I
..'".'
i!i
''
:;;<
!
i
!
,
Bewey
HD28
.M414
no.
Scale, Scope
and
Spillovers:
The Determinants of Research Productivity
in Ethical Drug Discovery
Rebecca Henderson, M.I.T.
& NBER
&
Iain
WP #
This paper
is
Cockburn, U.B.C.&
3738-94
a revised version of working paper
NBER
September 1994
number 3629-93,
orginally circulated in
Auguest 1993. This research was funded by four pharmaceutical companies and by the Sloan
Foundation. Their support is gratefully acknowledged. We would also like to express our appreciation
to all of those firms that contributed data to the study, to two anonymous referees and to the editor of
this journal, to Richard Caves and Zvi Griliches, and to seminar participants at Carnegie Mellon,
Columbia, MTT, the NBER, Northwestern, Stanford, UBC and UCLA for their helpful comments and
suggestions. The conclusions and opinions expressed in this paper are our own, and any errors or
omissions remain entirely our responsibility.
Abstract
We
examine the determinants of research productivity
in the
pharmaceutical industry
using detailed internal firm data. Larger research efforts are more productive, not only
because they enjoy conventional economies of scale, but also because they realize
economies of scope by sustaining an adequately diverse portfolio of research projects and
by capturing
internal
and external spillovers of knowledge. This
important as the industry has
discovery.
We
moved from "random"
ability has
become more
to "rational" techniques
also find surprisingly large and persistent heterogeneities
among
both in research productivity and in the structure of their research portfolios.
IMSSACHUSETTSIiMSTI.UTE
OF TECHNOLOGY
MAY 231995
LIBRARIES
of drug
firms,
Introduction
Economies of scale and scope
in the
production of knowledge play an important role
in the
theory
of the firm, and the belief that there are significant spillovers of knowledge across firms has assumed a
central role in
is
new
theories of economic growth. But although a considerable
body of historical evidence
consistent with the existence of scale and scope economies in research, econometric efforts to
their
importance have generated surprisingly inconclusive and sometimes contradictory
for the existence of spillovers, although quite compelling,
importance of their role
we
In this paper
modern
in
look inside the firm for evidence as to the importance of scale, scope and
some of
the problems that
literature
on firm
size and
and
we
on individual research programs obtained from the
may be
R&D.
responsible for the inconclusive nature of the existing empirical
Rather than drawing inferences from the behavior of firm aggregates,
measurements of the phenomena of
-
of research projects
also unsatisfactorily incomplete given the
major research-oriented pharmaceutical companies. This narrow focus mitigates
internal records of ten
direct
Evidence
theory.
spillovers in drug discovery, using detailed data
we have
is
results.
document
interest
-
the size and structure of firms' portfolios
allowing us to distinguish between economies of scope and economies of scale,
are able to control quite precisely for cross-sectional variation in appropriability and technological
opportunity conditions.
The pharmaceutical
study these issues since not only
scrutinized industries, but
"rational"
it
is
it
industry presents a particularly interesting arena in which to
one of the most research-intensive, innovative, and closely
has also recently undergone a dramatic shift from so called "random" to
drug discovery. This
shift
appears to have changed the nature of the returns to scope and scale
with potentially important consequences for the competitive dynamics of the industry.
in research
Our
results suggest that there are significant returns to size in pharmaceutical research, but that
only a small portion of these returns are derived from conventional economies of scale. The primary
advantage of large firms appears to be their
ability to realize
diverse portfolio of research projects, and to
This
may
number of
explain, for example,
relatively small
why
make use of
economies of scope: to sustain an adequately
internal
and external spillovers of knowledge.
the majority of firms in our sample sustain a surprisingly large
and seemingly "unproductive" research programs.
But while the size and scope of the research portfolio have significant implications for research
productivity, in these data, at least, most of the observed variation in the productivity of individual
research programs
"knowledge
capital"
is
:
accounted for by fixed firm and therapeutic class effects, and by accumulated
ex post, the most productive research programs are those that are in the "right" area
and the "right" firm and that have been successful
in the past.
1
Moreover, despite the evidence we find
for
significant
relationships
between the diversity and "focus" of the research portfolio and
productivity, the firms in our sample maintained
periods of time.
"random"
We
marked differences
observe similar heterogeneity
to "rational" techniques
in these
dimensions over extended
our firms' responses to the industry's
in
its
shift
from
of drug discovery, which appears to have significantly increased the
costs of pharmaceutical research, and to have altered the relative returns to scale, scope, and capturing
spillovers.
Thus our
results raise a
The paper begins with
number of
a short literature review as background to the development of our
hypotheses. Subsequent sections discuss
study
intriguing questions for further research.
some of the
estimation issues and describe the data on which the
based. Section (4) outlines the empirical results, and the paper closes with a brief discussion of
is
their implications.
Hypothesis Development and Literature Review.
Scale, scope, spillovers
and research
productivity.
In principle size confers three major advantages in performing
R&D
(Fisher
Panzar and Willig, 1981; Cohen and Levin, 1989; Schumpeter 1934, 1950). Firstly,
fully functioning
markets for innovation, larger firms
over a larger sales base. Secondly, large firms
may
may be
also
&
Temin, 1973;
in the
absence of
able to spread the fixed costs of research
have advantages
in the financial
markets over
smaller firms: to the degree that they are able to mitigate problems of adverse selection and moral hazard
may be
may
also be
able to exploit complementarities within the firm to increase the productivity of research,
where
in raising capital, they
complementarities
may
better positioned to fund risky projects. Lastly, larger firms
exist both
between research projects within any given research
effort,
and between
other functions of the firm such as marketing and manufacturing.
These
effects are
of long standing theoretical interest (Panzar, 1989). Economies of scale have
long been thought to play a central role in the determination of industry structure, and economies of scope
in
R&D
play an important role in the theory of the firm.
(1989) have pointed out, since there are several well
(Arrow, 1962), economies of scope
in research
As Teece (1980) and Cohen and
known problems
Helpman, 1991), and
R&D
market for information
and development are likely to be a particularly powerful
rationale for the existence of multiproduct firms. Externalities arising
knowledge have also assumed
in the
Levinthal
a central role in
from the public goods aspect of
modern growth theory (Romer, 1986; Grossman and
spillovers also play a key role in
many models of market
structure (Spence,
1984; Dasgupta and Stiglitz, 1980).
However
despite the theoretical importance of this topic, systematic empirical tests of the
relationship
between firm size and research productivity have been largely inconclusive. Although much
qualitative evidence
is
lack of
consistent with the belief that size confers significant advantages,
appropriately detailed data on
R&D
made
activity within the firm has
about the presence of economies of scale in performing
R&D,
it
difficult to
draw conclusions
or to distinguish between economies of
economies of scope (Freeman, 1982; Chandler, 1990; Mowery, 1989).
scale and
Baldwin and Scott (1987), and Cohen and Levin (1989), survey the contradictory
of the
results
extensive econometric literature which has attempted to establish a positive relationship between firm size
and
in
R&D
intensity
—
a stream of research which, as Fisher and
any case be taken as
tests
Temin (1973) have pointed
of these ideas. Unfortunately the results of
later
out, cannot
work exploring
appropriate relationship between research output and firm size have been similarly unconvincing.
(1984) found that there was evidence of constant returns to scale for
et al.
and
$100m when
measure of output found
Acs and Audretsch (1988), using
a
Bound
programs between
patents were used as a measure of research output, but their results
to specification assumptions.
were quite
smaller firms were likely to be
more
innovative.
A
$2m
sensitive
measure of "major innovations* as a
that in highly concentrated industries with high barriers to entry large firms
be the source of the majority of innovations, while
likely to
R&D
more
the
in less concentrated, less
study by Pavitt et
al.
were
mature industries
(1987), using a similar measure
of output, suggested that both very small firms and very large firms were proportionately more innovative
than
more moderate
sized firms.
Similarly, although the presence of economies of scope in production have been verified in
industries, including airlines, trucking, banking and advertising (Caves, Christensen and
1984; Friedlaender, Winston and
Wang, 1983; Glass and McKillop, 1992;
of appropriately detailed data has meant that with some notable exceptions
many
Tretheway,
Silk and Berndt, 1993), a lack
we have very
little
empirical
evidence as to their importance in research (Helfat, 1994). Empirical estimates of the impact of spillovers
have also been
difficult to obtain (Griliches, 1992).
On
the one hand
it
is difficult
to distinguish
between
the influence that spillovers from other industries have on research productivity through their unmeasured
effect
upon input
costs and the influence that they have directly
the total stock of knowledge available to researchers.
effects is
On
on research productivity by increasing
the other, the accurate estimation of spillover
dependent on adequate measures of technological "distance" and of R&D
that are exceedingly difficult to measure. Jaffe (1986) provides the best
capital, both constructs
example of a careful attempt
to
account for these problems, and his work suggests that spillovers have important effects on research
productivity, but there have been few attempts to duplicate his results using disaggregated data.
Studies of the relationship between size and productivity in the pharmaceutical industry have had
similarly inconclusive results.
While Comanor (1965), Vernon and Gusen (1974) and Graves and
Langowitz (1993) found evidence for decreasing returns
that there
beyond a
were
significant
economies of scale
in
to scale in
R&D, Schwartzman
(1976) suggested
pharmaceutical research, and Jensen (1987) found that
(quite small) threshold neither firm size nor the size of the research effort affected the marginal
productivity of research and development. Furthermore, while the qualitative discussions in these papers
acknowledge the probable importance of scope economies, the quantitative analyses do not distinguish
them from economies of
shown
scale,
and although Dranove and
that in aggregate private research productivity
knowledge by the public
spillovers
sector, there has
Ward
(1991) and Gambardella (1992) have
significantly correlated with the generation of
is
been no systematic study
in this industry
of the importance of
between or within firms.
In general, as
such as variations
in
Cohen and Levin (1989)
demand
suggest, a failure to control for specific industry effects
conditions, in technological opportunity, and in appropriability conditions,
coupled with an inability to use project level as opposed to aggregate firm level data
may be
responsible
for the inconclusive nature of existing results.
Measuring Research Productivity
in the
Pharmaceutical Industry
Pharmaceutical research takes place in two stages: drug discovery and drug development. The
goal of the drug discovery process
that
is
to find a chemical
mimics some aspect of a disease
state in
compound
that has a desirable effect in a "screen"
man, while the goal of the drug development process
is
to
ensure that compounds identified through the discovery process are safe and effective in humans.
Although our data base includes information about both drug discovery and drug development, since the
two require quite
different sets of skills
we
focus here upon the determinants of research productivity in
drug discovery as measured by grants of important patents. In contrast to prior research which has been
forced to rely on publicly available firm level data, our focus on drug discovery allows us to separate the
effect
of economies of scale
in
The measurement of
discovery from the effect of any economies of scale in drug development.
"true" research productivity
(Griliches, 1979). In the ideal case,
Unfortunately
in this industry
we would
economic returns
like to
to
is
a project fraught with well
known problems
measure the private economic return
R&D
to research.
investment are particularly difficult to estimate.
Returns to individual research projects are highly skewed since they are the final result of a lengthy and
uncertain process, driven both by the uncertainties of clinical testing and regulatory review, and by the
complexities introduced by marketing, competitive activity and the role of the forces that determine the
demand
for
new
therapies.
These problems are compounded by
substantial difficulties in estimating costs
of capital and in making appropriate adjustments for risk (Baily, 1972; Di Masi, 1991; Grabowski and
Vernon, 1990).
We
therefore narrow our focus to the determinants of "technical success" in drug discovery, as
measured by patent grants, and our
that
make one
research group
more
Economic productivity
another.
of additional factors, but
we
results are
is
most plausibly interpreted as giving insight
new
likely to generate plausible
ultimately determined
by
candidates for development than
this ability in
believe insight into the research process
into the factors
is
conjunction with a wide range
of interest in
itself.
Hypothesis Formulation
Prior
economies of
work and our
scale,
qualitative research allow us to
frame a number of hypotheses about
economies of scope and the role of spillovers
in
pharmaceutical research.
Economies of Scale
Conventional wisdom
circumstances there
is little
in the industry
to gain
has
it
that
from increasing the
beyond a minimum threshold, under most
size of an individual research
descriptive statistics (see below) are certainly consistent with this belief.
However
program, and our
since pharmaceutical
research often requires investment in substantial fixed costs and since the complexity of the underlying
science offers considerable scope for specialization,
level.
For example,
equipment are fixed
we
expect significant economies of scale
to the degree that inputs such as libraries,
costs,
we
at
the firm
computer resources, or large pieces of
hypothesize that a larger research effort will gain economies of scale by
spreading them over a larger base, and while a smaller firm might need to employ molecular biologists
who
can work across a relatively wide range of
to afford
more narrowly focused
specialists.
fields,
we
hypothesize that a larger one might be able
Thus:
HI:
Individual programs are not subject to returns to scale.
H2:
There are returns to scale in drug discovery at the firm
But
level.
Economies of Scope
We also expect there to be significant economies of scope in drug discovery.
exist
when
cost,
and given the nature of pharmaceutical research we expect them to be
a tangible asset or a
human
Economies of scope
resource can be used in more than one application
at
no additional
critically important in
shaping
research productivity. Consider, for example, the benefits of investing in a centralized laboratory devoted
to peptide chemistry.
Economies of scale
lab can serve a larger
and larger discovery
also exist
the laboratory can
if
specialization of
relevant to a
its
exist if the costs
effort for a less than proportionate increase in cost.
become more
work
in
may
one
it
has more
exist if the
wide range of applications, and can be
usefulness in the others. Economies of scope
qualitative
efficient as
members. Economies of scope
and the results of successful research
of the laboratory are partially fixed, and
field
if
have implications for work
work
implications for
out of
work
that
was
On the other,
is
potentially
its
there are internal spillovers of knowledge,
leads us to believe that pharmaceutical research
relevant across multiple fields.
is
will
any one of them without diminishing
is
in other fields.
characterized by both.
hand pharmaceutical research requires the input of a wide range of highly
whose work
They
the
to do, possibly through the
work of the peptide chemists
utilized in
also arise
work
if
On
skilled specialists,
Our
the one
much of
discoveries in one field often have important
important central nervous system therapies, for example, grew
in another. Several
on the cardiovascular system. Hence:
originally focused
There are significant returns to scope in drug discovery.
H3:
We are agnostic as to whether economies of scale and scope should increase or decrease with firm
size.
While formal treatments of firm
unilaterally beneficial (Panzar,
scale and scope suggest that increasing firm size should be
1989), several researchers have suggested that beyond a certain point
escalating coordination costs and problems of agency lead to diseconomies of size (Holmstrom, 1989;
Zenger, 1994).
The Role of Spillovers
Our
qualitative research also leads us to believe that spillovers
in research productivity.
literature,
The
industry
is
between firms play a major role
characterized by high rates of publication in the open scientific
and many of the scientists with
whom we
spoke stressed the importance of keeping
with the science conducted both within the public sector and by their competitors.
1
had a quite accurate idea of the nature of the research currently being conducted by
and they often described the ways
own
research.
H4:
in
which
their rivals' discoveries
Nearly
all
in
touch
of them
L.eir competitors,
had been instrumental
in
shaping their
Thus we hypothesize:
Research productivity
is
positively associated with spillovers
of knowledge between
firms.
Notice, however that the impact of the efforts of competing firms on
own
research productivity
1
Measuring the impact of "upstream" spillovers from the public sector, though very important,
beyond the scope of this paper, but will be addressed in future work.
is
is
ambiguous.
On
the one hand, in a world in which firms are in a first-past-the-post "race" with each
other to reach a particular target,
all
other things equal rivals' success imposes an "exhaustion externality"
on competitors and own research productivity
(Reinganum, 1989).
things equal,
if
On
the other hand, firms
be negatively correlated with competitors' efforts
will
may
benefit
from competitors' research
since,
all
other
there are extensive spillovers of knowledge between firms, the productivity of a research
team may increase as others work
found an orally active
ACE
to take advantage of the
in the
same
For example, when Squibb announced
field.
to focus their
In both theory and practice, of course, these
own
two
(Spence, 1984). In a related paper, "Racing to Invest?:
Discovery" (Cockburn and Henderson, 1994),
the analysis presented here,
we
we
hypothesize that,
had
competing firms were able
inhibitor, a potent hypertensive therapy, several
announcement
that they
research efforts (Henderson, 1994).
complex ways
effects interact with each other in
The Dynamics of Competition
in Ethical
Drug
explore this issue in more detail. For the purposes of
all
other things equal, a positive correlation between
research success in any given area across competing firms
is
consistent with the presence of significant
spillovers or research complementarities between firms, while a negative correlation
is
consistent with
a research environment in which rivals' research efforts are substitutes rather than complements,
spillovers are limited and research productivity
is
primarily driven by an "exhaustion externality".
Changes Over Tune
We hypothesize that the last thirty years have seen significant changes
and spillovers
in
in the roles
of scale, scope
shaping research productivity as the technology of drug discovery has changed
dramatically (Cocks, 1975; Caglarcan,1978; Gross, 1983; Spilker, 1989). Thirty years ago drug discovery
was principally
a matter of the large scale screening of naturally occurring
"mechanism of action" of most drugs
responsible for their therapeutic effects
-
compounds. In general the
the specific biochemical and molecular pathways that
- were not well
understood.
A
were
firm searching for a treatment for
hypertension, for example, might inject thousands of compounds into hypertensive rats in the hope of
finding something that reduced blood pressure.
Firms employed few
beyond
scientific specialists
pharmacologists and medicinal chemists, and the most important spillovers between firms took the form
of the knowledge that compounds of a particular type showed promise in the treatment of a particular
disease.
For example, when Sir James Black announced the discovery of the
ICI's competitors immediately started to search aggressively for
that
might have equally dramatic therapeutic
The
first
many of
beta blocker
compounds of related chemical
structure
effects.
biological revolution of the last twenty years has changed this process dramatically,
moving
it,
in
popular terminology, from a process of "random search" to one of "rational drug design."
Significant advances in the understanding of both the
of the biochemical roots of
screens.
that
to use
If,
of a key
many
diseases has
made
it
mechanism of action of many
existing drugs and
possible to design significantly
more
one common analogy, the action of a drug on a "receptor"
fitting into a lock,
of what the "lock"
may
look like and with
it
much more
a
The
precisely tuned
request "find
pressure in rats" can be replaced with the request "find
II
body
is
similar to
contemporary drug researchers often have a much better developed sense
synthesize potentially effective compounds.
angiotensin
in the
sophisticated
me
me
method
something that
to screen
will
and
lower blood
something that inhibits the action of the
converting enzyme." At the same time, an improved understanding of molecular kinetics,
of the physical structure of molecular receptors and of the relationship between chemical structure and
mechanism of
action have sharpened the search for suitable "keys".
This profound change
"technology" of pharmaceutical research
in the
may have
significantly
altered the relative importance of scale, scope, and spillovers in determining research productivity.
impact on scale economies
in activities
is
hard to determine ex ante. While the importance of scale economies inherent
such as the screening of thousands of compounds
may have
mechanism-based research has also greatly increased both the number of
central to drug discovery
The
declined, the transition to
scientific specialties that are
and the pace of knowledge generation within the industry. Modern drug
discovery relies upon the input of scientists skilled in a very wide range of disciplines,
are advancing at an extraordinarily rapid rate.
Thus we hypothesize
many of which
that the opportunities to exploit
economies of scale arising from specialization may well have increased:
H5:
Our
has
become
Economies of scale
in
beliefs about the relative
drug discovery may have increased in the
last thirty years.
importance of returns to scope are much firmer. As drug discovery
increasingly directed by an understanding of the fundamental physiological mechanisms,
knowledge acquired
in
one area has become more
likely to
have implications for research conducted
elsewhere within the firm:
H6:
Economies of scope
in
drug discovery have significantly increased in the
last thirty
years.
Finally,
we
believe that there
shaping research productivity.
When
may have been
a concomitant change in the role of spillovers in
one's competitors are merely screening compounds there
is little
to
be learned from them unless they find a particularly promising molecule. But when they are actively
investing in the generation of
false starts
and failures
may
new
physiological and biochemical knowledge, even knowledge of their
help to shape one's
own
8
research program.
Thus we hypothesize:
H7:
The impact of inter-Jinn
spillovers
on research productivity has
significantly increased
over the last thirty years.
of the Econometric Model
2. Specification
We test these hypotheses by
by Griliches (1984),
in
estimating a "knowledge production function" of the type proposed
which new knowledge
is
generated using both accumulated "knowledge capital"
and resources expended in the current period. Rather than look for economies of scope and scale in
restrictions
on the parameters of cost functions (Caves, Christensen and Tretheway, 1984; Friedlaender,
Winston and Wang, 1983; Glass and McKillop, 1992; Silk and Berndt, 1993), we use direct measures
of scope, scale and spillovers as
shift variables
modifying the relationship between research expenditures
and the generation of knowledge, where knowledge
is
measured as grants of "important" patents. 2
Pharmaceutical companies patent prolifically, and patents are, of course, a rather noisy measure of
research success, in part because the significance of individual patents varies widely.
by counting only "important"
two of the three major
patents,
where an "important" patent
jurisdictions: Japan,
a useful measure of the generation of
stages in drug development.
Y — f (X,(i),
where Y
vector of parameters.
is
4
We
Europe and the United
new knowledge, which
is
is
We control
for this
defined as one that was granted in
States.
3
We think of these
patents as
the "raw material" input to subsequent
hypothesize that patent counts are generated by a production function
patent counts,
X is
a vector of inputs to the drug discovery process, and
We have no priors about the
"true" functional form, so the
/3 is
a
model estimated should
be thought of as a local approximation.
Some previous studies have
looked
at the
dynamics of the R&D/patents relationship by estimating
a lag structure on the input variables. Rather than
2
make assumptions about
distributed lags (and
have
to
The use of structural models of this type to explore'the determinants of performance has been a
some debate (Rumelt, 1991). We employ it as a technique whose limitations are quite well
matter of
understood and as a
3
We
first
step towards exploring a
would have preferred
to
complex and multifaceted phenomenon.
have been able to use
citation
output. Unfortunately since the U.S. patent classifications do not
research programs
proved
*
to
we would have had
to
buy
weighted patents as our measure of
map
citation data directly
directly into our definitions of
from Derwent Publications. This
be prohibitively expensive.
work will explore
number of compounds obtaining the various levels of
suggests that Investigational New Drug Applications (INDs)
Patents are clearly only one possible measure of research output, and later
the use of alternative measures such as the
regulatory approval. Preliminary analysis
are highly correlated with "important" patents, suggesting that our output measure does indeed
capture an important dimension of performance.
throw out much of our data
in
order to have 4 or 5 lags present in every program)
we
include "stocks"
of the input variables as explanatory variables. The annual flows are reasonably smooth, so
equivalent in
many
senses to imposing a geometric lag structure, where
rate rather than estimated one.
Note
that given a
smooth
we have assumed
series for the flow variable,
as a practical matter to identify the estimated coefficient
it
this is
a depreciation
be
will
difficult
on the stock variable separately from the
depreciation rate, making our assumption of a particular depreciation rate a second-order problem.
Since the dependent variable in this relationship only takes on non-negative integer values, some
type of discrete dependent variable model
Poisson process, which
unknown
captures
appropriate
if
we
number of Bernoulli
(but large)
some
is
is
We
dictated.
assume
we
by a
are prepared to model research results as the outcome of an
trials
with a small probability of success. This model certainly
aspects of drug discovery, such as screening.
based research, and
that patent counts are generated
It
may be
less appropriate for
mechanism-
explore the robustness of some of our results to the use of alternative estimation
techniques below.
We
model the single parameter of the Poisson distribution function,
explanatory variables, X, and parameters
/S
= X, = exp(X„0)
to guarantee non-negativity of X, and estimate the parameters
that the choice of
implications in this model. If
by maximum likelihood
whether to use explanatory variables
we
use levels, the estimated
elasticity
in levels or logs
in the standard
has important
of output with respect to each
explanatory variable will vary with the magnitude of the variable by assumption. Conversely,
explanatory variable enters in logs
some
in the standard fashion:
E[YU]
way. Note
X, as a function of
we impose
the constraint that the elasticity
is
constant over
of variation. In the case of spending on research, for the results presented here
we
its
if
an
range
maintain the null
hypothesis of constant elasticity of Kwith respect to research expenditures, allowing easy comparison with
previous work.
3
In exploratory
work we
tested for non-linearities in this relationship.
We
found that
estimated coefficients on quadratic terms in the log of research were very small and insignificant. For
research spending measured in levels, additional terms were significant, but expansion suggested that the
elasticity
3
of
Y with
Since
log of the
respect to total research spending
we have
R&D
previous work,
a fair
number of observations
was indeed constant over the range of our
in
which the
variables introduces the complication that
we
deal with this by setting the log of
appropriately coded
dummy
we
R&D
R&D
variables are zero, using the
cannot take the log of zero.
Following
equal to zero in such cases, including an
variable to account for this in the regression.
10
data.
We
we
and
have no strong priors about the appropriate way
to include the other explanatory variables
report results obtained by entering these variables in levels.
substantial
numbers of zeros, and
collinearity of the
useful
way
when
all
of these variables also have
avoids numerical problems in the estimation caused by near
VARIABLE =0 dummy
obtained in exploratory work
A
this
Many
were
variables with other regressors, but very similar results
explanatory variables were entered in logs.
to think about this specification
classes: the research expenditure variables, R,
is
to divide the explanatory variables into
two
and other variables, Z. The estimated function has a direct
proportionate relationship between research expenditures and patent counts, mediated by a set of
multiplicative "shift variables"
Z
E[YU]
= X, = exp(j81og(i?
ft
=
where
Z
RS
•
)n^)
exp( 7 ZJ
includes both our measures of scope, scale and spillovers and a constant term and firm and
therapeutic class dummies. Re-writing the equation in logs,
logfXJ = 01og(J?,)
thus
we
yZ„
can interpret the coefficient on log(R) directly as the elasticity of
while the
of the
elasticities
The assumption
data of this type, the
Z additional
that the
Y with
respect to research,
variables are yZ.
dependent variable
is
distributed Poisson
mean = variance property of the Poisson distribution
of such overdispersion, although the parameters
/3
will
is
is
quite strong: like
most other
violated here. In the presence
be consistently estimated, their standard errors will
typically be under-estimated, leading to spuriously high levels of significance. Overdispersion
interpreted as evidence that the statistical
unobserved variables
is
well
distributed
(
is
known
Hausman,
(see
gamma, then
not truly
is
miss-specified in the sense that there
often
may be
in the equation for X,
E[Y„]
As
model
is
it
=X, -«p(XJJ+0
Hall, Griliches (1984), Hall, Griliches,
Hausman
(1986))
if c
is
can be integrated out giving Y distributed as a negative binomial variate.
If
gamma, however,
the
maximum
likelihood estimates of the coefficients of the model will
be inconsistent. Gourieroux, Montfort, and Trognon (1984) suggest using a quasi-generalized pseudo-
maximum
likelihood estimator based on the
consistent estimates for
e
drawn from
a
first
two moments of the distribution of
Y,
which gives
GMT estimator
is
just weighted
wide variety of distributions. The
11
non-linear least squares estimates of the model
exp(Xj3h Cjl
Y„ =
with weights derived from the relation VAR[Y]
/S.
Below we present
alternate estimates of
=
E[YJ (1+r? E[Y]) using
initial
consistent estimates of
some of our regression models using maximum-likelihood
estimation of the Poisson and Negative Binomial models, non-linear least squares (with robust standard
errors),
3.
and the
GMT estimator.
Sources and Construction of the Data Set.
This paper uses a data
pharmaceutical
industry.
qualitative study
interviews, and
The
set obtained as part
of a larger study of research productivity in the
larger study has both
draws upon the medical and
was designed both
quantitative components.
The
and upon a program of detailed
field
qualitative and
scientific literature
shape the choice of variables and hypotheses explored in the
to
quantitative study and to explore the role of less easily measurable factors such as organizational structure
in driving research productivity.
The
quantitative study
draws upon data about spending and output
at the
research program level
obtained from the internal records of ten pharmaceutical firms. Although for reasons of confidentiality
we
cannot describe these firms,
we
can say that they cover the range of major R&D-performing
pharmaceutical manufacturers, that they include both American and European firms, and that
that they are not
markedly unrepresentative of the industry
performance. Between them they represent about
25%
in
we
believe
terms of size or technical or commercial
of the pharmaceutical research conducted world
wide. This section offers a brief description of the important variables used in the econometric analysis,
and presents descriptive
The
year.
statistics for
the data set.
data set used in this paper
is
an unbalanced panel indexed by firm, research program, and
With a complete, rectangular, panel we would have
1
1,400 observations,
research programs, and up to 30 years of data. In practice not
average time period for which
all
firms are active in
all
we have complete
research areas.
data
is
all
made up of ten
firms, 38
of these observations are available: the
on average just under 20 years per firm, and not
Our working sample
is
drawn from a data base which has 5543
potentially useful observations. After deleting missing values, grossly problematic data and peripheral
research areas
1000
we
are
left
with 4930 observations.
to less than 100, with a
The number of observations per firm
mean of 489.8. For each observation we have
outputs to the research process.
Our measures of
varies
from over
data on both inputs and
input include person years and research spending in
12
NDAs, new drug
discovery and development, and our measures of output include patents, INDs,
introductions, sales and market share.
Assembling the data
in a consistent
and meaningful format required considerable
effort. In nearly
every case the process of data collection was an iterative one, involving close collaboration between the
researchers and key personnel from the participating companies.
specifically together for the purposes of this study.
usually from primary documents, and the
hard to ensure
that, as far as
was
data collection effort took nearly two years.
at the
same
way, and we took steps
worldwide research spending, not just
US
data were collected
Each firm spent some months assembling
level
its
We
data,
worked
program and of expense grouping were
possible, definitions of research
standard across firms, data were collected
treated in a consistent
full
The majority of the
of aggregation, and overhead expenses were
to ensure that,
wherever possible, the data included
facilities.
Data was collected by research program rather than by broad therapeutic class or by individual
project since
research.
research
we
believe that analyzing the problem in this
A grouping by therapeutic class is too general:
into
widely
hyperlipoproteinemia.
is it
different
However
areas,
difficult to assign effort to particular
productivity
is
many
than in particular candidates.
"cardiovascular research," for example, includes
cardiotonics,
by individual project
is
difficult
antiarrhythmics
and
and misleading. Not only
drug candidates with any accuracy, but the notion that research
some conceptual
best measured at this level raises
typically continues over
best reflects the dynamics of discovery
hypertension,
including
analysis of data
way
difficulties.
A
research program
years, and at the discovery stage, the firm invests in the program, rather
The
identification of a
drug development candidate
success of the program, and retrospectively assigning resources to
its
generation
is
an indication of the
may
introduce serious
biases into the analysis.
Classification of our data into therapeutic areas
is
an important factor in our analysis, since
drives both the fundamental organization of our data, and the notion of spillovers.
into therapeutic areas using the
is
given in the Appendix).
program"
level,
IMS Worldwide class
We
definitions.
distinguish between
two
tiers
it
We classified our data
(A more detailed discussion oi
this issue
of aggregation: the detailed "research
and a more aggregated "therapeutic class" level which groups related programs into
therapeutic areas. For example, the therapeutic class "Central Nervous System Drugs" includes the
research programs "Depression," "Anxiety" and "Analgesics."
Full details of the construction of the variables used in the study are given in the Appendix.
primary variables are
DISCOVERY,
compounds", which excludes
defined as "expenditure relating primarily to the production of
clinical
development work and
13
is
measured
in
Our
new
constant dollars; and
PATENTS,
We
a count of "important" patents.
development price index issued by the National
are matched by year and research program.
deflate discovery using the biomedical research and
Institutes
We
of Health. These measures of inputs and outputs
count patents by their year of application, and define
USA,
"importance" by the fact that the patent was granted in two of the three major markets: the
and the European Community. Applying
up
to
60%
were able
of the number
US
filed in the
to obtain this data
is
this criterion screens out large
any given year are discarded. The
1961, and because patents grants
years in the United States and six in Japan
We
in
we
may
first
lag applications
for
example
year in which
by as much
we
as four
use only observations for the years 1961-1988.
measure the scale of the firm's research
There are some grounds for believing
numbers of patents:
Japan,
effort
that the relevant
by SIZE,
total
measure of size
is
discovery spending that year.
the scale of the entire firm, as
captured by, for example, total sales, or total employees, and this measure has been used in a number
of previous studies of the industry.
but obtained poor results. This
such as
total
firm sales
is
is
likely to
We
experimented with these types of measures in exploratory work,
perhaps not surprising given that any size effect captured by a variable
be confounded with factors such as demand (Jensen, 1987; Graves and
Langowitz, 1993).
In order to test for economies of scope
we
constructed a variety of measures of the diversity of
the firm's research effort across therapeutic classes.
programs
in
qualitative
Ph.D.
which firm spent more than $500,000,
work
led us to believe that this level
level scientists
- was
—
SCOPE
is
in constant
a count of the
1986 dollars,
number of research
in a single year.
Our
roughly equivalent to the employment of three to four
the smallest size program such that the firm could reasonably hope to
maintain sufficient "absorptive capacity" in an area so as to be able to take advantage of results generated
elsewhere within the firm or
in the industry
(Cohen and Levinthal, 1989).
We
experimented with a
variety of alternative thresholds, but in general the use of different cut off points yielded similar or
insignificant results.
We also constructed SMALL,
a count of research programs in which the firm spent
diseconomies of scope could arise through the
less than this threshold to test the hypothesis that
multiplication of coordination costs.
To
explore a different dimension of diversification,
we computed H ERF,
of the research portfolio, as a measure of the "focus" of the firm's research effort.
the Herfindahl index
HERF
is
defined
as:
a
Herf^s
where sp
is
the share of the /th research
program
2
in the rth
14
firm in the year
t
and there are n research
programs. Note that
larger firms run
HERF
SCOPE
and
HERF
are not simple transformations of each other. In general,
more programs, and thus tend
SCOPE
have higher values of
to
by construction. Nonetheless some of the smaller firms
in the
and lower values of
sample have very diverse portfolios
with activities spread evenly over a wide range of programs and some of the larger firms concentrate their
efforts in highly focused portfolios: the
Pearson correlation coefficient between
SCOPE and HERF is only
0.66.
We
hypothesize that both
SCOPE
and
HERF
have non-linear relationships
to productivity: that
both very highly focused and very disparate efforts will be on average less productive than those that are
"appropriately" focused and "appropriately" diverse. Highly focused firms will be less productive since
they will be less able to take advantage of internal economies of scope, and very unfocused programs will
be
less productive since they will incur higher coordination costs.
SIZE
Note that SCOPE, HERF,
SMALL and
are unique only to firm and year, not to firm, year, and research program.
We then construct variables intended to capture the effects of spillovers both within and between
firms.
For each observation, we
patents.
start
with the basic data on each program's "own" annual flows of
We then construct spillover variables at
one of the mechanisms through which the firm
two
We
levels.
realizes
economies of scope
of all of the other research programs within the relevant broad therapeutic
of a program
programs. This
is
same therapeutic
We
Depression
in
we
-
capture spillovers internal to the firm
-
class.
by measuring the output
For example,
in the
case
look for spillovers from the firm's other Central Nervous System
equivalent to using a binary measure of technological distance: programs within the
class are
assumed
to
be plausible sources of spillovers whereas other programs are not.
construct two measures of spillovers between firms.
in the
same narrow research
in the
wider therapeutic
class.
the flows over time with a
and on the other
area,
6
Finally,
20%
we
we
On
the one hand
we
count competitors' output
count competitors' output in
construct "stocks" for
all
all
the other programs
of these variables by accumulating
depreciation rate, and also a "news" variable
by subtracting 20% of the
stock at the beginning of the year from the annual flow.
6
We chose
as the sample
from which
to construct
our measure of competitors' patents the ten
firms that have given us data together with 19 other firms
world wide pharmaceutical firms
in
terms of
R&D
who have been
dollars and sales.
Note
consistently in the top
40
that these 19 firms are only
a fraction of the population of other firms generating spillovers, and the estimated coefficient on our
external spillover variables
may
therefore tend to overstate the magnitude of this effect.
However we
believe that these firms are a representative sample of the industry as a whole, and account for the
majority of the industry's spillover pool.
15
,
Descriptive Statistics
Table(l
all
)
and Figures
(1) to (4) present
summary
statistics
firms, research programs, and years, our firms spent
describing the data. Averaging across
on average $ 1.06m 1986 dollars on discovery
per program per year for an average of 1.8 "important" patents. Each "important" patent, on average,
thus cost about $600,000 in 1986 dollars.
substantially
The average firm
(more than half a million dollars)
in just
show
visible in the time series aggregates
the
mean
a substantial
is
highly diversified, investing
over ten programs a year. In addition these firms
on average invest more than ten thousand dollars a year
All of the key variables
our sample
in
in a further six
programs.
amount of variation. Perhaps the most dramatic
effect
the continuing increase in discovery spending despite the fact that
is
cost per important patent rose dramatically
from 197S onwards. Figure
(1)
shows the evolution
of mean discovery spending per program and mean important patents obtained per program over time.
Mean
discovery spending per program almost tripled in real terms over our sample period, from just over
$0.6m
effort
in
1965 to around $1 .9m
for sophisticated equipment and
obtained per program
In combination, these
from a high of 2.4
This trend reflects both an expansion
in its nature, as the adoption
and a change
demand
in 1988.
fell
1978 to
to the level of the firm:
of the research
of rational drug discovery techniques has increased the
more highly
skilled researchers.
At the same time mean patents
dramatically after 1978, from a high of just over 3 to less than 0.S in 1988.
two trends imply
in
in the scale
that
mean important patents per
less than 0.S after 1986.
The same
million real discovery dollars
trends are evident in data aggregated
between 1965 and 1990 mean discovery spending per firm almost
terms, from just over $ 18m in 1965 to around
$54m
1990
in
7
fell
tripled in real
while mean patents obtained per firm
fell
precipitously.
The
decline in patenting rates
and European firms over the
last
may
decade,
reflect the general
some of which simply
resource constraints imposed on the U.S. Patent Office
the decline in our data
and
we
suspect that
One
change
is
significantly
more
more fundamental
possibility
is
in"
trend that has characterized
reflects institutional factors
factors
may be
and an increase
more
at
economy
US
such as the
the early eighties (Griliches, 1992).
precipitous than that characteristic of the
that the transition to
in patenting strategies
downward
However
in general,
work.
rational
methods of drug discovery has led
to a
of each patent: real research output
may
in the significance
This number may seem surprisingly small, given the size of the research budgets announced
by some of the larger firms in the industry in 1993. Recall that it is calculated as an arithmetic mean,
that it is measured in 1986 dollars and that it includes only resources spent on discovery, not clinical
development.
16
have remained constant or even
Another possibility
risen.
exhaustion (Grabowski, Vernon and Thomas,
investment in research
may be
that the industry is
is
approaching technological
1978; Grabowski and Vernon,
1990), but rates of
continuing to accelerate despite the decline in patenting (Figure (2» in
response to what Sutton (1991) has identified as the need to maintain differentiation in the product
market.
Looking next
at the
emerge, highlighting the
program
pitfalls
level data,
some
surprising characteristics of the research portfolio
of using data aggregated to the level of the firm. Figure (3) presents the
size distribution of discovery expenditures per research program.
of the program-years
in
"alive"
development
in the area.'
$0.2m 1986
in that the firm
Of
was
reflect intermittent expenditures
consistently obtaining patents or
involved expenditures of more than
90%
the best
over time, or programs
was engaged
$10m 1986
tail
dollars per
of the distribution, just over
program per
in clinical
2%
of cases
year.
the output side, the distribution of annual counts of important patents per
(4)) is also highly
To
year, and about forty percent spent less than our critical threshold
of $0.5m 1986 dollars per program per year. At the other
On
at all.
the remainder, about one quarter of cases involved spending less than
program per
dollars per
our sample invests
our sample, no expenditures on discovery were recorded
of our knowledge these are "genuine" zeros and
which were
in
number of much smaller programs. For over
heavily in a few large programs but also invests in a large
44%
The mean firm
program (Figure
skewed. About half of our observations show zero output per program-year, and almost
of the counts are less than 5 per year. Patent counts also vary significantly across firms, across
therapeutic classes and across time. While the
varies across firms
from
in the genito-urinary
3.
1
mean number of patents per program
to 0.09 per year, and across therapeutic classes
is
1.7 per year, this
from 0.7 per year for work
system to 3.7 per year for cardiac and circulatory products. There are also very
substantial differences in the average level of expenditure across therapeutic classes:
mean discovery
expenditures per program range from roughly $900,000 per year in alimentary tract and metabolic
research to over
4.
$4.0m per year
The Empirical
In this section
in anticancer research.
Results.
we
present our estimation results.
We
begin by presenting results obtained by
aggregating data to the firm level to allow comparison with previous work (Table
(2)).
We
then turn to
Deleting these observations from the data set does not substantially change either the
magnitude or the significance of our estimated coefficients
*
17
analysis at the
program
level.
Table
(3) presents results obtained
from
ths full sample,
Table
(4) explores
the robustness of the Poisson specification, and Table (5) explores alternative measures of scope. In
Tables (2) through (5)
we
control for any change in regime after 1978
intercept and in the time trend.
1978 was the year
after the publication
paper describing their synthesis of an orally active
ACE
of
by allowing
for changes in the
Cushman and
Ondettis' seminal
inhibitor, often described as
one of the
examples of successful "rational" drug design (Cushman, Cheung, Sabo and Ondetti, 1977).
It
was
first
also
the year in which the patents obtained per discovery dollar for the firms in our sample began
precipitous
fall.
splitting the
9
In Table (6)
we
test for
its
changes in the roles of scope, scale and spillovers over time by
sample into pre-1978 and post-1978 subsamples.
Analysis offirm level data.
Table
(2) presents the results
of estimating the knowledge production function discussed in section
(2) using data aggregated to the firm level.
This aggregation generates a rather small sample, with only
181 observations on our 10 firms, and results should be therefore be treated with caution. In
with previous
work we
find
no evidence of returns
common
to scale: the implied long run elasticity of "important"
patent output with respect to research spending in every model
is
between 0.4 and 0.5. This
is
somewhat
lower than the results obtained in previous studies of the pharmaceutical industry (Comanor, 965 Vernon
1
;
and Gusen, 1974; Graves and Langowitz, 1993; Schwartzman, 1976; and Jensen, 1987) and for much larger
samples of manufacturing firms by, for example, Bound
(1984) and Hall, Griliches and
Hausman
(1986).
et al. (1984),
However
our work and previous studies which make the results not
there are
strictly
Hausman,
Hall, and Griliches
two important differences between
comparable. In the
first
place
we have
distinguished between discovery and development expenditures, and in the second place our patent counts
are restricted to "important" patents.
These
results also indicate that firm effects are a very important determinant of innovative
performance: comparing equations (1) and (2)
we
likelihood function very substantially, which corresponds to an increase in
extent these firm
dummies
dummies
see that including firm
:
increases the log-
R from 0.29
to 0.79.
To some
are picking up factors such as systematic differences across firms in their
propensity to patent, accounting practices, labor market conditions in different countries and so forth, but
their statistical
importance also reflects more interesting types of heterogeneity among firms. For
The choice of 1978
some
alternative dates with
as a cutoff point
little
is
inevitably
impact on the results.
18
somewhat
arbitrary.
We
experimented with
example, in these firm level data
we
cannot control for the "mix" of investment across different
therapeutic classes, which varies substantially across firms and has important implications for aggregate
productivity. (At the request of the firms supplying the data,
on these dummies here or
we do
in the other Tables, but they are jointly
not report the estimated coefficients
and separately highly significant
in all
of the models estimated. The ranking of firms according to these dummies conforms to our beliefs from
our qualitative work about their relative innovative performance.)
Equations (3) and (4) investigate the presence of returns to scope. In equation (3) our measure
of scope enters the regression linearly, with an implausibly negative coefficient. Our preferred model
allows for diminishing returns to scope by including a quadratic term, giving us a more plausible
The
inverted-U relationship between scope and research productivity.
results conditioning
R&D
stock
If
falls
we
final
column of Table
on past innovative success by including the stock of past
patents.
The
(3) presents
coefficient
on
somewhat, but the coefficients for the scope variables are largely unchanged.
did not have access to program level data
there are no returns to scale per se in research
is
we might
stop here, noting that our finding that
consistent with the previous quantitative studies of the
industry and with the qualitative evidence that suggests that there are limited returns to increasing the size
of specific discovery programs.
Analysis of Program Level Data
Tables (3) through (6) present our program level results. Table (3) re-estimates the models of
Table
(2) at the
program
spillover measures.
level using the complete disaggregated data set
We find
no evidence for returns
and introduces some of our
to scale at the level of the research
program. Indeed
the estimated coefficients imply quite sharply decreasing marginal returns to increasing investment in any
single program.
However
Ceteris paribus programs
there do appear to be economies of scale and scope at the level of the firm.
embedded
in larger
and more diversified firms appear to be significantly more
productive.
Equations (1) and (2) demonstrate the importance of controlling for firm and therapeutic class
effects. Including
firm and class dummies in the regression leads to a very substantial increase in the log-
likelihood function, which corresponds to an increase in
10
R2
is
R2 from
0.10 to 0.46. 10
calculated as the squared correlation between observed values of
Y and
To
save space the
fitted
values of
from the Poisson regression. Since the firm and therapeutic class dummies are not orthogonal (the
first three canonical correlation coefficients between the two sets of variables are between 0.2 and
0.1) their separate contributions to explaining the dependent variable cannot easily be determined.
19
Y
coefficients
on the 16 therapeutic
class
dummies
are not reported here. IkA with minor exceptions they
The
are statistically distinguishable from each other as well as from zero.
are quite large:
all
else equal,
generate about 2.2 times as
(anti-infectives),
programs
many
in the
most productive area
coefficients
on these variables
and related disorders)
(arthritis
important patents as programs conducted in the least productive field
while programs conducted in the most successful firm in our sample are more than twice
as productive as those in the least productive firm.
Note though
that the
across therapeutic classes, and as stressed above coefficients on firm
economic value of patents
dummies pick up
differs
factors such as
firms' propensity to patent as well as "real" differences in research productivity.
While the firm dummies alone account for a
variable, the discovery elasticities
the coefficients
substantial
amount of the variance
do not change much when they are included
on the discovery variables
fall
by about a
when
third
the class
in the
in the
model.
dummies
dependent
By
contrast
are also included,
highlighting the importance of controlling for cross-sectional differences in technological opportunity at
this
very detailed level. Interestingly even the largest firms
classes,
in the
even though they clearly have the resources to do so
if
sample do not invest
they wished: an important part of the firm
effect identified in the firm level data appears to lie in the firm's choice
Equation
(3) introduces
all
other things equal programs
more productive than programs embedded
in smaller rivals.
of patent output with respect to SIZE
is
firm level results,
we
quite large: at the
mean
the
0.26, so that relocating the "mean program" for
is
10%
higher
is
associated with an
SCOPE
introduce our measures designed to capture the effect of scope. Just as in the
has a significant and non linear impact on research productivity. At the
the elasticity of patent output with respect to
to a firm that
15%. "News"
Firm
is
be
in larger firms will
program productivity of around 2.5%.
In equation (4)
mean firm
embedded
This effect
example, from the "mean firm" to a firm whose research budget
increase in
of research programs.
our measure of scale into the analysis. SIZE enters with a positive and
significant coefficient, suggesting that
elasticity
in all therapeutic
is
in the output
effects alone give an
SCOPE
running one more
is
"critical
0.30, so that
moving the mean program from
mass" program raises
its
the
productivity by around
of related programs within the same firm enters with a positive and strongly
R2
of 0.21, while class effects alone give an
R2
of 0.32.
Notice that with both firm fixed effects and therapeutic class fixed effects
to estimating a non-linear panel data
important to control for these effects
at
these levels
we
are quite close
While we believe that is
of aggregation, the large numbers of dummy
model with fixed (program)
effects.
variables (10 firms and 16 therapeutic classes) limits our ability to obtain stable estimates
for example, use time
mean
dummies
instead of trend variables.
20
if
we
try to,
significant coefficient," with an elasticity at the
given the assumption implicit
in
mean of 0.03. Although
our functional form,
programs become more successful, and the impact of
programs are generating additional patents
example, the
elasticity
jumps
we
In equation (5)
of knowledge
the
capital.
This variable
is
of one standard deviation above the mean, for
highly significant, and markedly improves the log-likelihood and
on discovery
this stock is innovative
fall
sharply. This
may be proxy ing
the discovery variables with respect to patents. Conditioning
we
In equation (6)
and from related
at the
it
and that the
may
also reflect
unobserved correlated
problem with the exogeneity of
on past success may simply be purging the
the endogenous response to past success.
introduce our measures of external spillovers in the narrow therapeutic class
Both enter with positive and significant coefficients, with
fields.
effects are
elasticities
of about 0.1
mean. Again, though the coefficients may seem quite small, they can generate quite significant
and together they account for roughly eight percent of the
effects,
firms.
is
a
capital,
However
for a variety of
may be
SCOPE
consistent with the hypothesis
by the available stock of knowledge
such as the quality of program-specific assets, or there
discovery coefficients of the part which
is
output rather than innovative input.
specification problems: the patent stock
effects,
larger. If related
condition on past success by including the stock of past patents, our measure
that successful patent applications are driven
two
becomes much
to 0.12.
essentially unchanged, the coefficients
measure of
small, recall that
magnitude as related
of the equation (R2 increases from 0.47 to 0.67). While the SIZE and
fit
better
may seem
this elasticity increases in
internal spillovers
at the rate
this
total
variance across programs and
At the mean, for example, a program whose competitors' programs
fields are
roughly
10% more
Taken together
supported by the data.
productive will be approximately
2% more
in the
same and
productive
in related
itself.
the results of Table (3) suggest that our first four hypotheses are strongly
As our
qualitative analysis suggested,
programs conducted within large firms are
work we included both flow and stock versions of these variables in the model,
but concerns about the presence of measurement error (in many cases the coefficients were equal but
opposite-signed suggesting that they were picking up the same underlying factor) led us to use the
"news" formulation presented here, in which news in X is given by N, = X, - 6Khl where K is the
11
In exploratory
stock of
X and
6
is
the depreciation rate. This construction reduces the measurement error problem
and has an informative interpretation:
own
research productivity
is
higher
when
the output of spillover
sources "spurts" beyond the level required simply to maintain their previous stock.
We
"news in spending" in related classes as an alternative measure of
spillovers. Entered alone it was positive and marginally significant, but became insignificant when
"news in related patents" was also included in the regression. We interpret this as suggesting that
also experimented with
patents are a better measure of the spillovers of knowledge than dollars of research spending.
21
at
a significant advantage,
where these advantages seem
economies of scope as they do economies of
Table
(4) explores the sensitivity
to flow as
from the previous table
Binomial estimates, where the variance
gamma
adding an unobserved
the opportunity to exploit
scale.
of our results to the use of the Poisson assumption by presenting
alternative estimates of equation (6) using the various statistical
(6) is duplicated
much from
in section 2.
comparison of results.) Equation
to allow easy
is
models discussed
(Equation
Negative
(7) gives
modelled as an increasing function of the mean, equivalent to
distributed
random program
effect to the
model.
12
The
coefficients are
broadly comparable to those obtained in the Poisson model, with slightly inflated standard errors, though
the coefficients
on stock of patents and current discovery
no longer
patents in related programs
is
linear least squares model,
Y =
significant.
+
exp(X&)
rise quite substantially
Equation (8)
is
and news
in competitors'
the consistent but not efficient non-
with robust (White) standard errors. In this model, the
c,
discovery variables lose their significance altogether, although by and large with the exception of the
coefficient
of News
in
Competitors' patents in this program, the other results carry through from the
Poisson specification. Finally, equation (9) gives the results from the weighted
which
if
the exp(X(3) part of the model
is
NLLS
GMT
estimator
correctly specified, will give consistent and efficient parameter
estimates. Results are broadly comparable with the Poisson estimates, with slightly larger standard errors.
In order to allow comparison with prior studies, in the remainder of the paper
results obtained using the consistent Poisson estimator, but neither the
results
continue to report
magnitude nor the direction of our
change significantly when we use the more computationally expensive estimation techniques.
In Table (5)
we
further investigate the effects of the diversity of the discovery effort
productivity. Recall that the measure that
programs operating
1986
we
dollars. In
at
we
in
Table
greater than "critical mass" which
Equation (10)
we
introduce
SMALL,
threshold, to test our hypothesis that these
coordination costs.
used
SMALL
things equal, programs
much
(3),
SCOPE,
we defined
the
is
on research
defined as the number of
as an annual investment of
number of programs
that
$500,000
were smaller than
smaller programs affect productivity by imposing
enters with a negative and significant coefficient, suggesting that
embedded
in
this
firms that have
many of
these small programs
may
all
suffer
other
from
increased coordination costs, consistent with our interpretation of the negative effect of SCOPE-squared.
In equation (1 1)
we
include the Herfindahl of the research portfolio,
HERF,
to test for the impact
of the "focus" of the research portfolio on program productivity. The coefficient of HERF
is
positive and
highly significant, while HERF-squared enters negatively and significantly, suggesting that there
12
Cameron and
Trivedi's (1986) Negbin
II
model.
22
is
a
substantial return to focus but,
(12) includes both
econometrically
measured
at the
variance in
it
HERF
and
is difficult
once again,
that this return is only available
SCOPE. While
SCOPE
and
HERF
Throughout Tables
is
SIZE,
move smoothly
SCOPE and HERF.
HERF,
These variables are
together over time. Moreover most of the
between firms rather than within firms.
(2) to (5), the continued significance
the time trend suggest that there
to a certain point. Equation
there does appear to be additional information in
to separate out the effects of
firm level, and in general
up
was a
significant snift of
of Dum78 and of Dum78 interacted with
regime in 1978. In Table (6)
we
investigate the
changing role of scale, scope and spillovers over time by splitting the sample into two: 1961-1978 and
1979-1988. Equation (12)
is
included for comparison. Not surprisingly given the large changes in the
estimated coefficients, tests for equality of
hypothesis. This result
therapeutic class
is
all
coefficients across the
two samples
easily reject this
driven to a large extent by substantial differences in the coefficients of the
dummies across subsamples, probably
reflecting changes in technological opportunity
across fields, but there are also significant differences in the estimated effects of our measures of scope,
scale and spillovers
Taken
on research productivity.
at face value, these results
suggest that any returns to scale per se disappeared after 1978.
In the short term there even appear to be negative marginal returns to increasing investment in any
particular program, and the coefficient
on SIZE
falls to
almost zero. Lower short run returns to
investment in research are consistent with the "retooling" necessary to adopt the techniques of rational
drug design, but the collapse on the coefficient
in
SIZE
is
implausible given the qualitative evidence about
the increasing importance of fixed costs and specialized scientific knowledge.
reflects
confounding between overall trends
in the data rather
between SIZE and research productivity. All the firms
and did so
at
approximately the same
and therapeutic classes (Figure
marginal effect of
for, for
The
(1)).
SIZE which might be
example, an increase
But while
rate,
all
in the
our sample increased their size over
while patents per program
fell
this period
dramatically across
strong negative correlation between these trends
all
firms
swamps any
discernable in the cross-section, despite our efforts to control
the firms in our sample increased their investments in discovery
were allocated
rather than widened, their research efforts, but
in quite different
some
ways.
On
13
more or
less in
average firms deepened,
firms grew by adding programs while others held
Exploratory calculations suggest that simple ways of controlling for changes in the
"significance" or "quality" of patents
in the
than any real change in the relationship
average "quality" of patents by including a time trend.
concert, these larger research budgets
13
in
We believe that this result
average number of
would have
little
impact: for example, there
citations per pharmaceutical patent
23
is
no trend evident
between 1970 and 1985.
the scope of their portfolio constant and altered
its
focus.
Thus we
relationship between the structure of the research portfolio and
phenomenon
coefficient
rather than an artifact of the
on
are
more confident
program productivity
measurement properties of the
SMALL and the substantial
data,
and
that the estimated
is
a real economic
that the greatly
reduced
increase in the coefficient of the internal spillover variable are
consistent with our hypothesis that the transition from
random
to rational
drug design increased the
economies of scope realized by the more diverse firms.
At the same time, the influence of external spillovers on research productivity
While prior
to
also changed.
1978 the benefits of spillovers from competitors were realized primarily within narrow
therapeutic areas, after 1978 the reverse appears to be true. Programs benefited primarily from
work
conducted by their competitors in related therapeutic areas, rather than from research focused on the same
disease targets. This
is
consistent with our beliefs about the
ways
design has changed the nature of spillovers within the industry.
as biochemistry and physiology has
research productivity has
become
become
in
which the transition
to rational
As fundamental knowledge
in fields
such
increasingly important to the process of drug discovery,
by the synthesis and cross
increasingly driven
fertilization
of related
bodies of knowledge. Understanding and being able to take advantage of competitors' advances in
areas of neurochemistry, for example, has
drug
become
relatively
more valuable
many
to a firm attempting to
develop drugs for the treatment of depression than information about the specifics of competitive research
in that particular field.
5.
Discussion and Conclusions
We
in the
find evidence for significant
conduct of pharmaceutical research. All other things equal, programs embedded
efforts are significantly
this
economies of scope and scale and significant spillover
more productive than
advantage was driven both by the
economies of scope, while
after
rival
programs embedded
ability to share fixed costs
effects
in larger research
in smaller firms. Prior to
and by the
1978
ability to exploit internal
1978 the primary advantage of larger firms appears to have been their
ability to sustain a sufficiently diverse research effort.
As pharmaceutical research has moved from
a
regime of "random" discovery to one of "rational" drug design and the evolution of biomedical science
has placed an increasing premium on the ability to exchange information within the firm, our results
indicate that the primary advantage of size has
-
become
the ability to exploit internal economies of scope
particularly the ability to exploit internal spillovers of
per
se.
As Cohen and
knowledge
--
rather than any
economy of
-
scale
Levinthal (1989) have pointed out, the benefits of spillovers can only be realized
by incurring the costs of maintaining absorptive capacity, which take the form here of large numbers of
24
small and apparently unproductive programs.
We
believe that
it
is
these effects which account for the
presence of very large research oriented firms, despite sharply decreasing marginal returns to research
spending
of the individual research program.
at the level
While our
work
results provide considerable empirical support for the theoretical
that has
suggested that the presence of economies of scope are more important for determining the boundaries of
the firm and relative firm performance than economies of scale, they also have relevance for theories of
economic growth. In
this industry at least, external
quite differently.
Knowledge generated within
programs than the
results of research
of firm boundaries
may
During the
conducting research
is
much more
efficiently transferred across
conducted outside the firm, and industry structure and the location
also changed significantly over time, as the industry has
transition
became much wider:
the firm
therefore have strong implications for patterns of economic growth.
knowledge externalities has
science.
and internal spillovers affect research productivity
from "random"
to "rational"
same
area, after
role of
closer to basic
drug discovery the industry's knowledge pool
prior to 1978 the most important spillovers to any given
in the
moved
The
program came from
1978 spillovers from competitive research
in a
rivals
broader pool
of related technological areas had the most significant effect on productivity.
Interestingly, although our measures of the scope
and scale of firms' research efforts are
consistently significant, most of the variance in research productivity
program, therapeutic class and firm
of patents obtained by the program
accounting for up to
dummies account
20%
across firms.
Our
-
is
in patenting across
is
capital
-
the accumulated stock
the single most important variable in our regressions,
30-40%, and some of
intriguing findings
results
Our measure of knowledge
in the past
of the variation
for a further
One of our most
effects.
attributable to idiosyncratic
is
programs. Firm and therapeutic class
their coefficients are
remarkably large.
the persistence of marked differences in research strategy
imply that quite small changes
in the scale
or scope of a research portfolio have
a significant impact on productivity at the margin, yet differences in the structure of the portfolio across
firms
show
great persistence over time. While the
mean value of SCOPE
is
quite close to the "optimal"
value predicted by our estimates, there are firms in the data set that invest in significantly more or
significantly fewer
would appear
programs for extended periods of time. Similarly nearly every firm
to benefit
from having a more focused program, yet differences
in
HERF
in the
sample
across firms are
remarkably constant. In combination with our finding that there are very significant firm effects
in
research productivity above and beyond the structure of the research portfolio, these results speak to the
continuing debate about the nature and importance of firm heterogeneity (Lippman and Rumelt, 1982;
Rumelt, 1991; Schmalensee, 1985).
25
They
also highlight the dangers inherent in using our estimates to speculate as to the "optimal"
scale and scope of a
modern pharmaceutical research
effort. Interpreted literally, the results
of Table
(6)
suggest that prior to 1978, for example, the "optimal" research portfolio had 9.8 programs of at least a
half million dollars each and a
HERF
8.3 "critical mass" programs and a
of about 0.33, while after 1978 the "optimal" portfolio had only
HERF
of about 0.32. But the effects that
we have
estimated are
marginal effects. While doubling the size of an existing program or adding an additional program to the
portfolio, for example,
might increase the productivity of adjacent programs through
its
effects
and scope, the net effect of such a change on the productivity of the entire research effort
estimate. In reality
one cannot add an average program to the average
not homogeneous, and their heterogeneity
in the past
-
-
particularly the degree to
portfolio. Research
on
scale
is difficult
to
programs are
which they have been successful
has important implications for productivity, so that determining the size and shape of the
optimal research portfolio requires solving a complex, non linear constrained optimization problem whose
parameters are not fully known.
To
the degree that our results are general izahle across other knowledge-intensive industries, they
suggest that moving beyond aggregate data to a
heterogeneity
may be
a fruitful approach to the
more
detailed exploration of the evolution of firm
development of a richer understanding of the role of
research and internal firm structure in industrial competition.
"stickiness" of research portfolios reflected in our result
strategy across firms that flow
capabilities, an issue
which
we
may
We
reflect
suspect,
that the
underlying differences in research
from more fundamental heterogeneities
in organizational structure
are actively exploring in our current research.
26
for example,
and
References
Acs, Zoltan
J.,
and D.B. Audretsch (1988) "Innovation
in
Large and Small Firms:
An
Empirical
Analysis." American Economic Review, September 1988, Vol. 78, No. 4, pp678-690.
Development Costs and Returns: The U.S. Pharmaceutical
Economy, January /February 1972.
Baily, Martin N. (1972) "Research and
Industry." Journal of Political
Baldwin, W.L., and J.T. Scott (1987) Market Structure and Technological Change, Chur: Harwood
Academic Publishers, 1987.
Bound, John,
(ed.)
Caglarcan,
et al.
R&D,
E.,
et
(1984)
"Who
does Research and Development and
who
Patents?", in Zvi Griliches,
Patents and Productivity, Chicago: Chicago University Press, 1984.
al.
(Ed.)
(1978)
Government Regulation,
New
The Pharmaceutical Industry: Economics,
York: John Wiley and Sons,
Performance and
Inc., 1978.
Cameron, A. Colin, and Pravin K. Trivedi (1986) "Econometric models Based on Count Data:
Comparisons and Appl ications of Some Estimatoi s and Tests" Journal of Applied Econometrics,
Vol 1, pp29-53.
,
Chandler, A. (1990) Scale and Scope, Cambridge,
Caves, Douglas
W,
Lauritis Christensen and Michael
economies of scale: why trunk and
Vol. 15, No 4., 471-489.
Cockburn,
MA: The MIT
Press, 1990.
Tretheway (1984) "Economies of density versus
Rand Journal of Economics,
local service airline costs differ."
and Rebecca Henderson (1994) "Racing to Invest? The Dynamics of Competition
Drug Discovery", Forthcoming, Journal of Economics and Management Strategy.
Iain,
Ethical
in
Cocks, Douglas, (1975) "Product Innovation and the Dynamic Elements of Competition in the Ethical
Pharmaceutical Industry." in Drug Development and Marketing, ed. R. Helms, pp 255-254.
Washington D.C.: American Enterprise
Cohen,
Institute, 1975.
W.M.
and R. C. Levin (1989) "Empirical Studies of Innovation and Market Structure" in
and R. Willig (eds.), Handbook of Industrial Organization, Volume 2,
Amsterdam: North Holland, 1989.
Schmalensee, R.
Cohen,
W.M.
and Levinthal. D.A, (1989) "Innovation and Learning: The
99: 569-596.
Two
Faces of
R&D."
Economic Journal,
Comanor, William S. (1965) "Research and Technical Change
of Economics and Statistics, May 1965, pp 182-90.
Cushman, D. W., Cheung, H.S., Sabo, E.F., and
inhibitors of angiotensin-converting
Ondetti,
enzyme."
27
in the
Pharmaceutical Industry," Review
M.A. (1977) "Design of potent
Biochemistry, 1977, 16,5484-5491.
competitive
Dasgupta, P., and
Stiglitz
J.
(1980) "Industrial Structure and the Nature of Innovative Activity."
Economic Journal, Vol 90, pp266-93.
DiMa.si. J., et
al.
(1991) "The Cost of Innovation in the Pharmaceutical Industry." Journal of Health
Economics, Vol. 10, ppl07-142.
Dranove, David, and Michael Ward (1991) "The Vertical Chain of R&D
Mimeo, Northwestern University, September 1991.
Fisher,
in the
Pharmaceutical Industry."
R.M., and P.Temin (1973) "Returns to Scale in Research and Development: What Does the
Schumpeterian Hypothesis Imply?" Journal of Political Economy, Vol 81., No. 1, pp 56-70.
Freeman, Christopher (1982) The Economics of Industrial Innovation,
Friedlaender, Ann, Clifford Winston and
Kung Wang
MIT
Press, Cambridge,
MA.
(1983) "Costs, technology, and productivity in the
U.S. automobile industry." Bell Journal of Economics, 14: 1-20.
Gambardella, Alfonso (1992) "Competitive advantages from in-house scientific research: The
pharmaceutical industry in
Glass, J.C. and
US
the 1980s." Research Policy, 1-17.
D.G. McKillop., (1992) "An empirical
analysis of scale and scope economies and
technical change in an Irish multiproduct banking firm." Journal of Banking and Finance, 16:
423-437.
Gourieroux,
C,
A.
Trognon (1984) "Pseudo-Maximum Likelihood Methods:
Poisson Models", Econometrica, Vol. 52, No. 3, pp70 1-720
Monfort,
Applications to
and A.
Grabowski, Henry G., Vernon, John M., and Thomas, Lacy Glen (1978) "Estimating the Effects of
Regulation on Innovation: An International Comparative Analysis of the Pharmaceutical
Industry," Journal of Law
and Economics, 1978, ppl33-163.
Grabowski, Henry G., and Vernon, John
R&D." Management
Pharmaceutical
Graves, Samuel B. and
Nan Langowitz,
M
(1990)
"A New Look
at
the Returns and Risks to
Science, Vol. 36, No. 7., 804-821.
(1993) "Innovative Productivity and Returns to Scale in the
Pharmaceutical Industry," Strategic Management Journal, Vol. 14., p593-605.
Griliches, Zvi (1979) "Issues in Assessing the Contribution of Research and
Growth." Bell Journal of Economics, 10
Griliches, Zvi (1984)
R&D,
Griliches, Zvi (1992)
Development
to Productivity
(1) pp 92-116.
Patents and Productivity, (ed.) Chicago: Chicago University Press, 1984.
"The Search for
R&D
Spillovers." The Scandinavian Journal of Economics,
94(Supplement): 29-47.
Gross, F. (1983) Decision Making in
Drug Research,
28
(ed.)
New
York: Raven Press, 1983.
Grossman, Gene and Elhanan Helpman (1991) Innovation and Growth
Cambridge, MA., MIT Press.
Hall,
Bronwyn H., Zvi Griliches, and
International Economic Review,
Hausman,
Jerry,
Bronwyn H.
Jerry
Hausman (1986)
"Patents and
in
the
R&D:
Global Economy,
Is
There a Lag?"
Vol. 27, No. 2, June 1986, pp265-283.
Hall, and Zvi Griliches (1984) "Econometric
Models
for
Count Data With
an Application to the Patents-R&D Relationship", Econometrica, Vol. 52, No. 4, pp909-938.
"Know-how Complementarities and Knowledge Transfer Within Firms: The
Case of R&D" Mimeo, Wharton Graduate School of Business.
Helfat, Constance (1994)
Henderson, Rebecca (1994) "The Evolution of Integrative Capability: Innovation
Discovery" Forthcoming in Industrial and Corporate Change.
Henderson, Rebecca and
Iain
in
Cardiovascular Drug
Cockburn (1994) "Measuring Competence? Exploring Firm Effects
Pharmaceutical Research." Forthcoming in the Strategic
in
Management Journal.
Holmstrom, Bengt.: "Agency Costs and Innovation." Journal of Economic Behavior and Organization,
December 1989, 12 3 p305-27.
Jaffe,
Adam
B. (1986) "Technological Opportunity and Spillovers of
Patents, Profits and
Jensen, Elizabeth
J.
R&D:
Evidence from Firms'
Market Value," American Economic Review, December 1986, pp984-1001.
(1987) "Research Expenditures and the Discovery of
Industrial Economics, September 1987. Vol
XXXVI
New
Drugs." Journal of
pp 83-95.
Lippman, S.A. and R.P. Rumelt. (1982) "Uncertain Instability: an analysis of interfirm differences
efficiency under competition." The Bell Journal of Economics, V. 13., No. 2, pp 418-438.
Mowery, David (1989) Technology and
Cambridge University
the
Pursuit of Economic Growth,
in
Cambridge, England:
Press, 1989.
Panzar, John C. and Robert D. Willig (1981) "Economies of Scope" The American Economic Review,
71(20), 268-272.
Panzar, John C. (1989) "Technological Determinants of Firm and Industry Structure," in Schnidlensee,
R. and R. Willig (eds.),
Handbook of Industrial Organization, Volume
2,
Amsterdam: North
Holland, 1989.
"The Size Distribution of Innovative Firms in the UK.: 1945-1983," Journal
of Industrial Economics, Vol. XXV, March 1987, pp297-319.
Pavitt, Keith et al. (1987)
Reinganum, Jennifer F. (1989) "The Timing of Innovation: Research, Development and Diffusion",
Schmalensee, R. and R. Willig (eds.), Handbook of Industrial Organization, Volume
Amsterdam: North Holland, 1989.
29
in
1,
Romer, Paul, (1986) "Increasing Returns and Long-Run Growth," Journal of Political Economy, XCIV,
1002-37.
Rumelt, Richard P.
"How much does
industry matter?" Strategic
Management Journal,
Vol. 12, 167-185
(1991)
Schmalensee, Richard (1985) "Do Markets Differ Much?" American Economic Review, Vol. 75, 341351.
Schwartzman, David (1976) Innovation
in the
Pharmaceutical Industry, Baltimore: Johns Hopkins
University Press, 1976.
Schumpeter, Joseph A. (1934) The Theory of Economic Development, Cambridge,
MA:
Harvard
MA:
Harvard
University Press, 1934.
Schumpeter, Joseph A. (1950) Capitalism, Socialism and Democracy, Cambridge,
University Press, 1950.
and Ernst R. Berndt (1993) "Scale and Scope Effects on Advertising Agency Costs"
Marketing Science, 12 (1) pp 53-72.
Silk, Alvin J.
Spence, Michael (1984) "Cost Reduction, Competition and Industry Performance." Econometrica, Vol.
52., No. 1. January 1984, pp 101-121.
Spilker, Bert (1989) Multinational
Drug Companies:
Issues in
Drug Discovery and Development. New
York: Raven Press, 1989.
Sutton, John (1991)
Sunk Costs and Market Structure, Cambridge,
MA. MTT
Press.
Teece, David (1980) "Economies of Scope and the Scope of the Enterprise" Journal of Economic
Behavior and Organization." 223-247.
Vernon, J.M., and Gusen, P. (1974) "Technical Change and Firm Size: The Pharmaceutical Industry."
Review of Economics and
Statistics,
August 1974, Vol xx, pp294-302.
Zenger, Todd R.: (1994) "Explaining Organizational Diseconomies of Scale in R&D: Agency Problems
and the Allocation of Engineering Talent, Ideas, and Effort by Firm Size." Management Science,
Vol. 40, No. 6, 708-729.
30
Appendix: Data Sources and Construction
The
data set used in this study
is
based on detailed data on
R&D
inputs and outputs at the
research program level for ten ethical pharmaceutical manufacturers.
Inputs
Our
data on inputs to the drug research process are taken from the internal records of
participating companies, and consist primarily of annual expenditures
on exploratory research and
research by research program. Several issues arise in dealing with these data.
Research vs. Development
(a)
We define
all
resources devoted to research (or "discovery," in the terminology of the industry) as
pre-clinical expenditures within a therapeutic class,
compound has been
particular
in
identified as a
and development as
development candidate.
program wherever possible, but exploratory research
overhead. Clinical grants are included
in
We
all
be so assigned was included
the figures for development, and grants to external
that could not
researchers for exploratory research are included in the total for research. In
supplied us with data already broken
others,
we had to
down by
some
cases, the companies
research versus development by research program. In
classify budget line items for projects/programs into the appropriate category. This
done based on the description of each item in the
structure of the company's reporting procedure.
(b)
expenses incurred after a
attributed exploratory research to a
was
original sources, and the location of items within the
Overhead
In order to maintain as
much
consistency in the data collection process as possible,
we
tried to
include appropriate overhead charges directly related to research activities, such as computing,
R&D
administration and finance etc., but to exclude charges relating to allocation of central office overhead
etc.
The overhead
also includes
some expenditures on
discipline-based exploratory research such as
"molecular biology" which appeared not to be oriented towards specific therapies. Overhead was allocated
across therapeutic classes according to their fraction of total spending.
(c)
Licensing
We treat up-front,
lump sum payments
in respect
of in-licensing of compounds, or participation
in joint programs with other pharmaceutical companies, universities or research institutes, as expenditure
on research. Royalty fees and contingent payments are excluded. Though increasing over time,
expenditures on licensing are a vanishingly small fraction of research spending in this sample.
Outputs
In this paper
we
use "important" patent grants as our measure of research output.
We
count
patents by year of application, where we define "importance" by the fact that the patent was granted in
two of the three major markets: the USA, Japan, and the European Community. These data were
31
provided by Derwent Publications Inc. who used their proprietary- classification and search software to
produce counts of "important* patents to us broken down by therapeutic class for 29 US, European, and
Japanese pharmaceutical manufacturers for the 1961 to 1990. These firms were chosen to include the ten
firms that have given us data together with 19 other firms chosen on the basis of their absolute R&D
R&D
expenditures,
intensity,
and national "home base"
exhaustive, assessment of world-wide patenting activity.
40 world wide pharmaceutical firms
Note
many of these
that
in
terms of
R&D
that
citation
we were
rather than
firms have been consistently in the top
dollars and sales.
patents will be "defensive" patents in that firms
they do not intend to develop in the short term but that
and
to try to get a representative,
The 19
may have
may
patent
compounds
competitive value in the longer term,
not able to exclude process patents. Alternative measures of "importance" such as
weighting and more detailed international
filing data
proved prohibitively expensive to construct.
Classification
Classification of inputs and outputs by therapeutic class is important because this drives our
measure of spillovers. There are essentially two choices: to defme programs by physiological
mechanisms, e.g. "prostaglandin metabolism", or by "indications" or disease states, e.g. "arthritis". We
have chosen to classify on the basis of indication, largely because this corresponds well to the internal
divisions used by the companies in our sample (which is conceptually correct), but also because
classification by mechanism is much more difficult (a practical concern.) We classified both inputs and
outputs according to a scheme which closely follows the IMS Worldwide classes. This scheme contains
two tiers of aggregation: a detailed "research program" level, and a more aggregated "therapeutic class"
level which groups related programs. For example, the therapeutic class "cardiovascular" includes the
research programs "anti-hypertensives", "cardiotonics", "antithrombolytics", "diuretics" etc.
There are some problems with
very difficult to classify.
serotonin,
it
procedure. Firstly,
it
is
some
projects and
compounds
either side of the blood-brain barrier.
believed to play important roles in mediating motor functions.
has a variety of effects on smooth muscle, for example
companies report expenditures
in areas
are simply
be indicated for several quite distinct therapies: consider
which has quite different physiological actions on
neurotransmitter
hormone
this
A particular drug may
which are very
it
As
a
As a systemic
functions as a vasoconstrictor.
Some
difficult to assign to particular therapeutic classes:
company doing research using rDNA technology might charge expenditure to an accounting category
listed as "Gene Therapy/Molecular Biology" which is actually specific research performed on e.g. cystic
a
we were forced to include these expenditures in "overhead". Secondly, our two-tier
classification scheme may not catch all important relationships between different therapeutic areas. We
believe that we are undercounting, rather than overcounting spillovers in this respect. Thirdly, where
firms supplied us with "pre-digested" data, they may have used substantively different conventions in
classifying projects. One firm may subsume antiviral research under a wider class of anti-infectives, while
another may report antivirals separately. Not surprisingly there are major changes within companies in
internal divisional structures, reporting formats, and so forth, which may also introduce classification
errors. After working very carefully with these data, we recognize the potential for significant missassignment of outputs to inputs, but we believe that such errors that remain are not serious. Using patents
(as opposed to INDs or NDAs) as the output measure should reduce our vulnerability to mis problem,
since we observe relatively large numbers, and a few miss-classifications are unlikely to seriously affect
fibrosis, but
our
results.
32
Matching
Data series on inputs and outputs for each firm were matched at the research program level. This
procedure appears to successfully match outputs and inputs unambiguously for the great majority of
programs. In a very few cases, however,
NDAs
were
filed,
we ended up
with research programs where patents,
but where there were no recorded expenditures.
coding errors or reflected dilemmas previously encountered
corrections were made. In other cases,
ostensibly in, for example, hypertension,
it
was
may
Of these
INDs or
the majority were obviously
in the classification process,
clear that these reflected "spillovers"
and appropriate
-
research done
generate knowledge about the autonomic nervous system
which prompts patenting of compounds which may be useful in treating secretory disorders (e.g. ulcers.)
In such cases we set "own" inputs for the program equal to zero, and included these observations in the
data base.
Deflation
Since our data sources span
terms.
at the
We
many
years,
it
is
important to measure expenditures in constant dollar
used the biomedical research and development price index constructed by James Schuttinga
National Institutes of Health.
The index
is
calculated using weights that reflect the pattern of
NIH
expenditures on inputs for biomedical research, and thus in large measure reflects changes in the costs
of conducting research
at
academic
institutions.
However
with academic research laboratories for scientific talent
since the firms in our sample compete directly
we
believe that this index
is
likely to the
most
appropriate publicly available index, and our results proved to be very robust to the use of alternate
indices. In a later paper
on
R&D
we
intend to exploit the information that
some companies were
able to give us
inputs in units of labor hours to construct an index specifically for research costs in the
pharmaceutical industry.
Construction of stock variables
Annual flows of research and expenditures were capitalized following the procedure described
(The R&D Masterfile: Documentation, NBER Technical WP #72). In brief, we first assume
a depreciation rate for "knowledge capital", 6, here equal to 20%. (This is consistent with previous
studies, and as argued above is not going to be very important in terms of its impact on the regression
results since no matter what number we chose, if the flow series is reasonably smooth we would still find
it difficult to identify 6 separately from the estimated coefficient on the stock variable.) We then calculate
a starting stock for each class within firm based on the first observation on the annual flow: assuming
by Hall
et al.
have been growing since minus infinity at a rate g, we divide the first observed
by 8+g. Each year, the end-of-year stock is set equal to the beginning-of-year stock net of
depreciation, plus that year's flow. For the cases where the annual flow was missing "within" a series
that real expenditures
year's flow
of observations,
we
set
it
equal to zero. In almost
all
expenditure flows have been declining towards zero:
not missing data which should be interpolated.
patents, based
We
instances, these missing values occur after the
we
are reasonably that these are "real" zeros and
used the same procedure to accumulate "stocks" of
on the flow variables described above.
33
::
Original
misnumbered.
Page 34 does not
exist
:"
Table
(1):
Variable
Descriptive Statistics: Selected variables at the research program level.
Table
(2):
Determinants of patent output at the FIRM level.
Foisson Regression. Dependent variable = Total Firm Patents, 181 observations.
Table
(3):
Determinants of patent output at the research program level.
Poisson Regression. Dependent variable = Patents, 4930 observations.
Table
<4.i:
Determinants of patent output
it the
research program lereL
Lxpiortof Econometric Issues
Pomon
Refressjoo- Dependent Tariabie
=
patents, 4934 obserraboos.
Notes
to
Table
(4)
Poisson model as in column (6) of Table
Negative Binomial variance modeled
Non-linear Least Squares:
GMT:
Y =
Weighted non-linear
as:
exp (X
(3).
VAR(Y) =
ff)
+
least squares, with
wM
=
E[Y](1
+ a
ETTl)
i
weights derived from Poisson estimates of
exp(Xj)
*
30
2
ft
exp(Xj) 2
in
column
(6):
Table
(5):
Determinants of patent output at the research program
Alternative measures of SCOPE.
Poisson Regression. Dep. variable
=
level,
Patents, 4930 observations.
Table
(6):
Determinants of patent output at the research program
level,
Exploring changes over time, Poisson Regression. Dep. variable
Sample
N
= Important
Patents
Fig (1):
Research, Pits/prog, over tin*
Research
$
—m—
Petnl Pro gram
E
*.£
dc
8.
t
£
c
Ytar
Fig (2): Salts
ft
research over time
Re»ea/ch
SmWa
!
i
Y.ar
Figure (3)
Research/program, Frequency
1-0 2
0-0
1
3-0 «
2-0 3
5-0
6
7-0 t
4-0 9 0«-0 7
0«-0»
9-1
1-1
129-19 179-2
29
1
9-1
2 5-3
79 2-2 9
Research/program. 1686
Dittrib.
3 9-4
3-3 9
$m
Figure (4)
Patent* 'program Frequency Dlatrlbutlon
Patents/program
4 9-9
4-4 9
9-10
>10
Date Due
«*
\5
998
•'
111
ii
2
-
-
==i
=
-
mi inn
r.:.
W9224459