The Effect of Content on Global Internet Adoption and the Global

MANAGEMENT SCIENCEArticles in Advance, pp. 1–22ISSN 0025-1909 (print) � ISSN 1526-5501 (online) http://dx.doi.org/10.1287/mnsc.2013.1875

© 2014 INFORMS

The Effect of Content on Global Internet Adoptionand the Global “Digital Divide”

V. Brian ViardCheung Kong Graduate School of Business, 100738 Beijing, China, [email protected]

Nicholas EconomidesStern School of Business, New York University, New York, New York 10012; and Haas School of Business,

University of California, Berkeley, Berkeley, California 94720, [email protected]

Acountry’s human capital and economic productivity increasingly depend on the Internet as a result of itsexpanding role in providing information and communications. This has prompted a search for ways to

increase Internet adoption and narrow its disparity across countries—the global “digital divide.” Previous workhas focused on demographic, economic, and infrastructure determinants of Internet access that are difficultto change in the short run. Internet content increases adoption and can be changed more quickly; however,the magnitude of its impact, and therefore its effectiveness as a policy and strategy tool, has until now beenunknown. Quantifying the role of content is challenging because of feedback (network effects) between contentand adoption: more content stimulates adoption, which in turn increases the incentive to create content. Wedevelop a methodology to overcome this endogeneity problem. We find a statistically and economically signif-icant effect, implying that policies promoting content creation can substantially increase adoption. Because it isubiquitous, Internet content is also useful to affect social change across countries. Content has a greater effect onadoption in countries with more disparate languages, making it a useful tool to overcome linguistic isolation.Our results offer guidance for policymakers on country characteristics that influence adoption’s responsivenessto content and for Internet firms on where to expand internationally and how to quantify content investments.

Keywords : Internet; technology adoption; economic development; two-sided markets; network effects;technology diffusion; digital divide; language

History : Received November 2, 2011; accepted October 5, 2013, by Lorin Hitt, information systems. Publishedonline in Articles in Advance.

1. IntroductionThe number of Internet users has exploded since itscommercialization in the early 1990s. From approxi-mately 10.1 million users in early 1992, the Internethad expanded to almost 1.6 billion by 2009.1 However,this growth has been very uneven across countries,with penetration rates varying from 90% to nearly 0%(see Figure 1). This global “digital divide” is of con-cern because Internet access is increasingly importantfor economic productivity and a well-informed citi-zenry as more information is accessed online.2 As aconsequence, there is a large literature examining eco-nomic and social determinants of cross-country Inter-net adoption but focuses almost exclusively on factors

1 World Development Indicators Database (2010b).2 For an aggregate study on the link between the Internet and pro-ductivity, see Litan and Rivlin (2001); compare with a critique byGordon (2000). Industry-specific studies include Goolsbee (2002)in health insurance and Scott Morton et al. (2001) in car retail.The International Telecommunication Union (ITU) (1999) providesa policy perspective on its economic and social role.

that are fixed in the short run. We focus on a factorthat can be changed quickly: Internet content.

It is well understood that more Internet content ina language will lead to more adopters who use thatlanguage. As a United Nations (UN) report asserts,“Availability of content, in an appropriate languagealso affects the diffusion of the Internet. After all ifyou cannot find content in your language and youdo not read other languages, how can you use theInternet?” (ITU 1999, p. 3, italics in original). Whatis not known is the magnitude of the effect of con-tent on adoption. This has important policy impli-cations. Because content is more easily altered thaneconomic, educational, or infrastructure conditions, itoffers governments and nongovernmental organiza-tions (NGOs) a means to more quickly influence Inter-net diffusion. The exact magnitude of the effect is alsorelevant for Internet firms that rely on a user base foradvertising or subscription revenue. It is importantin evaluating the trade-off between investing in con-tent creation to build the user base indirectly versusmarketing efforts to attract new users directly. Our

1

Dow

nloa

ded

from

info

rms.

org

by [

218.

189.

95.1

0] o

n 09

Jul

y 20

14, a

t 20:

46 .

For

pers

onal

use

onl

y, a

ll ri

ghts

res

erve

d.

Viard and Economides: The Effect of Content on Global Internet Adoption and the Global “Digital Divide”2 Management Science, Articles in Advance, pp. 1–22, © 2014 INFORMS

estimates quantify the effectiveness of this “build con-tent and they will adopt” strategy and allow us tooffer some guidance to Internet firms in making suchinvestments.

If content sufficiently stimulates adoption, the abil-ity to target content by language suggests a usefulstrategy to narrow the global digital divide. The UNhas suggested content’s role in reducing this divide,stating, “The dominance of European languages haslimited the spread of Internet use by excluding thosenot fully literate in those languages” (Mutume 2006,p. 14). It would also suggest that content productionis an effective strategy for firms to expand their userbase. However, the question remains how effectivelycontent stimulates adoption.

Content has a statistically and economically signif-icant effect on adoption, implying that it is an effec-tive policy and strategic tool. Our estimates explic-itly recognize language as the conduit from content toadoption, confirming that creating content in under-served languages is an effective policy to address theglobal digital divide. We quantify content’s effect onadoption in four different ways, but all indicate alarge effect. First, we find an elasticity of adoptionwith respect to content of 0.31—about three-fourthsthe price elasticity of adoption. Second, a country onestandard deviation above the mean level of relevantcontent has an adoption rate 2.0 percentage points or20% higher than the mean adoption rate of 9.9 per-centage points in the sample. Third, the magnitudeof content’s effect is about one-third that of the grossdomestic product (GDP) (the most significant driver)and is stronger or of similar strength to that of othereconomic, infrastructure, and demographic factorsthat significantly affect adoption. Fourth, the annualrate of content creation in our sample increased adop-tion by 6.0% to 7.8% annually.

To further inform policy-making and firm strate-gies, our model can identify country characteristicsthat affect adoption’s sensitivity to content. Con-tent has greater influence in countries with betterinfrastructure, as measured by the extensiveness ofthe domestic phone system and international gate-way speeds. This suggests Internet content providerswishing to access international markets should tar-get such countries, and infrastructure investment is ameans for governments to stimulate adoption. Con-tent also has more influence in countries with weakerintellectual property protection, consistent with lesscostly and more widely available content. Thus, con-tent providers who can sufficiently protect their con-tent will experience greater uptake in internationalmarkets with weaker protections, and governmentssetting copyright policies face a trade-off betweendynamic incentives to create content and its usageonce created.

We also identify an important role for the Inter-net in overcoming linguistic isolation. Content affectsadoption more in countries with more disparate lan-guages. This suggests that creating content targeted atpopulations that speak languages uncommon in theirsurroundings may reduce their isolation. The pre-dominance of English-language Internet content hasbeen cited as an important dimension of inequalitybetween social and linguistic groups (see DiMaggioet al. 2004). This result parallels that of Sinai andWaldfogel (2004), who find that the Internet helpsovercome racial isolation in the United States. Thisalso suggests an opportunity for Internet firms to tar-get such populations.

Internet service is a two-sided market—user adop-tion depends on content availability, and vice versa.This feedback makes it difficult to empirically iso-late content’s effect on adoption. Estimating the causaleffect of content is further complicated by the likelypresence of unobserved country-specific factors thatdrive both content production and adoption. In par-ticular, populations of countries with a high desire forInternet usage for unobserved reasons may also cre-ate more content for the same reasons. We develop amethodology to control for the endogeneity of con-tent with respect to the installed base of Internet userswhile controlling for a host of factors known to affectadoption. This approach also helps eliminate sourcesof spurious correlation that explain both content andadoption. To further reduce the possibility of spuri-ous correlation, we include an extensive set of fixedeffects in our estimation.

Our identification approach uses “large”-countrycontent as an instrument for relevant content whenestimating the effect of content on adoption for“small” countries, where we define small and largebased on the number of potential adopters in a coun-try. We argue and provide empirical evidence thatcontent production by large countries is exogenousto Internet adoption in small countries. We assumethat potential adopters value most content in theirown language. Therefore, to identify content relevantto a country’s potential adopters, we use the distri-bution of their language usage and measure contentbased on the storage capacity of computers hostingInternet content in those languages. Previous paperssupport our use of language to define Internet con-tent relevance. ITU (1999) uses aggregate Web-trafficstatistics to show that language determines the rele-vance of Internet content. Gandal (2006) shows thatlanguage usage heavily influences the languages ofwebsites visited during individual-level browsing andprovides evidence that English-language dominancein Internet content may continue based on bilingualusers’ online behavior.

Dow

nloa

ded

from

info

rms.

org

by [

218.

189.

95.1

0] o

n 09

Jul

y 20

14, a

t 20:

46 .

For

pers

onal

use

onl

y, a

ll ri

ghts

res

erve

d.

Viard and Economides: The Effect of Content on Global Internet Adoption and the Global “Digital Divide”Management Science, Articles in Advance, pp. 1–22, © 2014 INFORMS 3

Figure 1 Internet Penetration Across Countries in 2008

0

10

20

30

40

50

60

70

80

90

100In

tern

et p

enet

ratio

n (%

)

Icel

and

Ber

mud

a

Ger

man

y

Aus

tria

Hon

g K

ong

Irel

and

Jam

aica

Lith

uani

a

Mal

ta

Por

tuga

l

Col

ombi

a

New

Cal

edon

ia

Cos

ta R

ica

Ecu

ador

Tun

isia

Vie

tnam

Gre

nada

Dom

inic

an R

epub

lic

Kyr

gyz

Rep

ublic

Hon

dura

s

Pak

ista

n

Sud

an

Ken

ya

Uga

nda

Phi

lippi

nes

Nam

ibia

Erit

rea

d'Iv

oire

Kiri

bati

Nep

al

Tur

kmen

ista

n

Gui

nea

Cen

tral

Afr

ican

Rep

ublic

Côt

e

Source. World Development Indicators Database (2010a). Internet penetration (fraction of population with Internet access) for 197 countries is sorted fromhighest to lowest penetration. Not all country names are displayed due to a lack of space.

Using large-country content to instrument small-country-relevant content also helps eliminate sourcesof spurious correlation that might bias our results.Instrumenting sterilizes the estimates from unob-served factors that drive both content and adoptionwithin each small country. Any remaining spuriouscorrelation must be across small and large countries.We include an extensive set of fixed effects that makesthis unlikely. Country fixed effects remove country-specific time-constant unobservables, and year fixedeffects eliminate time-specific unobservables operat-ing across the small and large countries. Finally, lan-guage fixed effects remove language-specific unob-servables that drive adoption in small countries andcontent production by large countries with whichthey are instrumented.

Our results have implications for government poli-cies that affect Internet content production. Govern-ments directly create content, so much so that itsquantity has raised concerns about effective archiv-ing.3 Much of this is generated as a part of regular

3 See Carrell (2009). Also refer to National Archives: The Challengeof Electronic Records Management, Before the Subcommittee on Govern-ment Management, Information, and Technology, Committee on Govern-ment Reform (1999) (statement of L. Nye Stevens, Director, FederalManagement and Workforce Issues, General Government Division).

government business, but some is specifically tar-geted at underserved languages. Qatar’s govern-ment is developing digital archives of major Arabictexts to increase Arabic content (Pratap 2010). NGOshave also targeted underserved languages. One NGO,Canada’s International Development Research Cen-tre (IDRC), described content development efforts inUganda as “increasingly important and valuable tothe market” (Uganda Communications Commission2005, p. 27). Arab countries working with NGOshave established rewards for high-quality Arabic con-tent and encouraged collaboration between universi-ties and research centers to produce content.4 Otherefforts are targeted at underserved populations. Inthe United States, the Federal Communications Com-mission announced in late 2011 a policy to promotejob and education information relevant to householdsthat had not yet adopted broadband (Stelter 2011).

Perhaps more important than governments’ directcontent creation are the indirect effects of their poli-cies. Decisions on Internet technical standards havefar-reaching effects on content creation. Originallyarchitected in English, the Internet does not eas-ily accommodate developing or finding content in

4 The Emirates Center for Strategic Studies and Research (2008).

Dow

nloa

ded

from

info

rms.

org

by [

218.

189.

95.1

0] o

n 09

Jul

y 20

14, a

t 20:

46 .

For

pers

onal

use

onl

y, a

ll ri

ghts

res

erve

d.


languages using non-Latin characters. In response,the Internet Governance Forum approved a multi-year effort to allow non-Latin characters in websiteaddresses.5 Similarly, many Internet browsers will notproperly display Arabic content because of a lackof agreement among Arab countries on a uniformformat.6 Our results also offer guidance to firms inevaluating their content investments. The estimatesof adoption’s sensitivity to content can be used toquantify the trade-off between marketing investmentsto increase usage directly and content investments toincrease it indirectly. These are especially useful forfirms relying on user-generated content to evaluateinvestments in customer acquisition.

2. Identification StrategySimply relating adoption and content will overstatecontent’s effect because it will conflate content’s effecton adoption with the feedback effect of adoption oncontent. At the same time, unobserved heterogeneityacross countries may introduce spurious correlation.Our identification approach addresses both of theseissues.

To disentangle content’s effect on adoption, we usethe subset of content created by large (in terms ofnumber of language users but not necessarily geo-graphic area) countries as an instrument for rele-vant content when estimating the effect of content onadoption for small countries only.7

Identification relies on the assumption that con-tent creation by large countries is exogenous to adop-tion in small countries. Intuitively, we assume thatthe number of adopters in small countries is smallenough that content creators in the large countriesfocus only on the number of adopters in the largecountries.8 That is, we assume that content created

5 See Waters (2006). Methods of using non-Latin characters in web-site addresses emerged in 2003 but without standardization or offi-cial approval.6 The Emirates Center for Strategic Studies and Research (2008).7 We use the number of language users as a measure of potentialadopters in that language. We do not use the actual number ofadopters using the language because it is endogenous.8 Two previous papers use related identification schemes. Gowri-sankaran and Stavins (2004) estimate network effects in adoptionof the automated clearinghouse system (ACH) by clusters of U.S.banks. One method to isolate the network effect from a strong localpreference for ACH is to examine the effect of adoption by smallbranches of large banks on the adoption decisions of rival banks inthe same local markets. Identification relies on the fact that a bankmust implement ACH at all its branches simultaneously. Shriveret al. (2013) examine the effect of online content production on theformation of social ties and Internet usage among surfers. Theyuse exogenous wind speed changes as an instrument to break thefeedback loop between social ties and production of user-generatedcontent.

in large countries is relevant to and therefore con-sumed by those in small countries who share thesame language, even though the latter are typicallyignored by the content creators when choosing theprofit-maximizing level of content.

We justify this assumption based on two relatedarguments. First, even if content creators can col-lect revenues from users in small countries, theserepresent such a small fraction that they do notaffect content creation decisions. Second, it is fre-quently difficult to collect revenues from users insmall countries because of legal impediments, highfixed costs of collecting subscription fees across coun-try boundaries, and difficulty in targeting onlineadvertising to these small groups. Relevancy of con-tent to users in small countries has been reportedto create a financial conundrum for major contentproviders such as Facebook and YouTube, whichmust provide the additional bandwidth to sup-port these users despite difficulty in collecting rev-enues.9 Besides these qualitative arguments, we pro-vide quantitative evidence that this assumption holdswhen we present our data and results. At the sametime, the instrument’s inclusion restriction is metbecause relevant content consumed in small coun-tries is affected by large-country content, given itsubiquity. We must omit the large countries from esti-mation to maintain exogeneity. Therefore, our resultsmay not extrapolate to large countries; however,the combined population of our small countries is2.0 billion.

We assume that an Internet user is most interestedin content of her primary language and define smalland large countries accordingly. We identify countriesthat comprise a large percentage of the worldwideusers of a language as large. The remaining countrieswith small populations using that language we iden-tify as small. Identification requires languages with askewed distribution of users—a few countries repre-sent most of the worldwide users, and many countrieshave a small percentage of the users. This providesa large number of observations while satisfying theexogeneity assumption.

For each small country, relevant content includesworldwide content (produced by both small and largecountries) in the language(s) of its population. Sincea small country’s population may use a mixture of

9 Stone and Helft (2009) note that “Facebook is in a particularly dif-ficult predicament. Seventy percent of its 200 million members liveoutside the United States, many in regions that do not contributemuch to Facebook’s bottom line,” and quotes the chief executiveofficer of a San Diego-based video-sharing site who says of its usersin Africa, Asia, Latin America, and Eastern Europe: “They sit andthey watch and watch and watch. The problem is that they are eat-ing up bandwidth, and it’s very difficult to derive revenue from it.”

Dow

nloa

ded

from

info

rms.

org

by [

218.

189.

95.1

0] o

n 09

Jul

y 20

14, a

t 20:

46 .

For

pers

onal

use

onl

y, a

ll ri

ghts

res

erve

d.


languages, we construct a weighted-average mea-sure of the relevant content based on the fractionusing each language. For example, in Belgium, 38%of people speak Dutch, 33% French, 9% Walloon, 9%Flemish, 5% Limburgish, and 2% Italian as their pri-mary language.10 Relevant content for Belgium wouldequal 0.38 times the worldwide quantity of Dutchcontent plus 0.33 times the worldwide quantity ofFrench content and so on. As a by-product, the lan-guage usage distributions provide significant cross-sectional variation in relevant content. The instrumentfor each small country is constructed analogously—a weighted-average of large-country content basedon the language distribution of the small country’spopulation.

Identification may also be affected by the presenceof country-level unobservables that affect both con-tent production and adoption but are separate fromthe indirect network feedback loop. If unaccountedfor, these will induce correlation between relevantcontent and the error in our adoption equation andbias the coefficients on relevant content and the con-trol variables. Our instrumenting approach combinedwith the large number of controls we include makesthis unlikely. In our estimation we include year andcountry fixed effects in addition to a wide range ofcontrol variables. This means that any unobservedfactors cannot be common to countries within thesame year or result from country-specific character-istics. Thus, our estimation approach is robust to,among others, country-specific policies that promoteadoption or content production, changes in standardsthat promote adoption or content production Internet-wide, and secular trends in adoption or content pro-duction due to factors such as technological changesin storage or transmission of data.

The content variable, once instrumented, will onlybe correlated with the adoption error if the unob-served factors drive both large-country content pro-duction and small-country adoption. Moreover, ourinstrumenting approach groups small and large coun-tries based on the distributions of language usageacross countries. Bias would require that adoptionand content production be correlated within thesegroupings but in a way such that there is no commoncorrelation across the small countries and no commoncorrelation across the large countries because thesewould be absorbed by the year fixed effects. Impor-tantly, these languages, and therefore the set of largecountries within each group, differ for each smallcountry according to its language distribution, whichis exogenous with respect to Internet adoption and

10 The remaining 4% use languages that each represents less than1% of Belgium’s population.

content. Since the grouping of a small adopting coun-try with large content producers is mediated throughlanguage, a possible way for bias to enter is throughlanguage-specific unobservables. To address this, weshow that our results are robust to adding languagefixed effects.

3. Econometric ModelWe model the simultaneous determination of a coun-try’s content production in a language and adoptionin that country by people using that language. Thefraction of a language’s users adopting the Internetin a country is a function of the worldwide contentavailable in that language since Internet content isaccessible anywhere.11 Internet content produced bya country in a language is a function of the world-wide adopters using that language since the contentis accessible worldwide.12

Let i = 1121 0 0 0 1 I index countries, j = 1121 0 0 0 1 J lan-guages, and t = 1121 0 0 0 1 T years. We model adoptionand content production according to the simultaneoussystem of stochastic equations:

Adoptersijt

Usersij= �AXA

it +�AZAi +�A

t + �Ai

+�AI∑

k=1

Contentkjt + �̃Aijt1 (1a)

Contentijt = �CXCit +�CZC

i +�Ct + �C

i

+�CI∑

k=1

Adopterskjt + �̃Cijt1 (1b)

where Adoptersijt is the number of Internet adopterswho use language j in country i at time t, Usersij isthe number of users of language j in country i thatdoes not vary over time in our data, and Contentijt isthe content available in language j at time t producedby country i. XA

it and XCit include possibly overlapping

sets of time-varying factors affecting Internet adop-tion and content, whereas ZA

i and ZCi are the same for

time-constant factors.The parameters to be estimated are 8�A1�A1�A1

�C1�C1�C9. The latent year effects, �At and �C

t , captureunobserved time-specific factors affecting adoptionand content, respectively. The latent country effects,�Ai and �C

i , are time-invariant random variables thatcapture unobserved factors affecting adoption andcontent, respectively. We discuss the statistical prop-erties of these fixed effects below. The error terms,

11 We control for government restrictions on Internet access in ourestimation.12 As explained below, the content is not necessarily hosted on acomputer located physically within the country.

Dow

nloa

ded

from

info

rms.

org

by [

218.

189.

95.1

0] o

n 09

Jul

y 20

14, a

t 20:

46 .

For

pers

onal

use

onl

y, a

ll ri

ghts

res

erve

d.


�̃Aijt and �̃C

ijt, are independently and identically dis-tributed across countries, languages, and time peri-ods. We expect �A,�C > 0. This specification assumesthat content’s effect on adoption is the same acrosslanguages. Although in theory we could allow theeffect to vary by language, in practice there are insuf-ficient data to identify this.

If XAit and XC

it each contain at least one variablenot contained in the other, a system method of esti-mation for (1a) and (1b) may be feasible. Unfortu-nately, we do not have available any variables thoughtto affect content but not adoption. Instead, we esti-mate (1a) using limited-information estimation meth-ods and use Equation (1b) to inform our search foran appropriate instrument for the content variable inEquation (1a).

For a set of the most frequently used languages,JF , we divide countries into large 4i ∈ IL5 and small4i ∈ IS5 based on the number of language userswith I = 8IS1 IL9. Our identification assumption isthat content production by large countries is unaf-fected by adoption in small countries. More formally,∑

k∈ILAdopterskjt ≈

∑Ik=1 Adopterskjt so that

Contentijt = �CXCit +�CZC

i +�Ct +�C

i +�C∑

k∈IL

Adopterskjt

+ �̃Cijt ∀ i∈ IL1∀ j ∈ JF 0 (1b’)

If Equation (1b’) holds, then∑

k∈ILContentkjt

(large-country content) is a valid instrument for∑I

k=1 Contentkjt (worldwide relevant content) in Equa-tion (1a) estimated on the set of small countries:

Adoptersijt

Usersij= �AXA

it +�AZAi +�A

t + �Ai

+�AI∑

k=1

Contentkjt + �̃Aijt ∀ i ∈ IS 0 (1a’)

To preserve degrees of freedom, we use only theworld’s most pervasive languages to construct theinstrument. Enlarging this set involves a trade-offbetween decreasing available data and increasing theinstrument’s power. Including an additional languagereduces the available data because large content pro-ducers for that language must be excluded to main-tain the exogeneity assumption. On the other hand,it increases the instrument’s power since more lan-guages mean the instrument is more highly correlatedwith the small countries’ consumed content. In §5 weempirically assess the exogeneity and relevance con-ditions for our instrument. Our choice of languagesfor the instrument is discussed in §4.

Since we observe only the aggregate number ofInternet adopters in each country, we transform Equa-tion (1a’) into one which we can estimate. Multiplying

through by the number of users of language j andthen summing across all languages we obtain

J∑

j=1

Adoptersijt

= 4�AXAit +�AZA

i +�At 5

J∑

j=1

Usersij +J∑

j=1

6Usersij4�Ai + �̃A

ijt57

+�AJ∑

j=1

[

UsersijI∑

k=1

Contentkjt

]

∀ i∈ IS 0 (2)

Since∑J

j=1 Usersij = Populationi,

∑Jj=1 Adoptersijt∑J

j=1 Usersij=

J∑

j=1

Adoptersijt

Populationi

= Penetrationit1 (3)

where Penetrationit is the fraction of country i’s popu-lation that have adopted the Internet at time t, whichwe observe. Dividing both sides of Equation (2) byPopulationi, we get13

Penetrationit

=�AXAit +�AZA

i +�At +�A

i

+�A

∑Jj=1 6Usersij

∑Ik=1 Contentkjt7

Populationi

+�Ait ∀ i∈ IS 0 (4)

We call the weighted-average measure of content inEquation (4) the relevant content for small country iin year t:

relconit =

∑Jj=1 6Usersij

∑Ik=1 Contentkjt7

Populationi

1 i ∈ IS 0 (5)

This includes content produced worldwide in eachof the languages used within country i weightedby the proportion of the population using that lan-guage. Worldwide content includes content producedin country i as well as content in relevant languagesproduced outside of it. The instrument for relevantcontent is defined similarly but includes only contentproduced by large countries:

instrumentit

=

∑Jj=1 6Usersij

∑

k∈ILContentkjt7

∑Jj=1 Usersij

1 i ∈ IS 0 (6)

We denote a country-time period unobservable thataffects adoption in country i at time t by �A

it . We

13 Transforming Equation (1a’) into Equation (4) introduces country-level heteroskedasticity since the distribution of languages variesacross countries. This is difficult to accommodate in the Hausman–Taylor estimates (Hausman and Taylor 1981). However, our fixed-effects estimates in Table 5, which are consistent, are robust togeneral forms of heteroskedasticity and yield similar results to theHausman–Taylor estimates.

Dow

nloa

ded

from

info

rms.

org

by [

218.

189.

95.1

0] o

n 09

Jul

y 20

14, a

t 20:

46 .

For

pers

onal

use

onl

y, a

ll ri

ghts

res

erve

d.


distinguish, on a priori grounds, columns of X andZ that are asymptotically uncorrelated with �A

i fromthose that are not so that our assumptions about therandom terms in the model are

E4�Ait 5= E4�A

i �XA1it1Z

A1it5= 0

but

E4�Ai �XA

2it1ZA2it5 6= 01

Var4�Ai �XA

1it1ZA1it1X

A2it1Z

A2it5= �2

�1

Cov4�Ait1�

Ai �XA

1it1ZA1it1X

A2it1Z

A2it5= 01

Var4�Ait + �A

i �XA1it1Z

A1it1X

A2it1Z

A2it5= �2 = �2

� +�2�1

Corr4�Ait + �A

i 1�Ais + �A

i �XA1it1Z

A1it1X

A2it1Z

A2it5

= �= �2�/�

20

(7)

This error structure allows the Hausman and Tay-lor (1981; HT hereafter) estimator. HT refer to XA

1it astime-varying exogenous, XA

2it as time-varying endoge-nous, ZA

1it as time-invariant exogenous, and ZA2it

as time-invariant endogenous variables. We discussthese classifications and justify our use of the HT esti-mator vis-à-vis a fixed-effects and random-effects esti-mator in §5.

The exogeneity of large-country content suggestsa simpler estimation approach: regress small-countryadoption on large-country content. However, this hasan important practical limitation. Large countries forall included languages must be dropped from theanalysis to meet the exogeneity condition. To main-tain a sufficient sample size, not all languages can beincluded, and therefore it is not possible to include allexternal (outside the country) content for each smallcountry. This introduces an omitted variable bias, thesign of which depends on the correlation betweenthe included and excluded external content net ofthe effect of the control variables. Although the signof this correlation is theoretically indeterminate, it islikely negative since more included external contentfor a small country implies less excluded content.Our instrumenting strategy frees us from producinginconsistent estimates because now we can include allcontent (internal and external) for each small coun-try while solving the endogeneity problem. We stillmust exclude large countries from the analysis, butthis can be minimized because we need only enoughincluded languages to adequately satisfy the instru-ment’s inclusion restriction.

Ideally we would estimate adoption’s effect of con-tent using a similar strategy—use adoption rates inlarge countries as an instrument for small-countryadoption rates when predicting small-country content

production. This is not possible for two reasons—one methodological and the other practical. Large-country adoption rates as an instrument fails theexclusion restriction. Since content is ubiquitous, largeand small country content are substitutes. We alsoface a practical problem: we do not observe language-specific adoption rates. Therefore, only time-seriesvariation would identify adoption’s effect on content.

4. DataOur sample includes data on 176 small countriesand 31 large countries from 1998 to 2004.14 Table 1contains summary statistics on the main variables.Online Appendix B contains more details on vari-ables and their sources. (Online appendices avail-able at http://english.ckgsb.edu.cn/faculty_content/brian-viard.)

Internet Users: Our dependent variable is the frac-tion ofcountry i’s population with Internet access attime t (see Figure 2 for 2004 small-country adoptionrates in the sample). The ITU collects these data anddoes not distinguish speeds or modes of access. Dur-ing our sample years, virtually all access was throughone of three modes: narrowband (or dial-up) accessthrough a phone line, broadband (or digital subscriberline) access through a phone line, and broadbandaccess through cable lines. The ITU data measure allInternet users regardless of location.15 Unfortunately,the data do not allow us to control for access speedsince content may drive adoption of higher-qualityaccess. During our sample period, most relevant con-tent is textual minimizing this concern;16 however, ifnontextual content were significant, it would bias our

14 Online Appendix A contains a list of the small countries.These include 12 non-self-governing territories: overseas territo-ries (Bermuda), overseas regions (French Guiana, Guadeloupe,Martinique), overseas collectivities (French Polynesia, Mayotte),sui generic collectivities (New Caledonia), special administrativeregions (Hong Kong, Macao), disputed territories (PalestinianWest Bank and Gaza), unincorporated organized commonwealths(Puerto Rico), overseas departments (Reunion), and unincorporatedorganized territories (Guam, U.S. Virgin Islands). We include thesebecause we believe their social and economic conditions differ sub-stantially enough from their governing countries that they repre-sent independent observations. Content measures are not availablefor Hong Kong, Macao, and Mayotte, so they do not identify theeffect of content.15 ITU’s data distinguish between “estimated Internet users” and“Internet subscribers.” Users of Internet cafés, for example, wouldbe included in the former, which is our variable, but not in thelatter.16 Of file space for publicly available Internet data in 2003,text-related files (Excel, text, Word, Powerpoint, PDF, PHP, andHTM/HTML) represented 41%, image data 23%, and audio andmovie files 7%. The remaining 29% were of unknown type or exe-cutable files (Lyman and Varian 2003). Since images may also con-tain text a lower bound for text files as a fraction of known filetypes is 58% and of all (classifiable and unknown) is 41%. Since

Dow

nloa

ded

from

info

rms.

org

by [

218.

189.

95.1

0] o

n 09

Jul

y 20

14, a

t 20:

46 .

For

pers

onal

use

onl

y, a

ll ri

ghts

res

erve

d.


Table 1 Descriptive Statistics, 176 Small Countries, 1998–2004

Variable N Mean S.D. Min Max

Time-varying covariatesInternet Users (per 100 people) 11169 00099 00143 00000 00755Per-Capita GDP (US$ thousands) 958 80392 90323 00450 600249Telephone Infrastructure 776 00240 00209 00000 00908Log Normalized Internet Price (1998) 25 −50935 00578 −60644 −40496Log Normalized Internet Price (2000) 25 −60457 00601 −70167 −40981Log Normalized Internet Price (2001) 111 −40562 10517 −70279 −10526Fraction School Enrollment 653 00872 00160 00278 10000Civil Liberties Index 11055 40495 10794 10000 70000

Time-constant covariatesLiteracy Rate 753 00795 00197 00240 10000Gini Coefficient 688 00406 00106 00247 00743Age Below 20 11078 00424 00117 00196 00605Age 20 to 39 11078 00302 00034 00244 00480Age 40 to 64 11078 00208 00071 00110 00341Age Above 64 11078 00066 00044 00011 00182Fraction Urban Population 11168 00546 00241 00077 10000Household Size 678 40525 10413 20000 100500

Content measuresRelevant Content (millions of relevant hosts) 11114 30047 130442 00000 1720503Own Content (millions of hosts) 11114 00081 00343 00000 50434Large Country Content (millions of hosts) 926 220525 450152 00000 2060814Language Herfindahl 926 00868 00197 00378 10000

Supplementary variablesLog[Gateway Capacity] (gigabits per second) 11169 00331 10154 00000 80144IP Protection 321 50090 10973 00300 80600

Price instrumentsGovernment Tax Receipts (% of GDP) 519 160908 70153 00958 430705Corporate Tax Rate (%) 499 270783 90609 00000 540000Telephone Employees (per 1,000 fixed lines) 922 120000 200789 00068 1750385

Note. See Online Appendix B for a description of the variables and their sources.

estimate of content’s effect downward. There wouldtend to be less content created in languages whoseusers have slow connections. Since we allocate con-tent based on language usage, we will tend to over-allocate content to language users with slow con-nections and underallocate to those with fast. Sincelanguage users with slow connections actually haveless content available, they will have lower adoptionrates, biasing our results downward.

Content: We measure content by the number ofhost computers connected to the Internet in each yearfor each country. Host computers contain accessible

2003 is near the end of our sample period, text is likely an evenhigher fraction in earlier years as faster access speeds over timehave led to increased use of images and video. These data are forthe “surface” Web. There is a large amount of data in the “deep”Web, but most of it is not publicly accessible during our sampleperiod either because it is behind corporate firewalls or becauseis not indexed by search engines (see Bergman 2001). Our mea-sure of hosts includes the surface but not deep Web. For a laterperiod, Bohn and Short (2009) estimate that in 2008 Internet textcomprised 178 hours of usage for the average Internet user whilevideo comprised 2 hours. In terms of storage, they estimate that in2008 there were 8.0 exabytes of Internet text compared with 0.9 ofvideo. Video would play an even smaller role during our sampleperiod when Internet connections were much slower.

content, and the total quantity of content is propor-tional to the number of computers.17 This does notmeasure content quality; however, for our estimatesit need only be the case that quality is proportionalto storage capacity across different languages. We donot directly observe the language of these computers’content but rather infer it from the registration coun-try, as explained below.

Internet host numbers are based on data from theInternet Systems Consortium, Inc. (ISC). During oursample period, ISC took an annual census of hostcomputers connected to the Internet. ISC maintainedthe same sampling procedure throughout our sampleyears, ensuring comparability. However, since com-puter storage capacity may change over time, weinclude year dummies in all our estimates and alsoestimate separate yearly effects as a robustness check.

The ISC data also allocate each host to a coun-try, which allows us to allocate them to languages.Assignment of a host to a country does not necessarily

17 Host computers are connected to the Internet and contain acces-sible content. There are many more computers connected indirectlyto the Internet through local area networks (intranets). Computerson an intranet can access the Internet but cannot host content.

Dow

nloa

ded

from

info

rms.

org

by [

218.

189.

95.1

0] o

n 09

Jul

y 20

14, a

t 20:

46 .

For

pers

onal

use

onl

y, a

ll ri

ghts

res

erve

d.


Figure 2 Internet Penetrations in Sample Small Countries in 2004

0

10

20

30

40

50

60

70

80In

tern

et p

enet

ratio

n (%

)

Sw

eden

Ber

mud

a

Icel

and

Mon

aco

Bel

gium

Cyp

rus

Slo

vaki

a

Mar

tiniq

ue

Reu

nion

Kuw

ait

Isra

el

Rom

ania

Gua

delo

upe

Bul

garia

Jord

an

Mol

dova

Ukr

aine

Sou

th A

fric

a

Pan

ama

Phi

lippi

nes

Aze

rbai

jan

Geo

rgia

Bot

swan

a

Ken

ya

Gab

on

Leso

tho

Zam

bia

Sri

Lank

a

Erit

rea

Equ

ator

ial G

uine

a

Tur

kmen

ista

n

Gui

nea

Mau

ritan

ia

Mal

awi

Cen

tral

Afr

ican

Rep

ublic

Eth

iopi

a

Source. World Development Indicators Database (2010a). Internet penetration (fraction of population with Internet access) for 176 sample small countriessorted from highest to lowest penetration. Not all country names displayed because of a lack of space.

mean that the computer is physically located withinthe country; however, our estimation requires onlythat the computer contains content created within thatcountry. The rules for assigning hosts make this likely.Although the rules differ slightly across countries,most require a local presence requirement such ascitizenship, resident address, or local administrativecontact.18

Since more than one language is used in mostcountries, we allocate the total hosts to each lan-guage based on the fraction of the country’s pop-ulation using each language.19 This assumes thatall content is language-specific. As discussed ear-lier, Internet content during our sample period isprimarily textual; moreover, much nontextual con-tent is language-specific as images often contain textand videos language-specific dialogue. Nonetheless,Online Appendix D shows that our model accom-modates non-language-specific content assuming thateach country’s language-specific content is produced

18 Online Appendix C contains more detail on how ISC collects thehost data and allocates it to countries.19 This is not a major concern for our instrument because the pop-ulations of virtually all of the large countries are dominated by asingle language. We assume that all the host computers in largecountries pertain to that country’s dominant language.

in the same proportion as language users in thatcountry and that language-specific and non-language-specific content affect adoption equally on the margin.The former assumption is the same as that requiredwhen we allocate hosts to languages within coun-tries, so only the latter is potentially restrictive. Ifadoption is more responsive to language-specific thannon-language-specific content, then our estimates willunderstate content’s effect. If the opposite is true, wewill overstate it. Combining this measure of contentfor each country in each year with the language data,we construct the relevant content and instrument foreach country-year pair based on Equations (5) and (6).

Prior to 1999, if ITU could not find an independentestimate of Internet users in a country, it based itsestimate on a multiple of the number of host comput-ers in the country, which would pose problems forour estimation. After this, ITU used only surveys toquantify users (Minges 1999). To determine whetherthis is a problem, we reestimated our baseline esti-mates dropping the year 1998 data. The results werevirtually identical.

Language Users: Our source for language data is Eth-nologue (Gordon 2005), which offers the most com-prehensive catalogue of the world’s languages (forlinguistic reviews, see Campbell and Grondona 2008,

Dow

nloa

ded

from

info

rms.

org

by [

218.

189.

95.1

0] o

n 09

Jul

y 20

14, a

t 20:

46 .

For

pers

onal

use

onl

y, a

ll ri

ghts

res

erve

d.


Hammarström 2005, Paolillo and Das 2006). Ethno-logue provides detailed and comprehensive estimatesof first-language speakers of each language by coun-try.20 Its data are not complete enough to estimateusing second-language speakers.

Since Internet content was primarily textual dur-ing our sample period, we ideally would use thenumbers of literate users of each language to cre-ate our relevant content measure. Because we do notobserve language-specific literacy rates by country,we use the numbers of speakers of each language ina country and include the country’s overall literacyrate as a control variable. We combine spoken dialectswhose users employ the same written language. Forexample, we combine speakers of the many Chinesedialects that all utilize simplified Chinese for writing.Ethnologue is a thorough accounting of the world’slanguages. As a result, some are spoken by very fewpeople. To make data entry manageable, for eachcountry we added languages in descending order ofthe most spoken and kept adding until the next lan-guage would contribute less than 1% of the country’spopulation or all languages were exhausted. Acrossall countries, this comprised 811 languages or spokendialects.

To choose instrument languages, we apply the twocriteria discussed in §2: the language is spoken inmany countries and its usage distribution is skewedwith a few countries making up a significant fractionof total users. Based on these criteria, we use 14 lan-guages to construct our instrument:21

JF =

Chinese, Spanish, English, Hindi,Portuguese, Russian, Japanese,German, French, Hausa, Zulu,Nyanja, Pulaar, Pular

0 (8)

The first 8 are among the top 10 most spoken lan-guages in the world based on Ethnologue.22 French isthe 17th most spoken language. The usage of the lan-guages between the 10th and 17th (Javanese, Telugu,Marathi, Vietnamese, Korean, and Tamil) is eithernot widespread or fairly uniformly distributed acrosscountries. The last five languages were chosen toinclude African languages subject to meeting our two

20 Ethnologue does not distinguish between native and primary first-language speakers. This should be considered in interpreting ourresults.21 To check the robustness to the choice of instrument languages,we randomly divided these into two groups of seven languageseach and reestimated our baseline results in column 3 of Table 5,using each of these two sets to construct the instrument. The coef-ficient on relevant content for both estimates was within 1.6% ofour baseline estimate, and the coefficients remained significant atbelow the 0.01% level.22 Arabic (fourth) and Bengali (seventh) were not included becausetheir usage was not skewed enough.

criteria. Each of these five is spoken in at least fourcountries, and the two most populous countries usingthe language represent at least 82% of total users. Col-umn 3 of Table 2 shows the total number of users forthe 14 languages used to construct our instrument:2.7 billion people, or 44% of the 6.1 billion world pop-ulation in 2000.

We use the number of potential adopters (i.e., pop-ulation using a language) in a country to identifylarge and small countries. We choose the large coun-tries (the set IL) for our instrument (Equation (6)) bythe following procedure: For each language, sort thecountries in descending order by the number of users.Starting at the top, add countries until the last coun-try added brings us above 75% of worldwide users.There were three exceptions when we kept addingabove 75%.23 Column 5 of Table 2 shows the 31 largecountries chosen, and columns 6–8 show the numberof users in the large countries and as a percentageof worldwide users. Identification relies on the col-umn 8 percentages being large so that these countriesare unaffected by adoption in small countries (the31 large countries will be excluded from our analy-sis). The large countries represent 80% or more of theworld’s users of each language.

The last three columns of Table 2 show data foreach language’s largest small country. Columns 10and 11 show the number of users in the largest smallcountry and as a fraction of worldwide users. Identi-fication depends on the column 11 percentages beingsmall so that Internet adoption in these countriesdoes not affect large countries’ content production.The largest small countries represent eight percent orfewer of the world’s users for each language. The per-centages for all other small countries are below this.

Control Variables: We include as many control vari-ables from previous studies of Internet adoption aspossible so as to isolate content’s effect. Therefore,subject to preserving enough degrees of freedom todiscern content’s effect, our goal is to maximize thevariance explained by our regressions rather than thesignificance of individual coefficients. To identify con-trol variables, we rely on previous papers estimatingcross-country Internet adoption.

Time-varying factors include measures of wealth,education, infrastructure, cost of access, and freedom

23 The three exceptions were because there was an obvious largedrop between two countries. For Chinese, mainland China alonewould bring us above 75%, but we added Taiwan because it had5.2 times as many Chinese speakers as the next largest country,Malaysia. For English, the United States and the United Kingdomalone would bring us above 75%, but we added Canada and Aus-tralia because Australia was 4.9 times as large as the next largestcountry, New Zealand. For Portuguese, Brazil alone would bringus above 75%, but we added Portugal because it is 15.6 times aslarge as the next largest country, Paraguay.

Dow

nloa

ded

from

info

rms.

org

by [

218.

189.

95.1

0] o

n 09

Jul

y 20

14, a

t 20:

46 .

For

pers

onal

use

onl

y, a

ll ri

ghts

res

erve

d.


Table 2 Profiles of Large and Small Countries for Included Languages

Worldwide Large countries Largest small country

3 4 6 7 10Total no. of Total No. of Total no. of 8 Total no. of 11

1 2 users content 5 users users % of 9 users % ofRankinga Language (millions) (1,000s of hosts)b Country (millions) (millions) worldwide Country (millions) worldwide

1 Chinese 11204076 21674068 China 11171005 11193074 9901 Malaysia 4039 0036Taiwan 22069

2 Spanish 322030 111100000 Mexico 86021 256016 7905 Dominican Republic 6089 2014Colombia 34000Argentina 33000

Spain 28017Venezuela 21048

Peru 20000Chile 13080Cuba 10000

Ecuador 90503 English 309035 891300000 United States 210000 297078 9603 New Zealand 3021 1004

United Kingdom 55000Canada 17010

Australia 150685 Hindi 180077 37001 India 180077 180077 10000 Nepal 0011 00066 Portuguese 177046 21538054 Brazil 163015 173015 9706 Paraguay 0064 0036

Portugal 100008 Russian 145003 677062 Russia 145003 145003 10000 Ukraine 11034 70829 Japanese 122043 81370064 Japan 122043 122043 10000 Singapore 0002 0002

10 German 95039 51289015 Germany 75030 82080 8608 Kazakhstan 0096 1000Austria 7050

17 French 64086 21953088 France 64086 64086 10000 Belgium 4000 6017Hausa 24016 0026 Nigeria 18053 23053 9704 Chad 0010 0041

Niger 5000Zulu 9056 60072 South Africa 9020 9020 9602 Lesotho 0025 2059

Nyanja 9035 0035 Malawi 7000 8060 9200 Mozambique 0050 5032Zambia 1060

Pulaar 3024 0036 Senegal 2039 2065 8107 Guinea-Bissau 0025 7056Gambia 0026

Pular 2092 0012 Guinea 2055 2055 8704 Sierra Leone 0018 601221671058 1231003032 21563025 21563025 9509 32081 1023

All languages 61070050c 1381648022

aMost spoken languages by first-language speakers according to Gordon (2005). If blank, it is not ranked.bAverage number of hosts across six years of data.cBased on year 2000 data from United Nations (2004).

of expression. Per-capita GDP measures a country’swealth, which we expect to positively affect adoption.Internet access is likely more highly valued in coun-tries with more educated populations, so we includethe fraction of eligible children enrolled in primaryschool. We include the fraction of the populationwith fixed phone lines to measure telecommunica-tions infrastructure quality. Although there are otherways to access the Internet during this time, thesewere either rare (satellite and Wi-fi) or likely highlycorrelated with telephone infrastructure (cable televi-sion). We include a measure of citizens’ freedom toengage in expression, based on a measure used byNGO Freedom House, to control for the degree ofgovernment restrictions on content access. The mea-sure ranges from 1 to 7, with 7 being the most free.24

24 Freedom House defines 7 as the least free. We reverse the orderfor ease in interpretation.

We include average monthly Internet access pricesnormalized by per-capita GDP to control for cost ofaccess. Unfortunately, prices are available only forthree years (1998, 2000, and 2001) and not for allcountries. Since each year’s data measure a differ-ent type and amount of usage, we cannot pool themacross years. Internet access prices and adoption mayboth be higher in countries with higher unobservedaccess quality, which would bias the price coefficienttoward 0. We therefore instrument with variablesaffecting price but affecting adoption only throughprice. Since we include fixed effects in our final esti-mation, we use three time-varying instruments. Cor-porate tax rates directly affect the cost of providingInternet access. The ratio of government tax receiptsto GDP captures the regulatory atmosphere in whichthe Internet service providers operate. The numberof telephone employees per fixed line proxies for the

Dow

nloa

ded

from

info

rms.

org

by [

218.

189.

95.1

0] o

n 09

Jul

y 20

14, a

t 20:

46 .

For

pers

onal

use

onl

y, a

ll ri

ghts

res

erve

d.


productivity of or labor-capital ratio in the telecom-munications industry.

We also include a number of time-constant controls.These are “time-constant” in that we observe onlyone year of data, although they likely change slowly.The Gini coefficient of income controls for the wealthdistribution within a country, and we expect higherinequality (higher Gini coefficient) to negatively affectadoption. The fraction of a country’s population liv-ing in urban areas measures infrastructure, demand,or both. More densely populated areas can be servedmore cheaply on a per-customer basis than more dis-persed. At the same time, it may be that Internetaccess demand by urban residents differs from that byrural residents. Since familiarity with the Internet islikely age dependent, we control for the country’s agedistribution using the population fraction in four agebrackets. Average household size allows for potentialeconomies of scale in adopting Internet access withinhouseholds. Literacy rate controls for the ability of acountry’s population to read content.

These control variables are drawn from a varietyof papers. Wallsten (2007) explains broadband pen-etration for Organisation for Economic Co-operationand Development (OECD) countries, and Wallsten(2005) assesses the impact of regulation on develop-ing countries’ Internet adoption rates and prices. Fordet al. (2008) produce a broadband performance indexfor OECD countries based on the predicted valuesfrom an adoption regression. Chinn and Fairlie (2006)explain cross-country Internet and computer adoptionrates.

We know of only three papers that include lan-guage to explain Internet adoption, and all includeonly a single language (English) and do not con-trol for endogeneity. Hargittai (1999) explains Internetadoption by OECD countries and includes English-language usage as an explanatory variable becauseof its importance in the media and computing fields.The effect of language is not significant. Kiiski andPohjola (2002) estimate a diffusion model of Internetadoption by OECD countries and include English-language proficiency for the same reason but estimatea negative effect. Wunnava and Leiter (2009) alsoestimate a diffusion model of Internet adoption butwith more countries. They include English-languageproficiency to measure the accessibility of English-language content. They find a positive and significanteffect.

5. Main ResultsContent availability has a positive and statisticallysignificant effect on adoption. Since content is notdirectly measurable, there is no single right way toquantify its effect. We provide several ways and find

Table 3 Adoption/Content Correlation Matrix for Sample Countries,1998–2004 4N = 7795

Internet users Own content Relevant content

Own content 005324000005

Relevant content 00348 000334000005 4003665

Large country content 00293 00083 00517(instrument) 4000005 4000215 4000005

Note. Significance levels are in parentheses.

an important role regardless. First, we compute anelasticity of adoption with respect to content andcompare it to other markets. Second, we comparecontent’s effect to that of other adoption determi-nants. Third, we quantify the additional adoption thatresults from the annual average content production inour sample.

Before we formally estimate the effect of content onadoption, we examine the correlations between adop-tion and the various measures of content includingthe instrument. These are shown in Table 3. A coun-try’s own content is highly positively correlated withits own Internet adoption, consistent with a two-sidedmarket. Relevant content is also highly positively cor-related with adoption, consistent with the ubiquity ofInternet content (this content is produced both withinand outside the country but in the languages of itspopulation). However, relevant content and own con-tent are not significantly correlated. This is consistentwith a country’s own content production being deter-mined by two-sided market effects within the country,whereas relevant content is determined by two-sidedmarket effects across many countries sharing commonlanguages.

Finally, large-country content (the instrument) ishighly correlated with both adoption and relevantcontent but is much less correlated with a country’sown content. This is consistent with large-countrycontent influencing a country’s content productiononly indirectly through adoption. The low correlationbetween large-country and own-country content isinformal evidence that the exclusion restriction is met,whereas the high correlation between large-countryand relevant content is informal evidence of its rel-evance. We provide more formal tests of the instru-ment’s validity below.

5.1. First-Stage ResultsColumns 1–3 of Table 4 show the first-stage resultsfor Internet access prices. Given the small number ofobservations in each year, the coefficients are noisy.The number of telephone employees has a positiveeffect and is significant in two of the three years, con-sistent with higher prices from lower productivity.

Dow

nloa

ded

from

info

rms.

org

by [

218.

189.

95.1

0] o

n 09

Jul

y 20

14, a

t 20:

46 .

For

pers

onal

use

onl

y, a

ll ri

ghts

res

erve

d.


Table 4 First-Stage Regressions for Internet Access Prices and Relevant Content

1 2 3 4 51998 2000 2001 Relevant Relevant

log prices log prices log prices content content

Intercept −405397∗∗∗ −501697∗∗∗ −403932∗∗∗ 102422∗∗∗ 00551440056325 40070155 40020375 40030845 40044855

Large country content 000266 001379∗∗

40002505 40005665(Large country content)2 000007∗

40000045Government tax receipts (% of GDP) −000359∗ −000438∗∗ −000601∗∗∗

40001935 40002165 40001245Corporate tax rate (%) −000219 −000138 000030

40001445 40001955 40007075Telephone employees (per fixed line) 000692∗∗ 001493∗∗ 001371

000345 000588 001162R2 003000 003909 002563 001946 001803N 44 44 134 926 926

Notes. Standard errors are in parentheses. Standard errors for relevant content regressions are clustered at the country level. Dummyvariables for missing values are included for all variables in price regressions.

∗Indicates 10% significance; ∗∗indicates 5% significance; ∗∗∗inidcates 1% significance.

Government tax receipts has a significantly negativeeffect in all three years, consistent with greater subsi-dies for Internet access in countries with greater gov-ernment revenues.

We allow for a flexible functional form in the first-stage regression of relevant content. We use a second-order, Taylor-series expansion of the instrument asshown in column 4 of Table 4.25 Both the linearand quadratic terms are positive, although only thequadratic term is significant. Specification tests indi-cate the exclusion restriction and the relevance condi-tion are likely met. A Hausman specification test ofexogeneity yields a test statistic of 47.1 compared to acritical value of 0.1, and the F -value for our first-stageregression is 111, which greatly exceeds the criticalvalue of 10 specified in Staiger and Stock (1997) torule out weak instruments. To test the sensitivity ofour results to the quadratic functional form, we reesti-mated using the linear first-stage specification shownin column 5 of Table 4. Large-country content has asignificantly positive effect on relevant content, andvery similar second-stage results were obtained.

5.2. Panel Data ResultsAlthough we control for many factors thought toaffect Internet adoption, we also include country fixedeffects to control for country-level unobservables. Awithin-groups estimate of Equation (4) provides con-sistent estimates of the time-varying variables in themodel including content. To compare content’s effectto that of as many variables as possible, we alsoinclude time-constant variables. Since we believe that

25 A cubic term was not significant.

we have plausibly exogenous time-invariant factorsavailable, we use an HT estimator.

Of the time-varying variables, telephone infrastruc-ture and civil liberties are likely endogenous in theHT sense (i.e., correlated with country-level unobserv-ables). A country that invests heavily in technology(more than commensurate with its per-capita GDP)likely has high Internet adoption and high fixed-phone line penetration. A society with greater unob-served preferences for Internet access may also havea greater preference for civil liberties. Price and rel-evant content are exogenous by design. Neither per-capita GDP nor school enrollment is likely affected byunobserved preferences for Internet adoption in theshort run.

Of the time-invariant variables, all are likely exoge-nous in the HT sense except for literacy. The incomedistribution, age distribution, average household size,and urban density are not likely affected by unob-served preferences for Internet adoption. Measuringliteracy is subjective because there are no standardcriteria across countries. Countries with low literacyrates may report artificially high rates and also havea low unobserved preference for Internet access.

Column 1 of Table 5 shows the second-stage resultsof a random-effects specification with standard errorsclustered by country and robust to general het-eroskedasticity. The table is divided into four panelsclassifying the variables as time varying versus timeinvariant and exogenous versus endogenous. We willnot discuss the results in detail since this is rejected infavor of a fixed-effects specification, but relevant con-tent has a highly statistically significant effect (belowthe 0.01% level).

Dow

nloa

ded

from

info

rms.

org

by [

218.

189.

95.1

0] o

n 09

Jul

y 20

14, a

t 20:

46 .

For

pers

onal

use

onl

y, a

ll ri

ghts

res

erve

d.


Table 5 Effect of Content on Internet Adoption for All Sample Countries, 1998–2004, Second-Stage, Panel Data Estimates

1 2 3 4RE FE HT-GLS HT-GLS

Time-varying exogenousPer-Capita GDP 000101∗∗∗ 000115∗∗∗ 000104∗∗∗ 000105∗∗∗

40000085 40000145 40000105 40000105Log Normalized Internet 000131 000033 000063 000108

Price (1998) 40004235 40004165 40004175 40004165Log Normalized Internet −000363 −000447 −000413 −000399

Price (2000) 40002905 40002865 40002865 40002855Log Normalized Internet −000012 −000016 −000013 −000014

Price (2001) 40000625 40000615 40000615 40000615Fraction School Enrollment 000051 000022 −000006 −000023

40002185 40002245 40002225 40002225Relevant Content 000015∗∗∗ 000016∗∗∗ 000015∗∗∗ 000050∗∗∗

40000045 40000045 40000045 400001154Language HHI Above Mean5× −000039∗∗∗

Relevant Content 40000115Time-varying endogenous

Telephone Infrastructure −001579∗∗∗ −001723∗∗∗ −001745∗∗∗ −001716∗∗∗

40001675 40001715 40001695 40001685Civil Liberties Index 000028 000002 −000007 −000001

40000305 40000395 40000385 40000385Time-invariant exogenous

Gini Coefficient −000247 000015 −00016640007905 40010205 40010135

Fraction Urban Population 000582∗ 001018∗∗ 000920∗

40003425 40004995 40004895Age Below 20 001609 009908 009698

40041555 40063935 40063195Age 20 to 39 000720 002976 002820

40033575 40043685 40043615Age 40 to 64 005408 107489∗ 106693∗

40061585 40094015 40092895Household Size −000138∗ −000070 −000056

40000725 40001035 40001025Language HHI Above Mean 000117

40001785Time-invariant endogenous

Literacy Rate −000549 000936 00124740004505 40012795 40012495

�� 00045 00045 00045 00044� 00695 00827 00788 00785R2 00917N 1,169 1,169 1,169 1,169Wald �2-statistic 1,785.3 1,719.3 1,743.6Specification test 6907 2003

Notes. Standard errors are in parentheses. Year dummies and dummy variables for missing values are included for all variables in all regressions. Prices andrelevant content were instrumented in all regressions. Standard errors are clustered by country and allow for general heteroskedasticity in the random-effects(RE) and fixed-effects (FE) specifications. The HT estimates use the covariance matrix specified in Hausman and Taylor (1981).

∗Indicates 10% significance; ∗∗indicates 5% significance; ∗∗∗indicates 1% significance.

Column 2 of Table 5 shows the second-stage resultsof a fixed-effects regression with standard errors clus-tered at the country level and robust to general het-eroskedasticity. The regression yields an R2 of 0.917,consistent with a wide range of control variables.Only a few of the control variables are significant, butthere are two reasons why. First, given the country

fixed effects identification comes only from time-series variation. Second, we include more control vari-ables than previous studies (conditional on includingcountry fixed effects). Since the results are similar tothose obtained in the HT specification, we postponetheir discussion. The fixed-effects estimates are con-sistent even if included variables are correlated with

Dow

nloa

ded

from

info

rms.

org

by [

218.

189.

95.1

0] o

n 09

Jul

y 20

14, a

t 20:

46 .

For

pers

onal

use

onl

y, a

ll ri

ghts

res

erve

d.


the country-level unobservables, allowing a Hausmanspecification test for the consistency of the random-effects estimates. The null hypothesis of consistencyis rejected below the 0.01% level with a chi-squaredstatistic of 69.7, consistent with correlation betweenunobserved country-level effects and the regressors.

Column 3 of Table 5 contains HT estimates. Sincethe fixed-effects specification provides consistent esti-mates and our model is overidentified, we can per-form a Hausman specification test of the exogene-ity of our HT instruments. The null hypothesis thatour instruments are uncorrelated with the country-level unobservables is not rejected (16% significancelevel with a chi-squared statistic of 20.3). Thus, boththe fixed-effects and HT estimators provide consis-tent estimates of the time-varying factors; however,the HT estimator is more efficient and provides con-sistent estimates of the time-invariant factors. This isour preferred specification, although content’s effectis similar across both.

Per-capita GDP has a positive and highly significanteffect on adoption. An additional $958 in annual percapita GDP is associated with one percentage pointhigher adoption.26 A country one standard deviationabove the mean per-capita GDP has 9.7 percentagepoints higher adoption than one at the mean. This isa large effect given the mean adoption level of 9.9%in the sample.

Internet prices for two of the three years are nega-tive but only the year 2000 prices are borderline sig-nificant (at the 12% level). The lack of significance islikely due to the lack of data. A country one standarddeviation above the year 2000 mean log price has 2.5percentage points lower adoption (25.0% of the meanadoption level). The estimates imply a price elastic-ity of −0042 for Internet adoption. School enrollmentand civil liberties are not significant, although there islittle time-series variation in these. Telephone infras-tructure has a significant negative effect on adop-tion, inconsistent with prior expectations. This may bebecause countries with heavily-subsidized and ineffi-cient telephone industries have high Internet accessprices and poor telephone infrastructure. Consistentwith this, telephone infrastructure and instrumentedprices are significantly negatively correlated. We alsoshow below that the time-series impact from this vari-able is small.

Content has a positive and significant (below the0.01% level) effect on adoption. A country one stan-dard deviation above the mean in relevant contenthas 2.0 percentage points higher adoption or 20.0% ofthe mean adoption level. Countries with users of lan-guages with more worldwide accessible content have

26 The effects of changes in independent variables are calculated atthe mean values of all other variables unless otherwise noted.

higher adoption rates. The unreported coefficients onthe year dummies are consistent with higher Internetadoption rates over time (and all but year 1999 aresignificant); however, this should be interpreted withcaution since the content measure is not necessarilyconsistent over time.27

Of the time-invariant variables, only urbanizationis significant at the 10% level or better, although theage bracket dummies are jointly significant at the12% level. Fraction of urban population has a posi-tive and very significant effect, consistent with eithereasier construction of Internet infrastructure in moredensely populated areas, greater demand for accessrelative to more rural areas, or both. Each additional1% of population living in urban areas is associatedwith 0.1 percentage points higher adoption. A coun-try one standard deviation above the mean has 2.5percentage points higher adoption, or 24.8% of theaverage adoption rate in the sample. Although theage variables are not highly statistically significant,they have a large economic impact. Countries witha smaller fraction of people above 65 years of age(the omitted age category) have higher adoption lev-els, with the greatest effect both statistically and eco-nomically in the age 40 to 64 category. Increasing thefraction of population in the age 40 to 64 categoryby one standard deviation and spreading an equiva-lent decrease equally across the other three categoriesresults in a 9.3 percentage point increase in adoption(93.6% of the average adoption rate in the sample).Running the same experiment (increasing a categoryby one standard deviation and decreasing the otherthree categories equally by the same total amount)results in the following: below 20 category, a 36.1%increase; 20–39 category, a 21.2% decrease; and above64 category, a 73.0% decrease.

5.3. Interpreting the Effect of ContentContent has a large impact on the equilibrium levelof adoption. Our estimates imply an elasticity of 0.046of adoption with respect to relevant content. For anelasticity of adoption with respect to hosts, we needto estimate how much relevant content increases withone additional host as determined by the languagedistributions across countries. We measure this by theratio of relevant content to hosts across all countries(small and large) and all years, yielding 6.72. There-fore the elasticity of adoption with respect to hosts

27 If some countries add more hosts earlier when computers havesmaller capacity and other countries add more hosts later whencomputers have greater capacity, this could bias our results. Wereestimated Equation (4) using a three-year moving-average ofinstrumented relevant content. This allows relevant content todepend on both the current and previous stocks of host computers.We tried moving averages of 1/4, 1/3, and 1/2 and found resultsvery similar to our baseline estimates.

Dow

nloa

ded

from

info

rms.

org

by [

218.

189.

95.1

0] o

n 09

Jul

y 20

14, a

t 20:

46 .

For

pers

onal

use

onl

y, a

ll ri

ghts

res

erve

d.


is 0.31. This is more powerful than the indirect net-work effect estimated by Gandal et al. (2000) in thecompact disk (CD) market (an elasticity of CD playerswith respect to the number of CD titles of 0.033) butbelow that estimated by Dranove and Gandal (2003)for the digital video disc (DVD) market (an elasticityof DVD players with respect to fraction of movie titlesreleased on DVD of 1.13).28 Using our year 2000 priceelasticity of −0042, adoption is 0.74 times as respon-sive to content as price—above that in the CD market(the ratio of content and price elasticities is 0.54) butbelow that in the DVD market (a ratio of 1.2).29

We can also compare content’s impact to that ofother factors. Its effect is below that of GDP andsome age-group redistributions but is comparable tothe other significant control variables. A country onestandard deviation above the mean in relevant con-tent has 20.0% higher adoption. For time-varying fac-tors the effects of a one standard deviation increaseare the following: per-capita GDP, a 98.2% increase;year 2000 normalized prices, a 25.0% decrease; andtelephone infrastructure, a 36.8% decrease. For time-constant factors the effects are as follows: fractionurban population, a 24.8% increase; and age distribu-tion, a 73.0% decrease to a 93.6% increase, dependingon the age category that is increased.

This has important implications for countries wish-ing to stimulate Internet adoption. Increasing GDPwill increase Internet adoption dramatically, but thisis difficult. Similarly, short-run changes in the age dis-tribution would require dramatic immigration policychanges. Stimulating relevant content, either directlyor indirectly, is easier and less costly. In addition, gov-ernments and NGOs can influence adoption in othercountries by creating relevant content in the targetcountry’s languages.

There are two issues with this comparison. First,moving any of these variables by one standard devia-tion is a lot. Therefore, it is useful to estimate the effectof “reasonable” changes. Second, it assumes that it isequally easy to move the variables by one standarddeviation. Therefore, it is useful to gauge the speed atwhich these variables change over time. To do so, wecompute annual changes in the time-varying factorsand the effects such changes would have on adoption.Since we do not have a comparable price measureover time, GDP and telephone infrastructure are the

28 Although these two markets are not directly analogous to theInternet, they are similar in that a CD or DVD title is replicatedmultiple times just as a host of content is “replicated” by multipleusers accessing it.29 Other papers estimate the magnitude of two-sided networkeffects on firms’ market shares (Corts and Lederman 2009, Nairet al. 2004, Ohashi 2003, Park 2004). These results are not directlycomparable to ours since we estimate the effect on overall demandand they estimate the effect on firms’ residual demands.

only variables to which we can compare (although wecannot measure yearly changes in the age distributionor fraction urban population, these are likely small,implying small changes in adoption).

The top panel of Table 6 summarizes these changesfor the small countries. Adoption increased on aver-age 2.2 percentage points per year in small countries.The two rightmost columns compute the effect thatthe annual changes in each of the explanatory vari-ables would have on small-country adoption evalu-ated at the mean of all other variables. For example,per-capita GDP increased $398 per year on averagein the small countries. This would increase adoptionby 0.42 percentage points, or 19.1% of the averageyearly increase of 2.2 percentage points for the smallcountries. Similar calculations for telephone infras-tructure reveal a minimal 1.0% annual decrease. Rele-vant content for the small countries increased on aver-age by 885,000 hosts per year. This would increasesmall-country adoption by 6.0% of the average yearlyincrease in their adoption.

The bottom panel of Table 6 summarizes annualeffects based on the large countries. Per-capita GDPincreased $493 per year on average for these coun-tries. Such an increase would stimulate small-countryadoption by 23.6% of the 2.2 percentage pointannual increase in adoption for the small coun-tries. A similar calculation for telephone infrastruc-ture yields a 5.9% decrease. The annual increase inlarge-country content—the content produced by the

Table 6 Estimated Effects of Variables on Adoption by SmallCountries

Implied % of annualAverage increase in increase inannual adoption Internet usagechange for small by small

Variable Na 1998–2004 countriesb countriesc

Small countriesInternet Users 157 0.022Per-Capita GDP 136 0.398 000042 1901

(US$ thousands)Telephone Infrastructure 41 0.001 −000002 −100Relevant Content 152 0.885 000013 600

(millions of hosts)Large countries

Per-Capita GDP 28 0.493 000051 2306(US$ thousands)

Telephone Infrastructure 7 0.007 −000013 −509Own Content 29 1.150 0000174 708d

(millions of hosts)

Note. Large countries are identified in Table 2 and small countries in OnlineAppendix A.

aData are missing for some countries in some years.bMarginal effect evaluated at the means of all other independent variables.cRelative to the average annual increase in Internet users in small countries

(0.022).dAssumes all content is “relevant” as defined in the text.

Dow

nloa

ded

from

info

rms.

org

by [

218.

189.

95.1

0] o

n 09

Jul

y 20

14, a

t 20:

46 .

For

pers

onal

use

onl

y, a

ll ri

ghts

res

erve

d.


countries themselves—is 1.2 million hosts. This wouldincrease small-country adoption by 7.8% of the 2.2percentage points annual change in adoption for thesmall countries. Whether the top or bottom panel ofTable 6 is more appropriate depends on which moreaccurately predicts annual changes. However, theyare similar. In both, content is an important factor inaffecting adoption—it has about one-third the impactof GDP.

5.4. Linguistic IsolationInternet content may act as a substitute for or com-plement to isolation. If isolated populations use theInternet to access people with similar interests orcharacteristics, content would have a greater effect onadoption by more isolated groups. On the other hand,if people learn about the Internet through word ofmouth, and this is less likely if one is isolated, contentwould have a smaller effect on adoption by more iso-lated groups. We distinguish these alternatives usinglinguistic isolation, as measured by linguistic hetero-geneity. We implement this using a Herfindahl index(HHI) of language usage in each small country:

HHIi =J∑

j=1

( Usersij∑J

j=1 Usersij

)2

i ∈ IS 0 (9)

A country with an HHI close to 0 is linguisti-cally very heterogeneous, whereas a country with anHHI of 1 is completely homogeneous. To identifycontent’s importance in linguistically homogeneousversus heterogeneous countries, we interact instru-mented relevant content with a dummy variable indi-cating whether a small country has an above-averageHHI.

Column 4 of Table 5 shows the results. The base-line effect of language heterogeneity is insignificant.Relevant content has a positive and significant effect,but the effect is lower for countries above the meanlanguage HHI. Content has a smaller effect in coun-tries with more homogeneous language users. A smallcountry one standard deviation above the mean in rel-evant content has 5.2 percentage points higher adop-tion if it is below the mean language HHI but only1.5 percentage points if it is above. This is con-sistent with the Internet being a tool to overcomelinguistic isolation. In contemplating the future ofthe online encyclopedia Wikipedia, its founder, JimmyWales, asked in mid-2009: “Is it more important toget to 10 million articles in English, or 10,000 inWolof?” (Cohen 2009). Our results imply that in termsof adoption—the latter.

5.5. RobustnessTo see whether our relevant content measure simplyproxies for the small country’s own content produc-tion, we add a measure of the latter to our estimation:

owncontentit =

∑Jj=1 6UsersijContentijt7

Populationi

1 i ∈ IS 0 (10)

This differs from relevant content in Equation (5) inexcluding content produced outside the country. Sincethis variable is endogenous, its coefficient shouldbe interpreted with caution. Columns 1 and 2 ofTable 7 show the results. Relevant content’s coefficientand significance is very close to that in our baselineresults in column 3 of Table 5. This is consistent withinstrumented relevant content measuring content thataffects but is not affected by small-country adoption.The other coefficients are not greatly affected exceptthat the age variables are more significant. Own con-tent is associated with higher adoption and is highlystatistically significant, as would be expected in atwo-sided market. The magnitude is not interpretablesince it is endogenous, but it exceeds that of instru-mented relevant content since it reflects the feedbackbetween adoption and content.

Our main results assume that content’s effect onadoption is the same across years. In columns 3 and 4of Table 7, we relax this assumption and allow for dif-ferential effects in each year. The content coefficientsare all positive and jointly very significant (at the 1%level). The magnitudes are similar across years (theeffect of a one standard deviation increase in contentranges from 2.1 to 4.4 percentage points) and gener-ally greater than that obtained when restricted to beequal in all years (2.2 percentage points).

If there are language-specific unobservables thatdrive adoption and content production, this may biasour estimates. For example, if users of certain lan-guages have higher preferences for adoption not cap-tured by our control variables, this will lead to higheradoption in small countries whose populations usethat language and at the same time lead large coun-tries to produce more content in that language toserve the higher large-country demand. To addressthis, we add language along with country and yearfixed effects to Equation (1a). Once transformed intoEquation (4), this is equivalent to including as aregressor the fraction of each small country’s pop-ulation using each language. Since including fixedeffects for all languages is infeasible, we include themonly for the 14 instrument languages. These are thelanguages that link small and large countries in ourinstrumenting approach and are most likely to intro-duce endogeneity. The results are shown in columns 5

Dow

nloa

ded

from

info

rms.

org

by [

218.

189.

95.1

0] o

n 09

Jul

y 20

14, a

t 20:

46 .

For

pers

onal

use

onl

y, a

ll ri

ghts

res

erve

d.


Table 7 Effect of Content on Internet Adoption for All Sample Countries, 1998–2004, Second-Stage Estimates

1 2 3 4

Hausman–Taylor 5 6

Own content Year effects Language fixed effects

Coeff. S.E. Coeff. S.E. Coeff. S.E.

Time-varying exogenousPer-Capita GDP 000095∗∗∗ 000010 000104∗∗∗ 000010 000119∗∗∗ 000015Log Norm. Internet Price (1998) −000102 000419 000101 000414 000121 000416Log Norm. Internet Price (2000) −000408 000287 −000419 000281 −000387 000285Log Norm. Internet Price (2001) −000028 000061 −000015 000060 −000007 000061Fraction School Enrollment 000008 000222 000016 000218 000031 000224Own Content 000311∗∗∗ 000074Relevant Content 000014∗∗∗ 000004 000018∗∗∗ 000004Relevant Content (1998) 000093 000079Relevant Content (1999) 000059 000045Relevant Content (2000) 000033 000025Relevant Content (2001) 000027∗ 000016Relevant Content (2002) 000030∗∗ 000014Relevant Content (2003) 000023∗∗∗ 000008Relevant Content (2004) 000019∗∗∗ 000006

Time-varying endogenousTelephone Infrastructure −001720∗∗∗ 000169 −001721∗∗∗ 000166Civil Liberties Index 000001 000038 −000010 000038

Time-invariant exogenousGini Coefficient −000018 000988 −000066 001088Fraction Urban Population 001185∗∗ 000482 000539 000510Age Below 20 101436∗ 006161 003947 006444Age 20 to 39 004665 004194 000835 004626Age 40 to 64 109370∗∗ 008979 007969 009474Household Size −000038 000100 −000087 000106

Time-invariant endogenousLiteracy Rate 000856 001247 000715 001289

�� 0.044 0.045 0.045� 0.774 0.817 0.832R2 0.919N 1,169 1,169 1,169Wald �2-statistic 1,757.5 1,747.1

Notes. Prices and relevant content are instrumented in all regressions. Dummy variables for missing values are included for all variables in all regressions.Estimates in columns 2 and 4 use the covariance matrix specified in Hausman and Taylor (1981). Standard errors in column 6 are clustered by country andallow for general heteroskedasticity. Columns 1–4 also contain country and year fixed effects; columns 5 and 6 also include country, year, and language fixedeffects.


and 6 of Table 7 and are similar to our baseline esti-mates in column 2 of Table 5.30

6. ApplicationsOur model can be used to measure how coun-try characteristics influence adoption’s sensitivity tocontent. Including an interaction between country

30 An alternative explanation of our results is that countries affecteach other’s adoption through a direct network effect: a commonlanguage between countries leads to increased economic activityand therefore more communication via the Internet such as email orinstant messaging, resulting in increased adoption. Adding a trade-weighted measure of trading partners’ adoption rates to Equa-tion (4) as a proxy for the economic closeness between countrypairs has minimal effect on the estimated effect of content, consis-tent with indirect and direct network effects being orthogonal.

characteristics and instrumented relevant content inEquation (4) captures whether content plays a smalleror larger role as these characteristics vary. Althoughsome of these results are descriptive, others such asthose for international gateways are explicit hypothe-sis tests because “natural experiments” induce exoge-nous cross-country differences. Since we worry aboutthe quality of instruments available for these interac-tions in an HT specification, we use fixed-effects spec-ifications. Also, there are insufficient data to simulta-neously identify multiple interactions, so we estimateeach effect separately.31 Thus, the effects are not con-ditional on the other interaction effects.

31 In a regression combining all of the interaction effects, the coef-ficients on the interaction terms have similar magnitudes as in theseparate regressions, although not all of them are significant.

Dow

nloa

ded

from

info

rms.

org

by [

218.

189.

95.1

0] o

n 09

Jul

y 20

14, a

t 20:

46 .

For

pers

onal

use

onl

y, a

ll ri

ghts

res

erve

d.


Table 8 Effect of Content and Interactions Between Content and Country Characteristics on Internet Adoption for All Sample Countries, 1998–2004,Second-Stage, Fixed-Effects Estimates

1 2 3 4 5Per-Capita Telephone Gini IP Gateway

GDP interaction Infrastructure interaction Coefficient interaction Protection interaction Capacity interaction

Per-Capita GDP 000134∗∗∗

40000155Telephone Infrastructure −002064∗∗∗

40001865Log[Gateway Capacity] −000028

40000185Relevant Content 000027∗∗∗ 000013∗∗∗ 000028∗∗∗ 000020∗∗∗ 000017∗∗∗

40000055 40000045 40000055 40000045 40000045Per-Capita GDP × −000001∗∗∗

Relevant Content 40000005Telephone Infrastructure × 000071∗∗∗

Relevant Content 40000175Gini Coefficient × −000078∗∗∗

Relevant Content 40000195Intellectual Property Protection × −000003∗∗

Relevant Content 40000015Log[Gateway Capacity] × 000010∗∗

Relevant Content 40000055�� 00045 00045 00045 00045 00045� 00825 00837 00826 00826 00828R2 00918 00919 00919 00917 00918N 1,169 1,169 1,169 1,169 1,169

Notes. Standard errors, clustered by country and allow for general heteroscedasticity, are in parentheses. All control variables shown in column 2 of Table 5, yeardummies, and dummy variables for missing values for all variables are included in all regressions. Prices and relevant content instrumented in all regressions.


These results have important implications for pub-lic policy. Content plays a larger role in driving adop-tion in poor countries, suggesting that direct net-work effects may play a larger role in rich countries.The results suggest that lowering income inequal-ity enhances content’s effect on adoption. Develop-ing both a ubiquitous domestic telephone networkand high-speed international links appears to enhancecontent access and stimulate adoption.

These results also have important implications forfirm strategies. They inform which countries Internetcontent providers should target in expanding inter-nationally. A country in which adoption is more sen-sitive to content suggests that its population findscontent more appealing. If so, countries with well-developed telecommunications networks and lowerincome inequality are better targets. Poorer coun-tries are also better targets, although revenue recov-ery is likely problematic. Our results for intellec-tual property (IP) protection have interesting policyand strategy implications. Not unexpectedly, weakerIP protection allows content to more heavily influ-ence adoption. Therefore, regulators face a trade-offin strengthening IP protection—while increasing theincentive for content creation, it discourages con-tent dissemination. For firms, targeting countries with

weaker protections is a good strategy if the firm cansufficiently protect its own content.

Per-Capita GDP: Column 1 of Table 8 shows thatrelevant content’s effect on adoption declines in acountry’s wealth. A one standard deviation increasein relevant content increases adoption by 1.9 percent-age points less (19.0%)32 for a country one standarddeviation above the mean per-capita GDP than for acountry at the mean. Although there are alternativeexplanations, one possibility is that adoption in poorcountries relies more on externally produced content,whereas adoption in rich countries depends more ongreater direct network effects within the country.

Telephone Infrastructure: Column 2 of Table 8 showsthat relevant content’s effect on adoption increases ina country’s telephone infrastructure quality (althoughthe direct effect remains negative). A one standarddeviation increase in relevant content increases adop-tion by 2.4 percentage points more (24.7%) for a coun-try one standard deviation above the mean level oftelephone main lines in use than for a country atthe mean. Because the telephone network is the pri-mary means of Internet access during our sample

32 All comparisons in this section are to the average adoption level(0.099) in the sample.

Dow

nloa

ded

from

info

rms.

org

by [

218.

189.

95.1

0] o

n 09

Jul

y 20

14, a

t 20:

46 .

For

pers

onal

use

onl

y, a

ll ri

ghts

res

erve

d.


period, this is consistent with a more pervasive net-work allowing widespread content access to therebystimulate adoption.

Gini Coefficient: Column 3 of Table 8 reveals thatgreater income inequality in a country dampens con-tent’s influence on adoption. A one standard devia-tion increase in relevant content increases adoption by1.4 percentage points less (13.8%) for a country onestandard deviation above the mean Gini coefficientthan for a country at the mean. This is consistent withmore evenly distributed wealth leading to a broaderdesire to access content. Although we find no directeffect of income inequality on adoption, there is anindirect negative effect via content sensitivity.

Intellectual Property Protection: Column 4 of Table 8investigates the role of a country’s IP protection basedon the intellectual property rights (IPR) componentof the Intellectual Property Rights Index (Horst 2006).The index rates each country’s level of IP protectionon a 0 to 10 scale, with 10 being the strongest. Greaterprotection diminishes content’s influence. A one stan-dard deviation increase in relevant content increasesadoption by 0.9 percentage points less (9.4%) for acountry one standard deviation above the mean IPRthan for a country at the mean. This is consistent withgreater protection, making content less freely avail-able to stimulate adoption. Of course, weaker IP pro-tection reduces the dynamic incentives to create con-tent, but our results suggest that it stimulates usageof extant content.

High-Speed Infrastructure: During our sample period,more than 95% of Internet traffic between countriestraveled over submarine cables (Carter et al. 2009).Landing points for these high-speed cables must be incountries adjacent to the ocean. As a result, landlockedcountries connect through generally slower terrestrialcables to access external content providing exogenousdifferences in geographic advantage. This allows us toestimate the causal indirect network effect of interna-tional gateway capacity.

We identified the major telecommunications sub-marine cables, their years of operation, capacity, andlanding points (see Online Appendix B for sources).From this we calculated each country’s gatewaycapacity in each year and interacted it with relevantcontent. The results are shown in column 5 of Table 8.International gateway capacity has an insignificantdirect effect on adoption. This is consistent with inter-national gateways being located exogenously—basedon geography rather than Internet access demand.However, adoption in a country with greater capac-ity is more affected by relevant content than is acountry with lower capacity. A one standard devia-tion increase in relevant content increases adoption by1.9 percentage points more (19.1%) for a country one

standard deviation above the mean log capacity thanfor a country at the mean.33

Managerial: Our results provide some rough guide-lines for firms evaluating content investments. Weestimate an elasticity of 0.31 of adoption with respectto number of hosts. This is the effect on the extensivemargin (an increased number of adopters); assumingthat the usage of these additional adopters is spreadacross websites in proportion to their content, thistranslates into an elasticity of usage on a particularwebsite.34 If traffic is the goal (as it is for most Internetfirms), then increasing content by a given percentagehas about one-third the effect of increasing adoptionby the same percentage. Therefore, investments can beevaluated by comparing the marketing cost of increas-ing the user base by a certain percentage to the cost ofincreasing content by the same percentage. If this ratioexceeds about three, then the firm should focus oncontent production—otherwise, adoption. For firmsrelying on user-generated content, our estimates pro-vide a means for adjusting a user’s lifetime value tothe company. On average, each percentage increase inthe user base will ultimately yield roughly 1.3 timesthat because of the increased adoption from contentthat these users create. Of course, not all content iscreated equal. Higher-quality or more-targeted con-tent will lead to a greater elasticity and lower-qualityor less-targeted content to a smaller elasticity.

It would be preferable to have firm-level estimatesof the effect of content on usage. This is possible givenfirm-level data during episodes in which a firm addsdiscrete chunks of content but does not otherwisealter its marketing efforts to encourage usage. Suchestimates will reflect not only net increases in aggre-gate usage but also business-stealing effects (usagediverted from other Internet sites).

7. ConclusionInternet content plays a significant role in stimulat-ing Internet adoption. Its effect is on par with manyother important social, demographic, and economicfactors. Thus, content can play a crucial policy rolein encouraging Internet diffusion even in the shortrun, and some countries are already taking action.ITU, the UN body responsible for information tech-nologies, reports that “some countries are launching

33 An alternative explanation is that countries adjacent to the oceanwere easier to colonize and gained closer associations with largecountries, sharing the same language. We estimated our results incolumn 3 of Table 5 excluding small countries adjacent to the ocean.The results were virtually identical.34 The effect on the intensive margin—increased usage by preexist-ing adopters—means that this will understate the elasticity of usagewith respect to content. A firm-level elasticity of usage would bestill greater because it includes business-stealing effects (shiftingusage from other sites without increasing aggregate usage).

Dow

nloa

ded

from

info

rms.

org

by [

218.

189.

95.1

0] o

n 09

Jul

y 20

14, a

t 20:

46 .

For

pers

onal

use

onl

y, a

ll ri

ghts

res

erve

d.


initiatives to subsidize the production of local contentin its initial stages. Several of them are also revis-ing and upgrading key legal instruments that wouldallow them to protect and promote the production oflocal content” (ITU 1999, p. 121).

Governments and NGOs can influence adoption,and thereby encourage social change, in other coun-tries through this mechanism. In fact, this is implicitin our estimation strategy. Policymakers can use con-tent targeted at particular countries and in the appro-priate language to stimulate adoption in countriesadversely affected by the global digital divide. Inter-net content can also play an important role in over-coming social isolation. Countries with more dis-parate language usage are more affected by contentthan are those with more homogeneous. More tar-geted Internet content is likely to have even greatereffects than we find since we treat all content in agiven language as equally relevant.

For Internet firms wishing to predict Internet adop-tion at the country level, we provide estimates for amore comprehensive list of factors driving adoption—factors varying rapidly over time as well as thosemore slowly changing. For firms attempting to tar-get countries with high Internet adoption rates, ourresults suggest that content will influence adoptionmore heavily in countries with lower income inequal-ity, better telephone infrastructure, weaker IP pro-tection, and larger gateways connecting the coun-try to externally produced content. This last effect islikely to increase in importance over time as Inter-net information includes more video and audio. Ourresults also suggest that targeting linguistically iso-lated populations offer higher expected usage of afirm’s content.

Because of the need to ensure exogenous changesin content production, we are unable to estimate theeffects of content on large-country adoption. This alsoprevents us from distinguishing the effect of contentproduced within a country from that produced exter-nally. To estimate this would require different datathan are available to us. It would require an exoge-nous change in content or its availability in largecountries that affects adoption only via content. Wecan speculate on a few possibilities. The official useof non-Latin characters in Web addresses became fea-sible in 2010 because of a regulatory change. Thispotentially provides an abrupt exogenous change inthe availability of preexisting Internet content withinlarge countries such as China, Russia, India, andJapan. At the same time, it does not directly affect theincentive to adopt Internet access. Discrete changes incountries’ intellectual property laws or their enforce-ment may suddenly increase or decrease availabil-ity of preexisting content within the country withoutotherwise changing incentives for Internet adoption.

Government subsidies to produce content or bring itonline might provide exogenous geographic variationif they target local populations (such as schools orlocal governments). These and other possibilities willhave to await future research.

AcknowledgmentsThe authors thank Steve Berry, Avi Goldfarb, GuidoMeyerhans, Hongbin Cai, Yuxin Chen, Li Gan, Fiona ScottMorton, Stéphane Straub, Noam Yuchtman, the editor, andtwo anonymous referees for helpful comments. They alsothank seminar participants at the Cheung Kong Gradu-ate School of Business, Peking University, Yale University,Southwestern University of Finance and Economics, Zhe-jiang University, and University of California, San Diego, aswell as participants at the Insitut d’Economie IndustrielleConference on the Economics of the Software and InternetIndustries, the Second Annual Internet Search and Innova-tion Conference, and the 2009 International Industrial Orga-nization Conference. They thank Wang Xin and Qin Mianfor excellent research assistance. All errors are their own.

ReferencesBergman MK (2001) White paper: The deep Web: Surfacing hidden

value. J. Electronic Publishing 7(1), http://dx.doi.org/10.3998/3336451.0007.104.

Bohn RE, Short JE (2009) How much information? 2009 reporton American consumers. Report, Global Information Indus-try Center, University of San Diego, La Jolla, CA. http://hmi.ucsd.edu/howmuchinfo_research_report_consum.php.

Campbell L, Grondona V (2008) Ethnologue: Languages of the World(review). Language 84:636–641.

Carrell S (2009) Website archives to be fast-tracked. Guardian(December 27), http://www.theguardian.com/books/2009/dec/27/libraries-internet.

Carter L, Burnett D, Drew S, Marle G, Hagadorn L, Bartlett-McNeilD, Irvine N (2009) Submarine cables and oceans—Connectingthe world. UNEP-WCMC Biodiversity Series Report 31,ICPC/UNEP/UNEP-WCMC, Lymington, UK.

Chinn MD, Fairlie RW (2006) The determinants of the global digi-tal divide: A cross-country analysis of computer and Internetpenetration. Oxford Econom. Papers 59:16–44.

Cohen N (2009) Wikipedia looks hard at its culture. New YorkTimes (August 30), http://www.nytimes.com/2009/08/31/business/media/31link.html.

Corts K, Lederman M (2009) Software exclusivity and the scope ofindirect network effects in the U.S. home video game market.Internat. J. Indust. Organ. 27:121–136.

DiMaggio P, Hargittai E, Celeste C, Shafer S (2004) Digital inequal-ity: From unequal access to differentiated use. NeckermanK, ed. Social Inequality (Russell Sage Foundation, New York),355–400.

Dranove D, Gandal N (2003) The DVD-vs.-DIVX standard war:Empirical evidence of network effects and preannouncementeffects. J. Econom. Management Strategy 12:363–386.

Ford GS, Koutsky T, Spiwak LJ (2008) The broadband efficiencyindex: What really drives broadband adoption across theOECD? Phoenix Center Policy Paper No. 33 (May), http://www.phoenix-center.org/pcpp/PCPP33Final.pdf.

Gandal N (2006) Native language and Internet use. Internat.J. Sociol. Language 182:25–40.

Gandal N, Kende M, Rob R (2000) The dynamics of technologicaladoption in hardware/software systems: The case of compactdisc players. RAND J. Econom. 31:43–61.

Dow

nloa

ded

from

info

rms.

org

by [

218.

189.

95.1

0] o

n 09

Jul

y 20

14, a

t 20:

46 .

For

pers

onal

use

onl

y, a

ll ri

ghts

res

erve

d.


Goolsbee A (2002) Does the internet make markets more com-petitive? Evidence from the life insurance industry. J. PoliticalEconom. 110:481–507.

Gordon RG, ed. (2005) Ethnologue2 Languages of the World, 15th ed.(SIL International, Dallas).

Gordon RJ (2000) Does the “new economy” measure up to the greatinventions of the past? J. Econom. Perspect. 14:49–74.

Gowrisankaran G, Stavins J (2004) Network externalities and tech-nology adoption: Lessons from electronic payments. RAND J.Econom. 35:260–276.

Hammarström H (2005) Review: General linguistics: Gordon(2005)—Ethnologue: Languages of the World, 15th edition. Lin-guist List 16 (September 12) Article 2637.

Hargittai E (1999) Weaving the Western Web: Explaining differencesin Internet connectivity among OECD countries. Telecomm. Pol-icy 23:701–718.

Hausman JA, Taylor WE (1981) Panel data and unobservable indi-vidual effects. Econometrica 49:1377–1398.

Horst AC (2006) Intellectual Property Rights Index (IPRI): 2007Report. Study, Americans for Tax Reform Foundation, Wash-ington, DC. http://www.americansfortaxreformfoundation.org/userfiles/2007_IPRI.pdf.

International Telecommunication Union (ITU) (1999) Challengesto the network: Internet for development, 1999. Report, ITU,Geneva.

Kiiski S, Pohjola M (2002) Cross-country diffusion of the Internet.Inform. Econom. Policy 14:297–310.

Litan RE, Rivlin AM (2001) Projecting the economic impact of theInternet. Amer. Econom. Rev. 91:313–317.

Lyman P, Varian HR (2003) How much information? Study, Schoolof Information Management and Systems, University of Cal-ifornia, Berkeley, Berkeley. http://www2.sims.berkeley.edu/research/projects/how-much-info-2003/.

Minges M (1999) Measuring the diffusion of the Internet. Presen-tation at INET’99: Dimensions of Internet Diffusion, June 23,Internet Society, Reston, VA. http://www.itu.int/ITU-D/ict/papers/1999/MM-Inet99-Jun99.ppt.

Mutume G (2006) Harnessing the Internet for development: Africancountries seek to widen access, produce content. Africa Renewal20(2):14–15.

Nair H, Chintagunta P, Dubé J-P (2004) Empirical analysis of indi-rect network effects in the market for personal digital assis-tants. Quant. Marketing Econom. 2:23–58.

Ohashi H (2003) The role of network effects in the US VCR market.J. Econom. Management Strategy 12:447–494.

Park S (2004) Quantitative analysis of network externalities incompeting technologies: The VCR case. Rev. Econom. Statist.86:937–945.

Paolillo JC, Das A (2006) Evaluating language statistics: The ethno-logue and beyond. Contract report for UNESCO Institute forStatistics.

Pratap J (2010) Qatar initiative to increase Arabic content on Inter-net. Gulf Times (February 10), http://www.gulf-times.com/site/topics/printArticle.asp?cu_no=2&item.

Scott Morton F, Zettelmeyer F, Silva-Risso J (2001) Internet carretailing. J. Indust. Econom. 49:501–519.

Shriver SK, Nair HS, Hofstetter R (2013) Social ties and user-generated content: Evidence from an online social network.Management Sci. 59:1425–1443.

Sinai T, Waldfogel J (2004) Geography and the internet: Is the inter-net a substitute or a complement for cities? J. Urban Econom.56:1–24.

Staiger D, Stock JH (1997) Instrumental variables regressions withweak instruments. Econometrica 65:557–586.

Stelter B (2011) F.C.C. push to expand net access gains help. NewYork Times (November 9), http://www.nytimes.com/2011/11/09/business/media/fcc-and-cable-companies-push-to-close-digital-divide.html.

Stone B, Helft M (2009) In developing countries, Web gores withoutprofit. New York Times (April 26), http://www.nytimes.com/2009/04/27/technology/start-ups/27global.html.

The Emirates Center for Strategic Studies and Research (2008)Arabic content on Internet0 0 0Obstacles and solutions. Ac-cessed February 10, 2010, http://www.ecssr.ae/CDA/en/FeaturedTopics/PFFeaturedTopics/0,176.

Uganda Communications Commission (2005) Funding and Imple-menting Universal Access: Innovation and Experience from Uganda(Fountain Publishers, Kampala, Uganda).

United Nations (2004) World population to 2300. Report, UnitedNations, New York.

Wallsten SJ (2005) Regulation and Internet use in developing coun-tries. Econom. Development Cultural Change 53:501–523.

Wallsten SJ (2007) Whence competition in network industries?Broadband and unbundling regulations in OECD countries.Study, Technology Policy Institute, Washington, DC, https://www.techpolicyinstitute.org/files/s8.pdf.

Waters D (2006) International net domains ‘risky.’ BBC News (Octo-ber 30), http://news.bbc.co.uk/2/hi/technology/6099370.stm.

World Development Indicators Database (2010a) The World Bank:Internet users (per 100 people): International Telecommu-nications Union. Accessed July 10, 2010, http://databank.worldbank.org/data/views/variableselection/selectvariables.aspx?source=world-development-indicators#s_i.

World Development Indicators Database (2010b) The World Bank:Internet users (per 100 people): International Telecommu-nications Union; and The World Bank: Population (total).Accessed July 10, 2010, http://databank.worldbank.org/data/views/variableselection/selectvariables.aspx?source=world-development-indicators#s_i.

Wunnava PV, Leiter DB (2009) Determinants of intercountry Inter-net diffusion rates. Amer. J. Econom. Sociol. 68:413–426.

Dow

nloa

ded

from

info

rms.

org

by [

218.

189.

95.1

0] o

n 09

Jul

y 20

14, a

t 20:

46 .

For

pers

onal

use

onl

y, a

ll ri

ghts

res

erve

d.

Documents

The Effect of Content on Global Internet Adoption and the Global