It Takes Two To Tango

illbreedinggrouseΠολεοδομικά Έργα

16 Νοε 2013 (πριν από 3 χρόνια και 11 μήνες)

56 εμφανίσεις


It Takes Two To Tango

dual structural and temporal dynamics in inter
-
organizational

network evolution within the Dutch life sciences industry


Jan Faber
a

, Tom Poot
a

and Marius Meeus
b

a

Department of Innovation and Environmental Sciences, Utrecht Universit
y

b

Department of Organization Studies, Tilburg University








Paper prepared for presentation at the FIRB
-
RISC conference on

“Research and entrepreneurship in the knowledge
-
based economy”


September 7
-
8, 2009

KITeS
-
CESPRI, Bocconi University

Milan, Ita
ly







Jan Faber, Department of Innovation and Environmental Sciences, Utrecht University, P.O. Box 80125,

3508 TC Utrecht, The Netherlands. Phone: (+31)302531625. Fax: (+31)302532746. Email: J
.Faber@geo.uu.nl


Abstract


In order to avoid the problems of interpretation and biased estimates of the effects of dyad
specific combinations of two organizations’ individual characteristics on the existence of
their collaborative relation present in previous studies
of inter
-
organizational
collaboration, this study introduces a variance decomposition model of complex discrete
network dynamics. The model identifies the relative magnitudes of structural and temporal
dyad specific inducements as well as organization spec
ific opportunities to collaborate.
Furthermore, the evolution of collaborative mechanisms can be assessed by analyzing
successive waves of networks. The model has been applied in the analysis of the rather
volatile collaboration network of Dutch dedicated
life sciences firms (DDLSFs). The
results partly correspond with those obtained by Powell et al (2005) but in a more
articulated way. In the first period analyzed (2002
-
2004), structural preferential
attachment of DDLSFs, irrespective of their centrality,
to central partner organizations (i.e.
Dutch universities) and temporal mutual preferential attachment between the most central
DDLSFs and partner organizations account for 97% of the variance of collaborative
relations. In the second period analyzed (2004
-
2005), structural preferential attachment
and temporal mutual preferential attachment weakened while structural dyad specific
collaboration became dominant together accounting for 99% of the variance of
collaborative relations. Due to network membership a
nd the exploration of collaborative
relations via temporal mutual preferential attachment both DDLSFs and partner
organizations seemingly learned from which organization to obtain what they need. These
successive collaboration mechanisms also altered the s
tructure of the collaboration
network of DDLSFs with a stronger articulation of more central DDLSFs and a
diminishing articulation of central partner organizations because organizations relying on
structural dyad specific collaboration partly retreated fro
m the network. Consequently, the
opportunities of the remaining DDLSFs to establish beneficial collaborations deteriorated
and, accordingly, also the effectiveness of Dutch governmental policy stimulating the
development of a Dutch life sciences industry
1




I
ntroduction


Inter
-
organizational collaboration
has been
argued

to occur

due to

various
incentives
.
Incentives for such collaboration encompass
, inter
alia,

control over supply relations (Pfeffer
& Salancik, 1978), access to complementary resources (Hagedo
orn, 1993) and learning new
competences (Powell et al, 1996; Baum et al, 2000). Apart from these
organizational
i
ncentives to collaborate (i.e. inducements;
Ahuja, 2000
-
a
), actual inter
-
organizational
collaboration also depends on the prior position
s

of or
ganizations in the collaboration
network and
their
previous experiences with collaboration (Gu
lati & Gargiu
lo, 1999) (i.e.
opportunities;
Ahuja, 2000
-
a
). Both sources of effects on the collaborative behavior

of
organizations have been studied for various t
ypes of organizations but mostly from
a single

actor

perspective.
Much empirical research adopted this single node perspective in studying
dyadic collaborations, mostly for pragmatic reasons

(e.g. Powell et al, 1996; Eisenhardt &

Schoonhoven, 1996; Ahuja,
2000
-
a,b
; Sakakibara, 20
02
).
However none of these rearchers
would deny that a

collaborative relationship
results from

both

partnering organizations’
incentives and opportunities to collaborate
as ‘it takes two to tango’. Consequently,

features
of
both
no
des affecting

the occurrence of such a relationship should be investigated
simultaneously
because

the relationship itself
is
the subject
of interest.
So far,
there is

a
small number of

recent
studies
that actually did

incorporate

this perspective in their
research
of collaborative relations
among organizations

by taking dyads of organizations as
the unit
of analysis (
e.g.
Stuart, 1998;
Gula
ti & Gargiulo, 1999; Powell et al, 2005).

Unfortunately,

the
se

studies
lumped together

inducement and opportunity indic
ators of both partner
s

involved in a

dyad
in a
compound

form (difference, sum, ratio or product)

thereby obscuring
whether or not
the occurrence of

their collaborative relationship stems from the inducements
and opportunities of
each

or both of them
; for e
xample, ‘technical overlap of firms
i

and
j

and ‘sum of sales of firms
i
and
j
’ i
n Stuart (1998) and ‘joint centr
ality’ and ‘size’ in Gulati
& Gargiulo (1999).

Estimating effects of compound measures has methodological implications such as
biased estimate
s if the effects of the constituents of a compound measure are not estimated
as well (e.g. Fuguitt & Lieberson, 1974). Only if the effect of a compound indicator is
specified together with the effects of its constituent individual components then the
compo
und indicator represents a distinct concept. A nice case in point is the
analysis
conducted

by Powell et al (2005) that uses asymmetric features of partners and focal firms
2




such as their incentives to forge col
laborative relations. In their analysis they c
ontrolled

for
the opportunities to collaborate for every partner organization and the inducements to
collaborate of each focal firm in their analysis. This control was confined to the perspective
of the focal firms of interest trying to establish collabora
tive relations with partner
organizations. Controls for the opportunities to collaborate of each focal firm and the
inducements to collaborate of every partner organization are absent. This suggests as if only
a subset of incentives (inducements and opport
unities) of the focal firms and the partner
organizations are relevant in establishing collaborative relations between the focal firms and
their partner organizations. And in addition to these specification and estimation issues there
is the issue that the

effects of opportunities

to collaborate

on the establishment of
collaborative relations between organizations have been argued to reflect multiple
explanations simultaneously; these effects are regarded as “macro phenomena emerging out
of micro decisions
of organizations” (Gulati & Gargiulo, 1999: 1475). The question that
arises
from this discussion
is how repeatedly observed network relations should be analyzed
in order to avoid these problems of contamination of effects, biased estimates and interpreted
causalities.

The first step is to decompose the variation of dyadic network relations into mutually
exclusive sources of variation. Although such a variance decomposition is a descriptive
statistical analysis, it provides a complete set of distinctive inte
rmediate variables to be
explained by specific types of attributes derived from focused additional theory formation.
This is exactly what statistical social network analysis models do (e.g. Kenny & La Voie,
1984; Snijders & Kenny, 1999). The sources of var
iation specified in these models reflect
the multi
-
level character of networks, i.e. network, actors and dyad specific sources of
variation. Consequently, it
becomes

obvious how the effects of additional explanatory
network, actor(s) and linkage specific a
ttributes should be specified and also the effects of,
for example, compound dyad specific attributes constituted of actor specific attributes (e.g.
Duijn et al, 2004). These social network models are, however, specified for the analysis of
directed dichot
omous network
s

measur
ed at one moment in time. But

these models suffer
from three inadequacies regarding the analysis of repeatedly observed dichotomous non
-
directed networks like the inter
-
organizational networks analyzed by Gulati & Gargiulo
(1999) and P
owell et al (2005). First, the social network analysis models developed are
unable to uniquely identify the constituent variance components of non
-
directed networks.
Secondly, they lack specifications of autocorrelation in panel data on repeatedly observed

3




networks. Thirdly, they do not accommodate accepted insights into the role of the topology
of networks in establishing network relations (e.g. Barabási & Albert, 1999).

In order to circumvent these inadequacies in the analysis of repeatedly discretely
mea
sured non
-
directed networks, Faber & Poot (2009) developed a multi
-
level variance
decomposition model of structural and temporal dynamics in discrete non
-
directed networks.
This model does not only specify the multi
-
level variance components mentioned befo
re but
also their structural and temporal effects on the occurrence of dyadic network relations.
Additionally, the observed variance components are measured as topological properties of
networks (node degrees and density) so that the role of network topolo
gy in establishing
network relations is explicitly specified. Application of this model to two waves of data on
collaborative relations between Dutch dedicated life sciences firms (DDLSFs) and their
partner organizations not only provides statistical evide
nce of the existence of structural and
temporal collaborative relations between them and the effects of both partners’ inducements
and opportunities thereon but also of the succession of different collaboration mechanisms
over time.

This article is structu
red as follows. First, the model, the analyses conducted and results
obtained will be presented. In Section 2, the variance decomposition model representing
basic graph
-
theoretical concepts of discrete non
-
directed networks will be described together
with
its estimation from dichotomous data on the (non)existence of collaborative relations
among organizations. In Section 3, the data collection used, characteristics of the datasets
used and the results obtained from their statistical analysis will be present
ed. After that, the
implications of the results will be elaborated, discussed and related to previous studies. The
theoretical and policy implications of these results and their context will be explored in
Section 4. Section 5 ends this article with a disc
ussion of the research carried out and the
conclusions to be drawn.


1.

Variance decomposition of discrete network dynamics


1.1

Specification of the variance decomposition model


In this section,

first
the ideas and model of scale
-
free networks of Barabási & Alb
ert
(1999)

will be explained
.
A
nalysis of the assumptions behind their model reveals that there
4




is an important option to develop the
ir

ideas further. Next
,

it
will
be described

how our
variance decomposition model can deal with this challenge.

The startin
g point
for the development
of the variance decomposition model
was

the
study of Barabási & Albert (1999). Their assessment of hierarchically structured empirical
networks
being

reflected in
very right
-
skewed node degree distributions, i.e. power
-
law
distr
ibutions with
P(k)


k


and
k

representing the node degree, induced them to simulate
and investigate the preferential attachment
of new entrants to already
connected

nodes over
time
.
The results demonstrated that if



3

then the node degree distributions

become

independent of the size of the network;
which means that

a scale
-
free hierarchically
structured network appears.
Mitzenmacher (2004) proved mathematically that

right
-
skewed

node degree distributions result from
an underlying

geometric model whose p
arameters

determine the nature of the power
-
law distribution

as
, for example, exponential, log
-
normal
or Poisson.
Furthermore
, Clauset et al (2007) demonstrated that it is very hard to determine
from empirical data on node degree distributions which theore
tical distribution applies due
to noise and measurement errors in the data. They found
that
in many cases the empirical
distribution fitted equally well to
all

three theoretical distributions mentioned

above
,
especially for


<3
.
This implies

that the unde
rlying geometric model can be represented as a
linear model after log
-
transformation of the data because of their
close

correspondence with
the

log
-
normal distribution.

The underlying geometric model producing the results presented by Barabási & Albert
(19
99) can be specified as


N
ij
t

=


*
(N
i

t

)


1

*
(N

j
t

)


2

*
(N
∙∙
t

)


3

*

ij
t

(1)


with


=
1
,

1=

3=0
,

2>0

and

ij
t

=
1
.
In this model
N
ij
t

represents the (non)existence of a
collaborative relationship between organizations
i

and
j

at the moment
t
,
N
i

t

denotes the
node (out)degree of organ
ization
i

at
t
,
N

j
t

denotes the node (in)degree of organization
j

at
t
,
N
∙∙
t

denotes the total number of linkages in the network at
t
,

ij
t

represents the prediction
5




error in
N
ij
t

and


is

a constant.

N
i

t

and
N

j
t

reflect the
centrality

of organizations
i

and
j
,
respectively,
in

the network at
t
.
N
∙∙
t

reflects the density of the network at
t
. And

ij
t

reflects the effects of unidentified
dyad (i.e.
linkage or pair of actors
)

specific characteristics
at
t
.
Each effect

specified in eq(1)
is supposed

to oper
ate within the time period separating
two successive moments o
f observation (e.g. Fisher, 1970
);
a non
-
trivial issue that

wil
l be
readdressed
later
.
The geometric model
in eq(1)
specifies that the occurrence of a
collaborative relationship not only depends

on preferential attachment of organization
i

to
j

but also on that of organization
j

to

i

while both effects
are

controlled for the positive effect
of network density on that occurrence. Consequently, the Barabási & Albert (1999) model of
preferential att
achment in network dynamics is a partial
model of network dynamics
, which
can be developed further by allowing the other topological properties of the network
specified in eq(1) to exert their influences
.

Log
-
transformation of eq(1) results in


ln(
N
ij
t
) =
ln(

) +

1 ln(
N
i

t
) +

2 ln(
N

j
t
) +

3 ln(
N
∙∙
t
) + ln(

ij
t
)

or

L
N
ij
t

=
a

+
b1

L
N
i

t

+ b2 LN

j
t

+ b3 LN
∙∙
t

+ e
ij
t

(2)


where
a=ln(

)
,
b1=

1
,
b2=

2
,
b3=

3
,
e
ij
t
=ln(

ij
t
)

and
e
ij
t

~
N(0,

2
eijt
)
.
This linear model of
network dynamics specifies the

variance of the
(logarithm of the)
bilateral collaborative
relationship between any pair of organizations
i

and
j

at
the

moment
s

t

to consist of four
parts induced by
the distinguished

sources of variation

mentioned before
; in fact a two
-
way
analysis of v
ariance perspective of a graph is adopted

here
.
1

E
q(2) does
, however,

not differentiate

between time
-
independent or structural and time
-
dependent or temporal effects

yet
. Such effects cannot be derived directly from the
observations of
L
N
ij
t
,
L
N
i

t
,
L
N

j
t

and
L
N
∙∙
t
.

In order to differentiate between structural and



1

For the sake of the argumentation that follows the measurement of
N
ij
t
,
N
i

t
,
N

j
t

and
N
∙∙
t

on discrete scales
is left aside here but will be returned to in Section 2.2.

6




temporal effects of the sources of variation specified on the occurrence of a bilateral
collaborative relationship between any pair of organizations
i

and
j

at
t
,
the variance of
L
N
ij
t

i
s specified to consist of two parts, namely a structural and a temporal part. This is done by
representing the value of
L
N
ij
t

as the sum of the value of an unobse
rved structural latent
variable

M
i
j


and the value of an unobserved temporal latent variable
M
ij
t

(e.g. Wheaton
et
al
, 1977)
. In the same way, the values of
L
N
i

t
,

LN

j
t

and

LN
∙∙
t

are represented as the sum of
an unobserved structura
l and
an unobserved
temporal latent variable.

In order to estimate the variances and covariances of the unobserved st
ructural and
temporal
latent
variables
M

at least two successive moments
t

and
t
-
1

of the observed

variables
L
N

must be specified, namely
L
N
ij
t

=

M
i
j


+

M
ij
t

,

L
N
ij
t
-
1

=

M
i
j


+

M
ij
t
-
1

,
L
N
i

t

=

M
i


+

M
i

t

,

L
N
i

t
-
1

=

M
i


+

M
i

t
-
1

,
L
N

jt

=
M

j


+

M

j
t

,

L
N

jt
-
1

=
M

j


+

M

j
t
-
1

,
L
N
∙∙
t
=

M
∙∙

+

M
∙∙
t

and
L
N
∙∙
t
-
1
=

M
∙∙

+

M
∙∙
t
-
1
. These
equations

constitu
te the measurement

model of the unobserved
structural and temporal latent variables
M

on the observed variables
L
N
. The variance of each structural latent variabl
e represents the common variance of two
successively observed
and log
-
transformed
variables
(i.e. their auto
-
covariance)
whereas the
variance of each temporal latent variable represents the unique variance of
a
log
-
transformed
observed
variable. Additional
ly, in accordance with eq(2), the following influence relations
between

the latent variables are specified,


M
i
j

=

a

+

b1
M
i



+
b2
M

j


+
b3
M
∙∙


+
e
ij


(3
-
a)

M
ij
t

=
b4
M
i

t

+
b5
M

j
t

+
b6
M
∙∙
t

+
e
ij
t


(3
-
b)

M
ij
t
-
1

=
b7
M
i

t
-
1

+
b8
M

j
t
-
1

+
b9
M
∙∙
t
-
1

+
e
ij
t
-
1


(3
-
c)


with
e
ij

~
N(0,
2
eij

)
,
e
ij
t

~
N(0,

2
eijt
)

and
e
ij
t
-
1

~
N(0,

2
1

eijt
)
.

7




Eqs
(3
-
a,b,c) constitute the variance decomposition model of
dual
structural and temporal
network dynamics. Both the measurement and the influence relations model can be specified
straightforward in the computer program
LISREL
®

(Jöresko
g & Sörbom, 1993) in order to
estimate the variances and covariances of the specified latent variables
M

and the
values of
the
unknown constant parameters
a
,
b
1
-
b
9
,

2
eij
,

2
eijt

and

2
1

eijt
.

The latter th
ree
parameters

are called the unique between
-
linkages and within
-
linkage variances of
L
N
ij
t

and
L
N
ij
t
-
1
.
The
unique between
-
linkages variance repr
esents the effects of unidentified

structural
dyad

specific characteristics. The unique within
-
linkage varianc
es are conceived as random errors
at
t

and
t
-
1
.

But in LISREL the residuals
e
ij
,
e
ij
t

and
e
ij
t
-
1

can also be specified as latent
variables acting as dependent variables in additional analyses.
In order to comply with the
specification of
an

econometric pan
el data model (e.g. Johnston & DiNardo, 1997)
constraints are imposed, namely
b4=b7
,
b5=b8

and
b6=b9
, so that the effects of the
temporal latent variables apply to every moment
t

and
t
-
1

in
the observed
time period
T
.

The combined model specified in LISREL

can be estimated by means of, with increasing
accuracy, the Unweighted Least Squares (ULS), the Generalized Least Squares (GLS) or the
Maximum Likelihood (ML) method. All three methods estimate the unknowns in the
LISREL model from
an

input matrix
S

conta
ining the variances and covariances of the
observed variables
L
N

and their additionally provided

mean values
. Saris & Stronkhorst
(1984) recommend
to use the ULS method only when
S

is nearly not positive definite.
2

After
completion of the estimation LISREL

produces various measures of fit, of which the
Goodness of Fit Index (GFI), the Adjusted Goodness of Fit Index (AGFI) and the Root
Mean Square
Residual (RMSR
) will be presented
together
with the
model
estimates
.


2.2

Correlations from discrete data


The l
inear variance decomposition model of network dynamics specified in LISREL
assumes that all observed network indicators
N
ijt
,
N
ijt
-
1
,
N
i

t
,
N
i

t
-
1
,
N

jt
,
N

jt
-
1
,
N
∙∙
t

and
N
∙∙
t
-
1


are measured on
an interval or ratio scale
. But with discrete measurements of

network

indicators

they are not;
N
ijt

and
N
ijt
-
1

are even measured as dichotomous variables having a



2

Nevertheless, if
S

is nearly positive definite then boundary problems of convergence may occur resulting in
insignificant negative estimates of one (or more)
variance(s). Such problems can be solved by fixing the value
of each such a variance equal to zero.

8




value of either one or zero. To solve this discrepancy, the polychoric correlation coefficient
described

by Olsson (1979) and programmed in PRELIS


(Jöres
kog & Sö
rbom, 1995)
has
been applied.

The polychoric correlation coefficient represents the association of two normally
distributed
standardized continuous
latent variables, which lay under two observed discrete
variables;
in other words,

every observed di
screte variable is conceived as the discrete
realization of an underl
ying continuous
latent
variable. This approach

also

has been
successfully
applied in the analysis of
the
succession

of

individual discrete
directed
network
relations

(e.g. Schweinb
erger &

Snijders, 2007). Before estimation of the polychoric
correlation coefficient
, the binomial or multinomial distribution of each observed discrete
variable is

folded


under the normal distribution of a standardized
continuous
latent variable
.
Whether or no
t the binomial or multinomial distribution is skewed
and/
or peaked is irrelevant;
in case of skewed distributions the

folding


represents a nonlinear data transformation like
the

log
-
transformation.
Subsequently
,

the minimum variance unbiased estimate of
the
correlation of every pair of standardized
continuous
latent variables is
derived

by means of the
ML method
. In a Monte Carlo study of various measures of
the
association
for

observed
discrete variables this polychoric correlation
is

demonstrated to out
perform all other te
sted
measures of association (Jöreskog & Sö
rbom, 1995:

10
-
17
).

The estimated polychoric correlations between the observed
discrete
network indicators
N
ijt
,
N
ijt
-
1
,
N
i

t
,
N
i

t
-
1
,
N

jt
,
N

jt
-
1
,
N
∙∙
t

and
N
∙∙
t
-
1

are the

standardized covaria
nces

of the continuous latent variables underlying them, which are placed
in the input matrix
S

to
be analyzed by LISREL. The means and variances of
the standardized continuous latent
variables

are by
definition

equal to zero and one, respectively.


3.

Da
ta,
characteristics

and results


3.1

Data collection


The data used for testing the variance decomposition model and assessing the various
dynamics in an empirical discrete network have been obtained from the Dutch BioPartner
Program. This program was started
in 2000 by the Ministry of Economic Affairs to stimulate
entrepreneurship in the life sciences as well as the entire industry (Ministry of Economic
Affairs, 1999). The program provided
advisory services and
seed capital for life sciences
9




based new product
development by Dutch ventures and start
-
ups. Furthermore, a registration
of Dutch firms active in the life sciences industry was set up in order to monitor the
development of these firms by means of a yearly survey since 2002. Participation in these
survey
s was obligatory for firms that received funding from the BioPartner Program
. The
program ended in 2005.
As of 2002 till 2005

four
annual

surveys were sent out. In these
surveys, except
for the

2003

survey
, the Dutch dedicated life sciences firms (DDLSFs)

were
asked

to list

names of their five most important partner organizations. The maximum of 5
partners to be mentioned induces, however, a truncation of the total number of established
relations, which was also asked for in each survey.

This truncation
re
sults in

data on
the
actual collaborations of DDLSFs with individual partner organizations
that
represent
about

80% of all collaborative relations of DDLSFs

in 2002, 2004 and 2005. Due to the omission
of data on partner specific collaboration of each DDLSF

for 2003, the data for 2002, 2004
and 2005 represent two periods of unequal length. Therefore, the data have been analyzed as
two successive waves (2002
-
2004 and 2004
-
2005) and not as successive moments of
observation in one

set of panel data
.

Each wave h
as been analyzed separately via estimation
of the variance decomposition model specified.

The data on actual collaborations of DDLSFs with partner organizations differ in nature
from the data on announced collaborative agreements
between organizations
used

by
, among
others, Gulati & Gargiu
lo (1999) and Powell et al (2005). The data on actual collaborations
may reflect besides formal
ized

also informal activities. Furthermore, these data show clearly
whether or not actual collaboration is continued,
terminate
d

or disrupted after one year.
But
t
he dynamics of actual collaboration taking place within one year remain unobserved. In this
respect, the data on collaborative agreements give more insight into
the speed of
their
conclusions and by that
into

the intensi
ty of collaboration
. However,
these data provide no
insight into when collaborative agreements come to an end

(e.g. Ahuja, 2000
-
a,b)
.
Hence
, it
remains unknown whether successive collaborative agreements between two organizations
accumulate and strengthen
the
ir

collaboration or that they
succeed one another and
only
continue the collaboration or that they reflect the reestablishment of their collaboration

after

disruption.
And
what
if those organizations do not conclude successive collaborative
agreements o
r less than before
?

I
s that indicative of ended or less collaboration between
them or of successful continuation of a p
reviously concluded agreement? So, b
oth types of
data

on collaborative relations between organizations have their peculiarities. The adva
ntage
of the data on actual collaborations is that they
provide more information about

the ecology
10




of such relations than the data on announced collaborative agreements. But the latter type of
data
provide more information about

the speed and
changes in in
ten
sities of collaborations

than the data on actual collaborations.


3.2

Characteristics of the data


From the 2002 and 2004 surveys 108 DDLSFs were derived that at least in one of them
replied to the question on partner names. These 108 DDLSFs said to co
operate with 220
partner organizations (of which 11 and 8 DDLSFs in 2002 and 2004, respectively).
Additional desk research was done to find out when a DDLSF responded only to either the
2002 survey or the 2004 survey whether or not it already existed in 20
02 or ceased to exist in
2004. DDLSFs not existing in 2002 or 2004 were assigned missing values for that particular
year and left out in further analyses of the observed variables via pairwise deletion. From the
2004 and 2005 surveys 96 DDLSFs were derived

that at least in one of them replied to the
question on partner names. These 96 DDLSFs said to cooperate with 179 partner
organizations (of which 8 DDLSFs in both years). As in the previous wave, DDLSFs not
existing in 2004 or 2005 were assigned missing v
alues for that particular year and left out in
further analyses of the observed variables via pairwise deletion.

The status of DDLSFs, partner organizations and their collaborative relations and
changes thereof are presented in Table 1 together with the bi
rth and death of organizations
and their relations
. The column totals represent the situation of the DDLSFs, partner
organizations and their collaborative relations at the beginning of each observation period
whereas the row totals represent their situatio
n at the end of each observation period. The
reported percentages in Table 1 indicate the part of the total number of collaborative
relations reported by the DDLSFs in the surveys that are included in both data sets

From the figures on the organizational a
nd relational ecologies contained in
both

data
set
s

presented in Table 1
it can be derived that the population of
collaborating
DDLSFs
remains rather

stable
; 71% of the DDLSFs in 2004 already existed in 2002 and 88% of the
DDLSFs in 2005 already existed in

2004.

However,

the
ir

collaborative network is
very

volatile between
2002,
2004 and 2005;
75% of t
he relations in 2002, 69%

of the relations in
2004
and 58% of the relations in 2005
were changed within one year. Furthermore, the
density o
f this network is
extremely low
,

i.e.
0.8
%,
1.0
% and
0.8
%

in
200
2, 2004 and 2005,
respectively.
11




Table 1. Organizational ecologies of DDLSFs and partner organizations and the ecology of their collaborative relations



total

2002

2004


conn. disconn. org. death

total
2004

2005


conn. disconn. org. death

connected


DDLSFs disconnected


org. birth

64


18


33 26 5



18



26

77


26


45 32



16 10



3

DDLSFs total

2004 / 2005




77 26 5




64 32 10

connected

Partner Orgs.

org. birth

137


48 89



83

131


48 83



48

Partner Orgs total

2004 ./ 2005




131
89




96 83

establ. relations


Collab. new relations
from
:

Relations
-

disconn. orgs.


-

org. birth

189

(77%)


47 130 12




90



61

198


62 136




76



10

Collab. total

Relations 2004./.2005



198 130 12


(81%)



148 136


(90%)



12




Insight into the hierarchical structure of the network can be derived from the descriptive
statistics for the distributions of the node degrees > 0 of th
e DDLSFs and their partner
organizations in
2002,
2004 and 2005, respectively.


Table 2
. Weighted binomial moments
3

and power degree
of the node degree distributions


DDLSFs

partners


2002

2004

2005

2002

2004

2005

mean

2.95

2.57

2.31

1.38

1.51

1.54

varianc
e

2.
36

1.93

1.84

1.06

2.28

2.21

skewness

0.
30

0.4
5

0.
67

2.99

3.80

3.48

kurtosis


-
0.
85


-
0.99


-
0.80

8.46

16.25

12.47

power degree


0.13

0.55

0.86

2.08

1.84

1.77


N

64

77

64

137

131

96


As skewness and kurtosis are zero for a
n approximated

normal distri
bu
tion, it can be
derived from T
able 2

that the
centrality

of DDLSFs represented by their node degrees is
virtually normally distributed in
all ye
ars of observation

(e.g. Faber,1988)
.
The skewness
and kurtosis of the distributions of node degrees of the part
ner organizations in
2002,
2004
and 2005 show that there is a clear hierarchy in the
centrality

of partner organizations
during
the

years
of observation
with a few partner organizations having many collaborative
relations with DDLSFs and many partner organ
izations having only one or a few
collaborative relations with DDLSFs.
Although the power degree decreases over time it
remains

substantial
, which is

also indicative of the hierarchy in the
centrality

of partner
organizations.
From

these statistics it may
be derived that the network of collaborative
relations of DDLSFs is more likely to be structured by the partner organizations

than by the
DDLSFs themselves.

This preliminary conclusion is supported by the membership of organizations of the top
-
10 of node d
egrees of DDLSFs and partner organizations in the years 2002, 2004 and 2005.
Only one of the 10 most
central

DDLSFs in 2002 appears also in the top
-
10 of 2004.

The
same is true for the top
-
10 of DDLSFs in 2004 and 2005. Regarding the top
-
10 of most



3

As the moments of the multinomial distribution of a discrete variable with more than 2 categories are
undefined, they are approximated here by using the m
oments of the binomial distribution of every category and
weighting them by the number of categories identified (e.g. Evans et al, 2000).

13




central

partner organizations another
picture emerges:

six

of the partner organizations in the
top
-
10 of 2002 are also
in t
he top
-
10 of 2004. I
n the top
-
10 of 2004 and
that of
2005 seven
partner organizations are the same. These stable most
central

partner organi
zations are all
Dutch universities
with

life sciences departments.

The dynamics in the network of collaborative relations of DDLSFs
over time
will be
investigated in the next section based on estimation
s

of the variance decomposition model of
such dynamics

presented before.

T
he large volatility of collaborative relations after one
observation

period
reflects that the effects of inducements and opportunities of both parties
to collaborate
predominantly

operate
within
that period
.

Therefore,
these effects are

specified to occur at the same moment of observation as the best approximation of effects
operating within
any

period of observation

(Fisher,
1970
).


3.3

Results on
variance decomposition of
network dynamics


The estimations of the various unknown constan
t parameters in the variance
decomposition model specified in LISREL by means of one of the three methods mentioned
before
are based on the input correlation matrices
S

presented in the Appendix (N =
23760

and N =
17184
). The elements of
S

are polychoric c
orrelations estimated by means of
PRELIS. In the Appendix the correlations of the entire network indicators
N
∙∙
t

and
N
∙∙
t
-
1

are
not presented because with two years of observation
in each wave
there is only one
observation available for each indicator res
ulting in standardized variances and covariances
of these indicators equal to zero. So, the effects of
an
increasing or decreasing
density

of the
network on the establishment of a collaborative relation between a DDLSF and a partner
organization cannot be
estimated. But as the density of the network of collaborative relations
of DDLSFs hardly changes over the
four

years of observation, the effects of changes in
network density
may

be

conceived as

very small
. Furthermore, as standardized variables
have zero
means also the intercept
a

is by definition equal to zero and accordingly not
specified. These limitations affect the final specification of the variance decomposition
model in LISREL as follows. The measurement mod
el of unobserved structural and

14




temporal
latent variables contains the following equations
L
N
ij
t

=

M
i
j


+

M
ij
t

,

L
N
ij
t
-
1

=

M
i
j


+

M
ij
t
-
1

, L
N
i

t

=

M
i


+

M
i

t

,

L
N
i

t
-
1

=

M
i


+

M
i

t
-
1

, LN

jt

=
M

j


+

M

j
t

and

LN

jt
-
1

=
M

j


+

M

j
t
-
1

and the influence relations model in eq(3
-
a,b,c) reduces to

M
i
j

=
b1 M
i



+
b2 M

j


+
e
ij


(4
-
a)

M
ij
t

=
b
3

M
i

t

+
b
4

M

j
t

+
e
ij
t


(4
-
b)

M
ij
t
-
1

=
b
3

M
i

t
-
1

+
b
4

M

j
t
-
1

+
e
ij
t
-
1


(4
-
c)

Each

LISREL model containing
the reduced variance decomposition model

has been
estimated from
one of the input correlation matrices

S

in the Appendix. This is done by
means of the ULS method becaus
e both correlation matrices contain

four correlations that
are virtually zero
thereby
making
S

in each
analysis

nearly not positive definite as
mentioned

before. The ULS estimates of the various unknown constant parameters
in
the

models
for

the
periods 200
2
-
2004 and 2004
-
2005
are presented below in table
s

2
and 3
, respectively

(all
with
p < 0.001
).

The fit of the reduced variance decomposition model to each input matrix
S

is excellent as reflected by the values of the GFI, AGFI and RMSR measures.


Table 2.

LISREL estimates of the variance decomposition model of discrete
collaborative network dynamics of DDLSFs, 2002
-
2004


b1

=
-
0.2394

var(M
ij
t
)

= 0.2661

var(M
i

t

)

= 0.8403


b2

=

4.6613

var(M
ij
t
-
1

)

= 0.2661

var(M
i

t
-
1

)

= 0.8321


b3

=

0.4333

var(M
i
j

)

=

0.7339

var(M
i



)

= 0.1638


b4

=


0.3035


2
eijt

= 0.0178

var(M

j
t

)

= 0.9830




2
1

eijt

= 0.0224

var(M

j
t
-
1

)

= 0.9497




2
eij

= 0
*

var(M

j


)

= 0.0333

*
: fixed at zero after
an
insignificant negative est
imate
(p>0.10)

GFI = 0.999; AGFI = 0.998; RMSR

=
0.014

15





Table
3
.

LISREL estimates of the variance decomposition model of discrete
collaborative network dynamics of DDLSFs
, 2004
-
2005


b1

= 0.5623

var(M
ij
t
)

=
0.1977

var(M
i

t

)

=
0.9994


b2

= 1.9066

var(M
ij
t
-
1

)

=
0.1977

var(M
i

t
-
1

)

=
0.9616


b3

= 0.3351

var(M
i
j

)

=
0.8023

var(M
i



)

=
0.0191


b4

= 0.3018


2
eijt

=
0.0025

var(M

j
t

)

=
0.9104




2
1

eijt

=
0.0095

var(M

j
t
-
1

)

=
0.8803




2
eij

=
0.4167

var(M

j


)

=
0.1044

GFI = 0.999; AGFI = 0.999; RMSR

=
0.014


The polychoric variance of dichotomous network relations
(= 1.0)
contains a structural
variance
(
var(M
i
j

)
)

of 0.734 (73.4%) and 0.802 (80.2%) in 2002
-
2004 and 2004
-
2005,
respectively. For the independent
node degrees of DDLSFs and partner organizations another
picture appears. The structural variances of the node degrees of DDLSFs

(var(M
i



)
)

and
partner organizations
(var(M

j


)
)

are only 0.164 (16.4%) and 0.033 (3.3%) in 2002
-
2004 and
0.019 (1.9%) and
0.104 (10.4%) in 2004
-
2005
, respectively
. This implies that the node
degrees of DDLSFs and partner organizations are very volatile over time; with
an
increasing
volatility of the DDLSFs’ node degrees and
a
diminishing volatility of the partners
organizatio
ns’ node degrees.

The structural variance of dichotomous network relations
in 2002
-
2004
(var(M
i
j

)
=
0.7339
)

consists of a variance
equal to

0.009 induced by the structural variance of the node
degree of DDLSFs
(
b1
2
*

var(M
i



)
)
,

a variance
equal to

0.717 induced by the structural
16




variance of the node degree of partner organizations
(
b2
2
*

var(M

j


)
)

and a variance
equal to

0.0075 induced by the structural variance of unidentified
dyad

specific characteristics
(
var(M
i
j

)

-

b1
2
*

var(M
i



)

-

b2
2
*

var(M

j


)
)
. So, 98% of the structural variance of
dichotomous network relations is explained by the structu
ral variance of the node degree of
partner organizations in 2002
-
2004. This result reflects a very strong preferential attachment
of DDLSFs in 2002
as well as

2004, irrespective of their
centrality

in the network
, to a few
very central

par
t
ner organization
s over time.

During the period 2004
-
2005, the
structural part of the
polychoric variance of
dichotomous network relations (= 1.0)
equals 0.8023 (80.2%) and
consists of a variance
equal to

0.006 induced by the structural variance of the node degree of DDLSF
s, a variance
equal to

0.380 induced by the structural variance of the node degree of partner organizations
and a variance
equal to

0.417 induced by the structural variance of unidentified
dyad

specific
characteristics like, for example, mutual complementa
rities of specific
technologies
. So, in
2004 and 2005 only 47% of the structural variance of dichotomous network relations is
explained by the structural variance of the node degree of partner organizations and 52% by
the structural variance of unidentifie
d
dyad

specific characteristics. After comparison of
these results with those for 2002
-
2004 it can be derived that the preferential att
achment of
DDLSFs to a few
very central

partner organizations has become weaker and that
dyad

specific partnering, irresp
ective of the
centrality

in the network
of the DDLSF and partner
organization involved, has become stronger in partner selection for collaboration and the
structuring of the collaboration network of DDLSFs. The weakened influence of structural
preferential

attachment is also reflected in the smaller estimate of
b2

being

1.907
for

2004
-
2005
against

4.661
for

2002
-
2004.

The temporal variances of dichotomous network relations in the years of observation
(
i.e.
var(M
ij
t
)

for 2004 and 2005 and
var(M
ij
t
-
1

)

for 2
002 and 2004, respectively) are 0.266
(26.6%) in 2002
-
2004 and 0.198 (19.8%) in 2004
-
2005; that is a reduction of
nearly

7% of
the total polychoric variance
o
f dichotomous network relations.
These temporal variance
s

are
almost entirely induced by
both
a
re
lative
stronger effect of the temporal variance of the
node degree of DDLSFs and a
relative
weaker effect of the temporal variance of the node
degree of partner organizations
(
b3


b4)
. So, the temporal variance of network relations
17




seems to be induced by
u
nequal

mutual preferential attachment between more
central

DDLSFs and partner organizations. But the resulting temporal collaborative relations
between these typical DDLSFs and part
ner organizations do not last

and the influence of
temporal unequal mutual
preferential attachment diminishes over time
.


4. Exploration of i
mplications

4.1 Introduction


The results of the variance decomposition model reveal that this model distinguishes
structural as well as temporal mechanisms of network dynamics whereas in pr
evious studies
of dyadic inter
-
organizational collaborative relations (e.g. Gulati & Gargiulo, 1999; Powell
et al, 2005) only structural mechanisms of network dynamics are assessed. Furthermore, the
results show that by analyzing successive waves of data o
n
an inter
-
organizational
collaboration network also the succession of structural and temporal structuring mechanisms
can be assessed. The theoretical implications of these more detailed results and insights will
be
explored

in Section 4.2
.

However, the
es
timated
variance decomposition model is a descriptive statistical model
that distinguishes methodologically distinct sources of variation, which themselves are only
components and no explanation of the variation in dyadic collaborative relations;
otherwise

the components would not explain 100% of the variance of dyadic collaborations. This
implies that

the effects of the variance components (network, actors, dyad and linkage)
reflect only empirical regularities instead of an explanation (e.g. Faber & Schepe
r, 2003).
Consequently, the variance components themselves need to be
additionally
explained by
network, actor, dyad and/or linkage specific attributes in order to provide an explanation of
dyad specific collaborations. So, the results obtained from the va
riance decomposition model
help to focus additionally required theory formation. The question then arises from the 100%
explanation of the variance of dyad specific collaboration whether or not the results and
network structuring mechanisms identified from

them make sense or are artifacts of the
variance decomposition model specified. This question will be addressed in Section 4.
3

by
relating these results to the then prevailing policy context and results
obtained
from
qualitative research into the strategi
c behaviors of
some
DDLSFs.
Subsequently,

the
implications of the results for the policies pursued by the Dutch government are
explored
.


18




4.
2

Theoretical implications


In this section the theoretical implications of the results presented before will be
com
pared with those derived by
Stuart (1998),
Gulati & Gargiulo (1999) and Powell

et al
(2005)

as they also applied a relational perspective in their empirical research

of inter
-
organizational collaborations
. In
all

these
studies positive structural effects o
f previous ties
on the establishment of collaborative relations were found. This effect is confirmed in this
study by the structural variances of collaborative relations
(var(M
i
j

))

, which reflect

the
common variance
(i.e.

auto
-
covariance
)

of collaborative relati
ons over time.
Powell et al
(2005)
found that
this effect
operates
without an accompanying effect
of the collaborating
organizations’ network positions whereas
Stuart (19
98) and
Gulati & Gargiulo (1999)
found
that
this effect
operate
s

besides a
n independent

positive effect of their joint network position.
In this study, which disentangles the structural and temporal network position
s

of

individual
organizations participat
ing in bilateral collaborative relations, the structural positive
autocorrelation of successive collaborative relationships
, that is the positive effect of the
dyadic history of collaboration,

is found to depend strongly on the

structural

network
positions

of the partner organizations irrespective of the
structural
network positions of the
DDLSFs involved.
In other words,
the

large
structural
centrality

of the partner organization
s

stimulated

the
persistence

(or
unobserved repetition

within one
period of ob
servation
) of
bilateral collaborative relationship
s
. This result
reflects

the concept of ‘structure begets
behavior’ utilized in complex network analysis (e.g. Strogatz, 2001). As the structural
collaborations of the DDLSFs with especially Dutch universiti
es
containing

life sciences
departments are for 24% in 2002, 46% in 2004 and 40% in 2005 initiated by those
universities

(including spin
-
offs)

it can be derived that the preferential attachment of these
DDLSFs to the universities originated from both sides
; that is, universities searched for

and
established

DDLSFs and DDLSFs searched for universities to exchange
or combine
complementary resources. But as the DDLSFs’ structural network positions in terms of
centrality

play an insignificant role in this proce
ss in terms of explanatory power,
the
concept of

liability of dis
connectedness (Powell et al, 1996) receives little support from the
res
ults presented in this study.

The liability of disconnectedness does get some support with the results from a temporal
p
erspective.
Temporarily less
central

DDLSFs are less likely to collaborate with a partner
19




organization than temporarily more
central

DDLSFs. But
, simultaneously,

temporarily less
central

partner organizations are
also
less likely to collaborate with a DDLS
F than
temporarily more
central

partner organizations
.
These

results
can

also be interpreted as that
m
ore connected DDLSFs tend to collaborate temporarily with more connected partner
organizations

just
because both are more connected
, which allows them to
explore

mutual
collaborative relationship
s

without
severe penalties of a failure.
This

interpretation

coincides
with the finding of Powell et al (2005) that the most central organizations in the
ir

network

adopt
ed

a diversification mechanism in their collab
orative behavior next to a preferential
attachment mechanism.
The former mechanism is argued to be related to the exploration of
collaborative relationships
whereas

the latter mechanism is argued to be related to the
exploitation of collaborative relations
hips.
The mechanism of
diversification

is
also
argued
to incorporate novelties in the network and is therefore considered as

an important source

of
change within the network.

The mechanism
of preferential attachment
is also found in this
study

but

as struc
tural preferential attachment of DDL
SFs to Dutch universities.
T
he

results

on temporarily collaboration
also
do not
provide
much
support
for the concept of

liability of
dis
connectedness
hindering DDLSFs in establishing collaborative relations with partner
organizations
as

they are also searched for by partner organizations.

Another implication of the results presented in this study is that different collaboration
mechanisms operate simultaneously in the collaboration network of DDLSFs as found
before by Pow
ell et al (2005) for American
dedicated biotechnology firms

in the Boston area
but also that these mechanisms evolve over time. Structural preferential attachment and
temporal mutual preferential attachment both become weaker
whereas

structural dyad

specif
ic collaboration
becomes stronger
irrespective of
the participating organizations’

centrality
. This implies that
at
first
centrality

in the network

is crucial for an organization’s
access to information, knowledge or other complementary resources but that
later the
organization seems to have learned where to get them

from actual collaboration(s) an
d
information diffusion via indirect ties (Ahuja, 2000
-
b)
.
So
,
membership

of a network during
a longer period of time decreases the organization’s dependency on
i
ts

network position for
establishing collaborative relations with other organizations. Consequently, learning about
the network and its members
as a result of

continued membership of that network
plays

an
important role i
n the evolution of collaboration

me
chanism
s (de)structuring the network.
More i
nsight into the content and conditions facilitating this kind of
organizational
learning
requires
,

however,

further research.

20




From these insights the following image of the evolution of collaborative behaviors of

DDLSFs and their partner organizations appears. At the beginning when the collaboration
network emerged, DDLSFs established collaborative relations
, irrespective of their
centrality,

preferentially with Dutch universities occupying central positions in th
e network.
It seems as if

Dutch universities have selected DDLSFs to collaborate with quite at random
via trial and error. But the DDLSFs and universities having established multiple
collaborative relations
also
searched
for
each other

to explore more bene
ficial collaborative
relations as they were not
solely
dependent on each other

for immediate survival
. As a result
of the information gathering via this exploration and
the
established direct and indirect ties
,

several DDLSFs and universities learned quick
ly to establish collaborative relations

between
them

that were seemingly more beneficial in terms of contents exchanged
or combined
irrespective of their network positions. These
DDLSFs and universities decreased their
numbers of collaborative relations as

they (partly) found what they were looking for. With
their retreat from the tails of the node distributions of DDLSFs and partner organizations, the
node distribution of DDLSFs got more skewed to the right and that of the partner
organizations got less sk
ewed to the right as
is b
est

reflected
by

their, respectively
,

increasing

and decreasing power degrees (T
able 2)
. T
his
may
i
mply

that for the remaining
DDLSFs and partner organizations, especially new entrants, it may have become more
difficult to establis
h beneficial relations. For these DDLSFs the guidance of search based on
the centrality of partner organizations as an indicator

of opportunities to find a
beneficial
relation diminishes. At the same time, with a sharper articulation of the centrality of s
ome
remaining DDLSFs it becomes clearer for remaining partner organizations which DDLSFs
offer more opportunities for establishing a benefic
ial collaboration. Extrapolation

of these
tendencies over time indicates that

the collaborative behavior of DDLSFs a
nd partner
organizations based on (mutual) preferential attachment will have diminished further while
that based on structural
dyad

specific characteristics will have increased. As a result, it may
be expected that
eventually
the
growth

of the network will

slow down

due to

deteriorating

chances of survival for new and non
-
connected DDLSFs. This implies that the concept of
liability of disconnectedness (Powel et al, 1996)
was not

operational during the emergence
of the collaboration network of DDLSFs but
mor
e likely
later
when
the network

had
developed further
.

Nevertheless, it should be
recalled

that these structural aspects of the
evolution of collaborative behaviors of DDLSFs and their partner organization and the
resulting collaboration network are far fr
om deterministic. They exert their influences
on

a
21




rather bumpy
series

of very volatile
successive
collaborations with fast chang
ing
opportunities and uncertainties
for the
participating organizations.


4.
3

Policy context and implications


The two
-
way stru
ctural preferential attachment of DDLSFs to especially Dutch
universities
with

life sciences departments irrespective of the DDLSFs’ network position
s

may be related to two contextual situations, namely an already developed network of
collaborative relatio
nships of DDLSFs
and/
or specific Dutch governmental policies

implemented
. The first contextual situation of an already developed collaborative network
wherein the participants have become acquainted with each other is rather unlikely. 56%,
65% and 71% of t
he DDLSFs active in 2002, 2004 and 2005, respectively, were established
since 2000 when the BioPartner Program of the Ministry of Economic Affairs was
implemented. One might even say that after implementation
of the program the collaboration

network of DDL
SFs emerged as a result of the sharp increase of foundation
s

of DDLSFs
seemingly stimulated by the provision of seed capital subsidies

and
advisory services
. The

subsidies
might

even be
en

the reason for the low organization
al

death rate of DDLSFs
during th
e
observed period 2002
-
2005. So, the age of the co
llaboration

network of DDLSFs
seems not to be reason for the two
-
way structural preferential attachment within the network
an
d the absence of liability of dis
connectedness hindering DDLSFs to establish coll
aborative
relations.

The second contextual situation of specific Dutch governmental policies
implemented

seems more likely to have resulted in the two
-
way structural preferential attachment.
Frustration about good scientific IT research and virtually no IT

industry development
during the 80’s and 90’s of the last century, the virtual absence of an established
pharmaceutical industry that might trigger biotech
nology based

industry development and
good life sciences research reflected by the successful partic
ipation of several Dutch
academic

research groups in the Human Genome Project together induced the Dutch
government to take up an active role in life sciences ind
ustry development. T
his active role
consisted of the implementation of

the BioPartner Program
providing new ventures and
start
-
ups in the life sciences industry with seed capital and
advisory services
.
In the same
period
a public debate
developed
about the role of universities in society

and
the economy

and their participation in knowledge valoriza
tion.
After time
, all Dutch universities set up
22




incubators and technology transfer offices and adopted a policy of active knowledge
utilization. This resulted in a growing number of spin
-
offs, especially from the life sciences
departments, which also recei
ved funding from the BioPartner Program. Another result was
that the technology transfer offices began actively searching for DDLSFs that could benefit
from especially R&D collaboration with life sciences departments within the universities.
So, DDLSFs wer
e not alone in their search for beneficial collaboration as they were also
searched for by Dutch universities. In this context, DDLSF
-
university collaboration
gave

legitimacy to both their
activities

but also to the life sciences industry policy pursued by

the
Dutch government.

This policy has been successful in two
respects
, namely
a sharp increase of the number
of DDLSFs

founded and
the
survival of DDLSFs during thei
r first years after founding
.
The
DDLSFs operate
d

only
to a
limited

degree

outside their i
nitial technological field of
expertise (van der Valk et al, 2009) and change
d

after time from
either

a product
/service

or
technology

development strategy to a hybrid strategy of
technology development

on the
short term and product
/service

development on t
he long term (Willemstein et al, 2007)

in
order to increase their chances of survival
. The change after
time

to a hybrid strategy within
the initial technological field indicates that DDLSFs became more aware of how to increase
their (limited) chances of s
urvival and what they needed to improve them and where to get it.
Consequently, more content related dyad specific collaborations may be expected to have
developed as is reflected by the results of the variance decomposition model.

However,
the policy purs
ued by the Dutch government with the BioPartner Program
seems to have a paradoxical outcome on the long term for Dutch life sciences industry
development. The provision of seed capital and advisory services helped DDLSFs with their
founding and survival du
ring their first years of existence.

But as explained in the previous
section, after time
various

members of the collaboration network of DDLSFs found
seemingly more beneficial collaborations with established partner organizations, irrespective
of their ne
twork positions, that induced both of them to ret
reat partially from the network.

Thereby it
seems to have become

more difficult for the remaining DDLSFs and new entrants
to find and establish beneficial collaborations, which resulted

for them

in deteriora
ting
chances of survival. Consequently, it seems as if the initial success of the BioPartner
Program lost its effectiveness after time as a result of the dynamics of successive
collaboration mechanisms that it provoked.

23




An
other

intriguing issue to be addre
ssed is

the
large

volatility of the collaborative
relations of DDLSFs
.
Only 47 and 25 of the 189 collaborative relations of DDLSFs in 2002
survived in 2004 and 2005, respectively; that is 25% after two years and 13% after three
years. It seems as if
many

D
DLSFs
were

wandering around searching for
finance and/or
facilities in order to survive on the short term. Or
has

this collaborative behavior
(also)
been
induced by
the speed of
successive technology, service and product development
s

requiring
additional k
nowledge inputs from other organiz
ations

on short terms (e.g. Gay & Dousset,
2005)
? Consequently, one of the most intriguing questions to be addressed in future research
is not
‘W
hy
are

dyadic
collaborative relations

between DDLSFs and their partner
organi
zations established
?’
, which subject already

has
been investigated in
several

studies,
but
‘W
hy
are

dyadic
collaborative relations

between DDLSFs and their partner organizations
changed or not
?


And
did

this happen

for evolving reasons and in different way
s? These
latter
two questions are not only theoretically relevant in the light of the results produced by
the variance decomposition model presented before but are also

of practical relevance

for the
policies pursued by the Dutch government and other organ
izations

(including competitors)
in
order
to
address

more
effectively

the
development of
DDLSFs.


5.

Discussion and conclusions

5.1

Discussion


The variance decomposition model
of structural and temporal dynamics in
discrete non
-
directed

networks
explains

97%

and more

of the variance of collaborative relationships
between DDLSFs and partner organizations in the period 2002
-
2005. The model is
unprecedented and combines statistical methods developed in psychometrics (i.e.
correlations between normally distrib
uted continuous variables derived from dichotomous
an
d discrete data; Olsson, 1979
), sociometrics (i.e. structural and temporal c
omponents in
longitudinal data;
Wheaton et al, 1977) and economet
rics (i.e. panel data analysis;
Johnston
& DiNardo, 1997). But

the nature of the model is quite different from those utilized in
previous studies of
dyadic
collaborative behavior
and

relations
of

organizations. The latter
model
s

were
partially
specified on theoretical grounds and tested for their validity. The
varian
ce decomposition model
is a time
-
dependent

two
-
way analysis of variance model of a
graph represented as a matrix

and is specified on statistical grounds to decompose the
variance of relations into methodologically mutually exclusive components
.
The two
-
way

24




analysis of variance

model relates the values of the cells to the column and

row totals and the
overall total being mutually exclusive sources of variance

(see eq.2)
. As the row and column
totals correspond to the node (out)degrees of DDLSFs and the node
(in)degrees of their
partner organizations, respectively, and the overall total corresponds to the network density,
the correspondence between the graph of a network and the two
-
w
ay analysis of variance
model is

established. The effects of the column, row
and overall totals on the cell values
determine which part of the variance of cell values is due to each of them or cell specific (i.e.
due to unidentified
dyad

specific characteristics), which together account for all
the
variance
of cell values. However,

the column, row and overall variances and their effects and the cell
variance left over are empirical regularities to be explained themselves on theoretical
grounds (e.g. Salancik, 1995). So, the variance decomposition model helps to unravel
complex netwo
rk dynamics in order to better focus theo
ry formation
on

its largest
constituent
variance components; in other words, it increases insight into complex
empirical
network dynamics and helps to systemize (further) theory formation about it

but does not
repre
sent a theoretical model
. In this respect, Barabási & Albert’s (1999) mechanism of
preferential attachment is only a label of an empirical regularity instead of an explanati
on
(e.g. Faber & Scheper, 2003).

Another difference between the variance decomposit
ion model and earlier models of
collaborative behavior and relations is that the former model takes explicitly structural as
well as temporal effects into account whereas the latter models concentrate on structural
effects only

(e.g. Gulati & Gargiulo, 199
9; Powell et al, 2005)
. As demonstrated by the
results presented and discussed before, the temporal effects play an important role in the
evolution of collaborative behavior and relations within a rather volatile
inter
-
organizational
network. But in the mo
dels specified and tested in earlier studies such temporal effects and
the
effects of random errors collapse and
cannot

be separated from one another. In this
respect, the variance decomposition model
provides
important
additional information about
the dyn
amics of
inter
-
organizational

networks.

The differences between the data used in this study and those used in other studies have
been already discussed in Section 3.
These differences in the measurement of collaborative
relations between organizations and
their possible effects on the results obtained from data
analysis provide an interesting t
opic for further research but that

is beyond the scope of this
study.
Nevertheless,
the results obtained in this study
seem to

correspond

with
some

obtained by Powell

et al (2005) regarding the behavioral mechanisms of preferential
25




attachment and diversification applied by the most centra
l organizations
.
But in this study
both mechanisms are derived from the results obtained for one (meta) model specification of
comple
x network dynamics whereas in Powell et al
(2005)
separate models are

specified

and
tested

together
. Furthermore, the relative dominance of each mechanism can be assessed
from the results obtained in this study based on their contribution to the explained
variance
of established collaborative relations. And it can be derived whether these mechanisms are
structural or temporal in nature.

Another novelty of the data used in this study is that they seem to reflect the emergence
of the collaborative network of
DDLSFs
as a result of

the BioPartner Program of the Dutch
government since 2000. Accordingly, the results presented in this study provide specific
insight into the sources of variance and processes relevant during the
early development

of
an

industrial net
work and the industry itself, namely structural trial
-
and
-
error selection of
new firms by central established organizations and temporal
mutual
exploration of
relationships

between the most central new firms and established organizations.

Unfortunately
, th
e data used in this study cover only two periods of observation with
different le
ngths. Consequently, the effect

of changes in network density on the
establishment of collaborative relations between DDLSFs and their partner organizations
and its reducing i
mpact on the effects of other specified sources of variance could not be
estimated.
Because

the BioPartner Program ended in 2005 and data collection with it, it will
be impossible to gain insight into these effects from
future

research.


5.2

Conclusions


A
pplication of the variance decomposition model of complex network dynamics in the
analysis of inter
-
organizational collaborations of Dutch dedicated life sciences firms has
made clear that ‘it takes two to tango’ and that the individual network positions o
f both
collaborating organizations

in each dyad

must be explicitly taken into account. Compound
measures of these network positions and other organizational characteristics obscure the
effects inducing inter
-
organizational collaboration.
Additionally
, the
results show that a
distinction should be made between structural and temporal effects related to
, respectively,

th
e exploitation and exploration

of collaborative relations between organizations.
Furthermore
, as demonstrated, time series of observations of

collaborative relations between
26




organizations should be analyzed as successive waves (or windows of time) in order to gain
insight into the evolution of collaboration mechanisms structuring the network.

Although the variance decomposition model is a descr
iptive statistical model and not a
theoretical explanatory model it contributes to inter
-
organizational network analysis by
systemizing

‘the
groping in the dark’ and focusing theory formation on those variance
components that explain most of the variance o
f inter
-
organizational collaborative

relations.
Systemizing ‘
the
groping in the dark’ consists of
specifying

the possible effects of all
methodological mutually exclusive categories of sources of variation

on the (non)existence
of collaborative relations b
etween organizations
. But
significant

sources of variation
themselves need further explanation on theoretical grounds. Nevertheless, the estimates and
standard errors of the possible effects and related variance components also provide
descriptive statisti
cal insights into the dynamics of the network studied as discussed in
Sections 4.1 and 4.2

as inputs for future re
search
.

Finally, the variance decomposition model of discrete networks elevates the analysis of
(un)weighted (non)directed graphs to a multiva
riate statistical analysis. By using polychoric
correlations as measures of association between observed discrete phenomena the vast array
of
more informative

multivariate statistical methods comes available for
the analysis of
graphs
. But in every multiva
riate statistical analysis based on polychoric correlations
estimated from observed discrete network data their threefold panel data
dimensions (i.e.
i, j
and

t
)

should be taken into account.

Regarding the collaboration net
work of Dutch dedicated life scie
nces firms (DDLSFs)
the results

indicate a swift evolution of successive collaboration mechanisms
; i.e. from
dominant structural preferential attachment of DDLSFs to cent
ral partner organizations
(
Dutch universities)
and considerable

temporal mutual attach
ment between the most central
DDLSFs and partner organizations in the period 2002
-
2004 to dominant dyad specific
collaboration between DDLSFs and partner organizations and quite weaker temporal mutual
and structural preferential attachment mechanisms of co
llaboration in the period 2004
-
2005
.
This evolution of successive collaboration mechanisms occurred in the context of a very
volatile collaboration network of DDLSFs with much rewiring; of the collaborative relations
established
betw
een DDLSFs and their partner organizations

only 25% and 13% survived
after two and three years, respectively. Consequently, in order to improve our understanding
of the collaborative behavior of DDLSFs and the structuring of their collaboration network
the

question of ‘why do DDLSFs’ collaborative relation
s

change or not?’

becomes central
.

27




Future research into the
this

question will
further
improve
the
insight into the inducements
and opportunities affecting the collaborative behavior of DDLSFs

and establis
hed partner
organizations
.

28




Appendix: Polychoric correlations of the observed discrete network indicators


2002
-
2004 (N=23760)



0.10000D+01


0.31661D+00 0.10000D+01


0.41575D+00
-
0.45480D
-
07 0.10000D+01


0.73338D+00
-
0.41159D
-
01 0.18683D+00 0.10000D+0
1

-
0.37453D
-
01

0.16373D+00


0.44388D
-
07 0.32959D+00 0.10000D+01


0.12952D+00 0.25244D
-
07 0.22450D
-
01

0.48083D+00 0.10412D
-
06
0.10000D+01


N
ijt
-
1

N
i

t
-
1

N

jt
-
1

N
ijt

N
i

t

N

jt


2004
-
2005 (N=17184)


0.10000D+01

0.30423D+00 0.10000D+01

0.43943D+00
-
0.10901D
-
07 0.10000D+01

0.80231D+00 0.10814D
-
01 0.19875D+00 0.10000D+01

0.11755D
-
01 0.18736D
-
01 0.15564D
-
07 0.37329D+00 0.10000D+01

0.20027D+00 0.26
760D
-
07 0.10417D+00 0.49835D+00
-
0.89346D
-
07 0.10000D+01


N
ijt
-
1

N
i

t
-
1

N

jt
-
1

N
ijt

N
i

t

N

jt

legend:

-

N
ijt

: dichotomous relationship between DDLSF
i

and partner organization
j

in 2004 or 2005

-

N
i

t

: node degree of DDLSF
i

in 2004 or 2005

(
t
-
1

= 2002

or 2004
)

-

N

jt

: node degree of partner organization
j

in 2004

or 200
5

-

D
-
07 : * 10
-
7

in double precision

29




References


Ahuja, G., 2000
-
a. The duality of collaboration: inducements and opportunities in the formation
of interfirm linkages. Strategic Manag
ement Journal 21, 317
-
343.

Ahuja, G., 2000
-
b. Collaboration networks, structural holes and innovation: a longitudinal
study. Administrative Science Quarterly 45, 425
-
455.

Barabási, A.L., Albert, R., 1999. Emergence of scaling in random networks. Science 28
6, 509
-
512.

Baum, J.A.C., Calabrese, T., Silverman, B.S., 2000. Don’t go it alone: alliance network
composition and startups’ performance in Canadian biotechnology. Strategic
Management Journal 21, 267
-
294.

Clauset, A., Shalizi, C.R., Newman, M.E.J., 2007.

Power
-
law distributions in empirical data.
arXiv: o706.1062v1 (eprint).

Duijn, M.A.J. van, Snijders, T.A.B., Zijlstra, B.J.H., 2004.
p
2
: a random effects model with
covariates for directed graphs. Statistica Neerlandica 58, 234
-
254.

Eisenhardt, K.M., Scho
onhoven, C.B., 1996. Resource
-
based view of strategic alliance
formation: strategic and social effects in entrepreneurial firms. Organization Science 7,
136
-
150.

Faber, J., 1988. Consistent estimation of correlations between observed interval variables
wit
h skewed distributions. Quality and Quantity 22, 381
-
392.

Faber, J., Scheper, W.J., 2003. Social scientific explanations? On Quine’s legacy and
contextual fallacies. Quality & Quantity 37, 135
-
150.

Faber, J., Poot, T., 2009.
Multi
-
level auto
-
covariance ana
lysis

of structural and temporal
dynamics in discrete
non
-
directed
networks
: specification, application and extensions
.
Social Networks (
accepted
).

Fisher, F.M. 1970. A correspondence principle for simultaneous equation models.
Econometrica

38
, 73
-
92.

Fugu
itt, G.V., Lieberson, S., 1974. Correlations of ratios and difference scores having common
terms. In Costner, H.L. (ed). Sociological Methodology 1973
-
1974. Jossey
-
Bass, San
Francisco, 128
-
144.

Gay, B., Dousset, B., 2005. Innovation and network structural
dynamics: study of the alliance
network of a major sector of the biotechnology industry. Research Policy 34, 1457
-
1475.

30




Gulati, R., Gargiulo, M., 1999. Where do interorganizational networks come from? American
Journal of Sociology 104, 1439
-
1493.

Hagedoorn
, J., 1993. Understanding the rationale of strategic technology partnering: inter
-
organizational modes of cooperation and sectoral differences. Strategic Management
Journal 14, 371
-
385.

Johnston, J., DiNardo, J., 1997. Economeric methods. McGraw
-
Hill, New
York.

Jöreskog, K.G., Sörbom, D., 1993. LISREL
®

8: user’s reference guide. Scientific Software
International, Lincolnwood (IL).

Jöreskog, K.G., Sörbom, D., 1995. PRELIS


8: user’s reference guide. Scientific Software
International, Lincolnwood (IL).

Kenny,

D.A., La Voie, L., 1984. The social relations model. In: Berkowitz, L. (ed).
Advances in experimental social psychology (vol. 18). Academic Press: Orlando,
141
-
182.

Ministry of Economic Affairs, 1999. Action plan life sciences. Ministry of Economic
Affair
s, The Hague.
.

Mitzenmacher, M., 2004. A brief history of generative models for power law and lognormal
distributions. Internet Mathematics 1, 226
-
251.

Olsson, U., 1979. Maximum likelihood estimation of the polychoric correlation coefficient.
Psychometrika

44, 443
-
460.

Pfeffer, J., Salacik, G.R., 1978. The external control of organizations: a resource dependence
perspective. Harper & Row Publishers, New York.

Powell, W.W., Koput, K.W., Smith
-
Doerr, L., 1996. Inter
-
organizational collaboration and the
locus
of innovation: networks of learning in biotechnology. Administrative Science
Quarterly 41, 116
-
145.

Powell, W.W., White, D.R., Koput, K.W., Owen
-
Smith, J., 2005. Network dynamics and
field evaluation: the growth of interorganizational collaboration in the
life sciences.
American Journal of Sociology 110, 1132
-
1205.

Sakakibara, M., 2002. Formation of R&D consortia: industry and company effects. Strategic
Management Journal 23, 1033
-
1050.

Salancik, G.R., 1995. Wanted: a good network theory of organization. Ad
ministrative Science
Quarterly 40, 345
-
349.

31




Saris, W.E, Stronkhorst, L.H., 1984. Causal modeling in nonexperimental research.
Sociometric Research Foundation, Amsterdam.

Schweinberger, M., Snijders, T.A.B., 2007. Markov models for digraph panel data: Monte

Carlo
-
based derivative estimation. Computational Statistics & Data Analysis 51,
4465
-
4483.

Snijders, T.A.B., Kenny, D.A., 1999. The social relations model for family data: a multi
-
level approach. Personal Relationships 6, 471
-
486.

Strogatz, S.H., 2001. Ex
ploring complex networks. Nature 410, 268
-
276.

Stuart, T.E., 1998. Network positions and propensities to collaborate: an investigation of
strategic alliance formation in a high
-
technology industry. Administrative Science
Quarterly 43, 668
-
698.

Van der Valk
, T., Moors, E.H.M., Meeus, M.T.H., 2009. Conceptualizing patterns in the
dynamics of emerging technologies: the case of biotechnology developments in the
Netherlands. Technovation 29, 247
-
264.

Wheaton, B., Muthén, B., Alwin, D.F., Summer, G.F., 1977. Asse
ssing reliability and
stability in panel models. In: Heise, D.R. (ed). Sociological Methodology 1977.
Jossey Bass, San Francisco, 89
-
136.

Willemstein, L., Van der Valk, T., Meeus, M.T.H., 2007. Dynamics in business models: an
empirical analysis of medical
biotechnology firms in the Netherlands. Technovation 27,
221
-
232.