Embed Size (px)
Citation preview
Inter-Team Technical Communication
in Complex New Product
Development Projects
_______________
Manuel SOSA
Martin GARGIULO
Craig ROWLES
2012/65/TOM/OB
(Revised version of 2011/119/TOM/OB)
Inter-Team Technical Communication in Complex
New Product Development Projects
M a n u e l S o s a *
Martin Gargiulo**
Craig Rowles***
Revised version of 2011/119/TOM/OB
* Associate Professor of Technology and Operations Management at INSEAD, 1 Ayer Rajah
Avenue, Singapore 137686. E-mail: [email protected]
** Professor of Organisational Behaviour at INSEAD, 1 Ayer Rajah Avenue, Singapore
137686 . E-mail: [email protected]
*** Pratt & Whitney Aircraft
A Working Paper is the author’s intellectual property. It is intended as a means to promote research to
interested readers. Its content should not be copied or hosted on any server without written permission
from [email protected]
Find more INSEAD papers at http://www.insead.edu/facultyresearch/research/search_papers.cfm
1
Inter-Team Technical Communication in Complex New Product
Development Projects
Abstract
This paper investigates how the structure of the informal communication network that results from efforts
to coordinate task interdependence between design teams in complex new product development efforts
may itself moderate the role of task interdependence in explaining the occurrence of inter-team
communication. Based on theoretical mechanisms drawn from the social network literature and recent
empirical advances in exponential random graphs models of social networks, we investigate how the
presence of third parties that are common to the ―acquirer‖ and the ―provider‖ of technical
communication in a dyad of interdependent teams moderates the effect of task interdependence in
predicting the occurrence of inter-team communication. We test our hypotheses by examining the
network of task interdependencies and the technical communication network among teams designing the
components of a large commercial aircraft engine. We find that while task interdependence is a
significant predictor of communication between teams, task interdependence is less likely to predict inter-
team communication to the extent that the acquirer team may receive information from the provider team
indirectly through common third parties. However, task interdependence is more likely to predict inter-
team communication if the provider team shares common third parties with the acquirer team, albeit in
this case the moderating effect of common third parties is only apparent for strong interdependencies.
Keywords: inter-team communication, social networks, information redundancy, social control,
new product development
2
1. Introduction
Since the pioneering work by Allen (1977), organizational scholars have highlighted the crucial role of
informal communication networks in new product development organizations. Because informal
communication can help teams coordinate interdependent tasks, a proper understanding of such networks
has been recognized as a key element in improving the performance of product development
organizations (Tushman 1977, Tushman and Katz 1980, Ancona and Caldwell 1992, Brown and
Eisenhardt 1995, Reagans and Zuckerman 2001, Reagans et al. 2004, Lomi and Pattison 2006). The
importance of understanding communication networks in new product development organizations became
even more apparent once Henderson and Clark (1990:15) stated that both formal and informal
communication channels are ―the relationships around which the organization builds [product]
architectural knowledge‖ and elaborated on the consequences of a poor understanding of such channels
for incumbent organizations developing complex products while facing architectural innovation.
Despite the acknowledged importance of informal communication channels for the performance
of product development organizations, we still have a limited understanding of the factors responsible for
the observed communication networks among teams during the development of complex systems (Colfer
and Baldwin 2010). The difficulties are both theoretical and analytical. From a theoretical viewpoint,
existing research suggests that such networks result from a complex interplay of factors involving task
interdependence, spatial configuration, formal, and informal organizational structures. Analytically, the
fact that the occurrence of a certain tie in a communication network is not independent of the co-
occurrence of other ties in the same network poses significant challenges for the statistical analysis of
social networks (Contractor et al. 2006), which only recently scholars have been able to overcome (see
Robins et al. 2007a,b for review).
Task interdependence has been probably the factor most commonly invoked by scholars to
explain observed patterns of informal communication in organizations (Galbraith 1973, Thompson 1967,
Henderson and Clark 1990, Sosa et al. 2004, Gokpinar et al. 2010). In this view, informal communication
ties emerge from attempts to coordinate tasks among interdependent actors. Yet, scholars have also found
3
that other factors, such as spatial arrangements (Allen 1977, 2004; Van den Bulte and Moanaert 1998),
the formal structure (Nadler and Tushman 1997), and even the informal structure of the organization can
also influence communication patterns in the organization (Lomi and Pattison 2006, Rank et al. 2010).
Spatial arrangements and formal structures are designed to serve the information processing needs of the
organization (Allen 1977, Nadler and Tushman 1997), but both factors also have unintended
consequences that can shape communication networks. Research shows that physically distant teams are
significantly less likely to exchange information than proximate teams are, even in the presence of task
interdependence (Allen 2004, Sosa et al. 2002). Formal organizational boundaries facilitate
communication among actors within the boundaries (Sosa et al 2004), but they also discourage cross-
boundary exchanges, generating the well-known ―silo effect‖ that hinders coordination between different
organizational units (Lessard and Zaheer 1996, Gulati 2010).
An intriguing influence on observed patterns of communication is that of the very structure of the
informal communication network. Research on network analysis has shown that the occurrence of a
certain communication tie is not independent of the co-occurrence of other ties in the network (Contractor
et al. 2006). For instance, informal communication network in new product development organizations
tend to exhibit strong tendencies for reciprocation as well as to form transitive triads (Lomi and Pattison
2006, Rank et al. 2010). This property of network structures not only poses important challenges for the
statistical analysis of social networks (Wasserman and Pattison 1996) but also has substantive theoretical
significance, because it suggests that endogenous variables can be an important factor in explaining the
occurrence of informal communication ties. Although such ties may emerge as a response to coordination
needs between interdependent actors, making task interdependence one of the most salient determinants
of communication, the very structure of the emerging communication network may moderate the effect of
task interdependence in predicting communication. It is worth noting that this endogenous influence does
not represent a tautology or constitute a circular argument. Rather, it results from an underlying structural
tendency of social networks (Contractor et al. 2006:686).
4
This paper studies how the structure of the informal communication network that results from
efforts to coordinate task interdependencies between teams may itself moderate the role of task
interdependence in explaining the occurrence of inter-team communication. In developing our hypotheses,
we build on recent research on organization theory, product development, and network analysis, as well
as on methodological advances that make it possible to model endogenous network effects typically
present in social networks. We also benefit from a rich dataset that documents the product architecture,
formal organization arrangements, and network of informal communication among teams in charge of
designing the different components of a large commercial aircraft engine (Sosa et al. 2004). The data also
allows us to control for spatial influences: teams were co-located the same place within the facilities,
hence physical distance could not affect significantly observed communication patterns. Each team was in
charge of designing a specific component of the engine, whereas technical interfaces between components
determined the structure of task interdependencies between the teams.
Recent advances in our understanding of the mechanisms operating in social networks enable us
to formulate specific hypotheses on how the structure of the communication network among design teams
may affect the inter-team communication patterns. Specifically, we argue that the number of
communication parties shared by two interdependent teams (i.e., teams whose components are linked by a
directed technical interdependence) can affect the behavior of these teams in ways that moderate the
effect of task interdependence in predicting inter-team communication. Two findings are especially
relevant in this context. First, common third parties to two interdependent actors can enable indirect flows
of information between these actors, providing the actor seeking information with data on the actions of
the other actor even in the absence of direct communication. In a context of resource constraints, the more
an actor can receive indirect information through common third parties, the less likely this actor will be to
invest scarce resources in obtaining additional information directly from the source, because this
information is more likely to be perceived as ―redundant‖ (Burt 1992, 2005). Second, an actor is more
likely to collaborate with another actor if the relationship is embedded in a dense network of common
third parties (Reagans and McEvily 2003). The presence of common third parties entails concerns about
5
local reputation, which create normative pressures to respond to requests from other parties in the network
(Coleman 1990, Raub and Weesie 1990, Burt 2004, Gargiulo et al. 2009).
Building on these findings, we argue that when a design team (the acquirer) is in charge of
designing a component that can be affected through a technical interface by another component designed
by a second team (the provider), then the acquirer team is less likely to request technical information
directly from the provider when it can receive that information indirectly through third parties that get it
from the provider. We also argue that a design team is more likely to provide information if it shares
common communication parties with the team requesting that information, because the presence of such
teams raise reputational concerns in the provider. Hence, we expect that the structure of the
communication network moderates the extent to which the network of technical interdependences
between components predicts the inter-team communication network.
We test these ideas by analyzing the technical interdependences and the informal technical
communication network among design teams in charge of the development of the Pratt & Whitney
PW4098 commercial engine that powers the Boeing 777 twin-engine aircraft. Several factors justified the
selection of this project to study. First, it featured a complex design involving explicit decomposition of
the engine into subsystems and of those subsystems into components. Such an explicit product breakdown
made it easy to identify the technical interfaces between engine components. Second, the formal
assignment of a single design team to each engine component and the grouping of teams ―mirroring‖ the
subsystems of the engine facilitated the implementation of our research approach because it not only
allowed us to capture the formal and informal organizational structures present during the design phase of
the engine’s development but also because it allowed us to measure task interdependence between design
teams based on the technical interfaces between the engine components they designed. Third, the project
was the most recent engine program to complete design and development at Pratt & Whitney, so almost
all team members involved in the detail design phase were still accessible.
The nature of our arguments poses some significant analytical challenges. In essence, we need to
model the presence or absence of a communication tie between a pair of design teams as a function of the
6
presence (or absence) of communication ties between these two teams and common third parties, as well
as of other endogenous and exogenous covariates. In other words, we need to model the presence or
absence of ties in the informal communication network as a function of factors that are endogenous to that
same network—such as the presence of common third parties to a relationship. This renders traditional
statistical methods inappropriate, because observations are clearly not independent. For this reason, we
test our predictions using exponential random graphs models (ERGMs), also known as p* models, which
allow us to obtain estimates of the effects of endogenous network parameters on the probability of
observing communication ties between the design teams in our data (Wasserman and Pattison 1996,
Contractor et al. 2006, Snijders et al. 2006, Robins et al. 2007a, Hunter et al. 2008, Robins et al. 2009).
2. Theory and Hypotheses
Informal communication between teams is an important mechanism to coordinate interdependent tasks in
formal organizations (Nadler and Tushman 1997). Such organizations seek to structure tasks in ways that
minimize coordination requirements and put in place a number of formal mechanisms to meet these
requirements (Thompson 1967, Galbraith 1973). In the specific case of engineering organizations
developing complex products, efforts to reduce coordination requirements focus on minimizing the
number and complexity of the interdependences between product components without compromising
product functionality (Baldwin and Clark 2000). The remaining coordination requirements are met with a
mix of formal and informal organizational mechanisms. Formal mechanisms involve decisions such as
assigning the responsibility of designing product components to design teams, grouping these teams into
subsystems groups led by a manager, and introducing special teams to facilitate horizontal coordination
across teams designing related components. Informal mechanisms rely on the efforts of design teams to
seek task-related information from other teams. We call these informal exchanges inter-team technical
communications. These informal communications are particularly important in the development of new
and complex products: The very novelty and complexity of such products makes it harder to anticipate the
nature and the intensity of task interdependencies, which limits the effectiveness of formal mechanisms in
7
coordinating such interdependencies and strengthens the role of inter-team technical communications
(Henderson and Clark 1990, Sosa et al. 2004, Gokpinar et al. 2010).
Despite the importance of informal inter-team communication in helping coordinate task
interdependencies, our knowledge of the factors that shape the emergence of such communication
patterns remains vague. The basic tenet is that informal communication network mirrors to a large extent
the structure of task interdependencies between design teams (Henderson and Clark 1990). The mirroring
results from the fact that a team whose task is affected by the task of another team typically seeks to
acquire the relevant information from the second team, which is expected to provide such information to
the first. Empirical evidence confirms that this is indeed the case (e.g., Sosa et al. 2004, Gokpinar et al.
2010). Task-related informal communication between design teams is typically triggered by a request for
information from the ―acquirer‖ team to the ―provider‖ team (McCord and Eppinger 1993, Morelli et al.
1995, Terwiesch et al. 2002). Hence, holding the effects of formal mechanisms constant, inter-team
technical communications should simply mirror the structure of task interdependence in the organization.
The expected isomorphism between the structure of task interdependence and the structure of the
communication network provides the baseline hypothesis for our analysis:
H1: The probability of observing a directed technical communication tie from team j (the
provider) to team i (the acquirer) increases with the presence of a directed task interdependence from
team j to team i.
Although the relationship between the task interdependence and the inter-team technical
communication networks stated in H1 has been demonstrated in a number of occasions (see Colfer and
Baldwin 2010 for a review), the isomorphism is not perfect, making it important to identify the factors
that may be responsible for departures of the predicted relationship—that is, factors that may moderate
the effect of task interdependence in predicting informal inter-team technical communication. The
importance of these factors is both theoretical and practical. From a theoretical viewpoint, understanding
what factors moderate the effect of task interdependence on tie formation should allow for better models
of the occurrence of informal communication in organizations. From a practical viewpoint, such models
8
could help managers focus their attention on task interdependencies that are less likely to trigger the
informal communication that could help achieve the necessary coordination between the actors involved.
Previous work has shown that the formal organizational structure and the strength of the task
interdependence moderate the relationship between task interdependence and inter-team communication
(Sosa et al. 2004). In this paper, we focus on the moderating effects of the informal inter-team
communication network structure. Next, we build on recent developments in social network analysis to
formulate hypotheses on how the structure of the technical communication network between design teams
moderate the relationship between task interdependence and inter-team communication.
The Moderating Effects of the Communication Network Structure
The possibility that the very structure of the informal communication network that emerges out of the
efforts by actors to coordinate their task interdependence could affect the form of this same network is
fully consistent with our knowledge of the endogenous mechanisms that shape social networks. Research
on network analysis has shown that the occurrence of a given tie in a network is not independent of the
co-occurrence of other ties in this same network (Wasserman and Pattison 1996, Contractor et al. 2006,
Lomi and Pattison 2006, Rank et al. 2010). Insofar as the occurrence of ties in the informal
communication network are dependent on variables that are endogenous to this same network, such
endogenous influences can moderate the association between task interdependence and inter-team
communication. In this paper, we focus on a particular property of the inter-team communication network,
the number of third parties common to two design teams, and theorize how the presence of common third
parties moderates the effect of task interdependence in predicting inter-team communication by affecting
the behaviour of teams (as either acquirer or provider of technical information) responsible for designing
interdependent components of a complex system.
Social network theory suggests that the presence of common communication parties can affect
the behavior of a design team through two different mechanisms: information redundancy and social
control. On the one hand, network scholars have argued that the information an actor i receives from actor
9
j may be redundant if both i and j are connected to a common third party q, since the information might
already have reached i indirectly through q. Burt (1992, 2005) proposes this mechanism to argue for the
information benefits of sparse networks in which dyadic relationships do not share common third parties.
The more a relationship between two actors is embedded in common third parties, the higher the
redundancy of the information that might circulate between these two actors. On the other hand, research
has shown that common third parties also create incentives to cooperate because actors are mindful that
failure to do so can be observed by those third parties, which would have a negative impact on the
offending actor’s reputation (Granovetter 1985, Raub and Weesie 1990). The presence of common third
parties also facilitates the emergence and the enforcement of collaborative norms (Coleman 1990; Barker
1993, Gargiulo 1993), which becomes a form of ―social control‖ on the behavior of the involved players.
Research on knowledge transfer confirms that the presence of common third parties makes it more likely
that the provider will transfer knowledge to the acquirer (Reagans and McEvily 2003, Gargiulo et al.
2009) while Lazega (2001) provides a rich account of the role of common third parties in enforcing
cooperative norms.
Figure 1 illustrates how information redundancy and social control can moderate the effect of
technical interdependence on the occurrence of technical communication between two design teams. The
figure depicts a directed interdependence xij through which the component designed by team j can affect
the component designed by team i. As stated in H1, such interdependence predicts the occurrence of
technical communication flowing from team j to team i because team i is likely to seek to acquire task-
related information from team j, and team j is likely to provide such information to team i. Yet the
presence of common communication parties q1 and q2 may affect the behavior of the two teams in ways
that moderate the effect of task interdependence xij in predicting the occurrence of the inter-team
communication tie yij.
10
Figure 1: The moderating roles of common third parties
Because technical information circulates through communication ties between teams, a focal team
i sharing an interdependence xij with team j is likely to receive indirect information on this
interdependence if j sends information to team q1, which in turn sends information to i. The more that
team i’s members could get information about component j indirectly through other teams q1, the more
likely that they would feel confident that they know enough about that component, rendering efforts to
acquire additional information directly from design team j superfluous. This confidence should increase
with the number of common third parties q1 that receive information from j and send information to i. In
such a situation, the presence of a technical interdependence from j to i is less likely to prompt the
occurrence of direct communication from j to i.
It is worth noting that our argument here assumes that technical information flows only in the
direction of the technical communication ties. This assumption is consistent with findings of research on
new product development, which shows that the flow of technical task-related information follows the
i j
p
p
q1
INFORMATION REDUNDANCY
Common third parties q1, which fall in two-
step indirect-communication paths going
from j to i, can transmit information to i
about i’s interdependence with j, making i
less likely to request information from j
Technical interdependence(x)
Technical communication (y)
p
p
p
yiq
xij
yqjq1
q2
q2
SOCIAL CONTROL
Common third parties create
normative pressures to cooperate,
making j more likely to respond to
information requests from team i
i j
p
p
q1
INFORMATION REDUNDANCY
Common third parties q1, which fall in two-
step indirect-communication paths going
from j to i, can transmit information to i
about i’s interdependence with j, making i
less likely to request information from j
Technical interdependence(x)
Technical communication (y)
Technical interdependence(x)
Technical communication (y)
p
p
p
yiq
xij
yqjq1
q2
q2
SOCIAL CONTROL
Common third parties create
normative pressures to cooperate,
making j more likely to respond to
information requests from team i
11
direction of inter-team communication (e.g. Smith and Eppinger 1997, Terwiesch et al. 2002) and is
grounded in the very nature of the information exchanged. Because technical information is necessarily
complex, team q1 is more likely to be aware of the design choices made by team j when q1 acquires
technical communication from team j rather than when q1 only provides information to j, as this last
information is likely to refer to q1’s (and not to j’s) component. Hence, q1 is more likely to be viewed
by i as a ―reliable‖ source of technical information on j’s design choices only when q1 acquires technical
information from j. Because the presence of third parties q1 offer alternative ways to team i to get
technical information from team j, then the existence of a task interdependence from j to i is less likely to
predict technical communication from j to i. Thus, we offer the following hypothesis:
H2: The number of common third parties that fall into a two-step indirect communication
path going from j to i attenuates the positive effect of task interdependence xij in predicting the occurrence
of technical communication from provider team j to acquirer team i, yij.
While the information redundancy mechanism associated with the presence of indirect
communication paths through common third parties affects the behavior of design teams as acquirers of
information, the social control mechanism triggered by the presence of any common third parties affects
the behavior of teams as providers of information. The role of common third parties in facilitating the
emergence and enforcement of normative pressures to collaborate is a common theme in the sociological
literature (e.g., Granovetter 1985, Coleman 1990). The presence of common third parties to a relationship
makes actors aware that their actions can be monitored and that non-cooperative behavior can easily be
made known to (or observed by) those third parties, which would damage the reputation of the offending
actors (Raub and Weesie 1990, Ahuja 2000, Reagans and McEvily 2003, Gargiulo et al. 2009). If the
acquirer team i requests task-related information from j, the presence of common third parties to that
relationship should make j more likely to provide such information, because failure to do so may become
known by the common third parties. This, in turn, may affect j’s relationships with these common third
parties as well as with other teams that could become aware of the uncooperative behavior through the
communication network. Hence, the presence of common third parties should make team j more likely to
12
provide information to i, enhancing the positive effect of task interdependence in predicting the
occurrence of technical communication. This leads to our third hypothesis:
H3: The number of common third parties between teams i and j enhances the positive effect
of task interdependence xij in predicting the occurrence of technical communication from provider team j
to acquirer team i, yij.
While the redundancy mechanism proposed in H2 assumes that technical information flows in the
direction of the communication, such a restriction is not necessary for the social control mechanism
behind H3. In this case, the simple fact that the two teams get in touch should suffice for exchanging
comments about the behavior of other teams in the network: ―We still haven’t heard from j, that’s really
slowing us down!‖ Thus, the effectiveness of the control mechanism is independent from the direction of
the flow of technical communication between the teams. This difference between the network structures
behind the redundancy and control mechanisms allows us to construct different measures for the variables
used to test the two hypotheses predicated on these mechanisms, captured in the different position of third
parties q1 and q2 in Figure 1. Third parties q1 fall within two-step indirect communication paths going
from j to i, whereas third parties q2 include all other possible common third parties to the i,j dyad in the
communication network.
The Strength of the Technical Interdependence as a Contingency
The moderating effects of common third parties proposed in H2 and H3 are likely to be contingent on the
strength of the technical interdependence, which increases with the number and the criticality of design
dependencies between two given product components. The stronger the interdependence, the less likely it
is that the acquirer team would be satisfied with indirect information obtained through common
communication parties and the more likely it would be to seek first-hand information directly from the
source. In other words, the moderating effect of information redundancy proposed in H2 should be less
apparent in the presence of strong technical interdependences and more apparent in weaker
interdependences. Formally:
13
H4: The attenuating effect of the number of common third parties that fall into a two-step
indirect communication path going from j to i on the association between task interdependence xij and
inter-team communication yij decreases with the strength of this task interdependence.
In the same vein, the pressures on a provider j to behave cooperatively triggered by the presence of
common third parties should be more salient for stronger interdependences. In such cases, failure to
provide information to i about the interdependence xij will have a greater direct effect on i’s ability to
design its own component, which is more likely to trigger complaints about j’s neglect. Failure to provide
information about a weak interdependence may be viewed by both the rebuffed acquirer i and by the
common third parties as a tolerable breach of the cooperative norms, with only minor practical
consequences for the task at hand. However, failure to provide information on a strong interdependence
can have significant effects on all the parties involved, with serious consequences for the reputation of the
offending (non) provider. Thus:
H5: The enhancing effect of the number of common third parties on the association between
task interdependence xij and inter-team communication yij increases with the strength of this task
interdependence.
3. Data and Methods
We test our hypotheses by studying the detailed design phase in the development of a large commercial
aircraft engine. Figure 2 shows the product architecture and formal organizational structure of PW4098
engine. Figure 2(a) is a cross-sectional diagram of the engine that shows the eight subsystems into which
the 54 engine components were configured. Figure 2(b) shows the (flat) formal organizational structure
by which design teams were grouped into subsystems of teams, quasi-mirroring the architecture of the
engine.
We also captured network data associated with both the organization and the product. We
captured the technical communication network of the 60 teams (54 design teams plus 6 functional teams)
that designed the engine components and evaluated the overall engine performance as well as the
14
technical interface network of the 54 components comprising the engine (Sosa et al. 2004, Sosa et al.
2007). We collected data from multiple sources. Data on inter-team communications was gathered by
interviewing and surveying key members of the teams (Marsden 1990). Product network data was
constructed by interviewing several experienced engine architects.
Figure 2: (a) Aircraft engine studied and (b) the formal organizational structure of the project
Inter-team Communication Network Data
The informal communication network between the teams is defined by the presence or absence of task-
related technical communication between any two of the 60 teams (i.e., 54 design teams plus 6 functional
teams) involved in the ten-month detailed design phase of the development process. Our focus on
technical task-related interactions is akin to what Allen (1977) and Morelli et al. (1995) define as
―coordination-type communication.‖ The fact that the design of each engine component depends in
complex ways on the design of some other components and the stringent functional requirements of a
product that needs to operate faultlessly in extreme conditions highlights the importance of direct
communication between teams to adequately understand the effects that design choices might have on
product performance. Although we acknowledge that various types of inter-team technical information
flows are likely to exist during the design of complex systems, we did not try to distinguish between
specific types of technical exchanges and focused instead on an overarching measure that captures all
FAN
Low-pressure compressor (LPC)
High-pressure compressor (HPC)
Combustion chamber (CC)
High-pressure turbine (HPT)
Low-pressure turbine (LPT)
Mechanical Components (MC)Externals and Controls (EC)
(a) Cross-sectional diagram of the PW4098 engine (b) Formal organizational structure of the project
PW4098
Fan
7 D
esig
n T
ea
ms
LP
C7
De
sig
n T
ea
ms
HP
C7
De
sig
n T
ea
ms
CC
5 D
esig
n T
ea
ms
HP
T5
De
sig
n T
ea
ms
LP
T6
De
sig
n T
ea
ms
MC
7 D
esig
n T
ea
ms
EC
10
De
sig
n T
ea
ms
6 Functional Teams
FAN
Low-pressure compressor (LPC)
High-pressure compressor (HPC)
Combustion chamber (CC)
High-pressure turbine (HPT)
Low-pressure turbine (LPT)
Mechanical Components (MC)Externals and Controls (EC)
(a) Cross-sectional diagram of the PW4098 engine (b) Formal organizational structure of the project
PW4098
Fan
7 D
esig
n T
ea
ms
LP
C7
De
sig
n T
ea
ms
HP
C7
De
sig
n T
ea
ms
CC
5 D
esig
n T
ea
ms
HP
T5
De
sig
n T
ea
ms
LP
T6
De
sig
n T
ea
ms
MC
7 D
esig
n T
ea
ms
EC
10
De
sig
n T
ea
ms
6 Functional Teams
15
these possible technical exchanges in a single indicator. Doing so was both consistent with previous
research studying technical communication in engineering organizations (Allen 1977, McCord and
Eppinger 1993, Morelli et al. 1995) and important to avoid posing insurmountable cognitive hurdles to
our respondents that could jeopardize the quality of the data.
We measured the presence of technical communication between teams through a questionnaire
addressed to team leaders and validated (or corrected, when necessary) their responses by interviewing at
least one other key member of each team. Presentations describing the terminology and overall objective
of the data collection were made in two sessions to more than two thirds of the respondents; the others
were later briefed individually. Respondents were presented with the roster of the 54 design and 6
functional teams and asked to report if they have received non-trivial technical information from each of
the other teams during the design phase of the project. Respondents were encouraged to focus on
technical information exchanges that actually took place (i.e., ―how it was‖, not ―how it should have
been‖) and that contributed to ―fulfil the functional requirement of their component design or program
goals‖ (Rowles 1999, p. 51). We were able to obtain responses from 57 of the 60 teams. The three teams
whose direct responses were missing were the interface team at Boeing (one of the six functional teams),
one team in the High-pressure Compressor sub-system group, and one team in the Externals and Controls
sub-system group. Yet because these teams were identified as information providers by other teams in the
development process, we did not exclude them from the network. We estimated all our models excluding
the three missing teams, and after recalculating all the communication network variables without these
teams, the results were fully consistent with the ones reported here.
Consistent with previous research on technical communication (e.g., Allen 1977, Morelli et al.
1995, Lomi and Pattison 2006, Rank et al. 2010), we use a dichotomous definition of inter-team
communication and assume that relevant communication did take place insofar as team i (the acquirer)
acknowledged receiving some technical information from team j (the provider). Based on these data, we
reconstructed the informal technical communication network {y} among the 60 teams involved in the
project, where each element yij of {y} captures the transfer of technical communication from team j (the
16
provider) to team i (the acquirer). The resulting communication network contains 680 directed nonzero
inter-team communication ties (yij = 1) among the 60 teams—that is, 19% of all possible inter-team
communication ties.
Figure 3(a) and (b) exhibit the team communication network and its corresponding adjacency
matrix. In Figure 3(a), each node of the network is colored according to subsystem membership while
each link indicates technical communication between two teams. In Figure 3(b), a non-zero cell of the
adjacency matrix captures a directed inter-team technical communication. Because each team is part of an
organizational subsystem group (as shown in Figure 2(b)), then the elements of the adjacency matrix are
ordered such that teams that belong to the same subsystem group are sequenced together in the matrix so
that the blocks along the matrix’s diagonal represent subsystem group boundaries (i.e., non-zero cells
within the blocks represent inter-team communications within subsystem boundaries while non-zero cells
outside such blocks capture cross-boundary inter-team communications).
Figure 3: Inter-team technical communication network
Product Network Data
Our product data capture the breakdown of the engine structure into eight subsystems and 54 components,
as well as the five types of design dependencies among those components: spatial constraint, structural
FANLPC
HPC
CC HPT
LPT
MCEC
Functional Teams
(a) Communication network of 60 teams
(b) Communication Adjacency Matrix of 60 teams
FAN
LPC
HPC
CC
HPT
LPT
MC
EC
Functional Teams (FT)
FAN LPC HPC CC HPT LPT MC EC FT
FANLPC
HPC
CC HPT
LPT
MCEC
Functional Teams
FANLPC
HPC
CC HPT
LPT
MCEC
Functional Teams
(a) Communication network of 60 teams
(b) Communication Adjacency Matrix of 60 teams
FAN
LPC
HPC
CC
HPT
LPT
MC
EC
Functional Teams (FT)
FAN LPC HPC CC HPT LPT MC EC FT
17
constraint due to transfer of loads, exchange of material, exchange of energy, and exchange of
information, as identified by the system engineers. Even though such a detailed map of the product
architecture was not available at the project’s onset, design teams shared a common understanding of (i)
the engine’s division into subsystems and components and (ii) the relevant technical interfaces between
their own components and the components designed by other teams. Each technical interface xij is formed
by a vector of five design dependencies that capture the various types of technical linkages going from
component j to component i. Each type of design dependency x between components i and j was weighted
as either noncritical (xij,type = 1) or critical (xij,type = 2), depending on the impact on functionality of
component i due to its dependence on component j. We define the strength of technical interface xij as the
sum of its weighted design dependencies. There were 569 technical interfaces between the 54 components
identified by system architects. This means that 20 percent of the theoretically possible interfaces between
components were actually present in the project. Sosa et al. (2007) provide additional details on the scales
used to capture the product network data.
Figures 4(a) and 4(b) exhibit the component interface network and its corresponding technical
interface matrix. In Figure 4(a), each node of the 54-node network is colored according to subsystem
membership, with each link representing at least one directional design dependency between two
components. In Figure 4(b), a non-zero cell in the technical interface matrix represent a directed technical
interface between two engine components. The technical interface matrix shown in Figure 4(b) quasi-
mirrors the adjacency matrix shown in Figure 3(b) as their elements are identically sequenced so that the
first 54 elements of the adjacency matrix (the 54 design teams in the organization) mirror the 54 elements
in the technical interface matrix (the 54 engine components designed by each of the 54 design teams).
18
Figure 4: Technical interface network of 54 engine components
Variables
We model the observed network of inter-team technical communication {y} between the 54 design teams
(which is the dependent variable in the analysis) as a function of both exogenous and endogenous
variables. The key exogenous variable comprise the technical interface network, which captures the
presence of a directed technical interface xij between each pair of components in the product and therefore
indicates the existence of task interdependence between two teams. In addition, there are a number of
exogenous controls described below. The endogenous variables include measures of information
redundancy and social control, based on the presence of common third parties to a dyad in the inter-team
technical communication network.
Technical interfaces. We use a dichotomous measure of technical interfaces to test the first three
hypotheses of this paper. Hence, a technical interface going from component j to component i exists if
there is at least one design dependency (e.g., exchange of energy) going from component j to component i.
Because the presence of a technical interface indicates that the acquirer team has task interdependence
with the provider team we use this variable to test Hypothesis 1 (our baseline hypothesis) as well as to
construct the interaction effects to test hypotheses H2 and H3. However, because our fourth and fifth
hypotheses propose that the moderating effect of information redundancy and social control varies with
FAN LPC HPCCC HPT
MC
EC
LPT
(a) Technical interface network of 54 engine components
FAN
LPC
HPC
CC
HPT
LPT
MC
EC
FAN LPC HPC CC HPT LPT MC EC
(b) Technical interface matrix of 54 engine components
FAN LPC HPCCC HPT
MC
EC
LPT
FAN LPC HPCCC HPT
MC
EC
LPT
(a) Technical interface network of 54 engine components
FAN
LPC
HPC
CC
HPT
LPT
MC
EC
FAN LPC HPC CC HPT LPT MC EC
(b) Technical interface matrix of 54 engine components
19
the strength of the technical interface, we use a trichotomous version of interface strength to test these
hypotheses. This measure distinguishes between no interface (xij = 0), weak interface (xij < xij[mean]), and
strong interface (xij ≥ xij[mean]).
Information redundancy and social control. To construct measures of information redundancy
and social control, for each ordered dyad {i, j} formed by teams i (the potential acquirer) and j (the
potential provider), we counted the number of common third parties in the communication network as the
number of teams q that communicate on technical matters with both i and j. Two aspects of these
measures merit special attention. First, we calculate the number of common third parties for all possible
dyads and not just for dyads that share a technical interface. Although we model the network of technical
communication between the 54 design teams (for which could also identify the existence or not of
technical interfaces), we also include the six functional teams in the communication network to compute
the relevant network statistics, as these functional teams could act as ―third parties‖. However, our results
are robust to measures of information redundancy and social control that exclude functional teams as third
parties.
Second, the information redundancy and social control mechanisms have different sensitivity
with respect to the directionality of inter-team technical communications. Specifically, the information
redundancy mechanism assumes that technical information could pass from team j (the provider) to team
i (the acquirer) indirectly through q (common third parties); this requires a directed chain of technical
information going from j to q to i. In contrast, the normative nature of the social control mechanism
requires only the existence of common-third parties, irrespective of the directionality of the
communication with them (Gargiulo et al. 2009). This allows us to distinguish between two types of
common third parties, identified respectively as q1 and q2 in Figure 1. The q1 type can establish an indirect
chain of technical communication going from j to q1 to i, which can provide indirect information to i on j,
as proposed in the information redundancy mechanism. The q2 type comprises all other common third
parties to the i,j dyad.
20
Information redundancy for acquirer team i is measured as the number of third parties q1 that
create a two-step communication path going from team j to team i through q1. To determine this measure,
we use the binary asymmetric adjacency matrix, Y, shown in Figure 3(b) so that
( ) ∑
where yiq and yqj measure the existence (or absence) of technical communication from team q to team i
and from team j to team q, respectively.
Social control for provider team j is measured by counting exclusively the common third parties
q2 that do not create an indirect two-step communication path from j to i. This count provides a
conservative measure of the amount of social control that could be experienced by the provider, as these
third parties are not expected to affect the acquirer’s behavior. Because this measure excludes q1 third
parties, it allows us to accurately estimate the ―pure‖ social control effect of third parties on the provider’s
behavior. In order to construct this measure we first determine the ―overall‖ social control that includes
all common third parties between i and j. To determine this, we symmetrize the Y adjacency
communication matrix to construct the matrix Yb in which each element = 1 if either yji = 1 or yij = 1 in
the original communication matrix. Hence, overall social control is computed as follows:
( ) ∑
{ }
The ―pure‖ measure of social control is simply obtained by subtracting the value of the
information redundancy measure (that is, the number of common third parties q1) from this overall
measure of social control. Yet, our results are robust to using the measure of overall social control instead.
Our models also control for a number of exogenous component or team (node-level) attributes
that might have affected the likelihood of observing a directed technical communication between an
acquirer team and a provider team in the presence of a directed technical interface between their
respective components.
21
Component redesign. An important factor that affects both a team’s workload and the need to
communicate with other interdependent teams is the novelty of the team’s component. The more novel a
component with respect to the prior generation, the more likely it will involve extra work for the team and
the more likely it will affect interfaces with adjacent components. In contrast, components that carry over
a significant portion of design content from previous engine models decrease the demand for attention to
associated interfaces. We capture this factor by measuring component redesign as the percentage of novel
design content in a component relative to its design in the product’s previous version. Because it is
impossible to determine the exact amount of redesign in a component, we relied on estimates from the
design teams involved. We computed the percentage of redesign for the component of each member of
the acquirer-provider dyads
Component complexity. Another component attribute that can affect the team’s workload is its
internal complexity. We measure a component’s complexity in terms of its estimated number of distinct
parts. For this estimate we relied on the experience of one of the authors, who is a design expert with
substantial experience in similar engine programs and who also reviewed the design work for this
particular project. We control for the complexity of each of the two components in each dyad.
Component connectivity. This is simply the number of incoming and outgoing technical interfaces
of each component in a dyad, irrespective of their strength. For each component, the measure simply
counts the number of components with which it shares at least one interface, and it is thus equivalent to
the ―degree‖ of the component in the technical interface network. We also computed a similar measure
taking into account the strength of the interface. The results are robust to both specifications.
Team size. The team’s capacity to communicate with others might well be affected by resources
available to the team. Because managers—when allocating resources to design teams—may take into
account the workload entailed by component characteristics and by its interfaces with other engine
components, the effects communication network structures may be confounded by the effects of the
resources available to the team. To control for this possibility, we include a four-point ordinal variable
that accounts for the manpower resources allocated to teams. Although we were unable to collect precise
22
data on team size (which, in any case, varied throughout different stages of the process), we did obtain a
qualitative assessment of team size based on the direct experience of one of the authors in the project. We
enter the values of team size for each acquirer-provider dyad in our data.
Team degree. The effect of common third parties on a team’s behavior may be affected by its
overall number of communication partners. A similar number of common third parties may have a bigger
impact on the behavior of the team if the teams almost exclusively communicate with those third parties.
To account for this (and being consistent with the way we measure information redundancy and social
control), it is important to control for the number of providers an acquirer team receives information from,
and the number of teams a provider communicates with. Hence, we control for the in-degree for the
acquirer team and the full degree for the provider team in each dyad.
Same subsystem membership. Because each design team is formally assigned to a subsystem-
level group, teams in the same group might share some characteristics that affect their ability to attend to
technical interfaces. These teams will be also part of the same formal coordination structure, with their
leaders reporting to the manager responsible for the subsystem. For instance, the seven teams that formed
the fan group share technical expertise and experience relevant to the design of fan components, which is
different from that of design teams responsible for components of the low-pressure compressor or high-
pressure turbine subsystems. On the organizational side, teams in the same subsystem group share a
subsystem manager (responsible for planning and control of resources). All these factors create ―silo‖
effects that foster communications between teams within a sub-system while hindering communications
between design teams that belong to different sub-systems (Allen 1977, Sosa et al. 2004). To account for
this, we created dyad-specific dummies that are set to 1 if both the acquirer and provider team in a dyad
belong to the specific subsystem and to 0 otherwise.
Table 1 shows descriptive statistics and correlation coefficients of the key variables used in our
statistical analysis. All variables are defined for each directional dyad (N = 2,862).
23
4. Statistical Network Analysis
A standard tenet in network analysis is that the presence or absence of a relationship between two actors
in a network is not independent of the structure of that same network. This endogenous nature of the
statistical relationship between the network structure and the presence or absence of specific ties within
this same network makes traditional statistical techniques inappropriate to test our hypotheses. Hence, we
resort to exponential random graphs models (ERGMs), also known as p* models (Wasserman and
Pattison 1996, Robins et al. 2001, Snijders et al. 2006, Robins et al. 2007a, Hunter et al. 2008, Robins et
al. 2009). In essence, this technique allows for modeling an observed network—in our case, the binary
network of technical communication among the 54 design teams—as a function of both exogenous (i.e.,
the technical interface network, the attributes of both teams and components) and endogenous parameters
(e.g., the number of ties in the network, the number of reciprocal ties in the network, the number of
common third parties in the network). This approach allows us not only to obtain reliable estimates of the
effect of task interdependence on the probability of observing an inter-team communication tie but also to
examine how the presence of common third parties moderates such an effect, net of the effect of other
endogenous and exogenous factors that might also affect that probability (Contractor et al. 2006).
To estimate the effects of interest, we build a probabilistic model of the inter-team
communication network of design teams. This model considers inter-team communication as a random
variable. Hence, for i and j (which are distinct teams of the network of 54 design teams) we consider a
random variable Yij where Yij = 1 if design team i (the acquirer) receives technical information from team j
(the provider), and where Yij = 0 otherwise. We specify yij as the observed value of Yij and Y as the matrix
of all Yij variables while y is the matrix of observed ties (i.e. the binary adjacency matrix corresponding to
the inter-tem technical communication network of 54 design teams). In addition, w and x are realizations
of covariate matrices that contain the node-level controls and the value of the technical interface between
any two given teams, respectively. Hence, the model we estimate takes the following form:
24
( ) (
) ( )
(
) (∑ ( )
∑ ( )
∑ ( )
) (Eqn. 1)
where Y is a random network with possible ties Yij, y is a realization of Y, and is a normalizing quantity,
which is difficult to calculate analytically in networks of this complexity. A discussion of ERGMs and the
details of the statistical modeling approach we pursued are included in the Appendix.
Our model includes three types of parameters, which define the family of probability distributions
that can generate the observed inter-team communication network. (The coefficient parameters in the
model are estimated to maximize the fit between the random networks and the observed data.) First, we
include six basic network-level parameters (also called ―network statistics‖) present in a typical
organizational network (Contractor et al. 2006, Lomi and Pattison 2006, Rank et al. 2010). These network
statistics, captured by vector N(y), are: counts of ties, of reciprocated ties, of 2-stars (in-stars, out-stars,
and mixed stars), and of transitive triadic configurations in the inter-team communication network. (We
did consider including additional network statistics such as cyclical transitive triads, three-stars, and k-
triangles, but we ended up excluding them from our model because their coefficients were not significant.
Nonetheless, our results are robust to the inclusions of such additional network statistics.) Second, we
include ten node-level parameters, captured in vector P(y,w), corresponding to the effects of the five
(continuous) control variables, which are defined for both the acquirer and the provider of each
interacting dyad. Finally, we include seven dyad-level parameters, captured in vector L(y,w,x),
corresponding to four dyad-level effects and three interaction effects between some of them. The four
dyad-level parameter correspond to Lsame_subsystem, which captures whether a dyad is between teams that
belong to the same subsystem organizational group1; Lredundancy, which captures the counts of third parties
responsible for the information redundancy effect; Lcontrol, which captures the counts of third parties
responsible for social control effect; and Linterface, which captures the number of technical interfaces
responsible for the baseline effect of task interdependence.
1 Technically speaking, Lsame_subsystem is a categorical node-level effect that defines a dyad-level parameter
(Hunter et al. 2008).
25
We consider binary technical interfaces to test H1 to H3 and then consider trichotomous technical
interfaces to test H4 and H5. In the appendix we discuss how each of the parameters included in our
model are calculated. Again, the coefficients in our model are estimated by fitting the probabilistic model
to the observed inter-team communication network of design teams so that a significantly different than
zero coefficient indicates that the corresponding attribute is present in the observed network to a greater
or lesser extent than expected by chance, given other parameter values (Robins et al. 2007a).
In the case of binary technical interfaces, our model includes three interaction terms of interest.
The first interaction term ( ) takes into account the moderating effect that
formal organizational structures may have on the effect of technical interfaces in predicting inter-team
communication. Because Sosa et al. (2004) found this effect to be significant we control for it in our
model. The last two interaction terms, ( ) and ( ), test our
hypotheses concerning the moderating effects of common third parties. A negative and significant
coefficient indicates that the positive association of technical interfaces and inter-
team communications is reduced for dyads with higher values of information redundancy (H2). On the
other hand, a positive and significant coefficient indicates that the positive association
of technical interfaces and inter-team communications is increased for dyads with higher values of social
control (H3).
Finally, when considering trichotomous technical interfaces to test H4 and H5, then the model
specified in Equation 1 is augmented by including two (instead of one) dyad-level terms corresponding to
weak and strong technical interfaces and six (instead of three) interaction terms because each of the three
interaction terms mentioned above is estimated for both weak and strong interfaces.
5. Results
Tables 2 and 3 present ERGMs models estimating the effects of dyadic and node-level variables on the
probability of observing a directed technical communication tie between two design teams in the project.
The estimates were obtained using the Markov Chain Monte Carlo Maximum Likelihood Estimation
26
(MCMCMLE) procedure implemented in PNet (Wang et al. 2006). We also tested the robustness of our
results using the maximum pseudo-likelihood estimation procedure developed by Wasserman and
Pattison (1996) and Robins et al. (1999). Although this last procedure only yields approximate estimates
of the parameters of our models, the results (available from the authors upon request) are fully consistent
with the MCMCMLE results presented here. We examine the goodness-of-fit of our ERGMs by
comparing ―graph statistics from the observed network against a sample from the simulated graph
distribution‖ (Robins et al. 2009: 112). We compared two statistics of the observed communication
network (the in- and out- degree distributions and the in- and out- global clustering coefficients) with the
values of these statistics for a sample of directed networks simulated according to the fitted ERGMs
(Hunter et al. 2008, Robins et al. 2009). Our models provide a good fit to our data, as indicated not only
by the below 0.1 t-statistic of all the estimated parameters but also by the below 0.1 t-statistic of
additional network-level parameters such as the standard deviation of the in- and out- degree distributions
and the in- and out- global clustering coefficients. Similar to Robins et al. (2009), we use t-ratios to assess
whether the observed statistic is extreme compared to the distribution of a sample of simulated graphs
defined by our model.
We estimate two sets of ERGMs to test our hypotheses. The first set, reported in Table 2, shows
results of estimating the effect of the presence or absence of a technical interface on the probability of
observing an inter-team communication tie, as well as the moderating effects of information redundancy
and social control on the effect of technical interfaces (Hypotheses 1 to 3). The second set, reported in
Table 3, estimates whether the moderating effect of these two variables vary with the strength of the
technical interface. In all cases, the models include six endogenous network effects typically included in
ERGMs (Robins et al. 2007a), the node-level controls, and the main effects of same subsystem
membership, information redundancy, and social control.
An examination of the results reported in Table 2 indicates that all typical network statistics are
significant predictors of inter-team technical communication, confirming the strong endogenous effects
that are commonly observed in social networks (Contractor et al. 2006). Among the node-level controls,
27
only the size of the provider team and the degree of both the acquirer and the provider have significant
positive effects on the probability of observing an inter-team communication tie. The effect of team size
for the provider is consistent with the idea that the resources available to a team make it more likely to
respond to requests for technical information from other teams in the project, whereas the significant
degree effects suggest that the probability of observing communication ties is significantly affected by the
overall engagement in the communication network of both the acquirer and the provider teams
Model 1 shows that holding other variables constant, the probability of observing a directed
communication tie between two design teams is significantly higher in the presence of a technical
interface ( = 2.771, p < 0.001), offering strong support for our baseline hypothesis (H1). Hence,
the probability of observing a directed communication tie between two teams is 16.0 times higher if their
two components are connected by a directed technical interface (e2.771
= 16.0), confirming that the
presence of a technical interface is a strong predictor of whether or not two design teams communicate on
technical matters. The effect of the formal organizational structure is also noticeable: two teams whose
components were part of the same subsystem (and hence of the same formal organizational grouping)
were 3.24 more likely to communicate than those whose components belonged to different groupings
(e1.176
= 3.24), independently of whether or not their components shared a technical interface. These
results are fairly consistent across models in Table 2 and confirm the idea that both formal organizational
structures (i.e., the arrangement of the components into subsystems and the mirroring of such subsystems
into organizational groups) and the presence of directed technical interfaces are salient determinants of
inter-team communication. More importantly, the latter is indeed the strongest exogenous predictor of
inter-team communication, which highlights the importance of understanding what moderates such a
strong effect.
Of particular interest is the moderating effect that organizational boundaries have on the
relationship of task interdependence and inter-team communication (Sosa et al. 2004). This is examined
in Model 2. This model includes a significant negative coefficient for the interaction term
28
( ), which suggests that that the effect of a technical interface on the
likelihood of observing an inter-team communication tie is lower within boundaries than across
boundaries. In other words, the predictive power of technical interfaces gets diminished within subsystem
boundaries where other forces associated with the formal organizational structure play an important role
triggering inter-team technical communication. Such forces due to the formal organizational structure are
absent across subsystem boundaries where the association of technical interfaces and inter-team
communication is 167.7% stronger ( ).2 This is consistent with the fact that teams
that belong to the same subsystem are much more likely to communicate in the absence of a technical
interface than those from different subsystems do (Lessard and Zaheer 1996). Within subsystems, 20% of
the potential communication ties between teams whose components do not share an interface are actually
realized; the corresponding figure for teams across subsystems is only 2%.
The full model (Model 4) tests how information redundancy and social control moderate the
effect of technical interfaces in predicting the occurrence of technical communication between design
teams, holding constant all other factors in the model. This model shows a negative and significant
interaction effect of information redundancy and technical interface (
), which is line with H2. The higher the number of third parties falling in two-step
indirect communication paths going from the potential provider j to the potential acquirer i, the weaker
the effect of task interdependence as a determinant of inter-team communication. The moderating effect
of social control is positive but only marginally significant ( 0),
providing only marginal support for H3. As we will see in Table 3, however, the effect of social control
on the behavior of the providers is contingent on the strength of the interface. The sizes of the two
2 To estimate the size of this interaction effect we must calculate how much bigger the effect size of
technical interfaces across subsystem group boundaries (Lsame_subsystem_ij = 0) is versus the effect size of technical
interface within boundaries (Lsame_subsystem_ij = 1).
( )
( )
(
)
( )
( )
( )
.
29
interaction effects are noticeable: One-standard-deviation increase in information redundancy reduces the
association between technical interfaces and inter-team communications by 33.8% (
) while one-standard-deviation increase in pure social control increases the chances of a technical
interface predicting inter-team communication by 35.1% ( ).3 To check the
robustness of our results, we also estimated a full model (not shown) with an alternative measure of
overall social control that includes also third parties counted as contributing to information redundancy—
that is, a count that includes all third parties to a dyad in the communication network. The results for the
interaction terms are similar to the ones reported here obtained with the measure of pure social control.
Table 3 examines whether the moderating effects of information redundancy and social control
on the predicting effect of a technical interface are contingent on the strength of that interface (H4 and
H5). To do so, we distinguish between weak (below mean strength) and strong (above mean strength)
interfaces, with the lack of interface being the omitted category. For consistency, we follow the same
structure as in Table 2.
Model 1 shows that the likelihood of observing an inter-team communication tie is significantly
higher in the presence of a strong interfaces than in weak interfaces (3.143 vs. 2.575, p < .05), yet such
difference is not significant in the full model. Model 2 shows no significant difference in the negative
coefficients for the interaction between technical interface (weak and strong) and same subsystem (-1.088
vs. -1.066, n.s.), which is also true in the full model.
The full model (Model 4) tests H4 and H5. The pattern of results for the moderating effect of
information redundancy seems to provide support for H4, since the coefficient measuring the moderating
effect of information redundancy in strong interfaces is less negative (and non-significant) than the
3 To estimate the size of the interaction effect we must calculate how much smaller the
effect size of technical interfaces is when considering a one-standard-deviation increase in information redundancy.
( )
( ) ( )
( )
( )
( )
( ) ( )
( )
.
We estimate the size of the interaction effect in a similar manner.
30
coefficient for such a moderating effect in weak interfaces, suggesting that the overall effect presented in
Table 2 would be mainly caused by acquirers’ neglect to seek direct information on weak interfaces.
However, the difference between the two coefficients is not statistically significant (p < 0.366), failing to
provide strong support for H4. As we discuss in the final section of the paper, this ―non-finding‖ has
important practical implications, because it suggest that managers could not expect that the effects of
information redundancy theorized in this paper would be limited to weak interfaces.
Consistent with H5, the moderating effect of social control is only positive and significant for
strong interfaces. In this case, the coefficient for the interaction effect of social control with strong
interfaces is significantly greater than the corresponding coefficient for weak interfaces (0.010 vs. 0.330,
p < .01). This result suggests that the social control effects of common third parties are only apparent for
strong interfaces.
6. Discussion and Conclusions
The role of informal communication among the teams designing the components of a complex product is
crucial for product and process performance (Allen 1977, Brown and Eisenhardt 1995). Hence,
understanding the determinants of inter-team communication is critical for product development
organizations (Henderson and Clark 1990, Nadler and Tushman 1997, Sosa et al 2004, Gokpinar et al.
2010). Our findings confirm that task interdependence is a strong exogenous determinant of the
communication network among design teams, but we also show that this effect can be moderated by
factors endogenous to the communication network—namely, by the presence of common third parties to
the teams in this network.
The moderating effect of common third parties on the effectiveness of task interdependence in
predicting technical communication between two interdependent teams is contingent upon the position of
these third parties, the role played by each of the two teams, and the strength of the technical
interdependence between the two teams. Task interdependence is less likely to predict inter-team
communication to the extent that the acquirer team may receive information generated by the provider
31
team indirectly through common third parties. Task interdependence is more likely to predict inter-team
communication if the provider team shares common third parties with the acquirer team, albeit in this
case the moderating effect of the common third parties is only apparent for strong interdependencies.
Our paper makes several contributions to the understanding of how the structure of informal
inter-team communication networks—in particular, the presence of common third parties in the
communication network—can affect the behavior of design teams. First, we show that distinguishing
between the roles a team may play in a given task interdependence is crucial to revealing the effects of the
communication network structure on that team’s behavior. Specifically, we find that the presence of
common third parties to a technical interface can make teams more diligent providers but more lax
acquirers of information about this interface. This in turn moderates significantly the relationship between
task interdependence and inter-team communication. Hence, an analysis that fails to distinguish between
the two roles performed by teams in a design project may easily arrive at misleading conclusions. This
distinction between acquirer and provider roles in knowledge exchange networks is consistent with recent
analyses of the effects of network structure on outcomes. For example, Gargiulo et al. (2009) show that
distinguishing between acquirer and provider roles is crucial to understanding the effects of network
structure on the performance of members of knowledge exchange networks in the banking industry.
Second, our findings advance the understanding of the impact of information redundancy on the
behavior of actors in information exchange networks. Structural holes theory views information
redundancy as a cost associated with dense networks (Burt 1992). Because actors in such networks are
more likely to receive the same information through multiple channels, any additional channel is likely to
carry ―redundant‖ information, which diminishes its value. Sparse networks can connect actors to diverse
(i.e., non-redundant) sources of information and so are deemed to be more efficient. In contrast,
communication theory (Shannon 1948) views redundancy as a way to ensure an adequate understanding
of a message because redundant information reduces the ambiguity inherent in a single signal transmitting
that message. A similar principle holds in the design of complex engineering systems (Billinton and Allan
1992), where redundancy (understood as the duplication of critical components) is employed to avoid
32
systemic failures caused by malfunctioning of a critical component. Whereas these latter two approaches
focus on the value of the information for the potential receiver, we address the effect of information
redundancy on the acquiring behavior of the receiver. Our results suggest that, as the probability of
receiving indirect information about task interdependences increases, teams are less likely to seek direct
information about those interdependences. Although such neglect may be viewed as an efficient way to
economize on information search costs, it may also result in a biased representation of the decisions made
by the team responsible for this interdependence, which in turn could result in complications later in the
design process. We revisit this possibility later in the section when addressing the practical implications
of our findings.
Third, our results are consistent with the well-studied social control properties of triadic structures
(Coleman 1988, 1990, Gargiulo 1993): The presence of common third parties in the communication
network compels the provider team to be responsive to acquirer requests for technical information on
their technical interdependence. This is also consistent with previous studies that highlight the role of
common third parties in facilitating knowledge transfer (Reagans and McEvily 2003, Gargiulo et al.
2009) and collaboration in technical organizations (Allen 1977, Ahuja 2000, Obstfeld 2005). The
agreement of our findings with previous studies of knowledge transfer adds legitimacy to the results
reported here. At the same time, our results suggest that, at least in some cases, the availability of indirect
information might reduce the tendency towards network closure in social networks (Granovetter 1973).
Our study is based on comprehensive fieldwork that enabled us to collect rich quantitative and
qualitative data on the technical interfaces between components of an extremely complex product as well
as on the technical communication network among teams responsible for designing the components of
that product. The independent measurement of task interdependence between teams—based on the
observed technical interfaces between their respective components—provides a unique setting to
investigate how the structure of the communication network that emerges to address those technical
interfaces may itself moderate the effect of task interdependence in predicting inter-team communication.
Despite these advantages, the study has three limitations that result from the nature of our data.
33
First, our data on communication among teams captures only those exchanges that were explicitly
related to technical matters. Hence, our communication network may fail to capture informal social
interactions between members of different teams—that is, a social network that might exist in addition to
the observed technical communication network. However, we remark that unobserved social exchanges
pose a problem for our study only if they are largely unrelated to the observed exchanges of technical
communication. Such a possibility is not consistent with our qualitative observations of the
communication patterns between teams, which suggest that it was rare for social exchanges to occur
between teams that did not also communicate for technical reasons. This dynamic is consistent with the
idea that teams are preoccupied with their work and hence that social interactions in organized settings
will usually stem from encounters triggered by exogenous factors (in this case, interdependencies caused
by shared interfaces). In sum, these considerations suggest that the lack of information on purely social
exchanges between teams is unlikely to have biased our results. Also related to our measure of technical
communication, our data does not allow us to make fine-grain distinctions among various types of
technical communication. Unfortunately, given the complexity of the project studied, collecting the inter-
team communication data at a more granular level was impossible in our research site. Nonetheless, it is
important to acknowledge that distinguishing technical communication at a more granular level could
have allowed us to test further whether there are some types of technical flows more likely to be
responsible for the moderating effects of information redundancy and social control observed in our study.
Second, our measure of inter-team technical communication is based on the acquirer team’s
acknowledgment of having received technical information from the provider team. Strictly speaking,
failure to observe communication between the teams sharing a technical interface might have resulted
from failure of the acquirer team to request information, of the provider team to supply information, or
both. By the same token, observed communication might have resulted from proactive behavior by the
provider that is acknowledged by the acquirer, even in the absence of a specific request. Although our
data cannot conclusively discriminate among these situations, existing research on exchange of technical
information supports our interpretation, which assumes that communication is initiated by a request for
34
information by the acquirer team. Because teams are busy focusing on their own component, they are not
likely to proactively provide information about those components unless requested to do so. This
conjecture is consistent with a common observation in the product design literature that information
exchanges are typically triggered by a request from the acquirer team (Allen 1977, McCord and Eppinger
1993, Morelli et al. 1995, Terwiesch et al. 2002).
Third, estimating the likelihood of observing a communication tie as a function of parameters that
are partly endogenous to this very communication network poses significant methodological hurdles. We
tackled this difficulty by using exponential random graphs models (ERGMs) and resorting to a Markov
chain Monte Carlo maximum likelihood estimation procedure (Robins et al. 2007a). Although
computationally intensive and hence difficult to implement in estimations of complex models such as the
ones in this paper, MCMCMLE procedures are usually preferable to more traditional pseudo-likelihood
estimations of the regression coefficients (Snijders et al. 2006, Hunter et al. 2008, Robins et al. 2009). Yet,
regardless of the estimation procedure, ERGMs can only show that an observed association between the
presence of a tie yij between two actors i and j and a given network configuration (i.e., presence of
common third parties) is significantly different from randomness and independent from other exogenous
and endogenous factors that might also affect the presence of such a tie.
Our interpretation of the coefficients of interest is that the presence of common third parties
influences the behavior of interdependent teams. To assess the plausibility of this interpretation, it is
worth considering alternative behavioral sequences that might have resulted in a similar pattern of
common third parties. Regarding information redundancy, we argued that the presence of certain type of
common third parties makes the acquirer less likely to spend time and energy in seeking direct
information from the provider. Yet, acquirers might have also sought indirect sources of information if
they preferred to avoid (or have failed to establish) direct contact with the provider. In these cases, the
acquirer might have purposively approached the third party to obtain information on the provider’s
component. This, however, does not reduce the importance of common third parties as transmitters of
indirect information, nor the potential consequences for the accuracy of the information flowing in the
35
design project. Regarding the social control mechanism, ties with common third parties might have been
prompted by the direct tie between the two teams, reversing the sequence proposed in the social control
mechanism. While this sequence of events might be possible in friendship networks, it is unlikely to
happen in a technical communication network like the one studied in this paper, where ties are largely
driven by exogenous interdependences. Indeed, research using longitudinal network data has shown that
common third parties makes actors more likely to build a direct tie in the future (e.g., Gulati and Gargiulo
1999). In other words, while alternative interpretations of the observed association between common third
parties and dyadic ties are possible from a pure statistical viewpoint, they are either consistent with the
one privileged in this paper or unlikely to have occurred in this context.
Our findings show how the structure of the inter-team technical communication network might
affect the behavior of design teams involved in a complex new product development project, but we
cannot say anything regarding the desirability or undesirability of such behaviors. We can suggest that,
given the complexity of the technical interfaces in the PW4098 engine, indirect information about
interfaces might not have been sufficient to achieve an adequate understanding of how design decisions in
one component might affect other connected components. Yet, one might also argue that spending scarce
team resources in seeking direct information to corroborate already available indirect information might
have been an inefficient use of the team resources. Anecdotal evidence from our research site suggests
that some interfaces that were not directly attended did cause significant rework in subsequent stages of
the process, but we lack systematic data in this respect. These qualitative observations, however, are fully
consistent with the results of Gokpinar et al. (2010), which found a significant association between lack of
communication about technical interfaces and quality issues in the automobile industry. However, we
cannot say whether or not teams were ―talking too little‖ with other teams, as it is indeed possible that
omissions to establish a direct communication between provider and acquirer could have been sometimes
an efficient way to use scarce team resources. Neither can we say that they were ―talking too much.‖
While our data does show instances of communication in the absence of technical interfaces, which are
disproportionally present within group boundaries in the formal organizational structure, such
36
communication cannot be dismissed as superfluous, because it might have addressed unobserved
technical interfaces that were uncovered during the design process or other types of task-related
interdependencies not directly traceable to physical interfaces between components. What our results do
show is that although the technical interface network largely determines the presence or absence of
communication between design teams, deviations from this overall tendency are systematically associated
with specific configurations in the structure of the inter-team communication network.
Despite their limitations, our findings have important practical implications for managers seeking
to improve the coordination between design teams in complex product development projects. While our
study confirms the salience of task interdependence in prompting informal exchanges of technical
communication between teams, they also identify situations that can significantly alter this association.
This is important because managers (and engineering scholars) often assume that technical
interdependence should automatically trigger the communication necessary to coordinate design tasks
(Cataldo et al. 2006, Olson et al. 2009). Managers are aware of the unintended effect of formal
organizational boundaries in hindering communication between actors that sit across such boundaries.
However, the possibility that informal communication networks—often touted as a remedy to the failure
of formal organization—would also have unintended consequences that may affect the likelihood that
interdependence will lead to communication has not been considered. Yet, our analysis suggests that such
informal networks can alter the expected association between task interdependence and technical
communication. In particular, the possibility that indirect communication paths can substitute for direct
communication between interdependent teams should raise the attention of managers in charge of product
development projects, challenging the expectation that interdependence alone might suffice to prompt
teams to communicate.
While indirect communication may be viewed as an efficient way to coordinate tasks across a
relatively unimportant technical interface, it might lead also to design inadequacies that could affect the
performance or durability of the affected components and subsystems, causing significant warranty or
service expenses over the life of the product. In this sense, our inability to find significant differences for
37
the negative moderating effect of information redundancy across strong and weak interfaces makes the
risk of inadequate communication in product design projects even more apparent. Indeed, inadequate
exchanges of technical information could affect the performance or durability of the affected components
and subsystems, causing significant warranty or service expenses over the life of the product. For
example, if a critical component of an aircraft engine (e.g., a turbine airfoil) fails to reach its life
expectancy, the result is additional unplanned engine removals for maintenance. For an engine like the
PW4098, a single such removal could cost the customer as much as $150,000, in addition to the loss of
revenues associated with a grounded plane.
To conclude we refer to Simon (1996) who suggested that a complex system is difficult to
understand because the behavior of the whole depends in nontrivial ways on how its elements interact. In
studying the determinants of inter-team communications when designing a complex system, we have
learned that common third-party teams have a significant moderating influence on the predictive power of
the patterns of task interdependence. Moreover, assessing the extent of this influence is not trivial because
doing so depends on the role—as acquirer or provider of information—that the focal team plays during
the design process. We have shown that our results have important implications for both theory and
practice. They also raise a number of interesting questions. How does the presence of common third
parties to a focal technical interface influence the ability of interdependent teams to discover and attend to
other relevant interfaces that may also require inter-team communication? How do component
connectivity and team network structure relate to such important design decisions as component
outsourcing and component redesign? With the upraising of open innovation projects, how does the
relationship between task interdependence and inter-team communication change when development
projects are carried out across firm boundaries? By combining insights from operations management and
social networks, in this paper we have described a path that can lead to answering these important
questions.
38
7. References
Ahuja, G. 2000. Collaboration networks, structural holes, and innovation: A longitudinal study. Admin. Sci.
Quaterly. 45: 425–455.
Allen, T.J. 1977. Managing the Flow of Technology. Cambridge, Mass.: MIT Press.
Allen, T.J. 2004. Communication networks in R&D laboratories. Chapter 26 in The Human Side of Managing
Technological Innovation. Edited by Ralph Katz. Oxford University Press.
Ancona, D. and D. Caldwell. 1992. Bridging the boundary: External activity and performance in organizational
teams. Admin. Sci. Quart. 37, 634-665.
Baldwin, C.Y., and K.B. Clark. 2000. Design rules: Volume 1: The power of modularity. MIT Press.
Barker, J. R. 1993. Tightening the iron cage. Concertive control in self-managing teams. Admin. Sci. Quarterly,
38:408-437.
Billinton, R. and R.N. Allan. 1992. Reliability Evaluation of Engineering Systems: Concepts and Techniques.
Springer. Second Edition.
Brown, S.L., K.M. Eisenhardt. 1995. Product development: Past research, present findings, and future directions.
Academy of Manag. Review. 20, 343-378.
Burt, R.S. 1992. Structural Holes, The Social Structure of Competition. Harvard University Press.
Burt, R.S. 2004. Structural holes and new ideas. American Journal of Sociology 110: 349–399.
Burt, R.S. 2005. Brokerage and Closure: An Introduction to Social Capital. Oxford University Press.
Cataldo, M., P. Wagstrom, J. D. Herbsleb, K. M. Carley. 2006. Identification of Coordination Requirements:
Implications for the Design of Collaboration and Awareness Tools. Proceedings of the ACM Conference on
Computer-Supported Cooperative Work, Banff, Alberta, 353-362.
Coleman, J.S. 1988. Social capital in the creation of human capital. American Journal of Sociology, 94: 95-120.
Coleman, J.S. 1990, Foundations of social theory. Cambridge, Mass.: Belknap Press.
Colfer, L. and C. Baldwin. 2010. The mirroring hypothesis: Theory, evidence, and Exceptions. Harvard Business
School Working Paper, 10-058.
Contractor, N.S., S. Wasserman, and K. Faust. 2006. Testing multitheoretical, multilevel hypotheses about
organizational networks: An analytic framework and empirical example. Academy of Mgmt. Review 31(3): 681-703.
Galbraith, J. R. 1973. Designing Complex Organizations. Reading, Mass.: Addison-Wesley Publishing.
Gargiulo, M 1993. Two-step leverage: Managing constraint in organizational politics. Admin. Sci. Quart., 39:1-19.
Gargiulo, M., G. Ertug, and C. Galuni. 2009. The two faces of control: Network closure and individual performance
among knowledge workers. Admin. Sci. Quaterly 54: 299-333.
Gokpinar, B., W. Hopp, and S. Iravani. 2010. The impact of product architecture and organizational structure on
efficiency and quality of complex product development. Mgmt. Sci. 56(3): 468-484.
Granovetter, M.S. 1985. Economic action and social structure: The problem of embeddedness. American Journal of
Sociology 91: 481–510.
Gulati, R. 2010. Reorganize for Resilience: Putting Customers at the Center of Your Business. Boston: Harvard
Business School Press.
Gulati, R., and M. Gargiulo. 1999. Where do interorganizational networks come from? American Journal of
Sociology 104 1439-1493
Henderson, R., and K. Clark. 1990. Architectural innovation: The reconfiguration of existing product technologies
and the failure of established firms. Admin. Sci. Quart. 35(1): 9–30.
Hunter, D.R., S.M. Goodreau, and M.S. Handcock. 2008. Goodness of fit of social network models. Journal of the
39
American Statistical Association. 103(481): 248-258.
Lazega, E. 2001. The Collegial Phenomenon. The Social Mechanisms of Cooperation among Peers in a Corporate
Law Partnership. Oxford: Oxford University Press.
Lessard. D. R. and S. Zaheer. 1996. Breaking the silos: Distributed knowledge and strategic responses to volatile
exchange rates. Strategic Management Journal 17: 513-533.
Lomi, A., and P. Pattison. 2006. Manufacturing relations: An empirical study of the organization of production
across multiple networks. Org. Sci. 17(3) 313-332.
Marsden, P.V. 1990. Network data and measurement. Annual Review of Sociology. 16 435-463.
McCord, K. R. and Eppinger, S.D. 1993. Managing the Integration Problem in Concurrent Engineering. Sloan
School of Management Working Paper Number 3594. Massachusetts Institute of Technology.
Morelli, M.D., S.D. Eppinger, R.K. Gulati. 1995. Predicting technical communication in product development
organizations. IEEE Trans. Eng'g. Management 42(3): 215-222.
Morris, M., M.S. Handcock, and D.R. Hunter. 2008. Specification of Exponential-Family Random Graph Models:
Terms and computational aspects. Journal of Statistical Software, 24(4): 1-24.
Nadler, D. A. and M. Tushman. 1997. Competing by Design: The Power of Organizational Architecture. Oxford
University Press.
Obstfeld, D. 2005. Social networks, the Tertius iungens orientation, and involvement in innovation. Admin. Sci.
Quart. 50: 100–130.
Olson, J., J. Cagan, and K. Kotovsky. 2009. Unlocking organizational potential: A computational platform for
investigating structural interdependence in design. ASME Journal of Mechanical Design.
Rank, O.N., G. Robins, and P. Pattison. 2010. Structural logic of intraorganizational networks. Organization Science,
21(3): 745-764.
Raub, W. and J. Weesie 1990. Reputation and efficiency in social interactions: An example of network effects.
American Journal of Sociology, 96: 626-654.
Reagans, R., and B. McEvily. 2003. Network structure and knowledge transfer: The effects of cohesion and range.
Admin. Sci. Quart. 48: 240–267.
Reagans, R. and E. Zuckerman. 2001. Networks, diversity and performance: The social capital of R&D teams.
Organization Science, 12: 502-518.
Reagans, R., E. Zuckerman, and B. McEvily 2004 How to make the team: Social networks: Social networks vs.
demography as criteria for designing effective teams. Administrative Science Quarterly, 49: 101-133.
Robins, G.L., Elliott, P., & Pattison, P. 2001. Network models for social selection processes. Social Networks, 23:1-
30.
Robins, G.L., P. Pattison, Y. Kalish, D. Lusher. 2007a. An introduction to exponential random graph (p*) models
for social networks. Social Networks, 29: 173–191.
Robins, G.L., Pattison, P.E., and P. Wang. 2009. Closure, connectivity and degree distributions: Exponential random
graph (p*) models for directed social networks. Social Networks. 31(2): 105-117.
Robins, G., Pattison, P., and Wasserman, S. 1999. Logit models and logistic regressions for social networks, III.
Valued relations. Psychometrika. 64 371-394.
Robins, G.L. ,T.A.B. Snijders, P. Wang, M. Handcock and P. Pattison. 2007b. Recent developments in exponential
random graph (p*) models for social networks. Social Networks, 29: 192–215
Shannon, C.E. 1948. A mathematical theory of communication. Bell System Technical Journal, 27: 379-423, 623-
656.
Simon, H. A. 1996. The Sciences of the Artificial. 3rd
Edition, MIT Press, Cambridge, MA.
Smith, R.P., S.D. Eppinger.1997. Identifying controlling features of engineering design iteration.Manag. Sci. 43
276-293.
40
Rowles, C. M. 1999. System Integration Analysis of a Large Commercial Aircraft Engine. MIT Masters Thesis.
Snijders, T., P. Pattison, G. Robins, and M. Handcock. 2006. New specifications for exponential random graph
models. Sociology Methodology.
Sosa, M.E., S.D. Eppinger, M. Pich, D. McKendrick, S. Stout. 2002. Factors that influence technical communication
in distributed product development: an empirical study in the telecommunications industry. IEEE Trans. Eng'g.
Manag. 49(1) 45-58.
Sosa, M.E., S.D. Eppinger, and C.M. Rowles. 2004. The misalignment of product and organizational structures in
complex product development. Mgmt. Sci., 50(12): 1674–1689.
Sosa, M.E., S.D. Eppinger, and C.M. Rowles. 2007. A network approach to define modularity of components in
complex products. ASME Journal of Mechanical Design, 129(11): 1118-1129.
Terwiesch, C., C. H. Loch, A. De Meyer. 2002. Exchanging preliminary information in concurrent engineering:
Alternative coordination strategies. Org. Sci. 13(4): 402–419.
Thompson, J.D. 1967. Organizations in Action. New York: McGraw-Hill.
Tushman, M.1977. Special boundary roles in the innovation process. Admin. Sci. Quart. 22: 587-605.
Tushman, M and R. Katz. 1980. External communication and project performance: An investigation into the role of
gatekeeper. Manag. Sci. 26(11) 1071-1085.
Van den Bulte, C., R.K. Moenaert. 1998. The effects of R&D team co-location on communication patterns among
R&D, marketing, and manufacturing. Manag. Sci. 44 S1-S19.
Wang, P., Robins, G., Pattison, P., 2006. Pnet: A Program for the Simulation and Estimation of Exponential
Random Graph Models. University of Melbourne. http://www.sna.unimelb.edu.au/pnet/pnet.html.
Wasserman, S. and P. Pattison. 1996. Logit models and logistic regression for social networks: I. an introduction to
Markov graphs and p*. Psychometrika 61 401-425.
41
Table 1. Descriptive statistics and pairwise correlations of variables
N = 2,862 directed dyads. Correlations greater than |0.06| are significant at the .001 level. †These dyadic variables are calculated based on the network of 60 (design and functional) teams
Variables Mean S.D. 1 2 3 4 5 6 7 8 9 10 11 12 13 14
1 Inter-team
communication (yij) 0.15 0.35 1.00
2 Technical interfaces
(xij) 0.20 0.40 0.65 1.00
3 Component
redesign acquirer 0.49 0.33 0.00 -0.03 1.00
4 Component
redesign provider 0.49 0.33 -0.04 -0.06 -0.02 1.00
5 Component
complexity acquirer 41.59 65.58 0.03 0.13 -0.33 0.01 1.00
6 Component
complexity provider 41.59 65.58 0.12 0.17 0.01 -0.33 -0.02 1.00
7 Component
connectivity acquirer 13.07 6.37 0.18 0.23 -0.22 0.00 0.58 -0.01 1.00
8 Component
connectivity provider 13.07 6.37 0.17 0.25 0.00 -0.22 -0.01 0.58 -0.02 1.00
9 Team size acquirer 2.44 0.92 0.13 0.14 0.14 0.00 0.31 -0.01 0.53 -0.01 1.00
10 Team size provider 2.44 0.92 0.12 0.13 0.00 0.14 -0.01 0.31 -0.01 0.53 -0.02 1.00
11 Team in-degree
acquirer 9.89 7.11 0.30 0.17 0.03 0.00 0.07 0.00 0.56 -0.01 0.43 -0.01 1.00
12 Team degree
provider 13.17 7.07 0.16 0.17 0.00 -0.08 -0.01 0.39 -0.01 0.76 -0.01 0.55 -0.02 1.00
13 Same sub-system 0.11 0.32 0.49 0.46 -0.01 -0.01 0.03 0.03 0.05 0.05 0.01 0.01 0.02 0.03 1.00
14 Information
redundancy†
2.15 2.28 0.53 0.45 0.00 -0.07 0.00 0.23 0.30 0.32 0.23 0.20 0.58 0.30 0.36 1.00
15 Social control† 1.84 1.96 0.16 0.25 -0.13 -0.05 0.39 0.13 0.37 0.34 0.21 0.23 0.06 0.51 0.12 0.13
42
Table 2. Results of ERGM predicting inter-team communications with binary technical interfaces
Standard errors given in parentheses. *** < 0.01 ** < 0.05 * < 0.1 (two-tailed)
Variable Model 1 Model 2 Model 3 Model 4
Basic network effects
Baseline inter-team communication, ( ) -5.753*** (0.794) -5.886*** (0.833) -6.257*** (0.837) -6.092*** (0.805)
Reciprocity, ( ) 2.421*** (0.337) 2.428*** (0.326) 2.377*** (0.319) 2.379*** (0.330)
Two-star in, ( ) -0.160** (0.076) -0.158** (0.074) -0.165** (0.076) -0.172** (0.077)
Two-star out, ( ) -1.537*** (0.303) -1.494*** (0.303) -1.554*** (0.328) -1.541*** (0.303)
Two-star mixed, ( ) -0.255*** (0.049) -0.252*** (0.049) -0.255*** (0.055) -0.260*** (0.056)
Transitivity, ( ) 0.126*** (0.032) 0.120*** (0.032) 0.121*** (0.033) 0.124*** (0.033)
Node-level effects
Component redesign acquirer -0.354 (0.678) -0.358 (0.653) -0.359 (0.705) -0.283 (0.688)
Component redesign provider -0.250 (0.379) -0.292 (0.377) -0.294 (0.368) -0.310 (0.341)
Component complexity acquirer 0.005 (0.004) 0.005 (0.004) 0.005 (0.004) 0.005 (0.004)
Component complexity provider -0.001 (0.002) 0.000 (0.002) 0.000 (0.002) 0.000 (0.002)
Component connectivity acquirer 0.029 (0.053) 0.020 (0.052) 0.026 (0.054) 0.021 (0.058)
Component connectivity provider -0.007 (0.032) -0.009 (0.034) -0.008 (0.033) -0.011 (0.034)
Team size of acquirer -0.106 (0.266) -0.086 (0.273) -0.084 (0.277) -0.086 (0.288)
Team size of provider 0.314* (0.161) 0.334** (0.158) 0.332** (0.168) 0.328** (0.167)
Team in-degree of acquirer 1.408*** (0.248) 1.378*** (0.246) 1.425*** (0.269) 1.421*** (0.248)
Team degree of provider 0.218*** (0.055) 0.212*** (0.054) 0.22*** (0.057) 0.228*** (0.058)
Main effects
Same sub-system, ( ) 1.176*** (0.210) 1.829*** (0.316) 1.770*** (0.319) 1.729*** (0.321)
Info. Redundancy acquirer, ( ) -0.009 (0.064) -0.007 (0.065) 0.082 (0.079) 0.096 (0.080)
Social control pure, ( ) -0.144** (0.063) -0.120** (0.060) -0.136** (0.059) -0.221*** (0.081)
Presence of technical interface, ( ) 2.771*** (0.193) 3.047*** (0.237) 3.594*** (0.338) 3.334*** (0.368)
Interaction effects
-0.985*** (0.372) -0.874** (0.387) -0.843** (0.367)
-0.161** (0.073) -0.181** (0.074)
0.153* (0.092)
43
Table 3. Results of ERGM predicting inter-team communications with trichotomous technical interfaces
Standard errors given in parentheses. *** < 0.01 ** < 0.05 * < 0.1 (two-tailed)
Variable Model 1 Model 2 Model 3 Model 4
Baseline inter-team communication, ( ) -5.704*** (0.860) -5.825*** (0.858) -6.154*** (0.811) -6.144*** (0.849)
Reciprocity, ( ) 2.372*** (0.360) 2.389*** (0.335) 2.345*** (0.323) 2.369*** (0.325)
Two-star in, ( ) -0.157** (0.071) -0.153** (0.073) -0.159** (0.077) -0.179** (0.084)
Two-star out, ( ) -1.524*** (0.318) -1.512*** (0.311) -1.540*** (0.313) -1.532*** (0.327)
Two-star mixed, ( ) -0.252*** (0.052) -0.250*** (0.051) -0.252*** (0.049) -0.261*** (0.054)
Transitivity, ( ) 0.130*** (0.034) 0.122*** (0.032) 0.123*** (0.032) 0.132*** (0.033)
Component redesign acquirer -0.351 (0.682) -0.398 (0.656) -0.373 (0.701) -0.379 (0.658)
Component redesign provider -0.201 (0.376) -0.206 (0.355) -0.251 (0.373) -0.213 (0.402)
Component complexity acquirer 0.004 (0.004) 0.005 (0.004) 0.005 (0.004) 0.003 (0.004)
Component complexity provider 0.000 (0.002) 0.000 (0.002) 0.000 (0.002) 0.000 (0.002)
Component connectivity acquirer 0.023 (0.053) 0.012 (0.052) 0.015 (0.055) 0.009 (0.055)
Component connectivity provider -0.006 (0.033) -0.009 (0.034) -0.008 (0.035) -0.004 (0.035)
Team size of acquirer -0.097 (0.274) -0.082 (0.272) -0.089 (0.276) -0.045 (0.284)
Team size of provider 0.311* (0.161) 0.325** (0.153) 0.329** (0.168) 0.331** (0.164)
Team in-degree of acquirer 1.399*** (0.257) 1.395*** (0.254) 1.418*** (0.256) 1.424*** (0.267)
Team degree of provider 0.214*** (0.055) 0.207*** (0.053) 0.211*** (0.053) 0.222*** (0.060)
Same sub-system, ( ) 1.135*** (0.210) 1.830*** (0.305) 1.784*** (0.328) 1.775*** (0.332)
Info. Redundancy, ( ) -0.016 (0.066) -0.013 (0.071) 0.079 (0.078) 0.083 (0.076)
Social control pure, ( ) -0.136** (0.065) -0.117** (0.059) -0.126** (0.062) -0.208** (0.090)
WEAK interface, ( ) 2.575*** (0.214) 2.872*** (0.237) 3.575*** (0.364) 3.698*** (0.420)
STRONG interface, ( ) 3.143*** (0.265) 3.480*** (0.333) 3.855*** (0.448) 3.055*** (0.566)
-1.088** (0.423) -0.985** (0.435) -0.991** (0.428)
-1.066** (0.484) -0.975** (0.468) -0.970** (0.472)
-0.207*** (0.079) -0.230*** (0.083)
-0.111 (0.096) -0.105 (0.106)
0.010 (0.110)
0.330*** (0.118)
Appendix. Defining an Exponential Random Graph Model (ERGM) of our Inter-Team Communication Network
ERGMs differ from traditional statistical regression models (Contractor et al. 2006, Robins et al. 2007a).
In traditional models, the data typically consists of a set of (independent) observations measured on
various separate variables (one of which is the dependent variable and the other ones are the predictor
variables). Although these variables might be correlated, they are measured separately for each
observation. In a dataset like ours, the dependent variable corresponds to the presence or absence of a
directed communication tie between two design teams and the predictors are separately measured
attributes of the node (e.g., the size of the teams involved in the dyad), the dyad (e.g., whether the teams
design components of the same subsystem or not), and the tie (e.g., whether there is a directed technical
interface or not corresponding to the focal tie). Although some of the predictor variables are exogenous to
the inter-team communication network, ERGMs allow for the inclusion of predictors that are functions of
the structure of the same communication network being modeled. These predictors, called “network
statistics”, represent configurations of ties (for instance, pairs of teams with reciprocal ties or triplets of
teams that form transitive triads) that are expected to occur more often (or less often) than random in the
data. In that sense, ERGMs can be seen as autoregressive models, and this influences many aspects of the
models specification and estimation (Morris et al. 2008).
Considering the non-traditional nature of ERGMs, we carry out the statistical modeling in three
steps (see Robins et al. 2007a for a structured approach to build ERGMs of social networks). We build
random graphs of 54 nodes that try to predict the observed inter-team communication network. The
parameters used to build the graphs are network configurations that are hypothesized to be present in the
observed network. The coefficient parameters in the model are estimated to maximize the fit between the
random networks and the observed data.
Step 1: Formulate a baseline ERGM of the inter-team communication network
We build a probabilistic model of the 54-node network of design teams. This model considers inter-team
communication as a random variable. Hence, for i and j (which are distinct teams of the network of 54
design teams) we consider a random variable Yij where Yij = 1 if design team i (the acquirer) receives
technical information from team j (the provider), and where Yij = 0 otherwise. We specify yij as the
observed value of Yij and Y as the matrix of all Yij variables while y is the matrix of observed ties (i.e. the
binary adjacency matrix corresponding to the inter-tem technical communication network of 54 design
teams). Because it would be unrealistic to assume that design teams form an inter-team communication tie
independently of their other inter-team communication ties, this model assesses whether the presence of
observed communication ties is influenced by certain configurations of ties in the inter-team
communication network. Following previous work in modeling exponential random graphs for social
networks in technical organizations, we reduce our general model to a homogeneous Markov random
graph (Contractor et al. 2006, Lomi and Pattison 2006, Rank et al. 2010). Based on recent developments
in ERGMs (Robins et al. 2007b), we did consider including additional network statistics such as cyclical
transitive triads, three-stars, and k-triangles, we ended up excluding them from the base line model
because their coefficient were not significant. Nonetheless, our results are robust to the inclusions of such
additional network statistics. Hence, our first baseline ERGM takes the following form:
( ) (
) ( ( ) ( ) ( ) ( ) ( ) ( ))
(Eq. A1)
where Y is a random network with possible ties Yij, y is a realization of Y, and is a normalizing quantity
(which is difficult to calculate analytically in networks of this complexity). The terms L(y), M(y), S(y),
and T(y) denote, respectively, counts of ties, of reciprocated or mutual ties, of 2-stars (in-stars, out-stars,
and mixed stars), and of transitive triadic configurations in the inter-team communication network. These
are the basic “network statistics” in a typical organizational network. For instance, L(y) is a count of all
directed inter-team communication ties between teams in our network (423 ties) and M(y) is the count of
dyads with reciprocated ties in this network (141 dyads). These six network statistics correspond to the
six network-level effects included in vector N(y) of Equation 1 in the paper.
The relevance of the network configurations included in our base model is assessed by the
significance of the parameters , , , and , which are to be estimated from the data. A negative and
significant parameter indicates that there is a low probability of a communication tie existing between
two randomly chosen teams, controlling for other network configurations. Similar interpretations can be
made for other parameters that control for the tendency of teams to reciprocate ties (), receive or send
ties (), or to form certain types of triads ().
Step 2: Extend ERGM with node-level parameters
We extend the baseline model by including both exogenous and endogenous node-level attributes that can
influence the occurrence of the observed yij (cf. Robins et al. 2001). Consistent with Hunter et al. (2008),
our node-level effects fall into two main categories: continuous nodal attribute and categorical nodal
attribute. First we include continuous nodal attributes, also called nodal-factor effects. These nodal-factor
effects capture the extent to which higher levels of a nodal attribute (e.g. component redesign or team
size) increases the probability of inter-team communication (e.g. bigger provider teams tend to be more
responsive to requests). To capture the effect of continuous node-level attributes such as “team size” we
include two statistics (one for the acquirer and another one for the provider of each dyad). We do this to
be consistent with our basic premise that the role of the team as either acquirer or provider influences
their observed inter-team communication patterns. In the case of team size, these two statistics correspond
to the sum of the size of the acquirer team (and the sum of the size of the provider) of all interacting dyads.
We also test a model specification that, instead, includes (mutual) sum of the continuous node-level
attributes of all interacting dyads. We obtain similar results with such an alternative model specification.
In addition to exogenous nodal-factor effects, we also include endogenous nodal-factor effects to control
for the total number of teams from which the acquirer team i receives technical information from as well
as the total number of other teams with whom provider j communicates (either to seek or provide
technical information).
Our model includes ten nodal-factor effects (two for each of the five node-level continuous
control variables). The corresponding parameters to these ten effects are included in the vector P(y,w) of
Equation 1 in the paper. These parameters are function of both the observed communication matrix, y,
and a covariate matrix, w, which is a collection of 54x1 vectors containing the nodal-factor attributes of
interest. Hence, after including nodal-factor effects, the baseline model becomes:
( )
(
) ( ( ) ( ) ( ) ( ) ( ) ( )
) (Eq. A2)
Step 3: Extend ERGM with dyad-level attributes
Finally, we include four dyad-level effects and three interaction terms between some of them. The
corresponding parameters are contained in the L(y,w,x) of Equation 1 in the paper. The first dyad-level
effect is the result of a categorical nodal attribute defined by a team’s membership to one of the eight
subsystem organizational groups. Such a categorical node-level attribute allows us to define a uniform
homophily effect (Hunter et al. 2008), which captures the tendency for an inter-team communication to be
more likely to occur between teams that belong to the same subsystem group. Because such a categorical
nodal effect results in a dyad-level attribute we include it in this third step of our approach. The statistic
that captures this effect (Lsame_subsystem) counts the directed technical communications that occur between
teams that belong to the same subsystem group.
Next, we extend our model with endogenous dyadic parameters (i.e. the count of common-third
parties) and with an exogenous dyadic attribute (i.e. technical interfaces). To do so, we first include
dyadic parameters associated with our two main predictor variables Lredundancy(y) and Lcontrol(y), which
counts the total number of these two types of common third parties for all the pairs of interacting teams in
our inter-team communication network. Then, we define the covariate matrix X, which captures the
technical interfaces corresponding to the 54 design teams. For easy of exposition we consider here the
case in which Xij is binary and thus each element xij records the presence or absence of a technical
interface between the components designed by teams i and j. In order to test whether the presence of a
technical interface is associated with the occurrence of the corresponding inter-team communication (as
hypothesized by H1), our model must include a network statistic that captures the simultaneous
occurrence of Yij and Xij. Such a network statistic is ( ) ∑( ) = 349 inter-team
communications “matched” by a technical interface. Hence, our model becomes:
( )
(
) ( ( ) ( ) ( ) ( ) ( ) ( )
)
) (Eq. A3)
A positive and significant coefficient will indicate that task interdependence (as
measured by the presence of a technical interface) there is a significant determinant of inter-team
communication (in line with our baseline hypothesis H1). Yet, we are interested in testing how such a
baseline effect varies with the levels of information redundancy and social control of the acquirer and
provider teams involved in each interdependent dyad. Hence, we include interactions terms between
Linterface and our key predictor variables (cf. Robins et al. 2001). As a result, the full model with binary
technical interfaces, which correspond to Equation 1 in the paper, is:
( )
(
) ( ( ) ( ) ( ) ( ) ( ) ( )
( ) (
) ( )) (Eqn. A4)
