Escaping competition and competency traps: identifying how innovative
search strategy enables market entry*
Benjamin Balsmeier a, Gustavo Manso b and Lee Fleming b
a) ETH, Zurich, Switzerland
b) University of California, Berkeley, USA
December 2016
Abstract: Innovation is usually assumed to be a crucial component of firm performance, yet the
optimal strategy and progression from invention to performance remains unclear and poorly
identified empirically. Likewise the idea of a fundamental tradeoff between exploration and
exploitation has been extremely influential, however, the stages and causal linkages between
search strategy and performance have not been established. We first demonstrate that a variety of
simple patent based measures clearly load onto exploration and exploitation principal
components and illustrate the temporal relationship between exploration and new market entry.
To identify the effect of innovative strategy on entry and successful entry, we rely on exogenous
shocks that precede exploration (non-compete enforcement switch) and exploitation (anti-
takeover regulatory reform). Using these exogenous shocks with different and opposite
mechanisms but consistent effects on market entry, we isolate one pathway from invention to
performance and demonstrate how exploration enables market entry and increased sales in new
markets. Exploration strategies appear less effective when the firm’s competitors are closer in
technology space; closeness in market space appears to have no effect on the impact of
technology strategy.
Keywords: Exploration, Exploitation, Patents, Innovation, Strategy, Market Entry, Experiment
* The authors thank Guan Cheng Li for invaluable research assistance. We gratefully acknowledge financial support
from The Coleman Fung Institute for Engineering Leadership, the National Science Foundation (1360228), and the
Ewing Marion Kauffman Foundation. Errors and omissions remain the authors’.
1
Introduction
According to much popular press, we live in a knowledge economy and an age of innovation.
Academic research has only begun, however, to establish how the invention of technology – and
innovative search strategy – influences firm performance. Our current understanding of the
relationship comes mainly from economics, finance, and strategy. Griliches (1984) built a log
linear model of physical and intangible assets to estimate the value of a patent to be about
$200,000; Kogan and co-authors (forthcoming) built an event study that implies a median value
of $3.2 million in 1982 dollars. At the portfolio level, Hall, Jaffe, and Trajtenberg (2005) found
a positive correlation with Tobin’s q and future citations; past innovative efficiency (as measured
by patents/R&D dollar, see Hirschleifer, Hsu, and Li 2013) or R&D performance (Cohen,
Diether, and Malloy 2013) predict abnormal returns. Firms search locally (Stuart and Podolny
1996) and enter markets more proximal to their current technological capabilities and experience
(Silverman 1991, Nerkar and Roberts 2004, Helfat and Lieberman 2002). Greater innovation
capabilities facilitate entry and competition decreases entry (de Figueiredo and Kyle, 2006,
Cockburn and MacGarvie 2011). Modularity aids innovation when not taken to extremes
(Schilling 2017); similarly, moderate exploration, as measured by text analysis of news articles,
correlates with financial performance (Uotila et al. 2009). At the risk of over simplification,
innovation appears to improve future performance, tends to build cumulatively on past
innovation and success, and must contend with competitors’ innovation.
Research opportunity persists, however, in at least three areas. First, there is probably no single
path from an inventor’s inspiration through applied research, product development,
manufacturing, marketing, distribution, sales, and ultimately a firm’s financial success; many
contingent paths probably exist, some better than others, and successful journeys may well
branch and recombine a variety of intermediate strategies. Theory that jumps directly from
invention to performance misses this nuance, complexity, and variety of successful strategy
combinations. Second, much work proceeds empirically and regresses financial outcomes
directly on patent and citation counts. Better measures could capture intermediate richness and
outcomes and motivate more nuanced theory. Third, little work separates the endogeneity of
strategy choice from the impact of the choice. This is particularly important for a field where the
object of study is fundamentally wrapped up with numerous unobservable variables; while it is
2
easy to download patent data and measures of firm performance, it remains very difficult to
observe and measure the decision processes and wealth of inside information that went into the
strategy process and choices.
We begin to address these issues by 1) focusing on the link between search strategy, market
entry, and performance, 2) developing a principal components method that operationalizes
March’s (1991) exploration vs. exploitation strategy with easily available patent measures, and
3) using two exogenous experiments to isolate and identify the impact of exploration vs.
exploitation strategies on new market entry and performance. First demonstrating how eight
basic measures of patent portfolios load 79% of their variance onto two components, we present
lagged regression models that illustrate the temporal relationship of innovative search strategy
and new product market entry (exploration correlates positively with entry, exploitation
correlates negatively, and both effects weaken with time). To strengthen causal inference, we
use two exogenous shifts, in labor markets and governance, namely, the passage of the Michigan
Anti-trust Reform Act of 1985 and anti-takeover regulations and (we refer to these as MARA
and ATO, respectively). MARA moved firms towards more exploration and ATO moved them
towards more exploitation. Their mechanisms were different; for MARA it appears that
inventors moved into new technical fields, both within and across firms (Arts and Fleming,
2016), and possibly because firms undertook more risky research and development (for evidence
of the effect from changes in Texas and Florida but not Michigan law, see Conti, 2014); for ATO
we conjecture that opportunities for selling the firm declined and that the market pressure for
novelty decreased. Despite differences in mechanisms, however, the result of exploration on new
market entry is consistent and strong; MARA induced a 0.202 point increase in the exploration
measure that resulted in a 42% increase in the propensity to enter a new market, while ATO
induced a 0.105 point decrease that resulted in a -14.9% decrease in entry. Sales in new markets
changed in a similar manner. While the exogenous push towards exploration appears to have
been less beneficial for firms which had been operating in more crowded technological space, it
appears to have been unaffected by the firm’s crowding in market space. Results remain robust
across correlations, instrumental variables, and differences in differences models.
3
Innovation Strategy and Performance
Invention and innovation are often modeled as a search process (March and Simon 1958). (Here
we adopt the typical convention of referring to raw patents and technology as invention and the
commercialization of such as innovation; search strategy encompasses the choices and processes
of both invention and innovation.) Individuals, groups, or firms search for novel and creative
solutions to problems or societal and market needs. Novelty is often defined as a new
combination of things, ideas, or processes (one can call these the components of recombination –
see Gilfillan 1935, Schumpeter 1942, and Henderson and Clark 1990). To the extent that a
searcher combines familiar and well-understood and previously used components, they search
locally and exploit; to the extent that they use less familiar or previously unused components or
recombine them in new ways, they pursue distant search and explore. Exploitation is more
certain and likely to pay off and pay off sooner, though rewards may be smaller and incremental;
exploration is risky, more likely to fail completely or discover a breakthrough, and take longer to
bring to fruition. Local search is more accessible but ultimately often leads the searcher to a
competency trap and strands them on a local maximum (March 1991).
If this model of search is correct, it presents firms with a strategic and fundamental conundrum.
On the one hand, they can explore new areas of technology, for example, by hiring outside of
their current expertise, acquiring firms from new industries, and funding speculative projects that
seek breakthroughs in new areas. The reward to such a strategy will probably be a more skewed
distribution of outcomes, with a lower mean and more complete failures and breakthroughs. On
the other hand, the firm can stick to its knitting, refine current trajectories, and build on its
current expertise. The rewards to this strategy will be quicker and more assured successes, and
fewer completely failed projects, though also fewer breakthroughs. Ultimately it may also trap
the firm on a local maximum and competency trap.
While this model should generalize to a variety of search strategies, we focus on one obvious
path, namely, from invention to new market entry and competitive conditions for success with
that entry. Firms can enter new markets with a variety of strategies, for example, re-labeling an
existing product from an existing market, foreign expansion using extant products, superior
manufacturing and/or distribution, or acquisition, however, we focus on new technology based
4
entry. The hypothesis is simple; firms that explore will develop new technologies that provide
the opportunity to build new products, differentiated at least from their current product line and
possibly from competitors’ product lines. This capability will enable and facilitate entry and
should be observed in an increased probability of entry and number of new markets that are
entered - assuming that managers see value and pursue such a strategy. If the strategy succeeds,
one would expect sales from new markets to increase and the proportion of a firm’s sales in new
markets to increase. Exploration on average should move the firm further away from other
firms, as it enables differentiation with new to the world products; it is less likely that new
technology capabilities will precede follower entry strategies.
One would also expect reactions from competitors that would lessen the benefits of exploration
and market entry (Wang and Shaver 2016), especially when other competitors have similar
portfolios of technologies. In other words, when firms are “close” in technology space (Stuart
and Podolny 1996; Aharonson and Schilling 2016) and the focal firm resides in a crowded
technological neighborhood, the appropriability and effectiveness of an exploration strategy will
decrease, for a variety of reasons. This occurs because competitors can understand and respond
to the exploration strategy more easily due to more similar absorptive capacities (Cohen and
Levinthal 1990). Knowledge transfer will be easier, from diffusion of patents, papers, and other
codified knowledge, and from personnel transfer, as poached employees can more readily
suffuse their prior knowledge from the exploring firms to competitors. As a result, we would
anticipate decreased efficacy for an exploration strategy on new market entry and performance,
for firms that pursue such a strategy from a crowded starting point in technology space. We
would also assume a negative effect of crowding in market space, defined as firms that operate in
a similar set of industries. We propose similar mechanisms, that firms with similar market
knowledge would be able to more quickly follow, react, or even anticipate the focal firm’s
exploration strategy. We would also expect, however, that the market crowding effect would be
weaker than the technology crowding effect, because competitor’s are more likely to lack the
technical absorptive capacity that will enable and facilitate response.
5
Data and measures
The empirical analysis is based on all public US based firms that field at least one patent in a
given year between 1977 through 2001 as identified in the NBER patent data. Data on basic firm
characteristics comes from Compustat North America and market entry and sales from
Compustat’s Historic Segment File. Detailed information on each patent provides the raw
observations for our reduced measures exploration and exploitation, are taken from the United
States Patent and Trademark Office, the NBER patent database, and the Fung Institute database
at UC Berkeley (Balsmeier et al. 2016). Based on the year of application of a given patent, we
aggregate our measures to the firm level of analysis. As patent based measures have no obvious
value in case of non-patenting activity, the sample comprises only observations when a firm
applied for at least one patent in a given year (as such, the results do not generalize to firms
without a patentable innovation strategy). Table 1 shows the distribution of firm-year
observations over the sampling period.
Table 1 – Frequency count of firm-year patent portfolio observations. Year Frequency Percent Cum.
1977 718 2.97 2.97
1978 718 2.97 5.94
1979 747 3.09 9.03
1980 756 3.13 12.16
1981 765 3.17 15.33
1982 770 3.19 18.52
1983 763 3.16 21.67
1984 783 3.24 24.91
1985 831 3.44 28.35
1986 832 3.44 31.8
1987 855 3.54 35.34
1988 873 3.61 38.95
1989 844 3.49 42.44
1990 877 3.63 46.07
1991 937 3.88 49.95
1992 1,024 4.24 54.19
1993 1,106 4.58 58.76
1994 1,191 4.93 63.69
1995 1,348 5.58 69.27
1996 1,315 5.44 74.71
1997 1,347 5.57 80.29
1998 1,323 5.48 85.76
1999 1,232 5.1 90.86
2000 1,141 4.72 95.58
2001 1,067 4.42 100
Total: 24,163 100
6
The increasing number of observations over time reflects the increase of patenting firms during
the sampling period. In order to limit selection in and out of the sample we require firms to be
observed at least 4 times (results remain robust to 2, 3, 5, 6 or 7 years as the threshold value).
The following eight patent portfolio characteristics are used to assess the direction of innovation
pursued by companies in terms of exploration and exploitation (further detail and
characterization of the data and measures are provided in Manso et al. 2016). Table 2 shows
summary statistics of these measures.
1. Number of patents that are filed in a 3-digit technology classes where the given firm has
never filed beforehand in that class.
2. Number of patents that are filed in a 3-digit technology classes where the given firm has
filed beforehand in that class.
3. Number of new technology classes entered where the given firm has never filed
beforehand in that class.
4. Technological proximity between the patents filed in year t and the existing patent
portfolio held by the same firm up to year t-1 (the normalized correlation between two
years of the proportion of activity in a given class, calculated according to Jaffe 1989).
5. Number of prior art citations to other patents (‘backward citations’).
6. Number of prior art citations to patents held by the same firm (‘self-backward citations’).
7. Number of claims a patent makes.
8. Average age of the inventor(s) mentioned on the patent document as calculated by the
time difference between the first time an inventor occurs in the Fung Institute’s patent
database and the application year of a given patent.
7
Table 2 – Summary statistics – patent portfolio measures
Variable N mean Median sd min max
Patents 24163 34.26 4 137.2 1 4054
New tech classes entered 24163 2.539 1 4.559 0 89
Patents in new classes 24163 3.016 1 6.183 0 185
Patents in known classes 24163 31.25 3 134.4 0 4051
Technological proximity 24163 0.541 0.581 0.328 0 1
Av. age of inventors 24163 3.633 3.063 3.131 0 26
Backward citations 24163 331.0 41 1387 0 48540
Self-citations 24163 42.55 1 280.6 0 11413
Claims 24163 528.5 66 2357 1 85704
Patent stock 24163 312.7 23 1279 0 34942 Notes: This table reports summary statistics of patent portfolio variables used in the study. Patents is the total number of eventually
granted patents applied for in a given year. New classes entered is the number of technology classes where a firm filed at least one
patent but no other patent beforehand. Patents in new/known classes is the number of patents that are filed in classes where the
given firm has filed no/at least one other patent beforehand. Technological proximity is the technological proximity between the
patents filed in year t to the existing patent portfolio held by the same firm up to year t-1, calculated according to Jaffe (1989).
Average Age of inventors measures the average time difference between the first time an inventor occurs in the Fung Institute’s
patent database and the application year of a given patent. Backward citations is the total number of citations made to other patents.
Self-citations is the total number of cites to patents held by the same firm. Claims is the total number of claims on each patent.
Patent stock is the sum of all patents held by a given firm up to the year t-1.
To reduce the dimensions of these data, we run a principal components analysis based on the
eight variables (similar results are obtained with a count based approach, or running a PCA at the
patent level). Two components have an eigenvalue above one, suggesting that extracting two
components are sufficient to explain the joint variation of the variables of interest. It supports
mapping the theoretical focus of exploration vs. exploitation onto two dimensions of innovative
search.
The output shown in Tables 4 to 5 indicate that 79 percent of the joint variation of the eight
patent variables of interest can be explained by these two principal components. In order to
optimize the factor loadings and reflecting the idea that exploration and exploitation are two
distinct dimensions of innovative search, we apply a Varimax rotation of the two extracted
components (results are robust to other orthogonal rotations). Table 4 shows the corresponding
results and Table 5 shows how much and in which direction each variable loads on the two
components. Loadings below 0.2 are not shown for easier comparability. Patents in known
classes, technological proximity, inventor age, backward citations, self-backward citations, and
claims all positively load on component 1, from which we label component 1 as ‘exploitation’.
The number of new technology classes entered and patents in new to the firm technology classes
strongly and positively load on component two. Negatively related to component two is the
8
technological proximity and the age of the inventors. Consistent with characterizations that firms
are more likely to explore if we observe new technological areas, we label component 2 as
‘exploration’.1
Tables 4 and 5 – Principal Component Analysis
Component Variance Difference Proportion Cumulative
Comp1 4.02 1.72 0.50 0.50
Comp2 2.30 0.29 0.79
Notes: This table reports the results of a Principal Component Analysis after Varimax
Rotation. Only components with Eigenvalues above one are extracted. The 8 variables
that entered the PCA are: new classes entered, patents in new/known classes,
technological proximity, av. age of inventors, backward citations, self-citations, and
claims; all variables log-transformed.
Variable Comp1 Comp2 Unexplained
New tech classes entered 0.58 0.08
Patents in new classes 0.58 0.08
Patents in known classes 0.45 0.09
Technological proximity 0.39 -0.38 0.47
Backward citations 0.41 0.10
Self-citations 0.45 0.16
Claims 0.41 0.09
Av. age of inventors 0.31 -0.37 0.60
Notes: This table reports the results of a Principal Component Analysis after Varimax
Rotation. Only components with Eigenvalues above one are extracted. All variables
log-transformed. Variable definitions provided above.
Table 6 – KMO test Variable KMO
New tech classes
entered 0.70
Patents in new classes 0.70
Patents in known classes 0.85
Technological proximity 0.86
Backward citations 0.92
Self-citations 0.89
Claims 0.89
Av. age of inventors 0.87
Overall 0.83
Notes: This table reports the Kaiser-Mayer-
Olkin (KMO) measures on sampling adaquacy.
All variables log-transformed. Variable
definitions provided above.
1 Measures of originality and generality (Hall, Jaffe and Trajtenberg 2001 - does the patent cite a wide variety of
classes and is it cited in turn by a wide variety) do not load on either of our components (neither at the firm nor
patent level). The measures do not map clearly to our theory; a patent could cite a wide variety of classes that had
never been cited together before, or had been heavily cited together before. In other words, a highly ‘original’ patent
could be citing a previously uncombined set of classes or a very commonly combined set of classes.
9
The Kaiser-Mayer-Olkin measure of sampling adequacy, shown in Table 6, confirms that the
data can be summarized using a PCA analysis. The correlation between the two factors is 0.37.
While this correlation indicates that there are some firms working in areas that score high on
exploration and exploitation, the correlation is far from being perfect, implying substantial
independent variation. Figure 1 illustrates this by plotting the factor values of the exploration
component against the factor values of the exploitation component. Red lines represent the
median values of each component. In the multivariate empirical analyses below, the scores of the
exploration and exploration component, respectively, will be our main explanatory variables of
interest. In a simple robustness check (not shown) we find similar results when counting the
number of variables that score above or below the median value for each variable in a given year
(the score varies from 0 to +8, though the empirical range is 0 to +6).
Figure 1: Scatter Plot of PCA scores
Notes: This graph plots the component scores of ‘Exploration’ and ‘Exploitation’
extracted from the Principal Component Analysis shown above. Red lines mark the
median values of each factor. 19% of the observations are each in the upper left and
lower right quadrants, 31% in each of the other quadrants.
The impact of any strategy obviously depends on competitors’ prior strategies, capabilities and
reactions. In the current context of search this can be conceptualized – and visualized – as a
position in technological or market space (Stuart and Podolny 1996; Aghion et al. 2005,
Aharonson and Schilling 2016). The efficacy of a particular search strategy will depend on a
firm’s and its competitors’ positions in space. For example, if firms face competitors that are
-50
510
Explo
ita
tion
-4 -2 0 2 4 6Exploration
Exploitation vs Exploration Scores
10
active in the same technological or market areas, it might be harder for those firms to realize the
benefits from exploration as it may be easier for close competitors to anticipate or follow search
success.
To assess technology space empirically we calculate pairwise correlations between a given
firm’s patent portfolio and all other firms’ patent portfolios in a given year, following Jaffe
(1989). Specifically, we calculate the technological proximity TP between each firm i and j at
time t as:
𝑇𝑃𝑖𝑗𝑡 =∑𝑓𝑖𝑘𝑡𝑓𝑗𝑘𝑡
𝐾
𝑘=1
/ (∑𝑓𝑗𝑘𝑡2
𝐾
𝑘=1
)
12
∗ (∑𝑓𝑗𝑘𝑡2
𝐾
𝑘=1
)
12
where 𝑓𝑖𝑘𝑡 is the fraction of firm i’s patents that belong to the main 3-digit CPC patent class k at
time t. To detect firms that compete closely in technological space we counted for each firm in a
given year how many other firms are close in technological space as measured by a TP score
higher than 0.95 (results are robust to alternatively taking 0.9 or higher thresholds). Competitors’
positions in market space are calculated similarly with sales generated in 3-digit sales classes
instead of patents filed in 3-digit CPC technology class. Figure 2 illustrates that firms’ positions
in technology space needs not to overlap with firms’ positions in market space.
0
.05
.1.1
5.2
Tech
Pro
xim
ity
0 .05 .1 .15 .2Market Proximity
Tech vs Market Space
11
Figure 2: Technology proximity vs. market proximity.
Figure 3 plots the exploration and exploitation scores of IBM over time, as well as the number of
patents and technological proximity of competitors. IBM appears to have begun the 1970’s with
an exploration strategy but this decreases over time in favor of exploitation. The time series of
the number of patents and exploitation look similar, and by themselves would miss IBM’s
variation in search strategy. Consistent with the idea that exploration is less predictable and
harder to manage we see larger variation of exploration scores over time as compared to
exploitation scores. The firm’s near demise in 1993 is obvious in the outlier on the left of the
figure, as is a move towards exploitation in the latter years of CEO Lou Gerstner’s tenure. The
firm has attracted more similar market competition over time.
45
67
8
1975 1980 1985 1990 1995 2000Year
Pre Gerstner Gerstner
Exploitation0
12
34
1975 1980 1985 1990 1995 2000Year
Pre Gerstner Gerstner
Exploration
0
200
04
00
0
1975 1980 1985 1990 1995 2000Year
Pre Gerstner Gerstner
Patents
-.0
2-.
01
0
1975 1980 1985 1990 1995 2000Year
Pre Gerstner Gerstner
Technological Proximity
IBM
12
Figure 3: IBM innovation search strategy 1979-2001. 1993 marked the firm’s “near death”
experience as well as its lowest innovative exploration.
Figure 4 illustrates the same graphs for General Electric. Consistent with Jack Welch’s
reputation, we do see greater exploitation and lessened exploration during his tenure and in
particular, a step increase in exploitation in the 6th year after he became CEO. Figure 5 illustrates
how Intel appears to be relatively unique in its ability to increase exploration and exploitation
simultaneously. In contrast to both IBM and Intel, GE appears to have developed a more unique
market profile over time.
0.0
2.0
4.0
6.0
8.1
.12
.14
.16
1975 1980 1985 1990 1995 2000Year
Pre Gerstner Gerstner
IBM Market Proximity
13
Figure 4: General Electric innovation search strategy 1979-2001.
55
.56
6.5
1975 1980 1985 1990 1995 2000Year
Pre Welch Welch
Exploitation
12
34
5
1975 1980 1985 1990 1995 2000Year
Pre Welch Welch
Exploration6
00
100
01
40
0
1975 1980 1985 1990 1995 2000Year
Pre Welch Welch
Patents
0
.05
.1.1
5.2
1975 1980 1985 1990 1995 2000Year
Pre Welch Welch
Technological Proximity
GE
0.0
2.0
4.0
6.0
8
1975 1980 1985 1990 1995 2000Year
Pre Welch Welch
GE Market Proximity
14
Figure 5: Intel’s innovation search strategy 1979-2001. Intel appears to be relatively
unique in its ability to simultaneously explore and exploit.
02
46
8
1975 1980 1985 1990 1995 2000Year
Moore Grove
Barrett
Exploitation
-20
24
6
1975 1980 1985 1990 1995 2000Year
Moore Grove
Barrett
Exploration
0
750
150
0
1975 1980 1985 1990 1995 2000Year
Moore Grove
Barrett
Patents
-.0
25
-.0
1.0
05
1975 1980 1985 1990 1995 2000Year
Moore Grove
Barrett
Technological Proximity
Intel
0.0
2.0
4.0
6.0
8.1
.12
.14
1975 1980 1985 1990 1995 2000Year
Moore Grove
Barrett
Intel Market Proximity
15
Comparison of Figures 3, 4, and 5 invite a number of insights. It is first important to note the
differing scales. For example, while Intel appears to be one of the rare firms that increased its
exploration over time, it also dipped into a negative value of exploration in 1984. This may
illustrate the pressure on its DRAM business and transition towards microprocessors. And Intel
has simultaneously increased its exploitation over time, from scores near zero in the 1970s to
scores near seven in the 2000s. IBM demonstrates half as much change over the same time
period and GE half again as much. Perhaps this illustrates the transition of Intel from a relatively
small startup in the 1970s to a dominant manufacturer; in contrast, IBM and GE have been large
and established firms over the entire time period. The technological proximity measure also
varies greatly between firms and appears to correlate most closely to patenting and exploitation,
though it is important to note that it reflects competitors’ search strategies as well. Perhaps most
interesting are the differences between the measures; exploitation seems to keep a firm in more
crowded neighborhoods and exploration the opposite – though not always, as Intel manages to
increase exploration and compete in a more crowded neighborhood. Crowded technological
neighborhoods appear to make commercialization more difficult, as the regressions below will
demonstrate. Finally, the market position of a firm, at least as defined by 3 digit SIC codes,
bears little correlation to the technical position.
IBM and GE appear to exploit more as they age, and prompt the question of whether this is
typical of most firms. Figure 6 illustrates the relation between firm age (years since first
appearance in Compustat) and exploration and exploitation scores, respectively. Consistent with
the organizations and population ecology literature (Hannan 1998; Sorensen and Stuart 2001),
organizations typically appear to exploit more as they age.
16
Figure 6: Age and typical innovation search strategy.
Measures and outcomes: descriptions and correlations
We investigate how a firms’ innovation strategy influences product market entry and
commercialization success by assessing a firm’s likelihood of entering a new to the firm product
market, the number of markets entered, and the amount of sales in new markets. Financial data
comes from Compustat segment files for US public firms’ sales per 3-digit SIC industry class.2
We first consider a binary indicator if a given firm enters at least one new product market,
defined as the first time appearance of positive sales in a given 3-digit SIC industry where the
firm has not generated sales previously. Second, we measure the number of newly entered
industries, defined as the total number of industries where the firm generates sales for the first
time in a given year. The third variable is the total amount of sales generated in all new
industries where the firm did not generate sales beforehand.3 Table 3 provides descriptive
statistics on these variables (the number of observations reduces due to fewer availability of sales
2 Results are robust to considering 4-digit level sales data instead. 3 Compustat’s sales data come from firms which may not always be brake down generated sales by product
categories rather than geographical location. For the definition of entry in new product markets we just count each
time sales are generated in a new to the firm specific SIC code, regardless of whether the sales may have been
generated outside the US only. Further, firms often report sales data more than once year. We took the largest
number of sales reported in a given year for a given industry as often even the largest number does not capture all
sales a firm has generated in given industry and year. All results are robust to taking the average sales per industry
and year instead of the maximum.
-.5
0.5
11.5
2E
xplo
ratio
n/E
xp
loitation
facto
r score
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31Firm Age
95% confidence band Exploitation
Exploration
Fractional polynomial fit without log transformation
17
data). On average we observe that 9.7% of the firms in our sample enter a new to the firm
product market in a given year.
18
Table 7 – Summary statistics – firm level measures
Variable N Mean Median Sd Min Max
Exploitation 22897 0.00913 -0.349 2.014 -3.775 7.847
Exploration 22897 0.0210 -0.0719 1.516 -3.259 6.981
R&D int. 22897 0.0849 0.0376 0.176 0 9.753
log(age) 22897 2.124 2.303 0.865 0 3.466
log(total assets) 22897 12.61 12.49 2.252 3.807 20.02
Enrtry exp. 22897 1.098 0 1.843 0 15
HHI 22897 0.167 0.122 0.133 0.0296 1
Entry 0/1 22897 0.210 0 0.407 0 1
No. entries 22897 0.305 0 0.704 0 7
log(new sales) 22897 1.039 0 2.277 0 11.10
Prod. proximity 5800 0.630 0.602 0.161 0.142 1.231 Notes: This table reports summary statistics of firm level measures used in the study. Exploitation and exploration are the
component measures derived from the PCA described above. R&D int. is R&D expenditures divided by total assets. Age is the
number years since first time occurrence in Compustat. Entry 0/1 is a binary variable that indicates if a given firm enters at least
one new product market, defined as the first time appearance of positive sales in a given 3-digit SIC industry where the firm has
not generated sales previously. No. entries is the number of newly entered industries, defined as the total number of industries
where the firm generates sales for the first time in a given year. New sales is the total amount of sales generated in all new industries
where the firm did not generate sales beforehand. The latter three variables are measured as the sum over the years t+1 to t+3. Prod.
proximity is the median value of firms’ pairwise proximity scores based on textual analysis of firms’ 10k fillings by Hoberg and
Phillips (2015, 2010), multiplied by 100.
Transforming technological discoveries into new products takes time. Hence we consider years
one to three (all results are qualitatively robust to taking 2 to 4, 3 to 5 instead) after observed
patenting activity and product market entry. Specifically, we will regress the above mentioned
product market entry variables on the exploration and exploitation components observed one to
three years beforehand. With respect to the binary entry indicator variable we will use a new
binary variable as dependent variable that is one if a firm entered a new to the firm market in t+1,
t+2, or t+3. With respect to the number of industries entered and sales in new to the firm
industries, we sum up all sales generated in t+1 to t+3 and take the logarithm of it as the
dependent variable.
When the dependent variable is a binary indicator of new market entry we estimate a Probit
model instead of OLS.4 All regressions include controls for R&D intensity as measured by R&D
investment scaled by total assets, because more R&D intensive firms might be more inclined to
enter new markets. The logarithm of total assets controls for firm size as larger firms may find it
easier to diversify and enter new markets. The logarithm of a firm’s age addresses a potential
4 All results are robust to estimating a linear probability model instead.
19
focus on new markets after existing products have been exploited and the likelihood that firms
find exploration more difficult with age. Next, a sales-based Herfindahl Index measured at the
SIC 3-digit level enters the regressions to control for variations in competition across industries
as well as the logarithm of a firm’s patent stock (total number of patents accumulated over time).
Further controlling for firms’ capabilities and ability to enter new markets we add the logarithm
of the number of previously entered new industries plus one. Finally, a full set of industry and
year dummies control for heterogeneity of market entry rates across industries and time.
Table 8 – Correlations between Exploration/Exploitation and product market entry
a b c d
Dependent variable Entry 0/1 No. entries New sales Prod.
proximity
log(pat stock) -0.025* -0.004 0.001 -0.008***
(0.014) (0.003) (0.021) (0.003)
R&D int. -0.258 0.000 0.559*** 0.247***
(0.199) (0.029) (0.127) (0.025)
log(Age) -0.035* -0.011** -0.059** -0.019***
(0.019) (0.005) (0.028) (0.003)
log(Total assets) 0.065*** 0.022*** 0.267*** 0.013***
(0.014) (0.003) (0.020) (0.003)
Herfindahl ind. 0.229*** 0.046*** 0.214*** -0.030***
(0.029) (0.008) (0.051) (0.006)
Entry exp. 0.539** 0.109* 0.826** -0.078*
(0.222) (0.060) (0.355) (0.042)
Exploitation 0.005 -0.001 -0.001 0.006***
(0.012) (0.003) (0.017) (0.002)
Exploration 0.065*** 0.014*** 0.094*** -0.004**
(0.011) (0.003) (0.018) (0.002)
N 22897 22897 22897 5800
Industry and time fixed effects yes yes yes yes
R2 / Pseudo R2 0.134 0.150 0.202 0.498 Notes: All dependent variables are measured in t+1 to t+3. Model (a) is a Probit model where the dependent variable indicates if a
given firm enters at least one new product market, defined as the first time appearance of positive sales in a given 3-digit SIC
industry where the firm has not generated sales previously. Model (b) is an OLS regression of the logarithm of (no. entries + 1).
Model (c) is an OLS regression of the logarithm of (new sales +1), where new sales is the total amount of sales generated in all
new industries. Model (d) is an OLS regression of product proximity based on textual analysis of firms’ 10k fillings by Hoberg and
Phillips (2015, 2010), multiplied by 100. Patent stock is the cumulative number of patents applied for since 1976. Exploitation and
exploration are the component measures derived from the PCA described above. Heteroscedasticity-robust standard errors are
clustered at the firm level and shown in parentheses. ***, **, * indicate statistical significance at the 1%, 5%, 10% level,
respectively.
Table 8 demonstrates an exploration strategy is always statistically significant below the 1%
level and positively related to (1) a firm’s propensity to enter a new market, (2) the number of
20
new markets entered as well as (3) sales generated in new to the firm markets. In terms of
economic magnitude the results indicate that a one standard deviation increase in exploration is
associated with an 11.5% increase in the propensity to enter at least one new market in the next
three years and a 15.3% increase in sales generated in those new to the firm markets.
Consistent with these findings for exploration, we also find that exploitation correlates
insignificantly and in two out three cases negatively with new product market entry, the total
number of product markets entered, and the sales firms generate in those markets. This picture
stays qualitatively the same even if we consider the same product market entry measures
observed 4 or 5 years after the observed focus on exploration and exploitation (not presented).
Table A1a in the Appendix illustrates that all results are more pronounced in terms of statistical
significance and economic magnitude if only firm-years are considered when firms filed at least
10 patents in a given year and our PCA patent portfolio measure is based on a more solid basis.
In this setting we consistently find a significant negative relation between firms’ focus on
exploitation and market entry. Table A1b further shows that the results hold after controlling for
the number of patents, where the number of patents itself demonstrates only weak explanatory
power.
Figure 7 illustrates the temporal relationship between an increase in the innovative search scores
and the amount of sales generated in new to the firm industries (only for firms with at least four
patents in given year). Sales are calculated as three year moving averages starting with the first
three years after observation of firms’ exploration and exploration scores, respectively. Both
effects become weaker as the time from the search strategy to commercialization increases.
21
Figure 7: Temporal correlation between search strategy and sales in new industries.
Table 3, column d, shows regressions of Phillips and Hoberg’s (2015) measure of product
proximity between firms based on textual analysis of firms’ 10k fillings (the number of
observations drops because measure is only available for the years 1996 onwards, hence our
calculation of the SIC overlap). The measure ranges from 0 to 100 (rescaled), where 0 means
largest possible distance to other firms in the product market, while 100 indicates maximal
possible overlap of a firm’s products with its competitors’ products. Consistent with the previous
results we find that a focus on exploitation increases comparability with competitors in the
product market, while a focus on exploration helps firms to move away from their competitors
(results are again robust with longer time lags).
MARA as an instrument
Despite impressive uptake of the explore/exploit model of innovative search in the organizations
and strategy literatures, there has been little rigorous identification of the idea empirically or
causal evidence that connects exploration and exploitation strategies to performance. In order to
strengthen causal inference from innovative search to subsequent new market entry, we use the
Michigan Anti-Trust Reform Act (MARA) of 1985 and anti-takeover regulations (ATO).
MARA inadvertently made noncompete agreements enforceable and has been used previously as
-.25
-.15
-.05
.05
.15
.25
coe
ffic
ient
siz
e
1 2 3 4 5 6 7 8
Time to Entry
b-coefficients 95%-confidence-interval
Exploitation and Sales in New Industries
-.25
-.15
-.05
.05
.15
.25
coe
ffic
ient
siz
e
1 2 3 4 5 6 7 8
Time to Entry
b-coefficients 95%-confidence-interval
Exploration and Sales in New Industries
22
an instrument to study within state mobility (Marx, Strumky, and Fleming 2009), brain drain
(Marx, Singh, and Fleming 2014), human capital and acquisition (Younge, Tong, and Fleming
2014), and human capital and firm valuation (Younge and Marx 2015). Empirically it appears
that MARA increased both exploration and exploitation though the effect of MARA on
exploitation remains small and barely significant. We discuss why MARA might have both
effects but remain agnostic on exact mechanisms here, as our intent is only to isolate the impact
of strategy on commercialization outcomes.
MARA could arguably increase both exploration and exploitation. Because MARA decreased
the mobility of engineers (Marx, Strumsky, and Fleming 2009), firms’ work forces may have
become stale, as engineers stayed with current employers. This might have caused greater
exploitation if firms had previously depended on hiring for new ideas. MARA could also have
increased the influx of different ideas, because engineers that did move within Michigan had to
move farther from their former employer in technological “distance,” in order to avoid being
prosecuted for their noncompete agreement (Marx 2013). Engineers that moved within
Michigan after MARA therefore made more career detours into new areas and based on this,
they invented more novel patents (at the expense of decreased productivity, see Arts and
Fleming, 2016). Michigan firms may have also have performed more explorative projects given
the increased stability of their workforce, because firms might have been less concerned about
employee departure and competitor appropriation. Conti (2014) demonstrated such an effect
following noncompete changes in Texas and Florida but not Michigan.
Firms operating in Michigan are considered treated, while the control group comes from firms in
states that had similar laws as Michigan before and after the MARA law change. We estimate the
corresponding differences in differences (DiD) models based on firm data ranging from 1979 to
1993, i.e. six years before and after MARA. In a first step, the exploration and exploitation
measures are taken as dependent variables. Table 9, columns a and b, contain the corresponding
results for exploitation and exploration, respectively. Firms in Michigan scored higher on
exploitation and exploration alike, though the effect size is more significant for exploration and
almost three times larger. As such, we would expect to see increased product market entry by the
treated firms. We next run the same regressions with the previously used measures of product
23
market entry as dependent variables and presence inside Michigan after MARA as the treatment.
Consistent with a move towards exploration, all market entry variables are positively and
significantly related to the treatment interaction (Table 9, columns c, d, and e). This result also
holds when we alternatively identify the influence of exploration by an IV regression. In this
case model b serves as the first stage regression. Table 9, columns f, g, and h, present the results
of the second stage, i.e. ‘exploration’ are now the predicted values from model b that carry
exogenous variation caused by MARA. Again, we see a significant and positive influence of
exploration on all our market entry variables. The size of the coefficients in the IV and DID
models are comparable. Table 9, model e, indicates that firms subject to MARA increased their
sales in new to firm markets by 59%. The corresponding IV regression, Table 9, model h,
indicates an increase of 59.1%. The propensity to enter a new market increased by 42.0%
according to the DID model (c) and 32.3% according to the IV model (f).
It appears that Michigan firms took advantage of the increased focus on exploration with new
market entry and performance. One reason for the considerably large effect could be that the
treated firms increased their exploration at the right time, when good market opportunities
existed. Firms also simultaneously increased their exploration and exploitation. Increasing both
has often been suggested as a particular successful strategy (Tushman and O’Reilly 2004),
formally modeled through simulation (Fang, Lee, and Schilling 2010) and empirically confirmed
with patent citations by (Manso et al. 2016). However, the rather large magnitudes could also
point to an undetected estimation bias that led to an overestimation, for instance, because other
unobserved market entry enabling factors that are correlated with our exploration measure, e.g.
increased demand, are not perfectly controlled for. In order to replicate our findings we
investigated using the staggered imposition of anti-takeover laws as a second and also arguably
exogenous shock to firms’ exploration focus.
24
Table 9 – MARA experiment
a b C d e f g h DID DID DID DID DID IV IV IV
Exploitation Exploration Entry 0/1 No. entries New sales Entry 0/1 No. entries New sales
log(pat stock) 0.668*** 0.024 0.102*** 0.024** 0.163*** 0.050** 0.017 0.094 (0.051) (0.036) (0.030) (0.010) (0.040) (0.021) (0.011) (0.058)
R&D int. 0.794*** 1.007*** 0.743*** 0.158** 0.824*** -1.478** -0.136 -2.123* (0.080) (0.133) (0.234) (0.059) (0.200) (0.753) (0.103) (1.053)
log(age) -0.210*** -0.071* -0.159*** -0.025** -0.136** -0.003 -0.004 0.070 (0.029) (0.037) (0.056) (0.008) (0.043) (0.059) (0.010) (0.096)
log(total assets) 0.202*** 0.325*** 0.035 0.016* 0.224*** -0.681** -0.078* -0.726* (0.034) (0.019) (0.036) (0.009) (0.037) (0.271) (0.041) (0.374)
Herfindahl ind. -0.093 0.233 0.025 -0.124 -0.714 -0.488 -0.192 -1.395 (0.244) (0.378) (1.106) (0.201) (0.983) (1.262) (0.227) (1.194)
Entry exp. -0.167* 0.085 0.132 -0.010 -0.223 -0.054 -0.035 -0.470 (0.086) (0.054) (0.133) (0.031) (0.206) (0.169) (0.034) (0.274)
Exploitation 0.035* 0.004 0.011 -0.158* -0.022 -0.244* (0.021) (0.005) (0.033) (0.085) (0.014) (0.117)
MARA 0.061* 0.202*** 0.444*** 0.059* 0.590**
(0.031) (0.040) (0.163) (0.027) (0.235)
Exploration 2.205*** 0.291* 2.926**
(0.807) (0.134) (1.168)
N 3100 3100 3100 3100 3100 3100 3100 3100
Industry, Time and State FE yes yes Yes yes yes yes yes yes
R2 / Pseudo R2 0.735 0.436 0.279 0.297 0.343 0.279 0.299 0.348
Notes: Models (a) and (b) are OLS regressions of exploitation and exploration measures, respectively, as derived from the PCA described above. Model (a) and (b) are estimated
controlling for exploration and exploitation, respectively. Dependent variables of models c to h are measured in t+1 to t+3. Models c and f are Probit models where the dependent
variable indicates if a given firm enters at least one new product market, defined as the first time appearance of positive sales in a given 3-digit SIC industry where the firm has not
generated sales previously. Models d and g represent regression of the logarithm of (no. entries + 1). Models e and g represent regressions of the logarithm of (new sales +1), where
new sales is the total amount of sales generated in all new to the firm industries. Patent stock is the cumulative number of patents applied for since 1976. Heteroscedasticity-robust
standard errors are clustered at the state level and shown in parentheses. ***, **, * indicate statistical significance at the 1%, 5%, 10% level, respectively.
25
Antitakeover as an instrument
The staggered introduction of antitakeover laws by American states in the late 1980s and early
1990s had a surprisingly strong and unexpected influence on patenting and provides a second
natural experiment. Following Atanassov 2013, who found that these law changes led to a
reduction in overall patenting activity, we focus on the Business Combination laws that were
introduced in different years by most states (see also Bertrand and Mullainathan, 2003, who
analyzed the effect of antitakeover laws on corporate governance performance, and the years in
which each state introduced Business Combination laws). “Business Combination laws impose a
moratorium (3 to 5 years) on specified transactions between the target and the acquirer holding a
specified threshold percentage of stock unless the board votes otherwise before the acquiring
person becomes an interested shareholder.”5
In order to analyze the effect of antitakeover laws on exploration/exploitation and market entry
we run DID models where the treatment indicator is a binary variable that marks all years that a
particular state has had an antitakeover law in effect. Due to the staggered introduction of the
antitakeover laws, firms in states that eventually got treated can still serve as a control group. We
removed all firms situated in California and Massachusetts from the control group as these states
saw a huge increase in patenting activity at the same time many other states introduced their
antitakeover laws, which may lead to spurious correlations (Lerner and Seru, 2015). For our
empirical test we restrict the sample to the years 1981 to 1995, i.e. four years before the first
introduction and four years after the last introduction of a business combination law. State fixed
effects in all our regressions account for remaining time-constant unobserved differences across
States. As we also employ time fixed effects and basically estimate a classic DiD model. That
means under the assumption that firms in the control and treatment follow similar trends our
treatment variable “postBC” captures the causal impact of the introduction of the BC laws on the
dependent variable of interest.
Table 10 details the same regressions as previously used with MARA. First, we check the impact
of Anti-takeover law introduction on exploitation and exploration, followed by estimating the
5 Business Combination laws were arguably the most effective law changes that made takeovers harder or more
costly to carry out. Other less significant changes are reported in Atanassov (2013).
26
impact on our market entry variables. Next, we present IV regression results, where model b, our
exploration regression, serves as the first stage. Apparently, antitakeover regulation did not
affect firms’ exploitation focus (model a) but significantly decreased firms innovation search
towards exploration (model b). We conjecture that the decline in exploration is related to fewer
opportunities for selling the firm to competitors or other firms, and generally reduced market
pressure to present new discoveries that please investors’ (possibly biased, see Fitzgerald et al.
2016) attention on novelty.
As with MARA, we remain agnostic on the exact mechanism, and focus on the effect of
antitakeover regulation on market entry instead. Models c, d, and e, represent DID regressions of
market entry. Consistent with a decreased focus on exploration we see a significantly decreased
propensity to enter new markets, a significantly decreased number of markets entered, and
insignificantly decreased new market entry success as measured by sales generated in new to the
firm markets. In terms of economic magnitude model c implies a reduction in the likelihood to
enter a new market by -14.9%. This decrease stems from a reduction in the exploration score of
0.105 points. Hence, the magnitude of the effect seems to be broadly in line with estimations
based on MARA where firms were associated with an increase in their exploration score by
0.202 points and a corresponding increase in the propensity to enter a new market by 42.0%.
The IV regressions are consistent with these results. Model f suggests that an equivalent decrease
of 0.105 in the exploration score reduces the propensity to enter a new market by 15.4%. Effect
sizes are considerably smaller compared to MARA but still large in economic magnitude. While
this may still point to estimation issues it is reassuring to find consistent results across
experiments and increases the possibility that the identified influence of exploration on market
entry may be causal.
27
Table 10 – Anti-takeover experiment
a b c d e f g h DID DID DID DID DID IV IV IV
Exploitation Exploration Entry 0/1 No. entries New sales Entry 0/1 No. entries New sales
log(pat stock) 0.685*** 0.064*** -0.018 -0.003 0.006 -0.102*** -0.020** -0.068 (0.015) (0.023) (0.019) (0.005) (0.030) (0.036) (0.009) (0.062)
R&D int. 0.564** 0.335*** 0.194** 0.043** 0.329** -0.244 -0.045 -0.057 (0.246) (0.123) (0.087) (0.018) (0.126) (0.179) (0.047) (0.334)
log(age) -0.135*** -0.081*** -0.022 -0.009* -0.045 0.084 0.012 0.048 (0.021) (0.030) (0.023) (0.005) (0.029) (0.056) (0.014) (0.089)
log(total assets) 0.158*** 0.280*** 0.088*** 0.021*** 0.254*** -0.278* -0.053 -0.069 (0.016) (0.018) (0.017) (0.004) (0.026) (0.156) (0.038) (0.268)
Herfindahl ind. 0.449* 0.009 -0.104 -0.011 0.404 -0.116 -0.013 0.393 (0.225) (0.231) (0.484) (0.121) (0.722) (0.483) (0.121) (0.721)
Entry exp. -0.134*** 0.024 0.306*** 0.062*** 0.305*** 0.274*** 0.056*** 0.277*** (0.046) (0.042) (0.042) (0.010) (0.064) (0.045) (0.011) (0.066)
Exploitation 0.099*** 0.003 0.001 0.009 -0.126* -0.025 -0.104
(0.030) (0.017) (0.004) (0.029) (0.065) (0.016) (0.111)
postBC 0.035 -0.105*** -0.138** -0.028* -0.121 (0.050) (0.037) (0.060) (0.015) (0.103)
Exploration 1.308** 0.264* 1.152
(0.567) (0.138) (0.980)
N 9520 9520 9520 9520 9520 9520 9520 9520
Industry, Time and State FE yes yes yes yes yes yes yes yes
R2 / Pseudo R2 0.722 0.388 0.148 0.156 0.191 0.148 0.299 0.348 Notes: Models (a) and (b) are OLS regressions of exploitation and exploration measures, respectively, as derived from the PCA described above. Model (a) and (b) are estimated
controlling for exploration and exploitation, respectively. Dependent variables of models c to h are measured in t+1 to t+3. Models c and f are Probit models where the dependent
variable indicates if a given firm enters at least one new product market, defined as the first time appearance of positive sales in a given 3-digit SIC industry where the firm has not
generated sales previously. Models d and g represent regression of the logarithm of (no. entries + 1). Models e and g represent regressions of the logarithm of (new sales +1), where
new sales is the total amount of sales generated in all new to the firm industries. Patent stock is the cumulative number of patents applied for since 1976. Heteroscedasticity-robust
standard errors are clustered at the state level and shown in parentheses. ***, **, * indicate statistical significance at the 1%, 5%, 10% level, respectively.
28
The role of close competitors in technological or market space
The efficacy of any strategy depends on opponents (Cockburn and MacGarvie 2011). Here we
look for an interaction between exploration strategy and crowding or competitors’ positions in
“technology space” space (Stuart and Podolny 1996; Aghion et al. 2005, Aharonson and
Schilling 2016). If firms face competitors that are active in the same technological areas or
markets it might be harder for those firms to realize the benefits from exploration as it may be
easier for close competitors to anticipate or follow search success. Close competitors are more
likely to see the value of a firm’s exploration; they are also in a better position to hire away
engineers and/or marketing and sales people and compete more quickly and effectively.
Now we estimate our previously presented IV regressions based on MARA and the ATO
experiments, including a dummy that indicates close competition as measured by falling in the
highest quartile of the close competitors distribution (this applies to all firms that have more than
13 close competitors in technological space; results are robust to considering at least 10 close
competitors) and an interaction term between this dummy and firms’ exploration scores
instrumented by the respective regulatory changes of MARA and ATO. Similarly, we include a
dummy indicating close competition in market space (to stay consistent with the same threshold
of 13 or more close competitors) and an interaction term between this dummy and firms’
instrumented exploration score.
Table 11 shows that exploration has the previously identified positive influence on market entry
but that this positive effect is significantly reduced when firms face strong competition in
technological space. While individual coefficients of close competition, exploration, and the
interaction term are sometimes statistically insignificant, they are always jointly significant
according to χ2–tests (models a and d) and F-tests (models b, c, e, and f), respectively. As we are
looking at different sample compositions (different time, type and location of firms) it is not too
surprising that the sizes of the coefficients vary across models. While these differences in effect
size appear large they are not significantly different. We see no consistent impact of market
competition on innovation search strategies or their efficacy.
29
Table 11 – Exploration, market entry, and competition in technology and market space
a b c d e f
IV-MARA IV-MARA IV-MARA IV-AT IV-AT IV-AT
Entry 0/1 No. entries New sales Entry 0/1 No. entries New sales
log(pat stock) 0.028 0.014 0.081 -0.101*** -0.021** -0.078 (0.038) (0.014) (0.072) (0.034) (0.008) (0.059)
R&D int. -2.356** -0.246 -3.576** -0.224 -0.048 -0.119 (1.151) (0.145) (1.602) (0.188) (0.049) (0.345)
log(age) -0.045 -0.004 0.072 0.081 0.014 0.069 (0.052) (0.009) (0.089) (0.057) (0.014) (0.088)
log(total assets) -0.668** -0.079* -0.745* -0.292* -0.057 -0.082 (0.267) (0.036) (0.368) (0.165) (0.040) (0.278)
Herfindahl ind. -0.416 -0.185 -1.353 -0.127 -0.020 0.391 (1.176) (0.223) (1.169) (0.485) (0.120) (0.716)
Entry exp. -0.223 -0.069* -0.633* 0.237*** 0.048*** 0.246*** (0.192) (0.035) (0.296) (0.045) (0.010) (0.061)
Exploitation -0.140* -0.021 -0.238** -0.131* -0.026 -0.116 (0.082) (0.012) (0.105) (0.067) (0.017) (0.114)
Close tech comp. 0.461 0.053 0.649 0.155 0.038 0.111 (0.311) (0.051) (0.396) (0.172) (0.039) (0.270)
Close market comp. -1.005*** -0.182*** -0.791*** -0.421*** -0.078*** -0.209*
(0.171) (0.032) (0.223) (0.108) (0.020) (0.112)
Exploration 2.110** 0.299** 3.078** 1.372** 0.293* 1.377 (0.847) (0.131) (1.232) (0.608) (0.146) (1.028)
Close tech. comp. x Explor. -0.163*** -0.044*** -0.396*** -0.117* -0.029** -0.267**
(0.055) (0.010) (0.121) (0.067) (0.014) (0.113)
Close mark. comp. x Explor. 0.323* 0.007 -0.110 0.086 -0.009 -0.230
(0.179) (0.058) (0.468) (0.076) (0.018) (0.152)
N 3100 3100 3100 9520 9520 9520
χ2-test / F-test 178.86*** 95.99*** 23.19*** 28.25*** 4.97*** 4.29***
Industry, Time and State FE yes yes yes yes yes yes
R2 / Pseudo R2 0.281 0.299 0.348 0.150 0.157 0.193
Notes: All dependent variables are measured in t+1 to t+3. Models a and d are Probit models where the dependent variable indicates
if a given firm enters at least one new product market, defined as the first time appearance of positive sales in a given 3-digit SIC
industry where the firm has not generated sales previously. Models b and e represent regression of the logarithm of (no. entries +
1). Models c and f represent regressions of the logarithm of (new sales +1), where new sales is the total amount of sales generated
in all new to the firm industries. Exploration are the first stage estimated values based on MARA in case of models a, b, and c.
Exploration are the first stage estimated values based on the Antitakeover change in case of models d, e, and f. Close comp. is a
dummy that indicates if firms fall into the highest quartile of the distribution of close competitors in technological space as measured
by a pair-wise technological proximity score higher than 0.95. χ2-test (models a and d) and F-test (models b, c, e, and f) scores
belong to tests of joint of significance of the close competition, exploration, and the corresponding interaction term coefficients.
Heteroscedasticity-robust standard errors are clustered at the state level and shown in parentheses. ***, **, * indicate statistical
significance at the 1%, 5%, 10% level, respectively.
30
Methodological discussion
A caveat of our empirical identification strategy is that exploration might not be the only channel
through which firms increase market entry and entry performance. In the DiD setup this would
remain a problem of unobserved heterogeneity, i.e. unobserved variables that co-vary with
exploration and market entry alike. The IV setup can deal with unobserved heterogeneity, but the
validity of the instrument relies on the exclusion restriction being fulfilled, which we cannot test
as we have only one instrument per setup. In this respect it is particularly reassuring to see
consistent results in two quasi-experiments that move our main explanatory variable of interest –
exploration – in opposing directions, and finding that our main dependent variables of interest –
measures of market entry – move in the expected opposing directions, too. It increases
confidence in our findings because it is rather unlikely that two different regulatory changes that
affected exploration through different channels in opposing ways had the same opposing
influences on unobserved third factors.
These concerns remain endemic to the strategy literature, which faces a fundamental challenge in
establishing causality; it is difficult if not impossible to replicate the randomized lab experiment
in the field and equally, the randomized lab experiment cannot speak to the richness of the whole
phenomenon. We approached this challenge by finding two exogenous shocks that pushed
treated firms towards particular search strategies and then observed the impact of those strategies
on new market entry and success in that entry. Despite the consistent results from two quasi-
experiments, we remain concerned about the unobserved path from exogenous shock to success.
It is improbable that any executives in the treated firms recognized the possible impact of the
shocks on their search strategy at the time. This is even more likely if you accept the severe
uncertainty of research and development processes – executives rarely can predict or even
understand the output of their research labs. More likely, decision makers became aware that
their technology assets had changed; in the case of exploration, they might have realized that
they now had innovations that would enable new products and new market entry. This is not
unlikely, if one assumes that managers are typically on the look out for new product and market
opportunity. This is essentially the basis of our informal model, however, we must acknowledge
that the model rests upon these untested assumptions.
31
Patent data have all sorts of problems (Lerner and Seru 2015), however, one of their strengths is
that they allow relatively consistent comparison of inventions and portfolios across time, firms,
and industries. This allows us to include multiple industries in our analysis, and use entry and
sales into new industries to measure performance. Rather than defining and focusing on one
industry, the patent record enabled us to consider all patenting industries simultaneously and to
use financial data to observe market entry. Hopefully our relatively simple principal components
analysis, performed on readily available patent data, can be combined with additional
experiments in the future, and thus enable stronger causal inference.
Theoretical implications
Schumpeter’s “perennial gale” metaphor of creative destruction (Schumpeter 1942) spawned a
classic and central theme of the technology strategy literature; how do innovating incumbents
avoid obsolescence and irrelevance? The question reappears in many forms and the answer
rarely reassures the previously successful market leader. Incumbents rarely survive the era of
ferment (Abernathy and Utterback 1975), discontinuous and competency destroying change
(Tushman and Anderson 1986), developments outside their absorptive capacity (Cohen and
Levinthal 1990), seemingly simple architectural change (Henderson
and Clark 1990) and even incremental improvement (Christiansen 1998). Stuart and Podolny
(1996) and Sorenson and Stuart (2003) illustrate the phenomenon with patent data. While not
explicitly targeted at technology strategy, March’s “competency trap” image contributed another
appropriate and popular metaphor to the literature (March 1991). All of this work describes and
stresses the difficulty of successful and continuous innovation, of converting raw invention and
technology into successful products and profits.
The threat of competition is never absent from this genre, indeed, the biggest fear of any
incumbent is that someone else will invent a breakthrough that sweeps all others away. Under-
emphasized, however, is the interaction of others’ technologies with the focal firm’s search
strategy, and the difficulties that these interactions can cause the incumbent. In contrast, this
work demonstrates how the efficacy of a firm’s search strategy is strongly influenced by where
that firm sits in technology space. Firms in crowded neighborhoods appear to have a much more
difficult time in taking a technology from invention to innovation and commercial success. In the
32
parlance of landscape models of search, local hilltops are always difficult to move away from,
and especially when your neighbors are crowding your view.
Surprisingly, competitors with a similar market profile appear to have little effect on innovation
search strategy, though this might have resulted from reliance on the correlation in SIC code
distributions. Phillips and Hoberg’s (2015) measure of product proximity between firms based
on textual analysis of firms’ 10k fillings may be a closer measure to the idea of product
competition; the SIC measure is closer to a market distribution of profile similarity. Current
work aims to refine these measures of market and product competition.
One of the methodological strengths of this work is that it demonstrated consistent results from
the application of simple statistics to two quasi-experiments that exogenously nudged firms’
search strategies in opposite directions; this method enables and motivates new theory. Firms on
average tend to exploit more as they age and stop exploring. On the other hand, the principal
components analysis and illustrations clearly show that exploration and exploitation are not polar
opposites. Most intriguingly, MARA pushed firms to both explore and exploit and Intel appears
to be able to manage both. Furthermore, these strategies and their efficacy appear to be
uninfluenced by competitors who operate in a similar distribution of markets. These results
highlight the importance for continued theoretical development in how firms can balance
exploration and exploitation.
Conclusion
How does a firm’s technology strategy influence its commercial success? We contributed to
answering this question in a variety of ways. We began by reducing a variety of widely available
and common patent measures into two surprisingly complete principal components. These
components clearly loaded onto axes of exploitation and exploration and thus operationalized
March’s (1991) theory (given the simplicity of the statistical technique and the availability of
these measures for the past 20 years, we are surprised that this approach does not appear to have
gained traction yet). These components enabled us to identify and leverage changes in legal
regimes that exogenously varied firms’ search strategies. Armed with the component measures
and an exogenous shock, we could strengthen causal inferences about search strategy and
33
commercial success. In particular, we showed consistent results that exploration precedes new
product and market entry, that exploitation does the opposite, and that an exploration search
strategy is less efficacious when a firm’s patent portfolio correlates more closely with
competitors. In other words, and consistent with much of the literature, bold innovation strategies
are much more difficult when a firm starts from a crowded technological neighborhood.
The work speaks to three challenges in the technology strategy literature. First, it theorized
about one pathway (of many) from invention and innovation search strategy to commercial
success. Second, it offered measures, methods, and illustrations of this pathway. Taken
together, these demonstrated the pathway in greater detail and avoided jumping from patent
counts straight to profits or valuation. Finally, it applied two quasi-experiments that had
opposite effects on strategy but consistent effects on outcomes. These experiments enabled us to
strengthen causal inference.
The results suggest a number of follow on studies. The concepts and measures of technology
and market space need much refinement. Currently, the measures aggregate a firm’s entire
patent or market portfolio and calculate an overall correlation with other aggregated portfolios.
Especially for large and diverse firms, however, this measure obscures individual research
programs and interactions of individual programs. For example, GE owns patents in
photovoltaic and jet turbine technology, yet they probably have few if any competitors whose
portfolios also contain both areas (though Siemens perhaps might do so). While it will be
difficult to measure the commercial success of individual research programs, we can observe
some of the successful outputs of individual programs in the patent record. More nuanced
interactions between technology and market space can also be studied, though in the present
case, the negative effect of technological crowding illustrated does not appear to have been
influenced by the market.
The visualizations of exploration and exploitation suggest a number of studies. While Intel has
long been held up as a paragon of a well-run technology firm, the work illuminates specific
mechanisms of this success. In particular, how did Intel consistently simultaneously improve
both its exploration and exploitation? While at the same time, competing in a more crowded
34
market space? Ambidextrous organizations are certainly desirable (Tushman and O’Reilly 2004;
Fang, Lee, and Schilling 2010), but detailed and systematic evidence on how they are
accomplished might now be possible. The measures and modern visualization tools will enable
researchers to investigate how research and development portfolios evolve and how project
selection and management translate into the firm’s overall commercial success. Future research
should combine qualitative fieldwork with recent development in visualization and data
analytics. Now that we can clearly observe exploration and exploitation, it would be interesting
to find firms besides Intel that can do both, and more interestingly, to understand just how they
accomplish that.
35
References
Aghion, P. and N. Bloom, P. P. Blundell, R. Griffith, P. Howitt (2005). “Competition and
Innovation: an Inverted-U Relationship.” The Quarterly Journal of Economics 120 (2): 701-728.
Aharonson, B. & Schilling, M.A. 2016. Mapping the technological landscape: Measuring
technology distance, technological footprints, and technology evolution. Research Policy,
45: 81-96.
Cockburn, I. M. and M. MacGarvie (2011) "Entry and Patenting in the Software Industry,"
Management Science, INFORMS, vol. 57(5), pages 915-933, May.
Cohen, L. and K. Diether, C. Malloy. 2013. “Misvaluing Innovation.” RFS 26 (3):635-666.
Fang, C., Lee, J. & Schilling, M.A. 2010. Balancing exploration and exploitation through
structural design: advantage of the semi-isolated subgroup structure in organizational learning.
Organization Science, 21:625-642.
de Figueiredo, J. M. and Kyle, M. K. (2006), Surviving the gales of creative destruction: the
determinants of product turnover. Strat. Mgmt. J., 27: 241–264. doi:10.1002/smj.512
Hall, B.H. (1993) “The Stock Market’s Valuation of R&D Investment During the 1980’s.”
American Economic Review, 83(2):259-264.
Hannan, M. 1998. “Rethinking Age Dependence in Organizational Mortality: Logical
Formalizations.” American Journal of Sociology, Vol. 104, No. 1 (July 1998), pp. 126-164.
He, Z.L., & Wong, P.K. (2004). Exploration vs. exploitation: An empirical test of the
ambidexterity hypothesis. Organization Science, 15(4), 481–494.
Henderson, R. and K. Clark 1990. Architectural innovation: The reconfiguration of existing
product technologies and the failure of established firms. Administrative Science Quarterly 9-30.
Hirshleifer D., Hsu P., Li D. 2013. “Innovative efficiency and stock returns.” Journal of
Financial Economics 107 (2013) 632–654.
Lavie, D. and Rosenkopf, L. (2006). Balancing exploration and exploitation in alliance
formation. Academy of Management Journal, 49(4), 797–818.
Lavie, D., U. Stettner, and M. Tushman. (2010) "Exploration and Exploitation within and across
Organizations." Academy of Management Annals 4: 109–155.
Lin, Z.J., Yang, H., & Demirkan, I. (2007). The performance consequences of ambidexterity
in strategic alliance formations: Empirical investigation and computational theorizing.
Management Science, 53(10), 1645–1658.
Nerkar, A., & Roberts, P. (2004). Technological and Product-Market Experience and the
36
Success of New Product Introductions in the Pharmaceutical Industry. Strategic
Management Journal, 25(8/9), 779–799.
Podolny, J. M.. (1994). Market Uncertainty and the Social Character of Economic
Exchange. Administrative Science Quarterly, 39(3), 458–483.
Romanelli, E., & Tushman, M.L. (1994). Organizational transformation as punctuated
equilibrium: An empirical test. Academy of Management Journal, 37(5), 1141–1166.
Rosenkopf, L., & Nerkar, A. (2001). Beyond local search: Boundary-spanning, exploration,
and impact in the optical disk industry. Strategic Management Journal, 22(4), 287–306.
Rowley, T.B., Behrens, D., & Krackhardt, D. (2000). Redundant governance structures:
An analysis of structural and relational embeddedness in the steel and semiconductor
industries. Strategic Management Journal, 21, 369–386.
Schumpeter, J. (1942). Capitalism, Socialism and Democracy. London: Routledge. pp. 82–83.
Schilling, M. 2017. The Strategic Management of Technological Innovation. McGraw, NY., NY.
Silverman, B. (1999) “Technological Resources and the Direction of Corporate
Diversification: Toward an Integration of the Resource-Based View and Transaction Cost
Economics.” Management Science, 45(8):1109-1124.
Sorensen, J. and T. Stuart. 2001. “Aging, Obsolescence, and Organizational Innovation.”
Administrative Science Quarterly March 2000 vol. 45 no. 1 81-112.
Staw, B.M., Sandelands, L.E. and Dutton, J.E. (1981), “Threat‐ rigidity effects in
organizational behavior: a multilevel analysis”, Administrative Science Quarterly, 26:4:501‐ 24.
Stuart, T. E. and Podolny, J. M. (1996), Local search and the evolution of technological
capabilities. Strat. Mgmt. J., 17: 21–38. doi:10.1002/smj.4250171004.
Tushman, M.L., & Anderson, P. (1986). Technological discontinuities and organizational
environments. Administrative Science Quarterly, 31(3), 439–465.
Tushman, M. and C. O’Reilly (2004). The Ambidextrous Organization. Harvard Business
Review.
Uotila, J., Maula, M.V.J., Keil, T., & Zahra, S.A. (2009). Exploration, exploitation, and
financial performance: Analysis of S&P 500 corporations. Strategic Management
Journal, 30, 221–231.
Wang, R. and M. Shaver 2016. “The Multifaceted Nature of Competitive Response:
Repositioning and New Product Launch as Joint Response to Competition.” Strategy Science
Vol. 1, No. 3, September 2016, pp. 148-162.
37
Appendix
Table A1a – Exploration/Exploitation and product market entry, at least 10 patents filed
a b c d
Dependent variable Entry 0/1 No. Entries New sales Prod. proximity
log(pat stock) 0.112*** 0.031*** 0.208*** -0.022***
(0.034) (0.009) (0.063) (0.005)
R&D int. -1.524** 0.011 0.988** 0.340***
(0.664) (0.067) (0.445) (0.045)
log(age) -0.046 -0.018* -0.133* -0.011**
(0.036) (0.010) (0.070) (0.005)
log(total assets) 0.107*** 0.034*** 0.354*** 0.011**
(0.031) (0.009) (0.056) (0.005)
Entry exp. -0.035 -0.021 -0.142 -0.042***
(0.052) (0.015) (0.109) (0.008)
Herfindahl ind. 0.007 0.009 -0.070 -0.029
(0.373) (0.121) (0.815) (0.078)
Exploitation -0.085** -0.025*** -0.124* 0.013**
(0.035) (0.009) (0.065) (0.005)
Exploration 0.078*** 0.019*** 0.161*** -0.013***
(0.022) (0.006) (0.045) (0.003)
N 7274 7274 7274 2009
Industry and time fixed effects yes yes yes yes
R2 0.188 0.192 0.211 0.573 Notes: All dependent variables are measured in t+1 to t+3. Model (a) is a Probit model where the dependent variable indicates if a
given firm enters at least one new product market, defined as the first time appearance of positive sales in a given 3-digit SIC
industry where the firm has not generated sales previously. Model (b) is an OLS regression of the logarithm of (no. entries + 1).
Model (c) is an OLS regression of the logarithm of (new sales +1), where new sales is the total amount of sales generated in all
new industries. Model (d) is an OLS regression of product proximity based on textual analysis of firms’ 10k fillings by Hoberg and
Phillips (2015, 2010), multiplied by 100. Patent stock is the cumulative number of patents applied for since 1976. Exploitation and
exploration are the component measures derived from the PCA described above. Heteroscedasticity-robust standard errors are
clustered at the firm level and shown in parentheses. ***, **, * indicate statistical significance at the 1%, 5%, 10% level,
respectively.
38
Table A1b – Exploration/Exploitation and product market entry, at least 10 patents filed
plus control for number of patents filed.
a b c d
Dependent variable Entry 0/1 No. Entries New sales Prod. proximity
log(pat stock) 0.106*** 0.029*** 0.186*** -0.025***
(0.035) (0.009) (0.064) (0.005)
R&D int. -1.589** -0.004 0.788* 0.309***
(0.673) (0.071) (0.465) (0.044)
log(age) -0.044 -0.017 -0.124* -0.009**
(0.036) (0.011) (0.071) (0.005)
log(total assets) 0.100*** 0.032*** 0.330*** 0.007
(0.033) (0.009) (0.059) (0.006)
Entry exp. -0.035 -0.021 -0.140 -0.042***
(0.052) (0.015) (0.109) (0.008)
Herfindahl ind. 0.006 0.008 -0.077 -0.037
(0.371) (0.121) (0.810) (0.078)
log(patents) 0.067 0.019 0.248* 0.036**
(0.088) (0.021) (0.142) (0.015)
Exploitation -0.121** -0.035** -0.257*** -0.006
(0.060) (0.014) (0.098) (0.010)
Exploration 0.068*** 0.017** 0.127*** -0.017***
(0.024) (0.007) (0.048) (0.004)
N 7240 7240 7240 2009
Industry and time fixed effects yes yes yes yes
R2 0.165 0.192 0.211 0.577 Notes: All dependent variables are measured in t+1 to t+3. Model (a) is a Probit model where the dependent variable indicates if a
given firm enters at least one new product market, defined as the first time appearance of positive sales in a given 3-digit SIC
industry where the firm has not generated sales previously. Model (b) is an OLS regression of the logarithm of (no. entries + 1).
Model (c) is an OLS regression of the logarithm of (new sales +1), where new sales is the total amount of sales generated in all
new industries. Model (d) is an OLS regression of product proximity based on textual analysis of firms’ 10k fillings by Hoberg and
Phillips (2015, 2010), multiplied by 100. Patent stock is the cumulative number of patents applied for since 1976. Exploitation and
exploration are the component measures derived from the PCA described above. Heteroscedasticity-robust standard errors are
clustered at the firm level and shown in parentheses. ***, **, * indicate statistical significance at the 1%, 5%, 10% level,
respectively.
39
Table A2 – Summary statistics – MARA sample
Variable N Mean Median Sd Min Max
Exploitation 3100 -0.409 -0.743 1.851 -3.738 6.260
Exploration 3100 0.0428 0.0282 1.403 -3.053 5.144
R&D int. 3100 0.0849 0.0482 0.159 0 2.882
log(age) 3100 2.083 2.303 0.845 0 3.135
log(total assets) 3100 12.03 11.82 2.167 3.807 19.08
Enrtry exp. 3100 0.977 0 1.646 0 14
HHI 3100 0.167 0.125 0.115 0.0386 0.973
Entry 0/1 3100 0.193 0 0.395 0 1
No. entries 3100 0.288 0 0.703 0 7
log(new sales) 3100 0.922 0 2.123 0 10.21
Close comp. 3100 13.20 1 25.65 0 125 Notes: This table reports summary statistics of firm level measures used in the study. Exploitation and exploration are the
component measures derived from the PCA described above. R&D int. is R&D expenditures divided by total assets. Age is the
number years since first time occurrence in Compustat. Entry 0/1 is a binary variable that indicates if a given firm enters at least
one new product market, defined as the first time appearance of positive sales in a given 3-digit SIC industry where the firm has
not generated sales previously. No. entries is the number of newly entered industries, defined as the total number of industries
where the firm generates sales for the first time in a given year. New sales is the total amount of sales generated in all new industries
where the firm did not generate sales beforehand. The latter three variables are measured as the sum over the years t+1 to t+3. Close
comp. is the number of firms with a patent portfolio that correlates with a value of at least 0.95 with the given firm’s patent portfolio.
Table A3 – Summary statistics – Antitakeover sample
Variable N Mean Median Sd Min Max
Exploitation 9520 -0.174 -0.543 1.948 -3.774 6.910
Exploration 9520 -0.00836 -0.0716 1.410 -3.259 5.169
R&D int. 9520 0.0573 0.0269 0.157 0 9.753
log(age) 9520 2.271 2.565 0.857 0 3.258
log(total assets) 9520 12.61 12.57 2.190 4.533 19.34
Enrtry exp. 9520 1.284 1 1.873 0 15
HHI 9520 0.192 0.143 0.140 0.0386 1
Entry 0/1 9520 0.199 0 0.399 0 1
No. entries 9520 0.278 0 0.648 0 7
log(new sales) 9520 0.992 0 2.203 0 10.79
Close comp 9520 14.77 1 34.40 0 229 Notes: This table reports summary statistics of firm level measures used in the study. Exploitation and exploration are the
component measures derived from the PCA described above. R&D int. is R&D expenditures divided by total assets. Age is the
number years since first time occurrence in Compustat. Entry 0/1 is a binary variable that indicates if a given firm enters at least
one new product market, defined as the first time appearance of positive sales in a given 3-digit SIC industry where the firm has
not generated sales previously. No. entries is the number of newly entered industries, defined as the total number of industries
where the firm generates sales for the first time in a given year. New sales is the total amount of sales generated in all new industries
where the firm did not generate sales beforehand. The latter three variables are measured as the sum over the years t+1 to t+3. Close
comp. is the number of firms with a patent portfolio that correlates with a value of at least 0.95 with the given firm’s patent portfolio.