This article appeared in a journal published by Elsevier. The attached copy is furnished to the author for internal non-commercial research and education use, including for instruction at the authors institution and sharing with colleagues. Other uses, including reproduction and distribution, or selling or licensing copies, or posting to personal, institutional or third party websites are prohibited. In most cases authors are permitted to post their version of the article (e.g. in Word or Tex form) to their personal website or institutional repository. Authors requiring further information regarding Elsevier’s archiving and manuscript policies are encouraged to visit: http://www.elsevier.com/copyright
Transcript
This article appeared in a journal published by Elsevier. The attachedcopy is furnished to the author for internal non-commercial researchand education use, including for instruction at the authors institution
and sharing with colleagues.
Other uses, including reproduction and distribution, or selling orlicensing copies, or posting to personal, institutional or third party
websites are prohibited.
In most cases authors are permitted to post their version of thearticle (e.g. in Word or Tex form) to their personal website orinstitutional repository. Authors requiring further information
regarding Elsevier’s archiving and manuscript policies areencouraged to visit:
Usage and diffusion of cellular telephony, 1998–2004☆
Michał Grajek a,1, Tobias Kretschmer b,c,⁎a ESMT European School of Management and Technology, Schlossplatz 1, D-10178 Berlin, Germanyb Institute for Communication Economics, LMU Munich, Schackstrasse 4/III, D-80539 Munich, Germanyc Centre for Economic Performance, LSE, United Kingdom
a b s t r a c ta r t i c l e i n f o
Article history:Received 26 March 2007Received in revised form 14 August 2008Accepted 20 August 2008Available online 3 September 2008
We study the dynamics of usage intensity of second-generation cellular telephony over the diffusion curve.Specifically, we address two questions: First, can we draw conclusions about the underlying drivers oftechnology diffusion by studying usage intensity? Second, what is the effect of high penetration of previousgenerations and competing networks on network usage intensity? Using an operator-level panel covering 41countries with quarterly data over 6 years, we find that heterogeneity among adopters dominates networkeffects and that different technological generations are complements in terms of usage, but substitutes interms of subscription.
In this paper,we study thedynamics of usage intensity in the contextof second-generation cellular telephony. This is fairly unique as mostother studies on the success of newproducts or technologies look onlyatdiffusion speed (i.e. the slope of an S-shaped diffusion curve) andmaximum market size (i.e. the asymptotic number of adopters) as keyindicators for technological success. Studying usage intensity instead ofdiffusion speed has a number of advantages in addition to the scarcity of
research on this alternative measure of technological success. First,while the shapeof the diffusion curve almost always takes the formof anS, the dynamics of usage intensity is much more heterogeneous acrosscountries and operators. Second, the intensity with which a service isused is a more accurate indicator of the commercial success of atechnology than the mere number of subscribers, who may or may notbe using the service intensively. Third, usage intensity reveals informa-tion about the distribution of consumer preferences and the relativeimportance of network effects in away that the diffusion curve does not.Finally, examining usage intensity along the subscription diffusion pathhas the potential to indicate the existence of network effects withoutstrong assumptions on consumer heterogeneity.
Consequently, we address two specific questions in our paper:
(i) Can we draw conclusions about the underlying drivers oftechnology diffusion by studying usage intensity?
(ii) What is the effect of high penetration of previous generationsand competing networks on network usage intensity?
We use a panel with quarterly data from 41 countries and over 100cellular operators over a period of six years to address these issues. Ourresearch strategy is as follows: First, we run fairly standard diffusionregressions to determinea country's stage in the diffusionprocess at eachpoint in time because the answers to the two questions abovemay differdepending on the degree of penetration of mobile telephony.We use theresults from these diffusion regressions as input for the later stages of our
Int. J. Ind. Organ. 27 (2009) 238–249International Journal of Industrial Organization 27 (2009) 238–249
☆ We are grateful to Dr. Jan Krancke from T-Mobile International and Mark Burk fromInforma Telecoms & Media for making the data available to us and to seminarparticipants at Aston, CESPRI, ESMT, HEC Paris, Hong Kong University, LSE, Mannheim,LMU & TU Munich, Tanaka Business School, Tokyo University, WZB, the 2006International Industrial Organization Conference (Boston), the Telecom Paris Conferenceon the Economics of ICT and EARIE 2006 (Amsterdam) for comments. Editor DanAckerberg, Pedro Pita Barros, Jonathan Beck, Rafael Gomez, Pai-Ling Jin, KatrinMühlfeld,Martin Peitz, Konrad Stahl, Frank Verboven and an anonymous referee deserve specialthanks for their insightful comments. Financial support from the NET Institute, www.NETinst.org is gratefully acknowledged.MichałGrajek gratefully acknowledges financialsupport from the German Federal Ministry of Education and Research (Project01AK702A).⁎ Corresponding author. Institute for Communication Economics, LMU Munich,
j ourna l homepage: www.e lsev ie r.com/ locate / i j io
Author's personal copy
research. Second, controlling for potential problems of endogeneity, weregress usage intensity on three groups of potential drivers: i) the pricesof own, competing, and previous-generation services; ii) the installedbases of these three services; iii) a set of controls to capture the effects of acountry'swealth, the age of cellular service in general, and the proportionof prepaid users on the network. The use of operator-specific data lets ustease out the role of consumerheterogeneityandnetwork effects to somedegree. Specifically, ownnetwork size captures thenet effect of consumerpreferences andoperator-specificnetwork effects,while thenetwork sizeof competing operators proxies technology-wide network effects. As afinal step, we allow for time-varying coefficients on our variables ofinterest by interacting the results from our first regression with thevariables in question.
Our first major result is that consumer heterogeneity plays a muchmore important role than network effects in determining usageintensity for an individual operator. This is reflected in a negative effectof additional users in one's own network on average usage intensity.Second, network effects donot seem to operate across different operatornetworks, since the installed base of competing networks does not affectthe usage intensity in the focal network. Third, we find that fixed-linetelephony acts as a complement in the usage intensity of cellulartelephony, as evidenced by negative cross-price effects. Finally, we findevidence of fixed-mobile platform substitution, as lower fixed-linemarket penetration implies more cellphone usage. These findings areboth novel and illuminating for those eager to find more completemeasures of technological success.2
Ourpaper is oneof the few to considerusage intensityas ameasureofsuccess for an emerging technology.3 By contrasting our results on usagedynamics with results on diffusion speed in cellular telephony (e.g.,Doganoglu and Grzybowski, 2007; Grajek, 2007; Liikanen et al., 2004;Koski andKretschmer, 2005),we can see if ourmeasures of technologicalsuccess are correlatedwith those in the conventional diffusion literature.Our paper is also (to our knowledge) the first to use aggregate usage datato draw conclusions on the preference distribution of users. Further, byanalyzing complementarity and substitutability across different techno-logical generations, our work offers new insights on usage dynamics fornew technologies in the presence of imperfectly compatible technolo-gies (Grajek, 2007). Finally, by taking an explicitly dynamic approach inour analysis, we allow some effects to vary over time to differentiatebetween usage determinants in early and late adoption stages.
Apart from the existing work on cellular diffusion speed, our workalso complements more structural approaches of studying diffusionspeed and usage intensity (Ackerberg and Gowrisankaran, 2006; Ryanand Tucker, 2007), who use microdata at the adopter level andstructural models to estimate structural parameters, especially onnetwork effects (Ackerberg and Gowrisankaran, 2006) and consumerheterogeneity (Ryan and Tucker, 2007). While our reduced-formapproach does not yield precise estimates of the structural parametersof the market we study, our approach does not rely on detailedmicrodata to draw conclusions about the relative importance ofcompeting effects and the existence of cross-network effects. Wethus believe both approaches to be complementary.
Our paper is structured as follows: We describe the global cellulartelecommunications industry in Section 2 and discuss potentialdeterminants of usage intensity in Section 3. We then introduce ourdata and give descriptive statistics in Section 4. As part of thedescriptive statistics, we also run cellphone diffusion regressions,which facilitate subsequent empirical analysis. Our results arepresented in Section 5 and discussed in Section 6. Section 7 concludes.
2. The global cellular telecommunications industry
The general features and recent history of the cellular telecommu-nications industry are discussed in detail in Grajek (2007), Koski andKretschmer (2005) and Gruber and Verboven (2001). We thereforeonly provide a brief history of the technological improvements andcorresponding generation changes in cellular telephony over time.
In most countries, cellular phones were first available to endconsumers in the 1980s. The early technology was based on analoguesignal transmission, which was relatively inefficient and unreliable. Insome countries, first-generation (1G) cellular networks reached theircapacity relatively quickly, leading to lower service quality andcongestion, particularly for initiating calls. As soon as digital technology(second generation, 2G) had matured enough to present a crediblealternative to analogue cellular, it was introduced gradually across theworld (Dekimpe et al., 2000). Several different technological standardsexisted in different countries – for example, GSM in Europe, PDCS inJapan – and some countries – most notably the US – even introducedseveral standards in one country. Technological competition betweenstandards within countries has slowed down overall diffusion (Koskiand Kretschmer, 2005), but may have had the long-term effect offostering technological progress for future generations (Cabral andKretschmer, 2007). In addition to 2G's improved reliability and networkcapacity, 2G phones also had SMS functionality, which enabled users tosend short text messages to each other and was a huge success amongyounger users, especially in Asia and Europe.4 Following the success of2G, a third generation with more advanced data transmission facilitieswas developed and is currently being rolled out.
For our sample period 1998–2004, 2G cellular was dominant.Second-generation telephony itself displayed significant technologicalprogress, with handsets becoming smaller and containing an increas-ing number of additional functions (Koski and Kretschmer, 2007). Inaddition to ongoing technological innovations on the product side,pricing and services became increasingly sophisticated. First-genera-tion cellular phones were targeted at business consumers for severalreasons: First, handsets were rather heavy and used predominantly incars,5 which, combined with very high tariffs, appealed mainly tobusiness users. With the introduction of digital cellular telephony,however, operators focused on capturing themassmarket tomake thetechnology succeed commercially.
2.1. Penetration pricing
Early attempts by second-generation cellphone operators wereaimed at gaining a critical mass of consumers. Since later adopterswould be basing their adoption decisions on those of early adopters,operators were pricing aggressively to grow their installed base. Withlock-in contracts over one, sometimes two years, this strategy wasprofitable (Farrell and Klemperer, 2007).
2.2. Handset subsidies
Most cellularhandsetswere, and still are, heavily subsidized. Thiswasa strategy to get consumers to adopt in the first place, as handsets weretypically themostexpensivepart of getting a cellphoneconnected.6 Basichandsets are often given away “for free”7 if the consumer signs a long-
2 Comin et al. (2006) develop a framework of technological diffusion inwhich extensive(adoption) and intensive (usage intensity) dimensions of technology diffusion areseparated. Their results suggest that the existing preconception of S-shaped diffusionholds only if we look at the extensive margin, i.e. first adoption, while looking at theintensive margin, i.e. usage intensity, may generate quite different dynamics.
3 Exceptions include Ward and Woroch (2005) and Cabral (2006).
4 In the US, text messaging had not caught on and the average user was sending 203text messages a year compared to 651 in China in 2004, the end of our study period.(http://www.newsmax.com/archives/articles/2005/8/11/135257.shtml).
5 One of the largest cellphone retailers in the UK, founded in 1989, is still called“Carphone Warehouse”.
6 Operators used to charge a one-off connection fee, but competition amongoperators meant that this practice has mostly disappeared.
7 Most commonly, operators would charge £/$/€ 1 for a handset that typically cost about$100 to produce. There are even instances however of “paying” consumers to buya handset:In France, a Siemens S35was sold in connectionwith a contract for FFR190 and contained avoucher for a FFR200 reimbursement if sent to the mobile operator.
239M. Grajek, T. Kretschmer / Int. J. Ind. Organ. 27 (2009) 238–249
Author's personal copy
term contract. This is a particular form of product cross-subsidization toovercome the installed-base problem (Barros, 2006).
2.3. Prepaid contracts (pay-as-you-go)
Possibly the most successful strategy of moving cellular telephonyinto themassmarket was the introduction of pay-as-you-go contracts.These contracts involve no monthly fee, but a higher per-minute cost.Such contracts are especially attractive for low-frequency users forwhom a monthly fee would be too high given the few calls they makeor who do not have access to a bank account to set up amonthly debit.The introduction of pay-as-you-go tariffs coincided with a rapidincrease in diffusion speed, and most of the growth in later stages ofdiffusion came from prepaid users.8
2.4. Tariff proliferation
Finally, with an increase in competition and increasingly finemarket segmentation, the number of tariffs has proliferated enor-mously. This has two effects: First, it could serve as a collusive deviceby confusing consumers (Hörnig, 2005), and second, it could enableconsumers to make more fine-grained decisions based on theirexpected calling patterns (Miravete and Röller, 2004; Naranayan et al.,2007). The fact that consumers seem to switch quite readily betweencontracts to optimize their behavior (Miravete and Röller, 2004)suggests that consumers will have some degree of uncertainty abouttheir future calling patterns, but eventually settle on the contract thatsuits their consumption behavior best.
3. Determinants of cellular usage intensity
In this section, we identify and discuss potential determinants ofcellular usage intensity. Specifically, we examine the effects ofconsumer preferences, network effects, and substitute technologieson new technology usage.
3.1. Heterogeneous consumer preferences
The preference distribution of current users of a technologywill affectthe usage intensity of a particular technology at any given time. Interest-ingly, the twomost common theories of technology diffusion have differ-ent implications for usage intensity over the diffusion path (Cabral, 2006).Theepidemicmodel assumes that all usershave identical preferences foranew technology, and theS-shapedpath is generatedby thedifferent timesat which adopters learn about the technology (Geroski, 2000).9 Con-versely, theheterogeneitymodel postulates that adopters adopt accordingto their preferences,with the highest-preference adoptermovingfirst, thesecond-highest moving second and so on—the “rank effect” (Karshenasand Stoneman,1993). The rank effect implies a decrease in usage intensityover the diffusion curve if, as seems reasonable, the preference to adoptearly is correlated to using the technology intensively. We do not have adirect measure of consumer preferences that would let us measure therank effect directly, but if consumers are heterogeneous, later adopterswill have lower preferences than earlier ones, so that average usage willdecrease as more low-preference adopters join the network. We alsocontrol for operators' share of prepaid users since prepaid consumerstypically have lower usage intensity than postpaid ones.
3.2. Network effects
Network effects generally make usage of a technology moreattractive since there are more potential communication partners (inthe case of direct network effects) or awider variety (or cheaper supply)of complementary products. In cellular telephony, direct network effectsmay operate acrossmultiple operators and technologies (since users of aparticular network can call users from other networks and even fixed-line numbers), while indirect network effects may operate predomi-nantly on the operator level (via provision of operator-specific content,ringtones, services etc.). In general, more users of a technology will notonly make initial adoption more attractive (which has been widelydocumented in the literature), but also increase usage intensity ofexisting users. The degree of compatibility, or the extent to which usersview two competing networks as substitutes,will determine the relativemagnitudes of the effect of additional subscribers to one's own networkand to a competing network.
3.3. Substitute technologies
In network industries, substitutes in the product market may carry adegree of complementarity via the network effect, especially regardingusage intensity. Consider a product with direct network effects.Communication partners can either be users of the same network, of acompeting network, or of the previous generation's network. The more
Table 1Variable definitions and descriptive statistics
Variable Definition Observations Operators Mean Standard deviation Min Max
MoU Average monthly minutes of use 2146 114 174.52 115.54 51 960CellP Average revenue (US cents per minute) 2052 109 22.03 10.69 3.26 114.00FixedP Price of a local fixed-line connection (US cents per three minutes) 2839 157 8.33 5.21 0 19CellSubs(j) Own subscribers as population's share (%) 3110 150 12.70 12.11 0.002 54.58CellSubs(− j) Subscribers to competing operators as population's share (%) 3110 150 31.50 22.47 0.04 99.49FixedSubs Fixed-line subscribers as population's share (%) 3199 157 40.96 20.91 2.20 75.76Prepay Share of prepay users among own subscribers (%) 3110 150 43.54 29.95 0 100OnAir Time since the launch of service (quarters) 3444 150 15.94 13.34 0 50GDP GDP per capita (000′s US dollars) 3561 157 17.79 13.01 0.36 51.98Stage Diffusion stage indicator (1 after a country reached the inflection
point of the cellular telephony diffusion; 0 otherwise)3605 157 0.66 0.47 0 1
Table 2Descriptive statistics by year (variable definitions as in Table 1)
8 This process is currently in reverse as operators try to get pay-as-you-go users toswitch to monthly fee contracts.
9 This is the simplest version of the epidemic model, which assumes no externalsource of information and identical adopters. The resulting diffusion curve issymmetric around 50%, although advanced models can generate also asymmetricdiffusion patterns.
240 M. Grajek, T. Kretschmer / Int. J. Ind. Organ. 27 (2009) 238–249
Author's personal copy
overall users and potential communication links, the more intensivelyusers of any particular network will use the technology. However, thiswould not hold if the competing services were incompatible. In cellulartelephony, complete incompatibility is unlikely, although Grajek (2007)finds that inter-network compatibility is low, i.e. users do not viewcompeting operators as close substitutes, mostly due to different on-netand off-net prices.10 We therefore need to distinguish betweensubstitute technologies within the same and previous generations.
3.3.1. Intra-generational effectsThe literature on competition and its effects on product diffusion
finds that overall diffusion speed typically increases with competition(Koski and Kretschmer, 2005; Gruber and Verboven, 2001). This isattributed to price and non-price competition and increased (technol-ogy-wide) network effects. Further, an early build-up of consumerslocks in more consumers for the future.
The effect of competition on usage intensity is not obvious, however.First, since winning consumers often takes place through subsidizedhandsets, the marginal costs of making additional calls may well be thesame, suggesting no effect on the usage intensity of individualconsumers from increased competition. Second, since users will adoptif overall expected utility exceeds the costs of purchasing, loweringprices will attract adopters with lower preferences and therefore lowerexpected usage intensity. Hence, intense competition may have theeffect of the rapid adoption of a technology (i.e., a steep S-shapeddiffusion curve), but decreasing usage intensity because low-preferenceusers end up adopting more quickly than they would otherwise.11
3.3.2. Previous-generation substitutesExisting work on the effect of the installed base of an established
technologysuggests thata larger installedbase typicallyhinders transitionto a new technology. If users of the incumbent generation face some costof switching to the new generation, a large installed base may preventthem from doing so, and given network effects, the new technology maynot be adopted at all unless the degree of technological improvement inlarge enough (Farrell and Saloner, 1985; Shy, 1996). In markets withbackward compatibility, however, this result may be overturned. If earlyusers of the newgeneration can communicatewith “old” users, the start-up problem for the new generation may be alleviated. Koski andKretschmer (2005) show that in countries with a comparably largenumber of 1G mobile users, 2G cellular telephony diffused more quickly,which mirrors the results of Liikanen et al. (2004). On the other hand,Barros and Cadima (2000) and Sung and Lee (2002) show thatmobile andfixed-line telephony appear to be substitutes. Ward and Woroch (2005)find similar results, but study usage rather than adoption.
Overall, a complementary effect may only be relevant if a start-upproblem exists in the early stages of the new generation, whereas inlater stages, users may replace their fixed line with cellularconnections or new generations of buyers using only cellphones.Therefore, the effect of fixed-line prices and availability on cellularusage may vary over the diffusion curve.
4. Data
We draw our data predominantly from two sources: The InformaTelecoms & Media World Cellular GSM Datapack (Informa T&M) andMerrill Lynch's Global Wireless Matrix. Merrill Lynch, a US-basedinvestment bank, publishes aquarterly report on thedevelopmentof theglobal cellular telephonymarket as a service to their clients and industryobservers. Merrill Lynch reports, among other data, the total number ofminutes called per operator, which can be used to construct the average
usage per consumer. The Informa T&M data has been used in previousstudies (e.g. Koski and Kretschmer, 2005) and covers the number ofsubscribers for individual mobile operators, average prices andtechnological standards in considerable detail. Informa T&M is aprovider of market and business intelligence to commercial entities inthemobile and media industries. Their customers base commercial andmarketing decisions on the data, thus ensuring a high level of accuracy.To ensure that our data is reliable, we triangulated it with publiclyavailable sources (OECD's Communications Outlook, ITU's Telecommu-nications Indicators) and found that the common variables werecomparable. We are therefore quite confident that our data is accurate.
To complement ourmain data, we use IMF's International FinancialStatistics (for GDP) and World Bank's World Development Indicators(for population, fixed telephone lines, and average cost of a local call).Since the WDI only provides annual data, we linearly interpolate thevariables to obtain quarterly data.
4.1. Descriptive statistics
Data coverage varies by variable, but overall our sample coversmore than 100 network operators in 41 countries.12 Table 1 givesdescriptive statistics of our variables, and Table 2 tracks the variablesacross our sample period. The relevant time span (t) is one quarter.
4.1.1. Dependent variable (MoUijt)Our dependent variable is the average usage intensity per
consumer (measured in minutes) at the operator level.
4.1.2. Prices (CellPijt, CellPi(− j)t, FixedPit)Ourpricemeasures are the average prices for a one-minute-call for the
operator in question, the average of all competing cellular operators in thecountry, andfixed-lineprices, respectively. Cellularprices are calculatedasthe revenue from services divided by the total number of minutes. Fixed-line prices are given by the average cost of a local call (per 3 min).
4.1.3. Installed-base variables (CellSubsijt, CellSubsi(− j)t, FixedSubsit)We measure the installed bases by the number of subscribers for
one's ownnetwork, all othermobile operators, and the numberoffixedlines in the country all expressed in percentage of total population.
of a network's installed base to control for consumer heterogeneity tosome extent, the time an operator has been offering second-generation10 “On-net” prices refer to calls made to members of the same network, “off-net”
prices refer to calls to other cellular networks.11 This only holds, of course, if later adopters are indeed low-intensity users. Ourempirical results show that this is indeed the case, as can be seen in Section 5. 12 Our data doesnot contain informationonMVNOs (Mobile Virtual NetworkOperators).
Fig. 1. Canada, diffusion and average usage, 9/98–9/04.
241M. Grajek, T. Kretschmer / Int. J. Ind. Organ. 27 (2009) 238–249
Author's personal copy
cellular telephony (OnAirijt), GDP per head at the country level (GDPjt),and a country's stage in the diffusion process at time t (Stagejt).
The time trends in Table 2 show that the increased penetration ofcellphones in our sample coincides with a significant increase in theshare of prepaid consumers. We also find a downward trend incellular service prices, which flattens around 2002, and an upwardtrend in average usage. Contrary to cellular telephony, subscriptionof fixed lines decreases slightly over our study period, while fixed-line prices remain relatively constant. Stage is a dummy variablethat takes value 1 if cellular diffusion in a country is advanced andzero otherwise.13 As Table 2 reports, in the beginning year of oursample there were no advanced countries in terms of cellulartelephony diffusion, whereas in the last year almost all countrieshad reached an advanced stage.
Our data reflects some interesting dynamics in cellular telephonyin the late 1990s and early 2000s. Diffusion was rapid – averagepenetration rates increased almost four-fold over six years – andprepaid usage went from being an option chosen by one subscriber infour to the option preferred by half of all users. Of course, looking atsample averages will hide many idiosyncrasies, in particular some ofthe effects we are interested in. To illustrate the differences, weconsider the diffusion and usage patterns in two countries.
4.2. Two examples—Canada and Spain
The following figures plot diffusion and average usage in Canadaand Spain, respectively.
Canada's average usage intensity is constantly increasing in thetimeframe we study (Fig. 1), while average usage in Spain is firstdecreasing steadily and stabilizing around halfway through oursample (Fig. 2). OLS regressions confirm that a linear time trendyields a modest fit for Spain (Slope: − .821, adjusted R2= .404), whilethe fit is better for Canada (Slope: 8.189, adj. R2= .914). Including asquare term improves results for Spain to an adjusted R2 of .910, whilefit remains similar for Canada (adj. R2= .921).
The above descriptive statistics suggest that usage patterns varysignificantly across countries, despite the fact that diffusion is S-shaped in both countries.14 Clearly, these statistics should beinterpreted with caution since we do not control for important
country- and firm-level variables. For example, Canada operatedunder the RPP (Receiving Party Pays) principle, while Spain used CPP(Calling Party Pays), Canada had four competing operators comparedto Spain's three, and Canada's GDP per headwasmore than 10% higherthan Spain's. At any rate, the usage patterns shown suggest that usageintensity across countries is worth studying in more detail. Wetherefore first run a fairly standard diffusion regression for eachcountry before looking at determinants of average usage by cellularoperator in the next section.
4.3. Global diffusion of cellular telephony
The wide coverage of our sample means that we study countries atvery different stages of diffusion in our sample period. To deal withthis heterogeneity adequately, we generate an indicator about eachcountry's stage in the diffusion. To this end and to provide adescriptive summary of the diffusion process, we estimate acountry-level logistic diffusion equation of the form:15
SUBSt ¼ SUBS⁎
1þ exp −β t−τð Þð Þ ; ð1Þ
where SUBS⁎=γPOP.SUBSt denotes the number of subscribers at time t, and POP
measures the population of a country. The potential number ofadopters SUBS⁎, i.e. the saturation level to which SUBSt converges, isa fraction γ of the total country's population. The other twoparameters in Eq. (1), τ and β, denote timing and speed of diffusion,respectively. That is, τ indicates the inflection point of the logisticcurve, while β gives the growth rate of SUBSt relative to its distanceto the saturation level, i.e. dSUBSt
dt1
SUBSt¼ β SUBS⁎−SUBSt
SUBS⁎. The growth rate
reaches its maximum (β/2) at the inflection point t=τ. Table 3presents Nonlinear Least Squares (NLS) estimates of the country-specific regressions, where τ is measured in quarters starting from 0in the 1st quarter of 1960: the average τ is approximately 163,which corresponds to the 4th quarter of 2000—the average countryin our sample reaches its inflection point in late 2000. There aresignificant differences across countries, however: In Finland, ourestimates suggest that diffusion speed reaches its maximum about2 years earlier (τ=154.6) than the average, while in Russia, theestimated inflection point was in late 2004 (τ=178.9). To illustratethe different diffusion stages across countries in our sample, we pickthree country groups – leaders, followers, and laggards, based onour estimates of τ – and plot actual and fitted penetration levels forthe three groups in Figs. 3–5.
Our regression estimates fit the actual diffusion curve well acrossdifferent stages of diffusion.16 We report the fixed-line penetrationratios in the last year of our sample in Table 3. By contrasting it withthe estimated cellular penetration thresholds (γ), we see that fixed-line and estimated cellular penetration ratios are correlated. Inparticular, the countries with very low γ (Argentina, China, andVenezuela) exhibit very low levels of fixed-line penetration as well.Overall, the average estimated fixed-to-mobile penetration ratio isaround 2, which is intuitively appealing as there was typically onefixed line per household, whereas with cellular telephony twomembers of a household may own a cellular phone.
Since the logistic diffusion equation is symmetric around theinflection point τ, it naturally defines the stage of the diffusion. Asour sample countries are at very different stages of diffusion andsome of our determinants may have different effects along thediffusion curve, we account for this by allowing for time-varying
Fig. 2. Spain, diffusion and average usage, 9/98–9/04.
13 The variable Stage will be defined later on, when we describe country-wisecellphone diffusion.14 Note that we are not capturing a complete S-shape in our data. The two countrieschosen here for illustrative purposes have roughly linear growth during our studyperiod, i.e. we are capturing the linear part in the middle of the diffusion curve forthese countries.
15 Beck et al. (2005) discuss this diffusion equation and contrast it with the otherscommonly used in the literature.16 We do not report the high R2 obtained in Table 3, as they are common in such non-linear time-series models and do not necessarily indicate a good specification(Trajtenberg and Yitzhaki, 1989).
242 M. Grajek, T. Kretschmer / Int. J. Ind. Organ. 27 (2009) 238–249
Author's personal copy
effects in our regressions. We therefore define the variable Stage,which takes value 1 if a country's cellular diffusion has reached theestimated inflection point and zero otherwise, and interact it withour variables of interest.17 For two countries (Columbia and India)the nonlinear estimation procedure did not converge, as 2Gcellphones were just taking off in 2004. We then set the Stage toequal zero for them.18
5. Empirical specification and results
5.1. Usage regressions
We use average monthly minutes of use per subscriber asdependent variable. Note that our dataset lets us run usageregressions at the operator level. This is useful since operators in thesame country may have different characteristics, for example theproportion of prepaid users or the installed base of subscribers, both ofwhich are expected to affect the average usage of a particular operator.Also, by including average prices by operator we can uncover own-price and cross-price effects on usage intensity.
Our baseline specification of cellular phone usage reads as follows:
where subscripts i, j, and t stand for country, cellphone operator, andtime, respectively. The dependent variable is the average usage persubscriber. We consider both own- and cross-price effects oncellphone usage by including an operator's (j) own price, the averageprice of other cellphone operators in the country (− j), as well as theprice of a local fixed-line connection, in our regressions. Similarly, wedistinguish between an operator's own network of subscribers,subscribers to other cellphone operators, and fixed-line subscribers.To facilitate comparison across countries, all price variables are in UScents and the installed-base variables are in percent of the country'stotal population.
The vector Xijt contains a set of control variables: GDP per capita,the share of prepaid-card users in the own subscriber base, and thetime on air. Finally, the αs capture the unobserved heterogeneityacross countries and operators driven by different pricing regimes(Receiving Party Pays vs. Calling Party Pays), different tastes forcommunications services (Italians tend to talk more than Swedes),incumbents' first-mover advantages or operators' brand reputation,and other time-invariant country and operator-specific effects.19
5.2. Expected effects
Based on our discussion in Section 3, we now briefly summarizethe expected effects on usage intensity of the variables we use in ourestimations.
5.2.1. Own subscribers (CellSubsijt)The number of subscribers of one's own and substitute networks
are our main variables of interest to capture the effects of consumerheterogeneity and network effects on usage intensity along thediffusion path. Its sign depends on the presence (or absence) ofdifferent factors and the underlying diffusion mechanism. If diffusionis driven by consumer heterogeneity – accompanied by falling priceand/or increasing quality over time – we expect a negative coefficienton the network size variable since subscribers joining the networklater have a lower preference for the product and thus decreaseaverage usage. If, however, strong network effects are present,increasing communications opportunities due to growing networksize may offset the rank effect leading to increasing usage intensity(Cabral, 2006). Depending on which effect dominates – rank/consumer-heterogeneity effect or network effect – the own cellularnetwork variable will carry a negative or positive sign, respectively.20
5.2.2. Competing network size (CellSubsi(− j)t)Network effects will also be present for subscribers of other
operators (− j) within a country if they originate from cellular userscalling each other across different operator networks. This is becauseadditional subscribers to competing cellular networks increase overallcommunication opportunities while leaving the composition of theown subscriber base unchanged. Conversely, the competing networksize variable also captures the substitution effect between thetechnologies. Although holding fixed-line and cellular connectionsat the same time – a prerequisite for usage substitution betweenplatforms – is much more common than holding cellular connections
17 Experiments with defining “advanced diffusion” at a later (or earlier) stage yieldeffectively the same results.18 Omitting these countries does not change our results.
19 The operator-specific effects would also pick up systematically different consumergroups by operators. If, for example, one operator were especially successful inattracting the high-usage bracket of a particular consumer group, this would show upas a positive fixed effect.20 Word-of-mouth (or epidemic) diffusion models do not deliver any predictionconcerning the usage intensity of a diffusing technology.
a Missing coefficient indicate that the NLS estimation procedure did not converge.
243M. Grajek, T. Kretschmer / Int. J. Ind. Organ. 27 (2009) 238–249
Author's personal copy
with different operators, the latter is also observed, in particular inmature cellular telecom markets (Wireless Intelligence, 2006;Doganoglu and Wright, 2006). We therefore expect the sign on thecompeting network size to be determined by the relative importanceof the network and the substitution effects.21
5.2.3. Fixed-line network size (FixedSubsit)The arguments for competing network size also apply for fixed-
line network size. Here however, we expect substitutability to bemorepronounced as users are more likely to simultaneously hold a fixed-line and a cellular connection than two cellular connections. Anegative effect of fixed-line network size on cellular usage wouldsuggest shifting usage between the platforms, i.e. a substitutiverelationship.
5.2.4. Share of prepaid users (Prepayijt)Prepaid consumers face higher marginal costs and lower fixed
costs, which is consistent with a lower average usage (Miravete andRöller, 2004). We therefore expect a negative effect of the share ofprepaid users on usage intensity as prepaid consumers are likely to below-intensity callers.
5.2.5. Own prices (CellPijt)Clearly, the price of a product and its substitutes (measured in our
study as the average revenue per minute) will have an effect on usageintensity. Controlling for other factors that might shift demandintertemporally (e.g. network or learning effects, quality improve-ments), we expect own price to have a negative impact on usageintensity.22
5.2.6. Competitors' prices (CellPi(− j)t)On the one hand, we can expect a positive coefficient (i.e. positive
cross-price elasticity) on the price of competing cellphone operators,as the services offered by competing operators are substitutes forconsumers holding multiple phones. On the other hand, however, anincrease in the price of competing cellphone operators will suppress
the incoming usage on network j. The balance between these twoeffects will therefore determine the sign of the coefficient.
5.2.7. Fixed-Line prices (FixedPit)The relationship between fixed-line and cellular phones is shaped
by similar effects. The empirical literature finds evidence of bothsubstitution and complementarity between fixed and cellular tele-phony by looking at subscription rates (see Ahn and Lee, 1999; Barrosand Cadima, 2000; Sung and Lee, 2002; Rodini et al., 2003).
5.2.8. Time on air (Onairijt)This variable measures the operator-specific time since the launch
of the service, i.e. the “age” of a service. The expected effect of anestablished network and technology is a gradual increase in the usageintensity since users get in the habit of calling on the move, and thenetwork may develop over time in terms of quality and brandreputation.23
5.2.9. GDP (GDPjt)Finally, we expect cellphone usage to exhibit a positive income
effect, captured by a positive coefficient on the level of GDP in acountry.
5.3. Econometric issues and estimation results
Our estimation strategy is as follows: To strip out operator-specificeffects αij, we apply fixed-effects (FE) as well as first-differenced (FD)estimation. In these regressions we do not correct for the possibleendogeneity problems. Comparing results across these two estimationtechniques is, however, a useful exercise, as under endogeneity theestimators have different probability limits, which gives a simple testof endogeneity (Wooldridge, 2002). We also address potentialendogeneity by using an instrumental variables (IV) approachaccounting for unobserved operator effects at the same time. To testthe robustness of our results, we also use a log–linear and a log–logspecification as alternatives to the linear specification in Eq. (2). Thelog–log specification is also useful as its coefficients can be interpretedas elasticities.
5.3.1. IdentificationTo identify own- and cross-price effects, and thereby the possible
complementarity or substitutability among operators, we need toaddress the likely endogeneity of our price variables, as prices may beset in direct response to a change in usage intensity. Making use of thepanel nature of the data, we construct instrumental variables based onthe geographical proximity between countries (see Hausman, 1997).To the extent that there are some common cost elements in thecellular service provision across regions (e.g., costs of equipment andmaterials), we can instrument for prices in a given country by averageprices in all other countries of the region.24 For instance, prices in theUK can be instrumented with a cellular price index for the rest ofWestern Europe. An important identification assumption we make,however, is that while the unobserved cost shocks are correlatedacross countries in a given region the unobserved demand shocks arenot. We believe that this assumption is reasonable given language andcultural differences across our sample countries. In particular,advertising campaigns – a common example of correlated demandshocks across states in the U.S. – will typically be designed and run atthe national level, so they are uncorrelated across countries. Thestrength of these geographical instruments depends on the extent to
21 The substitution effect could also resemble a rank effect if a larger installed base bya competitor grows by disproportionately adding higher-valuation consumers, leavingthe focal operator with low-intensity users. Assuming that the relative composition ofusers remains unchanged across operators over time (i.e. premium and budgetoperators remain in their respective segments of the market), our fixed-effect andfirst-difference regressions are expected to strip out most of this “fake” rank effect.22 Endogeneity and other econometric issues are covered in the next section.
24 The classification of countries into regions we apply follows the Informa T&Mclassification and includes: USA/Canada, Western Europe, Eastern Europe, Asia/Pacific,Africa, and Americas.
Fig. 3. Leaders' mobile penetration diffusion.
23 This is of course highly collinear with any time trend, so the two cannot be usedtogether. Using a time trend in place of time on air does not change our results.
244 M. Grajek, T. Kretschmer / Int. J. Ind. Organ. 27 (2009) 238–249
Author's personal copy
which the cost structure of 2G operators is correlated across countries.The existence of a global input market for the telecommunicationsindustry suggests that cost structures will be significantly correlated.
To obtain an operator-specific instrumental variable, we furthercondition on the technological standard used by each operator. Forinstance, we instrument for the price of a Chinese GSM operator withprices of GSM operators from other Asian-Pacific countries, the priceof a Chinese CDMAoperator with prices of CDMAoperators from otherAsian-Pacific countries, and so on. This allows us to obtain separateinstruments for own operators' prices and competitors' prices, at leastin countries where multiple 2G standards coexist. The instrument forfixed-line operators' prices is constructed in an analogous way.
As the decision whether to subscribe and how much to call is ajoint one, cellular and fixed-line network variables are also likely tosuffer from endogeneity bias in our usage equation—any omittedeffects that encourage both more intensive usage and new subscrip-tions (e.g. advertising campaigns) will lead to correlation between ournetwork variables and the error term. Similarly, consumers will self-select into calling plans, so the share of prepaid customers is notexogenous either. To address these problems we use lagged values ofthe network and prepaid variables as instruments. Since thisprocedure improves the endogeneity problem only to the extentthat residuals are not serially correlated, we report the residual serialcorrelation test along with our estimation results.
5.3.2. Main resultsThe first set of results is reported in Table 4. Columns (1) and (2)
report fixed-effects (FE) and first-differenced (FD) regression results,respectively. A useful indicator for the likely importance of endo-geneity problems is the extent to which results change across thesetwo econometric specifications (Wooldridge, 2002, p. 284). Althoughour results are not drastically different between the FE and FDspecifications, the cluster-robust Hausman test comparing the tworejects the null of exogeneity at the 1% significance level.
We also run first-differenced instrumental variable (IV FD)regressions with the set of instruments including prices in neighbor-ing countries and lagged values of network and prepaid variables, asdescribed. We use levels rather than first-differences of the instru-ments and include further lags to improve their strength. Column (3)in Table 4 reports the results. In general, our instruments performverywell: The F-statistics from the first stage regressions testing jointsignificance of the excluded instruments are all significant at the 1%level.25 Further, Hansen's J statistic, a valid overidentification test in
the presence of heteroskedasticity (Hayashi, 2000; Baum et al., 2003)does not reject the null. Finally, to test for serial correlation we re-estimate the model including the lagged residual (Wooldridge, 2002,p. 176). The coefficient of the lagged residual in the IV regressionreported in Table 4 is insignificant suggesting no serial correlation.Thus the use of lagged values of the network and prepaid variables asinstruments seems justified.
The IV results are close to the OLS results, but the IV coefficients aregenerally larger. In particular, the difference between our results incolumns (2) and (3) is statistically significant yielding further evidenceof an endogeneity problem in the OLS regressions. The Hausman testcomparing the two – again, we use its cluster-robust version – rejectsthe null of exogeneity at the 1% significance level.
Our control variables – the share of prepaid users, time on air, andGDP – have the expected signs (negative, positive, and positive,respectively) and are generally significant. Reassuringly, we find asignificant negative own-price effect on cellphone usage in all threeregressions. The cross-price effect is insignificant suggesting a lowdegree of substitution, consistent with Grajek (2007). The IVspecification suggests that a decrease in own price by 1 US centleads to an increase in the average monthly usage of a customer by6.6 min. Looking at fixed-line prices, we find weak evidence of usagecomplementarity—the coefficient is negative and significant at the15% level in the IV regression. We explore this tentative result in thenext set of regressions with time-varying coefficients.
Turning to the subscriber network variables, we find strongevidence of the consumer heterogeneity effect dominating thenetwork effect. The coefficient on own market penetration is negativeand significant in all specifications, which implies that additionalsubscribers to one's own cellular network significantly decreaseaverage cellphone usage. The magnitude of this effect is not marginal:From the IV specification we can see that average usage decreases by7 min per month with an increase of the penetration by 1 percentagepoint. In other words, to offset the effect of an additional percentagepoint of low-intensity users to an existing network, an operator needsto drop the prices per minute by roughly 1 US cent (the average priceper minute drops from over 35 cent to below 20 cent in our sampleperiod). The results in Table 4 also suggest that network effects do notexist across networks. Contrary to our expectations, the coefficient onthe installed base of cellular competitors is negative and significant atthe 10% level in the IV regression. One explanation of this result is(cellular) multihoming. The coefficient indicates that the higher thepenetration of competing networks, the lower the usage intensity ofone's own network—users use their second cellphone tomake calls, ordivide calls between them (Doganoglu and Wright, 2006).26
The apparent absence of network effects in cellphone usage issomewhat surprising given previous results that find network effectsto significantly contribute to the speed of cellular diffusion process(Grajek, 2007; Koski and Kretschmer, 2005; Gruber and Verboven,2001; Liikanen et al., 2004). However, a number of remarks are inorder: First, our results do not imply that there are no network effects.They merely suggest that they do not outweigh the consumerheterogeneity effect, and that no significant network effects seem tooperate across different cellular networks. Further, we consider adifferent dependent variable than existing studies. Thus, whilenetwork effects may be weak regarding usage intensity, they maywell be strong for first subscriptions. Finally, network effects may belimited to a small network of relevant users (Birke and Swann, 2006),
25 See the first stage results in Table A1 in the Appendix.
26 An alternative explanation would be that operators that have grown disproportio-nately over the sample period (since we already control for time-invariant differencesin installed bases) have also attracted comparatively more lucrative consumers, leavinglow-preference users to the focal firm. Given the extent of price competition in theindustry, however, we believe that growing market shares are often achieved byattracting lower-preference consumers. We therefore believe that empirically, multi-homing is a more appealing explanation for this result.
Fig. 4. Followers' mobile penetration diffusion.
245M. Grajek, T. Kretschmer / Int. J. Ind. Organ. 27 (2009) 238–249
Author's personal copy
which would lead to a small estimated network effect if network sizewere measured at the economy or the operator level.
Finally, the coefficient on fixed-line market penetration is negativeand significant in all three regressions. This indicates a degree of fixed-mobile platform substitution. It is interesting to contrast this findingwith our previous result on fixed-mobile usage complementarity asevidenced by the negative impact of fixed-line prices on cellphoneusage. It appears that the two communication platforms are comple-ments to the extent that keeping both fixed-line and cellularconnection at the same time is viable and attractive from a householdperspective. If users do not have access to a fixed line, however (eitherbecause fixed-line penetration is low or because users have starteddisconnecting their fixed lines), they satisfy all communication needswith their cellphones. This coexistence of the complementary and thesubstitution elements is in line with the results in Sung and Lee(2002), who report that the number of Korean mobile subscribers ispositively (negatively) correlated with the number of fixed-linedisconnects (new connections), suggesting substitution. However,the stock of fixed lines is positively correlated with the number ofmobile subscribers, indicating complementarity.
5.3.3. Time-varying effectsTo further investigate the relationship between old and new
telecommunications technologies, we allow for price and installed-base effects to vary over time. The motivation behind time-varyingcoefficients is that the relationship between old and new telecom-munications technologies or between competitors might changedepending on the diffusion stage of the new technology. For example,when penetration of cellphones is low, most cellular communicationmay go to (and from) fixed lines as there are limited opportunities forcellular–cellular interaction. Once cellular penetration approaches itsmaximum however, all communication needs can in principle besatisfied on the cellular network alone and fixed lines becomeobsolete. Therefore, the two technologies may change from beingcomplements initially (as fixed lines help cellular overcome theinstalled-base problem) to substitutes (as cellular phones replacefixed lines) later on. We therefore interact fixed-line prices and ourother variables of interest with a diffusion stage indicator. Theindicator is constructed from the estimates of the country-wisediffusion regressions (Table 3) and equals 1 in periods after a countryreaches the inflection point of cellular diffusion (τ) and zerootherwise. The sum of the early-stage (with Stage=0) and the late-stage (with Stage=1) coefficients is the net effect in the later stages ofdiffusion, while for the early stage only the first coefficient matters.We report our results in Table 5.
Because the variable Stage is a generated regressor we reportbootstrapped standard errors in Table 5. The variable Stage can also beexpected to endogenous in our equations, as it is merely a nonlinearfunction of network size. We control for this by treating Stage as anendogenous variable and instrument it along with the otherendogenous variables in our IV regression. The instrument we applyis the stage variable in countries with the same 2G introduction date(i.e. the date inwhich a positive number of 2G users is reported for thefirst time). Gruber and Verboven (2001) and Koski and Kretschmer(2005) show that the timing of introduction is correlated withdiffusion speed, i.e. at which point the inflection point is reached and“stage” becomes 1.27 The Hansen J statistic, however, reported alongwith the IV results in Table 5 is significant at the 10% level suggestingsome endogeneity concerns.
We find that own-price sensitivity increases in the later stages ofdiffusion, consistent with the addition of more low-intensity, high-elasticity users (over and above prepaid consumers, which we controlfor). This result is not very strong, however, as the change in own-priceeffect is significant only in the FE specification, i.e. column (1).Interestingly, our intuition of the fixed-mobile relationship changingfrom complements early on to substitutes in later stages of the diffusionprocess finds empirical support in the FD regression. The coefficient onfixed-line prices is significant and equals −3.2 initially and becomessignificantly positive (−3.2+4.5=1.2) in later stages of diffusion.
We also find some evidence of (cellular) multihoming becomingsignificant in the later stages of diffusion, as can be seen from thenegative coefficient on CellSubsi(− j)t interacted with Stage, although itis only significant in column (1). In the same regression the coefficienton the installed base of cellular competitors is initially positiveyielding some indication of network effects operating across cellularnetworks. Finally, we find that the penetration of fixed lines isnegative in the early stages of diffusion, but this effect wears off asdiffusion progresses (since the interacted variable is positive andsignificant). This may be because households cutting their fixed linerelatively late had low overall usage in the first place.
Byallowingour results tovary bydiffusion stagewefindsome resultsthat corroborate our intuition of the fixed–mobile relationship changingfrom complements initially to substitutes in later stages of the diffusionprocess. The time-varying coefficients also reinforceournotion that lateradditions to the network consist of low-preference users—own-priceelasticity seems to increase later on, and late adopters migrating fromfixed to cellular do not affect usage intensity much. Our indication thatmultihoming may become an issue later in the diffusion process willfurther reinforce the tendencyof decreasingARPU (AverageRevenuePerUser) for individual network operators, since the revenues of high-preference adopters are now split among different operators.28
6. Discussion of the results
The relationship between mobile usage and the network size isdetermined by two countervailing forces: Network effects andconsumer heterogeneity effects. Network effects arise as the growinginstalled base of subscribers allows them to satisfy more communica-tions needs.Hence, the averagenumberof calls increaseswithnetworksize. Consumer heterogeneity effects imply that usage of telecommu-nications services decreases with the installed base of subscribers, as
27 This procedure led us to identify nine groups of countries. Starting with the earlyadopters these groups read as follows: 1. UK, Denmark, France, Finland, 2. Germany,Sweden, Austria, Portugal, Italy, 3. Switzerland, Australia, Norway, 4. New Zealand,Ireland, Greece, Belgium, Turkey, Hungary, 5. Japan, South Africa, Thailand, Nether-lands, 6. Malaysia, Spain, Singapore, Czech Republic, US, 7. Korea, Canada, Poland,China, Egypt, India, 8. Chile, Mexico, Israel, Venezuela, Argentina, Russia, 9. Brazil,Colombia. Results on the strength of these instruments (all significant at the 1% level)are available from the authors.28 Our results are robust to log–linear and log–log specifications. The results areavailable from the authors.
Fig. 5. Laggards' mobile penetration diffusion.
246 M. Grajek, T. Kretschmer / Int. J. Ind. Organ. 27 (2009) 238–249
Author's personal copy
less eager (or poorer) users subscribe to the service over time and“dilute” average usage intensity as the installed base grows.
One of the problems in estimating the relative strength of theseeffects is that adding subscribers has the dual effect of enlarging anoperator's network and adding lower-preference users to the network.Our regressions suggest that in all specifications the heterogeneityeffectstrongly dominates the network effect since the coefficient on ownnetwork penetration is consistently negative and significant. Assumingthat the relative composition of operators does not change over time (i.e.operators donot switchmarket segments), a potential strategy to isolatethe network fromthe composition effectwas to consider the subscribersof competing networks since the composition of an operator's ownnetwork does not change while overall network size grows. In ourregressions, however, we find that competing network size does nothave a significant positive effect on own usage intensity.While this doesnot imply that there are nonetwork effects,we can at least conclude thatthey do not outweigh the heterogeneity effect, and that no significantnetwork effects operate across different cellular networks. If networkeffects exist, they do not appear to originate from opportunities to callcellular users on other networks. One interpretation could be thatsignificant network effects exist from sending and receiving textmessages to other cellular users, but not from calling them. Thiswould allow us to reconcile the fact that network effects are regularlyfound in adoption studies (e.g. Koski and Kretschmer, 2005; Gruber andVerboven, 2001) with the apparent absence of strong network effects inour usage intensity regressions. That is, adopting a cellphone becomesmore attractive if there aremanyothers to exchange textmessageswith,but this does not imply that users will call each other more.29
There is a growing literature on substitutability of fixed-line andcellular telephony. Our regression results suggest that one important
point is whether we consider telephone usage alone (given thesubscription decision) or usage and subscription as a joint decision.Controlling for the installed base of subscribers, we find some evidenceof fixed-mobile complementarity. However, the two telecommunica-tions platforms seem to be substitutes in terms of subscriptions, asfixed-line network size is negatively correlated with cellphone usage,which implies that in countrieswith a largefixed-linenetwork, cellularusers user their phones less (possibly because they use fixed-lineinstead). This suggests that an incumbent technology like fixed-linetelephony may foster diffusion at the start of cellular diffusion, but islikely to be replaced eventually as the new technology matures.
Finally, we find some support for a steady increase in usage aftercontrolling for other factors, as the positive coefficient of time on air incolumns (2) and (3) in Tables 4 and 5 shows. This suggests learningeffects and habit formation not picked up by our network size variables.
7. Conclusions and further research
We study the usage patterns of 2G cellular telephony over timeusing data from 41 countries over the 1998–2004 time period. Ourreduced-form regressions have uncovered a number of interesting
29 One might even suggest that text messages serves as substitute for calling.Unfortunately, our data does not allow testing for this.
Table 5Cellphone usage estimation results with diffusion stage interaction terms
(0.549) (0.330) (2.429)Average price of mobile competitors CellP(− j) 0.510 0.120 −1.960
(0.350) (0.153) (1.915)Price of local fixed-line connection FixedP −2.575 0.679 −8.466
(1.735) (0.702) (5.767)
Installed base effectsOwn penetration CellSubs(j) −1.251⁎ −1.534⁎⁎⁎ −6.990⁎⁎
(0.741) (0.452) (2.841)Penetration of mobile competitors CellSubs(−j) 0.455 −0.207 −2.870⁎
(0.565) (0.250) (1.666)Penetration of fixed line FixedSubs −2.667⁎⁎ −1.385⁎ −5.621⁎⁎
(1.198) (0.791) (2.223)
ControlsShare of own prepay users Prepay −1.092⁎⁎⁎ −0.149 −2.199⁎
(0.223) (0.098) (1.243)GDP GDP 3.743⁎⁎⁎ 1.552⁎⁎⁎ 3.841⁎⁎⁎
(1.028) (0.298) (1.339)Time since the launch of service OnAir −0.179 1.134⁎ 8.890⁎⁎⁎
(1.189) (0.590) (3.068)Hansen J statistic(degrees of freedom)
– – 8.89(7)
Observations 1314 1220 1064Clusters 91 91 90Functional form Linear Linear LinearEstimation method FE FD FD IV
⁎ pb0.1, ⁎⁎ pb0.05, ⁎⁎⁎ pb0.01; robust standard errors in parentheses.In the FE specificationwe report cluster-robust standard errors, which account for serialcorrelation in the error term.Operator-specific effects in the FE regression suppressed.
247M. Grajek, T. Kretschmer / Int. J. Ind. Organ. 27 (2009) 238–249
Author's personal copy
findings. First, it seems that consumer heterogeneity is considerableand network effects are moderate in comparison. Second, we findsome evidence of fixed-mobile usage complementarity in the earlystages of diffusion. At the same time we observe substitution of fixed-line with cellular minutes driven by changes in the fixed-linesubscriber base. This effect seems to wear off later as cellulartelephony becomes more established. These results are consistentacross most specifications and benefit from the use of instruments,suggesting that endogeneity needs to be accounted for.
We now outline a number of potential avenues for future research:
7.1. Functional form of network effects and heterogeneity
Our reduced-form approach does not allow a separate and preciseinterpretation of the shape of the preference distribution or thefunctional form of network effects individually, but rather the neteffect of both. One way to further separate out composition andnetwork effects would be to assume a sufficiently general distributionof preferences (taken, e.g., from Rogers, 2003), a functional form fornetwork effects (e.g., Swann, 2002), and a degree of compatibilitybetween the networks of different operators (e.g. Grajek, 2007). Asimilar approach has been taken by Ryan and Tucker (2007) andAckerberg and Gowrisankaran (2006), who develop structural modelsof adoption and usage behaviour using individual adopter data. Suchan approach would be complementary to ours, as the previouslyassumed strength of network effects is called in question to someextent by our results, suggesting that further research is called for.
7.2. Role of prepaid consumers
We find that the proportion of prepaid consumers has a negativeeffect on average usage, as expected. We do not, however, study indetail the origins and effects of the number of prepaid consumers incompetition between operators. For example, persistent first-moveradvantages may imply that later operators can only catch up byoffering prepaid services, which may in turn affect the first mover'sexisting users' incentives to call. In other words, the use of prepaidusers as a competitive tool and their contribution to network effectsseems an interesting line of research to follow up.
Learning about the shape of consumer preferences has significantimplications for firm and policymaker behavior. Strong consumerheterogeneity suggests that early adopters are more profitable thanlater ones—assuming that their decision to adopt earlier alsorepresents a higher willingness to make calls.30 This would makeintroductory pricing a double-edged sword: On the one hand,securing these early customers is likely to have long-term benefits,while on the other hand these early consumers are likely to constitutea large fraction of a firm's profits.31 Similarly, diffusion policies will beassessed on their expected impact on consumer surplus and firmprofits, which depends on the distribution of consumer preferencesand the intensity of network effects. Our results indicate that networkeffects are not overwhelming in determining usage, in which casepenetration pricing by operators significantly benefits early consu-mers (who get lower prices) rather than later ones (who do not benefitmuch from a larger network).
This study is the first to our knowledge that empirically tries todisentangle consumer heterogeneity and network effects on techno-logical diffusion using aggregate data on diffusion and usage. Ourstudy is also the first to allow for time-varying effects of an incumbent
technology, which has implications for policymakers and firms intheir incentives to phase out existing technologies. We believe thatwhile there have been a number of recent studies on the diffusion ofmobile telephony (including our own), recovering some informationon the underlying parameters and the subsequent causes of diffusionis a crucial next step in the study of new technologies and their successand impact on society.
Appendix A
Table A1Cellphone usage IV estimation results (first stage)
ΔCellP(j) ΔCellP(− j)
ΔFixedP ΔCellSubs(j)
ΔCellSubs(− j)
ΔFixedSubs
ΔPrepay
Included instrumentsΔGDP 0.349⁎⁎⁎ 0.335⁎⁎⁎ 0.025 −0.006 −0.035 −0.000 −0.086
⁎ pb0.1, ⁎⁎ pb0.05, ⁎⁎⁎ pb0.01; robust standard errors in parentheses.Δ stands for the change in value from t−1 to t.(t−1), (t−2) and (t−3) stand for lagged vales of one, two and three periods (quarters),respectively.
References
Ackerberg, D., Gowrisankaran, G., 2006. Quantifying equilibrium network externalitiesin the ACH banking industry. Rand Journal of Economics 37, 738–761.
Ahn, H., Lee, M.-H., 1999. An econometric analysis of the demand for access to mobiletelephone networks. Information Economics and Policy 11, 297–305.
Barros, P. (2006). Handset subsidies—an empirical investigation. Mimeo, UniversidadeNova de Lisboa.
Barros, P., Cadima, N. (2000). The impact of mobile phone diffusion on the fixed-linenetwork. CEPR Discussion Paper 2598.
Baum, C., Schaffer, M., Stillman, S. (2003). Instrumental variables and GMM: estimationand testing. Boston College Working Paper 545.
Beck, J., Grajek, M., Wey, C. (2005). Hypermarket competition and the diffusion of retailcheckout barcode scanning. WZB Discussion Paper SP II 2005-19.
Birke, D., Swann, P., 2006. Network effects and the choice of mobile phone operator.Journal of Evolutionary Economics 16, 65–84.
30 This would not be the case if early adopters had a high preference for incomingcalls (e.g. for emergency purposes), but not outgoing ones. However, we believe such apattern to be the exception rather than the rule.31 There is an extensive literature in marketing science that concerns itself with theoptimal pricing path of new products based on assumptions about the S-shapeddiffusion curve, but not on the origins of the S-curve (see, e.g., Krishnan et al., 1999).
248 M. Grajek, T. Kretschmer / Int. J. Ind. Organ. 27 (2009) 238–249
Author's personal copy
Cabral, L., 2006. Equilibrium, epidemic and catastrophe: diffusion of innovations withnetwork effects. In: Antonelli, C., et al. (Ed.), New Frontiers in the Economics ofInnovation and New Technology. Essays in Honour of Paul A. David, Edward Elgar,Cheltenham, Chapter 15.
Cabral, L., Kretschmer, T., 2007. Standardsbattles andpublic policy. In:Greenstein, S., Stango,V. (Eds.), Standards and Public Policy. Cambridge University Press, pp. 329–344.
Comin, D., Hobijn, B., Rovito, E. (2006). Five facts you need to know about technologydiffusion. NBER Working Paper 11928.
Dekimpe, M., Parker, P., Sarvary, M., 2000. Global diffusion of technological innovations:a coupled-hazard approach. Journal of Marketing Research 37, 47–59.
Doganoglu, T., Wright, J., 2006. Multihoming and compatibility. International Journal ofIndustrial Organization 24, 45–67.
Doganoglu, T., Grzybowski, L., 2007. Estimating network effects in the mobiletelecommunications industry in Germany. Information Economics and Policy 19,65–79.
Farrell, J., Klemperer, P., 2007. In: Armstrong, M., Porter, R. (Eds.), Handbook of IndustrialOrganization, vol. 3, Elsevier, Chapter 31. 1967–2072.
Geroski, P., 2000. Models of technology diffusion. Research Policy 29, 603–625.Grajek, M. (2007). Estimating network effects and compatibility in mobile telecommu-
nications. ESMT Research Working Papers, ESMT-07-001.Gruber, H., Verboven, F., 2001. The evolution of markets under entry and standards
regulation—the case of global mobile telecommunication. International Journal ofIndustrial Organization 19, 1189–1212.
Hausman, J., 1997. Valuation of new goods under perfect and imperfect competition. In:Bresnahan, T., Gordon, R. (Eds.), The Economics of NewGoods. University of ChicagoPress.
Hayashi, F., 2000. Econometrics. Princeton University Press.Hörnig, S. (2005). Equilibrium confusion? Mimeo, Universidad Nova de Lisboa.Karshenas, M., Stoneman, P., 1993. Rank, stock, order and epidemic effects in the
diffusion of new process technology. Rand Journal of Economics 24, 503–528.Koski, H., Kretschmer, T., 2005. Entry, standards and competition: firm strategies and
the diffusion of mobile telephony. Review of Industrial Organization, 26, 89–113.
Koski, H., Kretschmer, T., 2007. Emergence of dominant designs in mobile handsets.Industry and Innovation 14, 305–324.
Krishnan, T., Bass, F., Jain, D., 1999. Optimal pricing strategy for new products.Management Science 45, 1650–1663.
Liikanen, J., Stoneman, P., Toivanen, O., 2004. Intergenerational effects in the diffusion ofnew technology: the case of mobile phones. International Journal of IndustrialOrganization 22, 1137–1154.
Miravete, E., Röller, L., 2004. Estimating markups under nonlinear pricing competition.Journal of the European Economic Association 2, 526–535.
Naranayan, S., Chintagunta, P., Miravete, E., 2007. The role of self selection and usageuncertainty in the demand for local telephone service. Quantitative Marketing andEconomics 5, 1–34.
Rodini, M., Ward, M., Woroch, G., 2003. Going mobile: substitution between fixed andmobile access. Telecommunications Policy 27, 457–476.
Rogers, E., 2003. Diffusion of Innovations, Fifth Edition. Free Press, New York.Ryan, S., Tucker, C. (2007). Heterogeneity and the dynamics of technology adoption. NET
Institute Working Paper 06-26.Shy, O., 1996. Technology Revolutions in the Presence of Network Externalities. Inter-
national Journal of Industrial Organization 14, 785–800.Sung, N., Lee, Y., 2002. Substitution effects between mobile and fixed telephones in
Korea. Review of Industrial Organization 20, 367–374.Swann, P., 2002. The functional form of network effects. Information Economics and
Policy 14, 417–429.Trajtenberg, M., Yitzhaki, S., 1989. The diffusion of innovations: a methodical
reappraisal. Journal of Business and Economic Statistics 7, 35–47.Ward, M.,Woroch, G. (2005). Usage substitution between fixed andmobile telephony in
the U.S. Mimeo, University of Texas.Wireless Intelligence, . Multiple devices: real penetration in Western Europe. https://
www.wirelessintelligence.com.Wooldridge, J., 2002. Econometric Analysis of Cross Section and Panel Data. The MIT
Press, Cambridge.
249M. Grajek, T. Kretschmer / Int. J. Ind. Organ. 27 (2009) 238–249
