The Review of Economic Studies, Ltd.
Competition between Networks: A Study of the Market for Yellow Pages Author(s): Marc Rysman Reviewed work(s): Source: The Review of Economic Studies, Vol. 71, No. 2 (Apr., 2004), pp. 483-512 Published by: Oxford University Press Stable URL: http://www.jstor.org/stable/3700635 . Accessed: 18/01/2012 18:20 Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at . http://www.jstor.org/page/info/about/policies/terms.jsp JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship. For more information about JSTOR, please contact
[email protected].
Oxford University Press and The Review of Economic Studies, Ltd. are collaborating with JSTOR to digitize, preserve and extend access to The Review of Economic Studies.
http://www.jstor.org
Review of Economic Studies (2004) 71, 483-512 ? 2004 The Review of Economic Studies Limited
0034-6527/04/00210483$02.00
Networks: CompetitionBetween A Study of the Market for Yellow Pages MARC RYSMAN Boston University First version receivedAugust2000; final version accepted February2003 (Eds.) This paper estimates the importance of network effects in the market for Yellow Pages. I estimate three simultaneousequations:consumerdemandfor usage of a directory,advertiserdemandfor advertisingand a publisher'sfirst-ordercondition(derivedfromprofit-maximizingbehaviour).Estimation shows that advertisersvalue consumerusage and that consumersvalue advertising,implying a network effect. I find that internalizingnetwork effects would significantly increase surplus. As an application, I consider whether the marketbenefits from monopoly (which takes advantageof network effects) or oligopoly (which reducesmarketpower). I find that a more competitivemarketis preferable.
1. INTRODUCTION This paper measures the importance of a positive network effect in the market for Yellow Pages directories. Publishers of Yellow Pages directories face a "two-sided market": consumers value directories for information and retailers value directories as a way to advertise to consumers. More advertising leads to more consumer usage which in turn leads to more advertising, so consumer behaviour and advertiser behaviour together create a positive network effect.1 In fact, telephone company directories tend to have much higher prices, larger books and more usage than independent producers, suggesting that network effects are important in determining market structure.2 Because data is available on consumer usage as well as on prices and quantities of advertising, data is available on "both sides" of the feedback loop. This feature allows for the explicit estimation of a feedback loop in a way that has not been done before. To accomplish this goal, I estimate two demand curves simultaneously: the first is a consumer demand curve for directory usage as a function of advertising. The second is an inverse advertiser demand curve for advertising as a function of consumer usage and the quantity of advertising. I find that the amount that consumers use a directory increases in the directory's level of advertising. I also find that retailer willingness-to-pay for advertising in a directory increases in the amount that consumers use the directory. These two results imply that a network effect exists. In order to calculate equilibrium outcomes, I estimate a first-order condition derived from profit-maximizing behaviour by the publisher. I measure the importance of the network effect by looking at how much potential surplus is forgone due to the market's failure to fully account for the network effect. I find that the amount is large relative to the amount of deadweight loss 1. Advertisersmay also experiencea negativenetworkeffect becausemorecompetingadvertisersmakesreaching a consumermore difficult. Although the focus of this paperis on measuringthe positive networkeffect, the paper also attemptsto accountfor this congestion effect. 2. Asymmetric market shares may also be explained with increasing returnsto scale, but increasing returns typically imply the large firmalso sets a low price. 483
484
REVIEWOF ECONOMICSTUDIES
resultingfrom imperfectcompetitionand relativeto the amountof surplusrealizedin the market equilibrium. To furtherexplore the model, I consider the welfare trade-off between competition and monopoly undernetworkeffects. Positive networkeffects imply that there is a welfare gain to coordinatingeconomic activity on the same standard.However, there may be high congestion costs for advertisersand if the standardis proprietary,the owner can wield significantmarket power.3In thatcase, it is an empiricalquestionwhetherwelfareis maximizedundercompetition or standardization.Using estimates from the structuralmodel, I calculate equilibriumoutcomes for differentnumbersof competitorsand test if welfare increases.I find that networkeffects are not so strongas to counteractthe benefitsof entry.The resultsshow thatthe entryof independent publishersimproveswelfare, so rules thatforce telephonecompanypublishersto facilitateentry are welfare enhancing.4 The methodology developed here is also interestingbecause it could be applied to other markets.The methodology is relevantto any industrycharacterizedby indirectnetworkeffects and incompatiblenetworks.A particularlytopical example is the marketfor operatingsystems. Like Yellow Pages directories,computeroperatingsystems exhibit indirect network effects in the sense that higher consumer usage of a particularoperatingsystem leads to more available softwareand vice versa.Also, directoriesand operatingsystems are both incompatiblenetworks in the sense thatan advertisementplaced in one directoryconfersno benefiton the user of another directory,just as softwaredesigned for one operatingsystem confers no benefit on the user of a differentoperatingsystem. 2. RELATEDLITERATURE Network effects and positive feedback loops are the subject of increasing attentionin both the academic and popular press. The theoretical literatureon network effects begins with Katz and Shapiro (1985), who introduce the concept of indirect network effects under the name hardware/software paradigm.Hardwarebecomes morevaluablewhen more compatiblesoftware is supplied, and the amount of software available depends on the amount of hardwarethat consumers purchase.For consumers of Yellow Pages, the "hardware"is the directoryand the "software"is the advertising.Similarly,for advertisers,the "hardware"is the directoryand the "software"is the consumers.The firstformalmodels of indirectnetworkeffects appearin Chou and Shy (1990) and Church and Gandal (1992). Chou and Shy (1990), and Economides and Flyer (1997) study the welfare implications of entry and find that the effects of entry depend on parametervalues. Surveys of the literatureon networkeffects by Katz and Shapiro (1994), Economides (1996) and Shy (2001) provide an excellent overview. A numberof recenttheoreticalpapersstudythe generaltopic of two-sided markets.Rochet and Tirole (2003) provide a widely applicablemodel and discuss marketsfor advertising,credit cards, software and web portal usage. Anderson and Coate (2001) and Stegeman (2002) study 3. A common way for people who work in the Yellow Pages industryto convey the profitabilityof theirproduct is to compare it to the profitabilityof illicit narcotics. In interviews with the author,one person said "We earn more money than anyone this side of the Cali drugcartel."Anothersaid "Welike to say thatwe are the second most profitable industryin the world."Statisticalevidence of profitabilityis presentedin Section 3.1. 4. This exercise is relevantfor policy because the decision aboutwhetheror not to open the marketto competition is now in question.Althoughindependentpublishershave existed in the U.S. almost since the beginningof the industry, telephone company publishershave an advantagebecause they have better access to customerdata. However, a recent U.S. Supreme Court ruling established that White Pages are not copyrightable,effectively opening the market to competition (Feist Publications,Inc. v. Rural Telephone Service Company,Inc., 499 U.S. 340, 1991). This outcome seems to have been widely expected andbusiness practiceadjustedwell in advanceof the ruling.Also, the Telecom Act of 1996 includes a provisionthattelephonecompaniesmaketheirlistings availableat "reasonablerates".Similarpolicies have recentlybeen enactedin Europeand Australia.See http: //www. accc. gov. au/media/mr1997/telstra. htm.
RYSMAN
COMPETITIONBETWEENNETWORKS
485
broadcastmarketsin which retailers pay for advertisingto reach consumers, where typically consumersdislike advertising.Armstrong(2002) surveysthis literatureand providesa synthesis. Armstrong shows that it is straightforwardto write a model of the Yellow Pages market in which directoriesare distributedfor free and retailerspay for advertisingbecause each consumer generatesmore advertisingrevenuethan the cost of manufacturinga directory. The empiricalliteratureon networkeffects can be tracedback to Greenstein(1993), Gandal (1994) and Salonerand Shepard(1995). These papersevaluatedthe importanceof installedbase or networksize, typically in reduced-formequations.Very little work studies positive feedback loops. For instance, Ohashi(2000) and Park(2000) study the dynamicsof networkeffects in the VCR marketbut abstractfrom the video rentalmarket,presumablybecause of data constraints. Also, most work on networkeffects has focused on high-technologyproductswhere, as Goolsbee and Klenow (2002) make clear with the case of personal computers, it is often difficult to distinguishbetween networkeffects and the effect of consumerslearningfrom each other. A paper that does explicitly estimate a positive feedback loop is Gandal,Kende and Rob (2000). They study the entry decisions of producersof compact disc players and producers of compact discs. A major difference from my work is that they study firms producingfor a compatible standard.Because the standardis non-proprietary,there are no benefits to having a small numberof firms so the applicationof the model studiedhere does not arise. Two empirical models that are similar to the one here are by Rosse (1970) and Berry and Waldfogel (1999). Although Rosse's model is designed to identify the cost curve of a newspaperas opposed to measure network effects, his model does allow him to measure the feedbackbetween readershipand advertising.However,one would expect readers'valuationof newspaperadvertisementsto be ambiguous.5One reasonthat I choose to focus on Yellow Pages directoriesis thatthey are valuableto consumersexplicitly because of the advertisements.Berry andWaldfogelanalysethe effects of entryby radiostations.They also estimateconsumerdemand for radio play and retailerdemandfor advertising,althoughthey lack some station-specificdata (e.g. listenership,price, geographiccoverage) and so are constrainedfrom addressingsome of the issues exploredhere. 3. INDUSTRYAND DATACHARACTERISTICS As statedin the introduction,the Yellow Pages industryis an excellent one for studybothbecause of the structureof the market and because of the availability of excellent data. Section 3.1 discusses industry characteristicsand how they motivate importantassumptions. Section 3.2 reviews the data-set. 3.1. Industrycharacteristics The Yellow Pages generated $11.5 billion in sales in 1997 (Elliott, 1998). Yellow Pages directories published by telephone companies seem to earn exceptional profits. A directory serving 240,000 people averages over $6000 per page in revenue from display advertisements alone. The average size of such a book is 621 pages so the book brings in around $3-8 million in revenuefrom display advertisements.Industrysources estimate that variablecosts of productionfor a book of this size would be less than $1 million. Indeed, Yellow Pages industry sources estimate that profitsrepresent35-45% of revenue. In the break-upof AT&T,the court assigned Yellow Pages publishersto local phone companies in order to hold down local rates. 5. In Rosse's paper,the parameterthatcapturesthe effect of advertisingon readershipis estimatedto be positive but not significant.
REVIEWOF ECONOMICSTUDIES
486
Numberof Publishers in a Sub-Market
Numberof Directories in a Sub-Market 35- 31 31.9 30- 1 25-
50- 41 411
1 20-
I
2151-
30117
420-
5lo
3-2 0.8 0.2 0-2
05I)
I
1
2
l
3
5 4 Number
I
--I
m
6
/
66
Io-
~
-
n
T
7
17
!
1
8
Numberof Directoriesper Person
5
50-
47
40
40-
f
3020-
10.2 2
3 4Nume Number
Numberof Publishersper Person
50>-39-1 38.1 40064 o30& 16 201o-
2
0-2
3
4 5 Number
0.2
0.1
0.1
6
7
8
0
1 1
2
3 Number
4
5
FIGURE 1
Directoryand publishercoverage
Public service commissions estimate that if local phone companies did not receive Yellow Pages profits, local rates would have to increase between $1-80 and $3.50 per line per month (NARUC, 1994-1995). While the Yellow Pages industryhas traditionallybeen dominatedby telephone companies that never enter each other's markets,there is some competitionfrom independentpublishers. Independentpublishersprint 38% of the directoriesin this data-set.Figure 1 shows that about 55% of the population in this data-set receives two or three directories. This number is not just due to publishers that distributeoverlappingdirectories:Figure 1 shows that more than 50% receive directories from more than one publisher. That is, the median person receives two directories from separate publishers. However, independent publishers have not been overly successful. Directories associated with telephone companies average 6.42 references per household per month, while the same number at independent directories is only 1-32. Directories associated with telephone companies are on average almost twice the size (746 pages compared with 413) and charge about twice as much ($2014 to $1221 for a doublequartercolumn advertisement)as independentpublishers. A better measure of the difference comes from comparingdirectorieswith distributionareas that perfectly overlap. Bell Atlantic and R. H. Donnelly both have directorieswith distributionareas that are exactly equal to the boundariesof WashingtonD.C. The phone company's (Bell Atlantic's) directoryis 1443 pages long, charges $3387 for a double-quartercolumn advertisementand collects 7-6 referencesper household per month. In contrast,the Donnelly directoryis 947 pages, charges $2352 for the same size advertisementand collects 1-4 references per household per month. Despite these numbers,some independentpublishersare successful. About 25%of the directoriespublishedby telephonecompaniesface a competitorthattakes 25% of the usage marketor more. In summary, telephone directoriescompany directoriesare very successful but face non-trivialcompetition from independents. Two assumptionsabout the behaviour of publishers follow from a priori observationof the industry.First, although Yellow Pages publishing is dominatedby regulatedutilities, this
RYSMAN
BETWEENNETWORKS COMPETITI'ON
487
papermodels publishersas profitmaximizers.PublishersvigorouslyprotectYellow Pages profits, often transferringthem to for-profitsubsidiariesor otherwisehiding profits(White and Sheehan, 1992). The second assumptionis thatthe numberof books distributedis exogenous to the priceand quantity-settingprocess. Telephonecompanies and independentsdistributeone directoryto every phone line in an area, so the numberof directoriesis determinedby the geographicscope of a directory.Because scoping is determinedbefore the sale of advertising,I take population coverage as exogenous.6 3.2. Data The datacomes froma numberof independentsources.NationalYellow Pages Monitor(NYPM), a proprietarydata company, collects usage data for individual directories.At the request of a client (normallya large Yellow Pages publisher),NYPM measuresthe numberof referencesper householdper month going to all directoriesin a given metropolitanstatisticalarea,taking each directoryareaas a unit of observation.7The NYPM set containsdataon 476 directories. The number of pages in a directoryproxies for the quantity of advertising.The Yellow Pages PublishersAssociation (YPPA),an industrytradegroup,maintainsa libraryof directories publishedby membersand the Boston ConsultingGroupcollected the numberof pages in each directoryfor a separateproject.The dataon page numbersincludesa check for directoriesthatare observablysmall. I convertpages into a variablecalled "advertising"by multiplyingpages by the numberof columns and multiplyingby an adjustment(0-8) for directoriesthathave small height and widths. YPPA membershipcovers over 6500 directoriesincluding almost every directoryin the U.S. The YPPA claims that its membershipaccounts for 95% of Yellow Pages advertising sales. Advertisingprices come fromthe Rate and Data data-setand directorystatisticscome from the IndustrialCharacteristicsdata-set,both from the YPPA. The pricing data is an especially rich source. The Rate and Data set containsprices for every size and colour of advertisementat every directoryin the YPPA. Directories in the data-setoffer an average of around80 choices of advertisementsize and colour. Unfortunately,there is only one observationon quantityto match with each set of prices. In estimation,I choose one price to representeach directory.The IndustrialCharacteristicsdata-setcontains directoryinformationsuch as the numberof people a directorycovers, and the numberof columns in a directory. Directory boundary data come from Claritas, Inc., a proprietarydata company. The boundarydatacome in the form of computermapsthatI matchedwith populationcentresof fivedigit zip codes. I assume that if a populationcentre of a zip code falls within the boundaryof a directory'sdistributionarea,then the publisherdistributesthe directoryto the entirezip code. Zip codes are a reasonablyclose approximationof directoryboundaries,and in some cases coincide exactly. Combiningpopulationdata at the zip code level with the mappingdata determinesthe choice set of directoriesfor consumers. All data are from 1996 except for the boundarydata which are from 1997. This difference creates some discrepanciesand, after matching all four data-setstogether,the data-setcontains428 observations.8 6. Telephonecompaniesarerequiredto distributeone White Pages directoryto every phone line andmost choose to publishthe Yellow Pages in the same book as the White Pages. Independentpublishersmatchtelephonecompaniesin both ways. 7. NYPM surveyrespondentsmaintaindiariesof theirYellow Pages usage for 1 week. NYPM normallysurveys 1000-3000 people per MSA, although NYPM uses 11,200 respondentsin Los Angeles, normally leading to a few hundredrespondentsfor even relativelysmall directories. 8. I removed directoriesthat were targeted for non-English speaking audiences, such as Spanish or Chinese languagedirectories.
488
REVIEWOF ECONOMICSTUDIES
I assign directories to their central or most populous counties and match them to demographic data from the USA Counties CD-ROM 1996. Industry sources suggest that educated,relativelywealthy people who own theirown home are likely to use the Yellow Pages, as are people who have recently moved. People who live in urban settings or regularly use public transportationuse Yellow Pages less. Table A.1 in the Appendix presents a description of demographicdatathatI use to capturethese features. Table 1 presentssimple statisticsfor all of the variables.Note that thereis a strongpositive correlationbetween usage and the quantityof advertising,suggesting a network effect.9 Also, there is a strong negative correlationbetween measures of competition and price, advertising and usage. Table 1 providestwo measuresof competition:"adsat competingbooks" is the total amountof advertisingin competingbooks, averagedover householdsin a directory'sdistribution area; and "usage at competing books" is the averagenumberof references per household in a directory'sdistributionarea that go to a competingbook. Competingbooks are those published by other publishers.These negative correlationsare robust to controlling for whether or not a directoryis associatedwith a telephonecompany,and to controllingfor demographics. 4. MODEL This section presentsa model of competitionin the Yellow Pages industrysuitablefor estimation. The model explicitly captures the interaction between directory advertising and directory usage. The model is a one period, simultaneousmove, quantity-settinggame.10 There are J publishersthat each produce one directory.The distributionareas are taken to be exogenous. The distributionareas may or may not overlap, and may do so for only a portion of their areas. The publishersface two interactingmarkets:a marketfor advertisingand a marketfor consumerusage. Consumersreceive directoriesfor free so there are no profitsdirectly from the usage market.But the amount of consumer usage affects demand in the advertisingmarket.11 Specifically,each publisherj, j = 1,..., J, faces two demandcurves:retailerinverse demand for advertising Pj(A1, U1,..., Aj, Uj) and consumer demand for usage Uj(A1, ..., Aj), where Aj is the amountof advertisingat j and Uj is the numberof uses per consumercovered. I expect that Pj /a Uj > 0 and a Uj / Aj > 0. These two conditionstogetherrepresentthe networkeffect. If usage increasesfor some exogenous reason,then the inversedemandcurve for advertisingshifts out because aPj/ Uj > 0. In equilibrium,the shift leads to an increasein the quantityof advertisingthat then implies a furtherincreasein usage because aUj/lAj > 0. The positive feedback loop between usage and advertisingmeans that an otherwise small change in eitherdemandcurve can lead to a large change in both usage and advertising. I furtherexpect that aPj/aAj < 0, representingthe "scarcity"effect, the standardeffect thatprice decreasesin quantity.I also expect that aUj/0Ak < 0, k $ j. I discuss aPj/aAk and a Pj /a Ukin Section 4.1. Each publishersimultaneouslychooses its quantityof advertisingAj to sell in its directory.12The solution concept is Nash equilibrium. 9. There is already some evidence that consumers value Yellow Pages for the advertising.Laband(1986) and Mixon (1995) show that Yellow Pages advertisementsare likely to be more informativeif they are for productsthat are "searchgoods" (i.e. goods that are expensive and are purchasedinfrequently).Equally as convincing is a recent Ameritech radio advertisementboasting that the Ameritech Yellow Pages have "the most ads and the most complete information". 10. Estimatinga price-settinggame instead of a quantity-settinggame presentsa numberof difficulties,although I still find empiricallythata positive feedbackloop exists in thatcase. Section 4.3 discusses these issues. 11. Presumably,if publishersdid charge for directories,the same issues would apply. That is, a low price for directorieswould increasethe demandfor advertising,which would then increaseconsumerdemandfor directories. 12. In the data, some publishers own overlappingdirectories.I address this feature in the constructionof the publishers'first-orderconditions.
RYSMAN
489
COMPETITI'ON BETWEENNETWORKS TABLE 1 Simple statistics Usage
Mean S.D. Correlations Usage Rate (DQC) Pages Advertising Pop. cov Telco Ads at competingbooks Usage of competingbooks % Urbanpopulation % Have not moved % Lived in diff county % Lived in diff state % Takepublic trans. % Own house % Gradhi school % Gradcollege Establishmentsper cap. Per cap. income House-buildingpermits Countypop. growthrate Pop. density Earningsper worker GTE Bell South Distributionarea Rate (full page) Rate (Bold)
4.85 4.17
Price (DQC) 1787 1155
Pages 630 479
Advertising*
Population coverage
2621 2190
386 411
1.00
0.37 1.00
0.46 0.73 1-00
0-52 0-68 0.98 1.00
0.08 0-76 0-71 0-66 1.00
-0.64 -0.61 -0.14 0.16 0.19 -0.09 -0.01 0-20 -0-04 -0.11 0.04 0.02 -0.01 -0.04 -0.08 -0.09 -0.05 0-22 0.09 0.57 0.44
-0-28 -0.38 0.14 0-22 -0.11 0-08 0.30 0-01 0.08 0.17 0.25 0-22 0.05 -0.13 0.23 0.27 -0.09 0.17 0.01 0.91 0.90
-0-27 -0.36 0.16 -0.03 -0.05 0.23 0.21 -0.06 0-12 0-17 0-18 0-16 0.02 -0.08 0.15 0.19 -0.06 0.07 0.06 0.80 0.66
-0.31 -0.41 0.17 -0.04 -0.02 0.21 0.20 -0-08 0.10 0.15 0.16 0.16 -0.02 -0.10 0.15 0.17 -0.01 0.05 0.06 0.81 0.63
-0-06 -0.12 0.24 0.04 -0.20 0-17 0-27 -0-18 0-01 0.11 0.12 0-08 -0.07 -0.16 0.27 0.21 0.09 -0-01 0.08 0.71 0-64
Telco dummy 0.64 0.48 0.62 0.33 0.33 0.41 006 1-00 -0.35 -0-82 -0.04 0-11 0-09 -0-07 0-00 0-15 0.04 0-01 0.07 0-06 0-10 0.01 -0.03 0-00 0-02 0-18 0-01 0.52 0.38
*Advertisingequals pages multipliedby columns and a size adjustmentfor small directories.
In summary,thereare two effects from an increasein advertising.An increasein advertising leads to a price decreasebecause it is a movementalong the willingness-to-paycurve. However, usage increases which shifts the willingness-to-paycurve out. The locus of points that account for both changes is the demand curve that the publisher faces. Willingness-to-paycurves are demandcurves with usage held constant-the slope of a willingness-to-paycurve is measured eU 13 apj aPj = aAj + aU' ajjuj 13 by aaj . The slope of the demandcurve thatthe publisherfaces is dAj Although most of the theoreticalresults that follow hold for general demand functions, Sections 4.1 and 4.2 presentfunctionalforms for usage and advertisingdesigned for purposesof estimation. Sections 4.3 and 4.4 discuss equilibriumand efficiency, and Section 4.4 presents a measureof forgone surplusdue to a failureto take advantageof networkeffects. 4.1. Demandfor advertising The functionalform for Pj (-) is drivenby surprisingresults from estimation.The function Pj(.) clearly should dependon advertisingandusage at book j, as well as demographicvariables.One 13. Economides and Himmelberg(1995) make a similar distinction and point out that in many cases, network effects imply an upwardsloping demandcurve for low levels of quantity.Understrong networkeffects, no one will pay for a productthat no one else uses.
490
REVIEWOF ECONOMICSTUDIES
might think that it should also depend on advertisingand usage at competing directories.But I estimate these coefficients to be close to zero and insignificant.Section 4.1.1 presents a model of individualadvertiserdemandthat generatesthis result and aggregatesto the functionalform for P(.), therebyallowing for a structuralinterpretationof the parametersin P(.). Section 4.1.2 discusses empiricalidentification. 4.1.1. A model of the advertising market. A representativeadvertiser chooses the amountof advertisingaj (a continuouschoice variable)to place in directoryj. The advertiser acts as a price taker and chooses its optimal level of advertising,aj(P1,..., Pj). The market size of advertisingis m, so total advertisingin book j is maj = Aj. By invertingthe aggregate demandcurve, we can constructan inverse demandcurve thatthe publisherfaces. In orderto derivethe factordemandfor advertisingaj (P1,...., Pj), I use the following setEach consumerneeds informationsome exogenous numberM times per month. Each time up: need information,they can use a Yellow Pages directoryin their areaor some otheroption, they such as the Internetor word-of-mouth.14I expect that a consumeris more likely to use a Yellow Pages directorywhen it is more informative,or equivalently,when it has more advertising.The first assumptionon consumerbehaviouris: Al Consumersuse at most one directoryper informationrequirement. Rochet and Tirole (2003) referto this featureas "single-homing".15This assumptionwould naturallyfollow from a scenarioin which therewas some cost to opening a second directoryor storingit in an accessible place, especially if consumers found directoriesto be close substitutes.NYPM does not keep track of the simultaneoususe of multiple directoriesexplicitly, but this conforms with casual observationas well as some relatedstatistics.Forinstance,when consumersreferencea directory, they contact an advertiser82-1%of the time, but make more thanone contactonly 36.2% of the time. These datafit with consumerswho use only one directory,althougharenot definitive.Note thatunderthis assumption,consumersmay use differentbooks for differentneeds. For instance, they may use a local book when they need a barberbut a regional book when they need a car dealer.AssumptionAl says that consumersdo not open up multiple directoriessimultaneously. Andersonand Coate (2001) make a similarassumption.As in theirwork,this assumptionimplies that directoriesare monopolists over access to theirreaders. Once a consumerhas opened a directory,the consumerlooks at a numberof advertisements and makescontactwith a portionof the advertisers.Advertisersareinterestedin how manylooks are generatedfrom an advertisement.Let the numberof times the averagepersonin directoryj's distributionarea looks at the representativeadvertisementin book j be Lj = L(aj, Uj, Aj), where aLj/aaj > 0, aLj/lUj > 0 and aLj/aAj < 0. The last inequality capturesthe fact that a given advertisementis less likely to be seen in a large book. This "congestioneffect" is a negative networkeffect. Holding usage constant,advertiserswould preferto be in a directory with fewer other advertisers.Note that the L function is identical for each directory,although directoriesdiffer in usage and advertising. A proportion of the consumers who look at an advertisementmake contact with the advertiser,and these consumercontactsgeneratesome level of profit.The second assumptionis: 14. The parameterM capturessomething fairly abstract.For instance, every time a person needs a haircut,they have an informationrequirementin the sense that they need to pick a hair salon. In orderto choose, they could look in the Yellow Pages or they could use an "outsideoption" such as going to the place they went to the last time, getting a recommendationfrom a friendor going into a salon they have seen in theirneighbourhood. 15. Conversely,advertisersmay place advertisementsin multipledirectoriesand so are said to "multi-home".
RYSMAN
COMPETITIONBETWEENNETWORKS
491
A2 Advertiserprofit per look is constant. These two assumptions imply that profit is separablein aj. Intuitively,the first assumptionimplies that advertisingin one book is not a substituteor a complementfor advertisingin anotherbook. As a result, there is no demand-side reason why the advertisers'choice of advertisingat one book should affect the choice at another book. The second assumptionsays thathaving many customersas a result of one advertisement does not affect the cost or benefit of serving customers generatedby another advertisement. So there is no cost-side reason why the choice of advertisingat one book affects the choice at anotherbook. Separabilityof the profitfunctionfollows. The advertiser'sprofitfunction fI can be writtenas n = t1L(al, U1, A1) - Pial + . + 7jL(aj, Uj, Aj) - Pjaj. The term fj capturesthe profitto the advertiserfrom the numberof looks per person received from book j's distributionarea.The term ftj capturesvariablessuch as the numberof people in the distributionareaand their demographics. Separabilityimplies the desired result that outcomes at each book do not affect each other directly.The rest of this subsectiondevelops functionalforms for estimation.Let L(aj, Uj, Aj) have the Cobb-Douglas form, so Lj = a'1 A2 Ul1. The parameterY1is expectedto lie between0 and 1, and capturesdecreasingreturnsto large advertisements.I expect thatthe parameterY2will be negative capturingthe business stealing effect, i.e. the fact that an advertisementmight get lost in a large directory.The parameteral should be positive because more usage of a directory increasesthe likelihood of consumerslooking at a given advertisement.16 Now the profitfunctioncan be writtenas EI= ila1AY'AU11 - Pial + ... + TjaY1A2 U1 - Pjaj. Y1 AY2Ut The advertiserpicks aj to maximize rjajYA Uj - Pjaj. The representativeadvertiseris too small to affect Aj and takes it as given. The optimalaj for the advertiseris -
Yj aj =
J >'
)
i
AU
Aggregating,we have Aj--= Aj
V
i
)
11
where 7j = 7-rj/Yl- 1 is an aggregate equivalent of :tj that accounts for the mass of advertisers
in the market.Solving for Pj, we obtain the inverse demandcurve:
Pj(Aj, Uj) = ylAyl+j2 (1) YI+Y2--1U;17j. UI lj. There are several importantfeatures of this demandcurve. First, it should increase in usage (I expect aq > 0). This feature,along with the fact that usage increases in advertising,represents the network effect. Second, the demand curve decreases in advertisingboth because there are decreasing returnsto individual advertisersfrom large advertisements(I expect yi < 1) and 16. Note that it is straightforwardto obtain these results with advertiserheterogeneity.Let there be a continuum of advertisers indexed by 1 E [0, m] distributed f(l). Denote the choice of advertiser I at book j as ajl, so Aj (P1,..., Pj) = f(o ajl(P, ..., Pj)f(l)dl. Instead of fri, let each L() function be augmented by *jl and let This change of structureleads to equation (1). The interpretationof Tjl rj = (1/fom(1/ljl)l/(Yl-l)f(l)dl)Yl-l. is of factors idiosyncraticbetween the directory and the advertiser,such as the location of the advertiserwithin the directory'sdistributionarea,and 7rj is an aggregateof these effects.
492
REVIEWOF ECONOMICSTUDIES
because advertisementswill be lost in a largebook (I expect Y2< 0), but these effects will not be distinguishablein estimation.In estimation,I referonly to estimatingy, where y = yi + Y2- 1.17 Third,the price and quantityof advertisingand the amountof usage in anotherbook do not affect book j directly.Holding usage at book j constant,advertisersare willing to pay the same amountto advertisein book j regardlessof what book k does. This result follows from Assumptions Al and A2. If consumerschose books at randomso the consumerside of the networkeffect is eliminated,then publisherswould have significantmarketpower as there is no other way to reach these consumers. But with the network effect, book k competes with book j to attract usage. The inverse demand curve Pj(A1,...,
A j, U (A, ..., Aj), ...,
Uj(A1
....
Aj)) can
be rewrittenas Pj(Aj, Uj(A1,..., Aj)). This featureis consistent with estimationresults and, as it is difficult to constructa model of how advertiserstradeoff between directories,makes a structuralinterpretationof parametersmuch easier. 4.1.2. Identification. A formal discussion of the estimation is delayed until Section 5. Here I discuss intuitivelythe identificationof the parametersin equation(1). In practice,I specify ln(Pj) = y ln(Aj) + al ln(Uj) + XP/P + vj.
(2)
That is, ln(7jr) is capturedby a linear function of observable variables and an unobservable term vj. As the equilibriumquantity of advertisingdepends on price, we expect ln(Aj) and ln(Uj) to be correlatedwith vj. For instance, if willingness-to-paywas high for unobservable reasons,we might also expect the quantityof advertisingto be high via the publisher'sfirst-order condition,and thereforeusage to be high as well. I address this problem with instrumentalvariables. As an instrumentfor usage, I use variablesthat capturethe numberof people who recently moved. People who recently moved tend to use Yellow Pages much more than long-time residents. This instrumentworks well as long as recent movers do not tend to be more valuablecustomers-that is, recent movers do not affectthe demandfor advertisingover andabovetheireffect on usage.18In practiceto identifya1, I use the percentageof people in the county who lived in a differentstate 5 years previous, the percentagewho lived in a differentcounty 5 years previousand the percentagewho lived in the same house 5 years previous. In orderto identify the effect of the quantityof advertising,I use variablesthat would be expected to move marginalcost and thereforeaffect the publisher'sfirst-ordercondition.Almost all publisherscontractpublishingto a single firm,R. R. Donnelly. However,Bell South and GTE maintainedtheirown printingfacilities. I use dummyvariablesfor being one of these companies as instrumentsto identify y. Wagesat the level of the publisherwould makeexcellent instruments but are unobservable.Instead I use the census measure of earnings level in a county, which approximatesa county-level averagehourly wage. Local wages are useful because the biggest cost to producinga directoryis sales, and salespeoplemust be based nearto the county wherethe directorydistributes.Local wages are problematicto the extent that they also affect advertising demand-I control for this issue by including county level income in X". Section 5 discusses the effectiveness of these instruments. 17. Note that Y2representsa negative networkeffect. Addressingthis networkeffect with the same rigouras the positive networkeffect would requireestimatingthe slope of the willingness-to-paycurve as Y1and allowing the curve to shift in as advertisingincreases (by some amountrelatedto Y2)in the same way that it shifts out due to the effect of usage. This issue would be interestingto addressbut requiresmore detaileddataon the choices of advertisers. 18. For instance,if recent movers are more likely to be forminglong-termrelationshipswith retailers,then recent movers might be valuableover and above their effect on usage. But in practice,most users of Yellow Pages are forming long-term relationships.Consumerswho already have retailerstypically have the telephone numberor use the White Pages.
RYSMAN
COMPETITIONBETWEENNETWORKS
493
4.2. Consumerusage In the model of the advertiserdemand, I assume that consumers choose one directory or an outside option each time they need information.Thatnaturallysuggests a discretechoice model for the consumerchoice of a directory.Once a consumerpicks a directory,they may look at any number of advertisements-their behaviour generates the look function L() discussed above. But instead of formally modelling how consumers choose among advertisers,I simply allow the utility from choosing a directoryto depend on total advertisingin a lpg-linearfashion. This simplificationis likely to be reasonablyaccurateand is difficultto improve upon without more disaggregatedata. Section 4.2.1 presentsthe discrete choice model and Section 4.2.2 discusses identification. 4.2.1. Model. I follow methods presented in Berry (1994) for applying a nested logit model to marketswith aggregatedata. As above, the total numberof times that a representative householdrequiresinformationof the kind thatcan be found in the Yellow Pages is an exogenous numberM. In orderto allow Yellow Pages directoriesto be closer substituteswith each other than with the outside option, I place Yellow Pages directoriesin one nest and the outside option in a separatenest. Following Berry (1994), let the utility to consumeri from directoryj be Uij = 02 ln(Aj)
+ XuBU
+ j
+ i(a)
(
1 - a-)Ej.
I expect to find that a2 is positive, which along with al capturesthe networkeffect. The vector xy contains demographicvariables associated with the central county of each directory.The variable ~j is a directory-specificvariable that captures characteristicsthat are unobservable to the econometrician.In this case, the characteristicscould be unobservablequality of the directoryor region-specificusage effects. The unobservablevariable i (a) capturesindividual i's preferencefor Yellow Pages and Eij capturesindividualpreferencefor a specific directory. The variableEij capturesissues such as the location of the consumerrelative to the location of the directory.I assume that Eij is distributedType I ExtremeValue and (i (a) is distributedsuch that (1 - a)Eij + (i(a) is also distributedType I Extreme Value. Berry (1994) discusses this issue and Cardell (1997) shows that 5i (a) exists and is unique. The parametera is restricted to lie between 0 and 1, and measures the correlationin unobserved(to the researcher)utility from differentYellow Pages directories.As a approaches0, correlationwithin the group goes to zero, the directoriesbecome more differentiated,and the model approachesa standardlogit model. The parametera will be estimated.The utility from the outside good is normalizedto be Uij = Sio(a)
+ (1 - a)Eio.
Let sj be the marketshareto directoryj in its distributionregion, so Uj = Msj. Let sj Iyp be the share of people who choose directory j given that they choose to use a Yellow Pages directoryand let so be the shareto the outside option in j's region (the j is suppressedas it will be obvious in context). It is useful to define the mean utility for directoryj to be 8j, so =j= O2 ln(Aj)+
XYU
+ ~j.
(3)
Berry (1994) shows thatunderthe nested logit assumptions,we have sj = esjsosj yp. This relationallows for the identificationof a2, Pu and a in a log-linear model. If this relation held true at every observation,I could estimatea2, lu and a by ln(sj)
- ln(so) = a2 ln(Aj) + XJ U + a ln(sj Iyp) + j.
(4)
Note thatwe observesj Iyp in the dataandcomputesj andso by makingan assumptionaboutM, the total numberof informationrequirements.Assuming thatpeople need informationthe same
494
REVIEWOF ECONOMICSTUDIES
amountof times seems an appropriatefirst approximation.An alternativewould be to allow M to dependon demographics,althoughthat would be asking a lot of aggregatedata.Instead,I try differentvalues of M and show thatthe value does not significantlyaffect results. A comment is in order on the assumptionthat Eij is i.i.d. We expect the preferences of households to be correlatedacross time and the average number of choices per households per month ranges as high as 25, so the i.i.d. assumption could be problematic. However, less restrictiveassumptionswould suffice. For instance, allowing for correlationin Eij within households but assuming Eij is uncorrelatedwith the number of times a household needs informationis sufficient. To see this, take the extreme case where each household chooses M times and draws the same Eij each time. Households choose the same directoryeach time but regardlessof whetherM is 1, 5 or 30, marketsharesremainthe same and so equation(4) holds. As long as households that prefer a particularbook do not also need informationmore or less often, we may proceedusing equation(4). Thereis a separatereasonwhy equation(4) is problematic.Equation(4) can only be applied to areas where all consumershave the same choice set whereas the marketfor Yellow Pages is characterizedby overlappingmarketswith distinct boundaries.19In principle, one could apply equation(4) to each areawith a uniformset of directories,but the data-setcontainsusage shares for directoriesonly in their whole area. One cannot tell how much usage of a directorycomes from the areawhere the directoryis a monopolistas opposedto fromthe areawhere the directory faces competition.Fortunately,it is possible to infer what usage must be in each sub-marketby combining data on total usage share, data on the extent of overlap and the assumptionsof the nested logit model.20 Briefly,usage sharesfor each sub-marketcan be constructedbasedon the nested logit model from a vector 8 and the parametera. I develop an algorithmthat finds the vector 8 that implies sub-marketusage sharesthataddup to observedmarketshares.While thereis no explicit function for 8, it is straightforward to nest a fixed-pointalgorithminto anoptimizationroutineas suggested by Rust (1987) and Berry, Levinsohn and Pakes (1995). With 8 in hand, we can estimate the desired parametersfrom equation (3). In order to increase identificationpower over a, I use equation(4) at the 56 observationswhere it is appropriate,i.e. the directoriesthat contain only one sub-marketbecause they are completely overlappedby all competingdirectories.Details of this routineappearin the Appendix. 4.2.2. Identification. Equation(4) introducestwo identificationproblems.21The first is the same as discussed before, usage shareand advertisingare determinedsimultaneously.As an instrument,I use the numberof people covered by a directory.Populationcoverage should not affect a household'sdecision aboutwhich directoryto choose but shouldhave a majorimpacton the demandfor advertising.That is, populationcoverage appearsin XP but not XU and so is an appropriateinstrument. Anotherissue is to identify a. Conceptually,the shareof consumersswitchingto the Yellow Pages group as the features of the group change identifies a. Such variationcan be a result of either changes in directorycharacteristicsor changes in the number of directories. However, 19. The term "overlappingmarkets with distinct boundaries"distinguishes the problem from more standard problemsof marketareas.For instance,two restaurantsthatare distantfrom each othermay have common customersbut they do not have distinctboundaries.All consumerscould choose to go to eitherrestaurant.The problemin this paperis one in which differentgroupsof consumershave differentchoice sets. Consumerscannotuse a directoryunless they live in the directory'sdistributionarea. 20. Sub-marketsare defined as areas served by a uniformset of directories.In this data,428 directoriesgenerate 660 sub-markets,with some sub-marketsbeing served by as many as eight directories. 21. The result of the fixed-point algorithm described at the end of Section 4.2.1 has the same identification problems.
RYSMAN
COMPETITIONBETWEENNETWORKS
495
there is a potentialendogeneityproblemwith this second type of identificationif directoriesare attractedto marketswhere usage is high for reasons that are unobservableto the researcher.In equation(4), if ~j is high, sj will be high and there will be more entrantsso sj yp will be low. The estimate of a will be biased downwardsunless sj Iyp is properlyinstrumentedfor. For this purpose,I use the squaremileage of the distributionarea of a directoryas an instrumentfor the usage equation.An empiricalfact is thatlargerdirectorieshave less of theirregion overlappedby otherdirectoriesand so shouldhave a higherwithin-groupshare.The simple correlationbetween the numberof directoriesin an area (weighted by population)and the size of the areais -0-09. Also, if publisherschoose the scope of a directoryto maximize usage, then the first-ordereffect of a directory'sdistributionarea on its average usage should be zero. Otherissues most likely affect the publishers'decision, but as long as those effects arerelativelysmall, it is reasonableto exclude the distributionareafrom the matrixXU. 4.3. Equilibrium The model has J sets of three endogenous variables, Aj, Pj and Uj, and two demand curves for each set: the inverse demand for advertising Pj (Aj, Uj) and the demand for usage Uj (A1, ..., Aj). The model is closed by the publishers'first-orderconditions.I assumemarginal cost is constant at MCj, so MCjAj is the variable cost to j of producing a directory with advertisinglevel Aj. Let K (j) index the set of directoriesowned by the publisherof directoryj. For each of its directories,the publisherof directoryj solves: maxAj,
kEK(j)
Pk(Ak, Uk(A1 ...,
Aj))Ak-
MCjAj.
The first-orderconditionis ln(MRj) = XjCfC + oj
where
MRj = Pj + AjJ
P aAj
P + A j j Ujj + aA aUj aAj
A APk PkaUk a keK(j) Ajak OAj
(5)
The firsttwo elements of MRj (marginalrevenue)representthe standardmarginalrevenueterm. The thirdterm capturesthe networkeffect and the fourthterm capturesthe effect of advertising at directoryj on revenue from other directoriesowned by the same publisher.The Appendix establishesthat an equilibriumin pure strategiesexists when networkeffects are not "too large" and one always exists when the directorydistributionareas perfectly overlap. An equilibrium in pure strategiesexists for the parameterestimates found in this paper.This equationdoes not introduceany new endogeneityproblems. Note that although identifying the first-ordercondition allows me to perform policy computations,my primaryinterest in this paper is to estimate the demand curves in order to show the positive feedbackloop. For this reason,I only apply enough instrumentsto just identify the first-ordercondition so it does not impactthe estimatesof the demandparameters,only their standarderrors.22Indeed,when I estimatethe demandfunctionswithoutthe first-ordercondition, I get almost identicalparametervalues. Before moving on, I briefly discuss the use of a quantity-settingmodel instead of a price-setting model. Estimating the first-ordercondition in the price-setting game introduces serious difficulties because usage depends directly on quantity. The primary problem is a computationalone. A price-settinggame would requirespecifying a demandcurve of the form A = A(P, U(A)). The quantityof advertisingshows up on both sides of the equation.Therefore, 22. In this sense, the first-ordercondition is just-identified.This approachdiffers from Berry et al. (1995) where partof the identificationof demandparameterscomes from over-identifyingthe first-ordercondition.
496
REVIEWOF ECONOMICSTUDIES
taking the derivativeof the demand curve with respect to price (in order to compute marginal revenue)would requiresolving a fixed-pointequation,furthercomplicatingan alreadyinvolved optimizationroutine.23 I circumventthis problem by estimating a quantity-settinggame. Note that the difficulty with the price-settinggame is only in the estimationof the first-ordercondition.As I am primarily interestedin the demandcurves, I also triedestimatingthe advertiserdemandequationwith price on the R.H.S. and quantityon the L.H.S., which would be a firststep towardsestimatinga pricesetting game. When I do so, I get very similarresultsto those reportedhere. 4.4. Efficiency Consumersobtaindirectoriesfor free andconsumersdo not have the optionto pay for a directory with extra advertisements.The lack of prices for the consumer side of the market makes it difficult to convert consumer utility into units that are easily comparableto retailer surplus. I analyse surplus formally only from the point of view of advertisersbut I discuss consumer surpluswhen I am able to do so.24 A social planner (who accounted for advertiserand publisher surplus but not consumer surplus) would choose the set of advertising levels Aj, j = 1,..., J, simultaneously to maximize Ej= f j Pj(s, Uj(A1,..., Aj))ds-C(Aj). The first-orderconditionsfor the social plannerare ?
s^p
fAk
+ k=
aPk(s, Pk(s, Uk(Al ,.,AJ)),
Uk( a Uk
aUk d ds MCj =
aAj
= j
1,..
J.
(6)
It will be rarefor equations(5) and (6) to hold simultaneously.25It cannotbe said for sure which regime will result in more advertising.The social planneraccountsfor the value of the network effect to the entire set of advertisersand the advertisersin otherdirectories.The publishertakes into account the value of the networkeffect on its marginalpurchaser.However, the publisher also takes into account the effect of downward sloping demand on marginalrevenue and in most reasonablecases, we have the standardresult that the size of the networkis too small. By "reasonable",I mean thatthe networkeffect of advertisingon a publisher'sown price is stronger aPk aaUk>OYA thanthe negativeeffect on its competitors'prices, so Yk=l aU-k > 0 VAj, j = 1, ..., J. In this case, the social plannerpicks Aj such that price is less than marginalcost. "Reasonable" also means that demand is downward sloping, even accounting for the network effect, so a ij + aup aAIj < 0 at the Nash equilibrium,which implies that price is greaterthan marginal cost in the Nash equilibrium.These conditions are testable and I returnto them in Section 6. In equilibrium,some of the forgone surplusis a resultof publishersexercisingmarketpower while some is due to the market'sfailure to account for the networkeffect between advertisers 23. A secondary problem involves multiple equilibria. Conditional on directory choices, a price-setting game admits multiple equilibriawhereas a quantity-settinggame does not. Considertwo symmetricdirectories.If they both choose the same price, it is reasonableto expect multiple equilibriawhere one directoryhas high advertisingand high usage while the other has low advertisingand low usage. In this case, determiningmarginalrevenue from a change in price requiresan arbitraryassumptionaboutwhich equilibriumis selected. Conversely,if two symmetricdirectoriespick the same quantity,then they have the same usage and thereforethe same price. The quantity-settinggame can still have multiple equilibriabut not conditional on quantities.Therefore,the first-ordercondition in the quantity-settinggame, which takes competitors'choices as given, is straightforward.In fact, for the parametersestimatedin this paper,thereis only one equilibrium. 24. Berry and Waldfogel(1999) take a similarapproach. 25. I have assumed that the publisherscan set only one price. If a publishercan first-degreeprice discriminate, it could achieve the socially optimal level of advertising.This result is the case in Liebowitz and Margolis (1994) who assume that a monopolist faces identical purchasers,which implies that the monopolist can charge consumers their valuationwith a single price.
RYSMAN
COMPE'l'l'IONBETWEENNETWORKS
497
and consumers. This paper uses the following definitions to differentiatebetween these two effects. Let classical deadweightloss be the differencebetween the amountof surplusgenerated at the Nash equilibriumand the amounta social plannerwould generateif it did not accountfor the networkeffect. Let networkdeadweightloss be the differencebetween the amountof surplus generatedwhen the social planneraccountsfor the networkeffect and when it does not. The following discussion defines these terms formally and Figure 2 demonstrates.Fix the competitors' outcomes at their Nash equilibriumchoices A_j and denote the equilibrium choice of advertisingfor publisherj as Ae. Withoutnetworkeffects, the social plannerwould take the demand curve to be the willingness-to-pay curve Pj(Aj, Uj(Ae, A_j)). Denote the social planner's choice of advertising level when usage is fixed at U(Ae, A_j) as Ao. Let (Ao - Ae)MCj be the cost of the extra production,assuming constant marginal cost (as is assumedin estimation).Classical deadweightloss is: IA
Pj(s, Uj(Ae, A_j))ds - (Ao - Ae)MCj. Ae
If a social plannerraisedadvertisingfrom Ae to Ao, usage would rise above U(Ae, A_j), which means that advertisingwould generatemore surplus,so the efficient level of advertisingwould be even higher.Denote the socially efficient level of advertisingby firm j and its competitorsas A, and A_j,. Networkdeadweightloss is definedas: A**
AAO
Pj(s, Uj(A*, A_j))ds
-
Cj(APj(s, Uj(Ae, A_j))ds
Ao).
In the figure, the space between the efficient willingness-to-pay curve and the equilibrium willingness-to-pay curve is network deadweight loss. As stated previously, it will often be efficient for the willingness-to-payof the last purchaserto be below marginalcost because of the networkeffect. Estimatesof structuralparametersallow us to measurenetworkdeadweight loss, classical deadweightloss and equilibriumconsumersurplus.Clearly,it is importantto use a model which distinguishesbetween -a a + a-U a |aA-in orderto performthis calculation.26 aA and aA 5. ESTIMATION To review, we estimate three equations simultaneously,equations (2), (4) and (5).27 While the derivation of the functional forms for estimation in Section 4 is involved, the final functionalforms are quite intuitive. Equation(2) is price regressed on a log-linear function of advertisingusage and demographicsand equation(4) is usage shareregressedon advertisingand demographics,with some adjustmentsso it can be interpretedas a nested logit model. 5.1. Methodology Because the presence of demand parametersin marginal revenue creates non-linear cross equationrestrictionson parameters,I estimatethis system simultaneouslyusing the Generalized Method of Moments (Hansen, 1982). For each equation, I construct a matrix of exogenous variablesZP, Zu and Zc, and assume E([v w(]I ZP, zu, zc) = 0. Each Z matrixcontains 26. Another reasonablechoice for the definition of Ao would be the level of advertisingAj that sets demand Pj(Aj, U(Aj, A_j)) equal to marginal cost. This choice has some appeal but using Pj(Aj, U(Aj, A_j)) clearly involves networkeffects. In orderto isolate network effects as much as possible, I define Ao with the willingness-topay curve Pj (Aj, U(Ae, A_j)), so usage is held constant. 27. Equation(4) is for the case where a directoryis overlappedby all competitors.In other cases, equation(4) is replacedwith the calculationdiscussed at the end of Section 4.2.
498
REVIEWOF ECONOMICSTUDIES
PJ(Aj,Uj(A*))
Ao
Ae
A
Aj
FIGURE 2
Classical deadweightloss vs. networkdeadweightloss
exogenous variablesin its respectiveequation(i.e. XP, Xu and XC), as well as the "excluded" variables discussed in Sections 4.1.2 and 4.2.2. Descriptions of the variables can be found in Table A.1 in the Appendix. The estimation problem is to choose parameters[a ,f y a] to minimize the criterion[m'Im], where I is a positive definiteweighting matrixand m=
-ZP'
v-
ZU' _Z'
~_
The (J x 1) matrices v, ~ and c are estimates of v, ~ and wcbased on estimates of a, fi, y and a. Hansen (1982) shows thatthis estimatoris consistentfor any positive definite 0, and thatthe estimatoris efficient if c is chosen to be the inverseof the correlationmatrixof the vector m. For price, I use the rate for a double-quartercolumn advertisement.This rate is the most closely watchedratein the Yellow Pages industryand is availableat practicallyall directoriesin my sample. When I used other rates, results did not change (note the high correlationbetween ratesin Table 1). Forthe quantityof advertisements,I used the numberof pages multipliedby the numberof columns. I multipliedthis numberby 0.8 if the book was observablysmaller than a standarddirectory.I focus on the resultsfor M = 75, althoughI also presentresults for M = 35 to show thatthe differencesare not substantial. 5.2. Instruments
Computationally,parameterson endogenous variables(al, a2, y, a) are identifiedby variables thatappearin the correspondingZP or ZU matrixbutdo not appearin the equationwhich defines the parameter.These exclusion restrictionsare reviewed in Table 2. Table 2 lists instruments next to each endogenousvariable.But note that in GMM, instrumentsare appliedto equations,
RYSMAN
COMPETITION BETWEEN NETWORKS
499
TABLE2 Orthogonalityrestrictions Dependentvariable sj /so Pj
Instruments
Endogenousvariable Aj sj Iyp
Populationcoverage Distributionarea Earnings Pub dummies(GTE, Bell South) % Switched county % Switched state % In same house
Aj Uj
TABLE3 Networkdeadweightloss Networkdeadweightloss Result Summaryvariable
S.E.
Equilibriumadvt (pages) Classical social optimum Social optimum Equilibriumsurplus($000's) Class. soc. opt. surplus Soc. opt. surplus Classical deadweightloss Networkdeadweightloss Ratio of NDWL to CDWL Ratio of total DWL to equ. surp.
110 506 1511 23,054 25,439 32,535 2541 7725 1.20 0.09
418 1784 3039 25,595 30,515 36,788 4920 6273 1.28 0-43
S.E. = standarderror.
not variables. So computationally,each instrumentidentifies both endogenous variablesin its equation.Equation(5) does not introduceany furtherendogeneityproblemsso I use Zc = Xc. In GMM, there is no equivalentto the first stage of Two Stage Least Squares,but Table4 replicatesthe "firststage" in orderto gain intuition about identification.Table 4 presents OLS regressions of endogenous variables on their respective exogenous variables. That is, ln(A) and ln(sj Iyp) are regressed on ZU, and ln(A) and ln(U) are regressed on ZP. Variablesin bold representexclusion restrictions.For the usage equation,advertisingincreasesin population coverage, and within-group share increases in distributionarea, both as expected. For the advertising demand equation, the quantity of advertising decreases in earnings and usage increases in the numberof people who lived in a different state and a differentcounty 5 years previous.The only unexpectedresultis thatusage increasesin the numberof people who owned the same house as 5 years previous. The R2 statistics are fairly high, always greaterthan 0.5. Also, for each regression, an F-test rejects joint insignificance of the excluded variables at a 95% confidence level.28
6. RESULTS Table 5 presents the main results. The effect of usage on advertising and the effect of advertising on usage are positive and significant, implying the existence of a network effect.29 The effect of 28. Note that there are two "first-stage"regressions for ln(Aj). This follows because I use separate sets of instrumentsfor each equationalthoughformallyall instrumentscould be appliedto all threeequations. 29. If doubling averageusage doubled demandfor advertising,we would expect al to be 1. There are a number of explanationswhy a1 is significantlyless than 1. It may be that in areas with high usage, consumersare more willing
500
REVIEWOF ECONOMICSTUDIES TABLE4 "Firststage" regressions.Theseare OLSregressionsof endogenousvariables on their respective instruments.Note that in GMM,there is no explicit 'first stage" Usage equationinstruments Within-groupshare Advertising
Dependentvariable:
(Sj I yp)
Population coverage Distribution area Constant % Urbanpopulation % Lived in diff county % Lived in diff state % Own house % Gradhi school % Gradcollege Per cap. income Telco book # Building permits County growthrate % Takepublic trans. % Have not moved Pop. density
0-645 0.078 2.164 0.005 0-045 0.035 0-020 -0-025 -0-012 0-050 0-681 -0-059 -0-010 0.008 -0-008 -2.80E-05
F-stat for bold variables R-squared
20-94 0-60
(0-041) (0-026) (0-72) (0-00) (0-01) (0-01) (0-01) (0-01) (0-01) (0-02) (0-06) (0-051) (0-012) (0-026) (0-012) (3-33E-05)
0.547 0.370 -5-238 0-001 0-002 -0-007 0-021 -0-047 -0-099 0-171 1-666 -0-203 0.010 0-149 -0-033 -2-54E-04 13-10 0.76
Price equationinstruments
Dependentvariable:
Usage
Advertising Earnings per worker Bell South GTE % Have not moved % Lived in diff county % Lived in diff state Constant % Urbanpopulation % Gradhi school % Gradcollege Per cap. income Telco book Populationcoverage Establishmentsper cap. Populationdensity F-stat for bold variables R-squared
(0-065) (0-039) (1-577) (0-006) (0-019) (0-024) (0-015) (0.016) (0-024) (0-038) (0-093) (0-089) (0-030) (0-040) (0-027) (4-65E-05)
-0-020 0.093 0-041 0.011 0.057 0.048 1.878 0.003 -0-017 -0-022 0.060 0.683 0.699 0.141 -2-94E-05 3.56 0.59
(0-009) (0-143) (0-087) (0-007) (0.009) (0-011) (0-574) (0-002) (0-008) (0-011) (0-019) (0-062) (0-039) (0-100) (2-68E-05)
-0-019 0-338 -0-163 0.040 0-087 0.065 -2-463 -0-005 -0-037 -0-039 0-078 1.547 0-342 0-110 -7-80E-05
(0-014) (0-218) (0-133) (0-010) (0-014) (0-016) (0-874) (0-004) (0-013) (0-017) (0-030) (0-095) (0-059) (0-152) (4-08E-05)
3.33 0.54
Bold variablesare "excluded". Standarderrorsare in parenthesis.
the quantityof advertisingon the price of advertisingis negative and significant.The estimate of a is high (0-803), suggesting that consumersview directoriesas similar products.Note that consumers have much strongerdemand for directories associated with telephone companies. Theredoes not seem to be a similareffect in advertiserdemand.Most of the othercoefficientsare insignificant,althoughmost of the ones thatare significantare of the expected sign (for instance, to searchthroughlistings and small advertisements,which mitigatesconsumers'impact on the demandfor advertising. Another possibility is that consumers that use directoriesmore use them for differentheadings, as opposed to using each headingmore often. This behaviourcould have potentiallycomplicatedeffects on demand.Both of these issues are difficultto addresswith this data-set.
RYSMAN
COMPETITION BETWEEN NETWORKS
501
TABLE5 Resultsfrom the generalizedmethodof moments Dependentvariable Price of advertising
Quantityof advertising(y) Usage (ai) Constant Populationcoverage Telco book Establishmentsper cap. % Urbanpopulation % Gradcollege % Gradhi school Per cap. income Pop. density
Usage (ln(s /so))
Advertising(a2) Constant % Urbanpopulation % Lived in diff county % Lived in diff state % Have not moved % Own house % Gradhi school % Gradcollege Per cap. income Telco book # Of house-buildingpermits Countypop. growthrate % Takepublic trans. Pop. density Constant Populationcoverage Earningsper worker Pop. density Bell South GTE
Marginal revenue
Correlation(sigma) P-value of exogeneity test, df Numberof observations
Mrg Efct
0.001 0.024 0.181 0.170 0.010 -0-104 -0-104 -0-133 0-065 2.536 -0-598 0-129 0-053 -4-20E-04
Coefficient
S.E.
-0-729 0-564 6.735 0.741 0.013 0.026 -2-02E-04 0-009 0.009 0-014 2-68E-05
0-193 0-131 0-715 0-100 0.138 0-064 2.25E-03 0-009 0-006 0-014 1-44E-05
0-154 -2-964 0-003 0-022 0-020 0-009 -6-05E-03 -0-012 -0-016 0-008 0-304 -0-072 0-015 0-006 -5-03E-05
0.073 0-878 0.003 0-008 0.008 0.010 0-007 0.007 0.008 0.013 0.093 0.032 0.010 0-015 1-88E-05
3-228 0.437 0-003 9-65E-05 -0-631 0-612
0-677 0.116 0.014 4.03E-05 0.529 0.129
0.803 0-88
0.079 4 428
S.E. = standarderror.
the per cent of people who lived in differentstates and the per cent in differentcounties 5 years previous are positive in usage, as expected). Two surprisingresults are that the percentageof college graduatesand the numberof new home-buildingpermitsare negativein usage. It may be that the opportunitycost of time for college graduatesoutweighs the effect that higher educated people are more likely to use Yellow Pages. Also, the numberof new homes as a percentageof county populationis supposedto capturegrowthbut may actuallyproxy for sparselypopulated areas,which often have low Yellow Pages usage. Table 6 presentsthree other specificationsin orderto explore the robustnessof the results. The first column presents results without instrumenting.In this specification,I hold a fixed at 0-8 in orderto focus on changes in the othercoefficients and also to speed up convergence.30As expected, the coefficient on quantityin price is closer to zero thanin Table5, -0-004 insteadof -0-729. Also, the coefficienton advertisingin usage (a2) is higher(0-230 relativeto 0-154) when 30. Fixing a fixes 8 which means the fixed point may be computedonly once at the beginningof the program.
502
REVIEW OF ECONOMIC STUDIES TABLE6 Robustnessresults No instrumenting Coefficient S.E.
Price equation Advertising Usage Constant Populationcoverage Telco book Establishmentsper cap. % Urbanpopulation % Gradcollege % Gradhi school Per cap. income Pop. density Ads at competingbooks Usage at competingbooks Usage equation Advertising Constant % Urbanpopulation % Lived in diff county % Lived in diff state % Have not moved % Own house % Gradhi school % Gradcollege Per cap. income Telco book No. of house-buildingpermits Countypop. growthrate %Takepublic trans. Pop. density Cost equation Constant Populationcoverage Earningsper worker Pop. density Bell South GTE Correlation(sigma) P-value of exogeneity test, df Numberof observations
Marketsize (M) = 35 S.E. Coefficient
w/Competitionvars. S.E. Coefficient
-0.004 0.116 4.435 0.392 0.216 0-063 -0.004 -0.004 0.002 0-025 2-33E-05
0-035 0-020 0.301 0-035 0-047 0-045 0-001 0-005 0-004 0-010 1.44E-05
-0-727 0.562 6-730 0.741 0.017 0.025 -1.84E-04 0.009 0.009 0-014 2-63E-05
0.192 0.130 0.712 0.100 0.137 0.064 2-24E-03 0.009 0.006 0-014 1.44E-05
-0-757 0.762 6.762 0-663 -0-362 0.061 0.001 0.004 0.011 0-024 3-27E-05 0.062 -0-103
0-501 0-521 0.976 0-194 0.402 0-094 0.004 0.011 0011 0-020 1-79E-05 0.080 0.158
0.230 -2-419 0.003 0.025 0.020 0.012 -0-009 -0.011 -0-015 3.30E-04 0-260 -0.094 0.019 0-006 -5-89E-05
0-034 0.639 0.003 0-008 0-010 0-010 0-007 0-009 0.009 1-43E-02 0-059 0-038 0-010 0-017 2-16E-05
0.158 -1.800 0-004 0-026 0-023 0-008 -0-006 -0-016 -0.016 0.006 0.329 -0-082 0.015 0.010 -6-35E-05
0.082 1.007 0.003 0.009 0.010 0.012 0.008 0.009 0.010 0.016 0.107 0-037 0.011 0.018 2.26E-05
0-146 -2-987 0.003 0-022 0.021 0-010 -0-006 -0-011 -0.015 0.004 0-304 -0-079 0-015 0-005 -4.87E-05
0.074 0.883 0.003 0.008 0.009 0.011 0.007 0.007 0.008 0-014 0.094 0.032 0.010 0-015 1.92E-05
4.333 0-521 -0-004 1-41E-05 0.587 0.121
0-367 0.057 0.011 2.06E-05 0-083 0.086
3-154 0.438 0.009 8.38E-05 -0.424 0.505
0.936 0.113 0.026 6.55E-05 1.272 0.203
3.140 0.445 0.006 8.75E-05 -0-555 0.632
0.890 0.121 0.015 4.71E-05 0.672 0-200
0.8
Fixed Just identified 428
0.0796 0.9125
0.091 4 428
0.807 0.91
0.079 2 428
S.E. = standarderror.
not instrumenting. However, an unexpected result is that when not instrumenting, the coefficient on usage in price (al) is lower (0.116 instead of 0.564), suggesting that endogeneity is biasing that coefficient towards zero. One explanation for this result may be that the publisher first-order condition implies that advertising is a concave function of price. Also, usage is endogenous only because it is a (non-linear) function of advertising. That is, the coefficient on usage in the regression without instrumenting may be lower than expected because it is capturing some of advertising's non-linear relationship with price. This argument implies that it is necessary to instrument for both advertising and usage to see the exogenous effect of either variable. In fact,
RYSMAN
COMPETITIONBETWEENNETWORKS
503
when I instrumentfor only advertisingor only usage, the coefficients do not change nearly as much as when I instrumentfor both. In a separateregression, I estimate the usage equation withoutinstrumentingfor sj Iyp, the within-groupmarketshare.As expected, the estimateof a is much lower, only 0-43 relativeto the estimate0-803 found in Table5. The second column of Table 6 presentsresults when M, the numberof times a household needs informationper month, is set down to 35. The results are essentially the same except, as expected, the constantterm in the usage equationis much higher. The parameterM cannot be set much lower as I observe areas that use directories up to 25 times per month. Results also do not change when M is set much higher than 75. The thirdcolumn tests the accuracyof AssumptionsAl and A2 by augmentingthe vector of explanatoryvariablesfor the advertising demandequation with the two "competitionvariables"from Table 1. Specifically, the demand equationincludes an index of the amountof advertisingat competingdirectoriesand the amount of usage going to competing directories,both in logs. The coefficients are close to zero and insignificantsuggestingthatoutcomesat competingdirectoriesdo not affect each other'sdemand for advertisingdirectly. This feature is consistent with advertisers'decisions being separable in their choices at different books, and with Assumptions Al and A2. Note that competition variablesmight be endogenouslydeterminedand I did not include them in the instrumentvector. These results are robustto includingthe measureof usage at competitorswithoutthe measureof advertisingat competitors,and vice versa. Returningto the results in Table 5, the model fits the data well. When I use the three equations to predict all three endogenous variables simultaneously, the simple correlation between predicted advertisingand observed advertisingis 0.621. The correlationfor price is 0.618 and the correlationfor usage is 0.794. Note that the effect of missing a predictionby a small amountin one equationis magnifiedby the positive feedback structure.As anothercheck, I compute the profitsof directoriespublishedby telephone companies to be 33-5% of revenue. The numberis very close to quotes made to me by membersof the industryof 35-45%. How does the model explain the asymmetricmarket shares in the data? The theoretical model can explain the asymmetricoutcomes either with asymmetricdirectorycharacteristics or with multiple asymmetricequilibria.In estimation, I find that there is a unique equilibrium (exploredin detail in the next subsection)but thatdirectoryasymmetriesare important,captured by the large coefficient on the dummy variable for being a telephone company directory in the usage equation. Given that telephone directories always seem to be the ones with high price-quantity-usageand independentsalways have low, it is not surprisingthat the estimation procedureexplains the datawith the dummyvariableas opposed to findingmultipleequilibria. In orderto determinethe importanceof the networkeffect, I calculateclassical and network deadweight loss as discussed in Section 4.4. Figure 3 draws the estimated demand curve for a single directoryusing parameterestimates from Table 5. The solid line (Pj(Aj, Ue)) is the willingness-to-pay curve for the equilibriumchoice of advertising Ae. A social planner who did not take advantageof the network effect would choose advertising level Ao. The actual optimalchoice is A*, where demand(Pj(Aj, Uj (Aj))-the dottedline) is below marginalcost. Choosing A* places the marketon the optimal willingness-to-paycurve Pj (Aj, U*) (the dashdotted line). The space between Pj(Aj, Ue) and Pj(Aj, U*) representsnetwork deadweight loss. Classical deadweight loss is the triangle between Pj (Aj, Ue) and marginalcost MC, to the right of Ae. The results for an average market are computed in Table 3.31 The ratio of networkdeadweightloss to classical deadweightloss is 1.26 (note the large standarderror:1-2). Total deadweight loss equals 0.43 of equilibriumconsumer surplus (with a standarderrorof
31. Standarderrorsin Tables3, 7 and 8 are calculatedusing the delta method.
504
REVIEWOF ECONOMICSTUDIES
500
900
1300
1700
2100
2500
2900
3300
Numberof Double-QuarterColumns FIGURE 3
Deadweightloss at estimatedparameters
only 0-09). In a marketwhere producersclearly exhibit strong marketpower, network effects still createa large amountof deadweightloss. 6.1. An application:entry Network effects are at least moderately important in the sense that the positive feedback parametersare statisticallysignificantand networkdeadweightloss is non-trivial.Are network effects so importantthat the benefits of monopoly outweigh the benefits of competition?What parametersdeterminethis outcome?This section presentsan entryexperimentin orderto answer these questions. The results from the entry experimentappearin Table 7. This table presents equilibriumoutcomes for differentnumbersof symmetriccompetitorsthat perfectly overlapin an averagemarket.Duopoly firmseach choose higher quantitythan a monopolist,reflectingthe strengthof the competitiveeffect. This resultoccursbecausethe estimateof a is so high. Because there is little differentiationbetween directories,a book that is slightly largercapturesmost of the usage market.So two competing directoriesdrive each other to high levels of advertising. Even so, usage at each book drops substantiallyfrom the monopoly case. As each publisher enters thereafter,advertisingand usage at each directoryshrink,so the benefits of the network effect are dissipated.32However,total advertisingand usage increase, reflectingthe benefits of competition. Table 7 also presentsprofit, advertisersurplusand the sum of the two for each numberof competitors.The results show that, ignoring fixed costs, total surplus(advertiserand publisher) improves in the numberof competitors.In this market,networkeffects are not strong enough to imply that the benefits of monopolizationoutweigh the benefits of competition.Note that the results of the model could have been differentif the networkeffect had been estimated to be stronger.Consider raising the network parameterin the advertisingdemand equation (al, the coefficient on usage) and recalculatinghow surpluschanged over the numberof competitors. Figure 4 maps mean surpluslevels when the parameteris 48 and 55% larger.The figures show 32. While the non-monotonicityin directory-leveladvertisingis clear in Table 7, it is difficult to verify in data because the overlappingstructureof the distributionareas means that there are almost no clear duopoly or triopoly markets.
RYSMAN
505
COMPETITIONBETWEENNETWORKS TABLE7 Equilibriumfor differentnumbersof competitors
No. of Advertising(pages) Refs./HH/mth. competitors 1 2 3 4 5 6 7
613 707 624 549 490 443 405
(578) (606) (533) (470) (420) (381) (349)
4.10 2.38 1.68 1.30 1.07 0.91 0.79
Price ($) (DQC ad)
(0-69) 2136 (0-38) 1416 (0.28) 1273 (0-22) 1212 (0-19) 1178 (0-16) 1156 (0.15) 1141
Profits($)* Advertisersurplus* Totalsurplus* (1 directory)
(1207) 5.16 (794) 2.85 (736) 1.97 (712) 1.53 (699) 1.26 (690) 1.08 (684) 0.95
(1.60) (1.00) (0.79) (0-68) (0-60) (0.55) (0-50)
21.45 16.40 13-03 10-91 9.45 8.38 7.57
(17.07) (13.10) (10-53) (8.94) (7-85) (7-05) (6.43)
26.61 38.50 45.00 49.74 53.55 56.79 59-62
(19.67) (29.45) (35-06) (39-39) (43.01) (46.18) (49-02)
*Profits and surplus are in millions. Profits and surplus are computed assuming there are no fixed costs of production. Standarderrorsare in parenthesis.
TABLE8 Private returnsvs. social returns No. of competitors 2 3 4 5 6 7
Surplusincrease minus profits(%) (no fixed costs)
Profits (incl. fixed costs)
Surplus increase (%) (incl. fixed costs)
Adjustedsurplus increase (%) (incl. fixed costs)
0.76 0-70 0.68 0.67 0.67 0.66
1.80 0.92 0.48 0.21 0.03 -0.10
0.42 0.15 0.09 0.06 0-05 0-04
0.26 0-07 0.03 0.01 0-00 -0.01
(0.17) (0.22) (0-25) (0.26) (0.27) (0.27)
(1-15) (0.98) (0-90) (0-85) (0-82) (0-80)
(0.11) (0.06) (0.04) (0.03) (0.03) (0-03)
(0.11) (0.08) (0-07) (0.06) (0-06) (0-06)
Surplusincreaseminus profits(%) is (incsurp(k,k - 1) -prof(k))/incsurp(k, k - 1). Surplus increase (%) is incsurp(k,k - )/surp(k- 1) where surp(k) equals surplus generated by k competitors.incsurp(k,k - 1) = surp(k) - surp(k - 1). prof(k) is profit when there are k competitors.Adjusted surplusis computed ignoring the upper tip of the demand curve. Standard errorsare in parenthesis.
thatfor largenetworkeffects, the model implies thatwelfare decreasesin the numberof competitors or could even be hump shaped.Again, these results do not take accountof any fixed costs. For the actual parameterestimates, surplus increases in the number of competitors.The crucial question for welfare purposes is: how does the increase in surplus due to an entrant compare to the profits of the entrant?Table 8 compares the social benefits of entry to the private benefits capturedby the firm. The first column of Table 8 shows that surplus from entrantsis considerablyhigher than their profits. The difference between the increase in total surplusand profitsis significantlydifferentthan zero for each entrant.This computationis done withoutconsideringfixed costs, but almost whereverfixed costs lie, there will be under-entryin equilibrium.This result suggests that currentlaws that allow entry in the Yellow Pages market should be encouraged. If we included consumers in the analysis, the result would be even stronger. In this model, seeing that total usage increases in the number of competitorsimplies that consumer welfare increases in competition.While we cannot convertutils into dollars without observing consumers' response to price, we can use the discrete choice model and parameterestimates to see how much welfare increases in competition. Going from one directoryto four makes consumer welfare from the Yellow Pages marketgo up by 22%, and going from one to seven increaseswelfare from the Yellow Pages marketby 35%.
REVIEWOF ECONOMICSTUDIES
506
EstimatedNetworkParameters
0o x
0
2
14 8 10 12 6 Numberof Independents
4
16
20
18
Usage Parameter * 1-48
1 1-2494 x
I
I
I
I
.
.
I
I
.
.
.
.
.
I
.
.
I .
.
.
.
I
I
1-2492
-
1-2490
1-2488 1i /?-)Ag,4IL~ QK
r
2
0
1
x
1-4693
I
I
I
I
.
.
.
.
.
14 6 8 10 12 Numberof Independents
4
I
I
I
.
Usage Parameter * 1-55 ...................
I
I
I
I
II.I
.
.
.
.
.
I
I.
.
.
16
I
18
I.
I.
I
I I
.
20
I
14691
c
1-4689
1-4687
.
0
2
4
I
I
14 6 8 10 12 Numberof Independents
.
16
18
.
20
FIGURE 4
Surplusvs. the numberof competitors
6.2. Equilibriumentryand the optimalnumberof directories If we knew the fixed cost of producinga directory,we could push the model furtherand calculate the numberof entrantsin equilibriumas well as the optimal numberof entrantsfor a market. While obtaining data on the fixed costs of setting up a Yellow Pages directory is difficult, I can obtain a crudeestimateby imposing a zero-profitcondition on independentdirectories.The estimate of the fixed cost for independentpublishers is $1,004,700 (with a standarderrorof
RYSMAN
COMPETITIONBETWEENNETWORKS
507
$765,030). Telephonecompanies, which are requiredto distributeWhite Pages to every phone line, most likely have substantiallylower fixed costs for publishing Yellow Pages. The second column of Table 8 computes profitsfor independentscounting this fixed cost. We cannot reject the possibility that only a telephone companydirectoryis profitablein an average-sizedmarket. This predictionseems largely accurate,as we observe entrymostly in largermarkets. In order to calculate the optimal number of entrants,Table 8, column 3, computes the proportionalincreasein surplusaccountingfor the fixed cost of setting up a directory.With 95% confidence, we can say that four directoriesare better than three or less.33 The data have very few marketswith four or more directoriesso I take these resultsto mean that a more competitive marketis preferableto the currentmarketstructure.As before, measuringconsumerutility would mean thatthe benefitsto entrywould be even higher. Why does entry benefit advertisers?When a new directoryenters the market,advertisers benefit in three ways. First, the directoriescompete and drive down prices. Second, consumers view directories as differentiatedproducts and so more directories attractmore consumers, benefiting advertisers.Third, advertisers view the directories as differentiatedproducts and similarly,advertisersvalue relief from congestion costs. Which of these three effects drives the result?The following computationsshow that the effect of competitionis the crucial one in this market. First, I removed the competition effect by recalculating the entry experiment assuming that all directorieswere owned by a single publisher (results not reported).In that case, even having two directoriesis not significantlybetterthanhaving one. Second, I mitigatedthe effects of product differentiationfor consumers by recalculatingthe entry experimentfor competing directoriesbut with higher values of a (also not reported).The optimal numberof directories decreases as this differentiationdecreases because fewer consumers switch from the outside option when a new directoryenters so advertisersderive less benefit from entry.However,more substitutabledirectoriesmeans that the first two directoriesdrive each other out to very high levels of advertising.So reducingdifferentiationfor consumersdecreasesthe optimalnumberof directoriesbut also increases the value of the second entrant,so the optimal numbernever goes below two.34 Therefore,the result that multiple directoriesare preferableto a single directory would not hold withoutcompetingdirectoriesbut would hold for directorieswith differentlevels of differentiation. Finally, a potential problem with the computationsin Table 7 is that a steep log-linear demand curve for advertisingimplies that there is a large group of retailers who place very high value on each directory.This surplus can be interpretedas coming either from product differentiation(from the point of view of advertisers)or from a reductionin congestion effects, both of which are capturedin the slope of the advertisers'demandcurve.35While these effects are presumablymeasuredaccuratelyin the range of the data, welfare calculationscould depend heavily on the top of the demand curve, where we also have very little data. As an ad hoc investigationof this possibility, I recomputewelfare from the outcomes in Table 7, but ignore the "tips"of the demandcurve. Thatis, for welfare calculations,I assume thatthe demandcurve is a horizontalline from the Y-axis over to a kink, and then log-linearaccordingto the estimated 33. Because surplusincreases in the numberof competitorsfor any given set of parameters,it is possible for the proportionalincrease statistic to be statistically different than zero even though 95% confidence intervals aroundthe surplusstatisticsoverlap. 34. A possible criticism of my approachis that the model capturesgeographicdifferentiationin Eij. If the model accountedfor geographicdifferentiationexplicitly (a difficultendeavour),the estimateof a might be higher.But this last result suggests that doing so is unlikely to reversethe result thata more competitivemarketis preferable. 35. As noted in footnote 17, I could betteraddressthe congestion issue if I could separatelyestimate the effect of congestion from the otherwisedownwardslope in the inverse demandcurve.
508
REVIEWOF ECONOMICSTUDIES
demandcurve thereafter.Therefore,I calculatewelfare as before and then subtract Aj P(s, U(Ae, A_j))ds
-
P(Aj, U(Ae, A_j))Aj
where Aj is the kink point for directoryj and Ae and A_j are equilibriumadvertisinglevels. The last column of Table 8 repeatscolumn 3 but with Aj set to 30% of equilibriumadvertising. For this case, we can no longer say thatfour directoriesare preferredto three.However,even for this case, two directoriesare preferredto one. In fact, to reverse the result that two directories are preferredto one, one would have set the kink at a very high point, well over 40% of equilibriumadvertisingand well into the range where we have data for most directories.These computationssuggest thatthe resultthatsome competitionis preferredto monopoly is not driven by assumptionsaboutthe demandcurve outside of the rangeof the data. 7. CONCLUSION This paperexamines the welfare trade-offbetween competitionand standardizationin a market characterizedby positive networkeffects. The paperpresentsa model of the marketfor Yellow Pages that explicitly captures the relationshipbetween advertising and consumer usage in a directory.The paperestimatesthe model by extendingthe techniquesof Berry(1994) to the case of overlappingmarketswith distinct boundaries.The results show that, for a given directory, retailer demand for advertising increases in consumer usage and that consumer demand for directoryusage increasesin the amountof advertising,implying a networkeffect. In equilibrium, forgone surplus due to unexploited network effects is high relative to the amount of surplus obtainedin equilibriumand also relative to the deadweightloss due to imperfect competition. However,the results show that despite the networkeffect, a more competitivemarketstructure is preferable to a more concentratedone. Strikingly, the paper finds that multiple entrants improve welfare but are unprofitablein the averagemarket.The results of the paperimply that encouragingcompetitionin the Yellow Pages (as a numberof recentpolicy changes do) improves welfare. APPENDIX Equilibriumin the publishers' game This section shows thatthereexists an equilibriumin pure strategiesin the publishers'game. I show thateach publisher's objective function is concave in its choice variable and then I appeal to Theorem 1.2 in Fudenbergand Tirole (1991), originallyproven by Debreu, Glicksbergand Fan. Publisherj, j = 1, ..., J, chooses its quantityof advertisingAj to maximizeprofits:Pj (Aj, Uj (A1 .., Aj))AJ - Cj (Aj). For this sub-section,I assume Pj = AY'U1 Xj. The variable Xj capturesdirectorylevel characteristics.To parametrizeU(.), I assume a logit model holds in each sub-market,where mean utility to directory j is Sj = a2 ln(Aj) + ln(Xj). Sub-markets(described in Section 4.2) are regions with a uniformset of directories.Let K(j) be the set of sub-marketsthat are covered by directoryj and let C(k) be the set of directoriesthat serve sub-marketk. Let *k be the shareof j's marketrepresentedby sub-marketk (the j will be obvious in context). Therefore AJ2Xj ksjk* sjk =+1 S1K k EK(j) + EieC(k) A72 X i Usage is definedby Uj = Msj, where M is the size of the market.The parametrizedversion of the publisherfirst-order (equation(5)) is Pj 1 + y +
s1
L Ek
)
kSjk(-
Sjk))
=
MCj.
The term in large parenthesesrepresentsthe price-cost markup.Allow rj to equal the markup,so Pj Fj = MCj. Concavityrequiresthatthe second derivativeis negative:
RYSMAN (Pj
COMPETITIONBETWEENNETWORKS
Uj ) + Pa +Pia
j Uj U
Aj
a
j
f
k
ksjk(
kEK(j)
Aj
509
<
(A)
Sj
The term in the first set of parenthesesis the slope of the demand curve and is assumed to be negative. The pricecost margin rj is positive for any reasonable parametervalues. In the next term, Pj, a1 and a2 are each positive. k kK() ksjk is difficultto sign. In the case of perfect overlap,that term becomes Unfortunately,the term aU ZkFK(j)~kSjk(l-sjk) a-a ( 51-sj) which is obviously negative, so thereis a solutionin pure strategies.In the case withoutperfect overlap,the term is
a
aAj
EkEK(j)IkSjk(1
-
_ [ EkK(j)
jk)
-
ikSjk (
Si
Sj k)]
-K(j) (1 2jk)(
?k
sj
-
Sjk)jk.
The firstterm on the R.H.S. is always negativebut the second term might be positive. I must assume that al and a2 are not "too large"to guaranteean equilibriumin pure strategies.Condition(A.1) easily holds at the parametersestimated in this paper.
Computationaldetails of the usage equation In orderto writedown the correctrelationshipbetween sj and Sj, let K (j) be the set of sub-marketsin the regioncovered by directoryj, let irjk be the portion of directoryj's marketin sub-marketk, let s0k be the share of referenceto the outside option in sub-marketk and let sj IYPk be the shareto j in the Yellow Pages group.In this case, we have sj = eBj EkK(j)
(A.2)
jkSOkSj| YPk'
I do not observe sj IYPk or SOk.However,the assumptionson functional form that are a part of the nested logit model imply eSj /(1 -ar)
Sj I YPk
/( C(k))
Y)
1
Ok
1+
e
(A.3)
(Y-1I/C(k) eSl/l_
C (k) is the set of directorieswhich compete in sub-marketk. The equationsin (A.3) come from Berry(1994) and Cardell (1997). We can computesj as a functionof 3 by plugging equation(A.3) into (A.2). I cannot solve for 3 explicitly in this case. I use a fixed-pointalgorithmto solve this problem.Let s be the vector of observedmarketsharesand let s(3) be the vector of predictedmarketsharesdefinedby equations(A.2) and (A.3). Define the functiong(.) by g(3) =- + s -s(3).
(A.4)
In what follows, I show that equation (A.4) is a contractionmappingfor any orbelow a certaincut-off, ensuringthat a unique 3 exists and that it can be found. I check that a is less than the cut-off before each implementationof the fixed point algorithm. In order to increase identificationpower over a, I use the fact that sj Iyp = j Iypk and so = SOkat markets with only one sub-market.Thus, I can combine informationfrom equations(4) and (A.2). I use the following approach in orderto construct~: startwith a given a, a1 and fBu.Use a and equations(A.2)-(A.4) to obtain3 via the fixed point algorithm.Let Kj =- 1 if directoryj has one sub-marketand at least one competitor,and Kj = 0 otherwise. (Note that marketswith no competitorsconfer no informationabouta.) Define ,j by - ln(so) - a ln(sj Iyp) - a ln(A)l_ Jn(sj) a aj ln(Aj) Xjfu
Xjpu
if Kj = 1 if Kj = 0.
In this data, 121 directoriesout of 428 have a uniform set of directoriesoffered across their entire market.Of those, 65 do not overlapwith any otherdirectoriesand 56 are completely overlappedby all competitors. J Now I turnto establishingthat equation (A.2) has a unique fixed point. I show that the equationg: NRJdefinedas g () = 8 +s -s(3) is a contractionby showing thatthe functionsatisfiesthe conditionsstatedin the theoremin AppendixI of Berryet al. (1995). The importantconditionsto show are thatg(3) is continuousin 8, that agj (8)/83i > 0 0 for all i and j, and that i/1 agj I show that agj (8)/aj >n (g)/l ai < 1. The functionis continuousby construction. last as this partof the proof requiresconditions on a. First,I establishthat agj (3)/l3i > 0 for i 0 j. aj (3)
i
- ese E vaa EK(j)
o
ijk
k j YPk
REVIEWOF ECONOMICSTUDIES
510
TABLEA. 1 Explanationof variables Directorylevel variables Description
Name
Instrument vector
Numberof pages times numberof columns times 0.8 if observablysmaller,logged. From YPPAlibrary,collected by Boston ConsultingGroup. Rate for a double-quartercolumn ad, logged. From YPPARate and Data CD-ROM. Numberof uses per householdper month, logged. FromNYPM Area Reports. Numberof people covered by a directoryarea,logged. From YPPAIndustryCharacteristicsCD-ROM. Dummy, 1 if directoryis associated with phone company. Constructedby observing publisher,books and some companycontact. Dummy variables Squaremileage of distributionarea From ClaritasPower Pages CD-ROM
Advertising Price Usage Pop. coverage Telco GTE, Bell South Distribution Area
U,P,C U,P P,C U
Countylevel variables-from USA CountiesCD-ROM 1996 % Of populationin urbansetting, 1990 % Of populationin differentcounty 5 years previous, 1990 % Of populationin differentstate 5 years previous, 1990 % Of populationin same house as 5 years previous, 1990 % Of housing that is owner occupied, 1990 Per capitaincome, 1993 Populationin 1995 divided by squaremileage Rate of populationincreasefrom 1990 to 1995 % Of populationregularlyusing public transportation(1995) Per capitaearnings
Urbanpop. Diff county Diff state Have not moved Ownerocc. house Income Density Growth Public trans. Earnings
Following from Berry (1994), it is easy to show that asj IYPk/ai asOk/ai = - Siks0k. Plugging in, we have that
ai(6)
es
In orderto show that i1 In order to show that
(S)/ai
agj YPk._10gj(l)/OSi 1 - s()
+
) Ej = -s(S) Ei=I1gi ai~=EkKj
+
agj
= l-eJ
= -sj IYPksi IYPk/(1
si IYPk+
IjksoksO YPk (
kK(j
ik)>
= (s IYPk < 1, note that as YPk/ai 1,note that Osj[YPk/3~ii = (sjlypk - s)/ I
eJ
kK(j
esj
kK(j)
) VjkSOkSj1 YPk
U,P U,P U,P U,P U,P U,P U,P,C U,P U,P U,P,C
a
(Si IYPk
or) and that
a) Therefore <-)/(1 (1)T 1)+ Sjk)
'keK(j) 1jksoks IYPk(1 - sok) jksOsj
YPk < 1.
Now I establish conditions that ensure that agj(5)/aSj > 0. It is equivalentto show that asj(8)/asj sufficientconditionis that asjk ()/asj < 1 for all sub-regionsk. Suppressingthe subscriptk, I show that as1
as1
asj I ypsyp
:j Iyp
as1
asyp
as1
+ Syp
asj I yp
as1
< 1. A
<1.
= sj yp( - sjlyp)/(l -a). Also, we have that Following Berry (1994), we have that asjlyp/abj = - a)). The derivativereduces to asyp/asj syp = exp(x)/(1 + exp(x)) where x = (1 - a)ln j exp(6/(l syp (1 -syp)sj yp. Plugging into asj /a5j and solving for a shows thatwe have sufficientconditionsfor a contraction whenever 1 - sj( - sj) (A5) 1-
j(sjIYP
-
X,x .,Y
j)
RYSMAN
COMPETITION BETWEEN NETWORKS
511
It is easiest to study this condition by converting sj = sj ypsyp, as the latter two terms can be moved independentlyof each other. The upper bound on a decreases in syp and is convex in sj yp, reaching a minimum at sj IYP = 0.5. The boundis always greaterthan0.75 and less than 1. Thereis actually some intuitionto the resultthat a must be less than 1. We are tryingto show that asj /a8j is less than 1. When a is high, all of the randomnessin utility is placedat the grouplevel, which means thatwithin-groupchoices are based almost entirelyon mean utility.When mean utility (Sj) moves slightly,it generatesa big responseandsj rises too quickly.Of course, if few people choose the group, a can be higher and sj still will rise at a reasonablerate. And as is typical in logit models, the within-groupderivativeis highest when the within-groupmarketshareis close to 0-5. I do not impose the bound in estimation. Instead, I check for a to satisfy (A.5) at iterationof my optimization routine, before searchingfor a fixed point. Note that I always startmy estimationprocedurewith a guess of a that is below 0.75, ensuringthat a fixed point exists at my startingvalues. If the true a is greaterthanthe bound, I expect my estimateof a to rise. When the estimateof a crosses the bound,my estimationroutinestops becauseestimatesof 8 might be meaninglessat thatpoint. Acknowledgements. I thank Peter Arcidiacono, Ray Deneckere, John Kennan, Sam Kortum, Kevin Lang, Aviv Nevo, Larry Samuelson, Mark Stegeman, numerous seminar groups and especially Phil Haile for advice and encouragement.Also, the comments of three anonymousreferees and the editor significantlyimprovedthis paper.This researchwould not be possible without the generosity of NationalYellow Pages Monitor,Claritas,Inc., and the Yellow Pages PublishersAssociation. The ChristensenAwardin EmpiricalEconomics providedimportantfinancialsupport. REFERENCES ANDERSON, S. and COATE,S. (2001), "MarketProvision of Public Goods: The Case of Broadcasting"(Mimeo, CornellUniversity). ARMSTRONG,M. (2002), "Competitionin Two-SidedMarkets"(Mimeo, OxfordUniversity). BERRY, S. (1994), "EstimatingDiscrete Choice Models of ProductDifferentiation",Rand Journal of Economics, 25, 242-262. BERRY, S., LEVINSOHN,J. and PAKES, A. (1995), "AutomobilePrices in MarketEquilibrium",Econometrica,63, 841-890. BERRY, S. and WALDFOGEL,J. (1999), "FreeEntryand Social Inefficiencyin Radio Broadcasting",Rand Journalof Economics,30, 397-420. CARDELL, N. S. (1997), "VarianceComponents Structuresfor the Extreme Value of Logistic Distributionswith Applicationto Models of Heterogeneity",EconometricTheory,113, 185-213. CHOU, C. and SHY,O. (1990), "NetworkEffects without Network Externalities",InternationalJournal of Industrial Organization,8, 259-270. CHURCH, J. and GANDAL, N. (1992), "Network Effects, Software Provision, and Standardization",Journal of IndustrialEconomics,40, 85-103. ECONOMIDES, N. (1996), "The Economics of Networks",International Journal of Industrial Organization, 14, 673-699. ECONOMIDES,N. and FLYER,F. (1997), "Compatibilityand MarketStructurefor Network Goods" (Mimeo, Ster School of Business at New YorkUniversity). ECONOMIDES,N. and HIMMELBERG,C. (1995), "CriticalMass and NetworkSize with Applicationto the US FAX Market"(Mimeo, Ster School of Business at New YorkUniversity). ELLIOTT,S. (1998), "Advertising",New YorkTimes, 17 April, C7. FEISTPUBLICATIONS,INC. V. RURALTELEPHONESERVICECOMPANY,INC. (1991) SupremeCourtReporter, 111 S.Ct. 1282. FUDENBERG,D. and TIROLE,J. (1991) Game Theory(Cambridge:MIT Press). GANDAL, N. (1994), "HedonicPrice Indexes for Spreadsheetsand an EmpiricalTest of Network Exteralities" Rand Journalof Economics,25, 161-170. GANDAL, N., KENDE, M. and ROB, R. (2000), "The Dynamics of Technological Adoption in Hardware/Software Systems: The Case of CompactDisk Players",Rand Journalof Economics, 31, 43-62. GOOLSBEE,A. and KLENOW,P. (2002), "Diffusion of Home Computers",The Journal of Law and Economics, 45, 317-343. GREENSTEIN, S. M. (1993), "Did an Installed Base Give an Incumbent any (Measurable)Advantage in Federal Rand Journalof Economics,24, 19-39. ComputerProcurement?", HANSEN, L. (1982), "LargeSample Propertiesof GeneralizedMethod of Moments Estimators",Econometrica, 50, 1029-1054. KATZ,M. L. and SHAPIRO,C. (1985), "NetworkExternalities,Competition,and Compatibility",AmericanEconomic Review,75, 424-440. KATZ,M. L. and SHAPIRO,C. (1994), "SystemsCompetitionand NetworkEffects",Journalof EconomicPerspectives, 8,93-115. LABAND, D. N. (1986), "Advertisingas Information:An EmpiricalNote",Review of Economics and Statistics, 68, 517-521.
512
REVIEW OF ECONOMIC STUDIES
LIEBOWITZ,S. J. andMARGOLIS,S. E. (1994), "NetworkExternality:An UncommonTragedy",Journalof Economic Perspectives,8, 133-150. MIXON, F. G. (1995), "Advertisingas Information:FurtherEvidence'"SouthernEconomicJournal,61, 1213-1218. NARUC, Compilation of Utility Regulatory Policy 1994-1995 (Washington, DC: National Association of RegulatoryUtility Commissioners)376. OHASHI,H. (2000), "NetworkExternalitiesand ConsumerWelfare:Home Video CassetteRecordersin the U.S., 19781986" (Manuscript,Universityof BritishColumbia). PARK, S. (2000), "QuantitativeAnalysis of Network Externalities in Competing Technologies: The VCR Case" (Manuscript,SUNY at Stony Brook). ROCHET, J.-C. and TIROLE, J. (2003), "PlatformCompetition in Two-Sided Markets' Journal of the European EconomicAssociation, 1 (4). ROSSE, J. N. (1970), "EstimatingCost Function Parameterswithout Using Cost Data: IllustratedMethodology", Econometrica,38, 256-275. RUST, J. (1987), "OptimalReplacementof GMC Bus Engines:An EmpiricalModel of HaroldZurcher"Econometrica, 55, 999-1033. SALONER,G. and SHEPARD,A. (1995), "Adoptionof TechnologiesWithNetworkEffects:An EmpiricalExamination of the Adoption of AutomatedTellerMachines"RandJournalof Economics,26, 479-501. SHY, O. (2001) The Economicsof NetworkIndustries(Cambridge:CambridgeUniversityPress). STEGEMAN,M. (2002), "CompetitiveProvision of Non-rivaland Pure Public Goods ThroughAdvertising"(Mimeo, VirginiaPolytechnicInstitute). WHITE, E. D. and SHEEHAN, M. F. (1992), "Monopoly,the Holding Company,and Asset Stripping:The Case of Yellow Pages",Journalof EconomicIssues, 26, 159-182.
