How Do Communication Costs Affect Scientific Collaboration? Exploring the Effect of Bitnet

August 2005

Ajay Agrawal and Avi Goldfarb * Abstract We examine the effect of connecting to Bitnet on research collaboration between US universities. Furthermore, we estimate the degree to which the quality of and distance between potential collaborators mediate any such effect. In other words, are benefits that accrue from Bitnet adoption spread uniformly across all adopters? We use publication data from seven engineering journals and Bitnet connection data to address these questions. Overall, we find that Bitnet adoption results in approximately an 85% increase in the likelihood of collaboration between pairs of institutions after they have both connected. In addition, medium-ranked universities seem to benefit most, largely by increasing their collaboration with top-ranked schools. Moreover, the effect is largest for co-located institutions. (JEL O33, R11, Z13) Keywords: collaboration, Internet, knowledge flows, communication technology, productivity

* University of Toronto (both authors). Corresponding author: ajay.agrawal@rotman.utoronto.ca. We thank seminar participants at the Canadian Economics Association meetings and the University of Toronto for useful comments. We also thank Raghav Misra, Swapnil Kotecha, Cara Saunders, and Alex Oettl, who provided excellent research assistance. Errors and omissions are our own. This research was funded by the Social Sciences and Humanities Research Council of Canada (Grant Nos. 410-2004-1770 and 538-02-1013). Their support is gratefully acknowledged.

1

1. Introduction

The objective of this research is to further our understanding of scientific collaboration, an important

component of the knowledge production function. Collaborations are important for a number of reasons.

First, under certain conditions, inter-organization collaboration is more efficient than single-organization

research since collaborators are able to draw upon a wider set of expertise, specialized equipment, and

other resources. Cockburn and Henderson (1998), who measure co-authorships across institutions, report

results suggesting that, at least in the pharmaceutical industry, firms that are more open to collaboration

with universities are more productive in terms of innovation.

Second, collaboration may facilitate greater knowledge flows, and knowledge flows, particularly

knowledge spillovers, are central to economic growth as characterized in endogenous growth models

(Romer, 1986 and 1990). For example, Singh (2005) provides evidence suggesting that network ties

between inventors, as inferred from their past collaborations, are important determinants of knowledge

diffusion patterns.

Third, there is a growing empirical literature on the relationship between geography and scientific

knowledge flows (Jaffe et al, 1993; Audrestch and Feldman, 1996; Zucker et al, 1998; Agrawal et al,

2003; Thompson and Fox-Kean, 2005). Many of these studies have challenged the notion of the “death of

distance” that some suggest might result from advances in communication technologies (Cairncross,

1997). Despite the significant work on this subject, the degree to which lower communication costs

change the nature of collaboration along the geographic dimension remains an open question.

Finally, as a result of various policy initiatives, significant public expenditures have been incurred

with precisely the objective of fostering inter-organization scientific collaboration. One such expenditure

was the creation of the Internet. However, the behavior of researchers in terms of collaboration patterns in

response to such initiatives remains largely unknown. We aim to address these issues here.

In particular, we address this question: To what extent does connecting to Bitnet increase the

propensity of researchers to collaborate with others who are also connected? We find that connecting to

Bitnet, and thus lowering the cost of communication, increases collaboration significantly. Overall, the

2

“Bitnet effect” results in approximately an 85% increase in the likelihood of collaboration between pairs

of institutions after they have both connected.

Next, we examine how the Bitnet effect is mediated by university quality, indicated by the

institution’s overall emphasis on research as measured by the Carnegie Foundation. Our results suggest

that medium-ranked universities benefit more than high-ranked and low-ranked universities from

connecting to Bitnet. We discuss the implications of this finding in our conclusion.

Finally, we turn our attention to the question of whether the Internet is a complement to or a

substitute for face-to-face interactions (Gaspar and Glaeser, 1998). With respect to agglomeration, the

former will encourage greater urbanization, the latter greater sprawl. Our findings suggest that the Bitnet

effect is greatest between nearby institutions, suggesting that connecting to the network is a complement

to face-to-face interactions. Again, we discuss the implications of this finding in the conclusion.

Although there is a significant literature concerning the economics of scientific collaboration

(Crane, 1969; Beaver and Rosen, 1978, 1979; Barnett, Ault, and Kaserman, 1988; Cockburn and

Henderson, 1998; Mairesse and Turner, 2005) and also concerning the social impact of lowering

communication costs via the Internet (Gaspar and Glaeser, 1998; Wellman and Gulia, 1997; Smith, 1999;

Olson and Olson, 2003; Van Alstyne and Brynjolfsson, 2005), there has been very little prior research

examining the effect of lowered communication costs on collaboration.

One notable exception is Hamermesh and Oster (1998), who explore the effect of the 1980s

“communications revolution” on collaborative research in economics. Using co-authored papers from

three major economics journals, they compare the citation counts of papers published in 1970-79 and in

1992-96. The authors find: 1) a significant increase in the fraction of distant co-authorships, 2) lower

quality of distant co-authored papers as compared to close co-authored papers, and 3) no decline in the

relative quality of distant co-authorships over time.

While this prior work is similar to our study in that both are interested in the effect of lowered

communication costs on research collaboration, our work differs in two fundamental ways. First, we

directly measure the relationship between institution-level Bitnet adoption and collaboration. They instead

3

show that general collaboration patterns changed from the 1970s to the 1990s without showing a direct

link to communications technology. Second, we focus on changes in the propensity to collaborate, using

institution-pairs as the unit of analysis. They focus on how citation counts of collaborative work changed,

using co-authored papers as the unit of analysis.

Another exception is Gaspar and Glaesar (1998). While collaboration is not the main focus of

their article, they show that there has been a rapid growth in local co-authorships in economics since the

1960s. They use this to support their argument that information technology is a complement to face-to-

face interaction. Using more rigorous econometric analysis, we confirm their result. Unlike Gaspar and

Glaesar, however, we examine the mediating role of university quality and show that adoption of the

technology leads to increased productivity among a particular set of schools because of their interaction

with more research-oriented local institutions. Furthermore, our focus is on how technology affects

knowledge flows rather than on how technology affects geographic agglomeration.

Our paper proceeds as follows. In Section 2, we provide a brief historical description of Bitnet

and how it may have facilitated research collaboration. In Section 3, we develop a simple framework of

research collaboration, which motivates our method and results. In Section 4, we describe the Bitnet

adoption data and the engineering publication data that form the basis of our key measures. In Section 5,

we present our empirical framework, and, in Section 6, we present the results, including the estimated

Bitnet effect and the estimated degrees to which the effect is mediated by the quality of and distance

between potential collaborators. Finally, in Section 7, we conclude by discussing the implications of these

findings.

2. A Brief History of Bitnet

Our research question is predicated on the assumption that Bitnet facilitates collaboration. In other words,

we assume a causal relationship between Bitnet adoption and increased collaboration. As such, it is

important to offer some context in terms of what Bitnet was used for, how it originated, what was

4

involved in the Bitnet adoption process, and how Bitnet evolved over time in order to provide some

intuition for our causality assumption.

Bitnet was an early leader in network communications for the research and education community.

It allowed communication via email, exchange of data through file transfer protocols, remote file

archives, Listserv, and compatibility with other operating systems such as UNIX.1 Archived emails of

Bitnet users in the 1980s show instances of individuals exchanging publications, sharing opinions,

conducting academic discussions, searching for research partners for collaboration, and facilitating

resource sharing.2

Our Bitnet data spans the period 1981-1990. The World Wide Web was invented near the end of

this period, in 1989; the first mass-market browser, Mosaic, was developed in 1993.3 Therefore, the

version of the network that we are examining in this paper predates the Internet as it is known today.

In the early 1970s, the Advanced Research Projects Agency (ARPA) created one of the foremost

networks of the existing Internet, ARPANET. This network facilitated exchange of computer data across

North America amongst ARPA-funded researchers for ARPA projects. ARPANET’s restriction to only

ARPA-funded researchers led to the development of several other networks, including CSNET,

USENET, and Bitnet. CSNET was founded in 1981 to serve the needs of computer scientists, while

USENET was created by computer science graduate students to help exchange information via

newsgroups.

Bitnet, on the other hand, was a network created to promote the tools of computer networking for

all scholars, not just select ARPA-funded researchers or computer scientists. Ira Fuchs, the Vice

Chancellor of University Systems at the City University of New York (CUNY), conceptualized Bitnet

(Because It's There NETwork) to take advantage of the existing supply of IBM mainframes on many

university campuses. It was intended as a low-cost and easy-to-setup computer network for

1 http://computing.dcu.ie/~humphrys/net.80s.html (Mark Humphrys, The Internet in the 1980s) 2 Copies of archived emails are available from the authors. 3 The invention of the Web is commonly credited to Tim Berners-Lee at CERN in Switzerland. The browser was developed at the University of Illinois National Center for Supercomputing Applications.

5

communication between university scholars. The Bitnet executive committee organized the effort and

established regulations and bylaws. The network was restricted to academia, and no commercial

communication was allowed.

In order to connect to Bitnet, each institution required a 9600 baud leased line between its

computer facility and another institution that was linked with the network as well as IBM networking

software.4 Bitnet setup costs were reasonably low because mainframes already existed on university

campuses, and communication relied on dial-up technology. Administratively, however, creating an

initial connection to the network seems to have been a reasonably involved process.5 The decision to

adopt was generally made by a university’s computing center managers and approved by higher-level

administrators.

The first Bitnet adopters were CUNY and Yale University in May 1981. Over the next three

years, the network expanded slowly to about 157 nodes and approximately 3000 users (Grier and

Campbell, 2000). It also established an executive committee to expand the network and oversee its

operations. In 1984, Bitnet received significant support through a monetary grant provided by IBM to

establish an operations and information center.

In 1986, NSFNET was founded for connecting supercomputer centers across North America. It

created the backbone of what would become the Internet, leading to the rapid growth of Bitnet over the

next few years. In 1987, the Bitnet Information Center introduced Listserv, an automatic mailing list

server for Bitnet. Listserv remains Bitnet’s lasting legacy and is still employed by many universities

around the world. By the end of the 1980s, Bitnet had become the largest academic network in the world

for computer-based communications; it was gradually replaced by the World Wide Web in the mid 1990s.

4 Bitnet was a "store-and-forward" network. Information originating at a given Bitnet-connected computer (node) was received by intermediate nodes and forwarded to its destination. 5 In a rich description of the National University of Singapore’s decision to connect, the Dean of Science at that institution describes the bureaucratic steps that he had to take, including obtaining the assistance and approval of a number of senior university administrators (who, in this case, were very supportive). (http://www.physics.nus.edu.sg/~phytanb/bitnet4.htm).

6

3. Framework

In this section, we develop a framework to explore the implications of Bitnet adoption on research output

and collaboration. First, our framework assumes that communications technology will increase total

research output and multi-institution collaborative output. Second, the framework allows better

communications technology to increase collaboration both among institutions that are close by and those

that are far away. The framework leaves ambiguous whether communications technology makes distance

relatively more or less important; we will address this open question empirically. Third, the framework

also leaves ambiguous which types of institutions benefit from improved communications technology.

High-quality institutions may collaborate more with each other but not with lower-quality institutions.

Alternatively, lower-quality institutions may benefit from increased access to researchers and equipment

in high-quality institutions. We estimate these relationships in Section 6.

We model the decision to collaborate in two stages. In Stage 1, a researcher receives an idea and a

potential collaboration partner. In Stage 2, the researcher decides whether to collaborate on the idea or to

pursue her own different research agenda. We first will describe Stage 2 and then detail Stage 1.

Stage 2

Researchers aim to maximize the probability of getting published in a top journal. Suppose researcher i

has an idea with potential collaborator j. Further suppose the researcher knows that the probability that

this collaborative effort will lead to a publication is:

(1) PrC=g(qi,qj)-C(D,θ)

We model g() as the benefit to collaboration. It is a function of the quality of institution i, qi and the

quality of institution j, qj. C() is the cost of collaboration. It is a function of the distance between

institutions i and j, D, and the current communications technology, θ.

Suppose the probability of publishing if researcher i does not collaborate on the project and

instead pursues her own agenda is:

7

(2) PrS=f(qi)

We model f() as the net benefit to not collaborating. It is a function of the researcher i's quality, qi. The

researcher will choose the collaborative project if PrC>PrS. Otherwise, the researcher will pursue her own

agenda. We impose the following assumptions:

A1) Costs increase with distance but decrease with communications technology:

0>∂∂DC

and 0<∂∂

θC

.

A2) The rate at which costs increase with distance is lower when the technology is

better: 02

<∂∂

∂θD

C.

A3) The probability of collaborative publication increases with the quality of both

institutions: 0>∂∂

iqg

and 0>∂∂

jqg

.

Assumption A1 implies that improved technology will lead to more collaboration. Assumptions

A1 and A2 combined imply that, in this stage, improved technology will lead to more long-distance

collaboration relative to short-distance collaboration.6 They do not, however, make firm predictions about

the impact of relative quality on collaboration probability. We estimate the relationship between

collaboration and distance in Section 6.3.

This model leads to ambiguous predictions about the impact of an improvement in

communications technology on collaboration across and within quality levels. It is possible that improved

technology will lead to more collaboration across quality levels but not within quality levels. In particular,

suppose there are two types of institutions: high-quality H and low-quality L. Furthermore, suppose a

researcher at a high-quality institution is deciding whether to collaborate on a given project with another

researcher.7 If the other researcher is also at a high-quality institution, then the first researcher will choose

to collaborate if

6 Gaspar and Glaesar (1998) suggest that the opposite effect is possible if communications technology is a complement to face-to-face communication. In this case, C12()>0. This idea is modeled in Stage 1. Overall, the full two-stage model suggests that the relationship between collaboration, technology, and distance is ambiguous, both with and without A2. 7 For ease of exposition, we focus on the decisions of researchers at high-quality institutions. The general implications on the effect of communication costs hold if we focus on only researchers from low-quality institutions.

8

(3) g(H,H)-C(D,θ)>f(H)

Similarly, if the other researcher is at a low-quality institution, then the first researcher will choose to

collaborate if

(4) g(H,L)-C(D,θ)>f(H)

By assumption A3, g(H,H)-C(D,θ)>g(H,L)-C(D,θ). Therefore, collaboration is more likely between the

high-quality institutions. However, the effect of an improvement in communications technology from θ to

θ’ will be ambiguous. If we assume that high-quality institutions almost always collaborate with other

high-quality institutions (i.e., g(H,H) is sufficiently large), then an improvement in technology will not

affect collaboration at the margin in equation (3). Alternatively, if we assume that the benefit to a high-

quality institution of collaborating with a low quality institution is sufficiently low (i.e., g(H,L) is

sufficiently large), then an improvement in communications technology will not affect collaboration at the

margin in equation (4). The effect of Bitnet adoption on collaboration depends on which potential

collaborations are on the margin and is therefore an empirical question that we examine in section 6.4.

Stage 1

Stage 1 models the arrival of ideas with potential collaborators. A given researcher at a high-quality

institution, i, receives a collaborative idea from a researcher at another high-quality institution with

probability π and an idea from a researcher at a low-quality institution with probability (1-π). Now

suppose that the collaborative offers arrive at a rate that depends on distance A(D).

A4) Offers arrive more frequently from nearby institutions than from faraway

institutions: 0<∂∂DA

.

Assumption A4 implies that people are more likely to collaborate with others who are nearby, perhaps

because they are more likely to know each other or perhaps because collaboration is much more effective

If f() is increasing in quality, then all collaborations accepted by high-quality institutions will be accepted by low-quality institutions.

9

if people can meet face-to-face. Therefore, the total collaboration probability for institution i at a given

point in time is:

(5) V=Α(D){π·1[g(H,H)-C(D,θ)>f(H)]+(1-π)·1[g(H,L)-C(D,θ)>f(H)]}

Where 1[] is the identity function. Improved communications technology has an ambiguous effect on

nearby collaboration relative to distant collaboration. The results from Stage 2 (resulting from A1 and A2)

still hold: Improved technology reduces the costs of distant collaboration more than nearby collaboration.

An opposing effect is now also relevant: The likelihood of collaborating at all decreases with distance.

This can be seen most cleanly by taking the derivative of equation (5) with respect to D and θ: Let

u(D,θ)={π·1[g(H,H)-C(D,θ)>f(H)]+(1-π)·1[g(H,L)-C(D,θ)>f(H)]}. Then

(6) θ

θθ

θθθθ ∂

∂=

∂∂

+∂

∂=

∂∂ ),()(),()(),()( DuDADuDADuDAV

(7) D

DuDADuDDA

DV

∂∂∂

+∂

∂∂

∂=

∂∂∂

θθ

θθ

θ),()(),()( 22

This term is ambiguous since 0)(<

∂∂

DDA

, 0),(>

∂∂

θθDu

, 0)( >DA , and 0),(2

>∂∂

∂D

Duθ

θ. Therefore,

equation (6) shows that improved communications technology (i.e., Bitnet) will increase the number of

collaborations. Equation (7) shows that whether distant or nearby potential collaborators are more

affected is an empirical question. Distant institutions may be more affected because their costs to

communicate fall more. Alternatively, nearby institutions may be more affected because they are more

likely to know each other and meet face-to-face. They thus may benefit more if connecting to Bitnet is

highly complementary to these knowledge flow mechanisms. Section 6.5 addresses this question.

4. Data

10

We use two units of analysis in our empirical work. First, we examine the effect of connecting to Bitnet

on individual universities and thus use the university as our unit of analysis.8 Then, we examine the Bitnet

effect on potentially collaborating pairs of universities and therefore use university-pairs as our unit of

analysis. We employ four types of data in our empirical work: 1) publication data, 2) Bitnet connection

data, 3) institution quality data, and 4) distance data. We describe these below.

4.1 Publication Data

Since we are interested in identifying the Bitnet effect on collaboration, we use publication data from

researchers in technical areas that are likely to be early adopters of this communication technology. We

do this because we exploit the variation in connection years for identification, and the publishing behavior

of these researchers is most likely to reflect these time differences. Therefore, we select a variety of

electrical engineering research topics that appeared in journals published by the Institute of Electrical &

Electronic Engineers (IEEE).

Specifically, we collect publication data from seven journals over the 15-year period 1977-1991:

1) IEEE Transactions on Aerospace and Electronic Systems, 2) IEEE Transactions on Nuclear Science, 3)

IEEE Transactions on Biomedical Engineering, 4) IEEE Journal of Quantum Electronics, 5) IEEE

Transactions on Electron Devices, 6) IEEE Transactions on Communications, and 7) IEEE Transactions

on Education. Each of these journals is considered among the top outlets for research in the specified

field. Since we focus only on these seven top journals, the total number of publications in our analysis

does not rise over time.

There are 28,312 papers published in these seven journals during the time period under

investigation.9 We parse out all unique author-affiliated institutions from each paper. An example of the

data structure is provided in Table 1. Papers are categorized as either single-institution or multi-institution

8 We only include US universities that have published at least one paper in our set of seven electrical engineering journals during the period 1977-1991. 9 The distribution of publications across journals is not uniform. In the order listed above, the number of publications per journal is 6174, 8505, 3418, 4976, 827, 3585, and 827.

11

(i.e., collaborative). Papers are classified as single-institution if all authors are from a single university.

So, in the example provided in Table 1, the left-hand-side paper is classified as single-authored since all

three authors are from Duke University. However, the right-hand-side paper is classified as multi-

institution since the authors have collaborated across institutional boundaries (Yale and Villanova).

After extracting the institutional information from the set of publications, we identify 739 unique

institutions, of which 289 are US universities - our institution-type of interest. These form the basis of our

unit of analysis. We focus on US universities because many of the international institutions and US non-

university research labs used other networks besides Bitnet. US universities are particularly likely to have

Bitnet as the first data communication network they adopted.10

These data allow us to construct two datasets, each of which is focused on the measurement of the

number of collaborative (multi-institution) papers. First, we construct a single institution dataset that

includes 15 years of publishing data (1977-1991) from the specified journals by the 289 institutions of

interest. This is therefore a balanced panel dataset that consists of 4335 observations. Then, we construct

an institution-pair dataset that includes the same 15 years of publishing data from the specified journals

by the 41,616 institution-pairs resulting in 624,240 observations.

4.2 Bitnet Connection Data

We use an online reference, Cyber Geography Research, for a record of Bitnet connections.11 This archive

lists the 1,054 institutions worldwide that connected to Bitnet by the end of 1990 and their connection

date.12 Of the 289 US universities that published at least once in the seven IEEE journals we examine, 225

connected to Bitnet during by this time. In other words, 64 US universities in our publishing sample had

not connected to Bitnet by the end of 1990. Figure 1 illustrates the connection rate. The variation in

10 http://computing.dcu.ie/~humphrys/net.80s.html (Mark Humphrys, The Internet in the 1980s) 11 http://www.cybergeography.org/atlas/bitnet_topology.txt 12 In our analysis, we use the year following adoption as the first year Bitnet is available at the university. Therefore, the effect of adoption is first measured in 1984 for a school that adopts in 1983. This is to account for the (relatively short) publication lag in engineering and any lag in intra-university adoption. Results are robust to using the year of adoption.

12

connection years is important for our econometric analysis since we exploit this to identify the effect of

Bitnet.

In addition, although the first three institutions connected were all on the east coast, by the second

year some universities further inland were connected, such as Ohio State, and two, UC Berkeley and UC

San Francisco, were on the west coast. By the third year, universities up and down the east coast were

connected, as well as institutions as far inland as the University of Missouri-Columbia (MO), in addition

to a third institution on the west coast (Stanford). Since we examine the Bitnet effect with respect to the

distance between collaborating institutions, a sense of the geographic distribution is also useful. As such,

the geographic distributions of connected universities over the first three years are illustrated in Figure 2.

4.3 Quality Data

We use the 1987 Carnegie Foundation classification system to classify the research “quality” of each

university in our dataset.13 We classify universities as Carnegie Type 1, 2, or 3 (CT1, CT2, CT3).

Carnegie Type 1 (CT1) is an aggregate of the Carnegie Foundation’s categories “Research University 1

and 2.” Thus, institutions with our CT1 classification offer a full range of baccalaureate programs, are

committed to graduate education through the doctorate degree, and give high priority to research. They

receive annually at least $12.5 million in federal support and award at least 50 Ph.D. degrees each year.14

Institutions with a CT2 classification are an aggregate of the Carnegie Foundation’s categories

“Doctorate-Granting Universities 1 and 2.” These institutions offer a full range of baccalaureate

programs, and their mission includes at least some commitment to graduate education through the

doctorate degree, such that they award annually 20 or more Ph.D. degrees in at least one discipline or 10

or more Ph.D. degrees in three or more disciplines, but do not meet the requirements for Carnegie Type 1.

All other universities are classified as CT3.

13 A Classification of Institutions of Higher Education (1987 Edition), A Carnegie Foundation Technical Report, Princeton University Press, Lawrenceville, NJ. 14 The years used in calculating average federal support were 1983, 1984, and 1985.

13

4.4 Distance Data

We generate distance data as follows. First, we search the Internet for the official website of each

university in our dataset to establish the primary location (city, state) of its research campus. Then, we

obtain latitude and longitude measures from the US Geological Survey based on the city-state data.15

Finally, we determine the distance between each university pair by employing the great circle method to

calculate the distance in kilometers between the two sets of geographic coordinates.16

5. Empirical Strategy

We present two types of results: institution-level and institution-pair-level. Our estimation strategy is

based on “difference-in-differences” identification. For the single institution data, we examine how total

publications and how multi-institution publications change after an institution adopts Bitnet relative to

other institutions. For the paired institution data, we examine how collaboration between institution-pairs

that both adopt Bitnet changes relative to other institution-pairs in which one or both have not adopted.

We observe the number of collaborative multi-institution publications at the institution and the institution-

pair levels.

5.1 Single Institution Estimation

We use the single institution data for two purposes. First, we explore whether Bitnet adoption is related to

an increase in total publications, and, second, we explore whether Bitnet adoption is related to an increase

in multi-institution publications specifically.

To estimate the relationship between Bitnet adoption and total publications, we estimate the

following fixed effects regression for institution i in year t:

(8) Total # publicationsit=βHas Bitnetit+µt+φi+εit

15 The US Geological Survey can be accessed at http://geonames.usgs.gov/ and a web query application exists at http://geonames.usgs.gov/pls/gnis/web_query.gnis_web_query_form. 16 The great circle formula used is: acos(cos(lat1)*cos(long1)*cos(lat2)*cos(long2) +cos(lat1)*sin(long1)*cos(lat2)*sin(long2)+sin(lat1)*sin(lat2)) * earthRadius

14

Where β measures the effect of Bitnet adoption, µt is a year fixed effect, φi is an institution fixed effect,

and εit is an idiosyncratic error term.

When estimating the effect of Bitnet adoption on multi-institution publications, we estimate a

similar equation:

(9) #Multi-instit. publicationsit=αXit+βHas Bitnetit+µt+φi+εit

Where Xit contains observed institution-year characteristics. In particular, Xit is a count of single-

institution publications, an observable measure of how institution quality may vary over time.

We use equations (8) and (9) to explore whether Bitnet adoption is related to an increase in both

total publications and in academic collaboration. Since the dependent variable is a count variable, our

main models estimate these using fixed-effects negative binomial regressions. We also show results using

OLS regressions.

For equations (8) and (9) to estimate the Bitnet effect on total publications and multi-institution

publications respectively, we assume that unobserved institution quality can be decomposed into an

additively separable fixed component and a time varying component. The time component is constant

across institutions (Athey and Stern, 2002). This assumption is slightly weaker in equation (9) than in

equation (8) because we have an observable measure of institution quality.

More generally, this assumption is questionable if Bitnet adoption is a signal of an unobserved

quality improvement. While the decisions of the administrators on adoption are likely removed from

changes in engineering faculties’ propensity to collaborate, this alternative hypothesis cannot be rejected

in the single-institution data. Therefore, the results based on equations (8) and (9) are merely suggestive.

They show that Bitnet may have increased productivity and collaboration.

The institution-pair results on Bitnet and collaboration are much less susceptible to criticisms of

this assumption. In the next section, we present an estimation strategy based on institution-pairs that is

less susceptible to criticisms of spurious correlation between adoption choice and collaboration.

15

Furthermore, it allows us to better understand the reasons Bitnet adoption may be correlated with

increased collaboration.

5.2 Institution-Pair Estimation

We are interested in Bitnet as the communications technology. Since Bitnet is only effective if both

institutions have the technology, we focus on the effect of both institutions adopting Bitnet on

collaboration. We interpret evidence that just one institution adopting Bitnet does not lead to significantly

more collaboration as support for our main findings. We label the first institution in the pair i, the second

j, and the year t.

We run linear regressions on the data to get the following equation:

(10) Collaborationijt=αXijt+βBoth Have Bitnetijt+µt+φij+εijt

Where φij now measures institution-pair fixed effects and Xijt is a vector of observable institution-pair-year

characteristics. This vector contains our proxy for observed pair quality in year t: the total number of

single-institution papers published by both universities in that year.17 In some specifications, it also

contains the distance between the institutions and whether any one of the pair has adopted Bitnet.

In our main models, we treat Collaborationijt as a dummy variable for whether institution i and j

have any collaborations in year t. We estimate equation (10) using a fixed effects linear probability (OLS)

regression. We treat collaboration as a dummy variable because 85% of all institution-pair-years with at

least one collaboration have only one collaboration. We also show results for a fixed effects OLS

regression on total number of collaborations, a fixed effects probit regression on only those pairs with at

least one publication, and a fixed effects negative binomial regression on total number of publications on

only those pairs with at least one publication.18

17 We also show that results do not change if the product of the single-institution publications at the two universities is used instead of the sum. 18 We limit the fixed effects probit and negative binomial regressions to only those pairs with at least one publication in order to overcome computational difficulties in estimating thousands of dummy variables.

16

For this linear equation to identify the average effect of Bitnet adoption on collaboration between

two given institutions, we again implicitly assume that unobserved institution-pair quality can be

decomposed into an additively separable fixed component and a time varying component. The time

component is constant across institution-pairs (Athey and Stern 2002). This is a weaker assumption than

that used for the single-institution data. In this case, unobserved quality is at the institution-pair rather

than the institution level. Furthermore, we present evidence that both institutions need to adopt for

adoption to be correlated with increased collaboration. If only one institution adopts, they are no more

likely to collaborate. The next section presents the estimates of these models.

6. Results

In this section, we examine whether Bitnet adoption influences research collaboration among university

scientists. In section 6.1, we begin with descriptive statistics that suggest that schools that adopt Bitnet do

publish more and collaborate more. Using a difference-in-difference type identification, Section 6.2

shows that after Bitnet adoption universities produce more papers in the top journals than they do before

Bitnet adoption relative to universities that do not adopt. Section 6.3 shows that inter-institution

collaboration in particular increased considerably. In Sections 6.4 and 6.5, we show that medium-ranked

universities seem to benefit most and that Bitnet especially facilitates collaboration between nearby

institutions.

6.1 Descriptive Statistics

We begin our analysis by examining the basic distributional properties of our key measures: 1) research

publications, 2) Bitnet adoption, 3) institution quality, and 4) distance between potential collaborators.

Table 2a presents descriptive statistics with institution-years as the unit of analysis, and Table 2b presents

descriptive statistics with institution-pair-years as the unit of analysis.19

19 As discussed in Sections 4 and 5, our chosen empirical methodology requires us to examine our data from two perspectives using two distinct units of analysis. First, we examine the effect of Bitnet adoption on the research

17

The average university in our dataset publishes 2.51 papers per year in our specified set of

journals over the particular time period under investigation. Of these, 1.36 are multi-institutional and 1.15

are single-institutional. In other words, on average these institutions publish approximately 18% more

multi-institutional papers per year.20 Overall, 37% of the institution-year observations have at least one

multi-institutional paper.

With respect to Bitnet adoption, while the first three institutions were connected to Bitnet in

1981, the average university in the dataset, conditional on being connected before or during 1990, is not

connected until halfway through 1985.21 Figure 1 illustrates the significant variation in adoption years

across universities, which is central to our identification strategy.

The first hint of a Bitnet effect is reflected in the comparison of multi-institutional papers per year

by institutions with and without Bitnet: 2.47 and 0.92 respectively. However, this difference could be

caused by other factors such as time (multi-institution papers are more likely in later years, perhaps for

reasons that are not related to Bitnet adoption, and universities are more likely to have adopted Bitnet in

later years) and quality (higher-quality institutions may be more likely to publish a greater number of

multi-institutional papers and may also be more likely to adopt Bitnet earlier). Also, it is interesting to

note that, at least at first glance, Bitnet adoption is correlated with overall output, including single-

institution paper production. As Table 2a illustrates, institution-years with Bitnet average 2.30 more total

papers per year, of which only 1.55 are multi-institutional. Thus, institution-years with Bitnet have 0.75

more single-institution papers.

Next, we compare average paper production by institutions in the three Carnegie research quality

categories. Overall, the research output from these institutions corresponds with what one would expect.

behavior of single institutions. For this we use “institution-year” as our unit of analysis. Then, we examine the effect of Bitnet adoption on the research behavior of pairs of institutions and use “institution-pair-year” as our unit of analysis. 20 It is important to note that since our dataset is constructed from a constant set of top journals in electrical engineering, the total number of publications in our sample does not increase steadily over time. In fact, there are a total of 1,702 papers published in the first year of observation (1977) and 1,646 papers published in the last year (1991). The total number of publications fluctuates from year to year due to the publication of special issues and occasional conference proceedings. 21 Of the 289 universities in our dataset, 64 were not connected by 1990.

18

Type 1 institutions produce more than six times as many multi-institutional papers as Type 2, and Type 2

institutions produce almost 1.5 times as many multi-institutional papers as Type 3. The ratios for single-

institutional papers are similar.

Finally, we turn to Table 2b to examine the basic properties of our institution-pairs. The average

institution-pair is separated by a distance of 1,742 kilometers (1,082 miles) and produces 0.00132 multi-

institution papers per year. Again, we see suggestive evidence of a Bitnet effect since university-pairs

generate, on average, more than four times as many multi-institution papers if both are connected to

Bitnet. However, for the same reasons as described above, other factors such as time and quality could be

confounding this relationship.

We now move on to a more rigorous examination of the effect of Bitnet. That is, following up on

the suggestive evidence presented in the descriptive statistics, we seek to estimate the extent to which

Bitnet adoption increases the propensity of universities to publish and to collaborate, after controlling for

likely confounding effects.

6.2 The Effect of Bitnet on Total Publications

We begin our examination of the Bitnet effect by examining whether Bitnet adoption is associated with an

increase in total publications. We simply regress our count of total papers per institution-year on whether

an institution has adopted Bitnet (Has Bitnet), including year and institution fixed effects. Column (1) of

Table 3 shows our main negative binomial specification with year and institution fixed effects. This

specification estimates a positive and significant coefficient on Has Bitnet, consistent with the descriptive

data presented above.22 Columns (3), (4), and (5) suggest that middle-level institutions especially benefit

from Bitnet adoption. We explore this idea further in Section 5.4.

Since total publications rise, this suggests that Bitnet adoption may increase the productivity of

the relevant university research group. While this result may be a function of spurious correlation between

22 The OLS specification, which is highly inefficient here, also estimates a positive coefficient as expected, but it is not statistically significant.

19

adoption and research productivity, it is at least suggestive of a correlation between adoption and research

output. The rest of the paper explores how Bitnet adoption might have increased productivity by focusing

on its role in increasing collaboration across institutions.

6.3 The Effect of Bitnet on Collaboration across Institutions

We conduct our estimation of the Bitnet effect on collaboration in two parts. First, we provide suggestive

evidence that Bitnet adoption increases collaboration using single-institution data. Then we use

institution-pair data to overcome the alternative possibility of spurious correlation between Bitnet

adoption and collaboration. We present each of these analyses in turn.

6.3.1 Single Institution Analysis

For this analysis, our dependent variable is a count of multi-institutional papers (Table 4 columns (1) and

(2)). We regress this dependent variable on Has Bitnet, again including year and institution fixed effects,

and now also include a count of single-institution papers to control for within-institution quality changes

over time. Both negative binomial and OLS specifications generate positive and significant coefficients

on Has Bitnet. The magnitude of this effect is also economically significant. Our results suggest that even

after controlling for year, institutional characteristics, and within-institution quality changes over time,

connecting to Bitnet results in almost a 30% increase in multi-institutional papers per year.23 This

provides preliminary evidence that Bitnet adoption increased academic collaboration.

6.3.2 Institution-Pairs Analysis

We now turn our attention to our second unit of analysis, the institution-pair. With the single institution

analysis above, even though we have controlled for year and institution effects as well as within-

university quality changes as measured by single-institution paper counts, one might worry about an

23 For robustness, we also have run Poisson, zero-inflated Poisson, and negative binomial regressions and find similar results.

20

unobserved variable bias. For example, perhaps young scholars have a taste for collaboration and also for

Bitnet (and are able to influence their universities to connect earlier than they otherwise would). In such a

case, we would observe universities that connect earlier also collaborating more, even though Bitnet

connection is not facilitating collaboration. Ideally, we would have an instrument that is correlated with

Bitnet adoption but not with publications. In the absence of such an instrument, we rely on evidence from

institution-pairs. It is much less likely that researchers in our sample who, at a stretch, might be able to

influence their universities to connect, also are able to influence other university administrations to do so.

As reported in Table 5 column (1), we find a significant effect on collaborative publications from

both universities in the pair being connected to Bitnet (Both Have Bitnet). In fact, the likelihood of a

collaboration between institutions increases by approximately 85% if both institutions have connected,

controlling for year and institution-pair effects as well as for within-pair quality changes over time

(measured by single-institution publications).

We also observe that this effect remains virtually unchanged when we control for whether only

one of the members of the pair has connected (Table 5 column (2)). We find this a particularly interesting

result. There is no increase in the likelihood of collaboration when only one joins the network (the

coefficient on One or more has adopted Bitnet is not significant), but there is a significant effect when

both connect. We interpret this result as suggesting that it is both universities in the pair being connected

that drives the increase in collaboration. This provides strong evidence that it is the connection between

universities that leads to increased collaboration. It does not appear to be a result of spurious correlation

between adoption and collaboration.

We conduct a variety of robustness checks on our institution-pair results. First, we employ a

count measure rather than a dummy for our dependent variable and generate results that suggest that both

institutions having Bitnet results in a 75% increase in the number of collaborations (Table 5 column (3)).

Next, we run probit and negative binomial specifications and find similar results. Finally, we drop the

fixed effects and conduct OLS with clustered standard errors by institution-pairs in column (6), include a

21

variation in the quality control measure in column (7), and omit the control measure in column (8). Still,

the main result persists.24

Now that we have established the general effect of Bitnet, we turn our attention to how it is

mediated by institution quality and distance. Are the benefits from Bitnet adoption spread uniformly

across all adopters, or do they vary with quality and distance?

6.4 Does the Bitnet Effect Vary with Institution Quality?

In Section 6.2, we show that medium-ranked schools seem to benefit most from Bitnet in terms of

increased collaboration. In particular, when we split our sample into the aggregated Carnegie research

strength categories, the coefficient on Has Bitnet is only statistically significant for universities in the

middle category. This result persists when we examine the relationship between multi-institutional papers

and Bitnet adoption in Table 4, although the coefficient is positive for all three categories in both tables.

We examine the relationship between university research quality and the Bitnet effect more

closely in Table 6. Here, using our institution-pair data, we divide our sample into six groups, reflecting

all possible combinations of quality types. While all coefficients on Both Have Bitnet are positive, except

for the low-low pairs, only the coefficient on high-middle pairs is statistically significant. For this sub-

sample, both universities in the pair being connected increases the likelihood of collaboration by over

150%.

We find these quality-related results striking. It appears that the benefits of Bitnet adoption,

measured by an increase in publications, accrue primarily to medium-ranked schools (Tables 3 and 4).

Furthermore, it seems as though these medium-ranked schools benefit by collaborating with top-ranked

schools (Table 6). Top-ranked schools seem neither helped nor harmed by Bitnet adoption.

We have two possible explanations for this result. First, it may be that top-ranked schools

collaborate with other top-ranked schools anyway, such that these collaborations are not on the margin.

24 In Table 5 column (7), we employ the product of the count of single-institution papers from each university in the pair rather than the sum as our control for within-pair quality changes over time.

22

However, collaboration between top-ranked schools and medium-ranked schools may be marginal such

that a decrease in the costs of collaboration leads to significantly more collaboration. This could be

professors at top schools collaborating with former students, which was not previously worthwhile, or a

number of other similar possibilities.

Second, top-ranked schools may have equipment that medium-ranked schools do not. When top-

ranked schools collaborate with each other, their communication needs are simple. Each researcher can

work on their own equipment, talk on the phone, and overnight drafts when necessary. Medium-ranked

schools, however, may not have the specialized equipment necessary to run certain experiments, even if

they have the idea. Therefore, there is a considerable amount of data sent back and forth in this type of

collaboration. Under this scenario, Bitnet adoption is more valuable in top-middle collaborations that

require “thick” data communications as compared to top-top collaborations that may require only “thin”

communications.

6.5 Does the Bitnet Effect Vary with Distance?

Here we examine whether the Bitnet effect varies with distance. As discussed above, if Bitnet adoption is

a substitute for face-to-face interaction, we would expect the benefits of adoption to be greatest for

universities that are furthest apart. On the other hand, if Bitnet adoption is a complement to face-to-face

interaction, we would expect the benefits of adoption to be greatest for universities that are co-located.

To address this issue, we employ a spline regression, grouping together universities that are: 1)

within 100km, 2) between 100 and 1000km apart, 3) between 1000 and 3000km apart, and 4) further than

3000km apart. Our results, which again include year and institution-pair fixed effects as well as a control

for within-pair productivity changes over time, suggest that Bitnet adoption is likely a complement for

face-to-face interactions. Bitnet adoption has the greatest effect on university-pairs that are co-located

(within 100km). We also find that collaboration increases regardless of distance, suggesting that Bitnet

does reduce the cost of collaboration even for distant institutions. Overall, the Bitnet effect is more than

twice as large for co-located universities as it is for those that are in the next category (100-1000km

23

apart). For robustness, we report various groupings using different distance ranges, but the main result

persists: Co-located universities benefit most from Bitnet adoption.

Finally, we examine the interaction effect of quality and distance on the Bitnet effect (Table 8).

Overall, our prior results persist: The greatest effect on multi-institutional paper production occurs for

medium-ranked universities that collaborate with co-located top-ranked universities. Indeed, medium-

ranked universities also increase their collaboration with non-co-located top-ranked universities, but the

effect of Bitnet is almost 10 times greater for those that are co-located. When the data are split this way,

we see that low-ranked universities also benefit from the Bitnet effect through collaboration with co-

located top-ranked institutions. Interestingly, pairs of top-ranked, co-located universities seem to reduce

their level of collaboration upon adopting Bitnet. We speculate that this may be due to substitution for

collaboration with medium-ranked schools but leave further investigation of this phenomenon for future

research.

7. Conclusions

Does connecting to Bitnet increase research productivity? If so, how? Are benefits evenly distributed

across all adopters? We offer evidence that connecting to Bitnet is related to an increase in overall

research productivity, particularly for medium-ranked schools. Moreover, our findings suggest that this

increase is mainly related to an increase in collaboration by medium-ranked universities with co-located

top-ranked schools.

Bitnet adoption increases productivity by lowering the cost of communication and thereby

increasing collaboration. Indeed, we find that both institutions being connected to Bitnet increases their

likelihood of collaboration by approximately 85%, after controlling for year and institution-pair effects as

well as for within-pair quality changes over time. As discussed in the introduction, collaboration allows

the sharing of knowledge as well as specialized equipment and other resources that may lead to a more

efficiently functioning market for ideas. Bitnet adoption may lead to collaboration simply by reducing the

costs of exchanging ideas. Alternatively, Bitnet adoption may make it easier for a researcher in one

24

location with an idea to collaborate with a researcher in another location with specialized equipment for

testing the idea.

Universities that are co-located seem to experience the greatest boost in collaboration as a result

of Bitnet adoption. We interpret this finding as indicating that low-cost electronic communication, while

certainly a substitute for face-to-face interactions under certain conditions, is also a complement.

Researchers communicate with people they know and are more likely to know those who work nearby.

The policy implications of this finding are clear. For communication infrastructure to increase

collaboration, it must connect people who want to communicate; presumably these are more likely to be

people who already know each other and have an existing social relationship.

Medium-ranked universities seem to benefit most. One might expect top-ranked universities to

benefit most from Bitnet adoption. On average, these institutions publish over six times as often as their

medium-ranked counterparts. However, for all permutations of the data, including total publications and

multi-institutional publications as well as for single institutions and institution-pairs, the statistically and

economically significant coefficient on Has Bitnet is primarily associated with medium-ranked

universities. Why?

The results presented in Table 6 offer some insight. They suggest that the main source of

increased collaboration is between top- and medium-ranked universities. We interpret this result as

indicating that researchers at medium-ranked universities offer something valuable to their colleagues at

top-ranked schools and that Bitnet adoption facilitates this exchange. Perhaps they have good ideas that

need testing on specialized equipment only available at top-ranked schools, or perhaps they are willing to

work harder as collaborators. In either case, from the perspective of the top-ranked institution

considering collaboration with either another top-ranked institution or with a middle-ranked institution,

the marginal benefit of Bitnet adoption may be greater for the latter.

At the same time, there is surprisingly little increase in collaboration within same quality groups;

there is no significant increase in collaboration amongst the prolific set of top-ranked institutions. We

speculate that this is because top-ranked universities have access to their own specialized equipment, and

25

therefore the data-sharing capability of Bitnet is not as important. Of course, that doesn’t mean these

universities don’t collaborate. They do. However, connecting to Bitnet does not result in an increased

propensity to collaborate. Low-ranked institutions do not benefit significantly from being connected to

top- and medium-ranked universities. Even though they have to be publishing institutions to be included

in the dataset, the research efforts of these universities may be too limited to attract collaborative partners.

Overall, these results suggest that welfare returns on public investments in scientific

communications infrastructure are highly sensitive to the demand, or latent demand, to communicate.

Bitnet seems to have achieved its main purpose of facilitating academic collaboration. In doing so,

middle-ranked universities are the main beneficiaries. Bitnet facilitates access to resources at co-located

top-ranked universities and consequently increases their productivity.

26

References Agrawal, Ajay, Iain Cockburn, and John McHale (2003) “Gone But Not Forgotten: Labor Flows,

Knowledge Spillovers, and Enduring Social Capital,” National Bureau of Economic Research Working Paper #9950.

Audrestch, David B., and Maryann P. Feldman (1996) “R&D Flows and the Geography of Innovation and

Production,” American Economic Review Vol. 86(3), 630-640. Barnett, Andy H., Richard W.Ault, and David L.Kaserman (1988) “The Rising Incidence of Co-

authorship in Economics: Further Evidence,” Review of Economics and Statistics 70(3), 539-543. Beaver, D. and R. Rosen (1978) “Studies in Scientific Collaboration. Part 1. The Professional Origins of

Scientific Co-authorship,” Scientometrics, 1, 65-84. Beaver, D. and R. Rosen (1979) “Studies in Scientific Collaboration. Part 2. Scientific Co-authorship,

Research Productivity and Visibility in the French Scientific Elite,” Scientometrics, 1, 133-49. Cairncross, F. (1997) The Death of Distance: How the Communications Revolution Will Change Our

Lives, Harvard Business School Press, Cambridge, MA. Cockburn, I., and R. Henderson (1998) "Absorptive Capacity, Coauthoring Behavior, and the

Organization of Research in Drug Discovery," Journal of Industrial Economics, Vol. XLVI, No. 2, pp. 157-182.

Crane, D. (1969) “Social Structure in a Group of Scientists: A Test of the ‘Invisible College’

Hypothesis,” American Sociological Review, XXXIV, 335-352. Gaspar, J. and E. Glaeser (1998) “Information Technology and the Future of Cities,” Journal of Urban

Economics, 48(1), 136-156. Grier, D.A., and M. Campbell (2000). “A Social History of Bitnet and Listserv, 1985-1991,” IEEE Annals

of the History of Computing, April-June, 42-41. Gurbaxani, V., 1990, Diffusion in Computing Networks: The Case of BITNET, Communications of the

ACM 33, 65-75. Hamermesh, D.S., and S.M. Oster (1998) "Tools or Toys? The Impact of High Technology on Scholarly

Productivity," National Bureau of Economic Research Working Paper #6761. Jaffe, A.B., M. Trajtenberg, and R. Henderson (1993) “Geographic Localization of Knowledge Flows as

Evidenced by Patent Citations,” Quarterly Journal of Economics 108, 577-598. Mairesse J. and L. Turner (2005) “Measurement and Explanation of the Intensity of Co-publication in

Scientific Research: An Analysis at the Laboratory Level,” in New Frontiers in the Economics of Innovation and New Technology: Essays in Honor of Paul David, eds. C. Antonelli, D. Foray, B. Hall, and E. Steinmueller, Edward Elgar Publishing, forthcoming.

Olson, G.M., and J.S. Olson (2003) “Mitigating the Effects of Distance on Collaborative Intellectual

Work,” Economics of Innovation and New Technology, Vol. 12(1), pp. 27-42.

27

Romer, P. M. (1986) "Increasing Returns and Long-Run Growth," Journal of Political Economy, Vol. 94, pp. 1001-37.

Romer, P. M. (1990) "Endogenous Technological Change," Journal of Political Economy, Vol. 98,

supplement to No. 5, pp. 71-102. Singh, J. (2005) “Collaborative Networks as Determinants of Knowledge Diffusion Patterns,”

Management Science, 51(5): 756-770. Smith, M.A. (1999) “Invisible Crowds in Cyberspace: Mapping the Social Structure of the Usenet” in

Communities in Cyberspace: Perspectives on New Forms of Social Organization, Routledge Press, London, UK.

Thompson, P. and M. Fox-Kean (2005) “Patent Citations and the Geography of Knowledge Spillovers: A

Reassessment”, American Economic Review, 95(1): 450-460. Van Alstyne, M. and E. Brynjolfsson (2005) “Global Village or Cyber-Balkans? Modeling and

Measuring the Integration of Electronic Communities,” Management Science 51, 851-868. Wellman, B. and M. Gulia (1997) “Net Surfers Don’t Ride Alone: Virtual Communities as Communities”

P. Kollock and M. Smith eds. Communities in Cyberspace, University of California Press, Berkeley, CA.

Zucker, L., M. Darby, and M.B. Brewer (1998) “Intellectual Capital and the Birth of U.S. Biotechnology

Enterprises,” American Economic Review 88, 290-306.

28

Table 1 – Example of Publication Data Single-Institution Multi-Institution

Title A VOLTAGE-TRIGGERED SYSTEM FOR ADAPTIVE

SAMPLING IN BODY-SURFACE MAPPING PARAMETRIC MODELING OF SOMATOSENSORY EVOKED-POTENTIALS

Author(s) BLANCHARD SM, BARR RC, SPACH MS JACOBS MH, RAO SS, JOSE GV Source IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING

29 (11): 726-730 NOV 1982 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING 36 (3): 392-403 MAR 1989

Addresses DUKE UNIV, DEPT BIOMED ENGN, DURHAM, NC 27706 USA<br> DUKE UNIV, DEPT PEDIAT, DURHAM, NC 27706 USA<br> DUKE UNIV, DEPT PHYSIOL, DURHAM, NC 27706 USA

YALE UNIV, DEPT ELECT ENGN, NEW HAVEN, CT 06520 USA<br> VILLANOVA UNIV, DEPT ELECT ENGN, VILLANOVA, PA 19085 USA

29

Table 2a: Descriptive Statistics—Single Institution Data Variable (by year) Mean Standard

deviation Minimum Maximum # of

observations Total papers 2.508 6.439 0 131 4,335 Multi-institution papers 1.360 3.147 0 39 4,335 Single institution papers 1.149 3.739 0 92 4,335 Any Multi-institution papers (dummy) 0.374 0.484 0 1 4,335 Average year online* 1985.573 2.054 1981 1990 225 Has Bitnet 0.282 0.450 0 1 4,335 Total papers if do not have Bitnet 1.862 5.724 0 106 3,114 Total papers if have Bitnet 4.16 7.738 0 131 1,221 Multi-institution papers if do not have Bitnet 0.923 2.521 0 31 3,114 Multi-institution papers if have Bitnet 2.473 4.153 0 39 1,221 CMU Type 1 Multi-institution papers 3.062 4.537 0 39 1,575 Single institution papers 2.696 5.697 0 92 1,575 CMU Type 2 Multi-institution papers 0.494 1.046 0 9 945 Single institution papers 0.368 0.837 0 9 945 CMU Type 3 Multi-institution papers 0.333 1.131 0 17 1,815 Single institution papers 0.212 1.263 0 9 1,815 *Conditional on being online by the end of 1990

30

Table 2b: Descriptive Statistics—Institution-Pair Data Mean Standard

deviation Minimum Maximum # of

observations # collaborative papers between the pair 0.00132 0.0466 0 6 624,240 Dummy if any collaborative papers that year 0.00107 0.0328 0 1 624,240 # collaborative papers if at least one has not adopted Bitnet 0.000852 0.0379 0 6 516,635 # collaborative papers if both have adopted Bitnet 0.00360 0.0756 0 5 107,605 Distance 1,742.732 1,281.485 0 8,293.748 624,240 Sum of # of single institution papers produced by the pair 2.297 5.295 0 122 624,240 Product of # of single institution papers produced by the pair 1.362 15.250 0 3,496 624,240 Dummy if at least one of the pair has adopted Bitnet 0.391 0.488 0 1 624,240 Dummy if both institutions have adopted Bitnet 0.172 0.378 0 1 624,240

31

Table 3: Bitnet Adoption and Total Publications in the Single Institution Data (1) (2) (3) (4) (5) Main Specification:

Negative Binomial Regression

Linear Regression Main Specification on Institutions in

CT1

Main Specification on Institutions in

CT2

Main Specification on Institutions in

CT3 0.196 0.178 0.0248 0.700 0.260 Has Bitnet

(0.0597)** (0.177) (0.0712) (0.193)** (0.166) # of Observations 4,335 4,335 1,575 945 1,815 # of Groups 289 289 105 63 121 Log Likelihood -5,346.82 N/A -3,169.20 -947.11 -1,179.61 Regressions include year and institution fixed effects

32

Table 4: Bitnet Adoption and Multi-Institution Publications in the Single Institution Data (1) (2) (3) (4) (5) Main Specification:

Negative Binomial Regression

Linear Regression

Main Specification on Institutions in

CT1

Main Specification on Institutions in

CT2

Main Specification on Institutions in

CT3 0.163 0.394 0.0497 0.586 0.107 Has Bitnet

(0.0654)* (0.0799)** (0.0779) (0.235)* (0.180) 0.0193 0.221 0.0159 0.159 0.227 Single institution papers

(0.00280)** (0.00975)** (0.00260)** (0.0500)** (0.0477)** # of Observations 4,035 4,335 1,560 915 1,560 # of Groups 269 289 104 61 104 Log Likelihood -4,105.85 N/A -2,527.05 -672.92 -857.99

Regressions include year and institution fixed effects

33

Table 5: Bitnet Adoption and Collaboration Using Institution-Pairs (1) (2) (3) (4) (5) (6) (7) (8) Main specification:

Linear regression with a dummy for any collaboration as the dependent

variable

Includes variable if just one institution

has adopted

Dependent variable is the

total # of collaborations

Probit regression with only

pairs with at least one

collaboration

Negative binomial regression on total #

of collaborations using only pairs with at least one

collaboration

No fixed effects linear regression;

errors clustered by

institution-pair

Alternative control for pair

productivity

No control for pair

productivity

0.000917 0.000918 0.00102 0.270 0.409 0.00247 0.000921 0.000907 Both have Bitnet (0.000156)** (0.000158)** (0.000221)** (0.113)* (0.219)+ (0.000335)** (0.000156)** (0.000156)**

-0.00000500 One or more has adopted Bitnet (0.0001645)

0.0000544 0.0000544 0.0000399 0.00777 0.0190 0.000430 Sum of # of single institution papers (0.0000128)** (0.0000128)** (0.0000181)* (0.00425)+ (0.00829)** (0.0000542)**

0.0000154 Product of # of single institution papers (0.00000340)** # of Observations 624,240 624,240 624,240 6,930 6,930 624,240 624,240 624,240 # of Groups 41,616 41,616 41,616 462 462 N/A 41,616 41,616 Log Likelihood N/A N/A N/A -1,251.76 -1,373.50 N/A N/A N/A

Unless otherwise specified, regressions include year and institution-pair fixed effects

34

Table 6: Bitnet Adoption, Collaboration, and Institution-Pair Quality (1) (2) (3) (4) (5) (6) CT1 and CT1 CT1 and CT2 CT1 and CT3 CT2 and CT2 CT2 and CT3 CT3 and CT3

0.000809 0.00149 0.000368 0.000620 0.000171 -0.000463 Both have Bitnet (0.00102) (0.000380)** (0.000241) (0.000391) (0.000155) (0.000213)* 0.000214 -0.0000617 0.00000380 0.000214 0.00000630 -0.0000335 Sum of # of single

institution papers (0.0000486)** (0.0000270)* (0.0000163) (0.0000990)* (0.0000282) (0.0000273) # of Observations 81,900 99,225 190,575 29,295 114,345 108,900 # of Groups 5,460 6,615 12,705 1,953 7,623 7,260

Regressions include year and institution-pair fixed effects. CT1, CT2, and CT3 define the Carnegie Foundation’s rankings of research focus.

35

Table 7: Bitnet Adoption, Collaboration, and Institution-Pair Distance (1) (2) (3) (4) Main

specification Alternative distance (1)

Alternative distance (2)

Main specification with total # of

collaborations as the dependent variable

0.00203 0.00203 0.00203 0.00236 Distance is under 100 km and Both Adopted Bitnet (0.000867)* (0.000867)* (0.000867)* (0.00123)+

0.000742 0.000743 0.000533 Distance is between 100 km and 1000 km and Both Adopted Bitnet (0.000225)** (0.000225)** (0.000319)+

0.000754 0.00112 Distance is between 1000 km and 3000 km and Both Adopted Bitnet (0.000201)** (0.000284)**

0.00155 0.00154 Distance is over 3000 km and Both Adopted Bitnet (0.000293)** (0.000415)**

0.000998 Distance is between 100 km and 500 km and Both Adopted Bitnet (0.000327)**

0.000559 Distance is between 500 km and 1000 km and Both Adopted Bitnet (0.000282)*

0.000977 0.000977 Distance is over 1000 km and Both Adopted Bitnet (0.000179)** (0.000179)**

0.0000555 0.0000547 0.0000546 0.0000409 Sum of # of single institution papers (0.0000128)** (0.0000128)** (0.0000128)** (0.0000181)* # of Observations 624,240 624,240 624,240 624,240 # of Groups 41,616 41,616 41,616 41,616 Regressions include year and institution-pair fixed effects

36

Table 8: Bitnet Adoption, Collaboration, Institution-Pair Quality, and Distance (1) (2) (3) (4) (5) (6) CT1 and CT1 CT1 and CT2 CT1 and CT3 CT2 and CT2 CT2 and CT3 CT3 and CT3

-0.0185585 0.0192371 0.0047773 -0.0001092 0.0017877 -0.0019598 Distance is under 100 km and Both Adopted Bitnet (0.0040994)** (0.0017607)** (0.0012771)** (0.0020295) (0.0009123)+ (0.0012405)

-0.0002927 0.0017980 0.0004862 0.0011414 -0.0002585 -0.0003163 Distance is between 100 km and 1000 km and Both Adopted Bitnet (0.0012254) (0.0004926)** (0.0003517) (0.0005452)* (0.0002358) (0.0003274)

0.0017114 0.0001902 0.0001236 0.0005301 0.0003123 -0.0005181 Distance is between 1000 km and 3000 km and Both Adopted Bitnet (0.0011632) (0.0004408) (0.0003262) (0.0004502) (0.0002021) (0.0003200)

0.0027329 0.0030993 0.0001980 -0.0001052 0.0004064 -0.0004536 Distance is over 3000 km and Both Adopted Bitnet (0.0014988)+ (0.0006476)** (0.0004367) (0.0007454) (0.0003035) (0.0003971)

0.0002157 -0.0000576 0.0000042 0.0002141 0.0000058 -0.0000334 Sum of # of single institution papers (0.0000487)** (0.0000270)* (0.0000163) (0.0000991)* (0.0000282) (0.0000273) # of Observations 81,900 99,225 190,575 29,295 114,345 108,900 # of Groups 5,460 6,615 12,705 1,953 7,623 7,260

Regressions include year and institution-pair fixed effects. CT1, CT2, and CT3 define the Carnegie Foundation’s rankings of research focus.

Figure 1 Cumulative Bitnet Adoption over Time

0

50

100

150

200

250

300

350

1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 after1990

Num

ber o

f Uni

vers

ities

Con

nect

ed

CT1CT2

CT3

TOTAL

38

Figure 2 Bitnet Connections – 1981, 1982, and 1983