................................32
II.8.5 Pace of technological transition .........................................................34
II.8.6 Towards a new predictive model .......................................................37

II.9 DISCUSSION............................................................................................39
II.9.1 Summary of findings..........................................................................39
II.9.2 Questions and further analysis ...........................................................40
II.9.3 Survival bias.......................................................................................42
II.9.4 Performance per unit price.................................................................44
II.9.5 Multi-attribute performance...............................................................46

II.10 IMPLICATIONS.......................................................................................47

II.11 LIMITATIONS AND DISCUSSIONS FOR FUTURE RESEARCH ......48

III CHAPTER 3: PERFORMING HYPOTHESIS TESTS ON THE
SHAPE OF FUNCTIONAL DATA ..............................................................50

III.1 ABSTRACT...............................................................................................50

III.2 INTRODUCTION.....................................................................................51

III.3 METHODOLOGY.....................................................................................55
III.3.1 Functional iid Tests of Shape (FITS).................................................56
III.3.2 Functional arbitrary covariance tests of shape: multiple curves
(FACTSN)..........................................................................................................61
III.3.3 Functional arbitrary covariance tests of shape: (FACTS
1
) ................65

III.4 SIMULATION STUDY............................................................................69

III.5 EMPIRICAL STUDY................................................................................76

III.6 DISCUSSION............................................................................................80

IV CHAPTER 4: MARKET RETURNS TO TECHNOLOGICAL
INNOVATIONS DURING NEW PRODUCT DEVELOPMENT ...........................82

IV.1 ABSTRACT...............................................................................................82

vi

IV.2 INTRODUCTION.....................................................................................83

IV.3 THEORY...................................................................................................84
IV.3.1 Evolution of Technologies .................................................................85
IV.3.2 Definitions..........................................................................................85
IV.3.3 Market Returns to Technological Innovations...................................86
IV.3.4 Hypotheses.........................................................................................87
IV.3.5 Technological Leader or Close Follower...........................................88
IV.3.6 Investment Portfolio...........................................................................90
IV.3.7 Research Productivity........................................................................91
IV.3.8 Size of Firms ......................................................................................92
IV.3.9 Frequency of Announcements ...........................................................93
IV.3.10 Age of Technology.........................................................................94
IV.3.11 Effect of Innovation on Competitors .............................................94

IV.4 METHOD...................................................................................................95
IV.4.1 Sample selection................................................................................96
IV.4.2 Sources...............................................................................................97
IV.4.3 Procedure...........................................................................................98
IV.4.4 Model.................................................................................................99

IV.5 RESULTS................................................................................................102
IV.5.1 Abnormal Returns............................................................................102
IV.5.2 Technological Leadership................................................................103
IV.5.3 Investment Portfolio.........................................................................105
IV.5.4 Research productivity.......................................................................108
IV.5.5 Size of Firm......................................................................................111
IV.5.6 Frequency of announcements...........................................................112
IV.5.7 Age of Technology...........................................................................114
IV.5.8 Effect of Innovation on Competitors ...............................................115
IV.5.9 Market Response to being First to Announce Innovations ..............116

IV.6 DISCUSSION..........................................................................................117
IV.6.1 Summary of findings........................................................................117
IV.6.2 Profile of Firms with High Returns..................................................118
IV.6.3 Robustness of Results......................................................................119
IV.6.4 Survivor Bias....................................................................................122
IV.6.5 Implications and Contributions to Practice......................................123

IV.7 LIMITATIONS AND DISCUSSION FOR FUTURE RESEARCH ......123

V CHAPTER 5: TOTAL MARKET RETURNS TO NEW PRODUCTS..........125

vii

V.1 ABSTRACT.............................................................................................125

V.2 INTRODUCTION...................................................................................125

V.3 THEORY.................................................................................................127
V.3.1 Definitions........................................................................................127
V.3.2 Market Returns to Innovation ..........................................................127

V.4 HYPOTHESES........................................................................................129
V.4.1 Return to Innovation in each Phase of Innovation Project...............129
V.4.2 Long Term Abnormal Returns to Innovation...................................132

V.5 METHOD.................................................................................................133
V.5.1 Model...............................................................................................133
V.5.2 Absence of a clear estimation period ...............................................134
V.5.3 Clustering of Announcements..........................................................135
V.5.4 Inaccuracy of estimated abnormal returns over a long period .........135
V.5.5 Long-Horizon Event Studies............................................................135
V.5.6 Calendar-time portfolio approach (Jensen-alpha approach) ............136
V.5.7 Procedure.........................................................................................138

V.6 RESULTS................................................................................................139
V.6.1 Average Market Returns ..................................................................139
V.6.2 Initiation Phase.................................................................................140
V.6.3 Innovation Phase..............................................................................141
V.6.4 Commercialization Phase.................................................................143
V.6.5 Announcements with positive abnormal returns..............................145
V.6.6 Long term abnormal returns to innovation.......................................146

V.7 DISCUSSION..........................................................................................148
V.7.1 Summary of findings........................................................................148
V.7.2 Degree of Underestimation ..............................................................148
V.7.3 Implications and Contributions to Practice......................................150

V.8 LIMITATIONS AND DISCUSSION FOR FUTURE RESEARCH ......151

VI References....................................................................................................153

VII Appendix A: Operating Principles of Sampled Technologies .....................162
VII.1.1 External lighting...............................................................................162
VII.1.2 Desktop Memory..............................................................................162
VII.1.3 Display Monitors..............................................................................162

viii
VII.1.4 Desktop Printers...............................................................................163
VII.1.5 Data Transfer....................................................................................163
VII.1.6 Analgesics........................................................................................163

VIII Appendix B: Classification of Announcements...........................................164
VIII.1 Announcements during Initiation Phase ..............................................164
VIII.2 Announcements during Innovation Phase............................................164
VIII.3 Announcements during Commercialization Phase ..............................164

ix
LIST OF TABLES
Table II-1: Primary Dimensions of Competition .....................................................24
Table II-2a: Test of Logistic Fit for Technologies with Single S-curves ..................28
Table II-3: Performance of new technology relative to old .....................................31
Table II-4: Number of crossings between new and old technologies ......................32
Table II-5: Emergence of Secondary Dimensions of Competition..........................34
Table II-6: Modeling Probability and Size of Jump in Performance.......................39
Table III-1: Average minimum equivalent degrees of freedom for γ ........................78
Table IV-1: Data Sample ...........................................................................................96
Table IV-2: Mean Abnormal return using OLS Market Model...............................102
Table IV-3: Mean Abnormal Returns to Technological Leadership
vs. Close Followers .............................................................................104
Table IV-4: Mean Abnormal Returns to firms with Large vs.
Small Investment Portfolio .................................................................106
Table IV-5: Regression results for effect on Average Abnormal Return ................107
Table IV-6: Mean Abnormal Returns to firms with High vs. Low
Research Productivity .........................................................................109
Table IV-7: Mean Abnormal Returns to Small vs. Large Firms..............................111
Table IV-8: Mean Abnormal Returns to firms with High vs.
Low Frequency of Announcements ....................................................113
Table IV-9: Mean Abnormal Returns to firms with Old vs.
New Technologies...............................................................................114
Table IV-10: Market returns to Competitors ...........................................................116
Table V-1: Sample Data........................................................................................140
Table V-2: Mean abnormal return using Event Study approach...........................143
Table V-3: Mean abnormal return using Calendar Time Approach for
announcements in Innovation and Commercialization Phase.............147

x
LIST OF FIGURES
Figure 1: Technological Evolution ............................................................................12
Figure 2: Technological Evolution in six categories..................................................25
Figure 3: Technologies with Single S curve ..............................................................27
Figure 4: Technologies with Multiple S-curves.........................................................27
Figure 5: Pace of Technological Transition...............................................................35
Figure 6: Performance per unit price .........................................................................45
Figure 7: Multi-attribute performance .......................................................................46
Figure 8: Plots of the evolutions of three different technologies over time
along with the best fitting “S” curves (dashed lines). ....................................52
Figure 9: Plots of FACTS
N
methodology applied to three different
simulated data sets..........................................................................................64
Figure 10: Three different simulated data sets and their corresponding
p-value curves ................................................................................................69
Figure 11: Plots of power as a function of signal to noise ratio using the
FITS methodology. ........................................................................................71
Figure 12: Power plots using the FACTS
N
methodology. .........................................73
Figure 13: (a)–(d) Power plots using the FACTS
1
methodology...............................75
Figure 14: Plots for (a, b) incandescent lighting, (c, d) arc-discharge
lighting, (e, f) cathode ray tube display monitors and (g, h)
dot-matrix desktop printers. ...........................................................................79
Figure 15: Theoretical Model.....................................................................................88
Figure 16: Timeline of Event Study.........................................................................101
Figure 17: Effect of Announcement of Technological Innovations
on Market Value...........................................................................................103
Figure 18: CAAR Plots for Technological Leaders vs. Close Followers ................105
Figure 19: CAAR Plots for Large vs. Small Investment Portfolio ..........................106
Figure 20: CAAR Plots for High vs. Low Research Productivity ...........................110
Figure 21: CAAR Plots for Small vs. Large Firms..................................................112
Figure 22: CAAR Plots for High vs. Low Frequency of Announcements ..............113
Figure 23: CAAR Plots for Old vs. New Technologies...........................................115
Figure 24: CAAR Plots using Mean Adjusted, Market Adjusted,
and OLS Market Models..............................................................................120
Figure 25: Timeline of Event Study.........................................................................133
Figure 26: Calendar time Portfolio approach...........................................................137
Figure 27: Market Returns in each Phase of Innovation Project .............................142
Figure 28: Phases of Innovation Project ..................................................................144
Figure 29: Percentage of announcements with Positive Abnormal Returns............145

xi
LIST OF APPENDICES
Appendix A: Operating Principles of Sampled Technologies .................................162
Appendix B: Classification of Announcements.......................................................164

xii
ABSTRACT
Research and Development activity of a firm is the probably the most
important source of its intangible value. Numerous studies in the past have
demonstrated the failure of failure of leading companies to stay as market leaders
when technologies or markets change. But, the evolution of technologies might be
quite unpredictable in the short term and even more so in the long term. This stresses
the need to develop a better understanding of the phenomenon of technological
evolution and radical innovation.
In essay 1, I explore the phenomenon of technological evolution by collecting
a rich historical data from a comprehensive examination of the evolution of 23
technologies in six markets from the point of first commercialization till date. The
analysis provides rich insights into the patterns of evolution, competitive dynamics
and sources of innovation. In essay 2, I use functional analysis to develop techniques
that allow us to test hypotheses drawn from theoretical background underlying
technological evolution.
Moreover, for the innovation process to be sustainable, it is necessary for
firms to derive expected returns from innovative activity. Managers are interested in
estimating the marketability of technologically innovative products prior to
commercialization. This stresses the need for better measures to assess the impact of
firms’ strategy regarding innovation.
In essay 3, I examine the stock markets’ response to announcements of
innovations during the new product development phase of innovation projects. Using

xiii
the event study method, I examine the market returns to more than 500
announcements of various types of innovations in three industries over a period of
more than 25 years. The results indicate that firms may not be able to reap financial
benefits of being the first to announce innovative products unless they have a strong
track record of technological leadership. In essay 4, I examine short term and long
term returns to announcements of technological innovations using techniques in
financial analyses. The results indicate that limiting the focus on only new product
announcements severely underestimates the total returns to innovation.

1
I CHAPTER 1: INTRODUCTION AND OUTLINE
In this section, I develop the motivation, objective and outline of my
dissertation.
I.1 MOTIVATION
Any sufficiently advanced technology is indistinguishable from magic.
Sir Arthur C. Clarke, "Profiles of The Future", 1961 (Clarke's third law);
English physicist & science fiction author (1917 –)
Research and Development activity of a firm is the probably the most
important source of its intangible value. Nakamura (2001) estimated that US private
companies currently spent at least 1 trillion dollars on intangible assets, and that the
capital stock of US stock markets is approximately 5 trillion dollars – approximately
half the market value of all US corporations. Numerous studies in the past have
demonstrated the failure of failure of leading companies to stay as market leaders
when technologies or markets change (Christensen and Bower 1995).
But, the evolution of technologies might be quite unpredictable in the short
term and even more so in the long term. Current players often react in predictable
ways and lose their dominant positions in the industry (Utterback 1994a). For
example, no firm successfully stayed atop the desktop memory industry for longer
than a single generation (Christensen, 1997). This stresses the need to develop a
better understanding of the phenomenon of technological evolution and radical
innovation. Such an understanding can help managers predict how old technologies

2
mature and how emerging technologies improve in performance. We explore these
questions in essays 1 and 2.
Moreover, for the innovation process to be sustainable, it is necessary for
firms to derive expected returns from innovative activity. Managers are often
interested in estimating the marketability of technologically innovative products
prior to commercialization. They also need to assess the total market returns to
innovation projects over the short and the long term. Prior research has also
examined the effect of innovation on firm performance measures like sales, profits or
market share. But these are lag measures subject to many strategic and
environmental factors and suffer from problems in establishing causality. This
stresses the need for better measures to assess the impact of firms’ strategy regarding
innovation and new product development on financial returns. We explore these
questions in essays 3 and 4.
I.2 OUTLINE
The first chapter of the dissertation summarizes the motivation, outline,
objectives and key findings of my research. Chapters 2, 3, 4 and 5 are four separate
essays examining different aspects of technological evolution and financial returns to
innovation.
I.3 OBJECTIVE
The key objective of the dissertation is to enhance our understanding of the
phenomenon of technological evolution, and the market response to firms’
innovation strategies. Although each essay explores a distinct issue, all four essays

3
are related to the same substantive field of innovation and technology management.
Each of the essays examines a different issue related to this central objective and
contributes to the existing literature. In this section, we briefly present the findings of
each of the studies. We also discuss the theoretical backgrounds for each of the
essays and the techniques used.
As a well-known proverb states, “Well begun is half done.” Thus, in essay 1
(chapter 2), we begin by exploring the phenomenon of technological evolution itself.
We comprehensively examine the evolution of 23 technologies in six markets from
the point of first commercialization till 2003. The rich historical data, spanning more
than the last 120 years, provide rich insights into the patterns of evolution,
competitive dynamics and sources of innovation, and serve as the bases for
subsequent essays as well.
In essay 2 (chapter 3), we develop techniques that allow us to test hypotheses
drawn from theoretical background underlying technological evolution. More
specifically, we use functional analysis to develop algorithms for testing hypotheses
related to patterns of technological evolution. The analyses support the findings of
essay 2 and provide a general framework to investigate similar hypotheses related to
shape of functional data in other areas as well.
In essay 3 (chapter 4), we examine the stock markets’ response to
announcements of innovations during the new product development phase of
innovation projects. Using the event study methods developed in the finance
literature, we examine the market returns to more than 500 announcements of

4
various types of innovations in three industries over a period of more than 25 years.
The results indicate that firms may not be able to reap financial benefits of being the
first to announce innovative products unless they have a strong track record of
technological leadership.
In essay 4 (chapter 5), we extend the analyses of essay 3 on both the
theoretical and the modeling fronts. On the theoretical front, we expand the focus of
analysis from specific phases of innovation projects (e.g. commercialization) to
include the entire innovation project. On the modeling front, we also examine long
term returns to announcements of technological innovations using techniques in
financial analyses. The results indicate that limiting the focus on only new product
announcements severely underestimates the total returns to innovation.
We use a variety of techniques over the essays. All the essays, except essay 2,
are primarily empirical in nature. For data collection in essays 1, 3 and 4, we use
techniques of historical analysis. For analyses, we use non linear regression
techniques in essay 1, event study methods in essay 3, and financial analyses in essay
4.
The essays also differ in the dependence on theoretical backgrounds. Essay 2
draws from the economics and technology management literature. Essay 3 is based
on the bootstrapping and functional analytic techniques developed in statistics. Essay
4 and 5 are based on finance, economics and technology management literature.

5
II CHAPTER 2: TECHNOLOGICAL EVOLUTION AND
RADICAL INNOVATION
1

II.1 ABSTRACT
Technological change is perhaps the most powerful engine of growth in
markets today. To harness this source of growth, firms need answers to key questions
about the dynamics of technological change: (1) How do new technologies evolve?
(2) How do rival technologies compete? (3) How do firms deal with technological
evolution? Currently, the literature suggests that a new technology seems to evolve
along S-shaped path which starts below that of an old technology, intersects it once,
and ends up above the old technology. This belief is based on scattered empirical
evidence and some circular definitions. Using new definitions and data on 14
technologies from four markets, the authors examine the shape and competitive
dynamics of technological evolution. Results contradict the prediction of a single S-
curve. Rather, technological evolution seems to follow a step function, with sharp
improvements in performance following long periods of no improvement. Moreover,
paths of rival technologies may cross more than once or not at all.

1
Two papers coauthored with Gerard J. Tellis were published based on this essay as
per details below. The study was also supported by a grant from the Marketing
Science Institute.
Ashish Sood and Gerard J. Tellis, “Technological Evolution and Radical Innovation”
Journal of Marketing, July
Ashish Sood and Gerard J. Tellis, “The S-curve of Technological Evolution:
Strategic Law or Self- Fulfilling Prophecy?” MSI Working paper Series 04-116

6
II.2 INTRODUCTION
Understanding technological innovation is vital for marketers for several
reasons. First, technological change is perhaps the most powerful engine of growth.
It fuels the growth of new brands (e.g., Gillette’s Mach III), enabling corporations to
grow and incumbents to defend established positions. Second, technological change
creates new growth markets (e.g., digital video recorders) through radical
innovations (Tushman and Anderson 1986). Third, technological change often
transforms small outsiders (e.g., Intel) into market leaders replacing leading firms
(Foster 1986; Christensen 1997; Chandy and Tellis 1998).
To date the topic of technological evolution has been studied primarily in the
technology management literature. A central premise is that performance of a new
technology starts below that of an existing technology, crosses the performance of
the older technology once, and ends up at a higher plateau, so tracing a single S-
shaped curve (see Figure II- 1). There is scattered empirical support for the premise
and limited theoretical support for various aspects of the S-shape curve (e.g., Foster
1986; Utterback 1994a; Christensen 1997). Belief in this premise is so strong that it
has almost become a law in the strategy literature from which authors have derived
strong managerial implications. For example, they have warned that even though
managers might be able to squeeze out improvement in performance from a mature
technology at the top of its S-curve, that improvement is typically costly, short lived,
and small. Thus, a primary recommendation in the strategy literature and the trade

7
press is that managers abandon a maturing technology and embrace a new one in
order to stay competitive (e.g., Foster 1986; Christensen 1997).
A central practical problem that faces managers is when to shift investments
from the current to the future technology. If the S-curve is indeed valid, then the
appropriate time would be the inflection point of the S-curve. After this point,
performance improves at a decreasing rate until maturity.
New product development and major investments in research depend upon a
correct understanding of technological evolution in general and of the S-shaped
curve in particular. To foster this understanding, this study seeks answers to the
following questions:
• How do new technologies evolve? Do they follow the S-shaped curve or some
other pattern? Are technological changes predictable? Is the rate of technological
change increasing?
• How do rival technologies compete? What are the performance dimensions of
competition? What are the transitions between technological changes?
• Which firms carry out and survive technological evolution? Who introduces
radical innovations? Do incumbents survive the change?
The primary focus of the current study is empirical. We test hypotheses
derived from prevailing theory and prior findings, and examine the evolution of 23
technologies in six markets or industries. In the next three sections, we present the
prevailing theory, method, and results. The last section discusses the findings,
limitations, and implications of the research.

8
II.3 PREVAILING THEORY
The field does not enjoy a single, strong, and unified theory of technological
evolution. To guide our empirical work, we reviewed available theory and derived
testable hypotheses about the path, shape, source, and speed of technological
evolution and the competition among rival technologies.
Theory in this area has been partly confounded by the use of circular
definitions. So we begin by defining types of technological innovations
independently of their effects.
II.4 DEFINITIONS
Beginning with an early study by Schumpeter (1939), researchers have used a
wide variety of terms to describe innovations. Many terms such as revolutionary,
disruptive, discontinuous, or breakthrough (Freeman 1974; Tushman and Anderson
1986; Garcia and Calantone 2002) are intrinsically problematic because they define
an innovation in terms of its effects rather than its attributes. If the definitions are
then used to predict market outcomes (e.g., new entrants displacing incumbents from
disruptive technologies), researchers run the risk of asserting premises that are true
by definition. To avoid such circularity, we define technological change in terms of
intrinsic characteristics of the technology. As such, we identify and define three
types of technological change: platform, component, and design.
We define a platform innovation as the emergence of an entirely new
technology based on scientific principles distinctly different from those of the

9
existing technologies. For example the compact disk (CD) used a new platform, laser
optics, to write and read data where the prior technology used magnetism.
We define a component innovation as one that uses new parts or materials
within the same technological platform. For example, magnetic tape, floppy disk,
and zip disk differ by use of components or materials although all are based on the
platform of magnetic recording.
We define a design innovation as a reconfiguration of the linkages and
layout of components within the same technological platform. For example, floppy
disks decreased from 14 to 8 inches in 1978, to 5.25 inches in 1980, to 3.5 inches in
1985, and to 2.5 inches in 1989, although each was based on the platform of
magnetic recording (Christensen 1993).
These definitions may be considered refinements of the technological
dimension of radical innovations as defined by Chandy and Tellis (2000).
In our study we use the term new technology synonymously with new
platform. Further, we note that platform innovation results in improved performance
due to component or design innovations. In the interests of parsimony, this study
does not explicitly identify the component and design innovations that improve
performance in new platforms.
II.5 THEORETICAL BACKGROUND
In the technology literature, a strong consensus has developed about the
phenomenon of technology evolution; a consensus is emerging about the major
explanation or theory for this phenomenon.

10
Regarding the phenomenon, prior research suggests that technologies evolve
through an initial period of slow growth, followed by one of fast growth culminating
in a plateau. When plotted against time, the performance resembles an S-curve (see
Figure II-1a). Support for this phenomenon comes primarily from the work of Foster
(1986), Sahal (1981), and Utterback (1994a). These authors address the progress of a
technology on some primary dimension that is most critical to consumers at the time
the innovation emerges. Examples are brightness in lighting, resolution in monitors
and printers, and recording density in desktop memory products. Subsequent authors
have either accepted the above view or found additional support for it.
Authors have not developed any single, strong, and unified theory for the S-
curve. However, an emerging, and probably the most compelling, explanation
revolves around the dynamics of firms and researchers as the technology evolves.
We call this explanation the technology life cycle theory because it explains the
occurrences of the three major stages of the S-curve of technological evolution:
introduction, growth, and maturity (see Abernathy and Utterback 1978 and Utterback
1994a). We describe these stages as emerging from interplay of firms and
researchers over the life of the technology.
II.5.1 Introduction Stage
A new technological platform makes slow progress in performance during
the early phase of its product life cycle. Two reasons may account for this. First, the
technology is not well known and may not attract the attention of researchers.
Second, certain basic but important bottlenecks need to be overcome before any new

11
technological platform can be translated into practical and meaningful improvements
in product performance. For example, the laser beam was a new platform that
required much time and effort to achieve the safety and miniaturization required for a
surgical tool.
II.5.2 Growth Stage
With continued research, the technological platform crosses a threshold after
which it makes rapid progress. This stage usually begins with the emergence of a
dominant standard around which product characteristics and consumer preferences
coalesce (Utterback 1974). That consensus prompts focused research on the new
platform and rapid increases in its performance. Further, publicity of the
standardization draws a large number of researchers to study the new platform. Their
cumulative and interactive efforts also lead to rapid increases in performance.
Finally, the rapid progress leads to increases in sales of products based on the new
technology which leads to increase in revenues and profits and greater support for
research. These added resources fuel further improvement in performance (Klepper
1996).
II.5.3 Maturity Stage
After a period of rapid improvement in performance, the new technology
reaches a period of maturity when progress occurs very slowly or reaches a ceiling
(see Foster 1986; Brown 1992; Utterback 1994b; Chandy and Tellis 2000). Authors
have put forth several reasons for this change. Foster suggests that maturation may

12
be an innate feature of each platform a technology is good for only so much
improvement in performance.
Figure 1: Technological Evolution
Figure II-1a: Technological S-curve
Time
inflection
Performance
Time
inflection
Performance

Figure II-1b: Theory of S-Curves
Single crossing
Time
Old
Performance
1.
2.
New
Single crossing
Time
Old
Performance
1.
2.
New

Utterback (1994b) and Adner and Levinthal (2001) suggest that, as the
markets saturate, the focus of innovation shifts from product to process innovation.

13
As such, performance increases are few and modest. Reinganum (1985) and
Ghemawat (1991) suggest that maturity occurs when there is less incentive for
incumbent firms to innovate due to fears of obsolescence or cannibalization from a
rival platform. Thus the rate of innovation reduces relative to the growth stage.
Perhaps the best explanation is by Sahal (1981). He proposes that the rate of
improvement in performance of a given technology declines because of limits of
scale (things either get impossibly large or small) or system complexity (things get
too complex to work flawlessly). Once these limits are reached, the only possible
way to maintain the pace of progress is through radical system redefinition--that is,
move to a new technological platform.
II.6 HYPOTHESES
Based on the above theory, we derive hypotheses about six aspects of
technological evolution: shape, path, and dynamics of technological change on a
primary dimension, progress on secondary dimensions, and the source of innovations
and pace of technological transition.
II.6.1 Shape of technological progress
The theoretical background above suggests that technological evolution on a
particular performance measure follows an S-shaped curve. However, as also
indicated above, research in this area is relatively new and sparse. Extant theory does
not indicate the slope of this S-curve, the duration of the early or growth period, or
the timing or steepness of the turning points. We will try to determine if one can

14
identify any patterns or generalizations about these parameters. However, the most
precise testable hypothesis regarding shape we can formulate is:
II.6.1.1 H
1
: Technological progress on a primary dimension follows a single S-
shaped growth curve.

II.6.2 Technological Transition and Performance of Competing
Technologies
Do the paths of two technologies ever cross? If so, how many times? A
crossing signals whether a new technology is robust and productive enough to
supplant the old one. The extant literature suggests that such paths do cross and they
do so only once. This conclusion in the literature is based on three implicit premises.
First, successive technologies each follow an S-curve. Second, the new technology
starts its performance below the old technology. Third, the new technology ends
above the performance of the old technology.
Support for the first premise follows from that for H
1
. Support for the second
premise comes from several authors. Utterback (1994a) asserts that at the time an
invading technology first appears, the established technology generally offers better
performance or cost than does the challenger, which still has major problems. Foster
(1986) says, “During the fast growth phase of the first technology, the performance
of alternative technologies rarely surpasses that of established technologies.” Adner
and Levinthal (2001) also confirm that “it is unlikely that a new technology will
initially dominate an established technology in its primary domain of application.”
Likewise, Christensen (1992a, b) and Anderson and Tushman (1990, 1991) support

15
the general phenomenon that any new technology provides much lower benefits than
the old technology at the time it appears.
Similarly, several authors provide arguments and examples in support of the
third premise. Utterback (1994a, p.160) states that “the new technology often has so
much more potential for better performance that it” ultimately “surpasses the old.”
Two common examples cited in support of these arguments are steam ships
replacing wind-powered ships (Foster 1986) and turbo jet engines replacing internal
combustion engines for airplanes (Constant 1980).
Based on the above premises, Foster (1986) and Christensen (1997, p. 398)
postulate the following chain of events in the evolution of competing technologies.
Sometime in the life of an old technology, a new technology emerges. Initially, it
also makes slow progress on the primary dimension. However, at some time, it
enters its growth phase and improves rapidly. In contrast, the old technology
improves at a much slower rate even though major commitments are made to
develop products using old technology. As a result, a point is reached when the new
technology crosses the old technology in performance (see Figure II- 1b). This
crossing of the old technology is a signal of the end of its efficient progress. Hence
the threat to the old technology on the primary dimension is always from below. This
pattern of inter-technology competition results in overlapping S-curves, with each
new S-curve starting below but ending above the old. In support of this rationale
regarding the relative performance of technologies, some prior studies show a single

16
crossing between any two technologies (Christensen 1997; Foster 1986). This line of
argument suggests that:
II.6.2.1 H
2
: (a) At the introduction of a new technology, its performance is lower
than that of the old technology. (b) At the maturity of a new technology, its
performance is higher than that of the old technology. (c) As a result, the
performance path of a pair of successive technologies intersects once, when
the new technology crosses the old technology in performance

II.6.3 Dimensions of Technological Competition
Past research suggests certain secondary dimensions become important as
technology evolves. Progress occurs systematically along the first dimension, then
moves to the second, then to the third, and so on. These dimensions form the bases of
inter-technological competition. They also form the bases by which consumers
choose among rival technologies or products. In particular, Christensen (1999) posits
that while the new technologies under-perform along traditional dimensions, they are
better than old technologies on some secondary dimension.
The literature also suggests that the basis for such competition is quite
standard and occurs in the same form across markets. For example, Christensen
(1999) points out four generic dimensions of inter-technological competition:
functionality, reliability, convenience, and cost. Product functionality is the primary
attribute on which consumers choose products in that category. Similarly, Moore
(1991) suggests that products start competing on higher reliability after subsequent
innovations increase functionality beyond a certain point. Reliability is the product’s
consistent performance over time. Christensen suggests that after product
functionality and reliability requirements are satisfied, firms become more willing to

17
customize product designs to meet customers’ specific requirements, such as
convenience. Abernathy and Clark (1985) propose that the product becomes a
commodity and progress occurs through price reductions once the technology has
progressed up the S-curve sufficiently on functionality, reliability, and convenience.
The occurrence of such generic dimensions can be important in guiding firms about
the path of evolution and the direction of the next competitive attack. This line of
reasoning suggests that:
II.6.3.1 H
3
: Technological evolution progresses through four generic dimensions of
performance: functionality, reliability, convenience, and price.

II.6.4 Pace of technological transition
By pace of technological change we refer to the rate at which innovations are
introduced in the market. The pace of technological change may be essentially
stochastic in nature, due to the uncertainties in both the frequency of improvements
and the magnitude of gain realized through each innovation.
However, some authors believe that innovations are occurring faster, for
three reasons. First, every year greater resources are being devoted to research and
development. Second, every year an increasing number of countries and peoples get
involved in this research and development. Third, the progress in one area (e.g.,
computers) enables greater efficiencies in another area (e.g., materials design).
At least two studies have found empirical support for this thesis. For
example, Qualls, Olshavsky, and Michaels (1981) found that the percentage of
products in the introductory and growth phases of the product life cycle was

18
increasing in the last 50 years. This finding suggests that the pace of technological
transition is increasing and new products are being introduced at a faster pace.
Similarly, Tellis and Golder (1996) find that the time to takeoff of new products is
shorter now than in earlier decades. This finding would imply that the rate of
innovation is higher now than in earlier decades. Kayal (1999) found an increasing
recency in the median age of the patents cited on the front page of a patent
document, in the past 25 years. This finding suggests that the cited patents are
relatively recent and that the technology is experiencing a frequent replacement of
one generation of inventions by another, which is a sign of a rapidly progressing
technology. On the other hand, Bayus (1994, 1998) believes that even though more
products and product variations are available in the market at any point in time, the
rate of change now is not higher than in earlier decades. Hence, we can propose the
following null and alternative hypotheses:
II.6.4.1 H
4o
: The pace of technological change is constant with calendar time.
II.6.4.2 H
4a
: The pace of technological transition is increasing with calendar time.

II.6.5 Source of new technologies
Which types of firms are more likely to introduce the platform innovations:
incumbents or new entrants, large firms or small firms? This topic has been the
subject of research for decades. The conventional wisdom and dominant view is that
platform innovations come primarily from small firms or new entrants. These firms
are ridiculed and ignored by incumbents in the beginning, but later grow to be
successful with the progress of the new technology. Scherer (1984) shows how new

19
entrants contribute to a “disproportionately high share” of revolutionary industrial
products.
Previous studies have also proposed many reasons for large incumbents
failing to introduce innovations, including incumbent’s technological inertia
(Ghemawat 1991), complacency (Robertson, Eliashberg, and Rymon 1995),
arrogance (Lieberman and Montgomery 1988), and unwillingness to cannibalize
their current products (Chandy and Tellis 1998). Thus the dominant view in the
literature is that:
II.6.5.1 H
5
: Platform innovations are introduced primarily by small entrants.

II.7 METHOD
A readymade database does not exist for the study of technological evolution.
We collected our own data using the historical method, following a growing trend in
marketing (see Golder and Tellis 1993; Golder 2000; Chandy and Tellis 2000). The
benefits of using the historical method include freedom from survival and self-report
bias, ability to assess causality through longitudinal analysis, and new insights from
a fresh look at history (Golder 2000). Below we detail our sample selection, sources,
and procedure for data collection.
II.7.1 Sample Selection
We used two criteria in selecting categories: some overlap with past research
and adequate number of platform innovations. We selected a portfolio of categories
such that it included some that had been investigated in past studies (e.g., memory

20
and lighting) and others that had not been researched. This coverage allows us to
compare our results with past studies and validate these findings in new categories.
However, the present study goes further than previous studies in one important
aspect--within each category we selected all technologies. We also required that the
category must have had at least two platform innovations.
On the basis of these criteria we chose external lighting, data transfer,
computer memory, desktop printers, display monitors, and analgesics. Note that the
first two are utilities, the next three are consumer electronics, and the last one is
medical. Thus the sample crosses a broad spectrum of products.
II.7.2 Sources
The information required for this study is technical data on product
performance for various technologies at different stages of its evolution. The primary
sources of our data were reports in technical journals, industry publications, white
papers published by R&D organizations, and annual reports of industry associations.
We sourced industry reports of market research firms (like Disk/Trend or Stanford
Resources), press releases, timelines of major firms, and records in museums which
profiled innovations and the development of industries. We recorded the current
performance of many technologies from product information bulletins released by
firms. We used Medline and Cochrane Database of Systematic Reviews for
information on the biomedical literature for the analgesics category.

21
II.7.3 Procedure
We followed the rules for data collection for the historical method (Golder
2000). Despite these efforts, it was difficult to avoid some problems. We explain
these problems and the rules we used to resolve them.
First, in some cases, we found conflicting levels of performance between
multiple sources for the same technology in the same period. In these cases, we used
two checks to record the appropriate performance level for each year. One, we
checked to make sure the recorded performance was of comparative models. For
example, lamps based on the same technology provide different levels of efficiency
depending on the wattage of the lamp; so we ensured that the recorded efficiency
was for lamps of the same wattage. Two, if two sources provided conflicting data for
a period within the series, we chose the one whose starting and ending values were
more consistent with the rest of the series.
Second, within a platform innovation, we found that technologies evolved
with component and design innovations. At any particular time, a platform
technology could be represented by alternate components and designs. In such cases,
we use the performance of the best component or design or combination of the two,
as the performance of that platform. For example, magnetic memory is available in
various designs--floppies, zip disk, tape, hard disc, etc. Among these, the hard disk
had the highest memory. So, we chose the memory of the hard disk as the level of
storage capacity of magnetic memory.

22
We collected intensive data on two entities in each category--the platform
technology and the firm that introduced it. In each category, we first identified all the
platform innovations. Then we recorded the maximum performance of products in
each platform technology for each year from its year of introduction until 2001.
None of the past studies makes an explicit statement of the starting point of the S-
curve. We use the date of first commercialization of a product based on each
technology as the standard starting point to compare the relative performance of any
two technologies. We also collected additional information regarding the
incumbency status and relative size of the firm at the time of introduction of the
technology.
The literature is quite consistent in recommending the use of performance as
the key dependent variable when testing the S-curves. For example, Christensen
(1999, p. 19) suggests, “A simple graph plotting product performance as it is defined
in the mainstream markets on Y axis and time on the horizontal axis.” Similarly
Foster (1986, p. 274) suggests that we “collect the data on how well the product
performed.” Utterback (1994a) also identifies the dependent variables as
performance when comparing two technologies: “The established technology
generally offers better performance or cost than does the challenger, which is still
unperfected” (p. 158).
II.8 RESULTS
We first present the identification of platform innovations and the
performance attributes in each category. Next, we present findings on the hypotheses

23
regarding the shape, path, and dynamics of technological changes. We then present
findings on the competition, rate of improvement, and source of new technologies.
We used nonlinear regression to test the first and primary hypothesis, the existence
of the S-curve. We used cross tabs, chi-square and binomial tests, and regression
analysis for the other hypotheses. We used Tobit analysis to predict the probability
and size of the performance jumps.
II.8.1 Identification of platform innovations and performance attributes
We identified various technologies in each of the markets, each of which was
initiated by a platform innovation: five in external lighting, four each in desktop
printing, display monitors, and analgesics, and three each in desktop memory and
data transfer. Appendix A describes these technologies briefly. (Detailed definitions
for these innovations are available in a technical note from the authors.)
We found that some of the platform technologies may not be readily
distinguishable to the customers for one major reason. Even when a new technology
differs radically from an old one, firms try to facilitate consumer adoption by
maintaining a uniform interface for the new product based on the new technology.
For example, lamps from different platform technologies such as incandescence or
arc discharge use standard screw ends. We considered the underlying technologies as
distinct if they are based on fundamentally different scientific principles. We adopted
this rule so as not to confuse differences in technologies based on their
characteristics with superficial differences based on derived products.

24
In each category, at a particular stage of technological evolution and
consumer needs, certain dimensions of performance assume primacy. We did not
have difficulty identifying these dimensions based on the historical description of the
technologies and of what consumers then considered important. Fortunately, each of
the dimensions has fairly clear performance metrics. In choosing metrics, we were
careful to take into account output per unit of input (see Table II- 1).
Table II-1: Primary Dimensions of Competition
Category Primary Dimension Metric
External lighting Lighting efficacy Lumens per watt
Desktop memory Storage capacity Bytes per square inch
Display monitors Screen resolution Dots per square inch
Desktop printers Print resolution Pixels per square inch
Data transfer Speed of data transmission Megabits per second
Analgesics
Absolute risk reduction
(ARR)
Numbers needed to treat
(NNT = 1/ARR)
II.8.2 Shape of technological evolution
H
1
hypothesized that technologies evolve through S-curves. We first plotted
the performance of technologies on the y-axis against time on the x-axis (see figures
II-2a-f). As postulated by the hypothesis, these figures reveal that technologies do
have a slow start and a sudden growth spurt.

25
Figure 2: Technological Evolution in six categories
Figure II- 2a: External Lighting Figure II- 2b: Desktop Memory*
0
20
40
60
80
100
120
140
160
180
200
1879 1899 1919 1939 1959 1979 1999
Lighting Efficacy
Incandescent
Arc Discharge
Gas discharge
LED
MED
0.1
1
10
100
1000
10000
100000
1971 1976 1981 1986 1991 1996 2001
Areal Density .
Magnetic
Optical
Magneto-Optical
Figure II- 2c: Display Monitors* Figure II- 2d: Desktop Printers*
0.0001
0.001
0.01
0.1
1
10
1971 1974 1977 1980 1983 1986 1989 1992 1995 1998 2001
Resolution
CRT
LCD
PDP
OLED
0.001
0.01
0.1
1
10
1978 1981 1984 1987 1990 1993 1996 1999
Resolution .
Dot Matrix
Ink Jet
Laser
Thermal
Figure II- 2e: Data Transfer* Figure II- 2f: Analgesics
1.0E+02
1.0E+03
1.0E+04
1.0E+05
1.0E+06
1.0E+07
1.0E+08
1.0E+09
1.0E+10
1.0E+11
1.0E+12
1.0E+13
1962 1966 1970 1974 1978 1982 1986 1990 1994 1998 2002
Bits per sec
Cu/Al w ires
Fiber optics
Wireless

0
0.1
0.2
0.3
0.4
0.5
0.6
1800 1825 1850 1875 1900 1925 1950 1975 2000
Efficacy (1/NNT)
Acupuncture
Opiods (Narcotics)
Non-opiods Anti-inflammatory (NSAIDs)
Non-opiods Non-Anti-inflammatory (Acetaminophen)
Series1
+

* Performance on y-axis is in log scale
+ Accurate records of efficacy of acupuncture unavailable before 1971

However, we found that the technology progresses along a single S-shaped
path with a single inflection point followed by a permanent plateau or maturity in
only 6 technologies. In 14 technologies we do not find a single S-curve. Rather we

26
find long periods of static performance interspersed with abrupt jumps in
performance. A visual examination of these plots suggests a series of S-curves or
multiple S-curves, with a new one starting at the point the earlier one seems to
plateau. The figures suggest a series of step functions, each of which could
approximate an S-curve.
Two technologies show no change in performance since introduction while
one technology has only two data points. So we exclude these three technologies
from the formal tests of shape. To formally test hypothesis H
1
that evolution follows
an S-shaped function, we carried out the following two tests:
First, we fitted the generalized logistic function to the six technologies that
reveal a single S-shaped curve:

) (
1
c t b
t
e
a
d Y
− −
+
+ =
(1)
Where Y
t
= performance of the technology in year t, and a, b, c, and d are
parameters to be estimated: b is the growth rate, c is the time of maximum growth or
the inflection point, and a + d is the upper asymptote of the S-curve. We used the
nonlinear regression techniques in SAS to estimate the model over the entire data.
Second, for technologies seeming to exhibit multiple S-curves, we fitted the
generalized logistic function to both the entire series of data and to a sub-sample of
data which exhibited an S-curve. We used two criteria to select a subset of data for
this purpose: (1) performance of the technology during the sub-sample had a upper
plateau that was longer in duration than the duration of the just preceding growth

27
phase and (2) the subsequent jump in performance in the one year immediately after
the plateau was almost double the performance during the entire plateau. In these
cases, our practical goal is to test how well an S-curve fits on the whole sample and
whether an S-curve on a sub-sample fits better than one on the whole data.
Figure 3: Technologies with Single S curve
Figure II- 3a: Optical Memory
Figure II- 3b: Wireless Data
Transfer
0
500
1000
1500
2000
2500
3000
1 2 3 4 5 6 7 8 9 10 111213 141516 1718 19 20
Areal Density
Actual
Full S
1
10
100
1000
10000
100000
1000000
10000000
100000000
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23
Bits per sec
Actual
Full S
Figure 4: Technologies with Multiple S-curves
Figure II- 3c: Incandescent Lighting Figure II- 3d: LED Lighting

0
5
10
15
20
25
30
1879 1889 1899 1909 1919 1929 1939 1949 1959 1969 1979 1989 1999
Lighting efficacy
Actual
Full S
Small S

0
5
10
15
20
25
30
1965
1968
1971
1974
1977
1980
1983
1986
1989
1992
1995
1998
2001
L ighting E fficacy
Actual
Full S
Small S
For the 6 technologies with an apparent single S-shaped curve, the
generalized logistic function provides a very good fit with the data (see Table II-2a
and figures II-3a-3b). For the remaining 14 technologies which seemed to exhibit
multiple S-curves, an S-shaped curve over a sub-sample of data fits better than one
over the whole data, even after taking into account degrees of freedom (see Table II-
2b and figures II-3c-3d). Table II-2b shows that the fit over the sub-sample gives an

28
average reduction of 96% in the mean square error (MSE) from that over the whole
sample.
Table II-2a: Test of Logistic Fit for Technologies with Single S-curves
Parameter Estimates
Technology
Upper Asymptote
(t-value)
Growth rate
(t-value)
Magnetic memory 4.28 (8.5) .50 (24.8)
Optical memory 1.20 (7.4E + 06) 30.95 (24.2)
Magneto-optical memory 3.51 (5.2) .51 (7.8)
Wireless data transfer 1.57 (234) 6.29 (5.3)
Opioid analgesics1 4.40 (59.2) .67 (15.9)
NSAIDs 1.93 (6658) 8.44 (26.6)
1. Acetaminophen, non-drug analgesics, and OLED displays were
excluded from analysis
Table II-2b: Test of Logistic Fit for Technologies with Multiple S-Curves
Technology
Improvement in fit* of sub-sample over full data
(* Measured as reduction in Mean Square Error)
Full data Sub-sample
# Years MSE # Years
%
Reduction
Incandescent lighting 123 .05 50 88%
Arc discharge lighting 94 .02 40 86%
Gas discharge lighting 70 .09 25 97%
Led lighting 37 .15 30 97%
Med lighting 13 .03 8 100%
Dot matrix printers 24 .04 14 95%
Inkjet printers 18 .05 13 97%
Laser printers 17 .02 10 100%
Thermal printers 11 .09 9 95%
CRT monitors 31 .09 27 97%
LCD monitors 19 .20 18 93%
PDP monitors 19 .14 7 100%
Copper/al cables 42 .57 21 95%
Fiber optics 27 .06 25 99%
Mean 38.9 - 21.2 96%

29
Further, Table II-2c shows that the parameter estimates of the fitted
generalized logistic function for the sub-sample are significantly different from the
parameter estimates over the whole sample, after adjusting for degrees of freedom.
For this test, we used the t-test for differences in parameters with unequal variances
over the two models. Most importantly, the upper asymptote in the sub-sample is
significantly and substantially different from that in the whole series, leading us to
reject hypothesis H
1
of a single S-shaped curve.
Table II-2c: Difference in parameter estimates for technologies with multiple S curves
Difference in Parameter Estimates
Upper Asymptote Growth rate Technology
Difference (t-value) Difference (t-value)
Incandescent lighting 2.6 (182) -.4 (-137)
Arc discharge lighting 2.9 (282) .2 (71)
Gas discharge lighting 3.2 (141) -7.4 (-36)
Led lighting 318 (43) -1.3 (-80)
Med lighting 95.5 (731) -20.4 (-52)
Dot matrix printers 2.2 (68) -4.9 (-42)
Inkjet printers 2.6 (79) -1.8 (-9)
Laser printers 1.6 (63) -20.8 (-80)
Thermal printers 2.5 (14) -7.2 (-14)
CRT monitors 97.3 (44) -.2 (-11)
LCD monitors 89.6 (21) -.4 (-11)
PDP monitors 1.1 (10) -28.6 (-222)
Copper/al cables 17.7 (425) .6 (14)
Fiber optics -1.6E+12 (-3.0E+08) -12.6 (-19)
To summarize, the hypothesis of a single S-shaped growth in performance is
supported for only 6 of the 22 technologies. For 2 technologies, performance does
not change. For the remaining 14 technologies, change in performance follows a
series of irregular step functions better approximated with multiple S-curves than a
single S-curve. Across these step-functions within a technology, estimates of growth

30
rate and especially performance at maturity (the upper asymptote) differ
substantially.
These findings have important implications. An analyst expecting an S-
shaped curve would conclude that the first curve (on the sub-sample) meets the
hypothesis. He or she would then wrongly conclude that the technology has matured
at the upper asymptote, when indeed it has not. As a result of this incorrect
conclusion, the analyst would suggest abandoning the old technology. The average
period for the sub-sample S-curves is 21 years compared to an average of 39 years
for the full period. Thus this error may result in premature abandonment of a
promising technology as early as at least 18 years before its life to-date. Substantial
improvements in performance after the first plateau suggest the gravity of the error.
II.8.3 Performance of competing technologies
H
2
hypothesizes three characteristics of technological competition--
performance of new technology at introduction and at maturity (points 1 and 2 in
Figure II-1b) and a single crossing when the new technology crosses the old
technology in performance. We tested each part of this hypothesis with a binomial
test of the expected frequencies against observed frequencies.
Our results are quite contrary to the hypotheses (see Table 3a). The binomial
test rejects both H
2a
and H
2b
(p < .001). A majority of new technologies performed
better than the old technology, right from the time they were introduced. Also, many
new technologies never improved over the old technology while others enjoyed brief

31
spells of dominance over the old technology before the old technology regained
dominance.
Table II-3: Performance of new technology relative to old
Technology Category
Proportion of new
technologies with low
performance wrt old at
introduction*
Proportion of new
technologies with high
performance wrt old at
maturity^
External lighting 1/4 ¾
Desktop memory 0/2 ½
Display monitors 3/3 1/3
Desktop printers 2/3 1/3
Data transfer 1/2 ½
Analgesics 1/2 1/2
Total 8/16 9/16
Binomial test - Probability
technology performs as per
H
2a
/H
2b

* p < .001 ^ p < .001

This unexpected pattern of evolution results in three distinct types of
crossings between any pair of successive technologies (see Table II-3b). First, 8 out
of 16 technology pairs showed no crossing at all. In these cases new technologies
either started higher than the old technology and remained higher or started lower
than old technology and never crossed the old technology long after their
introduction. Second, many technologies (4 out of 16) showed multiple crossings. In
such cases, the new technology passed an older technology but was not able to
sustain its advantage. Third, the expectation of a single crossing, of new passing the
old from below, was satisfied in only 4 out of 16 technologies--a mere 25%rejecting
the hypothesis H
2c
of a single crossing (p < .001) using the binomial test of
differences.

32
Table II-4: Number of crossings between new and old technologies
No. of Crossings
Technology Category
Single Multiple No Crossing
External lighting 0 1 (2)* 3
Desktop memory 1 1 (3)* 0
Display monitors 1 1 (3)* 1
Desktop printers 2 1 (2)* 0
Data transfer 0 0 2
Analgesics 0 0 2
Total 4/16 4/16 8/16
Binomial test - Probability No. of crossings
as per H
2c

* p < .001 * p < .001 * p < .001
* Figures in brackets indicate the total number of crossings in the technology pair
with multiple crossings
In summary, we find no support for any of the three sub-hypotheses on
performance of competing technologies. So the final status of each technology
cannot be determined solely from the direction of the attack or timing of
introduction. As such, it would be unwise for an incumbent to scan for competition
only among technologies performing worse than its current technology and to make
decisions on that basis.
II.8.4 Progress on secondary dimensions
We find that new technologies typically perform better than old technologies
on some secondary dimension. This new dimension also affected the basis of
competition in these markets. For example CRT monitors were emphasizing
resolution prior to LCDs. However, LCDs were inferior in resolution but superior in
lightness to CRTs. After their introduction, the two technologies competed on dual
dimensions of resolution and lightness. These findings are consistent with
Christensen’s (1999) assertion about emergence of new dimensions.

33
In contrast to H
3
which proposes only generic dimensions, our results suggest
a sequence of random, unpredictable secondary dimensions in each of the six
categories. Each platform technology offered a completely new secondary dimension
of competition while still competing on the primary dimension. These dimensions
are supposed to form the basis of inter-technological competition as well as the basis
of consumer choice.
For example, consider four successive technologies in monitors: CRT, LCD,
plasma, and OLED. CRT monitors were initially introduced on the basis of
resolution. The new LCD monitors were inferior on resolution but were stronger on a
new but important secondary dimension: thinness and weight. The new plasma
display panels were inferior in resolution to both the prior technologies, but were
superior on a tertiary dimension, large screen size. The new OLED displays were
inferior to the prior three technologies in resolution but were superior in efficacy.
Hence, each technology in display monitors emphasized a new dimension, on which
it was superior: resolution, compactness, screen size, and efficacy.
Similarly, we identified the emergence of a new secondary dimension in each
of the other categories (see Table II- 4). Again none of these have any resemblance
to the generic dimensions proposed by Christensen (1999) and represented by H
3
.
We also found that technologies that excel in a particular dimension cater to
particular segments that value that dimension. When the mass market focuses on one
old or new dimension, niche markets, interested in the other dimensions, might still
survive. For example, LED lighting has carved a niche in automobile, signage, and

34
contour lighting applications; MED lighting is the popular choice in applications
where lamp replacement is difficult because of accessibility.
Table II-5: Emergence of Secondary Dimensions of Competition
Market Secondary Dimensions
External lighting
Brightness, color rendition, light efficacy, compactness and
life
Desktop memory Areal density, reliability, and cost
Display monitors Resolution, compactness, screen size, and efficacy
Desktop printers Resolution, graphics, speed and continuous color rendition
Data transfer Transfer speed, bandwidth and connectivity/mobility
Analgesics
Analgesia, reaction speed, targeted-action, and reduced side
effects

In summary, we find that though new technologies perform better than old
technologies on secondary dimensions, the competition evolves in new,
unpredictable secondary dimensions instead of the standard four generic dimensions
proposed by literature.
II.8.5 Pace of technological transition
The alternative hypotheses H
40
and H
4a
state that the pace of technological
transition is either constant or increasing, respectively. Past findings in the literature
on this issue are neither consistent nor use direct measures. To test these hypotheses,
we used three direct measures of the rate of technological change:
• The pace of introduction of new technologies, for each of the six categories,
calculated as the number of years between successive introductions.

35
• The pace of technological improvements within each platform, calculated as the
number of years between (non-zero) improvements in technological
performance.
• The annual rate of improvement for each technology, calculated as the
percentage increase in performance over the past year.
Tests of all three measures support an increasing pace of technological
change (see figures II-5a-c). The negative slopes of trends for both the measures of
duration suggest declining duration between introductions of successive new
technologies as well as declining duration between successive improvements in each
technology. The positive slope of trend for the rate of improvement over last year
suggests an increase in the pace of technological change.
Figure 5: Pace of Technological Transition
Figure II- 5a: Declining Time between Introduction of new Platform Innovations
Ŷ = -9.1 Ln(t) + 29.9
t=-3.4 F=11.6 R
2
= 0.49
0
10
20
30
40
50
60
70
80
90
100
1 2 3 4 5 6 7 8 9 1011 1213 141516
Platform Innovation
# years since last introduction
No of years
Trendline

36

Figure II- 5b: Declining Duration between Successive Improvements within
Technology
Trendline Ŷ = -2.9 Ln(t) + 8.9
(t= -6.7 F=45.1 R
2
= 0.63)
0
0.5
1
1.5
2
2.5
3
3.5
4
4.5
5
1357 9 11 13 15 17 19 21 23 25 27
No. of Successiv e Improv ements
No. of Years
Lighting
Memory
Monitor
Printer
Analgesics
Data Transfer

Figure II- 5c: Increasing Average % Improvement over Last Year
Ŷ = 0.0118 Ln(X) - 2.0653
t=4.2 F= 18.1 R
2
= 0.2511
-3
-2.5
-2
-1.5
-1
-0.5
0
0.5
1
1.5
2
2.5
0 20 40 60 80 100 120 140
Years
Log (Avg. % Increase)
Average % Improvement
Trendline

To formally test the alternate hypotheses, H
4o
and H
4a
, we pooled the
categories and estimated the following regression equation for each of the three
measures.
Y
t
= α + β log (t) + ε
t
… (2)
Where Y
t
represents each one of the above three measures of pace of
technological change in year t, α and β are coefficients to be estimated, and εt are
the errors assumed to be IID normal. The coefficients are significantly different

37
from 0 Note that our test is in the same spirit as meta-analyses which pool
estimates across multiple categories (Assmus, Farley, and Lehmann 1984; Tellis
1988). Such pooling increases the power of the test and reduces the probability of a
Type II error.
II.8.6 Towards a new predictive model
The empirical findings above lead to a rejection of the widely accepted S-
shape of technological evolution. They strongly suggest that most technologies
evolve through irregular step functions with jumps in performance or sudden
improvements occurring after many years of static performance. Are the timings and
size of these jumps entirely random or is there a pattern to them? If so, is the pattern
explainable or predictable? To answer these questions, we modeled the probability of
a jump (
*
1
Y
) and the size of the jump in a particular year (Y2) as a function of some
explanatory variables (X), thus:

U
e
e
Y
X
X
+
+
=
δ
δ
'
'
1
*
1
(3)

V X Y + = β
'
2
(4)
])
1
[ , 0 ( ~ ) , (
2
σ ρσ
ρ σ
N V U

We observe Y1 = 1(
*
1
Y
> 0) and X always, but observe Y2 only when Y1 =
1. We estimated the joint outcome – probability and size of jump - using the Tobit
(Type II) model (Tellis 1988; Maddala 1983; Amemiya 1985). This formulation
eliminates the sample selection bias introduced from observing the size of a jump on

38
only those events where the probability of a jump is 1. The explanatory variables in
X are from two broad categories – factors directly related to the intrinsic
characteristics of the technology and factors related to the existing competition in the
category. We first standardized the performance in order to pool across categories
and then calculated the size of jump relative to the preceding year. The results are in
Table II-6. The Rho2 (reduction in uncertainty) is 62% for the first stage model and
the R2 (explained variation) is 61% for the second stage model. The best predictors
of the probability of a jump are length of time since last jump (-), a jump in a rival
technology (+), and the number of prior jumps (+). The best predictors of the joint
event of probability and size given a jump are current performance of the technology
(+), time since last jump (+), number of prior crossings by rival technologies (-),
order of entry in category (-), current year (+), and number of prior jumps in the
technology (-).
These results together with the earlier ones suggest that the improvement in
technological performance neither follows an S-shaped curve nor is it entirely
random. Rather, improvement in performance follows some patterns that are
somewhat predictable. In general, size of the jump increases with newer technologies
and with those that have higher performance.
The probability of a jump also increases with a jump in a rival technology.
Most important, a plateau in performance by itself does not imply a technology has
matured. Indeed, the longer the plateau the higher the expected net-jump in

39
performance. Thus, the past pattern of performance may have clues about the future
performance, though it may not follow the path of a single S-shaped curve.
Table II-6: Modeling Probability and Size of Jump in Performance
Variable Probability of Jump
Size of Jump in
Performance
Current performance -.006 (-.07) 1.9 (30)
No. of prior crossings -.17 (-1.4) -1.05 (-11.4)
Order of entry .26 (.5) -.6 (-10.9)
No of prior jumps .06 (3.1) -.18 (-11.3)
Current year .001 (.4) .01 (5.5)
Time since last jump -.96 (-9.6) .03 (3.5)
Jump in rival .46 (2.5) -.3 (-2.3)
Mills - .09 (2.6)
% reduction in
Uncertainty (Rho2)
62% -
Adj-R2 - 61%
Note: t-values in parentheses
II.9 DISCUSSION
This section summarizes the findings and discusses questions, implications,
and limitations of this study.
II.9.1 Summary of findings
The current research leads to six major findings:
• Technologies do not show evidence of a single S-shaped curve of performance
improvement. Rather they evolve through an irregular step function with long
periods of no growth in performance interspersed with performance jumps. A

40
jump in performance appears to be largest after a long plateau of no
improvement.
• New technologies may enter above or below the performance of existing
technologies. The performance curves of a pair of competing technologies rarely
have a single crossing.
• The path of technological evolution seems partially predictable. Past
improvement in performance of the same technology, improvement or crossing
by a rival technology, and especially crossing by a rival firm tend to signal
immediate improvement in performance.
• Each new technology introduces a sequence of random, so far unpredictable
secondary dimensions as a new basis of competition.
• The rate of technological change and number of new technologies is increasing
over the time.
• New technologies come as much from new entrants as from large incumbents.
II.9.2 Questions and further analysis
Our results are based on comparing a technology with the one introduced just
prior to it. One might ask if these results are in any way sensitive to the reference
point of the comparison technology. We redid all our analysis using the first
technology and the dominant technology in the category. The results were not
materially different from those reported here.
We also examined the gestation period of each technology, defined as the
time it takes for a firm to convert a patent to a commercial product. The average

41
gestation period for technologies is 15.8 years for external lighting, 14.5 years for
display monitors, 14.3 years for desktop printers, 9.7 years for desktop memory, 22.7
years for data transfer technologies, and 67 years for analgesics. The overall average
for all categories is 22.3 years. (We again had to exclude acupuncture from the
calculation of gestation period as it is difficult to ascertain the exact date of
introduction of this technology.) Given this long gestation period, our results show
that investors need to be patient and managers need to persevere in order to bring a
new technology to fruition.
Fourth, we examined whether the gestation period is shrinking over time
given the increasing pace of technological change. In order to examine this
hypothesis, we did a median split of the gestation period by the year 1970. Each of
the groups had technologies from all six categories. An analysis of the mean
gestation periods of the two groups revealed a large and significant difference
between the pre-1970 set (average 30 years) and post-1970 set (average 14 years) of
16 years (t = 2.10).
Finally, to check whether these results were at the cost of a censoring bias
from not allowing enough time for the new technology to improve, we compared the
time taken for the technologies that failed to cross old technologies with those that
did. The average number of years for new technologies to reach the point of first
crossing the old technology is 4.2 years. In contrast, categories in which the new
technology never crossed the old have been in existence for 18.3 years. Hence, the

42
lack of a crossing cannot be due to not leaving enough time or for a censored time
frame.
II.9.3 Survival bias
Does survival bias affect our results? It is impossible to rule out survival bias,
albeit we took great pains to minimize its role. First, we tried to include every
technology that was commercialized in the markets that we covered for the time
period that we studied. Second, to examine the possible effect of inadvertently
excluding any technologies from the analyses, we define two categories of
technological failures—non-starters and stagnant technologies. Non-starters are
those that were used in related fields, could have been used in the target markets with
some modification, but were never used. In other words, these are potential
technologies that were never commercialized. On the other hand, stagnant
technologies are those that fail to show improvements in performance, cost, and
features. Stagnant technologies were selectively used in these markets at first but
were never successful beyond the initial introduction because of this failure to
improve. Examples are combustion-based lighting technologies like candles or oil-
based lamps. The key issue is whether our exclusion of failures, so defined, biases
our results, just as survival bias upwardly biases the alleged advantages of market
pioneers (Golder and Tellis 1993).
We believe that non-starters do not affect any of our results for three reasons.
First, in the pioneering literature, failures are firms that did commercialize a product
but are no longer surviving. Our definition of non-starters is much more stringent--

43
they are technologies that were never commercialized. Second, most of our analysis
tracks the progress of individual technologies without averaging performance across
technologies. As such, the exclusion of non-starters does not bias computed
performance levels (as it does when computing market share performance of only
surviving market pioneers). Third, our entire analysis tracks the performance of a
technology given that it was commercialized. We do not make any predictions or test
any hypotheses about the productivity of R&D, in which case non-starters would
loom large.
To ascertain what insights might be gleaned from a consideration of stagnant
technologies, we explored the histories of the three such technologies from three
categories--arc lamps in external lighting, chain printers in desktop printers, and wire
recorders in desktop memory. (Details of these technologies are available from the
authors.) We find common factors in each of these cases. First, each of the stagnant
technologies failed to develop an acceptable standard or be included in a prevailing
standard, e.g., wire recorders were excluded from the standard-setting process in
favor of magnetic tape technology by the recording industry. Such exclusion in the
standards-setting process, also termed technological lockout (Schilling 1998), leaves
the technology in a weak market position (Shapiro and Varian 1999). Second, a new
and better technology was either introduced very early in the life of stagnant
technology (e.g., magnetic tape and wire recorders, dot matrix printers and chain
printers etc.), or the performance of new technology was exceptionally superior to
(or growing at a fast rate than) the stagnant technology (e.g., incandescent lamps and

44
kerosene lamps). Perhaps as a result of these two factors, the stagnant technology did
not show any improvement in performance on all dimensions that we tracked.
The exclusion of these technologies does not lend support to any the
alternative hypotheses that we tested and rejected, such as a single S-shaped curve,
single crossing, or generic dimensions of competition. However, since stagnant
technologies were excluded, it would be wrong to conclude from our results that
performance always improves over time.
II.9.4 Performance per unit price
Some authors propose that, when testing the S-curves, benefits per dollar
should be used as the key dependent variable instead of performance (Chandy and
Tellis 1998). Although all our current performance measures also have a
denominator for proper scaling (e.g., Lumens per watt) we investigate the sensitivity
of our results to using this alternate metric.
We collected data on benefits per dollar for three categories--desktop
printers, desktop memory, and display monitors. For each technology, we identified
the product offering the highest benefit per year and the price at which it was offered
at introduction. The results in figures II-6a-c do not provide support to any of the
hypotheses that we rejected. For example, we observe multiple crossings in all
categories, and new technologies being introduced with higher benefits per dollar.
Moreover we found that the evolution of technologies is not even a monotonic
function of benefits per dollar. One possible reason is that firms charge higher prices
for technologically advanced products until competition drives the price down.

45
Figure 6: Performance per unit price
Figure II- 6a: Desktop Printer Technologies--Resolution per
Dollar
0.000001
0.00001
0.0001
0.001
0.01
0.1
1
1978 1981 1984 1987 1990 1993 1996 1999 2002
Kdpi_sqr/$
Dot Matrix
InkJet
Laser
Thermal

Figure II- 6b: Desktop Memory Technologies – Storage
Capacity per Dollar
0.001
0.01
0.1
1
10
100
1980 1984 1988 1992 1996 2000
Mb/$
Magnetic
Optical
MO

Figure II- 6c: Display Monitor Technologies--Size per Dollar
0.0001
0.001
0.01
0.1
1
1987 1989 1991 1993 1995 1997 1999 2001
inches/$
CRT
LCD
PDP
OLED

46
II.9.5 Multi-attribute performance
Some might question whether our results hold when we take into account
multiple dimensions of performance simultaneously. To clarify this issue, we
repeated the analysis using multiple dimensions in two categories: desktop printers
and display monitors. For printers, we collected data on speed of printers measured
as pages per minute (PPM) in addition to print resolution. For monitors we collected
data on screen size measured in diagonal inches in addition to resolution. Note that
our findings on shape, path, and crossing patterns are quite robust to the use of this
second dimension (see figures II-7a-b).
Figure 7: Multi-attribute performance
Figure II- 7a
Performance of Desktop Printer
Technologies on Speed
Figure II- 7b
Performance of Display Monitor
Technologies on Screen Size
0
5
10
15
20
25
30
1978 1981 1984 1987 1990 1993 1996 1999 2002
Pages per minute
Dot Matrix
InkJet
Laser
Thermal
0
10
20
30
40
50
60
70
1987 1989 1991 1993 1995 1997 1999 2001
Diagonal inches
CRT
LCD
PDP
OLED

Figure II- 7c
Multi-attribute Performance of
Desktop Printer Technologies
Figure II- 7d
Multi-attribute Performance of
Display Monitor Technologies

-3
-2
-1
0
1
2
3
4
1968 1971 1974 1977 1980 1983 1986 1989 1992 1995 1998 2001
DM
Inkjet
Laser
Thermal
-3
-2
-1
0
1
2
3
4
5
6
1971 1975 1979 1983 1988 1992 1996 2000
CRT
LCD
PDP
OLED

47
We also calculated standardized values of performance on each dimension
over the category for each platform, computed the sum of these standardized values
over all dimensions of performance, and then plotted the latter index by time (see
figures II-7c-d). The use of multiple dimensions simultaneously using this crude
index fails to yield any patterns that might support the theory of S-curves.
II.10 IMPLICATIONS
This study has several implications for managers.
First, using the S-curve to predict the performance of a technology is quite
risky and may be misleading for two reasons: One, most of the technologies do not
demonstrate an S-shaped performance curve. Two, several technologies show
multiple S-curves, suggesting that a technology can show fresh growth after a period
of slow or no improvement.
Second, the continuous emergence of new technologies and the steady
growth of most technologies suggest that relying on the status quo is deadly for any
firm. Moreover, technological progress is occurring at an ever-increasing pace. As
such, paranoia rather than complacency is healthy. Vigilance for the emergence of
new technologies coupled with efforts to improve the old technology can help an
incumbent sustain and advance its position or even preempt competitors.
Third, the present findings indicate that the attack from below remains a
viable threat. Many new technologies start by offering low performance but later
threaten old technologies by improving at a much faster rate. Incumbents are prone
to disregard these new technologies initially, because they often cater only to a niche

48
and not to the mass market. However, these niches can grow into mass markets
eventually replacing the old technology. Thus, this disregard of such new
technologies is particularly dangerous. Further, some new technologies can perform
better than old technologies even at the time of introduction.
Fourth, another threat to incumbents is the emergence of secondary
dimensions of competition. Old technologies may be completely vulnerable to these
dimensions. Faced with such threats, incumbents need research to identify
technological solutions to improve the value of the old technology as well as to
identify market segments that value the contributions of the old technology.
Alternatively, incumbents need to explore R&D options on multiple dimensions in
order to react appropriately to threats posed by entrants.
Fifth, first-mover advantages may not be lasting since entrants introduce even
more innovations than incumbent firms. However, we found that old technologies
demonstrated high levels of improvement even after being dormant and static for
many decades. In some cases, old technologies regained dominance in performance
even after being surpassed by a new technology. In contrast, a misplaced belief in the
theory of S-curves might become a self-fulfilling prophecy and lead to the premature
demise of an old technology.
II.11 LIMITATIONS AND DISCUSSIONS FOR FUTURE RESEARCH
This study has several limitations. First we had to limit our analysis to only
six categories due to the time-consuming nature and difficulty of data collection.
Second, our analysis of performance did not include cost to buyers. Third, we did not

49
incorporate sales of products based on each technology within a category. All of
these limitations are potential opportunities for future research. In addition, future
research may also explore whether the theory of S-curves applies at the sub-platform
level, why there are long periods of no improvement in performance, and how firms
should compete given the pattern of technological evolution.

50
III CHAPTER 3: PERFORMING HYPOTHESIS TESTS
ON THE SHAPE OF FUNCTIONAL DATA
2

III.1 ABSTRACT
We explore different approaches for performing hypothesis tests on the shape
of a mean function by developing general methodologies both for the, often
assumed, iid error structure case as well as for the more general case where the error
terms have an arbitrary covariance structure. The procedures work by testing for
patterns in the residuals after estimating the mean function and are extremely
computationally fast. In the iid case we fit a smooth function to the observed
residuals and then fit similar functions to the permuted residuals. Under the null
hypothesis that the curve comes from a particular functional shape, the permuted
residuals should have a similar distribution to the unpermuted ones. So the fitted
curves will have the same distribution thus allowing significance levels to be
computed very efficiently.
In the more general case when several curves are observed one can directly
estimate the covariance structure and incorporate this into the analysis. However,
when only one curve is observed we adopt a graphical approach where one plots the
p-value for differing levels of potential complexity in the covariance structure. This
allows one to judge the degree of deviation from the assumed null distribution. We

2
A paper coauthored with Gareth James based on this chapter was published as per
details below.
James Gareth and Ashish Sood (2005) “Performing Hypothesis Tests on the Shape of Functional
Data” Journal of Computational Statistics and Data Analysis, Vol. 50, Iss. 1.

51
demonstrate the power of these methods on an extensive set of simulations. We also
illustrate the approach on a data set of technology evolution curves which current
theory suggests should have an underlying S shape. The developed techniques have
wide potential applications in empirical testing of the shape of functional data.
III.2 INTRODUCTION
Suppose we observe curves or functions Y
1
(t), . . . , Y
N
(t), and wish to
perform hypothesis tests on the shape of their mean function, µ(t) = EY(t), when
either one, or multiple curves are observed. In this paper, we develop general
methodologies for performing these hypothesis tests.
First, we describe an application of the method that we will discuss in this
paper. Consider the case of a manager faced with the issue of detecting common
patterns from prior information on the evolution of a technology, and devising
investment strategies for the future. The theory of S-curves from the technology
management literature (Foster 1986; Sahal 1981; Utterback 1994) suggests that all
technologies evolve in the shape of an S-curve, i.e. when plotted against time, the
performance of a technology exhibits an initial period of slow growth, followed by
one of fast growth culminating in a plateau. The manager would be interested in
determining whether their particular technology evolves as expected i.e. in the form
of an S-shaped curve. In Figure III-1, we plot the evolution of three technologies
overtime (Sood and Tellis 2005). Figure III-1(a) plots the evolution of incandescent
lighting over a 123 year period along with the best-fitting S-curve. Though the
monotonic growth in performance observed seems to roughly fit the expected S-

52
shaped growth, a number of deviations are observed, and it is unclear how best to
quantify the departure from the hypothesized shape. Similarly, Figures III-1(b) and
(c), plot the evolution curves for cathode ray tube (CRT) display monitors and dot-
matrix printers. While these plots also indicate some departures from the S-curve, the
disparities do not appear to be nearly as large as those for incandescent lighting. In
addition, the curves have been observed over significantly shorter period of time,
making it more difficult to assess whether the deviations are statistically significant.
This example highlights the need to develop techniques to perform formal hypothesis
tests on the shape of functional data. Some other examples where functional
hypothesis testing is useful include testing the average shape of human growth
curves, product life cycle, sales growth, and new returns or even weather patterns.
We develop a set of methods in this paper that allow one to answer these questions
with a minimal set of general assumptions.
Figure 8: Plots of the evolutions of three different technologies over time along with the
best fitting “S” curves (dashed lines).

53
In the standard finite dimensional setting performing a hypothesis test on the
mean of a population is a well-studied problem. However, the infinite dimensional
functional case possesses additional difficulties. Dealing with simple hypotheses,
where the null is completely specified, is comparatively easier to handle and offers
many alternatives like Pearson’s χ
2
test of goodness of fit (refer Darling 1957 and
Johnson and Kotz 1970 for a survey of this field). However, the more common
situation involving composite hypotheses, where some or all of the parameters of the
functional form are left unspecified is more challenging, and the existing methods
pose many limitations. One option is to modify the classical Pearson χ
2
test of
goodness-of-fit to this problem by estimating the unspecified parameters. However,
the assumed asymptotic χ
2
distribution of the test statistic becomes questionable
when the sample size is small. Other options like using the Kolgomorov–Smirnov
test statistic pose additional computational difficulties because of the use of Monte-
Carlo techniques (Srinivasan, 1970). Moreover, it may not be possible to adapt the
test to some functional forms. Past research has also proposed semi-parametric forms
of tests for such hypotheses (Robinson, 1989). The advantage of these tests is that
they can be extended to cases where the functional form is unspecified even under
the null, and hence are quite general in nature. However, the tests are dependent on a
restrictive set of assumptions like stationarity of time series data, and pose serious
limitations like lack of parametric estimates, dimensionality problems and sample
size problems. Still other researchers have approached the problem from a Bayesian
perspective (Bewley and Griffiths 2001; Montgomery and Bradlow 1999). While

54
these methods allow the inclusion of prior information especially when the sample
size is small, it is difficult to assess the confidence that may be placed on the
assumed model for the observed data.
When dealing with functional data, one must generally use a finite number of
observations to try to make inferences about an infinite dimensional curve. One must
necessarily assume some form of finite dimensional representation for the curves.
Often this is achieved by placing a smoothness constraint on each curve (James and
Silverman, 2005). A significant additional difficulty with testing the mean of
functions such as those in Figure III-1 involves the estimation of a covariance
structure. One option often taken in the literature involves assuming iid measurement
errors at the observed time points, e.g. Bewley and Griffiths (2001). This assumption
simplifies the problem considerably and we begin by exploring this model. However,
often in practice a more realistic model needs to allow for correlations in the error
terms. We first develop an approach that assumes a general covariance structure and
is applicable when several curves are observed from the same mean function. We
then extend this approach to the more difficult, but rather common, situation where
only one curve is observed. All three methods utilize permutation tests on the
residuals between the observed curve and the estimated mean function. The iid
approach computes a test statistic on the observed residuals and then permutes the
residuals multiple times to compute a sample distribution under the null hypothesis.
The general correlation approach works in a somewhat similar manner but makes
adjustments for the estimated covariance structure. Our methods address the

55
common difficult of computational feasibility with functional data, which often
contain a large number of observations. All three methods are extremely fast but
have also demonstrated high power levels in simulation studies.
In Section III-3, we develop our methods for performing hypothesis tests on
µ(t). The first involves the situation where errors are assumed to be iid at the
observation times. We call this approach “Functional IID Tests of Shape” or FITS.
The other two methods allow for a far more general covariance structure in the errors
and are called “Functional Arbitrary Covariance Tests of Shape” or FACTS.
Different procedures are required in the cases where multiple curves are observed
(FACTS
N
) and the more difficult situation where only one curve is sampled
(FACTS
1
). The rest of the paper is organized as follows. Section III-4 provides a
detailed simulation study where we evaluate the performance of these methods under
a number of scenarios. In particular, we examine the significance levels of the
methods under the null hypothesis and their power under alternatives with varying
levels of signal. These methods are then used to perform an empirical study on a data
set of 20 technology evolution curves (Sood and Tellis, 2005) in Section III-5.
Finally, a general discussion of the methodology and their potential future extensions
is provided in Section III-6.
III.3 METHODOLOGY
Let Y
1
(t), . . . ,Y
N
(t) represent the observed values of N separate curves where
the i
th
curve is measured at times t
i1
, . . . , t
ini
. Suppose we wish to test the null
hypothesis

56
H
0
: E Y(t) = µ
0
(t), for all t
versus the alternative
H
A
: E Y(t) ≠ µ
0
(t), for at least one t.
Then we begin by modeling each curve using
Y
i
(t) = µ(t)+ ε
i
(t), i = 1, . . . ,N (1)
where E ε (t) = 0. (1)
In Section III-3.1, we explore the FITS methodology for the simpler situation
where the ε
i
’s are taken to be uncorrelated measurement errors so that
t s if
t s if
t s Cov
≠
=
⎭
⎬
⎫
⎩
⎨
⎧
=
,
,
0
)) ( ), ( (
2
σ
ε ε (2)
We then extend these methodologies to the more general FACTS situation
where the ε
i
’s are assumed to consist of a combination of smooth systematic
deviations from the mean function and uncorrelated measurement errors. In this case
t s if
t s if
t s t s Cov
≠
=
⎭
⎬
⎫
⎩
⎨
⎧
+ Γ =
,
,
0
) , ( )) ( ), ( (
2
σ
ε ε (4)
where ) , ( t s Γ is a suitably smooth function of s and t.
It should be noted that none of the algorithms in this paper require that the
curves be measured at the same time points.
III.3.1 Functional iid Tests of Shape (FITS)
Let Y
ij
=Y
i
(t
ij
), µ
ij
= µ
i
(t
ij
) and e
ij
= ε
i
(t
ij
). Then under the restricted covariance
structure given by (2), we can express (1) in the form
Y
ij
= µ
ij
+ e
ij
i = 1, . . . ,N (5)

57
where the e
ij
’s are iid with E(e
ij
)=0 and Var(e
ij
)= σ
2
. Our approach to
performing a hypothesis test is based on the following heuristic argument. Let
ij ij ij
Y r
0
ˆ µ − = be the residual at time t
ij
where
ij 0
ˆ µ is the best estimate for µ(t
ij
) under
the null hypothesis. Let θ
T
t b t s ) ( ) ( = be a q-dimensional function where b(t) the
basis function and θ are the corresponding coefficients. We utilize a cubic B-spline
basis which partitions the interval into a series of cubic functions joined at “knot
points” (Hastie et al., 2001, Chapter 5). However, in general we have found these
types of methods to be robust to the exact choice of basis function. One can use any
finite dimensional basis provided it contains enough flexibility to model whatever
pattern the residuals may display. Suppose that we choose s(t) so as to provide the
best fit, in the least-squares sense, to the residuals, r
ij
. Then under the null hypothesis
it should be the case that s( t) ≈0 or correspondingly that
∑
≈ = 0
2
l
T θ . In addition,
under the null hypothesis, the residuals will be approximately iid. Hence if we
permute the ordering of the residuals, refit s (t) and compute the sum of the squared
basis coefficients, T  (b)
, then it should be the case that T and T  (b)
will have
approximately the same distribution. Alternatively, if the null hypothesis is false
there will be a systematic pattern in the residuals, causing T to increase but having
less effect on T  (b)
 because of the randomizing of the residuals. Hence we reject H
0
if
T is larger than a significant majority of the T  (b)
’s. Below we formally outline the
algorithm.

58
III.3.1.1 FITS algorithm
∑
∑
∑
=
=
=
−
≤
=
=
= =
=
B
b
b
q
l
b
l
q
l
l
T
T
T T I
r X X
1
) (
(B) (1)
1
2
) ( (b)
(b)
ij
1
2
ij
1 T T
oij ij ij
0
). ( (1/B) to s correspond value - p estimated The . 8
. T ,..., T obtain to 6 and 5 steps Repeat . 7
ˆ
T Compute . 6
.
ˆ
obtain to residuals permuted
the to s(t) refit and , s ' r the randomize curves, all over residuals the Pool 5.
ˆ
T statistic test the Compute 4.
. ) b(t to ing correspond rows h matrix wit basis a is X
where , ) (X
ˆ
squares least via residuals the to b(t) s(t) Fit 3.
. ˆ - Y r residuals the Compute 2.
. hypothesis null under the of estimate squares least the , ˆ Compute 1.
θ
θ
θ
θ θ
µ
µ µ

Note that the FITS algorithm can be applied to either a single curve or to
multiple curves. Figure III-2 illustrates the various components of this procedure,
using a null hypothesis that µ (t) corresponds to a logistic curve, on three different
simulated data sets. The first column relates to data generated from the hypothesized
logistic curve plus iid error sampled at 200 time points. Figure III-2(a) plots the raw
data along with the true mean that they were generated from. Figure III-2(b) shows a
plot of the residuals. The grey line through the center is the least squares fit of a
smooth regression spline to the residuals while the ten black lines are the
corresponding fits resulting from permuting the ordering of the residuals ten times.
Notice that the curves are all close to zero. Figure III-2(c) plots the values of T  (b)

generated from 500 random permutations of the residuals along with a grey
horizontal line which corresponds to the observed value of T. As we would hope, T
is near the middle of the T  (b)
’s with a p-value of approximately 70%. There is no

59
evidence here to reject the null hypothesis. The second column contains data that has
been very slightly perturbed from a logistic curve between times 1 and 20. Even
though the perturbation is slight, the fitted spline in Figure III-2(e) contains
significantly more pattern than most of those generated from the permuted residuals.
This difference is evident in Figure III-2(f) where T is significantly increased from
Figure III-2(c) giving a p-value around 5%. Finally, the last column further increases
the perturbation, which induces a value of T larger than all 500 T  (b)
’s. Notice that the
perturbations have no noticeable effect on the T (b)
’s but cause T to rise rapidly.
Ultimately the usefulness of a hypothesis test depends on its ability to correctly
specify the significance level under the null hypothesis and to produce high power
under the alternative. In Section III-4, we demonstrate that the FITS method fulfills
both criteria.
Figure III 2: Plots of FITS methodology applied to three different simulated data sets

60
III.3.1.2 Adjustment for correlation in residuals
The FITS algorithm does not account for the correlation in the residuals that
may exist even if the underlying error terms are independent. For example, in a
standard linear regression the residuals are generally correlated with each other as a
consequence of the least squares fitting procedure. Hence, the permuted and
unpermuted residuals may have slightly different distributions even if the null
hypothesis is correct. A simple addition can be made to the algorithm to adjust for
this effect. We replace step 8 in the FITS algorithm by the following:
1. Generate pseudo errors according to an iid Gaussian distribution with
mean zero and standard deviation estimated from the observed residuals.
2. Fit a simple linear regression to the pseudo errors and record the resulting
pseudo residuals.
3. Calculate the test statistic T
p
based on fitting a spline to the pseudo
residuals.
4. Randomize the pseudo residuals to calculate T
p
(b)
and record the
difference d
(b)
= T
p
- T
p
(b)

5. Repeat steps 8a. through 8d B times and let
∑
=
b
b
B d d /
) (

6. The estimated p-value corresponds to
∑
=
≤ −
B
b
b
T d T I B
1
) (
) ( ) / 1 (
This addition estimates the average increase in T over T  (b)
i.e. d that one
might expect as a result of correlation in the residuals. The observed value of T is
then reduced by d to adjust for any correlation effect. We induce correlation in the
residuals using a simple linear regression because it can be computed extremely

61
efficiently. However, one could also generate pseudo observations and re-estimate
the entire mean curve to produce the residuals. In practice we have found that the
correlation effect is too low to cause problems unless the function is sparsely
observed.
III.3.2 Functional arbitrary covariance tests of shape: multiple curves
(FACTSN)
The iid error model given by (2) provides a simple and often effective means
to test for a given mean structure. However, an iid assumption may be unrealistic in a
number of real world situations. A few examples include the evolution of
technologies as discussed in the introduction, product life cycle, sales growth and
new product diffusion curves as well as human growth curves. In this Section III.3.2
we develop an approach that assumes a general covariance structure and to estimate
the covariance effect. In Section III-3.3 we deal with the situation where only one
curve is observed. Let γ
1
(t), . . . , γ
N
(t) be a sequence of random curves with E γ(t)=0
and Cov( γ(s), γ(t)) = Γ(s, t). Then under the general covariance structure given by
(3), we can express (1) in the form
Y
ij
= µ
ij
+ γ
ij
+ e
ij
i = 1, . . . ,N
where γ
ij
= γ
ij
(t
ij
) and the e
ij
’s are iid with E e
ij
= 0 and Var(e
ij
) = σ
2
. When
using this model, the FITS approach where the residuals are permuted makes little
sense because even under the null hypothesis the covariance structure is not
preserved. Instead we use an alternative test statistic, namely
dt t t T
2
0 0
)) ( ˆ ) ( ( µ µ
∫
− =

62
where ) ( ˆ
0
t µ is the least-squares estimate of µ under the null hypothesis and
) ( ˆ t µ is also a smooth estimate of µ but without the constraint of the null hypothesis
on its shape. To test the significance of T we use a bootstrap approach (Efron and
Tibshirani, 1993). We first estimate the γ
i
’s and compute
ij ij ij ij
Y r γ µ ˆ ˆ − − = . To
preserve the covariance structure in our bootstrapped sample we then take a sample
) ( ) (
1
ˆ ,..., ˆ
b
N
b
γ γ , permute the residuals, r*
ij
and let
* ) ( ) (
ˆ ˆ
ij
b
ij ij
b
ij
r Y + + = γ µ . Under the null
hypothesis the Y
(b)
ij
should have a similar distribution to the observed data. The
bootstrapped test statistics are then computed using
dt t t T
b b 2 ) ( ) (
)) ( ˆ ) ( ˆ ( µ µ
∫
− =
s t
b
' Y on based of estimate smooth a is ) ( ˆ where
(b) ) (
µ µ . Under the null
hypothesis, ) ( ˆ t µ will be similar to ) ( ˆ
0
t µ and T
(b)
will have approximately the same
distribution as T. Finally, we compare T
(1)
, . . . ,T
(B)
to obtain a p-value. Below we
provide details of the algorithm.

63
III.3.2.1 FACTS
N
algorithm
∑
∫
∫
=
≤
− =
+ + =
…
= =
=
=
− =
…
…
B
b
b
b b b
b
ij
i
ij i ij
i
T T I
s
t dt t t T
t
dt t t T
1
) (
(B) (1)
ij
(b)
) ( 2 ) ( ) (
*
ij
) (
ij
(b)
ij
*
ij
(b)
N
(b)
1
ij ij 0 ij ij
ij ij
ij ij
2
0
0 N 1
N 1
0
) (
B
1
to s correspond value - p estimated The 11.
T ,..., T get to times B 9 to 6 steps Repeat 10.
' Y the
fit to smooth a is ) ( ˆ where )) ( ˆ ) ( ˆ ( Calculate 9.
r ˆ µ ˆ Y Let 8.
r produce to residuals the Permute 7.
ˆ ˆ sample
ed bootstrapp the produce t to replacemen with s ’ ˆ the Resample 6.
) ( ˆ where ˆ - µ ˆ - Y r residuals the Estimate 5.
) (t µ ˆ µ ˆ where
1...N i for µ ˆ - Y to fits smooth using (t) ˆ functions the Estimate 4.
)) ( ˆ ) ( ˆ ( Calculate 3.
. µ ˆ calculate
to assumed not is H that Note . Y Y fit to smooth a using µ ˆ Estimate 2.
Y Y curves N all using
, hypothesis null under the µ of estimate squares least the µ ˆ Compute 1.
µ µ µ
γ
γ γ
γ
γ γ γ
γ
µ µ

Note that in step 9 we use ) ( ˆ t µ rather than ) ( ˆ
0
t µ because otherwise when H
0

is false T
(b)
would still have a similar value to T and hence the test would have no
power. To implement this algorithm in practice we approximate the integrals in steps
3 and 9 by evaluating the integrands over a fine grid of time points ranging from the
smallest to the largest t under consideration. Both ) ( ˆ t µ and the s t
i
)' ( ˆ γ are estimated
using smoothing splines (Hastie et al., 2001, chapter 5) which are high dimensional
cubic splines that have been regularized using a penalty term on the squared second
derivative.

64
Figure 9: Plots of FACTS
N
methodology applied to three different simulated data sets.

Figure III-3 illustrates this approach on three simulated data sets with non
independent covariance structures. Each data set contains ten curves sampled from a
common mean function plus curve specific γ
i
’s and iid measurement error. All three
data sets have the same covariance structure. The curves in Figure III-3(a) were
generated from a common logistic mean function which is shown in grey. The
covariance structure induced by the γ
i
’s means that each of the individual curves
differs in shape from a logistic. However, the average over all curves does have the
correct form so the null hypothesis is satisfied for this data. Figure III-3(b) illustrates
the value of the test statistic in grey along with 500 values of T
(b)
. As we would hope,
given the null hypothesis is true, there is no evidence to reject H
0
with a p-value of
only 0.414. Alternatively, Figure III-3(c) illustrates data generated from a mean

65
function that has been slightly perturbed from a logistic, as shown by the grey curve.
Here the null hypothesis is false but the signal is fairly low relative to the covariance
in the γ
i
’s. Despite the low signal to noise ratio the test statistic, T, illustrated by the
grey line in Figure III-3(d), has increased markedly and the corresponding p-value
has declined to 0.05. Finally, Figures III-3(e) and (f) illustrate a slightly higher signal
to noise ratio. The p-value for this data is zero. Notice that in all three cases the
distribution of the T
(b)
’s is very similar but the value of T increases with larger
departures from the logistic mean curve.
III.3.3 Functional arbitrary covariance tests of shape: one curve
(FACTS
1
)
The methodology outlined in the previous section works well when there are
enough curves that one can reasonably hope to produce an accurate estimate of the
underlying covariance structure using a bootstrap analysis. However, for certain
types of data one may wish to infer the underlying shape of the mean function based
on observing only one curve. An example of this situation is the technology
evolution curves illustrated in Figure III-1. In this case a bootstrap analysis of the γ
i
’s
must fail because one can estimate at most one covariance term. One option involves
the approach taken in Section III-3.1 where we assumed an iid covariance structure
but, as mentioned previously, such an assumption may be overly simplistic. Of
course when only one curve is observed there is a fundamental identifiability
problem between the mean curve, µ, and the covariance, γ. Consider for example the
data in Figure III-2(g). This data clearly suggests a deviation from a logistic curve

66
for the early time periods. However, based on one curve we cannot identify whether
this effect is caused by the mean curve or whether it is simply a result of the
covariance structure as in Figure III-3. In other words, if we had observed multiple
curves perhaps the difference from a logistic curve would have evened out.
Nevertheless, one can still ask “How complicated would the covariance
structure need to be to explain this large a deviation from a logistic curve?” For
example, it may be the case that for a given curve using, for example, the iid FITS
methodology it is possible to reject a logistic curve hypothesis. But that after
including a very simple covariance structure in the model it is no longer possible to
reject the null hypothesis. In this situation one could not state with a high level of
confidence that the mean curve differed from a logistic. Alternatively, if you were
still able to reject the null hypothesis even after allowing for a complicated
covariance structure one would state with more confidence that the mean curve
seemed to differ from a logistic.
We use the following approach to implement this strategy. We first compute
ˆµ0, the least squares estimate of µ under H
0
. We then obtain a “low complexity”
estimate of the covariance by fitting a smooth function, γ, with a given “equivalent
degrees of freedom” (edf) (Green and Silverman, 1994) to
j j
Y
0
ˆ µ − . (Note we have
dropped the subscript i because here N = 1.) The concept of edf was developed to
measure the flexibility of a given functional form. It is easy to measure the flexibility
of simple functions such as an n
th
degree polynomial which has n+1 parameters and
hence n+1 degrees of freedom (dof). In this case the edf and the dof coincide.

67
However, directly calculating dof of more complicated functions such as a
smoothing spline (which involves many parameters but also a penalty term to ensure
smoothness) is more difficult. In these cases we use edf which gives the same answer
as dof in simple cases but generalizes to more complicated situations. We calculate
the residuals by subtracting the estimated values of µ and γ from the observed data.
Then we calculate a p-value by using the FITS procedure applied to the residuals we
have just computed. This entire procedure is repeated for increasingly complex
covariance structures. Finally, we plot the different p-values versus the equivalent
degrees of freedom of γ. A curve that has low p-values even for high degrees of
freedom in γ strongly suggests a true deviation from the null hypothesis. However,
one that does not have any low p-values or only low values for very simple γ curves
provides no significant evidence of a deviation from the null hypothesis. Below we
provide details of the algorithm.
III.3.3.1 FACTS
1
Algorithm
1. Compute
o
µ ˆ , the least squares estimate of µ under the null hypothesis.
2. Select a possible range of equivalent degrees of freedom for γ, edf
1
,…, edf
K
.
3. Estimate
) (
ˆ
l
edf
γ by fitting a smooth function with equivalent degrees of freedom
no greater than edf
l
to
j j
Y
0
ˆ µ − .
4. Compute the residuals
) (
0
ˆ ˆ
l
edf
j j j
Y r γ µ − − =
5. Perform steps 3-8 of the FITS algorithm to obtain the p-value, p
l
.
6. Repeat steps 3-5 for l = 1, . . . , K to obtain p
1
, . . . , p
K
.

68
7. Plot p versus edf.
As with the FACTS
N
algorithm we estimate γ ˆ using a smoothing spline. We
illustrate the algorithm on three simulated data sets in Figure III-4. The data in
Figure 4(a) was generated from a logistic curve and10 iid errors. In Figure III-4(b)
we have plotted the p-values as a function of the various different equivalent degrees
of freedom (edf) used to generate the γ curves. Here zero edf corresponds to the
FITS procedure from Section III-3.1 while two edf corresponds to a fit involving a γ
curve that is restricted to be linear. As we would hope, given that this data was
generated from a logistic curve with iid errors, there is no evidence to reject the null
hypothesis for any value of edf. The data in Figure III-4(c) is generated from a mean
curve with a slight perturbation from a logistic. We now see in Figure III-4(d) that
there is clear evidence to reject the null hypothesis when edf equals zero (the iid
case). However, the p-value increases rapidly even with relatively little flexibility in
the γ curve. The grey line in Figure III-4(c) illustrates the mean curve plus the least
flexible γ such that we can no longer reject the null at the 5% significance level.
Notice that only a very slight adjustment to the mean curve is required to explain the
deviation from a logistic. This all suggests that even a relatively simple covariance
structure accounts for the observed effect. In other words the evidence to reject a
logistic distribution is weak. Finally, the data in Figure III-4(e) exhibits significant
deviation from a logistic function. Correspondingly, Figure III-4(f) shows very low
p-values even for complicated covariance structures which suggest strong evidence
that the data does not have a logistic mean function.

69
Figure 10: Three different simulated data sets and their corresponding p-value curves
for different levels of flexibility in the covariance structure. The black lines in (a),
(c) and (e) correspond to the best fitting logistic function. The grey lines illustrate the
least flexible γ curve that would explain the deviations from a logistic curve.

III.4 SIMULATION STUDY
In this section we present results from an extensive set of simulations
designed to test the true significance levels, under the null, and power, under the
alternative, of the three methodologies from Section III-3.
III.4.1.1 FITS Results
For the FITS simulation we generated data sets using the iid error distribution
illustrated in Figure III-2. These data come from a mean function which is a logistic
curve of the form
) ( exp( 1
) (
d t c
b
a t
L
− − +
+ = µ

70
where a = b = 1,c = 0.5 and d = 50 plus a perturbation of the form At(20-t),
0 ≤t ≤20. For A = 0 this gives a pure logistic while as A increases, the deviation from
a logistic curve becomes more pronounced. In addition Gaussian noise with a
standard deviation of 0.05 was added to each measurement. A total of 200
observations from 100 equally spaced time points between 1 and 100 were generated
for each data set. To perform the simulation we chose a range of values for A starting
at zero. For each value of A, 200 data sets were generated. We then applied the FITS
procedure from Section III-3.1 to each of the data sets to calculate a p-value. An
important part of this approach is the choice of the basis for s(t), the curve that is fit
to the residuals. We used three different B-spline bases with five, eight and fourteen
degrees of freedom. Larger degrees of freedom correspond to more knots and hence
greater flexibility.
The results are summarized in Figure III-5. Figures III-5(a)-(c) are plots of
the proportion of p-values less than 1%, 5% and 10% respectively as a function of
the signal to noise (SN) ratio. We calculated the SN ratio by taking the average
deviation over time of the mean curve from a logistic function divided by the
standard deviation of the error terms after adjusting for the fact that there are
multiple observations at each time. A signal to noise ratio of zero corresponds to the
null hypothesis. So, ideally the curves should respectively take on values around 1%,
5% or 10% at this point but then rapidly increase towards one for larger ratios. In
fact we see that, for all three basis choices, at zero the methodology is if anything
producing slightly conservative p-values.

71
Figure 11: Plots of power as a function of signal to noise ratio using the FITS
methodology.
(a)–(c) respectively illustrate 1%, 5% and 10% significance levels corresponding to
fits using bases with 5 (—-), 8 ( −−), and 14 (· · ·) degrees of freedom. (d) illustrates
the average power for the 1% (—-), 5% ( −−), and 10%(· · ·) significance levels. Grey
horizontal lines plot the corresponding 1%, 5% or 10% values.

However, the power increases rapidly as the signal increases with, for
example, the power with a basis of five degrees of freedom and significance level of
5% reaching 1 with a signal to noise ratio of only 0.4. All three bases exhibit a
similar pattern but it is clear that the basis with five degrees of freedom dominates
those with eight and fourteen. It is interesting to note that the increased power is not
at the expense of a mis-specified significance level under the null. The other two
bases gave similar performance to each other. Figure III-5(d) plots the average power
over the three bases for each of the three different significance levels to provide a
direct comparison of the effect of significance on power.
III.4.1.2 FACTS
N
Results
To test the performance of the FACTS
N
approach we generated data from the
distribution illustrated in Figure III-3. The data consisted of a logistic curve given by
(6) with a = b = 1, c = 0.2 and d = 50 plus perturbations of the form At(20-t), 0 ≤t ≤20

72
and -At(100-t), 81 ≤t ≤100. In addition Gaussian noise with standard deviation 0.01
was added to each measurement. For each data set we generated ten curves with a
different, random, value of A for each one. The A’s were generated from a Gaussian
distribution with mean m and a variance of one. Data sets with m = 0 corresponded
to the null hypothesis because the average over all curves had a logistic shape.
Alternatively, as m grew there was an increasing deviation from a logistic mean
curve. One key component of the FACTS
N
approach is the flexibility of the
unconstrained estimate of the mean function. We used smoothing splines with
equivalent degrees of freedom of eight, ten and twelve on each data set to test the
sensitivity of the results to this parameter. As with the previous simulation we
generated 200 data sets for each one of a range of values for m, beginning with m =
0.
The results are summarized in Figure III-6. Figures III-6(a)-(c) are plots of
the proportion of p-values less than 1%, 5% and 10% respectively as a function of
the signal to noise ratio. The SN ratio was calculated in a similar fashion to that for
the previous FITS simulation. Note that the SN ratio is measuring signal per curve
rather than total signal over all N curves. This makes the results comparable to those
for FITS and FACTS
1
. The first point we note is that there is very little difference
between the fits using eight, ten or twelve effective degrees of freedom with all three
curves almost identical. This is comforting because it suggests a low sensitivity of
the results to any reasonable fit for the mean function. Secondly, the true significance
levels under the null hypotheses, i.e. when the SN ratio is zero, are equal to the

73
nominal 1% significance level and slightly conservative for the 5% and 10% levels.
As with the FITS methodology considered in the previous simulation, there is a
steady increase in power as the SN ratio increases. It is noticeable that the FACTS
N

SN ratio needs to be approximately four times as large as the corresponding FITS SN
ratio to achieve the same level of power. This difference is not surprising given that
the general case requires estimation of the γ’s i.e. the covariance structure while the
iid case does not. However, it does provide useful information about the increased
strength in the signal that may be needed before one can remove the iid assumption
and still reasonably hope to obtain significant results. Finally, Figure III-6(d) gives a
direct comparison of the power levels for the three different significance levels.
Figure 12: Power plots using the FACTS
N
methodology.
(a)–(c), respectively, illustrate 1%, 5% and 10%significance levels corresponding to
fits using smoothing splines with equivalent degrees of freedom of 8 (—-), 10 ( −−),
and 12 (· · ). (d) illustrates the average power for the 1% (—-), 5% and 10% (· · ·)
significance levels.

III.4.1.3 FACTS
1
Results
Finally, we tested the performance of the FACTS
1
methodology where a
general covariance structure is assumed but only a single curve is observed. The data

74
sets were generated in an almost identical fashion to those illustrated in Figure III-4
and described for the previous simulation. The only differences were that one rather
than ten curves was generated, a standard deviation of 0.02 was used for the
measurement error and A was chosen as a fixed rather than random value. As in the
previous simulations, a grid of values for A was chosen and 200 data sets were
generated for each value. The FACTS
1
methodology was then applied to each data
set. As with the first simulation three different bases were tested for s(t)
corresponding to splines , with q = 5,8 and 14 degrees of freedom. For each data set
we recorded whether there was significant evidence to reject the null at the 1%,5%
and 10% significance levels as well as the least flexible γ such that we could no
longer reject the null at each significance level.
The results are summarized in Figure III-7. Figures III-7(a)-(c) respectively
give the fraction of datasets for which the null hypothesis was rejected for at least
one choice of edf on the γ’s for significance levels of 1%, 5% and 10%. As with the
other simulations, the observed significance levels are approximately the same as the
nominal ones and the power increases rapidly with the SN ratio. For any particular
SN ratio the power is slightly lower than that for FITS but considerably higher than
for FACTS
N
. In addition, for this data, the q = 5 and q = 8 curves both give similar
power while the q = 14 curve is clearly inferior. The relatively superior performance
of the methodology with q = 8 is likely a result of the added complexity of the
deviations of curves from a logistic in this data relative to that in the first simulation
(compare Figure III-2 with Figure III-4). This suggests that ideally one should use

75
more flexible bases for s(t) when the underlying curves have a more complicated
structure to them. Figure III-7(d) provides the average power level as a function of
SN ratio for the three different significance levels.
Figure 13: (a)–(d) Power plots using the FACTS
1
methodology
(e)–(h) corresponding plots of degrees of freedom. (a)–(c) and (e)–(g), respectively,
illustrate 1%, 5% and 10% significance levels corresponding to fits using bases with
5 (—-), 8 ( −−), and 14 (· · ·) degrees of freedom. (d) and (f) illustrate overall
averages for the 1% (—-), 5% ( −−), and 10% (· · ·) significance levels.

Figures III-7(e) through (h) provide identical information to (a) through (d)
except they plot the average minimum degrees of freedom for the γ curve, among all
data sets with a significant p-value, such that the null hypothesis was no longer
rejected for each significance level. For example, a value of ten would indicate that a
γ curve with at least ten degrees of freedom was required to explain the departure
from a logistic curve. Even after the power levels have reached 100% the minimum
degrees of freedom continue to rise with the SN ratio. Interestingly the highest
degrees of freedom are obtained with q = 8 rather than q = 5 providing further

76
evidence for the need for a somewhat more flexible spline s(t) when the underlying
curves are more complicated.
III.5 EMPIRICAL STUDY
In this section we revisit the application problem discussed in Section III-2
regarding testing whether technologies evolve in the shape of S-curves. Frequently,
in firms from high-technology industries, managers are involved in making decisions
regarding allocation of funds for research and development on various existing and
emerging technologies. In such cases, they need to understand the patterns of
technological evolution in order to predict the future growth of existing and new
technologies to maximize their competitive advantage and return on investments. As
stated in the introduction, theory from the technology management literature
suggests that all technologies evolve in the shape of an S-curve. If true, this theory
provides a good benchmark for estimating the ultimate decline of existing
technologies, and for predicting the growth pattern of emerging technologies.
However, in reality, some technologies seem to evolve through irregular step
improvements instead of a continuous S-shape, and the problem is to test whether an
observed pattern is from an S curve. Sood and Tellis (2005) test this theory on 20
technologies from six markets - external lighting, desktop memory, display monitors,
desktop printers, data transfer and analgesics. Using historical analysis techniques,
they plotted the improvement in performance of all these technologies from the year
they were first introduced in commercial applications. Their research cast doubt that
these curves evolved according to a single S curve, instead suggesting a series of

77
irregular step functions may more accurately reflect the true structure. However,
Sood and Tellis did not use formal hypothesis tests. Instead they computed the MSE
from a fit of a logistic function to a subset of the observations versus the MSE from a
fit to all the observations and concluded that an S curve fit better over a subset rather
than the full data if the latter value was greater. We are interested here in
reexamining their data using our more formal hypothesis testing methodology. A
logistic curve of the form given by (6) was used to model the S curve in Sood and
Tellis (2004) where the parameters a, b, c and d were estimated from the data using
non-linear least squares. We take the same approach so our null hypothesis is that
µ(t) =µ
L
(t) with an alternative that µ(t) ≠ µ
L
(t). Since a single curve was observed for
each technology we applied the FACTS
1
approach to each of the 20 technologies.
We wish to choose q so as to maximize the power of our hypothesis tests. In general,
as q increases one can detect finer departures from the null hypothesis and hence
increase the power. However, for curves sampled at a small number of time points
using a large value of q will tend to overfit the data and can actually decrease the
power. After some experimentation we opted to use q = 5 for technologies measured
at 30 or fewer points, q = 12 for those measured at 70 or more points and q = 8 for all
others.
Using this approach eight of the technologies showed significant departures
from an S curve at 1% or 5% significance levels. We have summarized the results
for these eight technologies in Table 1. The values in the table correspond to the
minimum degrees of freedom required in the γ curve, used to generate the covariance

78
structure, such that the null hypothesis was no longer rejected at the given
significance level. A value of NA indicates the null hypothesis was not rejected for
any level of degrees of freedom. Notice that certain markets appear to exhibit
stronger departures from an S-curve than others. Four of the five lighting
technologies and two of the three data transfer technologies showed strong evidence
that they did not come from an S curve. However, only one each of the display
monitor and printer technologies and none of the desktop memory or analgesics
technologies showed significant departures from an S-curve even at a 10%
significance level. The incandescent lighting shows the most significant departures
from the null hypothesis with p-values less than 1% for all edf up to 15.
Table III-1: Average minimum equivalent degrees of freedom for γ
required before we fail to reject the hypothesis of an S shaped mean curve for each
the eight technology evolution curves that showed significant departures from the
null hypothesis
p-value Market Technology
0.01 0.05 0.1 0.2
Incandescent 15.2 16.0 16.7 17.8
Arc-Discharge NA 4.6 6.4 7.5
Gas-Discharge 5.7 10.5 10.8 11.9
Lighting
LED 5.7 7.9 8.3 8.6
Display Monitors Cathode ray
tube
6.4 9.0 10.5 12.3
Printers Inkjet NA 4.0 5.0 5.9
Copper NA 2.0 4.9 6.0 Data Transfer
Wireless 5.3 6.0 7.2 7.5

Figure III-8(a) illustrates the incandescent lighting data. The grey line
indicates the least complicated γ curve for which we would no longer reject the null
hypothesis at the 1% significance level. Figure III-8(b) plots the corresponding p-

79
values vs. edf. Based on Figures III-8(a) and (b) there is strong evidence for a step
function shape rather than a single S curve. In comparison the arc-discharge lighting
data, illustrated in Figure III-8(c), while still significant for low edf, has p-values
which climb rapidly towards 1 beginning at about 5 edf (see Figure III-8(d)). This
indicates that the departures from an S curve could be explained using a moderately
simple covariance structure. We also provide the data for the CRT monitors, Figure
III-8(e), (moderate evidence) and dot-matrix printers, Figure III-8(g), (weak
evidence) for comparison.
Figure 14: Plots for (a, b) incandescent lighting, (c, d) arc-discharge lighting, (e, f)
cathode ray tube display monitors and (g, h) dot-matrix desktop printers.
The top row provides the technology evolution curves (black dots), the best fitting S
curve (black line) and the minimum covariance structure that would need to be
assumed to explain the deviation from an S curve (grey line). The second row gives
the corresponding plots of p-value vs. edf.

In comparing with the results of Sood and Tellis (2004) we find that twelve
of the twenty technologies produced identical conclusions. Seven of the remaining

80
eight technologies differed in that Sood and Tellis concluded that an S curve did not
fit the data while our FACTS
1
approach failed to reject this hypothesis. The
differences are not surprising since our formal hypothesis test requires a stronger
standard of evidence to reject the null than the more exploratory approach of Sood
and Tellis. Most of the data sets for which differences in the two methods were noted
contained relatively few years of observations, less than 30, making it more difficult
to prove statistical significance.
III.6 DISCUSSION
We have developed three alternative but related methods to perform formal
hypothesis tests on the shape of functional means. We see several advantages to our
approach over those suggested previously. First, it allows one to place different
levels of restriction on the covariance term. One can adopt either the FITS approach
which imposes an iid restriction on the covariance function but has high power even
for low signal to noise ratios or use the FACTS approach which imposes few
assumptions on the underlying model. Second, even though we have concentrated on
a null hypothesis of a logistic mean function, the methodology is completely general
to any hypothesized form. Third, the permutation procedure we use is extremely
computationally efficient which can be an important consideration when dealing
with high dimensional datasets. Finally, in the case of a single curve with a general
covariance structure, one can not only calculate statistical significance levels but also
observe how complicated the covariance structure would need to be to explain the
observed deviation from the null hypothesis.

81
There are several areas for future research. One involves the choice of q, the
dimension of the basis for s(t). The simulations in Section III-4 suggest that, while
any reasonable value for q gives the correct significance level, certain values provide
higher power than others. In general it seems that the optimal value of q increases
with the number of observations and complexity of the underlying curves. One
promising solution would be to perform the tests for several values of q and then
apply a correction to the p-value to account for multiple testing. In addition, we have
used regression splines both to model s(t) and γ. Potentially, one may wish to use an
automated procedure to adjust the basis that is used to maximize power for a given
data set. Another possible extension involves the p-value plots illustrated, for
example, in Figure III-8. Here the p-values are plotted as a function of the degrees of
freedom of γ which we are using as a proxy for the complexity of γ. However, there
are other possible ways to measure this quantity which may also provide meaningful
insights. Finally, we have only dealt with one dimensional functional data. However,
the methodology could be easily extended to perform tests on the shape of
multidimensional data. In this situation the computational advantages would become
even more pronounced.

82
IV CHAPTER 4: MARKET RETURNS TO
TECHNOLOGICAL INNOVATIONS DURING NEW
PRODUCT DEVELOPMENT
IV.1 ABSTRACT
Firms need good estimates of the market acceptance of technologically
innovative products and the expected returns from pursuing new technologies.
Traditional measures of market returns to innovation like sales, profits or market
share can not be used to estimate returns to products still under development phase.
However, stock market returns to announcements of technological innovations
reflect the discounted present value of firm’s net cash flows in response to
technological innovations, and may offer some estimates for such purpose.
The authors develop a comprehensive database of all announcements related
to technological innovations made by 52 firms in three industries over a period of
more than two decades to examine market returns to firms during new product
development projects before commercialization. They examine how market returns
vary by technological leadership, investment portfolios, research productivity and
communication strategies of firms, and age of technologies. Contrary to findings in
the prior literature, the results suggest that the technological leaders derive lower
returns than close followers unless they have a strong track record of developing and
announcing innovations. Market returns are higher for firms that focus on smaller
investment portfolios and demonstrate higher research productivity.

83
IV.2 INTRODUCTION
A good estimation of the marketability of technologically innovative products
and the expected returns of pursuing specific emerging technologies is critical for
two reasons. First, the basis of competition in many industries is technological
superiority (Sorescu, Chandy and Prabhu 2003), and firms continually invest in
costly research and development (R&D) projects in order to maintain their
competitive advantage. However, a relatively small percentage of all research and
development projects are profitable e.g. cumulative success rate in pharmaceutical
industry for FDA approval is less than 10%. Second, each new breakthrough or
discontinuous innovation has historically sparked a series of intangibles across
industries e.g. software, biotechnology and telecommunications. These innovations
alter existing industry dynamics and create opportunities for market dominance
(Chandy and Tellis 1998). Third, low intellectual property protection mechanisms in
most industries necessitate that firms maximize the returns of their investments as
early as possible.
While there is abundant research on market returns to innovation at launch of
new projects e.g. investments, alliances etc. (Das, Sen, Sengupta 1998) and outcomes
of these projects e.g. new products announcements (Chaney, Divenney and Winer
1991; Eddy and Saunders 1980) including a diverse array of firm strategic issues
during this time e.g. pre-announcements, delays in introduction, success/failure of
new products (Wooldridge and Snow 1990, Chan, Kensinger and Martin 1992), there
is no research on market returns to innovations during the new product development

84
phase which examines the impact of technological leadership, and effect on
competitors.
We aim to further our understanding of this important phenomenon in the
current study, and to answer the following questions:
• What are the market returns to firms during new product development projects
before commercialization?
• How do these returns vary for technological leaders vs. close followers?
• Do firms that focus on many technologies attract higher returns than firms that
focus on few technologies?
• What is the profile of firms with high market returns - in terms of size, newness
of technologies, research productivity, and communication strategy?
• What is the impact of innovation in one firm on the market returns of its
competitors / other firms in the industry?
• What are the implications for firm strategy, and marketing decisions?
The rest of the paper is organized as follows: The next two sections present
the theory and method. The last section discusses findings, limitations and
implications of the research. A brief description of the operating principles of
technologies in the three product categories is included in the Annexure A.
IV.3 THEORY
This section explains the theory of technological evolution and the market
response to information flow during the various development phase of new product
development. We then develop hypotheses about firm characteristics and strategies

85
towards innovation and their effect on the market return to technological
innovations.
IV.3.1 Evolution of Technologies
Traditional theory suggests that a new technology initially improves slowly
after its introduction. R&D efforts of a firm result in resolving various bottlenecks
and start yielding faster returns after a threshold. The technology then makes rapid
progress. Eventually the technology matures and progress occurs very slowly (Foster
1986; Utterback 1974). When plotted against time, the progress appears as an S-
shaped curve. However, recent historical analysis of Sood and Tellis (2005) suggest
that instead of a continuous and smooth S-shaped curve of evolution, most
technologies evolve through an irregular step function with long periods of no
growth interspersed with short periods of growth. This implies that improvements in
technology are sudden and sporadic. Moreover, the rate of technological change and
number of new technologies is increasing over the years. Hence, firms are frequently
faced with the issue of announcing interim outcomes of their efforts in improving the
technical performance of existing products.
IV.3.2 Definitions
We define a technological innovation as the outcome of all research
activities, either basic or applied, and all development activities of a firm during a
new product development project directed toward the production of new products but
prior to commercialization. Technological innovations include discovery of new
parts and materials, patents, development of prototypes and demo products used for

86
exhibitions. All these innovations are the outcome of research and development
activities of the firms but are still not ready for sale in markets. Examples are Sony’s
demonstration of a 50-inch prototype plasma display in an exhibition show in 2001
or development of a water-cooling system for computer chips by Cooligy in 2003. A
technological innovation may lead to an immediate improvement in technical
performance of products offered by the firm, or enable new product or process
competencies that enable new products to be developed or manufactured in future. A
technological innovation usually constitutes an intellectual property of the firm.
We define a technological announcement as release of information
regarding technological innovations. A technological announcement may be made
anytime during the product development project and provides advance and additional
information to the market.
IV.3.3 Market Returns to Technological Innovations
Prior research in technology management, marketing and strategy literature
report numerous benefits of pursuing innovation e.g. immediate financial gains
(Chaney, Devinney and Winer 1991), first mover advantages (Kerin, Varadarajan
and Peterson 1992) and long term competitive advantages (Lieberman and
Montgomery 1988). Various measures like improvement in sales, increase in
customer preference and increase in market share have been investigated. However,
it is difficult to establish a clear relationship between these measures and return to a
specific innovation.

87
We propose that announcements of technological innovations provide a more
accurate measure for assessing market return than either of these measures for three
reasons. First, the event date can be identified more accurately. Announcements are
generally made close to the date of the actual R&D achievement – hence reflect
market returns to only that event. Second, the announcement is more in control of the
firm and all the content of the innovation can be covered in one announcement. Third
the announcement can be made in journals, newspapers that are monitored by the
market increasing the reaction speed and accuracy.
The efficient market hypothesis (Fama 1969) states that at any given time,
security prices fully reflect all available information. Thus, new information is fully
incorporated into the price as soon as it is released. Considerable empirical research
has found substantial support for this hypotheses (refer Fama 1991 for a review).
Although recent research has also detected some exceptions to this hypothesis (De
Bondt and Thaler 1985; 1987), these exceptions seem to be minor and insignificant
(Malkiel 2003; Ross 1977; Fama 1991).
IV.3.4 Hypotheses
We derive hypotheses about key factors affecting the market returns to
technological innovations: technological leadership status, investment portfolio, size,
and research productivity of the firm and newness of technology. We also examine
the effect of strategic and tactical choices of the firm like frequency of
announcements on these returns (see Figure IV- 1).

88
IV.3.5 Technological Leader or Close Follower
Firms continually develop new products and make announcements related to
superior performance, features or uniqueness of these products at frequent intervals.
Sometimes these products are new to the customers and these firms are the first to
introduce such products in the industry, and other times the products have already
been introduced by other firms before the firms’ announcement. For example, firms
like Fujitsu, Intel and IBM have historically developed and introduced new products
based on state of the art technologies ahead of their competitors. On the other hand,
firms like Dell, Matsushita, AMD and Seagate have historically laid a greater focus
on incorporating the latest innovations in new products at a fast pace instead of being
the first to introduce these products.
Figure 15: Theoretical Model

Market Return to
Technological Innovations
Firm Characteristics
Competition Characteristics
Research
Productivity
Investment Portfolio
Technological
Leadership
Age of Technology
Technological
Announcement
Size
Frequency of
Announcements

89
Prior research has largely relied on order of market entry to identify
technological leaders (Lieberman and Montgomery 1988; Kerin, Varadarajan and
Peterson 1992; Szymanski, Troy and Bharadwaj 1995). However, identification of
true technological leader or first mover using this definition is problematic in many
industries where more than one firm closely vies for such spot and each firm
frequently introduces state-of-the-art products ahead of others. For example, Fujistu
and IBM have routinely competed on being the first to announce the hard disk drive
with higher maximum areal density. Hence, it is unclear who is the technological
leader is in such cases if we rely on order of entry as the definition.
In order to avoid this problem in identification of a technological leader, we
define technological leader as the first firm to announce a particular technological
innovation and a follower as the subsequent firm to announce the same innovation.
This definition of technological leadership is more suitable for the present research
questions for two reasons. First, it is based on outcomes of research and development
prior to commercialization rather than based on first entry in a market. Second, this
definition allows us to identify firms which introduce specific innovations ahead of
others irrespective of their order of entry. For example, Google was a late entrant in
the web search market but has consistently introduced innovative technologies ahead
of its rival in the last few years.
Technological leaders can gain sustainable competitive advantages through
learning and patents (Lieberman and Montgomery 1988). Kerin, Varadarajan and
Peterson (1992) suggest that the sources of first mover advantages accrue from

90
economic, strategic, technological and behavioral factors. Technological leaders
enjoy scale and experience economies, higher marketing cost asymmetries, ability to
acquire critical assets and factor inputs essential for production at a cost advantage,
and differentiation from product and process innovations (Robinson and Fornell
1985; Kerin, Varadarajan and Peterson 1992). They also benefit from switching costs
in consumers, acquisition of innovators segments leaving the largely skeptical
follower segments to followers, and information and consumption experience
asymmetries. Technological leader firms introduce cutting edge technologies and
innovative products (Chan, Kissinger and Martin 1992), and usually enjoy consumer
loyalty through preference formation, lock-in effects and dominance in niche
markets. Robinson, Fornell and Sullivan (1992) find a strong link between order of
entry, market share and returns on investment.
Hence, the benefits to a technological leadership strategy are substantial and
we expect market returns from technological leadership to be high and we propose:
IV.3.5.1 H
1
: Market returns for technological innovations are higher for
technological leaders than for subsequent followers.
IV.3.6 Investment Portfolio
Firms in most industries rely on more than one technology to develop
products and satisfy consumer demands. For example, in the display monitor
industry, firms offer products based on CRT, LCD, Plasma, OLED, and other
emerging technologies. The choice of investing in a small vs. large portfolio of
technologies is a strategic decision for any firm that reflects its strategy towards
managing capital costs, technological forecasting, risk preference and compatibility

91
with other competencies. These choices affect decision on resource allocation and
project selection (Cooper, Edgett, and Kleinschmidt 1999). Scarce and limited
resources like engineering, finance, and marketing need to be allocated among the
various technologies in order to optimize these investments.
We define the investment portfolio of a firm as the total number of
technologies that it invests in. A larger portfolio increases the firms’ capability to
offer a diverse set of products, reduces its overall risk, and increases options to offer
a wide range of products to their customers.
Findings from prior research support expectations of higher returns from a
larger investment portfolio. Robinson, Fornell and Sullivan (1992) found that a
broader product portfolio significantly increases the probability of survival of firms
and enhances their competitive advantages. Firms with a broad range of products can
selectively target the lucrative segments and leave less profitable segments to other
firms. Hence,
IV.3.6.1 H
2
: Market returns for firms with a large portfolio of technologies are
higher than for firms with small portfolio.
IV.3.7 Research Productivity
A firm introducing innovations at a frequency higher than the industry
average is perceived as successful / competent and attracts higher market returns.
The high pace of introduction of innovation has two main effects. First it affects
customer expectations. John, Weiss and Dutta (1999) demonstrate how customer
expectations about the pace, significance and certainty of improvements affect the
repeat purchase behavior for products in technological intensive markets. Second, it

92
also affects the overall momentum of innovation within the industry. Godoe (2000)
uses the telecommunications industry to demonstrate how a high R&D intensity
within that industry has complemented the absorption of both radical and
incremental innovations.
Good metrics for innovation are important for not only increasing the
efficiency of the innovation process within the firm but also to justify investments
and affect behavior of employees (Hauser and Zettelmeyer 1997). Some of the
measures used in prior research are based on inputs to a firm like annual R&D
budget as a percentage of annual sales, number of engineers in a firm or percentage
of total time devoted to R&D (Cohen and Levin 1989). Other researchers have
focused on measures based on processes within a firm to measure its innovativeness
like number of patents, citation counts of patents owned (Dutta and Weiss 1997) etc.
We define research productivity in terms of the frequency that a firm
introduces new innovations. High research productivity enables firms to make
frequent improvements to existing products as well as to introduce new products in
the market. A firm introducing innovations at a frequency higher than the industry
average is perceived as successful / competent and attracts higher market returns.
Hence, we expect
IV.3.7.1 H
3
: Market returns for firms with higher research productivity are higher
than returns to firms with lower research productivity.
IV.3.8 Size of Firms
It has been well established in technology management and strategy
literatures that higher resources of large firms enable economies of scale and scope

93
(Galbraith 1952; Comanor 1965) and increase the risk taking ability of the firms
(Cohen, Levin and Mowery 1987). Both affect the innovation capabilities of firms
positively. While some studies have reported small firms with higher research
productivity (Bound et al 1984), it is unclear whether the higher productivity is a
result of higher efficiency in converting inputs into outputs or just because of
underreporting R&D expenditures.
However, we expect market returns to technological innovations be higher
for larger firms for two reasons. First, large firms also have more opportunity and
resources for developing and introducing technological innovations (Ettlie and
Rubenstein 1987; Chandy and Tellis 2000). Second, large resources allow firms to
benefit more from these technological innovations through faster product
development, higher sales etc. Hence we expect,
IV.3.8.1 H
4
: Market returns for technological announcements are higher for large
firms than for small firms.
IV.3.9 Frequency of Announcements
Firms also vary in terms of their communications strategy and choose varying
levels of openness regarding their innovation projects. On one hand, some firms
adopt the taciturn approach and only announce significant advances and
achievements. On the other hand, some firms make regular, frequent announcements
about the progress of the projects. Similarly firms take different approaches to the
content of announcements – some aggregate many events in one announcement
while others make several separate announcements. Still others stress on certain

94
aspects of announcements as being more salient e.g. focus on later stages of NPD
projects rather than the initial stages.
While the impact of these diverse strategies on the market returns is not clear
and needs to be resolved empirically, we do know that the risk premium to new
information increases with the preciseness of the new information released Veronesi
(2000). We also know that from the efficient market hypothesis that markets react
only to new information and ignore announcements which repeat information
already released in prior announcements. Thus, the market returns to a mere increase
in number of announcements is expected to be null. Hence,
IV.3.9.1 H
5
: Market returns for firms with a high number of announcements are
equivalent to returns for firms with low frequency of announcements.
IV.3.10 Age of Technology
Older technologies usually attract numerous applications and improvements
in these technologies benefit many markets. But, technology management literature
also tells us that older technologies suffer from slower growth due many reasons like
technological inertia, maturity and reduced incentives (Foster 1986; Chandy and
Tellis 2000; Sahal 1981; Reinganum 1985). In contrast, new technologies have
higher long term potential. Hence we expect,
IV.3.10.1 H
6
: Market returns for new technologies are higher than returns for
old technologies.
IV.3.11 Effect of Innovation on Competitors
Technological change creates turmoil in the industry changing existing
competitive positions and spurring competition. Announcement of the start of a new

95
project, success in developing a working prototype or launch of a new product can
all affect other firms in the industry in many ways. Some innovations benefit many
firms in the industry by helping to set standards or to resolve bottlenecks related to
evolution of technologies. For example, the discovery of tri-phosphor materials for
florescent lamps or electronic ballasts for lamps were beneficial for all firms in the
industry since they improved the lighting efficacy of existing lamps and enabled
growth of an ancillary demand for these products. Other innovations like the word
processing computers were unfavorable in nature for the typewriter industry.
However, given the broad common goal of technological competition of enhancing
competitive advantage of a firm, we expect the overall effect of innovations to be
negative on the other firms in the industry. Hence we propose,
IV.3.11.1 H
7a
: The average abnormal return to all competitors on the day of
announcements is less than or equal to zero.
IV.3.11.2 H
7b
: The average difference in abnormal returns between firms that
announce and their competitors is greater than zero.
IV.4 METHOD
We created our own database of technological announcements using the
historical method (Golder and Tellis 1993; Golder 2000). One key advantage of
using the historical method is freedom from survival and self-report bias (Golder
2000). The next subsections detail our sample selection, sources, and procedure for
data collection.

96
IV.4.1 Sample selection
We used two criteria in selecting product categories: a reasonable number of
emerging technologies and data availability. We selected product categories where a
number of technologies have emerged in the last decade, and the key global players
are in the US markets. The first requirements is essential to ensure that we have a
large sample of announcements, and the second requirement is essential since we
require the firm to be listed on US stock markets in order to assess the market value.
While it is possible to conduct the same analysis on foreign market exchanges, the
availability of data, efficient stock markets and familiarity enhance the feasibility of
conducting the analyses in US markets. The present study goes further than previous
studies in one important aspect – we collected data on all major firms within each
industry and selected all technologies in that industry.
Table IV-1: Data Sample
Category External Lighting Display Monitors Desktop Monitors
Number of Firms 18 16 18
Number of
Announcements
141 159 210
Sample Period 1977-2003 1983-2003 1982-2003
5 5 5
Number/ Type of
Platform
Technologies

Incandescent, Arc-
discharge, Gas-
discharge, LED
and MED
CRT, LCD,
Plasma, Display
panels and OLED
Magnetic,
Magneto-optical
and Optical

97
On the basis of these criteria we chose three product categories - external
lighting, display monitors and computer memory. We identified 18 major firms in
the lighting industry listed on US exchanges and collected 141 announcements from
1977 till 2003 (see Table IV-1).
The sample comprises a total of 5 platform technologies in the lighting
industry – incandescent lamps, arc discharge lamps, gas-discharge lamps, microwave
electrodeless lamps and light emitting diodes. We also have all announcements of
innovations in fixtures and accessories required for these lamps. Similarly we
identified 16 major firms in the display monitor industry and collected 159
announcements from 1983 till 2003 (see Table IV-1). The sample comprises 5
platform technologies in the display monitor industry – cathode ray tube monitors
(CRT), liquid crystal display monitors (LCD), plasma display monitors (PDP) and
organic light emitting diode monitors (OLED). We identified 18 major firms in the
computer memory industry and collected 210 announcements from 1982 till 2003
(see Table IV-1). The sample comprises 3 platform technologies in the computer
memory industry – magnetic memory, magneto-optical memory and optical memory.
IV.4.2 Sources
Although many studies limit their focus on Wall Street Journal (WSJ) as the
only source of announcements, we feel the phenomenon of technological innovation
is covered by a much broader cross-section of sources, and limiting the source to
only one may not capture the date when information is first released to the markets.
Indeed, Glascock, Davidson and Henderson (1987) show how WSJ does not publish

98
all the news. They also show that there is a lag of 3 days between information about
bond rating between Moody’s and WSJ. Hence, in the interests of accuracy and
comprehensiveness, we include all sources of information.
The primary sources we include are Factiva, Lexis-Nexis and Company
websites for press releases/announcements regarding technological innovations. We
also included all newswire services like PR Newswire, Business Newswire and
Reuters. We collect company background information from General Business File
ASAP and Yahoo Finance, and data on stock market returns from WRDS.
IV.4.3 Procedure
For each selected industry, we identified the all major players in the industry,
and collected information on each player. We used a variety of key words to narrow
down on all the announcements e.g. (company name or ticker) and (display or
monitor or name of technology). We sorted the results based on oldest first to
identify the first release of information to the market. The following types of
announcements were deleted from the sample:
• Announcements made by firms whose data are not available from CRSP (firms
not traded on NYSE, AMEX or NASDAQ).
• Announcements appearing in non-daily publications because of inherent
inaccuracy of determining the exact date of release of information (Hendricks
and Singhal 1997).
• We do not include information on product reviews

99
• Repeat announcements appearing after the first announcement unless the content
of announcement is distinctly enhanced. When such an announcement was
identified, it was treated as a new announcement for the incremental information
contained in it.
We examined each announcement for incremental information contained in it
and discarded general information regarding the firm’s business operations that is
available prior to the announcement as well. We then classified the announcement as
per the classification system developed earlier. More specifically, we identified
whether the announcement contained any information regarding new material,
development of prototypes or demonstration models for exhibitions, granting of
patents. We examined whether the announcements contained any information about
improvement in performance, reliability, reduction in cost as a results of research
and development efforts of the firm.
Based on this information, we identified the firms that were first in
announcing the key innovations. We calculated the age of the firm and scope of
research and development efforts of the firm from the set of announcements. We also
calculated the age of technologies at the time of announcement, its absolute
performance and magnitude of jump in performance described in the announcement.
IV.4.4 Model
This section describes how market returns can be estimated through the
abnormal returns in the event window. Event study method is an attractive method
for investigating the magnitude of the effect that an unexpected event has on the

100
market returns of a firm. The selection of appropriate event is critical to the
measurement. We define the event as the actual day the market receives information
of technological innovation, or could have reasonably anticipated the news.
One of the most common approaches to estimating the market return to the
new information is also known as the residual approach. This involves modeling
returns to new information, also known as abnormal returns, as residuals from some
benchmark model of normal returns (Fama, French, Jensen and Roll 1969). Thus
estimation of the conditional return in absence of the new information is the first step
in the process. We used the market adjusted model (equation 1) which assumes a
linear relationship between the return of any security to the return of the market
portfolio, and the expected return of a firm is calculated from estimated linear
coefficients from a previous estimation period. The market adjusted model is
expected to provide the highest ability to detect effect effects because of its ability to
reduce the variance of the abnormal return by removing the variation in the market
return (MacKinlay 1997; Henderson 1990).
it mt i i it
R R ε β α + + =
… 1
where R
it
and R
mt
are the period t returns on security i and the market
portfolio respectively and ε
it
is the zero mean disturbance term. α
i
, and β
i
, are the
parameters of the model to be estimated from a pre-event estimation period. The
estimation window for equation 1 and 3 is a period preceding the event (refer Figure
IV- 2). Abnormal returns are the difference in observed returns of the firm during the

101
event window and the ‘normal’ returns and reflect the returns attributable to the
event.
Figure 16: Timeline of Event Study
T
0
T
1
T
2
T
3
0
Estimation
Window
Event
Window
Post-event
Window
T
0
T
1
T
2
T
3
0
Estimation
Window
Event
Window
Post-event
Window

For each firm i and event date t, we have
] [
it it it
R E R AR − = … 2

Where
AR
it
= abnormal return for stock i on day t
R
it
= actual return for stock i on day t
E[R
it
] = expected return for stock i on day t predicted by equation (3)
We calculate the cumulative abnormal return (CAR
i τ
) by cumulating the
abnormal returns for firm i over a period of time τ thus:
it
t
i
AR CAR
∑
=
=
τ
τ
0
… 3
These abnormal returns are then used as dependent variables in a cross
sectional multiple regression model to examine the source of the extra returns.

102
IV.5 RESULTS
We first present the findings on technological announcements in the three
product categories and the calculation of abnormal returns. We then present the
results of regression analysis for each of the hypotheses.
IV.5.1 Abnormal Returns
We calculated the abnormal returns using the model in equation 2. The
estimation period used was an average of 270 days. For some new firms which were
listed on the stock exchange for a short period before the announcements, the
estimation period was shorter. However, we removed any announcement where we
did not have an estimation period of at least of 100 days.
Table IV-2: Mean Abnormal return using OLS Market Model
External Lighting Display Monitors Desktop Memory
Event Day Mean t-val Mean t-val Mean t-val
-5 0.25% 1.07 -0.29% -0.68 0.25% 1.09
-4 -0.31% -1.34 0.42% 1.00 -0.30% -1.30
-3 0.05% 0.20 -0.24% -0.56 0.06% 0.27
-2 -0.05% -0.20 0.69% 1.64 -0.06% -0.24
-1 -0.17% -0.73 0.31% 0.73 -0.18% -0.77
0 0.93% 4.01 1.20% 2.85 0.88% 3.83
1 0.12% 0.52 0.07% 0.17 0.16% 0.70
2 0.20% 0.85 0.21% 0.49 0.18% 0.80
3 -0.33% -1.41 0.01% 0.03 -0.32% -1.37
4 0.06% 0.26 0.00% -0.01 0.06% 0.25
5 0.20% 0.88 -0.11% -0.26 0.21% 0.89
CAR(-1,+1) 0.88% 2.19 1.58% 2.16 0.87% 2.17
CAR(-3,+3) 0.75% 1.87 2.25% 3.08 0.74% 1.86

103
Table IV-2 presents the mean abnormal return (AR
t
) for the announcements
with t-statistic. Table IV-2 shows that the mean abnormal return is about 1% across
categories and significantly different from 0 on the event day. Plots of cumulative
average abnormal returns (CAAR) for these categories are in Figure IV- 3.
Figure 17: Effect of Announcement of Technological Innovations on Market Value
-0.50%
0.00%
0.50%
1.00%
1.50%
2.00%
2.50%
-5 -4 -3 -2 -1 0 1 2 3 4 5
Event Day
CAAR
External Lighting
Display Monitors
Desktop Memory

These findings are similar to Kelm, Narayanan and Pinches (1995), who
report 1.14% over a 2-day period to such announcements. We also measured the
CAAR over the three day period around the event day (-1, +1), and find significant
returns over the longer window as well. However the t-values are the highest on
event day indicating that the identification of the event day is sufficiently precise.
IV.5.2 Technological Leadership
H
1
predicts that market return to technological leaders are higher than those
to close followers. We measured technological leadership on an announcement level
as the first firm to announce a technological innovation. All other firms introducing

104
the same innovations were classified as close followers. However, in many cases,
followers also introduce some additional incremental improvements in the products.
We classified the firm as a follower if the key improvement in the product was not
this incremental innovation but the main innovation introduced earlier by another
firm.
In all three categories we find technological leaders attract lower returns than
close followers (see Table IV-3 and Figures IV-4a-4c). Plots of cumulative average
abnormal returns (CAAR) also demonstrate the difference in returns for these two
strategies. Hence, we fail to find support for H
1
that technological leaders have
higher returns than close followers.
Table IV-3: Mean Abnormal Returns to Technological Leadership vs. Close Followers
Technological Leader Close Follower
Category
Mean Return* N t-val Mean Return* N t-val
External Lighting 0.01% 41 0.03 1.79% 76 4.72
Display Monitors 1.21% 1162.33 1.61% 26 3.22
Desktop Memory 0.10% 73 0.36 1.00% 90 2.78
* abnormal returns on event day
These findings are in contrast to findings of Miller, Gartner and Wilson
(1989) who report positive returns to first movers and lower returns to imitators.
However, Miller et al examine new product introduction announcements and not
announcements prior to commercialization. There are two reasons why market
returns to technological leaders might be lower for technological innovations. First,
the marketability of technological innovations is not evident at the time of innovation

105
and subsequent cases of other firms imitating similar innovations supports market
demand for such products. Second, the higher variability of stock prices closer to
first announcement of technological innovations reflect the higher risk or uncertainty
associated with its success.
Figure 18: CAAR Plots for Technological Leaders vs. Close Followers
-4.00%
-3.00%
-2.00%
-1.00%
0.00%
1.00%
2.00%
3.00%
-5 -4 -3 -2 -1 0 1 2 3 4 5
CAAR
tech_leader follower
-1.00%
-0.50%
0.00%
0.50%
1.00%
1.50%
2.00%
2.50%
3.00%
-5 -4 -3 -2 -1 0 1 2 3 4 5
CAAR
tech_leader follower
Figure IV- 4a: Lighting Category Figure IV- 4b: Monitors Category
-1.00%
-0.50%
0.00%
0.50%
1.00%
1.50%
2.00%
2.50%
3.00%
-5 -4 -3 -2 -1 0 1 2 3 4 5
CAAR
tech_leader follower

Figure IV- 4c: Memory Category
IV.5.3 Investment Portfolio
H
2
predicts that firms with a large portfolio of technologies have higher
returns than firms with small portfolio. We first calculated the investment portfolio
for each firms as the total number of technologies that a firm makes technological
announcements for, and then divided the total sample by a median split. We then

106
compare the abnormal returns to firms with small portfolio to those with large
portfolios.
Table IV-4: Mean Abnormal Returns to firms with Large vs. Small Investment
Portfolio
Small Portfolio Large Portfolio
Category
Mean Return* N t-val Mean Return* N t-val
External Lighting 2.12% 46 3.00 0.38% 95 1.68
Display Monitors 1.48% 123 3.04 0.24% 36 0.31
Desktop Memory 2.06% 58 3.38 0.49% 152 2.52
* abnormal returns on event day
Figure 19: CAAR Plots for Large vs. Small Investment Portfolio
-2.00%
-1.50%
-1.00%
-0.50%
0.00%
0.50%
1.00%
1.50%
2.00%
2.50%
3.00%
3.50%
-5 -4 -3 -2 -1 01 2345
CAAR
Small Portfolio
Large Portfolio
-1.50%
-1.00%
-0.50%
0.00%
0.50%
1.00%
1.50%
2.00%
2.50%
3.00%
3.50%
-5 -4 -3 -2 -1 0 1 2 3 4 5
CAAR
Small Portfolio
Large Portfolio
Figure IV- 5a: Lighting Category Figure IV- 5b: Monitors Category
-1.00%
-0.50%
0.00%
0.50%
1.00%
1.50%
2.00%
2.50%
3.00%
-5 -4 -3 -2 -1 0 1 2 3 4 5
CAAR
Small Portfolio
Large Portfolio

Figure IV- 5c: Memory Category

107
Table IV-5: Regression results for effect on Average Abnormal Return
Lighting Monitors Memory Pooled
est t p est t p est t p est t p
Intercept 0.029 1.03 0.310.006 0.18 0.860.061 4.29 0.00 0.041 4.20 0.00
Investment
portfolio
0.010 1.60 0.11 -0.003 -0.63 0.53 0.001 0.38 0.70 0.002 1.08 0.28
Research
productivity
0.074 4.54 0.00 -0.069 -2.68 0.01 -0.017 -0.77 0.44 0.034 3.60 0.00
Technological
Leadership
-0.016 -1.39 0.17 -0.007 -0.93 0.36 -0.002 -0.46 0.65 -0.009 -2.16 0.03
Size -0.008 -1.83 0.070.000 -0.060.95-0.004 -2.44 0.02 -0.003 -3.57 0.00
Age of
Technology
0.000 -0.36 0.72 0.000 -0.60 0.55 0.000 0.28 0.78 0.000 -1.43 0.15
Freq. of
announcements
0.001 0.20 0.85 0.008 2.37 0.02 0.000 0.06 0.95 0.001 0.85 0.40
New material -0.028 -1.60 0.11 0.002 0.25 0.80 -0.008 -1.34 0.18 -0.011 -2.12 0.03
Prototypes 0.002 0.16 0.88-0.009-1.280.20-0.012 -2.14 0.03 -0.008 -1.90 0.06
Exhibits -0.021 -1.82 0.07 0.009 1.18 0.24 -0.007 -1.47 0.14 -0.007 -1.76 0.08
Patents 0.015 0.48 0.63-0.006-0.420.680.000 -0.020.99 -0.001 -0.120.90
N 126 112 206 444
R
2
0.22 0.35 0.17 0.14

We find no support for this hypothesis. Table IV-4 and Figures IV-5a-5c
present a comparison of firms with large and small investment portfolios. Market
returns to firms with smaller portfolios were approximately 2% for all three
categories and approximately 0.5% for firms with larger portfolios. Again these

108
findings are against conventional wisdom advocating firms to develop broader
product lines (Robinson, Fornell and Sullivan 1998; Cooper, Edgett and
Kleinschmidt 1999).
Table IV-5 presents the results in a multiple regression framework. We also
control for characteristics of announcements and innovations e.g. new material, new
exhibits, prototypes and patents. There are negative returns to increasing investment
portfolio in display monitors and desktop memory categories. However the
coefficient is positive and significant in external lighting category, which is a more
mature category than the other categories. The results suggest that it is profitable for
firms to have a smaller portfolio and focus on a fewer number of emerging
technologies that they can excel in instead of investing in a number of technologies.
However, as technologies and markets mature, firms need to enhance market
presence and develop products in more technologies.
IV.5.4 Research productivity
H
3
predicts that higher research productivity leads to higher market returns
and vice versa. Because of the small number of announcements made by each firm
each year, we make the assumption that the average research productivity of a firm
remains relatively constant during the period of analysis. While this assumption
implies markets have advance information of future innovations, and is not correct, it
is certainly plausible that markets use additional signals other than prior
announcements (e.g. announcements outside the period of analysis) in their
assessment of the research productivity of the firms. We find strong support for this

109
hypothesis (see Table IV-6 and Figures IV-6a-6c). We compared the firms with high
research productivity to firms with low research productivity in all three categories.
The results indicate that the former have higher returns of approximately 2%
compared to less than 1% for the latter. Plots of cumulative average abnormal returns
(CAAR) support these results. Hence, we find support for H
3
that larger portfolios
have higher returns than smaller portfolios.
However some differences in the three categories appear in the multiple-
regression framework (see Table IV-5). While there are positive returns to increasing
research productivity in external lighting, the coefficients are negative in the other
two categories. A possible explanation emerges from a closer examination of the
technologies in these categories. Most of the predominant technologies in lighting
had already matured decades before the period of data collection and all new
platform technologies introduced in this period were targeted towards niche markets
and emerging applications.
Table IV-6: Mean Abnormal Returns to firms with High vs. Low Research
Productivity
High Research
Productivity
Low Research
Productivity
Category
Mean Return* N t-val Mean Return* N t-val
External Lighting 1.86% 332.05 0.94% 108 4.01
Display Monitors 1.87% 78 3.04 0.56% 81 0.96
Desktop Memory 1.24% 843.21 0.72% 126 3.01
* abnormal returns on event day

110
On the other hand, almost all platform technologies were introduced during
or just prior to the start of the data collection period in monitors and memory
categories. Hence all technologies witnessed major improvements in this period.
This also affected the consumer knowledge about the technologies and market
acceptance of innovations. The impact of this factor could drive the differences in
market returns to variations in technological leadership and improvement in
performance.
Figure 20: CAAR Plots for High vs. Low Research Productivity
-2.00%
-1.50%
-1.00%
-0.50%
0.00%
0.50%
1.00%
1.50%
2.00%
2.50%
3.00%
3.50%
-5 -4 -3 -2 -1 012 345
CAAR
High
Low

-2.00%
-1.00%
0.00%
1.00%
2.00%
3.00%
4.00%
5.00%
-5 -4 -3 -2 -1 0 1 2 3 4 5
CAAR
High
Low

Figure IV- 6a: Lighting Category Figure IV- 6b: Monitors Category
-1.00%
-0.50%
0.00%
0.50%
1.00%
1.50%
2.00%
2.50%
-5 -4 -3 -2 -1 0 1 2 3 4 5
CAAR
High
Low

Figure IV- 6c: Memory Category

111
IV.5.5 Size of Firm
The alternative hypotheses H
4o
and H
4a
state that the market returns are
higher for larger firms than for small firms and vice versa respectively. We used
annual sales of the firm in the year of announcement as a measure of the size of firm.
Table IV-7: Mean Abnormal Returns to Small vs. Large Firms
Small Firms Large Firms
Category
Mean Return* N t-val Mean Return* N t-val
External Lighting 1.98% 43 2.05 0.70% 74 2.62
Display Monitors 2.87% 62 2.71 0.06% 80 0.24
Desktop Memory 1.57% 602.88 0.03% 103 0.16
* abnormal returns on event day
We conducted a median split and compared the average returns of small
firms to those of larger firms. For all three categories, we find that smaller firms have
higher returns than larger firms (see Table IV-7 and Figures IV-7a-7c). Market
returns to small firms were close to 2% for the three categories and less than 1% for
large firms. Plots of CAAR support the difference in returns. Hence, we only find
support for H
4a
that small firms have higher returns than large firms.
We also verified the results using a different measure of firm’s size - the
number of employees in the firm in the year of announcement. However the results
were not materially different from those reported above.

112
Figure 21: CAAR Plots for Small vs. Large Firms
-2.00%
-1.00%
0.00%
1.00%
2.00%
3.00%
4.00%
5.00%
-5 -4 -3 -2 -1 0123 45
CAAR
small
big
-2.00%
-1.00%
0.00%
1.00%
2.00%
3.00%
4.00%
5.00%
6.00%
-5 -4 -3 -2 -1 0 1 2 3 4 5
CAAR
small
big

Figure IV- 7a: Lighting Category Figure IV- 7b: Monitors Category
-1.00%
-0.50%
0.00%
0.50%
1.00%
1.50%
2.00%
2.50%
3.00%
3.50%
4.00%
-5 -4 -3 -2 -1 0 1 2 3 4 5
CAAR
small
big

Figure IV- 7c: Memory Category
IV.5.6 Frequency of announcements
H
5
hypothesizes that firms do not gain by making a larger number of
technological announcements per year. We find support for this hypothesis. For each
of the measures we first calculated an average frequency of announcements per year
for all firms in the industry. We then conducted a median split and compared the
average returns of firms making less than the average number of announcements to
those making more than the average number of announcements. In each of the
categories, there are a significant number of firms making from less than one
technological announcement per year to almost one announcement every month.
Thus, there is a large variation in announcement strategies employed by the firms.

113
Table IV-8: Mean Abnormal Returns to firms with High vs. Low Frequency of
Announcements
High Frequency Low Frequency
Category
Mean Return* N t-val Mean Return* N t-val
External Lighting 0.39% 472.05 1.69% 70 3.96
Display Monitors 0.11% 830.41 2.95% 59 2.78
Desktop Memory 0.95% 692.82 0.34% 94 1.01
* abnormal returns on event day

Figure 22: CAAR Plots for High vs. Low Frequency of Announcements
-2.00%
-1.50%
-1.00%
-0.50%
0.00%
0.50%
1.00%
1.50%
-5 -4 -3 -2 -1 0 1 2 3 4 5
CAAR
High Frequency
Low Frequency
-2.00%
-1.00%
0.00%
1.00%
2.00%
3.00%
4.00%
5.00%
6.00%
-5 -4 -3 -2 -1 012345
CAAR
High Frequency
Low Frequency
Figure IV- 8a: Lighting Category Figure IV- 8b: Monitors Category
-1.50%
-1.00%
-0.50%
0.00%
0.50%
1.00%
1.50%
2.00%
2.50%
-5 -4 -3 -2 -1 0 1 2 3 4 5
CAAR
High Frequency
Low Frequency

Figure IV- 8c: Memory Category

We find that firms making fewer announcements have higher returns than
firms making frequent announcements (see Table IV-8 and Figures IV-8a-8c) in all

114
categories. A comparison of firms making different frequency of announcements is
shown in the CAAR plots of these different strategies. Hence, we find support for H
5

that firms gain by making fewer announcements.
IV.5.7 Age of Technology
H
6
hypothesizes that market returns to technological announcements related
to new technologies are higher than those related to old technologies. We measured
the age of technology as the difference between the year of announcement and the
year of inception of the first product based on that technology. We measured the year
of inception as the year in which the first commercial product based on that
technology was introduced in the market. For example, the age of ink-jet printing
technology is measured relative to the first introduction of a desktop printer based on
ink jet printing in the market by Hewlett Packard.

Table IV-9: Mean Abnormal Returns to firms with Old vs. New Technologies
Old Technology New Technology
Category
Mean Return* N t-val Mean Return* N t-val
External Lighting 0.94% 842.61 1.74% 33 4.04
Display Monitors 2.17% 742.65 0.32% 68 0.96
Desktop Memory 0.27% 811.15 0.92% 82 2.16
* abnormal returns on event day
We find mixed results for this hypothesis (see Table IV-9 and Figures IV-9a-
9c). Market returns to older technologies was higher only in display monitors

115
category. Plots of cumulative average abnormal returns (CAAR) also demonstrate
the difference in returns. Hence, we find only partial support for H
6
that new
technologies have higher returns.
Figure 23: CAAR Plots for Old vs. New Technologies
-3.00%
-2.50%
-2.00%
-1.50%
-1.00%
-0.50%
0.00%
0.50%
1.00%
1.50%
2.00%
-5 -4 -3 -2 -1 01 2345
CAAR
old technology
new technology

-2.00%
-1.00%
0.00%
1.00%
2.00%
3.00%
4.00%
-5 -4 -3 -2 -1 0 1 2 3 4 5
CAAR
old technology
new technology
Figure IV- 9a: Lighting Category Figure IV- 9b: Monitors Category
-1.50%
-1.00%
-0.50%
0.00%
0.50%
1.00%
1.50%
2.00%
2.50%
3.00%
-5 -4 -3 -2 -1 0 1 2 3 4 5
CAAR
old technology
new technology

Figure IV- 9c: Memory Category
IV.5.8 Effect of Innovation on Competitors
H
8
suggests that the average abnormal return to all competitors on the day of
announcements is negative or equal to zero. First we created daily portfolios of all
firms that made an announcement on that day. We then created another portfolio of
competitor firms that comprised all firms that did not make any announcement on
that day. We then calculated difference in daily abnormal returns to both portfolios.
A positive difference in abnormal return indicates that the firm making the

116
technological announcement attracts higher abnormal returns than the competitors
firms, and vice versa.
We find support for the hypotheses – the difference between average market
returns to firms making technological announcements was positive and significantly
different from zero (see Table IV-10). Moreover, the average market returns to all
firms that did not make any technological announcement was not significantly
different from zero. We discuss the implication of these findings in the discussion
section.
Table IV-10: Market returns to Competitors
Market return to competitors Difference
Category
Mean t-val Mean t-val
Lighting -0.01% -0.03 2.04% 0.46
Monitors -0.13% -0.88 1.05% 0.19
Memory 0.21% 1.36 0.71% 0.17

IV.5.9 Market Response to being First to Announce Innovations
The finding that there are lower returns to technological leadership is
counter-intuitive. On one hand it reflects the higher uncertainty associated with
innovative activity and the usually higher rate of failure that besets pioneering firms.
On the other hand, it is not clear why any firm would choose to pioneer new
technologies if the returns to such ventures were lower than following other firms.
In order to explore this question, we compared the returns to firms that
introduced innovations more frequently before others in the industry. We first

117
calculated the percentage of all announcements made by a firm where it was the first
to announce technological innovations. We then divided the entire sample on a
median split based on the frequency by which firms lead or follow others. We then
examined the difference in market returns to these two sets of firms.
We found that firms that announced more ‘firsts’ compared to the industry
average had a high market return of 2.31%. On the contrary, firms that announced
more innovations after other firms had already announced similar innovations had a
lower average market return of 1.06%. This indicates that while markets might not
be enthusiastic of innovation by all firms, there are positive returns to firms which
frequently pioneer technological breakthroughs. This also suggests why firms often
choose between a predominant technological leadership strategy and a close follow
strategy and stick with it.
IV.6 DISCUSSION
This section summarizes the findings and discusses implications, and
limitations of this study.
IV.6.1 Summary of findings
The current research leads to the following major findings:
• Market returns to announcements of technological innovations prior to
commercialization are strongly positive and significantly different from zero for
all three categories.
• On average, the market returns to an innovation that is the first in the industry is
lower than the return to a similar innovation that is announced later.

118
• However, market returns to firms with a stronger track record of introducing
innovations ahead of all others are higher than firms that frequently announce
similar innovations with a time lag.
• Market returns are higher for smaller firms and firms focusing on small set of
technologies.
• Market returns to firms making technological announcements are higher than the
returns to their competitors.
IV.6.2 Profile of Firms with High Returns
We calculated the mean returns to all announcements of each firm and then
segregated all firms based on a median split over average abnormal returns in order
to compare the profile of firms with high returns to those with low returns. An
average firm with low returns is 59.7 years old, has approx $36000 Million annual
sales, makes approximately 3.7 announcements per year and invests in 3.5
technologies but has a low research productivity of 0.15 innovations per year. On the
contrary, an average firm with high returns is only 30.5 years old, has approx $8000
Million annual sales, makes approximately 5.9 announcements per year, specializes
in only 2.2 technologies and has a higher research productivity of 0.28 innovations
per year. Thus, smaller, younger firms focusing on fewer technologies with higher
research productivity obtain higher market returns.

119
IV.6.3 Robustness of Results
We carried out a number of analyses to test the robustness of the results
including alternative estimation techniques, choice of market index, and clustering
problem.
IV.6.3.1 Choice of market return model
We used two more techniques to estimate the normal returns in order to
verify the robustness of our results. First, we used the mean adjusted return model
(equation 6) where the firm is expected to generate the same return that it averaged
during a previous estimation period. Second, we used the market adjusted return
model (equation 7) where the firm is expected to generate the same return as the rest
of the market.
...(7)
...(6)
it mt it
it i it
R R
R R
ε
ε
+ =
+ =

where R
it
and R
mt
are the period t returns on security i and the market
portfolio respectively and ε
it
is the zero mean disturbance term. The estimation
window for equations 8 and 9 is the same as used for equation 1. Similarly, abnormal
returns are calculated as the difference in observed returns of the firm during the
event window and the ‘normal’ returns. For each firm i and event date t, we have
...(9)
...(8)
ˆ
*
*
mt it it
i it it
R R AR
R R AR
− =
− =

Where
i i it
R AR α ˆ ,
ˆ
,
*
and
i
β
ˆ
are the abnormal return, mean firm return and
parameter estimates of market adjusted mode respectively. Figures IV-10a-c present
the plots of CAAR using the all models – mean adjusted, market adjusted and OLS

120
market models. The average abnormal returns to all categories are positive on event
day and significantly different from zero. Moreover, there is no significant difference
between estimations using the other models.
IV.6.3.2 Choice of market index
We used the equally weighted market index to estimate the abnormal returns
in equation 6 as per recommendation of Brown and Warner (1980). We also re-
estimated the abnormal returns using the value-weighted market index to ensure
robustness. The results are not materially different from the ones presented.
Figure 24: CAAR Plots using Mean Adjusted, Market Adjusted, and OLS Market
Models
-0.60%
-0.40%
-0.20%
0.00%
0.20%
0.40%
0.60%
0.80%
1.00%
1.20%
-5 -4 -3 -2 -1 01 2345
Event Day
CAAR
OLS Market
Mean Adjusted
Market Adjusted
-1.00%
-0.50%
0.00%
0.50%
1.00%
1.50%
2.00%
2.50%
3.00%
-5 -4 -3 -2 -1 0 1 2 3 4 5
Event Day
CAAR
OLS Market
Mean Adjusted
Market Adjusted

Figure IV- 10a: Lighting Category Figure IV- 10b: Monitors Category
-0.60%
-0.40%
-0.20%
0.00%
0.20%
0.40%
0.60%
0.80%
1.00%
1.20%
-5 -4 -3 -2 -1 0 1 2 3 4 5
Event Day
CAAR
OLS Market
Mean Adjusted
Market Adjusted

Figure IV- 10c: Memory Category
IV.6.3.3 Non-Parametric Tests
We also estimated the mean abnormal returns using nonparametric tests
which do not assume security returns to be normally distributed (Brown and Warner

121
1980). If this assumption is not supported, it can lead to false inferences. We used
the Wilcoxon sign rank test to test the null hypothesis that the proportion of observed
sample securities having positive returns is equal to 0.5. This would be true if
markets do not respond favorably to positive news of technological innovations, and
observed abnormal returns would be almost evenly distributed around 0. The results
reject the null (p = .03) and support our findings.
IV.6.3.4 More Announcements in Estimation Period
An assumption intrinsic to the market adjusted model is that the estimation
period used to estimate market parameters prior to the event are clean – there is no
other announcement made by the firm in that period. Since we examine multiple
announcements made by the same firm over the entire research and development
projects, this assumption is violated. We re-estimated the abnormal returns by
removing the dates of all prior announcements made by the firm and the succeeding
day from the estimation period. This is in line with recommendation made by
Peterson (1989) and Brown and Warner (1985) to handle missing returns.
IV.6.3.5 Alternative definitions of technological leadership
We investigated whether the returns to technological leaders are lower
because of the operationalization of technological leadership at the announcement
level instead of the firm level. We had earlier defined technological leadership on an
announcement level whereby any firm which is the first to announce the particular
innovation is the leader. However no firm is the first to realize all innovations ahead
of others, and firms vary in the percentage of announcements they make where they

122
are the first in the industry. Hence we did a median split of the percentage of times
that the firm is a leader in all its announcements and then compared the top 1/3
rd
of
the firms in the sample with the bottom 1/3
rd
of the firms. We found that results do
not change and the average returns to followers are still higher than the returns to
technological leaders.
In order to investigate whether the lower returns to technological innovations
for technological leaders was because of the use of average returns as a metric, we
repeated the analysis on a firm level. Possibly the technological leaders have higher
‘total’ returns because they make more announcements in the earlier phases. We
compared the total returns of the top 1/3
rd
of the firms in the sample with the bottom
1/3
rd
of the firms. In this case the difference in returns between the two categories
was not significantly different from zero.
IV.6.4 Survivor Bias
We took possible care to avoid any source of survivor bias in the analysis.
First, we took care to include almost all firms that had entered the industry at any
point in time even if they had ceased to exist at the time of data collection. This
ensured that it is not only survivors that are included in the sample. Second, we took
care to ensure all technological announcements that were related to the technologies
being investigated even though some of them may not have ever reached the
commercialization stage. Thus we included all innovation projects irrespective of
size, focus or stage of project.

123
IV.6.5 Implications and Contributions to Practice
This study has several implications for managers. First, it shows that there
are significant and substantial returns to innovative activity even during the research
and development period. Managers need to be aware of the magnitude of such
returns and the factors affecting the returns.
Second, we found that markets may react negatively to announcements of
innovations if the firm does not have a strong track record of frequently being ahead
of its competitors in developing and announcing innovations. The implication for
current players is to strive to maintain a lead in research and development and strive
to be foremost in introducing new innovative products. The implication for new
players is to strive to build a strong history and not be disheartened by low returns in
the beginning. An example of how NEC has successfully maintained high market
dominance by following a follower approach in Japanese markets in contrast to
Fujitsu which has traditionally stressed on being a technological leader (Methe,
Toyama and Miyabe 1997).
Third, we found that innovations in new, emerging technologies by small
firms attract higher returns. This finding provides validation to the common strategy
of creating new spin-offs to pursue emerging technologies employed by many big
multinational firms.
IV.7 LIMITATIONS AND DISCUSSION FOR FUTURE RESEARCH
A reasonable estimate of the marketability of technological innovations is
critical and a precursor to managing the assets of the firm. Our study is the first step

124
in that direction. But the study has several limitations. First we had to limit our
analysis to only three product categories due to the time and difficulty of data
collection. Second, the analyses of abnormal returns are dependent on the
assumption of market efficiency. But since we base the calculation of ‘normal’
returns’ on the asset pricing model, errors in estimation of the abnormal return can
not be solely attributed to the inadequacy of the efficient market assumption or the
correctness of the model. Third, we limit the analyses to only those firms that are
listed on the US stock exchanges. Thus we can not ascertain the returns to privately
held firms or firms traded on foreign exchanges. Finally, we had to limit analysis to
only those announcements that firms chose to make regarding the research and
development projects. Hence, the total reported returns are neither maximum nor
optimal.
Future research is required in finding the ongoing processes within firms
regarding managing market communication on these issues. All of these limitations
are potential opportunities for future research. In addition, future research may also
explore the long term returns to innovation.

125
V CHAPTER 5: TOTAL MARKET RETURNS TO NEW
PRODUCTS
3

V.1 ABSTRACT
Reliable measures of returns to technological innovations enable firms to
manage their research and development projects better. However, limiting the
analysis to only the commercialization severely under-estimates the total returns to
innovation. Using new data on 3 industries and 15 technologies, we examine the
market returns to innovation projects. We find that the response of stock markets to
announcements related to innovations projects is positive. Market returns to
innovation phase are the highest in all categories.
V.2 INTRODUCTION
The rate of technological evolution has increased substantially in the last few
decades (Sood and Tellis 2005; Tellis and Golder 1996). The increased pace of
change also enhances the need and importance of assessing the returns from
innovative activity accurately. Firms are continually interested in better measures for
returns to innovation – a fact substantiated by this issue being in the Marketing
Science’s list of top research priorities for the last few years. Prior researchers have
used the event study method to assess the market return to launch of new products,
and have reported positive returns to such events (Chaney, Divenney and Winer
1991). The event study method measures the market reaction to new information and

3
This essay won the John Funk Research Fellowship 2003 from Center for Research in Technology
and Innovation, Kellogg School of Management, Northwestern University.

126
the returns are proportional to the net present value of the incremental information
released by the firm in each announcement. However, announcements are made all
through the innovation project and all publicly available information is already
incorporated in the prices.
Hence, limiting the scope of study to only announcements made at the end of
new product development projects and ignoring all information released to the
market prior to the event may result in erroneous estimates of the true return to
innovation. As far as we know, there is no study done on the impact of
announcements made at all phases of the innovation project on market returns of
firms.
We aim to further our understanding of this important phenomenon in the
current study, and to answer the following questions:
• What is the total market return to innovation projects?
• What is the market return to specific phases of innovation projects?
• Do firms realize all returns to innovation over a short period of time or are the
returns spread over a longer window?
• What are the implications for marketing communication strategies?
The rest of the paper is organized as follows: The next two sections present
the theory, model and method. The last section discusses findings, limitations and
implications of the research. A brief description of the operating principles of
technologies in the three product categories is presented in the Annexure A.

127
V.3 THEORY
This section explains the phases of an innovation project and the market
response to information flow during the various phases of the project. We then
develop hypotheses about the phase of innovation project, immediate and long-term
response to innovation, and the effect of innovations on competitors.
V.3.1 Definitions
We define an innovation project as the firm’s activity related to developing
all products based on any single platform technology. For example, Philips’ research
efforts in developing products based on incandescent technology or optical memory
technology comprise two separate innovation projects for the purpose of this study.
We define a technological announcement as release of information regarding an
innovation project. A technological announcement may be made anytime during the
innovation project. Market return refers to stock market return that can be attributed
to the announcement.
V.3.2 Market Returns to Innovation
Prior research has examined the effect of innovation on firm performance
measures like sales, profits or market share. But these are lag measures subject to
many strategic and environmental factors and suffer from problems in establishing
causality. Conversely, changes in stock price of firms in response to specific events
related to innovation projects signal more direct gains to technological innovations.
These changes reflect the expected discounted present value of firm’s net cash flows
in response to new information on technological innovations.

128
Researchers in the past have investigated market returns to announcements
related to both the start of new projects and the commercialization of new products.
Regarding initiation of innovation projects, researchers have found positive returns
to announcements of R&D expenditures, joint ventures and strategic alliances
(Chauvin and Hirschey 1993; Chan, Kissinger and Martin 1992; Das, Sen and
Sengupta 1998). Similarly other researchers have focused on the commercialization
phase and examined the effect of announcements of new product launches (Eddy and
Saunders 1980; Wittink, Ryans and Burrus 1982; Chaney, Devinney and Winer
1991), announcements of success in R&D (Sharma and Lacey 2004). A common
finding in these papers was that a high percentage of announcements attract negative
market returns, and on average the announcements during this phase provide a weak
indicator.
However, there is relatively little research focusing on the innovation phase
of the innovation project (Kelm, Narayanan and Pinches 1995; Mishra and Bhabra
2001). Thus, the effect of information released about the progress of innovation
project prior to introduction of the new products is not considered. This is especially
problematic since the event study method usually used by these studies is based on
the efficient market hypothesis, which states that all available information is
immediately incorporated in the current prices of stock. This limitation in scope of
enquiry may result in under-estimating market-returns to innovation.

129
V.4 HYPOTHESES
In this section, we derive hypotheses about the effect of timing of
technological announcement relative to the phase of innovation project on the market
return, and how these returns vary over time. We also develop hypotheses related to
the use of these returns on market signals and the effect of technological innovations
on returns to competitors.
V.4.1 Return to Innovation in each Phase of Innovation Project
It is common to view the new product development activity comprising of
several consecutive and distinct phases from the start of innovation projects with
allocation of resources till the end with commercialization of new products. Kelm,
Narayanan and Pinches (1995) identify two main phases during the new product
development period in their assessment of – one prior to new product introduction
called the innovation phase and second subsequent to introduction called the
commercialization phase. We extend the phases of innovation projects investigated
to also include a phase prior to innovation phase called the Initiation phase. In each
of the phases, firms release different types of information.
In the Initiation phase, firms prepare for pursuing research and development
activities, and announcements refer usually to developing structure and resources
allocated to the project. For example, firms may announce formation of joint
ventures, strategic alliances and allocation of financial resources to the projects.
These alliances may be between firms with complementary competencies or between
firms with different value propositions in the value chain.

130
Market returns to any announcement reflect the net discounted present value
of the implication of these strategies. On one hand, announcements in the initiation
phase drain firm’s current resources and reduce the short term earning potential of
the firm (Chan, Kessinger and Martin 1992). Moreover, longer the delay between the
time of investment and time of completion of the project, higher the net present value
of this ‘loss’. This reasoning suggests that returns to such announcements are
negative. However many prior studies have consistently found that markets react
positively to such announcements and recognize that investments in developing new
products can enhance the long term value of the firm and its competitiveness
(Chauvin and Hirschey 1993; Chan, Lakonishok and Sougiannis 2001; Doukas and
Switzer 1992). This reasoning suggests that returns to such announcements are
positive and firms with higher outlays stand to gain more. Hence, we expect,
V.4.1.1 H
1
: Market returns to announcements in initiation phase are positive.
In the Innovation phase, firms conduct research and development activities
using the resources allocated in the initiation phase. The level of uncertainty reduces
marginally in the innovation phase in comparison to the initiation phase, and some
signals of success become visible in the form of prototypes, patents, or
identifications of new materials leading to improved performance and / or reduced
cost. However the products in this phase are still not ready for commercialization
and may not achieve production cost or quality targets. The announcements made by
firms during this period sometimes reflect this uncertainty explicitly as well. Thus,

131
while some of the announcements might elicit positive abnormal returns, the returns
would be distributed approximately evenly around zero. Hence, we expect that
V.4.1.2 H
2
: Market returns to announcements in innovation phase are zero.
In the Commercialization phase, the activities of prior phases result in final
products that can be commercialized. During this phase, announcements related to
various types of innovations are made e.g. improvement of product performance on
primary or secondary dimensions (Sood and Tellis 2005), launch of products with a
higher performance than its competitors. Announcements in this phase include early
commercialization of emerging technologies e.g. initial shipments and identification
of new applications of emerging technologies. We summarize the various types of
announcements based on content and the phase of innovation project in Annexure A.
In the commercialization phase, the announcements are the most positive and usually
include news of launched products. Firms spend considerable resources in launching
these products, enhancing consumer information and distribution of new products.
Hence we expect that,
V.4.1.3 H
3a
: Market returns to announcements in commercialization phase are
positive.
V.4.1.4 H
3b
: Market returns to announcements in commercialization phase are
higher than returns to either innovation or initiation phases.
Prior research also suggests that it is only a fraction of all announcements that
attract positive returns on the day of the announcement or during a short window
around the event date. For example Chaney, Devinney and Winer (1991) and
Markovitch and Steckel (2004) report that only about half of all announcements in
their sample experienced positive abnormal returns. Our data allows us to investigate

132
this phenomenon in more detail and we expect market returns to get increasingly
more positive as the innovation project becomes closer to completion. Hence we
propose,
V.4.1.5 H
4
: The overall percentage of announcements with positive abnormal
returns is the highest in commercialization phase followed by innovation
and initiations phase respectively.
V.4.2 Long Term Abnormal Returns to Innovation
It is possible that markets do not react fully to the announcements related to
innovative products at the time of announcement. This could happen for various
reasons. First, investors may not be fully aware of the significance of technological
advances and may take time to fully adjust prices to reflect them. The pace of
change, especially in high technology industries, increases the uncertainty of future
direction of research and market evolution. This increases the probability that there is
a delay in release of full information by the firms. Second, the information may not
travel efficiently from the scientific journals to investors and general consumers.
Third, the process of standards formation especially in emerging technologies is
unclear and slow, and the returns to innovations might reflect this uncertainty.
Moreover, this delay in appropriating relevant returns to announcements may
be more prominent in the innovation phase than in the commercialization phase. The
real value of advances in technology is more ambiguous in the innovation phase than
in commercialization phase when tangible products are already available to
consumers. Hence we expect that,

133
V.4.2.1 H
5
: The long term abnormal returns to announcements in innovation phase
are positive and significantly different from zero.
V.4.2.2 H
6
: The long term abnormal returns to announcements in
commercialization phase are positive and significantly different from zero.
V.4.2.3 H
7
: The long term abnormal returns to announcements in innovation phase
are higher than long term returns to announcements in commercialization
phase.
V.5 METHOD
This section describes the models for estimating short term and long-term
effect of announcements related to innovation projects, our sample selection,
sources, and procedure for data collection.
V.5.1 Model
We use the standard event study approach (McKinlay 1997; Fama, French,
Jensen and Roll 1969) using the OLS market model (see equation 2) as the
benchmark model for estimating normal returns and to estimate the market returns to
announcements in the innovation projects. We defined the event as first release of
information by a firm regarding an innovation project. We estimated the parameters
of the OLS market model using the 240 trading days ending 30 days prior to the
announcement (see Figure V- 1). We used event windows of varying widths centered
on the date of announcement by including one, three and five days before and after
the event window.
Figure 25: Timeline of Event Study
T
0
T
1
T
2
T
3
0
Estimation
Window
Event
Window
Post-event
Window
T
0
T
1
T
2
T
3
0
Estimation
Window
Event
Window
Post-event
Window

134

For each firm i and event date τ, we have
] [
*
it it it
R E R − = ε
… 1
where ε
it
*, R
it
and E(R
t
) are the abnormal, observed and normal returns
respectively. For any security i, we have
it mt i i it
R R ε β α + + =
… 2
where R
it
and R
mt
are the period t returns on security i and the market
portfolio respectively and ε
it
is the zero mean disturbance term. α
i
, and β
i
, are the
parameters of the model. We calculate the cumulative abnormal return (CAR
i τ
) by
cumulating the abnormal returns for firm i over a period of time τ thus:
it
t
i
AR CAR
∑
=
=
τ
τ
0
… 3
However there are three limitations of this model in the context of the present
study. We describe the limitations below and discuss a method to address it.
V.5.2 Absence of a clear estimation period
As described above, we need to estimate the coefficients in order to estimate
the ‘normal’ returns from an estimation period. The estimation period needs to
satisfy two minimum requirements: first the period needs to be long enough to
provide sufficient degrees of freedom for proper estimation. Second, there should be
no event, or announcement of technological innovation in the estimation period. The
presence of an announcement in the estimation period can affect the estimates and
hence the estimated normal returns.

135
V.5.3 Clustering of Announcements
Our sample consists of many announcements made by a single firm over a
period of time and also many announcements from firms in the same industry
occurring on the same day. Thus the assumption that the event window of the
included securities does not overlap is violated. This problem of clustering of event
dates results in non-zero covariance between abnormal returns. This affects the
accuracy of standard t-tests to make assessments regarding the level of market return.
V.5.4 Inaccuracy of estimated abnormal returns over a long period
The traditional event study model cannot be used to predict the long-term
effect of an announcement with reasonable accuracy. A limitation inherent to the
event study approach as defined above is its dependence on a method of defining the
return generating process and its use for estimating the return unconditional on the
event on question. In a typical setting we use a long enough estimation period with a
relatively short event period and assume that the intercept and slope coefficients of
the regression equation remain constant during the event period. However, this
assumption becomes more untenable with longer event or forecast periods. The
problem becomes even more acute when we also account for new events happening
in the period after the event day and the effect of these events on the coefficients
defining the normal returns.
V.5.5 Long-Horizon Event Studies
Researchers have long been interested in using event study approach to detect
the long-term effect of events on stock performance (Fama, Fisher, Jensen, and Roll

136
1969). Fama (1998) and Kothari and Warner (1997) document many issues inherent
in such techniques including risk adjustment, expected/abnormal return modeling,
the aggregation of security-specific abnormal returns, and the testing of the statistical
significance of abnormal returns. These issues become critically important with long
horizons.
The calendar time portfolio approach described below can be used to detect
long-term returns to technological innovations and minimizes many of the problems
identified above. The method does not depend on a pre-event period for estimation
of normal returns and hence also avoids some of the limitations of the classical event
study approach.
V.5.6 Calendar-time portfolio approach (Jensen-alpha approach)
We follow the approach introduced by Jaffe (1974) and Mandelker (1974) in
the financial-economics literature, and recently used by Fama (1998) and Mitchell
and Stafford (2000). The approach has been used to detect long-term abnormal
returns to both one-time (e.g. announcements of technological innovations) and
recurring events (e.g. earning announcements). We first identified all announcements
made by firms and segregate them as per the phase of innovation project. We then
created portfolios of firms making announcements within a certain period and
calculated the average observed returns to the firms in the portfolio (see Figure V-
2). Since the number of event firms is not uniformly distributed over the sample
period, the number of firms included in a portfolio is not constant through time. As a
result, some new firms are added each month and some firms exit each month.

137
Accordingly, the portfolios are reformed each month and mean portfolio return is
calculated.
Figure 26: Calendar time Portfolio approach
0 days
0 days to + 150 days
0 days to + 300 days
Portfolio Formation Initialization Phase
Initiation Innovation Commercialization
Initiation Innovation Commercialization
Initiation Innovation Commercialization
0 days
0 days to + 150 days
0 days to + 300 days
Portfolio Formation Innovation Phase
1900
2003
0 days
0 days to + 150 days
0 days to + 300 days
Portfolio Formation Initialization Phase
Initiation Innovation Commercialization
Initiation Innovation Commercialization
Initiation Innovation Commercialization
0 days
0 days to + 150 days
0 days to + 300 days
Portfolio Formation Innovation Phase
1900
2003

We then calculate the abnormal return by regressing the portfolio return in a
multifactor Fama – French model (see equation 4). The resulting time series of
monthly portfolio returns is regressed on the three-factor Fama-French model. The
estimated intercept from the regression of portfolio returns against factor returns is
the post-event abnormal performance of the sample of event firms.
HML SMB R R R R
j j ft mt p p ft pt
λ δ β α
ˆ ˆ
) (
ˆ
ˆ + + − + = −
… (4)
where α
p
: average monthly abnormal return on the portfolio of event
R
pt
: Return on a portfolio of stocks
R
mt
: Return on market portfolio, also known as systematic risk
R
f
: Rate of a "risk-free" investment e.g. one month T-bill
SMB : Return on a portfolio of small stocks minus return on large stocks

138
HML : Return on a portfolio of stocks with high book-to-market ratio
minus the return on a portfolio of stocks with low book-to-market ratio
α
p
is also called Jenson’s alpha in the finance literature and measures the
average return on a portfolio over and above that predicted by the factors in the
model. Inferences about the abnormal performance are on the basis of the estimated
α
p
and its statistical significance.
V.5.7 Procedure
We used the same sample as the study of market returns to new product
developments in chapter 4. The criteria, product categories and firms selected for the
study were similar to that study. There were two main extensions to the data
collected for the present study. First, we collected all announcements related to the
innovation project made by the firms for each phase of the project. In addition to just
the innovation phase of the projects, we also collected all announcements made in
the initiation phase and the commercialization phases for each firm. The sources
used for this data were also similar to the previous study (refer chapter 4 for details).
Hence, we had external lighting, display monitors and computer memory product
categories for this study. There was substantial innovative activity in all the
categories during this period and the sample comprises announcements from new
platform technologies (Sood and Tellis 2005) introduced during the period. There
were a total of five platform technologies in the lighting industry, five platform
technologies in the display monitor industry and three platform technologies in the
desktop memory category respectively.

139
We extracted incremental information from each announcement and
classified the announcement as per the classification system in Appendix B. More
specifically, we identified whether the announcement contained any information on
joint ventures, strategic alliances, new financial commitments, setting up new
infrastructure like production plants etc. We identified whether the announcement
contained any information regarding new material, development of prototypes or
demonstration models for exhibitions, granting of patents. Based on details of
technological innovation in the announcement, we classified the innovations as
platform, design or component. We examined whether the announcements contained
any information about improvement in performance, reliability, reduction in cost as a
results of research and development efforts of the firm. Finally we also investigated
the announcements for news related to new product announcements, initial
shipments, or quality awards, and information of a crossing in performance
compared to rival technologies.
V.6 RESULTS
We first present the findings on the overall market returns in the three
categories and then the results related to each of the hypotheses in this section.
V.6.1 Average Market Returns
We calculated the average market returns using the event study method as per
equations 2 and 3. The estimation period used was from 300 days prior to the
announcement till 30 days prior. We did not include announcements where the
estimation period was less than 100 days. This was essential for some new firms in

140
the sample which had been listed on the stock exchange for only a short period
before the announcements. We identified a total of 52 firms in the three industries
and collected a total of 1431 announcements from 1977 till 2003 (see Table V-1 for
details for each category).
Table V-1: Sample Data
Category External Lighting Display Monitors Desktop Monitors
Number of firms 18 16 18
Total Number of
Announcements
382 423 626
Sample Period ’77 – ‘03 ’83 – ‘03 ’82 – ‘03
Initiation Phase 60 54 129
Innovation Phase 141 159 210
Commercialization
Phase
242 210 287
5 5 5
Number/ Type of
Platform
Technologies

Incandescent, Arc-
discharge, Gas-
discharge, LED
and MED
CRT, LCD,
Plasma, Display
panels and OLED
Magnetic,
Magneto-optical
and Optical
V.6.2 Initiation Phase
H
1
predicts that market reacts positively to announcements in the initiation
phase. We find no support for H
1
. Using the event study techniques on all
announcements of initiation phase, we find that the average return was negative for
lighting category and positive for monitors and memory categories. However in all
the cases the average return was not significantly different from zero (see Table V-2
and Figure V- 3a). This indicates that firms did not gain by announcing their future
plans, or that the announcements were discounted by the market.

141
These findings are contradictory to prior findings and we also calculated the
average return per firm for all the announcements it made in the initiation phase. We
found only 1 in 52 firms had average abnormal return significantly different from
zero. A possible explanation of these results could be the high rate of innovation in
all three categories. In such a case where technologies are evolving at a fast pace,
markets could anticipate such investments in advance and incorporate expansion
plans in the current price prior to the event. Hence, the abnormal return to
announcements regarding initiation of new projects would not attract high returns.
V.6.3 Innovation Phase
H
2
predicts that market does not react to announcements in the innovation
phase. Our results are quite contrary to the hypothesis. We segregated all
announcements from this phase and estimated the average abnormal return using the
event study techniques. We find strong response to announcements in this phase
(see Table V-2 and Figure V- 3b). In all cases, the response to announcements was
significantly positive and different from zero. We discuss the implications of these
findings in the discussion section.

142
Figure 27: Market Returns in each Phase of Innovation Project
-5.00%
-4.00%
-3.00%
-2.00%
-1.00%
0.00%
1.00%
2.00%
3.00%
-5 -4 -3 -2 -1 0 12345
CAAR
Lighting Monitors Memory

Figure V- 3a Initiation Phase
-1.50%
-1.00%
-0.50%
0.00%
0.50%
1.00%
1.50%
2.00%
2.50%
-5 -4 -3 -2 -1 0 1 2 3 4 5
CAAR
Lighting Monitors Memory

Figure V- 3b Innovation Phase
-1.00%
-0.50%
0.00%
0.50%
1.00%
1.50%
2.00%
-5 -4 -3 -2 -1 0 1 2 3 4 5
CAAR
Lighting Monitors Memory

Figure V- 3c Commercialization Phase

143
Table V-2: Mean abnormal return using Event Study approach
Stage N
Event
Day
t-val
CAAR
(-1,+1)
t-val
CAAR
(-3,+3)
t-val
Initiation 60 -0.36% -0.78 0.04% 0.04 0.26% 0.21
Research 141 0.93% 4.01 0.87% 1.39 0.35% 0.52 Lighting
Realization 242 0.89% 3.66 0.06% 0.14 0.02% 0.04
Initiation 54 0.09% 0.14
-
0.23%
-0.21
-
0.25%
-0.15
Research 159 1.20% 2.85 0.71% 0.68 0.48% 0.52 Monitors
Realization 210 0.59% 2.27 0.23% 0.52 0.09% 0.13
Initiation 129 0.55% 1.85 0.08% 0.16 0.07% 0.09
Research 210 0.88% 3.83 0.59% 0.47 0.31% 0.29 Memory
Realization 287 0.69% 3.16 0.37% 0.98 0.14% 0.25
V.6.4 Commercialization Phase
H
3a
predicts that market reaction to announcements in commercialization
phase is positive. Our findings support this hypothesis. In all three categories the
returns are positive and significantly different from zero (see Table V-2 and Figure
V- 3c). The estimated abnormal return to announcements in commercialization phase
of approximately 0.8% is also comparable to findings in prior literature. Chaney,
Divenney and Winer (1991) report 0.25% returns for a similar period. Lee, Smith,
Grimm and Schomburg (2000) found 0.72% returns, and Sharma and Lacey (2004)
report 0.52% returns. These findings support the hypothesis that firms profit by
introducing new products to the markets.

144
Figure 28: Phases of Innovation Project
-0.36%
0.09%
0.55%
0.93%
1.20%
0.88% 0.89%
0.59%
0.69%
-0.60%
-0.40%
-0.20%
0.00%
0.20%
0.40%
0.60%
0.80%
1.00%
1.20%
1.40%
Lighting Monitors Memory
Initiation Research Realization

Figure V- 4a: Average Market Return in each phase with
all announcements of Innovation Phase
-0.36%
0.09%
0.55%
0.59%
0.54%
0.40%
0.89%
0.59%
0.69%
-0.60%
-0.10%
0.40%
0.90%
1.40%
Lighting Monitors Memory
Initiation Research Realization

Figure V- 4b: Average Market Return in each phase with
only announcements of Prototypes and Exhibition
Products in Innovation Phase
H
3b
predicts that the market returns are highest in commercialization phase.
We find no support for this hypothesis (see Figure V- 4a). In all three categories, we
find that the innovation phase has higher returns than the other phases. While it is
plausible that firms experience higher returns for announcements of success in
research and development, the finding that the returns in innovations stage are higher

145
than the commercialization stage is unexpected. We discuss the implications of this
finding in the discussion section.
In summary, we find that while the returns to announcements in
commercialization stage are positive, the returns are much lesser than returns to
announcements in the innovation stage.
V.6.5 Announcements with positive abnormal returns
H
4
predicts that the percentage of announcements that elicit positive
abnormal returns increases as the innovation projects get closer to commercialization
phase. We calculated the abnormal return on event day for all the announcements of
each stage and segregated them as per their valence and phase of innovation project.
Figure 29: Percentage of announcements with Positive Abnormal Returns
30.0%
40.0%
50.0%
60.0%
70.0%
Initiation Innovation Commercialization
Percent Positive
Lighting Monitors Memory

We then calculated the percentage of all announcements in each category that
elicited a positive abnormal return for each stage. We find support for this hypothesis
(see Figure V- 5). While in all three categories we observe the lowest percentage of
announcements with positive abnormal returns in initiation stage, the percentage of

146
announcements with positive market returns increases consistently in all three
categories as the innovation project moves differences in innovation stage and
commercialization stage are not very clearly different.
V.6.6 Long term abnormal returns to innovation
H
5
predicts that the long term abnormal returns to announcements in
innovation phase are positive and significantly different from zero. We tested the
long term abnormal return to innovation using portfolios of varying lengths – three
long-term portfolios over a period of 7, 30 and 300 days from the day of
announcement. We found no support for this hypothesis (see Table V-3). In all
categories, the returns are not significantly different from zero indicating that firms
receive full returns close to the actual announcement. These findings are consistent
with Chan, Lakonishok and Sougiannias (2001) who also failed to find significant
link between R&D and long term abnormal returns.
The lack of significantly positive abnormal returns could occur for two
reasons. First, it implies that the valuation process is efficient and markets are able to
assess the full market value of the innovations close to the event. Second, it also
implies that the firms release complete information in one time and there is no
evidence of gradual learning in the market. Since we calculate the abnormal returns
using the calendar time portfolio approach, we also circumvent the problems
intrinsic to event time approach like clustering, lack of clean estimation period and
need to estimate a ‘normal’ return over the long period.

147
H
6
predicts that the long term abnormal returns to announcements in
commercialization phase are significantly positive. We tested this hypothesis also
using the equation 4 with long term portfolios as in H
5
. We fail to find support for
this hypothesis (see Table V-3). Long term portfolios of any length failed to reveal
evidence of any abnormal returns significantly different from zero in any of the three
categories.
Table V-3: Mean abnormal return using Calendar Time Approach for announcements
in Innovation and Commercialization Phase
Innovation Phase Commercialization Phase
Category Period
α
p
t-val N R
2
α
p
t-val N R
2

0 - 7 days 1.21% 1.41 71 11.8% 0.83% 0.87 131 5.2%
0 - 30 days 0.13% 0.12 90 13.9% 0.20% 0.31 154 16.2% Lighting
0 - 300 days 0.73% 1.54 161 50.0% 0.51% 1.46 237 47.3%

0 - 7 days 2.63% 1.70 97 4.2% 0.88% 0.59 125 2.8%
0 - 30 days 1.61% 1.31 128 13.8% 0.45% 0.44 161 7.2% Monitors
0 - 300 days 0.86% 1.25 210 27.7% 0.06% 0.11 220 27.3%

0 - 7 days 1.13% 1.55 137 7.1% 1.18% 1.75 142 6.8%
0 - 30 days -0.18% -0.25 180 18.9% 0.80% 1.00 178 12.0% Memory
0 - 300 days -0.47% -1.00 250 29.7% -0.12% -0.21 217 32.0%

H
7
predicts that the long term abnormal returns to announcements in
innovation phase are higher than long term returns to announcements in
commercialization phase. However we fail to find support for this hypothesis since

148
the long term returns of both the phases are not significantly different from zero (see
Table V-3). Hence, the differences do not provide any meaningful interpretation.
V.7 DISCUSSION
This section summarizes the findings and discusses questions, implications,
and limitations of this study.
V.7.1 Summary of findings
The current research leads to five major findings:
• The response of stock markets to announcements related to innovations projects
is positive.
• Limiting the analysis to only the new product introductions under-estimates the
total returns to innovation. Market returns to innovation phase are the highest in
all categories.
• The percentage of all announcements with positive abnormal returns improves as
the innovation project moves closer to commercialization phase.
• The entire market return to announcement is observed within a short period of
the announcement and the long term abnormal returns are negligible.
V.7.2 Degree of Underestimation
In order to estimate the degree of underestimation, we calculated the total
returns to innovation as the sum of returns to all three stages for each of the
categories. We find that the total returns are 1.46% for lighting, 1.88% for monitors
and 2.13% for memory products respectively. This indicates that the prior focus of
researchers on only new product announcements to assess market returns to

149
innovation underestimates the market returns to innovation. For example, the
estimated return by Chaney, Divenney and Winer (1991) is a small fraction of the
total returns. In the present sample, the estimated return for only commercialization
phase is only approximately 42% of total returns. Alternatively, the estimated total
returns to new products are 2.6 times the returns to new product announcements.
However, a simple addition of all returns to announcements in innovation
phase to returns to announcements in commercialization phase might not be correct.
Some innovations apply to more than one product in commercialization phase. For
example, discovery of a new coating material to increase the storage capacity of
floppy disks was used for increasing the storage capacity of other memory products
as well. In such a case, the above estimate would represent the upper bound to the
total returns. Hence, a conservative estimate needs to count only announcements of
innovations which directly apply to the target commercialization and no other.
We removed all announcements in innovation phase which could not be
directly linked to specific product announcements in later phase. Hence we retained
only announcements of working prototypes and exhibition models in the innovation
phase, and removed announcements related to new material and patents. The total
number of announcements reduced to 1366 announcements. The total returns using
this revised metric are 1.12%, 1.22% and 1.65% for lighting, monitors and memory
categories respectively (see Figure V-3b). Hence, the estimated return for only
commercialization phase is a more conservative 57% of total returns. Alternatively,

150
the estimated total returns to new products are 1.9 times the returns to new product
announcements.
These findings were supported by many examples in the database. For
example, EMC Corp experienced negative returns of -0.94% when it introduced new
optical disks if we only take into account the announcements during
commercialization phase. The picture changes completely and the total returns to
EMC were 1.3% when returns to announcements made during earlier stages of
research are also included. Similarly, actual returns to California based Hill Dot
Systems were 3.17% when all the stages were included compared to a low 1.17%
when the scope was limited to only the product launches in the commercialization
phase.
Hence, we find that ignoring announcements from prior stages in estimation
of market returns to innovation severely underestimates the total returns to
innovation.
V.7.3 Implications and Contributions to Practice
This study has several implications for managers especially in a firm in
technology-intensive markets at both strategic and tactical levels.
First, it is inappropriate to limit analyses to only new product announcement
to estimate the market returns to innovation. The returns to announcements in the
commercialization phase ignore all information released about the product during the
early phases of the project. The impact of a wrong estimation could lead to erroneous
investment decisions to push certain products / innovations over others.

151
Second, we found that though the overall returns to innovation are positive, a
significant percentage of all announcements attract negative returns. Moreover, only
few firms in our sample had a positive average market return for all their
announcements. This implies that a better understanding of marketing
communication strategy for managing customer expectations and market information
might be helpful in reducing the number of announcements that elicit negative
reactions.
Third, the absence of long term abnormal returns to announcements of
technological innovations indicates that markets are quite efficient in estimating the
returns to innovation. We failed to find market returns significantly different from
zero for any period longer than 3 days for any category. This implies that the net
present value of even the technical information released in announcements is
incorporated into the stock fairly quickly and efficiently. The various biases and
anomalies of efficient markets discussed by DeBondt and Thaler (1985; 1987) do not
seem to be present in the valuation of new information related to technological
innovations.
Fourth, it is wrong to assume that the highest returns are observed at market
launch of products. In all categories, the highest average returns were during the
innovation phase.
V.8 LIMITATIONS AND DISCUSSION FOR FUTURE RESEARCH
This study has several limitations. First, we had to limit our analyses to only
three industries due to the difficulty in collecting comprehensive set of

152
announcements related to new product development and innovation. Second, the data
does not include firms not listed on the stock markets. These firms are not obliged to
release information to the market and the present method is not appropriate to
estimate returns to innovation for such firms. Third, the event study method – both
short term and long term – suffer from limitations intrinsic to the method e.g. the
joint hypotheses problem. Moreover, the estimates of market returns using the event
study methods are still an underestimate of the total returns since this method only
captures the return to the unexpected portion of the announcement, and ignores the
increase, or decrease, in stock price over time. Future research is required to solve
these limitations. In addition, future research may also examine optimal market
communication strategies aimed at maximizing the market returns to
announcements.

153
VI References
Abernathy, William J and K. B. Clark (1985), "Innovation--Mapping the Winds of
Creative Destruction" Research Policy 14 (1), 3-22.
Abernathy, William J., and J. M. Utterback (1978), "Patterns of Industrial
Innovation" Technology Review 80 (7), 40-7.
Acs, Z.J. and Audretsch, D.B. (1987), “Innovation, Market Structure, and Firm
Size,” The Review of Economics and Statistics, Vol. 69, No. 4, pp. 567-574,
Nov.
Adner, R., and D. Levinthal (2001), “Technology Evolution and Demand
Heterogeneity: Implications for Product and Process Innovation.”
Management Science, 47, 611-28.
Amemiya, T. (1985), Advanced Econometrics. Cambridge, Mass.: Harvard
University Press.
Anderson, Philip, and M. L. Tushman (1991), "Managing through Cycles of
Technological-Change." Research Technology Management 34 (3), 26-31.
Anderson, Philip, and Tushman, M. L. (1990), "Technological Discontinuities and
Dominant Designs—A Cyclical Model of Technological-Change"
Administrative Science Quarterly 35 (4), 604-33.
Assmus, Gert, John U. Farley, and Donald R. Lehmann (1984), “How Advertising
Affects Sales: Meta-Analysis of Econometric Results.” Journal of Marketing
Research 21 (1), 65-74.
Banz, R.W. (1981), "The Relationship Between Return and Market Value of
Common Stocks," Journal of Financial Economics 9, 3-18.
Bayus, Barry L. (1994), "Are Product Life-Cycles Really Getting Shorter." Journal
of Product Innovation Management 11 (4), 300-8.
____ (1998), "An Analysis of Product Lifetimes in a Technologically Dynamic
Industry" Management Science 44 (6), 763-75.
Bewley, R. and Griffiths, W.E., (2001), “A forecasting comparison of classical and
Bayesian methods for modeling logistic diffusion”, Journal of Forecasting
20, 231–247.

154
Bound John, Cummings Clint, Griliches Zvi, Hall, Bronwyn H., and Jaffe Adam
(1984), “Who Does R & D and Who Patents?” in R&D, Patents, and
Productivity, ed. Z. Griliches, NBER.
Brown Mark G and Svenson, Raynold A, (1998) “Measuring R&D productivity”,
Research Technology Management, Washington: Nov/Dec, Vol. 41, Iss. 6; p.
30 (6 pages).
Brown, Rick (1992), "Managing the 'S' Curves of Innovation." Journal of Consumer
Marketing 9 (1), 61-73.
Chan, L., Lakonishok, J. and Sougiannis, T. (2001), ‘The Stock Market Valuation of
Research and Development Expenditures’, Journal of Finance, Volume 56 ,
pp.2431-56
Chan, Su, John Kissinger and John Martin (1992), “The market rewards promising
R&D and punishes the rest”, Journal of Applied Corporate Finance, 5:59-66.
Chandy, Rajesh K. and Tellis G.J. (1998), "Organizing for radical product
innovation: The overlooked role of willingness to cannibalize," Journal of
Marketing Research, 35 (4), 474-87.
Chandy, Rajesh K. and Tellis G.J. (2000), "The Incumbent's Curse? Incumbency,
Size, and Radical Product Innovation" Journal of Marketing 64 (3), 1-17.
Chaney Paul K., Devinney Timothy M. and Winer, Russell S. (1991), “The Impact
of New Product Introductions on the Market Value of Firms”, Journal of
Business 64 (4): 573-610.
Chauvin, K. and Hirschey, M. (1993), “Advertising, R&D Expenditures and the
Market Value of the Firm’, Financial Management, Winter, pp.128-140.
Christensen, Clayton M. (1992a), "Exploring the Limits of the Technology S-Curve,
Part I: Component Technologies." Production and Operations Management,
Vol. 1 (4), 334-57.
Christensen, Clayton M. (1992b), "Exploring the Limits of the Technology S-Curve,
Part II: Architectural Technologies." Production and Operations Management
Vol. 1 (4), 358-66.
Christensen, Clayton M. (1993), "The Rigid Disk-Drive Industry--A History of
Commercial and Technological Turbulence" Business History Review 67 (4),
531-88.

155
Christensen, Clayton M. (1997), "The Innovator's Dilemma: When New
Technologies Cause Great Firms to Fail." Boston, Mass.: Harvard Business
School Press.
Christensen, Clayton M. (1999), "Innovation and the General Manager" Boston,
Mass.: Irwin/McGraw-Hill.
Cohen, W.M, Levin, R.C. and Mowery, D.C. “Firm Size and R&D Intensity: A
Reexamination,” Journal of Industrial Economics, 35, pp. 543-563. 1987.
Cohen, W.M. and Levin, R.C. “Empirical Studies of Innovation and Market
Structure,” in Handbook of Industrial Organization, eds R. Schmalensee and
R.D. Willig, New York: North-Holland, 1989
Comanor W.S. (1965), “Research and Technical Change in the Pharmaceutical
Industry", Review of Economics and Statistics 47: 182-190.
Constant, Edward W. (1980), “The Origins of Turbojet Revolution.” Baltimore, Md.:
John Hopkins University Press.
Cooper Robert C., Edgett Scott J., and Kleinschmidt Elko J. (1999), “New Product
Portfolio Management – Practices and Performance” The Journal of Product
Innovation Management, Vol. 16, No. 4 July.
Darling (1957), “The Kolmogorov–Smirnov, Cramer–Von Mises tests,” Annals
Math. Statistics 28, 823–888.
Das, S., P. K. Sen, and S. Sengupta (1998), “Impact of Strategic Alliances on Firm
Valuation.” Academy of Management Journal 41 (February): 27-42.
David M. Szymanski, Lisa Collins-Troy and Sundar G. Bharadwaj (1995), “Order of
Entry and Business Performance: An Empirical Synthesis and
Reexamination,” Journal of Marketing, 59 (October), 17-33.
De Bondt, W. F. M., and Richard Thaler (1985), “Does the Stock Market
Overreact?,” Journal of Finance, 40, 793-805.
De Bondt, W. F. M., and Richard Thaler (1987), “Further Evidence of Investor
Overreaction and Stock Market Seasonality,” Journal of Finance, 42, 557-
581.
Doukas, John and Lorne Switzer (1992), “The Stock market’s Valuation of R&D
Spending and Market Concentration,” Journal of Economics and Business,
44, 95-114.

156
Eddy, Albert R. and Saunders George B. (1980), “New Product Announcements and
Stock Prices”, Decision Sciences 11 (1): 9097.
Efron, B. and Tibshirani, R., (1993), “An Introduction to the Bootstrap” Chapman
and Hall, London.
Ettlie JE and Rubenstein AH (1987) “Firm Size and Product Innovation” Journal of
Product Innovation Management 4: 89-108.
Fama, E., Fisher, Lawrence, Jensen, Michael C. and Roll, Richard (1969), "The
Adjustment of Stock Prices to New Information", International Economic
Review, 10(1), pp. 1-21.
____ (1969), "Efficient Capital Markets: A Review of Theory and Empirical Work."
Journal of Finance, 1970, 25(2), Papers and Proceedings of the Twenty-
Eighth Annual Meeting of the American Finance Association New York,
N.Y. December, 28-30, pp. 383-417.
____ (1991), "Efficient Capital Markets II." Journal of Finance; 1575-1617.
____ (1998), “Market efficiency, long-term returns, and behavioral finance”, Journal
of Finance 49, 283-306.
Foster, Richard (1986), "Innovation: The Attacker's Advantage," New York: Summit
Books.
Freeman, C. (1974), "The Economics of Industrial Innovation." London, U.K.:
Printer.
Galbraith JK. (1952), “American Capitalism: The Concept of Countervailing
Power”, Boston, Houghton Mifflin.
Garcia, Rosanna, and R. Calantone (2002), "A Critical Look at Technological
Innovation Typology and Innovativeness Terminology: a Literature Review"
Journal of Product Innovation Management 19 (2), 10-32.
Ghemawat Pankaj (1991), "Market Incumbency and Technological Inertia"
Marketing Science 10 (2), 161-71.
Glascock, J.L., W.N. Davidson III, and G.V. Henderson Jr. (1987), “Announcement
Effects of Moody's Bond Rating Changes on Equity Returns", Quarterly
Journal of Business and Economics 26: 67-78.
Godoe, H. 2000. "Innovation Regimes, R&D and Radical Innovations in
Telecommunications" Research Policy 29:1033-1046.

157
Golder, Peter N. and G. J. Tellis (1993), "Pioneer Advantage - Marketing Logic or
Marketing Legend," Journal of Marketing Research, 30 (2), 158-70.
____ (2000), "Historical Method in Marketing Research with New Evidence on
Long-term Market Share Stability" Journal of Marketing Research 37 (May),
156-72.
Green, P.J. and Silverman, B.W., (1994), “Nonparametric Regression and
Generalized Linear Models: A Roughness Penalty Approach” Chapman and
Hall, London.
Hastie, T.J., Tibshirani, R.J. and Friedman, J., (2001), “The Elements of Statistical
Learning” Springer, Berlin.
Hauser, John R and Zettelmeyer, Florian (1997), “Metrics to evaluate R, D&E”,
Research Technology Management, Washington: Jul/Aug, Vol. 40, Iss. 4; p.
32 (7 pages).
Henderson, G. V. (1990), "Problems and solutions in conducting events studies",
Journal of Risk and Insurance, 42, 282-306.
Hendricks, K.B. and Singhal, V.R. (1997), “Delays in New Product Introduction and
the Market Value of the Firm: The Consequences of Being Late to the
Market”, Management Science Vol. 43 No.4. Apr.
Jaffe, J.F. (1974), “Special information and insider trading”, Journal of Business 47,
410-428.
James, G.M. and Silverman, B.W. (2005), “Functional adaptive model estimation”,
Journal of American Statistical Association 100 (470), 565–576.
James, Gareth and Ashish Sood (2005), “Performing Hypothesis Tests on the Shape
of Functional Data,” Journal of Computational Statistics and Data Analysis,
Vol. 50, Iss. 1.
John, George, A.M. Weiss and S. Dutta (1999), "Marketing in Technology-Intensive
Markets: Towards an Integrative Framework," Journal of Marketing, 63, 78-
91.
Johnson, N. and Kotz, S., (1970), “Distributions in Statistics”, Vol. III: Continuous
Univariate Distributions. Houghton-Mifflin, New York.
Kayal, Aymen (1999), "Measuring the Pace of Technological Progress: Implications
for Technological Forecasting." Technological Forecasting and Social
Change 60 (3), 237-45.

158
Kelm Kathryn M., V.K. Narayanan, and George E. Pinches (1995), “Shareholder
Value Creation During R&D and Commercialization Stages,” Academy of
Management Journal, 38 (June), 770–86.
Kerin Roger A., Varadarajan, P. Rajan and Peterson, Robert A. (1992), "First-Mover
Advantage: A Synthesis, Conceptual Framework, and Research
Propositions", Journal of Marketing, 56 (October), 33-52.
Klepper, Steven (1996), “Entry, Exit, Growth, and Innovation over the Product Life
Cycle” American Economic Review 86 (3), 562-83.
Kothari, S. and Warner, J. (1997), Measuring Long-Horizon Security Price
Performance, Journal of Financial Economics 43, 301-339.
Lee Hun, Smith Ken G., Grimm Curtis M., Schomburg August (2000), “Timing,
order and durability of new product advantages with imitation”, Strategic
Management Journal, Jan; 21,1.Page 23.
Lieberman, Marvin B. and D. B. Montgomery (1988), "1st-Mover Advantages,"
Strategic Management Journal, 9, 41-58.
Maddala, G.S. (1983), “Limited Dependent and Qualitative variables in
Econometrics.” Cambridge, U.K.: Cambridge University Press.
Malkiel, Burton G. (2003), “The Efficient Market Hypothesis and Its Critics”,
Journal of Economic Perspectives, Vol. 17, No. 1, pp. 59-82, October.
Mandelker, G. (1974), “Risk and return: the case of merging firms”, Journal of
Financial Economics 1, 303.
Markovitch Dmitri G. and Joel H Steckel (2004), “New product forecasting: the
stock market as crystal ball”, Working paper.
McKinlay, 1997, “Event Studies in Economics and Finance,” Journal of Economic
Literature, 35, pp. 13-39.
Methe, D. T., R. Toyama, J. Miyabe, (1997), "Product Development Strategy and
Organizational Learning: A Tale of Two PC Makers." The Journal of
Product Innovation Management, September.
Miller, A., Gartner, W. B., and Wilson, R. (1989) "Entry Order, Market Share, and
Competitive Advantage: A Study of Their Relationships in New Corporate
Ventures." Journal of Business Venturing, 4:3, May, 197-209.

159
Mishra, Debi Prasad and Harjeet S. Bhabra (2001), “Assessing the economic worth
of new product pre-announcement signals: theory and empirical evidence”,
Journal of Product and Brand Management, Vol. 10, No. 2, 75-93.
Mitchell, M.L. and E. Stafford (2000), "Managerial Decisions and Long-Term Stock
Price Performance," Journal of Business 73, 287-320.
Montgomery, A.L. and Bradlow, E.T., (1999), “Why analyst overconfidence about
the functional form of demand models can lead to overpricing” Marketing
Science 18, 569–583.
Moore, G. A. (1991), “Crossing the Chasm: Marketing and Selling High-Tech Goods
to Mainstream Customers.” New York, N.Y.: Harper Business.
Qualls, William, R. W. Olshavsky, and R. E. Michaels (1981), "Shortening of the
PLC--An Empirical Test" Journal of Marketing 45 (4), 76-80.
Reinganum, Jennifer F, (1985) "Innovation and Industry Evolution," The Quarterly
Journal of Economics, Vol. 100 (1) pp. 81-99.
Reinganum, Marc R. (1981) "Abnormal Returns In Small Firm Portfolios,"
Financial Analyst Journal, v37(2), 52-56,71.
Robertson Thomas S., J. Eliashberg, and T. Rymon (1995), "New Product
Announcement Signals and Incumbent Reactions" Journal of Marketing 59
(3), 1-15.
Robinson, P.M., (1989), “Hypothesis testing in semiparametric and nonparametric
models for econometric time series” Review of Econometric Studies 56 (1),
511–534
Robinson, W.T. and Fornell, C. (1985), ``Sources of Market Pioneer Advantages in
Consumer Goods Industries’’, Journal of Marketing Research, August, Vol.
22, pp. 305-17.
Robinson, W.T., Fornell, C. and Sullivan, Mary (1992) "Are Market Pioneers
Intrinsically Stronger Than Later Entrants?" Strategic Management Journal
13 (November), 609-624.
Ross, S. (1977), "The Determination of Financial Structure: The Incentive-Signaling
Approach." Bell Journal of Economics (Spring) 23-40.
Sahal, Devendra, (1981), “Alternative conceptions of technology” Research Policy
10 (1), 2–24.

160
Scherer, F. M. (1984), "Innovation and Growth: Schumpeterian Perspectives."
Cambridge, Mass.: MIT Press, p. 297.
Schilling, MA (1998), “Technological Lockout: An Integrative Model of the
Economic and Strategic Forces Driving Technology Success and Failure.”
Academy of Management Review 23 (2), 267-84.
Schumpeter, Joseph A. (1939), "Business Cycles: A Theoretical, Historical, and
Statistical Analysis of the Capitalist Process." New York, N.Y.: McGraw-
Hill.
Shapiro, Carl, and H.L. Varian (1999), “The Art of Standard Wars” California
Management Review (Winter), 41–2.
Sharma, Anurag and Nelson Lacey (2004), “Linking new product outcomes to
marker valuation of the firm the caser of the U.S. Pharmaceutical Industry”
Journal of Product Innovation Management, 21: 297-308.
Sood, Ashish and Gerard J. Tellis (2005), "Technological Evolution and Radical
Innovation," Journal of Marketing, 69, 3, 152-168.
Sorescu, Alina, Rajesh K. Chandy, and Jaideep Prabhu (2003), "Sources and
Consequences of Radical Innovation: Insights from Pharmaceuticals."
Journal of Marketing 67 (October), 82-102.
Srinivasan, (1970), “An Approach to Testing the Goodness of Fit of Incompletely
Specified Distributions”, Biometrika 57, 605–611.
Tellis, Gareth J. and P. N. Golder (1996), "First to Market, First to Fail? Real Causes
of Enduring Market Leadership" Sloan Management Review 37 (2), 65-75.
Tellis, Gerard J. (1988), “The Price Elasticity of Selective Demand: A Meta-
Analysis of Econometric Models of Sales.” Journal of Marketing Research
25 (November), 331-41.
Tushman, Michael L., and P. Anderson (1986), "Technological Discontinuities and
Organizational Environments" Administrative Science Quarterly 31 (3), 439-
65.
Utterback, James M. (1974), “Mastering the Dynamics of Innovation.” Science, New
Series 183 (4125), 620-6.
Utterback, James M. (1994a), Mastering the Dynamics of Innovation. Boston, Mass.:
Harvard Business School Press.

161
Utterback, James M. (1994b), "Radical Innovation and Corporate Regeneration"
Research Technology Management 37 (4) 10.
Veronesi, Pietro, (2000), “How does Information Quality Affect Stock Returns?”
Journal of Finance 55, 807-837.
Wittink D., A. Ryans, and N. Burrus (1982), “New Products and Security Prices,”
Working Paper, Ithaca, N.Y.: Cornell University.

162
VII Appendix A: Operating Principles of Sampled
Technologies
Technology Principle
VII.1.1 External lighting
1 Incandescent
Generate light by heating up thin metallic wires with an electric
current
2 Arc-discharge
Emit light by arc formed between two electrodes oppositely
charged by an electric current in a high-pressure gas chamber
3 Gas-discharge
Electrons excited by passing an electric current in a low-
pressure gas chamber emit light
4
Light emitting
diode (LED)
Emission of the light in n-p transition zone under influence of
an electric potential
5
Microwave
electrodeless
discharge
(MED)
Emission of light by microwaves from induction coil inside the
bulb to excite the gas.
VII.1.2 Desktop Memory
1 Magnetic
Records data by passing a frequency modulated (FM) current
through the disk drive's magnetic head thereby generating a
magnetic field that magnetizes the particles of the disk's
recording surface.
2 Optical
Stores data using the laser modulation system and changes in
reflectivity are used to store and retrieve data.
3
Magneto-
optical
Records data using the magnetic-field modulation system but
reads the data with a laser beam.
VII.1.3 Display Monitors
1
Cathode ray
tube (CRT)
Form an image when electrons, fired from the electron gun,
converge to strike a screen coated with phosphors of different
colors
2
Liquid crystal
display (LCD)
Create an image by passing light through molecular structures
of liquid crystals
3
Plasma display
panel (PDP)
Generate images by passing a high voltage through a low-
pressure electrically neutral highly ionized atmosphere utilizing
the polarizing properties of light

163

Technology Principle
VII.1.4 Desktop Printers
1 Dot matrix
Create an image by striking pins against an ink ribbon
to print closely spaced dots that form the desired image
2 Inkjet
Form images by spraying ionized ink at a sheet of paper
through micro-nozzles
3 Laser
Form an image on a photosensitive surface using
electrostatic charges, then transfer the image on to a
paper using toners, and then heat the paper to make the
image permanent
4 Thermal
Form images on paper by heating ink through
sublimation or phase change processes.
VII.1.5 Data Transfer
1 Copper/aluminum
Transmit data in the form of electrical energy as analog
or digital signals.
2 Fiber optics
Transmit data in the form of light pulses through a thin
strand of glass using the principles of total internal
reflection.
3 Wireless
Encodes data in the form of a sine wave and transmits it
with radio waves using a transmitter-receiver
combination.
VII.1.6 Analgesics

1 Opioids (narcotics)
Reduce generation of pain signals by inhibiting the
action of Cox enzymes responsible for inflammation.
2
Non-opioids anti-
inflammatory drugs
(NSAIDs)
Reduce brain sensitivity to pain by imitating the body’s
own painkilling chemicals and binding to pain-sensing
sites throughout the brain.
3
Non-opioids non-anti-
inflammatory drugs
(acetaminophen)
Preferential inhibition of pain by disrupting the
activation of Cox enzymes.
4
Non-drug pain
treatments
(acupuncture)
Alleviates pain by correcting the imbalance of qi
(pronounced "chee"), a type of life force, with needles
inserted at points along the energy pathways in the
body.

164
VIII Appendix B: Classification of Announcements
VIII.1 ANNOUNCEMENTS DURING INITIATION PHASE
• Joint ventures for R&D: between firms with complementary competencies
• Strategic alliances between firms
• Financial resources: Grants, advance orders, funded development contracts etc.
VIII.2 ANNOUNCEMENTS DURING INNOVATION PHASE
• Identification of new materials, processes, equipment
• Demonstrated working prototypes – before commercialization
• Demonstration products e.g. in exhibitions etc.
• Patents
• Design innovations / Component innovations
• Higher performance on primary or secondary dimensions of customer preference
• Increased reliability / Reduced cost
VIII.3 ANNOUNCEMENTS DURING COMMERCIALIZATION PHASE
• Product launches – with improved performance
• Initial shipments of products based on new platform innovations / shipments of
improved products of existing platform
• Crossing - higher performance relative to competition
• Identification of new applications
• Awards of excellence / recognition of quality

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

An empirical investigation of everyday low price (EDLP) strategy in electronic markets

PDF

Essays on new product preannouncements

PDF

Essays on technology innovation and new product development strategy

PDF

Influencing individual innovation through technology features that support cross -departmental understanding

PDF

Essays in supply chain management

PDF

Essays on regulation of public utilities and the provision of public goods

PDF

Information systems capabilities and indirect relational value: Sustaining networked organizations

PDF

Essays on organizational forms and performance in California hospitals

PDF

Consumer choice under budget constraint: Why consumers overspend

PDF

An investigation of repeat visit behavior on the Internet: Models and applications

PDF

Customer satisfaction, customer bargaining power, and financial performance

PDF

Individual heterogeneity and program evaluation

PDF

Compliance study of second-generation antipsychotics on patients with schizophrenia

PDF

Does adding nonfinancial value drivers to a summary financial measure improve the learning and performance of managers?

PDF

Board involvement in monitoring and strategy making: Antecedents and consequences

PDF

Information diffusion and the boundary of market efficiency

PDF

Effects of redundancy in media coverage on nonprofessional investors' earnings forecasts

PDF

Corporate ethical orientation and crisis management before and after September 11, 2001

PDF

Beyond skill development: The effects of training and development on the attitudes and retention of employees

PDF

Essays on commercial real estate and commercial mortgages

Asset Metadata

Creator Sood, Ashish (author)

Core Title Essays on technological evolution and financial returns to innovation

School Graduate School

Degree Doctor of Philosophy

Degree Program Business Administration

Publisher University of Southern California (original), University of Southern California. Libraries (digital)

Tag business administration, marketing,Economics, Commerce-Business,education, technology of,OAI-PMH Harvest

Language English

Contributor Digitized by ProQuest (provenance)

Advisor Tellis, Gerard (committee chair), James, Gareth (committee member), Officer, Micah (committee member), Ridder, Geert (committee member), Siddarth, Sivaramakrishnan (committee member)

Permanent Link (DOI) https://doi.org/10.25549/usctheses-c16-620919

Unique identifier UC11335930

Identifier 3222046.pdf (filename),usctheses-c16-620919 (legacy record id)

Legacy Identifier 3222046.pdf

Dmrecord 620919

Document Type Dissertation

Rights Sood, Ashish

Type texts

Source University of Southern California (contributing entity), University of Southern California Dissertations and Theses (collection)

Access Conditions The author retains rights to his/her dissertation, thesis or other graduate work according to U.S. copyright law. Electronic access is being provided by the USC Libraries in agreement with the au...

Repository Name University of Southern California Digital Library

Repository Location USC Digital Library, University of Southern California, University Park Campus, Los Angeles, California 90089, USA