Interactive, Option-Value and Domino Network Effects in

Technology Adoption

February 2, 2006

Abstract

The network benefits of new technologies are often modeled as depending on thetotal number of users in a network. However, it is common in networks for onlysmall subsets of network users to interact. Therefore for network benefits to dependon the total network size requires network benefits which are derived outside of actualinteractions. One possibility is that network users place an option value on calling thosethey do not ultimately interact with, creating an“option network effect”. Alternatively,a “domino network effect” may occur if network users anticipate that having moreusers in the network increases the likelihood of those they want to talk with adopting.Therefore, the extent to which potential adopters value adoption by users they do notinteract with is an empirical question.

Studying network effects at the aggregate level does not permit a distinction be-tween interactive and non-interactive (option and domino) network effects. Therefore Iuse extensive micro-data on all potential adopters of a firm’s internal video-messagingsystem and their subsequent video-messaging patterns to examine the role of differentkinds of network effects in technology adoption.

The technology can also be used to watch TV. Exogenous shocks to the benefitsof watching TV are used to identify the causal (network) effect of changes in theinstalled base on adoption decisions. I find evidence that network effects are based oninteractions, and that potential adopters only react to adoption by people they wish tocommunicate with. This implies that the network benefits to adding a user for a newtechnology could be more restricted in scope than previously supposed.

1

1 Introduction

There are no network benefits to adopting a network technology such as a fax or

video-messaging unit if no-one else adopts it. Analysis of network and communication

technologies commonly assumes that all potential adopters put equal weight on the

probability and value of interacting with members of the network. This symmetry

implies that network effects only depend on network size: a convenient assumption

when modeling complex equilibria.1 Like any theoretical abstraction, however, it does

not wholly describe observed behavior. People may want to use a communication

technology to communicate with only a few other people in the network.

This paper uses an extensive dataset on the characteristics of potential adopters of

an internal video-messaging system within an investment bank and their subsequent

calls to analyze whether limited patterns of interaction are reflected in the size of

network effects. The video-messaging’s dual use for TV-watching allows an unusual

identification strategy. Some employees adopted the technology to watch one-off TV

events such as the 2002 Soccer World Cup. The reaction of other potential adopters

to this exogenous adoption identifies causal network effects.

My empirical results suggest that potential adopters only react to changes in the

installed base if they wish to communicate with the person installing. They do not

react particularly to the adoption of users with whom they do not communicate. As

not everyone communicates with each other, adding one more user to the network does

not make all other users equally more likely to adopt. If network effects are limited

to interactions in this manner, then the coordination issues associated with network

effects problems are more limited in scope than previously supposed.2

Katz and Shapiro (1994) define direct network effects as the result of an increase

in value of a good from the physical presence of more users in a network. Direct

1See for example (Katz and Shapiro 1985), (Farrell and Saloner 1986)2For example, network externalities become a less plausible justification for Universal Service in telecom-

munications provision (Riordan (2001)).

2

network effects are intuitive when they are based on interactions (“interactive” network

effects.) The extent to which a potential adopter values a technology such as a fax or

video-messaging unit depends on who they can send messages to.

Option-value and domino network effects, however, are types of direct network

effects which do not require actual interactions. Potential adopters may value the

option of talking with a wider group of users than they ultimately talk to. For example,

the option of calling the emergency service, may give phone-users utility, even if they

don’t actually call them. I call this an “option-value network effect”. Alternatively,

potential adopters may value additional subscribers even though they have no intention

of talking to them, because they value the potential for a domino effect on the adoption

decisions of those they do want to talk to. I call this anticipation a “domino network

effect”. Since option-value and domino network effects are plausible, whether network

effects are solely based on interactions is an empirical question.

Information constraints about adoption decisions and network connections could

prevent option/domino network effects from having significant impact. It is therefore

desirable to study them in a setting with few information constraints. The lack of

empirical evidence of domino or option network effects in this information-rich within-

firm setting would suggest that they are less likely in other settings where adopters

know less about other adopters.

This paper makes two contributions to the literature on network effects. First,

it provides the first empirical study of network effects to include network topology

directly in its estimation. The empirical results support recent theoretical models such

as Sundararajan (2004)’s model of multiple equilibria with local network effects and

Gale and Kariv (2003)’s models of social learning with myopia in social networks. It

also echoes the findings of Mobius (2001) on early small niche telephone markets.

Second, analyzing micro-level data on network interactions informs other empirical

work on network effects based on aggregate data. If network effects do not increase

linearly with network size, there are two possible empirical approaches. A first approach

3

is to use a non-linear functional form such as a probit or logit to ensure that network

effects decline in size with the network’s size, for example Nair, Chintagunta, and Dube

(2004). However this still requires symmetry in network relations. A second approach

is to break down national networks into smaller networks which more closely resemble

the social networks where interactions take place. For example, models of banking

technology adoption break up payment networks into geographic regions (Saloner and

Shepard (1994), Gowrisankaran and Stavins (2004)). Alternatively, as in Goolsbee and

Klenow (2002)’s models of computer adoption, network effects for friends/family can

be quantified separately.

2 Technology and Data

2.1 Technology

Installing video-messaging can improve the effectiveness of internal firm communica-

tion, by adding visual communication cues to the audio communication cues provided

by telephones (Marlow (1992)). Older video-messaging systems failed because they

were based on rarely used video-conferencing rooms. This research studies a new

video-messaging technology attached to an employee’s workstation. The end-point

technology consists of three elements: Video-messaging software; a media compressor;

and a camera fixed on top of the computer’s monitor. Using the language of Farrell and

Saloner (1985), the video-messaging technology has a “network use” and a “stand-alone

use.” The network use is television-quality video-messaging calls. The stand-alone use

is watching TV on a desktop computer.

The video-messaging technology can only be used for internal communication within

the firm. This makes it attractive for empirical studies because there are comprehensive

data on all potential adopters.

4

2.2 Firm Setting

After this bank chose this technological standard to conduct internal video-messaging,

they invested in an extensive network architecture. To encourage employee buy-in,

the bank decided to decentralize installation decisions to each employee. The bank

publicized the technology to employees and each employee decided if and when to

order a video-messaging unit from an external sales representative. The bank made

employees eligible to adopt the technology if they held a position of Associate or higher

(85 percent of full time employees). The equipment’s supplier had excess capacity,

so capacity constraints did not affect the timing of individual employee installation

decisions.

This decentralization focuses analysis on the private benefits to installation for em-

ployees, as opposed to firm-level productivity benefits. Firms find it hard to monitor

and reward improved communication (Lazear 2000). Information asymmetries there-

fore mean that employees’ installation benefits may be small relative to firm-level bene-

fits from the video-messaging system. I cannot, unfortunately, quantify these firm-level

productivity benefits.

2.3 Data

There are complete personnel records for each employee in the bank from March 2004.

Data are available for both those who adopted video-messaging and those who did not.

Employees were associated with two main products: Equities (with 60 percent of em-

ployees) and Derivatives. There were 46 further work groups within the bank, with 8 to

176 employees within them. Examples of work groups are “European equities research

for chemical industries” and “Japanese equities sales web database management.” The

precise street address is known for each employee, in the 33 cities the firm operates

in. For simplicity, cities are assigned to the British, North American, European or

Asian/Sub-Equatorial region.

5

A call database recorded each of 2.4 million calls made using video-messaging tech-

nology from January 2001 to August 2004, within the bank. The call database has two

types of call data. For two-way video-messaging calls, the database records who made

the call, to whom they made it, when they made it and how long it lasted. For one-way

TV calls, the database records who made the call, to which TV channel, when and for

how long. Further details of the data can be found in Appendix B.

Using this call data in conjunction with the personal data overcomes Rohlfs (1974)’s

warning that the studying the topology of network effects would be impossible because

“in any practical problem we could never hope to have a complete empirical list of

principal contacts.”3

3 Modeling Technology Adoption

Only installation decisions insti,t for employee i at time t are observed and not instal-

lation benefits inst∗i,t

insti,t =

1 if inst∗i,t > 0

0 if inst∗i,t ≤ 0

Each month, an employee chooses whether or not to install the technology. This

decision maximizes their utility given the set of users in the installed base, the stand-

alone TV benefit, their own idiosyncratic net benefits captured by controls Xi,t, and

unobserved heterogeneity (εi,t). I assume that each employee i takes the adoption

choices of other employees as given, but do not formally model the equilibrium selection

mechanism.

Though the firm bears the monetary costs of installing video-messaging, the em-

ployee bears non-monetary costs. These non-monetary costs are the time the employee

3Rohlfs based his pioneering work on the failure of an early video-messaging technology, the AT&TPicturephone ,to achieve critical mass.

6

spends away from their regular work while the desktop video-messaging unit is installed

and they learn how to use it. Since these time costs are sunk, I treat installation of

video-messaging as irreversible.

3.1 Interactive, Option and Domino Network Effects

This paper asks whether people’s installation decisions are only affected by those they

talk to, or whether installation decisions are also affected by people they do not talk to

but whose installation they may value because of an option or domino network effect.

Empirically, domino or option network effects cannot be identified at the aggregate

level from increases in the total number of adopters, because this would be confounded

with shocks to adoption benefits which affected everyone in the firm at that time. To

evaluate option and domino network effects it is necessary to go to the micro-level and

exploit network topology.

A “’contact” is someone employees would video-message with, if both they and the

contact installed the technology. Adoption by those who are closer in the network

topology to a potential adopter’s contacts have larger option and domino network ef-

fects, than adoption by those who are more degrees removed. An option network effect

occurs when a potential adopter places an option value on calling someone in a network

even they do not ultimately call them. Empirically, an employee calls a contact’s con-

tacts more often than a contact’s contact’s contacts or a contact’s contact’s contact’s

contacts. This suggests that when a contact’s contact adopts, this has a greater op-

tion value than adoption by someone more degrees removed from the employee in the

network. Similarly, domino network effects occur when people anticipate that others’

adoption will make those they want to talk to more likely to adopt. A contact’s con-

tacts are precisely the adopters whose adoption an employee would anticipate having

the greatest domino effect on their actual contact’s adoption decisions.

7

3.2 Other Factors Influencing Adoption

The extent to which employees value the TV-watching capacity of the video-messaging

technology varies both over time and across region. There are two types of television

employees can watch: News TV programming on CNN and CNBC, which covers stories

of interest to financial markets; and local TV programming (often non-news) broadcast

by country-specific channels. The percentage of adopters watching TV in employee i’s

region r in month t (TVr,t) captures this TV benefit. News events (such as extreme

weather in New England) are correlated with adoption in the month they occur, while

non-news locally broadcast events (such as the 2002 Soccer World Cup) are correlated

with adoption in the month prior to the month they occur. The vector TVr,t therefore

contains both the percentages of adopters watching “News TV” in month t and the

percentage of adopters watching “Local TV” in month t + 1.

Xi,t is a set of controls which captures the differences in net installation benefit as-

sociated with the employee’s observable characteristics. Xi,t contains 33 city dummies,

46 work-group dummies, and 4 hierarchical position dummies to capture variation in

net benefits for employees. The non-monetary costs of adoption for employees may

vary over time. To allow for a flexible modeling of this baseline hazard, the results re-

ported in this paper allow the time trend to vary by the bank’s two products (equities

and derivatives). These product-specific time effects are also included in Xi,t.

Not all changes to net benefits for employees are observable. For example, an

employee’s workload might vary over time, which affects the amount of free time the

employee has to install new technology. Such a shock εi,t is likely to be correlated across

employees. Therefore correlation between changes in the installed base and adoption

cannot be taken as causal. The next section discusses an identification strategy for

measuring a causal network effect.

8

4 Identification of Network Effects

I want to quantify the effect of an employee adopting on the adoption decisions of other

employees. Correlation between adoption behavior overstates the extent of network

effects, however, if all the employees receive similar unobserved shocks to their benefits

from adopting (Manski (1993)). Consider two employees who are both instructed to

install the technology by their boss; I need a clear identification strategy to avoid

interpreting the subsequent correlation in their adoption decisions as a causal network

effect.

I use shocks to the technology TV-watching benefit as an instrumental variable

to identify how changes in the installed base causally affect an employee’s adoption.

This exploits three types of variation in the data: Regional variation in the benefit

of watching TV; time variation in the benefit of watching TV; and variation in which

regions similar employees have contacts in. On average, 19 percent of employees in a

work group had an identical regional composition of contacts. Each month, the installed

base of two employees in the same work group and location will receive different shocks,

because they want to video-message with contacts in different regions.

Instrumental variables have been used to identify empirically network effects be-

fore - see for example Rysman (2004). What makes this instrumentation strategy

unusual is that the instrument is calculated each month for each employee, which al-

lows employee-level consideration of network effects. The TV benefit (TVr,t) for each

employee’s contacts, weighted to reflect the region of each contact, is the instrument

for the number of contacts who have installed the technology. Similarly, the instrument

for each employee’s installed base of contacts of contacts is the TV benefit (TVr,t) for

that employee’s contacts of contacts, weighted to reflect the region of each contact of

contact.

The Soccer World Cup in June 2002 illustrates my identification strategy. Figure 1

shows how the percentage of employees who watch local TV programming varies across

9

the US and UK in 2002. Far more employees in the UK watched the Soccer World Cup

in June 2002 than in the US. There was a spike of installations in the UK at the time of

the World Cup and also a smaller spike of installations in the US. Figure 2 shows that

the spike in installations in the US in June 2002 consists of employees in the US reacting

to TV-inspired installation by their contacts in the UK. This anecdote illustrates the

identification strategy. I do not count all earlier adoption by i’s contacts as necessarily

causing i’s installation. Instead I use variation in adoption by i’s contacts or contacts’

contacts that can be predicted by variation in the stand-alone (TV) benefit.

5 Estimation

5.1 Adopters’ and Non-Adopters’ Contacts

The identification strategy of using TV events as exogenous shocks to employee’s con-

tacts’ adoption decisions requires information on who each employees’ contacts are. It

would be ideal to establish contacts using data on existing communication networks

such as e-mail records or telephone records, but I was not able to obtain such data.

In the absence of external data, I use the last 12 months of call data from August

2003-August 2004 on video-messaging to establish contact network topology. Repeat-

ing the estimation using (March 2003-June 2003) and (March 2003-December 2004)

as representative communications periods does not significantly change the estimates.

This suggests that video-messaging did not alter the firm’s social network by reduc-

ing geographical separation. Figure 3 shows that the number of other employees each

installed employee videomessaged only increased slightly, if at all, from year to year.

The call database provides data on whom the 1,294 adopters video-messaged with,

but not on whom the 824 non-adopters would video-message with if they adopted. If

I used data only for those who adopted, my estimates would have selection bias. To

avoid this, I predict the links of non-adopters to contacts using the communication

10

patterns of similar adopters. The methodology for this is discussed in the Appendix.

5.2 AGLS Estimation

Since adoption decisions are irreversible, only observations where the employee adopts

video-messaging in that month or has not yet adopted it are used for estimation. The

installed base, however, includes all contacts’ installation decisions up to and including

month t. I use the first 2.5 years of data (January 2001- July 2003) to look at employee-

level installation decisions. The dependent variable insti,t is whether in that month

the employee used video-messaging technology for the first time. An observation is

an employee who did not adopt the technology in the previous months. Table 1 gives

summary statistics for the dependent and independent variables.

inst∗i,t = ∆1ContactsInstalledi,t +∆2ContactsContactsInstalled

i,t +λTVr,t +βXi,t + εi,t (1)

Equation (1) is estimated using Amemiya Generalized Least Squares (AGLS) es-

timators for probit with endogenous regressors (Amemiya (1974) and Newey (1987),

eq. 5.6.). I use an AGLS threshold model as opposed to a hazard model, because

the econometric theory for dealing with endogeneity is better established for AGLS

(Bijwaard 2002). Hazard models allow for heterogeneity of the baseline hazard func-

tion. Time-product group interactions capture such heterogeneity in probit estimation.

There are many different ways of specifying these interactions, and the results in Table

2 are representative of a wide variety of month/role/work-group specifications of these

baseline hazard dummies. Hazard models’ ability to explicitly deal with observations

entering and exiting the data set at different times is not crucial in this research, since

all observations are observed for each month.

Both the installed base of direct contacts (ContactsInstalledi,t ) and the installed base

of contacts’s contacts (ContactsContactsInstalledi,t ) are instrumented. The instruments

11

are the average TV benefit TVr,t for each employee’s contacts and contacts’ contacts.

This varies by the which regions these contacts and contacts’ contacts work in. The

other independent variables are TV shocks in i’s region r, and the series of dummies

in Xi,t, which controls for month-product, title, work-group, and city effects.

6 Results and Interpretation

Table 2 displays the estimates for equation (1). The first column displays probit esti-

mates of the correlation between changes in the installed base and adoption decisions.

These estimates are larger than the estimates in the second column, where the installed

base is instrumented. This suggests that without the instrumentation strategy, corre-

lated effects would wrongly be identified as network effects. The marginal estimates for

the instrumented installed base suggest that if a direct contact installs the technology

this increases the propensity of that employee to install video-messaging in that month

by 0.012. If, however, a contact’s contact installs the technology, then the point esti-

mate suggests that this only increases the propensity to install by 0.0005. This point

estimate is not significantly different from zero.

Therefore option-value and domino network effects had a negligible effect on adop-

tion of video-messaging by employees. This suggests two things. First, users had

predefined social networks and did not place an option value on calling outside these

networks. Second, adopters discounted the potential for contacts to be positively in-

fluenced by the adoption of contacts’ contacts in subsequent periods. An alternative

explanation is that users did not know that their contacts’ contacts had adopted, and

so could not adjust their behavior. If, however, users lacked this knowledge in a single

firm setting where information is readily available, then it seems likely that this ”local

network myopia” would be more widespread in other contexts.

The installation patterns shown in Figure 4 indicate no single global tipping point

for the technology. Instead, the disjointed and jerky patterns of adoption in Figure 5

12

support the empirical results in Table 2, which suggest that smaller social networks

tipped in succession.

The stand-alone benefits for TV watching are significant for an employee’s instal-

lation decision. The estimated coefficients suggest that compelling news TV has a

greater influence on employee’s installation than compelling local TV.

Space constraints mean that the coefficients for time-product, work group and city

effects are not reported in Table 1. Employees were more likely to install the technology

in Edinburgh and least likely to install in Johannesburg. Employees were most likely

to install the technology in the first two months, and were less likely to install during

August and December, presumably due to absence from the office due to vacations and

the holidays. The product-time interactions for equities and derivatives indicate that

employees in equities installed the technology both earlier and more often. Estimates

for work group effects show that employees in European equities sales adopted most

frequently, and employees who worked on North American trading floors installed the

least.

6.1 Specification Checks

The regressions reported in Table 2 are only one of many conceivable ways of estimating

equation (1). Since this is the first study of network effects using this scale of micro-

level data, it is important to verify that the estimates and interpretation are robust to

other specifications.

To check that the IV-Probit model approximates a more standard duration time

model, I estimated a reduced form hazard model, replacing the endogenous installed

base variables with their instruments. The coefficients echoed earlier results in signifi-

cance, though scale effects make direct comparison of magnitudes difficult.

Estimates of network effects which allow for within-group correlation amongst dif-

ferent work-groups and titles in the bank did not change in magnitude, though were

13

less precise. This suggests that the individual instrumentation strategy controlled for

a lot of these correlations. Without multiple adoption spells, it is not possible to use

standard panel techniques to evaluate the potential for serial correlation over time.

This paper evaluates empirical evidence of interactive, option-value and domino net-

work effects. These are direct network effects since they are associated with potential

adopters valuing the direct physical effect of having more people to talk to in the net-

work. There is a wide literature, however, which deals with indirect network effects. For

example (Gowrisankaran and Stavins 2004) study whether an increase in the number

of installers increases the level of knowledge and familiarity of an electronic payments

technology. A possible alternative explanation of the findings, therefore, could be that

people are aware of the technology when their contacts adopt the technology, but not

their contacts’ contacts.

Though this dataset is uniquely placed to distinguish between different kinds of

direct network effects because it can exploit regional variation in the TV benefit for

contacts, it is not well placed to evaluate indirect network effects, since there is no

such regional variation in sources of indirect network effects. Pure probit estimates

of correlations of adoption between workers in the same physical office, though, give

small, and not significantly different from zero point estimates of indirect network

effects, which suggests that local information flows are not a compelling alternative

explanation for my findings.

6.2 Network Effects and Policy

Network externalities occur when network participants do not internalize network ef-

fects inherent in adoption decisions. Proponents of universal service in telecommuni-

cations have cited network externalities as a justification for policy intervention. Rior-

dan (2001), for example, argues that by encouraging marginal users to adopt landline

telecommunication services, universal service regulation internalizes network externali-

14

ties. Cross-subsidization by existing subscribers to these marginal subscribers therefore

is efficient. The findings of this paper suggest, though, that for this to be a compelling

argument, there would have to be evidence that the existing subscribers would actually

interact with the marginal subscribers.

Network externalities have also been marshalled in the competition policy case sur-

rounding termination charges for cellphone networks (Armstrong 2005) and (Valletti

and Houpis 2005). Wireless networks have argued that competition authorities should

allow high termination charges because they facilitate internalization of network exter-

nalities. By offering low-cost handsets, they effectively cross-subsidized new subscribers

and internalized the network externalities inherent in having a large subscriber base.

If, however, as suggested by the results in this paper, people care more about having

the people they interact with in the network as opposed to a large subscriber base,

then existing subscribers may receive no network benefits from these new subsidized

network participants.

Furthermore if network effects are a purely interactive phenomenon then they are

confined to social networks, in which there may be social mechanisms which help par-

ticipants internalize network effects. Within these social networks, network effects are

less likely to become network externalities.

7 Conclusion

Using an extensive dataset on the adoption of an internal video-messaging system

within one firm, this paper finds that the empirical structure of network effects re-

flects interaction patterns within networks. Adoption cascades were confined to small

subsets of people who interact with each other. I find no evidence of either ‘option-

value network effects”, where users value the option of calling other users they don’t

ultimately call, or “domino network effects”, where users value adoption by friends of

friends because it has the potential to influence adoption by their friends.

15

The video-messaging technology’s stand-alone use of TV-watching provides an at-

tractive means of identifying causal network effects. This stand-alone use received a

series of exogenous shocks such as the Soccer World Cup which allow identification of

causal network effects. Studying a network technology’s diffusion within a firm is also

attractive because there are few information constraints for potential adopters. Given

this information-rich setting, the lack of empirical salience of option-value/domino net-

work effects seems likely to be repeated for other network technologies, in situations

where information about other adopters is less easily available.

Often policy assumes that network effects for communications and network tech-

nologies depend on the total number of subscribers. My estimates suggest that only

the smaller subset of people with whom a potential adopter interacts, play a signifi-

cant part in the adoption decision. If these results hold for other technologies, then

this suggests that the scope for network externalities in network technologies may be

relatively small, reducing the need for policy intervention.

16

A Appendix: Data Documentation

A.1 Call Database

Employees made 1,680,120 two-way user to user video-messaging calls. The dataset

includes only the 1,052,110 video-messaging calls where the callee accepted the call.

Accepted calls lasted on average 5 minutes 46 seconds. Calls could be made to more

than one employee at a time. Multi-party calls were simplified into their pairwise

equivalents. For example, a three-way call is treated as three calls (one call between

each two of the participants).

Employees made 754,327 one-way caller to media device calls. 741,926 of these calls

were successful and included in the dataset. The empirical work makes the distinction

between local and global news channels for users in the four regions. Global news

channels were CNN and CNBC. Local channels for Europe were ZDF (German), ARD

(German), Kanal (Swedish), ORF (Austria) and Eurosport. Local channels for Britain

were ITS, Seaports, Channel 4 and BBC. Local channels for the US were CSPAN, FOX,

NBC and CBS. Local channels for Asia were NT (Nippon TV), CATV (Japanese), TV-

Asia, and BBC 24 World Service.

A.2 Personnel database

A complete personnel database gives the rank, role, product group and street address

for both adopters and non-adopters of video-messaging, in March 2004. These per-

sonnel records omit data on employees who left the firm before 2004. Observations

of these employees’ calling patterns are excluded from the dataset. The dataset also

excludes personnel records of employees who joined the firm after January 2001. In-

cluding them in estimation would be problematic, as they did not actively choose not

to install video-messaging in earlier months. The dataset excludes the 127 employees

marked as recent recruits or trainees. The unfavorable business climate from 2001-2003

17

led the bank to make few new appointments.

Only Associates or higher were eligible to order video-messaging independently.

Therefore the dataset excludes executive secretaries to managing directors, as they

did not have discretion over the timing of their installation. The dataset excludes 18

employees in Moscow, Bangkok and Athens, since the video-messaging infrastructure

did not connect to these cities.

B Appendix: Predicting Contact Networks

This is a sparse network. For those who adopted, out of 1.5 million potential links,

there were only 23,805 actual links. Given this sparsity I use directed links to predict

contacts. A “directed link” is whether employee i initiates a video-messaging call to

employee j. Data on whether i called j or j called i are used only for link-prediction, not

for estimating installation decisions. This methodology resembles procedures used by

computer scientists to predict hypertext links between different web-pages ((Popescul

and Ungar 2003) and (Zhou and Scholkopf 2004))

I regress an indicator variable for whether or not i video-messaged j on a vec-

tor of interaction dummies for i and j’s characteristics.4 None of the coefficients are

interpreted. These interaction dummies include an indicator variable for all possible

combination of caller city and callee city, title, work group, and role in the firm. The in-

teraction variable between a caller who is a managing director in European convertible

swaps sales and a callee who is an associate in American automobile equities research

captures the likelihood of an outward link between employees with these attributes.

I use these estimates to first calculate for each employee, the integer value of the

total number of outward links Di to all employees in the firm. For non-adopters the

total number of unobserved links is Di. For adopters the extra number of unobserved

4Though the predictions use a linear probability model because there over 10,000 variables to be estimatedand 1.5 million observations, a probit model gives similar results on a sub-sample.

18

links due to non-adoption is the difference between Di and the number of contacts they

actually have. I take the most likely Di employees as the unobserved outward links for

non-adopters. For adopters, I take the additional most likely non-adopters for them to

call, until Di and their number of links are the same. I symmetrize this outward link

network so that if an outward link is projected in either direction, then I consider two

employees contacts.

To evaluate how well this procedure predicts contacts, I redid my empirical analysis,

using data for adopters from August 2001-August 2002 to predict contacts for those who

adopted in August 2002-August 2003. The results suggested that contacts are predicted

correctly approximately 60 percent of the time. This suggests that time to adoption

is not correlated with the social network’s size or nature. The lack of external call

data prevents easy validation of the assumption that non-adopters have identical social

networks to adopters and introduces measurement error. When I exclude non-adopters

from the regression analysis, I get larger estimates of network effects for contacts than

before, but the estimates for contacts’ contacts still remain insignificantly different

zero. This suggests that the papers’ conclusions are unchanged by any measurement

error prediction introduces.

References

Amemiya, T. (1974, July). The nonlinear two-stage least-squares estimator. Journal

of Econometrics 2, 105–110.

Armstrong, M. (2005, May). Competition in two-sided markets. Industrial Organi-

zation 0505009, Economics Working Paper Archive EconWPA.

Bijwaard, G. (2002, April). Instrumental Variable Estimation for Duration Data: A

Reappraisal of the Illinois Reemployment Bonus Experiment. Technical Report

0204001, Economics Working Paper Archive at WUSTL.

Farrell, J. and G. Saloner (1985). Standardization, compatibility, and innovation.

19

RAND Journal of Economics 16, 70–83.

Farrell, J. and G. Saloner (1986, December). Installed base and compatibility: In-

novation, product preannouncements, and predation. American Economic Re-

view 76 (5), 940–955.

Gale, D. and S. Kariv (2003, 11). Bayesian learning in social networks. Games and

Economic Behavior 45 (2), 329–346.

Goolsbee, A. and P. Klenow (2002, October). Evidence on learning and network

externalities in the diffusion of home computers. Journal of Law and Eco-

nomics 45 (2), 317–344.

Gowrisankaran, G. and J. Stavins (2004). Network externalities and technology adop-

tion: lessons from electronic payments. RAND Journal of Economics 35 (2),

260–276.

Katz, M. and C. Shapiro (1994, Spring). Systems competition and network effects.

Journal of Economic Perspectives 8 (2), 93–115.

Katz, M. L. and C. Shapiro (1985). Network externalities, competition, and com-

patibility. American Economic Review 75 (3), 424–40.

Lazear, E. (2000, May). Power of incentives. American Economic Review 90 (2),

410–414.

Manski, C. (1993, July). Identification of endogenous social effects: the reflection

problem. Review of Economic Studies 60 (3), 531–542.

Marlow, E. (1992). Corporate television programming: applications and technologies.

Knowledge Industry Publications, Inc.

Mobius, M. (2001). Death through success: The rise and fall of local service compe-

tition at the turn of the century. Harvard Working Paper.

Nair, H., P. Chintagunta, and J.-P. Dube (2004). Empirical analysis of indirect net-

work effects in the market for personal digital assistants. Quantitative Marketing

and Economics 2 (1), 23–58.

20

Newey, W. (1987). Simultaneous estimation of limited dependent variable models

with endogenous explanatory variables. Journal of Econometrics 36, 231–250.

Popescul, A. and L. H. Ungar (2003). Statistical relational learning for link predic-

tion. IJCAI03 Workshop on Learning Statistical Models from Relational Data.

Riordan, M. H. (2001). Handbook of Telecommunications Economics, Chapter 10:

Universal Residential telephone service. Amsterdam: Elsevier Science.

Rohlfs, J. (1974, Spring). A theory of interdependent demand for a communications

service. Bell Journal of Economics 5 (1), 16–37.

Rysman, M. (2004, April). Competition between networks: A study of the market

for yellow pages. Review of Economic Studies 71 (2), 483–512.

Saloner, G. and A. Shepard (1994, Autumn). Adoption of technologies with network

effects: An empirical examination of the adoption of automated teller machines.

RAND Journal of Economics 26 (3), 479–501.

Sundararajan, A. (2004, July). Local network effects and network structure. Mimeo.

Valletti, T. and G. Houpis (2005). Mobile termination: What is the right charge?

Journal of Regulatory Economics 28 (3), 235–258.

Zhou, D. and B. Scholkopf (2004). A regularization framework for learning from

graph data. In Workshop on Statistical Relational Learning, International Con-

ference on Machine Learning.

21

Figure 1: Relationship between New Installations and TV-watching in the US and UK, 2002

Figure 2: Relationship between New Installations and US employees having any UK contacts

22

Figure 3: Stability in Video-messaging Behavior over Time

23

Figure 4: Increase in Network Adoption by Function

Figure 5: Increase in Network Adoption by Work Group

24

Table 1: Variable Description and Summary Statistics

Label Variable Description Mean S.D. Range

insti,t Whether employee i makes avideo-messaging call for the firsttime in month t

0.041 0.169 0-1

ContactsInstalledi,t Number of i’s contacts who

have installed video-messaging bymonth t

2.550 3.040 0-22

ContactsContactsInstalledi,t Number of i’s contacts of con-

tacts who have video-messagingby month t

4.015 3.678 0-120

TV Newsr,t Percent of installers in i’s region

watching TV news in month t0.253 0.211 0-1

TV Localr,t Percent of installers in i’s region

watching local TV in month t + 10.085 0.132 0-1

25

Table 2: Interactive vs. Option/Domino Network EffectsDependent Variable: Whether employee participated in a video-messaging call for the first timein that month from January 2001-July 2003.Installation is permanent - only those who have not installed previous included as dependentvariables.

Sub-set of observations used in regressionsVariable Description Probit Probit-IVInteractive Network Effects

Installed contacts (IV)0.0241 0.0124

(0.0002)∗∗∗ (0.0025)∗

Option/Domino Network effects

Installed contacts’ contacts (IV)0.0072 0.0005

(0.0013∗∗) (0.0156)

Variables for TV Benefit, WrkGrp, location, product-time interactions included

Observations 35,207 35,207Log-Likelihood -2640.1789 -2843.2567

AGLS estimates of marginal effects for probit with endogenous regressors.Instruments are percentage of installers watching news-TV and local-TV in contacts’ regions inthat month and the next month, weighted to reflect regional distribution of contacts for eachemployee.Standard errors in parenthesis below.* Significant at the 10 percent level ** 5 percent level *** 1 percent level

26