Upload
others
View
2
Download
0
Embed Size (px)
Citation preview
Management of Science, Serendipity, and ResearchPerformance: Evidence from a Survey of Scientists in
Japan and the U.S.
Kota Murayamaa,∗, Makoto Nireib, Hiroshi Shimizub
aDepartment of Economics, Northwestern University, 2001 Sheridan Road, Evanston, IL60208, United States
bInstitute of Innovation Research, Hitotsubashi University, 2-1 Naka, Kunitachi, Tokyo186-8603, Japan
Abstract
In science, research teams are increasing in size, which suggests that science
is becoming more organisational. This paper aims to empirically investigate
the effects of the division of labour in management and science on serendip-
ity, which has been considered one of the great factors in science. Specifi-
cally, in examining the survey of scientists conducted in Japan and the U.S.,
this paper treats the following questions: Does pursuing serendipity really
bring about better scientific outcomes? How does the division of labour
in science influence serendipity and publication productivity? The empir-
ical results suggest that serendipity actually brings about better research
quality on average. It also finds that if the managerial role is played by a
leading scientist in the team, this is positively associated with the quality
of the paper through allowing researchers to pursue serendipitous findings.
In contrast, if the managerial role and leading research role are played by
different members, this has a positive association with the number of papers
published, as the project size becomes larger. These results indicate there is
∗Corresponding author. E-mail: [email protected]
Preprint submitted to Research Policy October 11, 2014
a trade-off between serendipity and publication productivity in science via
who plays the leading role in research and management.
1. Introduction
Would Alexander Fleming have discovered penicillin if he had been part
of a large research team? Would he have changed his research plan on
influenza to explore a culture contaminated with a fungus in 1928, if his
research project had been managed by an efficient project manager? By fo-
cusing on serendipity and productivity in science, this paper aims to explore
the relation between the management of science and its research outcomes,
in three steps.
The first step explores the nature of serendipity in science. Serendipity
is regarded as one of the most important aspects of science. Fleming’s
discoveries of the enzyme lysozyme in 1923 and penicillin from the mould
Penicillium notatum in 1928 are frequently cited examples of serendipity.
The cosmic background radiation identified by the Bell Lab scientists Arno
Penzias and Robert Wilson; the circular structure of benzene discovered
by Friedrich Kekule; X-rays developed by Antoine Henri Becquerel; and
Hans Christian Ørsted’s finding that electric currents create magnetic fields,
are also well-quoted examples of serendipity. It seems that many major
discoveries have been made by people who were looking for something very
different.
Much of the anecdotal evidence suggests that serendipity does indeed
have a positive effect on the quality of research. However, is serendipity a
good thing to pursue? One might think that serendipity does not necessar-
ily bring about better results because unintended findings occur randomly.
2
Or one might think that serendipity brings about better results because a
scientist does not change their research plan and pursue unintended findings
unless they expect that the change would be worth pursuing. This might
also be thought because such anecdotes in the history of science only illus-
trate successful results. Thus, considering serendipity as a significant factor
in science might be biased. It is therefore necessary to make a neutral defi-
nition of serendipity, and conduct empirical investigations which specify its
relations to the quality of research; such studies have been lacking in the
literature. By following the definition of serendipity provided by Stephan
(2010) as “the act of finding answers to questions not yet posed,” this paper
investigates whether pursuing serendipitous findings has a positive effect on
the quality of research.
The second step considers the effect of the division of labour in research
and management on serendipity. How can unintended findings be explored
when management and coordination are of importance in science? As will
be reviewed in the next section, the size of research projects has been in-
creasing. Inter-disciplinary and inter-organisational research has been of
significance to the performance of research and development (Agrawal and
Goldfarb, 2008). Prioritising in scientific discovery has also increased (Elli-
son, 2002; Stephan and Levin, 1992). Research is increasingly accomplished
in teams across nearly all fields (Wuchty et al., 2007). This indicates that
management, such as setting a research goal, planning the research proce-
dure, organising the research team, coordinating the members’ efforts, and
managing a research schedule, is increasingly important so as to achieve the
research goals effectively and efficiently.
Serendipity apparently happens in a random manner, implying that it
is not manageable. However, management studies have indicated that cer-
3
tain managerial settings, such as a close relation between a corporate R&D
laboratory and the business divisions, transferring managerial power to the
on-site manager, and the use of external managerial resources, can promote
exploration and allow the flexible pursuit of business opportunities (Ches-
brough, 2003; Nonaka and Takeuchi, 1995). This suggests that a certain
managerial setting in science can promote a flexible pursuit of serendipitous
findings.
When a scientist encounters a serendipitous finding, they are faced with
an important choice: to be flexible and change the research plan to pursue
the serendipitous event, or to stick closely to the initial plan. A serendip-
itous finding comes unexpectedly in the form of a very crude and nascent
condition. Thus, the scientist is forced to make an intuitive decision whether
to pursue it or not. As is reviewed in the following section, this choice is
difficult, particularly when the scientist is working as part of a research team
managed by a competent and efficient project manager. This situation is
seen not only in science but also in business management. This issue is
related to the classical managerial challenge of whether to use top down
or bottom up management. If managerial power is transferred to the im-
mediate director, they can fully desterilise uncodified and tacit knowledge,
and use managerial resources in the context of the actual situation. How-
ever, if a hierarchical managerial role is played top down, findings based on
ground level intuition are seldom used. A centralised bureaucracy cannot
readily adopt new ideas or easily adapt to environmental changes, due to
its formalisation (Gouldner, 1954; Merton, 1957; Selznick, 1949). Directing
its attention to the allocation of managerial and leading research roles, this
paper explores the effects of the division of labour on serendipity.
The third step concerns the effect of the division of labour in research
4
and management on research productivity. One of the advantages of the
division of labour is the increased efficiency resulting from specialisation and
concentration on a single subtask. Thus, if a leading scientist is separated
from a managerial role, they can focus on research and increase productivity.
A specialised project manager can also be fully responsible for the progress
of a research project. Top down hierarchical management facilitates the
completion of the original research goal. In other words, the second and
third steps touch on the dilemma that exists in management: exploration
versus exploitation (March, 1991).
Through exploring a survey of scientists, this paper reports the follow-
ing results. First, serendipity has a positive association with citation of
papers. This suggests that the management of science needs to seriously
take serendipity into consideration because it is one of the important factors
in scientific discovery. Secondly, the integration of a managerial and leading
research role has a positive association with serendipity. This is consis-
tent with the coordination cost framework, which indicates that integration
reduces the costs of the coordination between management and actual re-
search, and provides scientists with flexibility in their research. Thirdly, the
separation of management from research has a larger, positive association
with the number of papers, as the project size becomes larger. It must
be noted that our empirical results are patterns of associations between
serendipity, research quality, and management. However, our empirical re-
sults suggest a trade-off between serendipity and productivity in science via
considering who plays the managerial and leading research roles in research
management. The findings of this paper provide managerial and policy im-
plications for the management of science. Since the size of research projects
in science has been growing, the role of the research manager is of increasing
5
importance to research performance. The findings of this paper suggest that
a bureaucratic and formalistic research manager can block a leading scientist
from approaching an initial research plan flexibly and pursuing serendipitous
findings, which are a source of quality scientific discoveries.
The remaining part of this paper is organised as follows. Section 2 de-
fines serendipity and reviews the previous literature on the management of
science, serendipity, and productivity. In Section 3, we introduce the hy-
potheses, data description, definitions of the variables, estimation models,
and some estimation issues. Section 4 presents the estimated results and
robustness checks. Section 5 summarises the findings, considers managerial
and policy implications, and discusses three limitations for future research.
2. Management and Serendipity
Research is rarely undertaken in isolation; it is increasingly carried out
by a team. The mean number of authors per paper has increased from 2.8
in 1981 to 4.2 in 1999 and team size in science has increased by 50% over a
19-year period (Adams et al., 2005).
There are several factors behind this trend in increasing team size. Sev-
eral studies have shown that collaborative research produces better out-
comes with higher citation rates (Andrews, 1979; Presser, 1980; Sauer, 1988;
Wuchty et al., 2007). The internet and institutional change have decreased
communication costs and promoted increasing team size (Agrawal and Gold-
farb, 2008). The increase of team size in scientific research in the U.S. has
been attributed to the deployment of the National Science Foundation’s
NSFNET and its connection to networks in Europe and Japan after 1987
(Adams et al., 2005). Advances in research equipment (e.g., cyclotron, parti-
cle accelerators, and high-flux research reactors) have increased both collab-
6
oration and team size. Experimental design has also changed from table-top
experiments to large-scale projects. This, too, accompanies changes in the
pattern of collaboration among researchers because the use of one of these
new experimental tools requires several different sets of expertise simulta-
neously.
Another trend in science, discussed in the literature on the management
of science, is related to diversity. Many researchers have suggested that di-
versity in a research team can lead to a greater level of creativity (Allen,
1977; Garvey, 1979; Kasperson, 1978; Pelled et al., 1999). Singh and Flem-
ing (2010) argued that collaboration reduces the probability of very poor
outcomes due to more rigorous selection processes and greater recombinant
opportunities in the creative searches. Zuckerman (1977) showed that nearly
two-thirds of the 286 Nobel Prize winners named between 1901 and 1972
were honoured for work they did collaboratively. By investigating the con-
ditions under which major discoveries or fundamentally new knowledge oc-
cur in science, Hollingsworth (2006) demonstrated that scientists are likely
to develop new and alternative ways of thinking when they interact with
other scientists with diverse areas of expertise and backgrounds. With the
advances in information and communication technology, and institutional
changes, scientists can obtain relevant but different knowledge by collabo-
rating with other scientists in areas outside their own specialties. Accessing
external complementary knowledge and expertise through networking be-
comes significant when promoting innovation, not only in business, but also
in science (Fleming et al., 2007; Hagedoorn, 2002; Heinze et al., 2009; Powell
et al., 1996).
Furthermore, competition in science becomes fiercer. The race to be first
in science has intensified (Ellison, 2002; Stephan and Levin, 1992). There is
7
competition not only for scientific priority in scientific discovery, but also for
research funding. Thus, it is increasingly important for a research team to
choose a research area and method and to set a research goal to minimise the
threat of being “scooped” (Dasgupta and David, 1994; Stephan and Levin,
1992).
As research becomes large scale, requires a high level of technical and
scientific knowledge, and competition becomes fiercer, the management of
science has become increasingly important. Managing and coordinating the
research processes and the different sets of expertise, and synchronising the
efforts towards a team goal, do not happen naturally (Barnard, 1938; Si-
mon, 1976). As research teams become larger, research becomes more inter-
disciplinary and inter-organisational, and competition becomes fiercer, the
role played by research management will be greater.
Even though the literature on the economics of science has grown (Stephan,
2010), the study of the management of science has been quite limited. The
management of science has not yet been well investigated in management
studies; nevertheless, studies have accumulated on organisational structure,
managerial patterns, and the performance of firms. One of the central points
in management studies involves the allocation of authority and responsibil-
ities. For example, an on-site manager has better access to information on
demand, created through the interaction between customers and front-line
workers. Therefore, if customer needs and market trends frequently change,
the decentralised allocation of authority and responsibilities are suitable for
adapting flexibly to a slippery demand. A decentralised and less formalised
management, which allows a high degree of flexibility, is suitable when an
organisation faces many exceptional problems and problem solving is not
easy (Perrow, 1967; Woodward, 1965). This suggests that the decision-
8
making should be carried out where the important information is gathered
and knowledge is created if environmental change is uncertain but highly
frequent.
Why is centralised and hierarchical management not considered to be ef-
fective in a less stable environment? This is related to the division of labour
and its coordination costs (Becker and Murphy, 1992; Lawrence and Lorsch,
1967; Thompson, 2003). If the division of labour and the specialisation of
individual units are highly developed in an organisation, this increases pro-
ductivity because it allows individuals to concentrate on narrowly defined
tasks and to accumulate specialised knowledge about them. However, ac-
cording to the coordination cost framework, the advantage of the extensive
use of the division of labour is diminished by environmental complexity and
coordination costs. If an environment frequently changes, it is necessary for
an organisation to re-engineer the nature of the subtasks and reconfigure the
relations between these subtasks in order to adapt flexibly to the environ-
mental change. If the division of labour and the specialisation of subtasks
are highly developed, the specialised subtasks need to be coordinated in or-
der to achieve overall efficiency. However, according to the coordination cost
framework, costs are increased if individuals do not have a general knowl-
edge of the subtasks complementary to theirs and of the hierarchical subtask
relations (Arora and Gambardella, 1994).
In spite of the fact that the division of labour and its management in
science have become more important since the sizes of research teams have
increased, this aspect of the management of science has not yet been ad-
dressed by research. Thus, taking into account the division of labour and
using the framework of coordination costs, this paper, which considers the
research team to be closely linked with specialisation and the division of
9
labour, explores the effects of specialisation in science and the role of man-
agement in serendipity and the productivity of scientific research.
The primary focus of this paper is on exploring the effects on serendipity
of the division of labour between management and science research. The
word “serendipity” was coined by the novelist Horace Walpole, who was
inspired by the Persian fairy tale, “Three Princes of Serendip”. Merton
and Barber (2004) explored how the word was unexpectedly popularised
without a clear definition from its 1754 coinage to the twentieth century.
In scientific circles, the word has been used since the nineteenth century,
when the importance of unplanned and accidental factors in the making of
scientific discoveries gained recognition. Serendipity has been noted for its
role in the work of inventors and entrepreneurs, by persons such as George
W. Merck, a president of Merck & Co., and Willis Whitney, a director of
research of the General Electric Laboratories.
In the colloquial sense, serendipity is the making of happy and unex-
pected discoveries. Many anecdotal stories reveal how unintentional find-
ings have yielded unexpectedly fortunate results. Many great discoveries,
such as penicillin, X-rays, celluloid, and artificial sweetener, have been ut-
terly fortuitous, making the concept of serendipity not well-operationalised
(Roberts, 1989; Shapiro, 1986). It is uncertain whether the accidental na-
ture of serendipity is linked to the nature of the discovery process or the
unexpected impact of the discovery. However, upon closer examination, it
is obvious that the unplanned and accidental nature of serendipity is con-
nected only with the discovery process. This is reflected in official defini-
tions of the word. For instance, The American Heritage Dictionary of the
English Language, fourth edition, defines serendipity as “the faculty of mak-
ing fortunate discoveries by accident.” Furthermore, distinguishing between
10
the unexpected and the accidental is difficult, especially when research in-
volves exploration of the unknown. In order to operationalise the concept of
serendipity, it is therefore appropriate to think of serendipity as “the act of
finding answers to questions not yet posed” (Stephan, 2010). This definition
focuses not only on the discovery process, but also on the relation between
discovery and the specific research question. Even though this definition
directs our attention to the extent to which the discovery answers a ques-
tion not yet posed, the present paper adopts this definition and explores the
relation between management and serendipity in science.
Serendipity has been considered one of the most important character-
istics of science. As mentioned above, many great discoveries in science
such as penicillin, X-rays, insulin, and the pulsar, have been regarded as
serendipitous findings. Since all images of serendipity have been created by
such great discoveries, serendipity is generally recognised as something that
scientists should explore further, when they encounter serendipitous find-
ings. In other words, it is thought that serendipity has a significant positive
effect on research quality.
Is serendipity good to pursue? One might suppose that great anecdotes
in the history of science cover only the successful results of serendipity. There
might have been many cases in which scientists explored unintended find-
ings and ended up with nothing. Furthermore, endogenous effects may exist
in the relation between research quality and reported serendipity. This is
because scientists explore a finding further only when they expect it will be
worth pursuing. Therefore, whether they obtained the finding as intended
or by chance may affect the likelihood of the eventual publication of the
finding. No study has empirically examined the extent to which pursuing
serendipitous findings brings better scientific outcomes. In addition to the
11
investigation of the effects of the division of labour on serendipity and pro-
ductivity, this study empirically investigates the extent to which pursuing
serendipity contributes to the quality of research by introducing instrumen-
tal variables.
3. Estimation Strategy
3.1. Hypotheses
This paper directs its attention to the managerial role in a research team
in order to explore the effect of management on serendipity and productivity
in science research, focusing on three points. The first point examines the
relation between serendipity and research performance. Based on anecdo-
tal evidence, we assume that serendipity improves the quality of research.
Hence, our first hypothesis to be explored is the following:
H-1: The existence of serendipity has a positive effect on the quality of
research.
The second point is related to the discussion about information asymme-
try and the coordination costs between management and research, which is
closely related to the discussion of serendipity. Scientists possess specialised
and domain-specific expertise. As the previous literature on scientific discov-
ery has explained, the nature of scientific discovery is highly unpredictable
(Polanyi, 1962), and tacit and uncodified knowledge plays an important
role in research, even though the outcomes of research are usually codi-
fied and published (Collins and Harrison, 1975; Polanyi, 1967). Learning is
highly situated in an on-site context (Brown and Duguid, 1991; Kogut and
Zander, 1992; Lave and Wenger, 1991). When scientists are committed to
actual research, they often encounter unexpected observations and findings.
12
Thus, if the managerial role in the research project team and the leading
role in the actual research are taken by different individuals, the research
project will have an information asymmetry between management and the
actual research. When a scientist observes unexpected but potentially cre-
ative serendipitous findings or encounters a serendipitous idea, they need to
encourage the person who plays a managerial role to change the initial re-
search plan in order to pursue the serendipitous. Presenting a serendipitous
encounter to a manager may be risky, particularly when the new idea or
observation is contrary to accepted ways of doing or thinking about things
(Pelz and Andrews, 1966). Thus, even if a surprising fact or relation is
observed, there may be a case in which it is not (optimally) investigated
by the discoverer (Barber and Fox, 1958; Van Andel, 1992). In contrast, if
a core scientist is also responsible for the management of the project, the
coordination and communication costs for shifting the research to pursue
the serendipitous findings will be decreased. Hence, this paper investigates
the following hypothesis:
H-2: Serendipity is positively related to the integration of core-scientists
into management.
However, if a core scientist plays a managerial role, the advantage of
the division of labour in science will not be fully realised. Efficiency is
increased by specialisation and concentration on a single subtask. Managing
a research team and conducting research require different sets of expertise.
Thus, it is possible that if a core scientist is separated from a managerial role,
they can focus on the research. This is important, particularly for a large
scale research project, which requires many bureaucratic procedures, much
paper work, and many managerial tasks. This paper, therefore, explores the
13
following hypothesis:
H-3: Research productivity is positively related to the separation of core-
scientists from management.
3.2. Data Description
We use data from the scientists’ survey conducted in Japan and the
U.S. in the time period 2009–2011 by Hitotsubashi University, the National
Institute of Science and Technology Policy (NISTEP) of the Ministry of
Education, Culture, Sports, Science, and Technology, and Georgia Institute
of Technology.1 The survey aimed to characterise the knowledge creation
process in science. It sampled a population of articles and letters which
had been recorded in the Web of Science database of Thomson Reuters
from 2001 to 2006 (database years). Review papers were excluded from the
population. Because this survey was particularly interested in characterising
high-performing research projects, it divided the population into two groups:
highly cited papers and a control group, based on the number of citations as
of the end of December 2006. The highly cited papers (H paper) consisted
of the top 1% of the highly cited papers in each journal field (22 fields in
total) and from each database year, while the normal papers (N paper) were
randomly selected from each journal field and from each database year from
the population of the survey, excluding highly cited papers. The sampling
rates were different between the two groups: roughly one-third of the samples
fell in the H group (the top 1% in the world). The survey sent a questionnaire
to the corresponding author of the sampled paper (“focal paper”) and asked
about the research projects which generated their paper. The respondent
1The full questionnaire can be found in Appendix 2 of Nagaoka et al. (2010).
14
was asked to define the project as an entire body of continuous research that
generated the focal paper and related papers. The response rate was 26%
in the U.S. (2,329 authors responded) and 27% in Japan (2,081 authors).
The questions asked by the survey include the following topics: the
knowledge sources which inspired the projects; uncertainty in the knowl-
edge creation process; research competition; composition of the research
team; sources of research funding; the research outputs, including papers,
patents, and licenses; and the profile of the scientists. The basic findings
from the survey characterised the research inputs, the knowledge creation
process, and the research outputs in H and N papers in both countries.2
One of the findings, for example, is that both the main result of the paper
and the research process were as initially expected or planned only for 11%
of the H papers in Japan and 14% in the U.S. (17% of the N papers in
both countries) (Nagaoka et al., 2011). This empirically demonstrates that
science confronts high uncertainty in general. By using the questions in the
survey about managerial roles and serendipity in particular, this paper aims
to explore the effects on serendipity of the division of labour in management.
In the Appendix, we provide the exact questions asked and the responses.
It must be noted that the survey was not constructed as a panel dataset,
even though the focal papers are sampled from a multi-year time period.
3.3. Definitions of the Variables
This section introduces the definitions of the variables used in our estima-
tions. Table 1 presents a complete list of the variables and their definitions,
and Table 2 shows summary statistics for all variables.
2See Nagaoka et al. (2011) for the detailed results.
15
3.3.1. Dependent Variables
The H-1 hypothesis uses the total number of citations observed by 2009,
which is denoted by number of citations, as a proxy for research quality.
Even though the division between H and N groups is determined by the ci-
tations in 2006, we chose the year 2009 so that the time window for citations
should be as long as possible in our database. The time window may still
be relatively short for some papers. However, since 70% of papers indicate
an increase in number of citations only by less than 10 from 2007 to 2008,
we believe that this variable is a reasonable proxy for research quality. Note
that the summary statistics in Table 2 show that the distribution of number
of citations is heavily skewed to the right. In order to deal with the skewed
distribution and the effects of outliers, we take the logarithm of number of
citations in the following estimations.
For the H-2 hypothesis, the dependent variable, serendipity, indicates
whether the respondent reported that the main finding of the focal paper
was obtained through serendipity. More specifically, this survey asked: “Has
the research output found the answers to questions not originally posed (in
other words, was the research output serendipitous)?” Approximately 55%
of the respondents answered yes. The highest was 61% in Computer Science,
and the lowest was 42%, in the Social Sciences.
For the H-3 hypothesis, the dependent variable, published papers, is the
number of refereed articles published by the entire research project. The
variation in published papers captures the variation in research productivity
in the regression where the research inputs such as project funds and the
number of researchers are controlled.
16
3.3.2. Independent Variables
Management structure is measured by two mutually exclusive variables.
Integration is a binary variable that takes the value one if the researcher
executed the central part of the research and contributed the most to the
research output, and at the same time, took a leading role in the research
management, designing the research project, organising the research team,
and/or acquiring research funds. Separation is a binary variable that takes
the value one if the researcher executed the central part of the research and
contributed the most to the research output, but took no managerial role.
There are observations that take zero for both integration and separation.
These observations, which exhibit a management structure in the middle
between clear integration and clear separation, account for 36% of the sample
for the H-2 model and 20% for the baseline H-3 model.
Other variables that describe research project characteristics are project
size, project duration, project funds, skill diversity, inter-lab community,
knowledge diversity, and competitor threat. Project size is the number of peo-
ple involved in the project, which includes collaborative researchers (includ-
ing coauthors), graduate students, undergraduates, and technicians. Since
not all projects had been terminated by the time of the survey, project du-
ration is calculated by subtracting the year when the project started from
the year of the most recent corresponding publication. project funds is the
total sum of research funding provided to the project. Since project funds is
heavily skewed to the right, we take its logarithm in estimations. Skill diver-
sity, inter-lab community, knowledge diversity, and competitor threat are all
binary variables. Skill diversity takes the value one if the researcher stated
that it was important for conceiving the research project to communicate
17
with researchers who have different research skills, for example, experimen-
tal researchers communicating with theorists. Inter-lab community takes
the value one if the researcher built a research community beyond their
own laboratory. Knowledge diversity takes the value one if the researcher
stated that it was important for conceiving the research project to communi-
cate with visiting researchers or postdoctoral researchers. Competitor threat
takes the value one if the researcher considered the possibility of competitors
who may have had priority in the research results.
Scientists are classified by the following variables: past publications, years
in paper, age, PhD, award, university, country, theory, and experiment. Past
publications measures the number of refereed papers published by the re-
searcher in the past three years.3 Years in paper measures the number of
years taken to publish the focal paper. Age is the respondent’s age at the
time of the survey. PhD, award, university, country, theory, and experiment
are all binary variables. PhD takes the value one if the researcher has a
Ph.D. or equivalent degree. Award takes the value one if the researcher re-
ceived a distinguished paper award or a conference award. University takes
the value one if the respondent works for a university. Country takes the
value one for respondents in the U.S. and zero for respondents in Japan.
Theory and experiment take the value one if the researcher specialises in
theory, respectively, experiments.
We control the respondent’s research field based on the survey’s classi-
fication. Table 3 shows the correspondence between its classification and
the 22 ESI journal fields. All scientific areas are divided into ten fields:
3The three year period corresponds to 2006–2008 in the Japanese survey and 2007–2009in the U.S. survey.
18
Chemistry, Materials Science, Physics & Space Science, Computer Science
& Mathematics, Engineering, Environment/Ecology & Geosciences, Clinical
Medicine & Psychiatry/Psychology, Agricultural Sciences & Plant & Animal
Sciences, Basic Life Sciences, and Social Sciences.4
3.4. Estimation Models
We employ the following estimation models to examine our three hy-
potheses. The H-1 model investigates the relation between serendipity and
research quality. In estimating the effect of serendipity on quality, we pay
particular attention to the possibility that serendipity has an endogenous
relation to the quality. As we reviewed above and will discuss in the fol-
lowing section, a researcher pursues publication of a scientific finding only
when the researcher believes that it is worth pushing forward, while the re-
searcher’s ex ante evaluation of the finding may differ depending on whether
the finding emerged from serendipity. If this is the case, the coefficient of an
ordinary least squares (OLS) regression may overestimate or underestimate
the effect of serendipity. To circumvent the endogeneity bias, we use a Two-
Stage Least Squares (2SLS) regression, in which we first regress serendipity
on the instrumental variables, and then the number of citations is regressed
on the predicted serendipity of the first stage. In this way, the coefficient
in the second stage captures the effect on the number of citations of the
exogenous variation in serendipity.
The H-2 model asks whether the integration of core-scientists with man-
agement affects serendipity. Since serendipity is a binary variable in our
dataset, we conduct a probit regression. An important issue with probit
4Social Sciences may be a fairly broad field compared to other fields. However, sinceabout 95% of the respondents are natural scientists, this makes no significant difference.
19
regressions is their fragility to any heteroskedasticity in the error terms.
Hence, we test whether our results are robust to misspecification of the
error term, and claim that our hypothesis still holds.
For the H-3 model, which examines the relation between management–
research separation and research productivity, we use a Negative Binomial
(NB) regression under the assumption that the variance of the dependent
variable takes a quadratic form. Since research productivity is measured
by the number of papers produced by the research project (controlled for
inputs), our dependent variable is necessarily discrete. Moreover, we only
have data on research projects that published at least one paper. For these
reasons, we use a zero-truncated NB regression model. An alternative would
be a Poisson regression, but that is not suitable in this case, since the equi-
dispersion hypothesis is strongly rejected at the 1% significance level. Since
the estimator of the NB regression coincides with that of quasi-maximum
likelihood, it is robust to misspecification of the distribution of the dependent
variable. That is, the NB regression yields a consistent estimator as long as
the specification of the conditional expectation of the regressand is correct.5
The unit of analysis for the H-3 model is the entire research project,
whereas the unit of analysis for the H-1 and H-2 models is the focal paper.
This is because the H-3 model is concerned with the effect of management
structure on the productivity of the entire project. In estimating the H-
1 and H-2 models, we restrict the observations to those cases where the
respondent is the first author of the focal paper, while for the H-3 model
we use observations for which the respondent was the researcher who took
5See Cameron and Trivedi (2005) for further discussion and Ding et al. (2010) for anapplication in a related context.
20
a central part in the research and contributed the most.6 Thus, the author
characteristics controlled in those two models are those of the first author,
while in the H-3 model they are those of the principal investigator. The
project funds variable is modified as the amount per paper for the H-1 and
H-2 models.
3.5. Estimation Issues
3.5.1. Sampling Bias
As a consequence of the survey method, one-third of the samples were
chosen from those researchers who wrote one of the top 1% highly cited
papers. Hence, our samples are not randomly drawn from the entire popu-
lation. We must consider this problem in order to obtain a consistent esti-
mator for the H-1 model, since the stratification depends on the regressand
(i.e., the number of citations).
A straightforward way to address this problem is to use the weighted least
squares method. In our regressions, we have two strata: highly cited papers
and others. Each observation i is weighted by the ratio between the popula-
tion frequency and the sample frequency of the stratum to which i belongs.
The weights are 0.032 for the highly cited papers and 1.433 for the others.
Under reasonable regularity conditions, this weighted least squares estimator
is consistent and asymptotically normally distributed (Wooldridge, 2010a,
2010b). Moreover, a consistent estimator of covariance matrix is obtained
by slightly modifying White’s (1980) heteroskedasticity-consistent covari-
ance matrix.
6In the H-1 and H-2 models, we excluded samples who are the corresponding but notfirst author of the focal paper to control the main author’s characteristics. About 30% ofthe sample was dropped after this selection.
21
3.5.2. Selection Bias
A number of anecdotes have casually reported that serendipity was im-
portant for successful research. Clearly, this cannot be taken as evidence
for a general tendency, since dramatic and successful anecdotes tend to be
highlighted and selected. As noted above, this paper adopts a definition of
serendipity, following Stephan (2010), as neutral as possible to the conse-
quential value of the serendipitous finding. Our survey contains many cases
in which scientists report that their finding was serendipitous in the sense
that it answered questions not yet posed, and yet the value of the finding,
measured by the number of citations, was not that large. By comparing
serendipitous findings to intentional findings with both successful and un-
successful findings, this paper aims to estimate the difference in the potential
values of findings that are found intentionally and accidentally from the per-
spectives of scientists. The neutral definition of serendipity enables us to
avoid the issue of selection bias that arises from overlooking the cases of
serendipitous but not highly valued findings.
3.5.3. Endogeneity Bias
The H-1 model concerns the population difference in the value of findings
between serendipitous findings and intentional ones. A regression of value on
serendipity may not yield a consistent estimator for this difference, however,
if the value difference affects the observation of serendipity in our data. In
other words, the estimate is biased if serendipity is an endogenous variable.
The serendipity observed in the data can be affected by the consequen-
tial value of the finding in the following sense. The value of the finding and
the serendipitous event are observable only conditional on the publication
of the finding. Moreover, a research team pursues publication only when
22
it considers it to be valuable. Hence, the observed citation rate might re-
flect not only the intrinsic value of a finding uncovered by a serendipitous
event, but also the research team’s evaluation of that discovery. There can
be two directions of endogenous effects from the quality of the finding to its
serendipity. The first possibility is that the researcher is less familiar with
the topic pertaining to the serendipitous finding than the topic originally
pursued; hence, the finding seems more novel to the researcher, who overes-
timates the value of the serendipitous finding. The second possibility is that
the researcher pursues their serendipitous finding only if it is highly valu-
able. This is because diverting the direction of research from the original
plan seems undesirable or risky. While we cannot determine which effect of
endogeneity is dominant ex ante, both effects imply that a simple regression
would result in underestimation or overestimation of the effect of serendipity
on research quality. Indeed, in the estimation of the H-1 model, a variable
addition test rejects the hypothesis that serendipity is an exogenous variable
at the 1% significance level (see Table 4).
We use instrumental variables to deal with the endogeneity bias. The
instrumental variables must correlate with the existence of serendipitous
findings, and they must not affect the ex ante evaluation of the findings. To
satisfy this criterion, we use two instrumental variables, skill diversity and
inter-lab community.
It is plausible that serendipity correlates with these two variables, since
complementarity in knowledge and skills are key to enhancing creativity.
For example, Heinze et al. (2009) pointed out that communication with spe-
cialists who have different knowledge or skills constitutes one of the most
important factors in inspiring a researcher’s creativity. Furthermore, we
assume that our instruments do not directly affect the ex-ante evaluation
23
of the serendipitous finding for the following three reasons. First, by con-
struction, these two instruments characterise the entire project rather than
the focal research paper. Second, even if the instruments affect the qual-
ity of the findings, the effect occurs mainly through enhancing creativity.
For example, skill diversity, which indicates whether communication with
researchers with different skills was important for conceiving the research
project, improves the value of findings by increasing the chance of valuable
serendipitous findings. In fact, we find that the variation of number of cita-
tions explained by skill diversity through serendipity in our 2SLS estimate
almost exhausts the correlation between number of citations and skill diver-
sity. Third, we conducted Sargan and Bassman over-identification tests to
see whether these two instruments are exogenous.7 The tests did not reject
the hypothesis that all the instruments are exogenous at the 20% significance
level. See Table 4 for the detailed results.
4. Results
4.1. Baseline Estimates
Table 4 summarises the estimation result for the H-1 model. We observe
that serendipity has a positive correlation with the logarithm of citation
counts at the 5% significance level in the 2SLS regression. This confirms our
hypothesis H-1: the findings through serendipitous events exhibit a higher
value in terms of the number of citations than findings which were made
intentionally. Note that this effect is insignificant in the OLS regression.
Namely, our sample does not exhibit a significantly positive correlation be-
tween serendipitous events and greater citation counts directly. In other
7Since our sample is drawn from two groups, highly cited and others, with differentsampling rates, we conducted over-identification tests for each subsample.
24
words, there are many observations which report serendipitous events and
low citations. When the serendipitous events are instrumented by exoge-
nous variables, however, a positive effect of serendipity is identified. We
interpret this estimation result as follows. Researchers tend to overestimate
the value of a serendipitous finding, and thus pursue the publication of the
finding with a less stringent criterion than intentional findings. As a result,
relatively more publication of serendipitous findings with low value are ob-
served, which masks the intrinsic positive quality of serendipitous findings.
However, the estimate may be open to other interpretations.
Among the control variables, we note that the existence of a threat from
competitors and the number of past publications of the first author have sig-
nificant positive effects. This is natural, as the competitor threat proxies the
potential value of the research topic, and the publication record represents
the researcher’s ability. We observe that the project size is negatively re-
lated to the quality on average. We robustly find that university researchers
exhibit less citation counts, which may reflect the fact that more of them
publish in a field with fewer researchers, than do researchers in industry.
The country effect shows that the average number of citations is higher in
the U.S. than in Japan.
The H-2 model examines the connection between management structure
and serendipity. In Table 5, the left column exhibits the result of the base-
line probit regression for H-2. We observe a positive correlation between
management–research integration and serendipity. The baseline model is a
heteroskedastic probit, because the homoskedastic specification is statisti-
cally rejected at the 1% significance level. The variance of the probit index
depends on project duration, which is likely to be an exogenous variable in
this model.
25
Table 5 also indicates that serendipity is reported more often when the
researchers were more engaged with an inter-laboratory community (a re-
search community beyond their own labs) or when the conception of the
project benefited from communication with researchers with different skills.
This result justifies our use of these variables as instrumental variables in the
estimation of the H-1 model. In addition, we note from the country dummy
that the Japanese respondents reported serendipity more often than the U.S.
respondents.
The regression results of the H-3 model are shown in Table 6. Our
hypothesis is that the separation of management from research increases
research productivity, measured in the number of publications generated by
the entire project. The first column shows that the separation exhibits a
positive effect on the number of publications. When the interaction with
project size is included as in the second column, separation has a negative
coefficient, while the interaction term exhibits significant positive effects.
Thus, the estimates imply that the separating management style yields a
higher productivity for a large project team. This result conforms to our
hypothesis. Quantitatively, the baseline estimate predicts that the adoption
of a separating management style increases the number of publications for a
project with more than three researchers. The average increase is 1.9 for a
project with 6 researchers and 4.8 for a project with 10 researchers, whereas
the projects with 6 and 10 researchers correspond to the median and the
75th percentile in project size, respectively. Considering that the median
number of papers published by a project is 7, the increase of publications
by adopting a separating management style seems non-negligible.
The coefficients for the control variables indicate that the number of pa-
pers generated by the project is large when the project fundsing is large,
26
when the project duration is long (though the effect is attenuated by the
squared term), when threatening competitors exist, and when the principal
investigator has a good publication record or a Ph.D. Projects led by univer-
sity researchers generate fewer papers than the other projects, indicating the
same pattern as for the H-1 model, where university researchers receive less
citations. U.S. projects also tend to generate fewer papers than Japanese, in
contrast with the result in H-1, which showed that U.S. researchers received
more citations than Japanese.
4.2. Robustness Checks
We conducted various robustness checks. First, we tested the validity of
our choice of instrumental variables. Three different choices were examined:
skill diversity (labeled “H-1-1”), knowledge diversity (labeled “H-1-2”), and
knowledge diversity and inter-lab community (labeled “H-1-3”). Table 7
shows the results. All three models suggest that serendipity and the number
of citations are positively correlated, so that our hypothesis is maintained.
Secondly, for the H-2 model, we tested different specifications of the error
term. The estimation for the homoskedastic probit is shown in the right
column of Table 5. We observe that the estimated coefficients are qualita-
tively similar to the baseline results. We also confirmed that an alternative
specification with variance depending on country as well as project duration
generates a similar pattern. Finally, we restricted the sample in the H-3
model to the projects which clearly stated their management structure, by
dropping the observations stating that management was not necessary or
choosing “other” in the survey. The estimates in Table 6 show that all the
major results are still qualitatively unchanged. Moreover, dropping the out-
liers (top 1%) of project size, published papers, and past publications in the
27
H-3 model did not alter our estimation result qualitatively.
5. Conclusion
This paper investigated the effects of division of labour between manage-
ment and research on serendipity and productivity in scientific research. The
major estimated results reveal the following three points. First, the estima-
tion shows that on average serendipity brings about better research quality.
Much of the anecdotal evidence has suggested serendipity plays a critical
role in science. The estimation from the survey empirically demonstrates
that serendipity is a key feature of science not only in great discoveries but
also in science in general. This finding suggests the importance of a man-
agement that gives scientists the flexibility to pursue a serendipitous finding
when they face the unintended and unexpected. Second, the integration of a
managerial role with a leading research role has a positive effect on serendip-
ity. Following the discussion of the division of labour and its coordination
costs, it can be interpreted that integration reduces the coordination costs
between management and research and provides scientists with flexibility in
research. When a scientist finds unintended and unexpected findings, the
findings are usually still too crude and uncertain to be articulated. Thus, if
the front-line scientist is delegated decision-making in the research project,
they can fully desterilise any uncodified tacit knowledge and use managerial
resources in the context of the actual situation. However, if a hierarchical
managerial role, which increases the information asymmetry and incommen-
surability between management and actual research, is implemented in a top
down fashion, it rebuffs research efforts to pursue a serendipitous finding.
Thus, the serendipitous findings based on ground level intuition are seldom
pursued. Thirdly, the results show that if a project is managed not by a
28
scientist who actually leads the scientific research, but by a person who spe-
cialises in research management, it increases the productivity of the research
when measured by the number of papers. These three findings imply that
there is a trade-off between pursuing serendipity and achieving research ef-
ficiency in science, via who plays the managerial role and who the leading
research role.
Returning to the example of penicillin, the findings of the paper suggest
that Fleming would have faced difficulties in changing his original research
plan to pursue his serendipitous finding if he had been working in a large
laboratory and his research had been led by a competent project manager.
In other words, Fleming would not have pursued the serendipitous finding,
but he would have delivered more papers concerning the original research
project if a managerial role had been played by a specialised director.
Serendipity plays an essential role in discoveries not only in science, but
also in technology, management, business practices, art, and daily life (Ja-
cobs, 2010; Svensson and Wood, 2005; Van Andel, 1992). The findings
of this paper have implications for corporate R&D and university research
administrators in particular. First, these results about the effects of the
division of labour between research and management on serendipity and
productivity in science are consistent with the contingency theory of firms
between the complexity of environment (e.g., demand, strategic positioning,
and technology) and their organisational structure (Burns and Stalker, 1961;
Lawrence and Lorsch, 1967; Scott, 1981), which have indicated that decen-
tralised and less formalised management allows a high degree of flexibility.
This is suitable when an organisation faces many exceptional problems and
problem solving is not easy (Perrow, 1967; Woodward, 1965). It suggests
that if corporate R&D is involved with embedded and uncodified knowledge
29
and if the firm needs alert responses to rapidly changing demand and supply
conditions, greater autonomy should be given to the R&D unit (Birkinshaw
et al., 2002). This may explain why it is difficult for corporate R&D over-
seen by a central business director to profit from serendipitous findings in
the laboratory. The findings of this paper suggest that decisions should be
made where the important information is gathered and knowledge is created
if unexpected findings are important.
Secondly, the findings have implications for university research adminis-
trators. Since the size of research projects is increasing and the competition
for priority of discovery in science is becoming fiercer, university research
administrators who are responsible for planning and managing research ac-
tivities and promoting research outcomes are serving very important roles in
science (Kaplan, 1959). According to the findings of this paper, the research
outcomes of the team depend highly on the extent to which the leading re-
search role and the managerial role are divided in the team. Operational
administrators are usually trained to complete the project’s goal. In fact,
they attempt to manage in a way that will eliminate uncertainty in their af-
fairs so that they can meet budgets and target deadlines (Udwadia, 1990). If
a managerial decision is made by such an operational research administrator,
it is highly likely that they would adhere closely to the initial research plan
even when scientists in the team encounter something serendipitous. Our
results do not suggest that the division of labour in research and manage-
ment is always inappropriate in science. The findings of this paper highlight
the way that the separation of research from managerial role allows the team
to achieve higher productivity as measured by the number of publications
for a given level of inputs. It would also work well if the research project
aims to make a thorough investigation of the possible combinations of ma-
30
terials with a door-to-door-check method, for example. Additionally, it is
quite important for a university research administrator to fully understand
the nature of discovery in science and the trade-off between serendipity and
productivity in science via who plays the managerial role and who the lead-
ing research role in research management (Kaplan, 1959; Kulakowski and
Chronister, 2006).
For the purpose of thinking about future research, it is important to note
three limitations to the present study. The first is related to the time scope
of the research project. A research project does not stand alone for a scien-
tist. Instead there is usually a sequence of continuous and related projects.
Even when the project is defined by the respondent themselves, as an entire
body of research that generated the focal paper and related papers, there
still remains, to some degree, a continuity of the research projects. This
problem is also related to the measurement of the quality of the research.
The scientists’ survey, which aimed to explore the nature of high performing
research projects compared with that of randomly selected research projects,
measured the quality of a research project by the number of citations that
the focal paper of the project received. Even though the measurement of
research quality by the number of citations has been widely used, there are
debates over the validity of this method. One of the issues related to the
present paper lies in the base year of the citation. Although it is necessary
to set a certain base point for measuring the number of citations, one might
always question whether some research might become highly valued after
the data cut-off point, for instance, due to the advancement of complemen-
tary technology. Since the scientists’ survey adopted a relatively short data
cut-off point because it aimed to distribute the questionnaires properly to
the corresponding authors, there is a possibility that this data cut-off point
31
underestimates the quality of research that takes a long time to be valued. In
addition, we used cross-sectional data, rather than a panel dataset. There-
fore, it should be noted that our empirical results are patterns of associations
between variables. Even though we carefully introduced instrumental vari-
ables in order to eliminate endogenous bias and tested our hypotheses based
on the patterns of correlations in the regression models, it is important to
collect panel data for deducing causal relations. These time scope issues are
an endemic problem in the study of science.
The second point is related to the country effects. The main aim of this
paper was to explore the effects of the division of labour between manage-
ment and research in science on serendipity and publication productivity.
Thus, although this paper introduces a country dummy variable to control
for country effects, it did not address the international comparison of the
management of science. However, as the estimations show, the country vari-
ables show significant effects on serendipity and publication productivity. It
is reasonable to assume that the ways in which scientific research is organ-
ised and managed are different across countries. Since the sample size of the
study did not allow robust estimation in the international comparison, it is
potentially important to collect a larger sample of the management of science
across countries so that one can make detailed international comparisons.
The third point is linked to the quality of the project managers. A
key result suggests that if scientific research is bureaucratically controlled
in a research organisation, serendipitous encounters will not be realised. In
other words, even when a managerial role and a leading research role are
played by different people, serendipity will be realised if a manager shares
tacit and domain-specific knowledge with leading scientists and understands
the nature of scientific discovery. This paper presupposes a certain degree of
32
incommensurability, which was proposed by “Kuhnian paradigm arguments”
(Kuhn, 1970) between a manager and leading scientists. However, the degree
of incommensurability depends on a manager’s expertise and capabilities.
Since the scientists’ survey does not allow the investigation of a manager’s
capabilities, this paper does not explore the quality of managers in a research
organisation. Organisations for university research administrators such as
SRA (Society of Research Administrators) and NCURA (National Council of
University Research Administrators) in the U.S., and ARAM (Association
of Research Managers and Administrators) in the U.K., were established
in the 1960s. Not only these organisations but also governments (e.g., the
Development of a Research Administration System program launched by the
Ministry of Education, Culture, Sports, Science, and Technology in Japan)
are beginning to understand that a managerial role should be played by a
specialist who can share tacit and domain-specific knowledge with leading
scientists: scientists could then focus on large-scale research projects, which
could have the managerial flexibility for realising serendipitous encounters.
The previous literature on how scientists with different sets of expertise
and paradigms communicate has indicated that scientists communicate in
groups called “trading zones,” where they can agree on the rules of exchange,
share the same learned languages, and share tacit knowledge (Collins et al.,
2007; Galison, 1997, 1999). However, since the extent to which managers
and scientists can reduce the degree of incommensurability depends on a
manager’s ability, it is important to explore the manager’s expertise and
capabilities for the research outcome in detail.
33
Table 1: Definitions of Variables
Variable Definitionnumber of citations Cumulative number of citations in 2009.serendipity Equals 1 if their research output found the answers to ques-
tions not originally posed.published papers The total number of refereed publications produced by the
research project.integration Equals 1 if the researcher executed the central part of the
research and contributed the most to the research output,and at the same time, took a leading role in the researchmanagement, designing the research project, organising theresearch team, and/or acquiring research funds.
separation Equals 1 if the researcher executed the central part of theresearch and contributed the most to the research output,but on the other hand, played no managerial role.
project size Sum of the number of collaborative researchers (includingcoauthors), graduate students, undergraduates, and techni-cians involved in the project.
project duration Years between the launch of the research project and thelatest publication by the project.
project funds The total sum of research funds prepared for the project.skill diversity Equals 1 if the researcher states that communication with
researchers who have different research skills was importantfor conceiving the research project.
inter-lab community Equals 1 if the researcher built a research community be-yond own laboratory.
knowledge diversity Equals 1 if the researcher states that communication withvisiting researchers or postdoctoral researchers was impor-tant for conceiving the research project.
competitor threat Equals 1 if the researcher considered the possibility of com-petitors who may have obtained priority in the researchresults.
past publications The number of refereed papers that the researcher publishedin the three previous years.
years in paper Years between the launch of the project and the publicationof the focal paper.
age Respondent’s age at the time of survey.PhD Equals 1 if the researcher had a Ph.D. or equivalent degree.award Equals 1 if the researcher received a distinguished paper
award or a conference award.university Equals 1 if the researcher works for universities.country Equals 1 for the respondents in the U.S. and 0 for respon-
dents in Japan.theory Equals 1 if the researcher specialised in theoretical work.experiment Equals if the researcher specialised in experiments.
34
Table 2: Summary Statistics
H-1 Model (Obs.= 1629) H-3 Model (Obs.= 1892)Variable Mean SD Min Med Max Mean SD Min Med Max
number of citations 51.0 102.0 0 12 1367serendipity 0.59 0.49 0 1 1published papers 19.3 42.5 1 7 590integration 0.58 0.49 0 1 1separation 0.07 0.25 0 0 1 0.06 0.23 0 0 1project size 9.66 25.21 1 6 603 9.95 21.7 1 6 601project duration 7.23 4.99 0 7 46skill diversity 0.29 0.46 0 0 1inter-lab community 0.45 0.5 0 0 1knowledge diversity 0.24 0.43 0 0 1project funds (M$) 0.14 1.12 0 0.02 31.3 0.15 1.17 0 0.02 31.3competitor threat 0.26 0.44 0 0 1 0.31 0.46 0 0 1past publications 20.6 41.5 0 10 750 26.4 50.8 0 12.5 750PhD 0.78 0.42 0 1 1 0.82 0.39 0 1 1award 0.34 0.47 0 0 1 0.41 0.49 0 0 1age 48.5 10.2 31 47 91 50.0 9.93 32 49 91years in paper 3.14 3.48 0 2 36 3.18 3.41 0 2 36university 0.73 0.45 0 1 1 0.77 0.42 0 1 1country 0.39 0.49 0 0 1 0.32 0.47 0 0 1theory 0.24 0.42 0 0 1 0.2 0.4 0 0 1experiment 0.62 0.49 0 0 1 0.68 0.47 0 1 1
Note: Sample statistics of H-2 model are similar to those of H-1 model, and hence omitted.
35
Table 3: List of Fields
22 ESI Journal Fields 10 Fields Obs. %Agricultural Sciences Agricultural Sciences & 349 7.9
Plant & Animal Science Plant & Animal ScienceBiology & Biochemistry Basic Life Science 910 20.6
ImmunologyMicrobiology
Neuroscience & BehaviourPharmacology & Genetics
Chemistry Chemistry 441 10.0Clinical Medicine Clinical Medicine & 710 16.1
Psychiatry/Psychology Psychiatry/PsychologyComputer Science Computer Science & 208 4.7
Mathematics MathematicsEconomics & Business Social Sciences 250 5.7Social Science, general
Engineering Engineering 368 8.3Environment/Ecology Environment/Ecology & 308 7.0
Geosciences GeosciencesMaterial Science Material Science 214 4.9Multidisciplinary (Journal field was assigned based on 13 0.3
the analysis of the backward citations)
Physics Physics & 639 14.5Space Science Space Science
Total 4410 100
36
Table 4: H-1 Model (Effect of Serendipity on Research Quality)
OLS 2SLSserendipity 0.043 (0.061) 1.066∗∗∗ (0.395)integration 0.112∗ (0.067) 0.037 (0.079)separation 0.013 (0.126) -0.108 (0.137)project size -0.002∗∗ (0.001) -0.002∗ (0.001)project funds (log) 0.024∗ (0.014) 0.015 (0.015)competitor threat 0.333∗∗∗ (0.077) 0.271∗∗∗ (0.087)past publications 0.002∗∗∗ (0.001) 0.002∗∗∗ (0.001)years in paper -0.019 (0.017) -0.038∗ (0.021)(years in paper)2 0.000 (0.001) 0.001 (0.001)age -0.060∗∗ (0.025) -0.047 (0.029)(age)2 0.000∗ (0.000) 0.000 (0.000)PhD 0.221∗∗∗ (0.082) 0.186∗∗ (0.091)award 0.004 (0.065) 0.014 (0.073)university -0.143∗∗ (0.072) -0.178∗∗ (0.081)country 0.309∗∗∗ (0.074) 0.543∗∗∗ (0.122)theory -0.059 (0.111) -0.035 (0.124)experiment 0.113 (0.094) 0.130 (0.106)field fixed effects Y YObservations 1629 1629F -statistic 7.759 6.587Variable addition test: F (1, 1602)=8.66∗∗∗
Sargan test: χ2(1)= 0.95, 1.29 (p-value = 0.33, 0.25)Basmann test: χ2(1)=0.93, 1.21 (p-value = 0.34, 0.27)
Standard errors in parentheses∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01Note: Over-identification tests are conducted separatelyfor N (normal papers) and H (highly cited papers).
37
Table 5: H-2 Model (Effect of Management Structure on Serendipity)
Heteroskedastic Probit Probitintegration 0.059∗∗ (0.025) 0.048∗ (0.026)project size -0.000 (0.000) -0.000 (0.001)project funds (log) 0.007 (0.005) 0.005 (0.006)skill diversity 0.081∗∗∗ (0.027) 0.097∗∗∗ (0.026)inter-lab community 0.113∗∗∗ (0.030) 0.131∗∗∗ (0.027)competitor threat 0.034 (0.029) 0.048∗ (0.028)past publications -0.000 (0.000) 0.000 (0.000)years in paper -0.002 (0.007) 0.011 (0.008)(years in paper)2 0.000 (0.000) -0.000 (0.000)age -0.003 (0.009) -0.016 (0.011)(age)2 0.000 (0.000) 0.000 (0.000)PhD 0.022 (0.034) 0.029 (0.033)award 0.011 (0.026) -0.008 (0.027)university -0.002 (0.028) 0.001 (0.029)country -0.245∗∗∗ (0.036) -0.274∗∗∗ (0.037)theory 0.027 (0.047) 0.029 (0.048)experiment 0.018 (0.039) 0.021 (0.042)field fixed effects Y Ylnσ2
project duration -0.061∗∗∗ (0.023)Observations 1474 1474Log Likelihood -915.7045 -919.700LR test for homoskedasticity: χ2(1) = 7.99∗∗∗
Standard errors in parentheses ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01
Note: All coefficients report average marginal effects exceptfor project duration.
38
Table 6: H-3 Model (Effect of Management Structure on Research Productivity)
All Sample Projects Sample with Management Needsseparation 0.287∗∗ (0.132) -0.137 (0.212) 0.294∗∗ (0.131) -0.134 (0.208)sep. × project size 0.051∗∗ (0.022) 0.052∗∗ (0.022)project size 0.012∗∗∗ (0.002) 0.012∗∗∗ (0.002) 0.012∗∗∗ (0.002) 0.012∗∗∗ (0.002)project duration 0.152∗∗∗ (0.015) 0.152∗∗∗ (0.015) 0.141∗∗∗ (0.016) 0.142∗∗∗ (0.016)(project duration)2 -0.003∗∗∗ (0.001) -0.003∗∗∗ (0.001) -0.002∗∗∗ (0.001) -0.002∗∗∗ (0.001)project funds (log) 0.115∗∗∗ (0.012) 0.114∗∗∗ (0.012) 0.115∗∗∗ (0.012) 0.115∗∗∗ (0.012)competitor threat 0.304∗∗∗ (0.068) 0.284∗∗∗ (0.068) 0.265∗∗∗ (0.069) 0.245∗∗∗ (0.069)past publications 0.005∗∗∗ (0.001) 0.005∗∗∗ (0.001) 0.005∗∗∗ (0.001) 0.005∗∗∗ (0.001)age -0.018 (0.030) -0.017 (0.030) -0.024 (0.032) -0.022 (0.032)(age)2 0.000 (0.000) 0.000 (0.000) 0.000 (0.000) 0.000 (0.000)PhD 0.373∗∗∗ (0.094) 0.388∗∗∗ (0.094) 0.321∗∗∗ (0.098) 0.337∗∗∗ (0.097)award 0.111∗ (0.067) 0.103 (0.067) 0.074 (0.068) 0.066 (0.068)university -0.262∗∗∗ (0.073) -0.279∗∗∗ (0.073) -0.309∗∗∗ (0.074) -0.328∗∗∗ (0.075)country -0.486∗∗∗ (0.072) -0.491∗∗∗ (0.072) -0.505∗∗∗ (0.074) -0.511∗∗∗ (0.074)theory 0.125 (0.133) 0.142 (0.132) 0.012 (0.139) 0.030 (0.138)experiment -0.105 (0.111) -0.078 (0.111) -0.056 (0.114) -0.029 (0.113)field fixed effects Y Y Y YObservations 1892 1892 1708 1708Log Likelihood -6630.021 -6627.055 -6069.145 -6065.998
Standard errors in parentheses ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01
Note: All coefficients are semi-elasticities.
39
Table 7: H-1 Model (2SLS with Alternative Instruments)
H-1-1 H-1-2 H-1-3serendipity 1.281∗∗ (0.574) 1.953∗∗ (0.930) 0.982∗∗ (0.478)integration 0.014 (0.086) -0.034 (0.108) 0.036 (0.079)separation -0.141 (0.150) -0.218 (0.188) -0.107 (0.138)project size -0.002∗∗ (0.001) -0.002 (0.002) -0.002∗∗ (0.001)project funds (log) 0.075∗∗ (0.036) 0.053 (0.047) 0.084∗∗∗ (0.032)competitor threat 0.237∗∗∗ (0.092) 0.202∗ (0.111) 0.253∗∗∗ (0.087)past publications 0.002∗∗ (0.001) 0.002∗ (0.001) 0.002∗∗∗ (0.001)years in paper -0.050∗∗ (0.022) -0.060∗∗ (0.029) -0.045∗∗ (0.021)(years in paper)2 0.001 (0.001) 0.001 (0.001) 0.001 (0.001)age -0.046 (0.031) -0.037 (0.038) -0.050∗ (0.029)(age)2 0.000 (0.000) 0.000 (0.000) 0.000 (0.000)PhD 0.177∗ (0.096) 0.154 (0.114) 0.187∗∗ (0.090)award 0.006 (0.077) 0.015 (0.090) 0.002 (0.072)university -0.148∗ (0.088) -0.180∗ (0.106) -0.134 (0.083)country 0.591∗∗∗ (0.154) 0.741∗∗∗ (0.237) 0.524∗∗∗ (0.134)theory -0.034 (0.129) -0.017 (0.151) -0.042 (0.122)experiment 0.111 (0.109) 0.126 (0.128) 0.104 (0.103)field fixed effects Y Y YObservations 1629 1629 1629F -statistic 6.490 4.801 7.348
Standard errors in parentheses ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01
40
Appendix: Selected Survey Questions and Responses
Question: Has the research output found the answers to questions not origi-nally posed?
Response RatesAnswers Highly Cited NormalYes 59.9% 54.0%No 40.1% 46.0%
Question: Please indicate which of the following best describes your role in the man-agement of the research project.
Response RatesAnswers Highly Cited Normal(1) A leading role in the research management, design-ing the research project, organising the research team,and/or acquiring research funds
70.9% 69.2%
(2) A member of the research management, but a roleless than that of the leader
14.1% 14.8%
(3) No managerial role 7.2% 5.8%(4) Management was not necessary 5.8% 8.0%(5) Other 2.1% 2.3%
Question: Please indicate which of the following best describes your role in the re-search implementation.
Response RatesAnswers Highly Cited Normal(1) I executed the central part of the research and con-tributed the most to the research output
64.4% 65.5%
(2) I took part in the central part of the research, but mycontribution was not as substantial as that of the centralresearcher
20.8% 21.9%
(3) I implemented the research under the guidance of theabove members
2.1% 3.0%
(4) I contributed to the research through the provisionof materials, data, equipment, or facilities
2.7% 2.8%
(5) Other 10.0% 6.8%
41
References
Adams, J. D., Grant, C. B., Clemmons, R. J., Stephan, P. E., 2005. Scientific
teams and institutional collaborations: Evidence from U.S. universities.
1981–1999. Research Policy 34 (3), 259–285.
Agrawal, A., Goldfarb, A., 2008. Restructuring research: Communication
costs and the democratization of university innovation. American Eco-
nomic Review 98 (4), 1578–1590.
Allen, T. J., 1977. Managing the Flow of Technology: Technology Trans-
fer and the Dissemination of Technological Information within the R&D
Organization. MIT Press, Cambridge, MA.
Andrews, F. M., 1979. Scientific Productivity: The Effectiveness of Research
Groups in Six Countries. Cambridge University Press.
Arora, A., Gambardella, A., 1994. The changing technology of technological
change: General and abstract knowledge and the division of innovative
labour. Research Policy 23 (5), 523–532.
Barber, B., Fox, R. C., 1958. The case of the floppy-eared rabbits: An
instance of serendipity gained and serendipity lost. American Journal of
Sociology 64 (2), 128–136.
Barnard, C. I., 1938. The Functions of the Executive. Harvard University
Press, Cambridge, MA.
Becker, G. S., Murphy, K. M., 1992. The divison of labor, coordination costs,
and knowledge. Quarterly Journal of Economics 107 (4), 1137–1160.
42
Birkinshaw, J., Nobel, R., Ridderstrale, J., 2002. Knowledge as a contin-
gency variable: Do the characteristics of knowledge predict organization
structure? Organization Science 13 (3), 274–289.
Brown, J. S., Duguid, P., 1991. Organizational learning and communities-
of-practice: Toward a unified view of working, learning, and innovation.
Organization Science 2 (1), 40–57.
Burns, T. B., Stalker, G. M., 1961. The Management of Innovation. Tavis-
tock Publications, London.
Cameron, C. A., Trivedi, P. K., 2005. Microeconometrics: Methods and
Applications. Cambridge University Press.
Chesbrough, H., 2003. Open Innovation: The New Imperative for Creat-
ing and Profiting from Technology. Harvard Business School Press, Cam-
bridge, MA.
Collins, H., Evans, R., Gorman, M., 2007. Trading zones and interactional
expertise. Studies in History and Philosophy of Science 38, 657–666.
Collins, H., Harrison, R. G., 1975. Building a tea laser: The caprices of
communication. Social Studies of Science 5 (4), 441–450.
Dasgupta, P., David, P., 1994. Toward a new economics of science. Research
Policy 23 (5), 487–521.
Ding, W. W., Levin, S. G., Stephan, P. E., Winkler, A. E., 2010. The
impact of information technology on academic scientists’ productivity and
collaboration patterns. Management Science 56 (9), 1439–1461.
43
Ellison, G., 2002. Evolving standards for academic publishing: A Q-R the-
ory. Journal of Political Economy 110 (5), 994–1034.
Fleming, L., Mingo, S., Chen, D., 2007. Collaborative brokerage, generative
creativity, and creative success. Administrative Science Quarterly 52 (3),
443–475.
Galison, P., 1997. Image and Logic: A Material Culture of Microphysics.
University of Chicago Press.
Galison, P., 1999. Trading zone: Coordinating action and belief. In: Biagioli,
M. (Ed.), The Science Studies Reader. Routledge, New York, pp. 137–160.
Garvey, W. D., 1979. Communication, the Essence of Science: Facilitating
Information Exchange among Librarians, Scientists, Engineers and Stu-
dents. Pergamon, Oxford.
Gouldner, A. W., 1954. Patterns of Industrial Bureaucracy. The Free Press,
Glencoe, IL.
Hagedoorn, J., 2002. Inter-firm R&D partnerships: An overview of major
trends and patterns since 1960. Research Policy 31 (4), 477–492.
Heinze, T., Shapira, P., Senker, J. M., 2009. Organizational and institutional
influences on creativity in scientific research. Research Policy 38 (4), 610–
623.
Hollingsworth, R. J., 2006. The dynamics of American science: An institu-
tional and organizational perspective on major discoveries. In: Beckert, J.,
Ebbinghaus, B., Hassel, A., Manow, P. (Eds.), Transformationen des Kap-
italismus: Festschrift fur Wolfgang Streeck zum sechzigsten Geburstag.
Campus Verlag, Frankfurt, pp. 361–380.
44
Jacobs, B. A., 2010. Serendipity in robbery target selection. British Journal
of Criminology 50 (3), 514–529.
Kaplan, N., 1959. The role of the research administrator. Administrative
Science Quarterly 4 (1), 20–42.
Kasperson, C. J., 1978. An analysis of the relationship between information
sources and creativity in scientists and engineers. Human Communication
Research 4 (2), 113–119.
Kogut, B., Zander, U., 1992. Knowledge of the firm, combinative capabili-
ties, and the replication of technology. Organization Science 3 (3), 383–
397.
Kuhn, T. S., 1970. The Structure of Scientific Revolutions, 2nd Edition.
University of Chicago Press.
Kulakowski, E. C., Chronister, L. U., 2006. Research Administration and
Management. Jones and Bartlett, Sudbury, England.
Lave, J., Wenger, E., 1991. Situated Learning: Legitimate Peripheral Par-
ticipation. Cambridge University Press.
Lawrence, P. R., Lorsch, J. W., 1967. Organization and Environment:
Managing Differentiation and Integration. Division of Research Gradu-
ate School of Business Administration Harvard University, Boston, MA.
March, J. G., 1991. Exploration and exploitation in organizational learning.
Organization Science 2 (1), 71–87.
Merton, R. K., 1957. Social Theory and Social Structure. The Free Press,
Glencoe, IL.
45
Merton, R. K., Barber, E. G., 2004. The Travels and Adventures of Serendip-
ity: A Study in Sociological Semantics and the Sociology of Science.
Princeton University Press.
Nagaoka, S., Igami, M., Eto, M., Ijichi, T., 2010. Knowledge creation process
in science: Basic findings from a large-scale survey of researchers in Japan.
IIR Working Paper No.10-08, Hitotsubashi University, Kunitachi, Japan.
Nagaoka, S., Igami, M., Walsh, J. P., Ijichi, T., 2011. Knowledge creation
process in science: Key comparative findings from the Hitotsubashi–
NISTEP–Georgia Tech scientists’ survey in Japan and the US. IIR Work-
ing Paper No.11-09, Hitotsubashi University, Kunitachi, Japan.
Nonaka, I., Takeuchi, H., 1995. The Knowledge-Creating Company: How
Japanese Companies Create the Dynamics of Innovation. Oxford Univer-
sity Press.
Pelled, L. H., Eisenhardt, K. M., Xin, K. R., 1999. Exploring the black box:
An analysis of work group diversity, conflict and performance. Adminis-
trative Science Quarterly 44 (1), 1–28.
Pelz, D. C., Andrews, F. M., 1966. Scientists in Organizations: Productive
Climates for Research and Development. Wiley, New York.
Perrow, C., 1967. A framework for the comparative analysis of organizations.
American Sociological Review 32 (2), 194–208.
Polanyi, M., 1962. The republic of science. Minerva 1 (1), 54–74.
Polanyi, M., 1967. The Tacit Dimension. Routldge & Kegan Paul, London.
46
Powell, W. W., Koput, K. W., Smith-Doerr, L., 1996. Interorganizational
collaboration and the locus of innovation: Networks of learning in biotech-
nology. Administrative Science Quarterly 41 (1), 116–145.
Presser, S., 1980. Collaboration and the quality of research. Social Studies
of Science 10 (1), 96–101.
Roberts, R. M., 1989. Serendipity: Accidental Discoveries in Science. Wiley,
New York.
Sauer, R. D., 1988. Estimates of the returns to quality and coauthorship in
economic academia. Journal of Political Economy 96 (4), 855–866.
Scott, W. R., 1981. Organizations: Rational, Natural and Open Systems.
Prentice-Hall, Englewood Cliffs, NJ.
Selznick, P., 1949. TVA and the Grass Roots. A Study in the Sociology of
Formal Organization. University of California Press.
Shapiro, G., 1986. A Skeleton in the Darkroom: Stories of Serendipity in
Science, Harper & Row, San Francisco.
Simon, H. A., 1976. Administrative Behavior: A Study of Decision-Making
Processes in Administrative Organization, 3rd Edition. The Free Press,
New York.
Singh, J., Fleming, L., 2010. Lone inventors as sources of breakthroughs:
Myth or reality? Management Science 58 (1), 41–56.
Stephan, P. E., 2010. The economics of science. In: Hall, B. H., Rosenberg,
N. (Eds.), Handbook of the Economics of Innovation. Vol. 1. Elsevier,
Amsterdam, pp. 217–273.
47
Stephan, P. E., Levin, S. G., 1992. Striking the Mother Lode in Science:
The Importance of Age, Place, and Time. Oxford University Press.
Svensson, G., Wood, G., 2005. The serendipity of leadership effectiveness in
management and business practices. Management Decision 43 (7), 1001–
1009.
Thompson, J., 2003. Organizations in Action: Social Science Bases of Ad-
ministrative Theory. Transaction, New Brunswick, N.J., London.
Udwadia, F. E., 1990. Creativity and innovation in organizations: Two
models and managerial implications. Technological Forecasting and So-
cial Change 38 (1), 65–80.
Van Andel, P., 1992. Serendipity “expect also the unexpected.” Creativity
and Innovation Management 1 (1), 20–32.
White, H., 1980. A heteroskedasticity-consistent covariance matrix estima-
tor and a direct test for heteroskedasticity. Econometrica 48 (4), 817–838.
Woodward, J., 1965. Industrial Organization: Theory and Practice. Oxford
University Press.
Wooldridge, J. M., 2010a. Econometric Analysis of Cross Section and Panel
Data, 2nd Edition. MIT Press, Cambridge.
Wooldridge, J. M., 2010b. Asymptotic properties of weighted M-estimators
for standard stratified samples. Econometric Theory 17 (2), 451–470.
Wuchty, S., Jones, B. F., Uzzi, B., 2007. The increasing dominance of teams
in production of knowledge. Science 316 (5827), 1036–1039.
48
Zuckerman, H., 1977. Scientific Elite: Nobel Laureates in the United States.
The Free Press, New York.
49