effects of wom vs traditional marketing

8/3/2019 Effects of WOM vs Traditional Marketing

1/24

1Journal of Marketing

Vol. 73 (September 2009), 1000

2009, American Marketing Association

ISSN: 0022-2429 (print), 1547-7185 (electronic)

Michael Trusov, Randolph E. Bucklin, & Koen Pauwels

Effects of Word-of-Mouth VersusTraditional Marketing: Findings from

an Internet Social Networking SiteThe authors study the effect of word-of-mouth (WOM) marketing on member growth at an Internet social networkingsite and compare it with traditional marketing vehicles.Because social network sites record the electronic invitationsfrom existing members, outbound WOM can be precisely tracked. Along with traditional marketing, WOM can thenbe linked to the number of new members subsequently joining the site (sign-ups). Because of the endogeneityamong WOM, new sign-ups, and traditional marketing activity, the authors employ a vector autoregression (VAR)modeling approach. Estimates from the VAR model show that WOM referrals have substantially longer carryovereffects than traditional marketing actions and produce substantially higher response elasticities. Based on revenuefrom advertising impressions served to a new member, the monetary value of a WOM referral can be calculated;this yields an upper-bound estimate for the financial incentives the firm might offer to stimulate WOM.

Keywords: word-of-mouth marketing, Internet, social networks, vector autoregression

Michael Trusov is Assistant Professor of Marketing, Robert H. SmithSchool of Business, University of Maryland (e-mail: mtrusov@rhsmith.

umd.edu). Randolph E. Bucklin is Peter W. Mullin Professor of Marketing,Anderson School of Management, University of California, Los Angeles(e-mail: [email protected]). Koen Pauwels is an associate pro-fessor, Ozyegin University, Istanbul, and Associate Professor of BusinessAdministration, Tuck School of Business, Dartmouth College (e-mail:[email protected] and [email protected]). Theauthors thank the three anonymous JM reviewers and participants of the2006 Marketing Dynamics Conference and the 2007 DMEF ResearchSummit for helpful comments.The authors are also grateful to the anony-mous collaborating firm for providing the data used in this study.KatherineN. Lemon served as guest editor for this article.


2/24

Word-of-mouth (WOM) marketing has recentlyattracted a great deal of attention among practi-tioners. For example, several books tout WOM as

a viable alternative to traditional marketing communicationtools. One calls it the worlds most effective, yet leastunderstood marketing strategy (Misner 1999). Marketersare particularly interested in better understanding WOMbecause traditional forms of communication appear to belosing effectiveness (Nail 2005). For example, one surveyshows that consumer attitudes toward advertising plum-

meted between September 2002 and June 2004. Nail (2005)reports that 40% fewer people agree that advertisements area good way to learn about new products, 59% fewer peoplereport that they buy products because of their advertise-ments, and 49% fewer people find that advertisements areentertaining.

Word-of-mouth communication strategies are appealingbecause they combine the prospect of overcoming con-sumer resistance with significantly lower costs and fastdeliveryespecially through technology, such as the Inter-net. Unfortunately, empirical evidence is currently scantregarding the relative effectiveness of WOM marketing inincreasing firm performance over time. This raises the need

to study how firms can measure the effects of WOM com-munications and how WOM compares with other forms ofmarketing communication.

Word-of-mouth marketing is a particularly prominentfeature on the Internet. The Internet provides numerousvenues for consumers to share their views, preferences, orexperiences with others, as well as opportunities for firmsto take advantage of WOM marketing. As one commentatorstated, Instead of tossing away millions of dollars onSuperbowl advertisements, fledgling dot-com companiesare trying to catch attention through much cheaper market-ing strategies such as blogging and [WOM] campaigns(Whitman 2006, p. B3A). Thus, it is important to under-

stand whether WOM is truly effective and, if so, how itsimpact compares with traditional marketing activities.One of the fastest-growing arenas of the World Wide

Web is the space of so-called social networking sites. Asocial networking site is typically initiated by a small groupof founders who send out invitations to join the site to themembers of their own personal networks. In turn, newmembers send invitations to their networks, and so on.Thus, invitations (i.e., WOM referrals) have been the fore-most driving force for sites to acquire new members. Associal networking sites mature, they may begin to increasetheir use of traditional marketing tools. Therefore, manage-ment may begin to question the relative effectiveness ofWOM at this stage.

The objective of this research is to develop and estimatea model that captures the dynamic relationships among newmember acquisition, WOM referrals, and traditional mar-keting activities. In doing so, we offer several contributions.First, we are among the first to link observed WOM directlyto new customer acquisition. Second, we show how toincorporate both the direct effects and the indirect effects ofWOM and traditional marketing actions (e.g., a marketingaction increases WOM activity, which in turn increases newmember acquisition). We empirically demonstrate, for our

2 / Journal of Marketing, September 2009


3/24

data set, the endogeneity among new member sign-ups andthese marketing variables. This highlights the need toaccount for these indirect effects to avoid biased estimatesfor both WOM and traditional marketing effects. Third, wequantify and contrast the immediate and long-term elas-ticities of WOM and traditional marketing actions. In par-ticular, we document strong carryover effects for WOM inour data. Finally, we attach an estimated monetary value toeach WOM referral, providing an upper bound to the finan-cial incentive management might consider offering for

WOM referrals. Indeed, the practice of seeding or stimulat-ing WOM has grown rapidly, but quantifying the effective-ness of this activity remains difficult (e.g., Godes and May-zlin 2004).

We organize the remainder of this article as follows: Webegin by summarizing previous research to help put ourcontributions in perspective. We then describe our modelingapproach. Next, we present our empirical analysis of thedata from a collaborating Internet social networking siteand offer implications for theory and managers. In particu-lar, we find that WOM referrals strongly affect new cus-tomer acquisitions and have significantly longer carryoverthan traditional forms of marketing used by the firm (21

days versus 3 to 7 days). We estimate a long-term elasticityfor WOM of .53approximately 2030 times higher thanthe elasticities for traditional marketing.

Research BackgroundTo help put the intended contribution of this study in con-text, we briefly review previous empirical research on theeffectiveness of WOM marketing. In Table 1, we present acomparison chart of selected prior work. As the tableshows, researchers have used a variety of means to capture,infer, or measure WOM. The table also outlines findings forthe effect of WOM on customer acquisition, comparisonswith traditional marketing, and incorporation of indirecteffects.

The earliest study on the effectiveness of WOM is sur-vey based (Katz and Lazarsfeld 1955) and was followed bymore than 70 marketing studies, most of them also inferringWOM from self-reports in surveys (Godes and Mayzlin2004; Money, Gilly, and Graham 1998). Researchers haveexamined the conditions under which consumers are likelyto rely on others opinions to make a purchase decision, themotivations for different people to spread the word about aproduct, and the variation in strength of peoples influenceon their peers in WOM communications. Moreover, cus-tomers who self-report being acquired through WOM addmore long-term value to the firm than customers acquired

through traditional marketing channels (Villanueva, Yoo,and Hanssens 2008).

Social contagion models (e.g., Coleman et al. 1966)offer an alternative perspective, typically inferring WOM/network effects from adoption behavior over time. Althougha social contagion interpretation of diffusion patterns isintuitively appealing, recent studies have pointed out thatsuch inferences can be due to misattribution. For example,when Van den Bulte and Lilien (2001) reestimated Colemanand colleagues (1966) social contagion model for physi-

Word-of-Mouth Versus Traditional Marketing / 3


4/24

cian adoption of tetracycline, they found that the contagioneffects disappeared when marketing actions were includedin the model. This raises the question whether WOM effectswould have been significant in the model had there beendata available on the actual transmission of informationfrom one physician to another.

Both of these research approaches do not observe actualWOM but infer it from self-reports or adoption. ExaminingWOM on the Internet can help address this limitation byoffering an easy way to track online interactions. Godes and

Mayzlin (2004) suggest that online conversations (e.g.,Usenet posts) can offer an easy and cost-effective way tomeasure WOM. In an application to new television shows,they link the volume and dispersion of conversations acrossdifferent Usenet groups to offline show ratings. Chevalierand Mayzlin (2006) use book reviews posted by customersat Amazon.com and Barnesandnoble.com online stores as aproxy for WOM. They find that though most reviews werepositive, an improvement in a books reviews led to anincrease in relative sales at the site, and the impact of anegative review was greater than the impact of a positiveone. In contrast, Liu (2006) shows that both negative andpositive WOM increase performance (box office revenue).

Although the foregoing three studies observe the post-ing of reviews (i.e., sending WOM), they do not directlyobserve the reception of WOM. In contrast, De Bruyn andLilien (2008) observe the reactions of 1100 recipients afterthey received an unsolicited e-mail invitation from one oftheir acquaintances to participate in a survey. They find thatthe characteristics of the social tie influenced recipientsbehaviors but had varied effects at different stages of thedecision-making process. They also report that tie strengthexclusively facilitated awareness, perceptual affinity trig-gered recipients interest, and demographic similarity had anegative influence on each stage of the decision-makingprocess. However, this study does not compare the effec-

tiveness of WOM with that of traditional marketing actions,nor does it quantify the monetary value of WOM to thecompany.

The current article differs from these studies in bothresearch objective and application. A key research objectivein this study is to compare the effects of observed WOMreferrals with those of traditional marketing efforts. Quanti-fying the full effects of WOM referrals and marketingrequires us (1) to account for the potential endogeneityamong these communication mechanisms and (2) toaccount for their potential permanent effects on customeracquisition. First, WOM may be endogenous because it notonly influences new customer acquisition but also is itselfaffected by the number of new customers. Likewise, tradi-

tional marketing activities may stimulate WOM; theyshould be credited for this indirect effect and the possibledirect effect on customer acquisition. Second, all thesecommunication mechanisms may have permanent effectson customer acquisition. For example, WOM may bepassed along beyond its originally intended audience andthus increase the sites potential to recruit sign-ups in thefuture.1 Network externalities can also imply that sign-upgains today may translate into higher sign-up gains tomor-row, even in the absence of marketing actions.



5/24

Internet Social Networking SitesIn the past few years, social networking sites have becomeextremely popular on the Internet. According to comScoreMedia Metrix (2006), every second Internet user in theUnited States has visited at least one of the top 15 socialnetworking sites. Approximately 50 social networking Websites each have more than one million registered users, andseveral dozen smaller, though significant, sites also exist(e.g., Wikipedia 2008). Compete.com (2008), a Web traffic

analysis company, reported that the largest online socialnetworking site (as of November 2008) was MySpace, with56 million unique visitors per month, closely followed byFacebook, with 49 million unique visitors.

Typical social networking sites allow a user to build andmaintain a network of friends for social or professionalinteraction. The core of a social networking site consists ofpersonalized user profiles. Individual profiles are usually acombination of users images (or avatars); lists of interestsand music, book, and movie preferences; and links to affili-ated profiles (friends). Different sites impose differentlevels of privacy in terms of what information is revealedthrough profile pages to nonaffiliated visitors and how farstrangers versus friends can traverse through the net-work of a profiles friends. Profile holders acquire newfriends by browsing and searching through the site andsending requests to be added as a friend. Other forms ofrelationship formation also exist.

In contrast to other Internet businesses, online commu-nities rely on user-generated content to retain users. Thus, acommunity member has a direct benefit from bringing inmore friends (e.g., by participating in the referral program)because each new member creates new content, which islikely to be of value to the inviting (referring) party. Typi-cally, sites facilitate referrals by offering users a convenientinterface for sending invitations to nonmembers to join.Figure 1 shows how a popular social networking site,

Friendster, implements the referral process.The social network setting offers an appealing context

in which to study WOM. The sites provide easy-to-use toolsfor current users to invite others to join the network. Theelectronic recording of these outbound referrals opens anew window into the effects of WOM, giving researchers anunobtrusive trace of this often difficult-to-study activity.When combined with data that also track new member sign-ups, it becomes possible to model the dynamic relationshipbetween this form of WOM and the addition of new mem-bers to the social networking site. In a real sense, thesemembers are also customers of the social networking siteinsofar as their exposure to advertising while using the site

produces revenue for the firm.Our empirical application to an Internet social network-ing site provides a set of substantive findings based onactual consumer WOM activity. In previously analyzed set-tings, such as movies and television shows, the Internetgives a partial view of interpersonal communication. Forsome product categories though, the vast majority of WOMmay transfer online. For online communities, the electronicform of spreading the word is the most natural one. This



6/24

leads us to suggest that online WOM is a good proxy foroverall WOM in our Internet social network setting.

Modeling ApproachIn this section, we describe our approach to modeling theeffects of WOM and traditional marketing on new membersign-ups. First, we test for the potential endogeneity amongWOM, marketing, and new customer acquisition and for thepotential permanent effects of these communication mecha-

nisms. Second, we specify a vector autoregressive (VAR)model that accounts for endogeneity and the dynamicresponse and interactions between marketing variables andoutcomes (Dekimpe and Hanssens 1999). Next, we com-pare the in-sample and out-of-sample fit of the VAR modelwith several alternative models. Finally, we estimate theshort-term and long-term responses of sign-ups to WOMand traditional marketing actions and compute the corre-sponding elasticities. Although we describe our approach inthe context of WOM referrals and social networking, theprocedure should be applicable on a more general basis(e.g., when data on new customers, tracked WOM, andother marketing activity are available over time).

The first step in our approach is to test for the presenceof endogeneity among new sign-ups (our measure of cus-tomer acquisition), event marketing (directly paid for by thesocial networking site), media appearances (induced bypublic relations), and WOM referrals. As Figure 2 shows,we anticipate that WOM referrals lead to new sign-ups and(following the reverse arrow) that new sign-ups lead tomore WOM referrals and, thus, indirectly to more new sign-ups. We anticipate a similar pattern of causality for newsign-ups and traditional marketing activity. It is also likelythat traditional marketing will stimulate WOM referrals,leading to another indirect effect on new sign-ups. We alsoinclude lagged effects of traditional marketing, new sign-ups, and WOM referrals in the model (as the curved arrowsindicate).

The links represented in Figure 2 can be tested by inves-tigating which variables Granger-cause other variables(Granger 1969; Hanssens, Parsons, and Schultz 2001). Inessence, Granger causality implies that knowing the historyof a variable X helps explain a variable Y, beyond Ys ownhistory. This temporal causality is the closest proxy forcausality that can be gained from studying the time series ofthe variables (i.e., in the absence of manipulating causalityin controlled experiments). We perform a series of Granger-causality tests on each pair of key variables.2 If sign-upsindeed Granger-cause (some of) the marketing variables,we need to capture the complex interactions of Figure 2 in a

full dynamic system.Next, we test for the potential of permanent effects of

the communication mechanisms on new sign-ups. If any ofthese mechanisms expand the intended site audience orinduce network externalities, sign-up gains today implyhigher sign-up gains tomorrow. In this case, the time seriesfor sign-ups would be classified as evolving. The oppositeclassification, that of stationary, implies that sign-upshave a fixed mean and that changes (including those causedby marketing actions) do not have a permanent impact (e.g.,



7/24

Dekimpe and Hanssens 1995, 1999). (The Appendix pro-vides details on testing for evolution versus stationarity ofnew sign-ups.)

Tests for both endogeneity and evolution enable us tospecify the VAR model in Equation 1. In this model, newsign-ups and marketing actions are endogenousthat is,they are explained by their own past and the past of theother endogenous variables (Dekimpe and Hanssens 1999).Specifically, the vector of endogenous variablessign-ups(Y), WOM referrals (X), media appearances (M), and pro-

motional events (E)is related to its own past, whichallows for complex dynamic interactions among thesevariables. The vectors of exogenous variables include foreach endogenous variable (1) an intercept, C; (2) adeterministic-trend variable, T, to capture the impact ofomitted, gradually changing variables; (3) indicators fordays of the week, D; and (4) seasonal (e.g., holidays)dummy variables, H (Pauwels and Dans 2001). Instanta-neous effects are captured by the variancecovariancematrix of the residuals, . The VAR specification is givenby

where t indexes days, J equals the number of lags included

(to be determined on the basis of the Akaike informationcriterion), D is the vector of day-of-week dummies, and tare white-noise disturbances distributed as N(0, ). Theparameters , , , and are to be estimated. Because VARmodel parameters are not interpretable on their own (Sims1980), effect sizes and significance are determined throughthe analysis of impulse response functions (IRFs) and elas-ticities computed on the basis of the model (for details, seethe Appendix).

The last step of the approach uses the model to estimatethe short-term and long-term elasticities of customer acqui-sition to WOM and traditional marketing actions. Fromthese results, several managerial implications can be drawn.We also show how the model can be used to estimate thepotential monetary value of each WOM referral (describedfurther in the Appendix).

Empirical Analysis

Data Description

We applied our model to data from one of the major socialnetworking sites, which prefers to remain anonymous. Thedata set contains 36 weeks of the daily number of sign-ups

( )1

Y

X

M

E

C

C

C

C

t

t

t

t

Y

X

M

E

=

+

+

Y

X

M

E

Y

X

M

E

T

+

+

H D

Y

X

M

E

j

11 122 13 14

21 22 23 24

31 32 33 34

41

j j j

j j j j

j j j j

j

42 43 44

1

j j j

j

J

t j

t

Y

X

=

+j

t j

t j

Y t

X t

M tM

E

,

,

,

E t,

,



8/24

and referrals (provided to us by the company) along withmarketing events and media activity (obtained from third-party sources). The data cover the period in 2005 from Feb-ruary 1 to October 16. Table 2 provides descriptive statisticsfor the variables.

During the observation period, the daily sign-ups andWOM referrals showed a positive trend. We observed some-what lower activity in referrals over the summer season(June 20September 5 [Labor Day, as observed in theUnited States]). Over the 36 weeks, the company organized

or cosponsored 101 promotion events. On some days, mul-tiple events occurred in different locations. Overall, 86 daysin the observation period had some promotion event activ-ity. Finally, we identified 236 appearances (on 127 days) ofthe company name in the media. We considered 102 differ-ent sources, both electronic and traditional media, as pro-vided by Factiva News and Business Information Service(www.factiva.com). We did not use the content of thesepublications, which makes our measure of media activitysomewhat coarse. More generally, it could be important toaccount for the valence of the message (e.g., Godes andMayzlin [2004] report for television shows). Given the rela-tively young age of the company, we had no reason to

believe that a significant share of the media reports wouldhave a negative tone. We removed a few negative suspectsfrom the sample, as judged by the title of the publication,but found no significant impact on the results. In summary,the number of media appearances is a useful measure forour research purpose.

Test Results for Evolution and Endogeneity

We begin our empirical analysis by first testing for station-arity versus evolution in each time series. The unit-root testsindicated trend stationarity in all series (i.e., all seriesappeared to be stationary after we controlled for a determin-istic trend). This indicates that model estimations can be

performed with all variables in levels (as we depict inEquation 1).To investigate whether endogeneity is present among

the variables, we conducted Granger-causality tests. Wesummarize the results in Table 3. Each cell gives the mini-mump-value obtained from the causality tests as conductedfrom 1 lag to 20 lags.

The results clearly show that endogeneity is presentamong the variables in our data. As we expected, Grangercausality is detected for WOM referrals, media, and eventson sign-ups (the direct effects). In addition, Granger causal-ity is found for many of the other pairings. For example,sign-ups Granger-cause WOM referrals (the interdependenteffect we noted previously), events (indicating managementperformance feedback; see Dekimpe and Hanssens 1999),and media (indicating that spikes in sign-ups may receivemedia attention). Moreover, events Granger-cause media(indicating that media covers events), and media Granger-cause events (indicating that management may try to timeevents to match pending media coverage). However, WOMreferrals do not Granger-cause events or media appearances(because the media do not observe referrals directly), andmedia appearances do not Granger-cause WOM. In sum-mary, the results from the Granger-causality tests indicate



9/24

the need to consider the full dynamic system, as in a VARmodel, and to account for the indirect effects of marketingactions.

Model Estimation

We estimated the VAR model of Equation 1 with two lags(the optimal lag length selected by the Akaike informationcriterion) and found good model fit (e.g., R2 = .89). We alsoassessed the significance of including WOM and the tradi-tional marketing variables for events and media in the

model at the same time. This is intended to address Van denBulte and Liliens (2001) concern about the significance ofWOM when marketing variables are incorporated. In par-ticular, we estimated two VAR models nested in Equation 1.In the first, we estimated the model without the WOMvariable. In the second, we estimated the model withoutevents and media. In both cases, the full model providedsuperior fit in-sample and out-of-sample based on both rootmean square error (RMSE) and mean absolute deviation(MAD). Thus, WOM and traditional marketing remain sig-nificant contributors to model fit.

To illustrate the ability of the VAR system model to rep-resent the data, we plot predicted versus actual values of

daily sign-ups and display this in Figure 3. The predictedvalues (labeled VAR) closely track the actual number ofsign-ups.

Short-Term and Long-Term Effects for WOM andTraditional Marketing Actions

To gauge the impact of WOM and the two traditional mar-keting variables on new sign-ups over time, we computeIRFs on the basis of the estimated VAR system parameters(for details, see the Appendix). The IRFs trace the incre-mental effect of a one-standard-deviation shock in WOM,events, and media on the future values of sign-ups. Theseenable us to examine the carryover effects of each activity

on sign-ups while fully accounting for the indirect effects ofthese activities in a dynamic system. Figure 4 plots the threeIRFs for the effect of WOM referrals, media, and events onnew sign-ups over time.

The top panel in Figure 4 shows that the WOM effect onsign-ups remains significantly different from zero forapproximately three weeks. In contrast, the effects of mediaand events (the middle and bottom panels of Figure 4) losesignificance within just a few days. Compared with tradi-tional marketing activities, the WOM referrals induce botha larger short-term response and a substantially longercarryover effect. These results highlight the need forresearchers to employ models that can also account for thelonger-term effects of WOM marketing.

We also calculate several short-term elasticities and along-term elasticity for WOM, events, and media. (Detailson the computations appear in the Appendix.) Table 4 pre-sents these estimated short- (one day, three days, and sevendays) and long-term elasticities.

The immediate (one day) elasticity of WOM (.068) is8.5 times higher than that of traditional marketing actions(.008). Moreover, this discrepancy grows over time. Indeed,the long-term elasticity indicates that WOM referrals are



10/24

akin to the gift that keeps on giving, especially comparedwith the performance of traditional marketing activities.Table 4 shows that the long-term elasticity of WOM refer-rals (.53) is approximately 20 times higher than the elastic-ity for marketing events (.53 versus .026) and 30 timeshigher than the elasticity for media appearances (.53 versus.017). The estimated WOM elasticity of .53 substantiallyexceeds the range of values for advertising elasticitiesreported in the literature (e.g., Hanssens, Parsons, andSchultz 2001). Thus, our empirical findings support the

notion that WOM may be among the most effective market-ing communication strategies.

Comparison with Alternative Models

We compare the in-sample and out-of-sample fit of the VARmodel with five alternative models (for details and specifi-cations of each model [Equations A1A5], see the Appen-dix).3 First, an autoregressive (AR) model captures theimmediate effects of marketing actions on sign-ups and thedynamic effect of past sign-ups on current sign-ups. Sec-ond, an autoregressive-distributed lag (ARDL) model addsdynamic effect of each communication mechanisms, with-out accounting for endogeneity and indirect effects among

these mechanisms. Third, social contagion models (e.g., theBass diffusion model) allow for several dynamic processesin the growth of online communities. We estimate versionsboth with and without marketing covariates. Finally, theeffects of WOM and marketing variables could vary overthe period of our data. Thus, a fifth model incorporatestime-varying coefficients in a state-space specification.Table 5 compares the in-sample and out-of-sample fit ofeach model in terms of RMSE and MAD. Both in-sampleand out-of-sample, the VAR model produces the lowest val-ues for RMSE and MAD, indicating superior fit to the data.

Implications for Marketing Theory

Our investigation of WOM in an Internet social networksetting offers several potential contributions for advancingthe fields knowledge and understanding of WOM effects.First, the electronic tracking of WOM, along with newmember sign-ups, enables us to provide a concrete andmeasurable link between observed WOM activity and cus-tomer acquisition. This bodes well for future efforts toquantify WOM effects and to incorporate them into theplanning of marketing communication strategies.

A second set of theoretical implications comes from thedynamic performance of WOM marketing. Note that thestrong carryover effects of WOM are the key factor drivingits high long-term elasticity in our data. Notably, when esti-mated from standard regression models (the AR and ARDLmodels in the Appendix), the direct WOM elasticity is closeto the average advertising elasticity of .10 to .20 (e.g.,Hanssens, Parsons, and Schultz 2001). Accounting forcarryover by the VAR system model produces a long-termelasticity that is several times higher. Our results providefurther impetus for models of advertising to go beyondinvestigating the direct effects of advertising to incorporatea variety of potential indirect benefits, such as increasing



11/24

retailer support (e.g., Reibstein and Farris 1995) andinvestor awareness (e.g., Joshi and Hanssens 2006).

Third, our findings highlight the importance of account-ing for the indirect effects among WOM, traditional market-ing actions, and performance. Conceptually, we illustratethis in Figure 2. Empirically, the Granger-causality testsindicated the presence of significant indirect and feedbackeffects in our data. Taken together with the superior fit pro-vided by the VAR model compared with other approaches,the findings demonstrate the importance of handling the

endogeneity that is likely to be present with WOM. Ourframework and model also enable traditional marketingactions to be credited for their WOM-generating abilities.To the extent that modeling WOM alters estimated effectsizes for traditional marketing activities, the productivity ofthose activities could change, with potentially importantimplications for budgeting and resource allocationdecisions.4

Though still a relatively new phenomenon, online socialnetworking sites have begun to attract the attention of mar-keting scholars. For example, Ansari, Koenigsberg, andStahl (2008) have developed an approach for modeling mul-tiple relationships of different types among users of a social

networking site. Dholakia, Bagozzi, and Pearo (2004) studytwo key group-level determinants of virtual community par-ticipation: group norms and social identity. Kozinets (2002)has developed a new approach to collecting and interpretingdata obtained from consumers discussions in onlineforums. Trusov, Bodapati, and Bucklin (2008) propose amodel that enables managers to determine which users arelikely to be influential and, therefore, important to the busi-ness for their role in attracting others to the site. Ourresearch contributes to this emerging area of study by eluci-dating the role of WOM in the recruiting of new membersand the growth of these social networks.

Managerial Implications

Managers should be able to make use of our research inseveral ways. These include obtaining improved metrics forboth WOM and traditional marketing, testing changes inonline WOM referral content, and evaluating the extent towhich financial incentives can be used to stimulate WOM.

A practical benefit from applying the proposed model islikely to come from better measures of the effects of bothWOM and traditional marketing. As we noted previously,the VAR model incorporates the potential endogeneitylikely to be present with WOM, incorporates indirect effectsamong marketing variables and WOM, and allows eachvariable (WOM and traditional marketing) to have differ-entand potentially long-lastingcarryover effects. Thegeneral result is a potentially different estimate of the returnon the firms investments in traditional marketing.

To examine this, we compared the elasticity estimatesfor events and media obtained from the VAR model (Table4) with those from the best-fitting standard regression (theARDL model). The results from the standard regressionmodel show neither events nor media registering significantshort-term or long-term effects. Specifically, to three deci-mal places, we estimate the elasticities for events and mediaat .002 and .000, respectively. In the VAR model, however,



12/24

these effects, though still relatively small, are larger and sig-nificant (see Table 4).

Managers can also use our approach to test changes toreferral content. When an existing member wants to refer afriend to join the site (the outbound referrals we study), themember completes a preformatted message, which is thensent out by e-mail. At many social networking sites (includ-ing the major ones, such as Facebook and MySpace), mem-bers have relatively little control over the content andformat of this invitation. Although there is room for a short

text message, many of these are left blank. The upshot isthat the firm is largely responsible for the appearance andcontent of the referral message. This provides the opportu-nity to test and refine the message. The firm could alsoconsider targeting different messages according to what isknown about the referrer (e.g., demographics, usage his-tory). Our modeling approach should be useful to managersin these and analogous settings as a means to gauge theeffect of different content executions while controlling forother marketing activity. For example, this could be done byestimating the model on the data for different test periods ortarget samples.

Finally, our approach may help managers answer the

question regarding the extent to which the firm should beprepared to invest in stimulating additional WOM. A grow-ing practice in both offline and online markets is to offerfinancial incentives to existing customers to provide WOMreferrals (e.g., Netflix has incentivized current subscribersto recruit new ones). Authors of several studies in the aca-demic WOM literature have also suggested that companiesshould try to create WOM communication (e.g., Godes andMayzlin 2004; Liu 2006; Rosen 2000). Important input forsuch a referral program would be the value WOM commu-nication provides to the firm. This can be obtained from ourdata and model through simulation, and we describe oneapproach in the Appendix. Using some assumptions about

current advertising rates, the simulation shows that eachoutbound referral sent out is worth approximately $.75 peryear in increased advertising revenue. By sending out tenreferrals, each network member could bring in approxi-mately $7.50 of incremental ad revenue for the site.

An important caveat to the simulation is that it does notaccount for the possible difference between the quality ofWOM that occurs naturally and WOM stimulated by thefirm. When a company stimulates WOM activity, it is nolonger organic WOM. Indeed, it could be called fertil-ized WOM. We do not know whether fertilized WOMwould produce the same elasticity as the organic WOMobserved in our data. If the paid nature of WOM activity isknown to prospective members, fertilized WOM may be

substantially less effective than organic WOM. In thisrespect, our monetary value calculations represent an upperbound of the money that could be generated by stimulatingWOM.

Nonetheless, managers can use figures derived from thistype of approach as an upper bound when contemplatingthe offer of financial incentives for WOM. For example, ifthe firm cannot effectively generate additional referrals atless than $.75 each, it should not pursue firm-stimulatedWOM programs.



13/24

ConclusionThe purpose of this study was to increase the understandingof the effects of WOM marketing by taking advantage ofnew, detailed tracking information made possible by theInternet. Using data from an online social networking site,we quantify the effect of WOM referrals, which arerecorded electronically, on new member sign-ups to the site(i.e., customer acquisitions). We also compare the effect ofWOM with traditional marketing activity and examine its

carryover dynamics. This permits us to speak to the relativeeffectiveness of WOM both in the short run and in the longrun.

We find that WOM referrals have a strong impact onnew customer acquisition. The long-term elasticity of sign-ups with respect to WOM is estimated to be .53 (substan-tially larger than the average advertising elasticitiesreported in the literature). The elasticity for WOM isapproximately 20 times higher than that for marketingevents and 30 times that of media appearances. In addition,an important feature of the proposed modeling approach isthe ability to handle the endogeneity and indirect effectsamong WOM, marketing activity, and customer acquisition.

Our work has several important implications for practic-ing managers. Our approach offers managers a tool toimprove the metrics they use for assessing the effectivenessof traditional marketing when WOM effects are present. Wealso conducted a simulation analysis to illustrate the poten-tial monetary implications from inducing additional WOMby offering financial incentives to existing customers.

We should also note the limitations to our research. Ourdata come from one large social networking site, cover aperiod of less than one year, and are at the aggregate level.Thus, data limitations prevent us from analyzing the effectsof WOM forand marketing actions bycompeting sites.Though unfortunate, this is common to this type ofcompany-specific data set. Because our data set tracks new

sign-ups and WOM at the aggregate level, another limita-tion is that we are unable to model heterogeneity. Therefore,we do not address segmentation in WOM response orpotential differences in wear-out or saturation effects at theindividual level. In addition, site members attracted in dif-ferent ways (i.e., through WOM, events, or media appear-ances) could differ in visit frequency and pages viewed,thus yielding different revenue benefits to the site (Vil-lanueva, Yoo, and Hanssens 2008). Because we lacked suchindividual-level data, we could not make this distinction inour revenue calculations.

Another data-related limitation to our work is that thestationary nature of this companys sign-up series means

that none of the changes to sign-ups (including those causedby WOM and other marketing actions) have a permanentimpact. This is consistent with marketing effect findings forlarge, established brands across industries (e.g., Nijs et al.2001; Srinivasan et al. 2004) but not with those for smallerbrands (Slotegraaf and Pauwels 2008). In light of ourdescription of how social networking sites start out (with acrucial mix of user-generated content and WOM offounders to friends), sign-up growth/evolution for smallsites may be driven by network externalities (more user-



14/24

generated content makes membership more valuable) andWOM transmission beyond the initial audience of thefounders friends. Moreover, smaller social networks maynot have sufficient funding to engage in traditional market-ing events and therefore must focus on cost-effectiveoptions, such as blogging and WOM campaigns (Whitman2006). Thus, our estimates of WOM elasticities for anestablished site are likely to be conservative for smallersites. In general, further research is needed to determinewhether our substantive findings generalize to other compa-

nies and settings. Our results are consistent with the spiritof the results that East and colleagues (2005) report; in areview of 23 service categories, they find that WOM hadgreater reported impact on brand choice than advertising orpersonal search.

Finally, our model is in a reduced as opposed to struc-tural form. This implies that the long-term impact calcula-tions are subject to the assumption that the basic data-generating process does not change. This is appropriate forinnovation accountingthat is, identifying and quantify-ing the effects of WOM and traditional marketing on sign-ups in the data sample (Franses 2005; Van Heerde,Dekimpe, and Putsis 2005). However, the modeling

approach is not suited for revealing structural aspects ofsubscriber and company behavior.

In summary, our goal was to shed new light on the trueeffectiveness of WOM marketinga vital but still poorlyunderstood element of the marketing communications mix(Misner 1999). We do this by quantifying the dynamic per-formance effects of WOM in an Internet social network set-tinga setting that offers researchers and managers a directway to observe this phenomenon. Thus, we are among thefirst to link WOM to new customer acquisition. Our elastic-ity results add further evidence to the view that WOM com-munication is a critical factor for firms seeking to acquirenew customers and that WOM can have larger and longer-

lasting effects than traditional marketing activity.

Appendix

Testing for Evolution or Stationarity: Unit-RootTests

We perform unit-root tests to determine whether each of thevariables in our data set is stable (i.e., fluctuate temporarilyaround a fixed mean or trend) versus evolving (i.e., candeviate permanently from previous levels). We use both theaugmented DickeyFuller test procedure that Enders (1995)recommends and Kwiatkowski and colleagues (1992) test.

AR and ARDL ModelsThe AR specification in Equation A1 relates sign-ups toevents, media, and WOM and controls for deterministiccomponents, such as a base level (constant), a deterministic(time) trend, seasonality, and lags of the dependent variable(e.g., Box and Jenkins 1970):



15/24

Equation A1 includes the same variables as Equation 1, andJ is the number of lags of the dependent variable (thus, theterm autoregressive) needed to ensure that the residuals tare white-noise errors (no residual autocorrelation). We alsocheck whether the interaction effects were significant, butwe did not find any.

Equation A1 directly models only the immediate effectsof marketing actions on sign-ups. Although this immediateeffect may carry over to future periods through the ARparameters, the model assumes that such carryover does notdepend on the marketing action that caused the immediategain in sign-ups. To include dynamic effects specific to themarketing action, we can add lags of the marketingvariables, obtaining the following ARDL model (e.g.,Hanssens, Parsons, and Schultz 2001):5

where L, M, and N are the number of lags for the predictorvariables WOM referrals, number of media appearances,and number of promotional events variables, respectively.

Social Contagion Model (Bass Diffusion Model)

Social contagion models (e.g., the Bass diffusion model)allow for several dynamic processes in the growth of online

communities and have recently been applied to this setting(e.g., Firth, Lawrence, and Clouse 2006). An importantissue is that stable and robust parameter estimates for theBass model are obtained only if the data include the peakfor the noncumulative adoption curve, which is difficult toassess a priori (Heeler and Hustad 1980; Srinivasan andMason 1986). To address this problem, we assist the Bassmodel by using information about the number of site mem-bers observed in the time since the detailed original datawere collected (specifically, two years later). We alsoensure that the selected value for m (market capacity) pro-vides the best possible model fit to the in-sample data. TheBass model and extended Bass model are both estimatedwith nonlinear least squares (e.g., Srinivasan and Mason1986).

Our specification for the social contagion model followsthe Bass diffusion model but also includes indicators forday of week and the summer holiday:

where

( ) ( ) ( ) ,A Y m Z p qZ

md Ht t

ti i

i

t3 11

1

6

= + + +

=

( ) , , ,A Y X Mt t

L

m t m

m

M

n2 11

2

1

3= + +=

=

l l

lEE

C T d H Y

t n

n

N

i i

i

t j t j t

=

=

+ + + + + +

1

1

6

jj

J

=

1

,

( )A Y X M E C T d H

Y

t t t t i i

i

t

j t

1 1 2 31

6

= + + + + + +

+

=

=

+ j tj

J

1

.



16/24

t = day index,m = number of eventual members (market capacity),p = coefficient of innovation (external influence),q = coefficient of imitation (internal influence

[WOM]),Yt = number of sign-ups (new members),Zt = cumulative number of adopters (new members) at

time t,di = indicators for days of the week (using Friday as

the benchmark),

H = holiday indicator (summer vacation), and 16 areparameters to be estimated.

We extend this social contagion model by augmentingEquation A3 with the marketing variables for media appear-ances and promotional events. This gives the specificationfor the social contagion model with marketing covariates:

where

Mt = number of media appearances,Et = number of promotional events, and 12, 16, and

are parameters to be estimated.

Varying Coefficients Model

If the effects of WOM and marketing variables could varyover time, a logical alternative to the VAR model is a state-space specification with time-varying coefficients (e.g.,Dekimpe et al. 2006; Kao and Allenby 2005). This giveswhere the coefficients for WOM and marketing variablesbecome time varying:

where the coefficients for WOM and marketing variablesbecome time varying and and are parame-ters to be estimated. We estimate this model using Kalmanfilter recursions (Harvey 1991).

In-Sample and Out-of-Sample Fit Analysis forModel Comparison

For the six models in Equations 1 and A1A5, we performan in-sample and out-of-sample fit analysis. We use the sta-tic forecasting technique for the out-of-sample analysis,treating both exogenous and any lagged endogenousvariables as observed for every observation in the holdout.This one-step-ahead forecasting procedure is common inmodel comparison because it makes use of the same infor-mation set across competing models. We reserved the lasttwo months of the data (61 observations) for a holdout testand reestimated all models on the first 197 daily observa-tions (approximately 6.5 months). For each model, we com-puted two fit statisticsRMSE and MADfor both the

1 3 1 6 t , , ,

( )A Y X M E C T d Htt

tt

tt

t i i

i

t5 1 2 31

6= + + + + + + +

= tt ,

( ) ( )A Y m Z p qZ

mM

E

t tt

t

t i

4 11

1

2

= + +

+

+

dd Hii

t

= +

1

6

,



17/24


18/24

approximately $7.50 to the firm. Practitioners should viewthese results with caution because the measures used mayvary. In addition, other online advertising models, such aspay-per-click, pay-per-lead, and pay-per-sale could be ana-lyzed by substituting appropriate conversion rates.

REFERENCESAnsari, Asim, Oded Koenigsberg, and Florian Stahl (2008), Mod-

eling Multiple Relationships in Online Social Networks,

working paper, Columbia Business School, ColumbiaUniversity.Box, George E.P. and Gwilym M. Jenkins (1970), Time Series

Analysis: Forecasting and Control. San Francisco: Holden-Day.

Chevalier, Judith and Dina Mayzlin (2006), The Effect of Wordof Mouth on Sales: Online Book Reviews,Journal of Market-ing Research, 43 (August), 34554.

Coleman, James S., Elihu Katz, and Herbert Menzel (1966),Medi-cal Innovation: A Diffusion Study. Indianapolis: Bobbs-Merrill.

Compete.com (2008), Site Analytics. Profile: MySpace.com andFacebook.com, (accessed December 22, 2008), [available athttp://siteanalytics.compete.com/facebook.com+myspace.com/?metric=uv].

comScore Media Metrix (2006), Total Number of Unique Visi-tors to Selected Social Networking Sites, as of March 2006,(accessed May 26, 2009), [available at http://www.emergencemarketing.com/2006/05/05/social-networking-sites-draw-tremendous-traffic/].

De Bruyn, Arnaud and Gary L. Lilien (2008), A Multi-StageModel of Word of Mouth Through Viral Marketing, Inter-national Journal of Research in Marketing, 25 (3), 143225.

Dekimpe, Marnik G., Philip Hans Franses, Dominique M.Hanssens, and Prasad A. Naik (2006), Time-Series Models inMarketing, ERIM Report Series Research in Management,Erasmus Research Institute of Management.

and Dominique M. Hanssens (1995), The Persistence ofMarketing Effects on Sales,Marketing Science, 14 (1), 121.

and (1999), Sustained Spending and PersistentResponse: A New Look at Long-Term Marketing Profitability,

Journal of Marketing Research, 36 (November), 397412.

Dholakia, Utpal M., Richard P. Bagozzi, and Lisa Klein Pearo(2004), A Social Influence Model of Consumer Participationin Network- and Small-Group-Based Virtual Communities,

International Journal of Research in Marketing, 21 (3),24163.

East, Robert, Kathy Hammond, Wendy Lomax, and Helen Robin-son (2005), What Is the Effect of a Recommendation?Mar-keting Review, 5 (2), 14557.

Enders, Walter (1995), Applied Econometric Time Series. NewYork: John Wiley & Sons.

Firth, David R., Cameron Lawrence, and Shawn F. Clouse (2006),Predicting Internet-Based Online Community Size and Timeto Peak Membership Using the Bass Model of New ProductGrowth,Interdisciplinary Journal of Information, Knowledge,and Management, 1 (1), 112.

Franses, Philip Hans (2005), On the Use of Econometric Modelsfor Policy Simulation in Marketing, Journal of MarketingResearch, 42 (February), 414.

Godes, David and Dina Mayzlin (2004), Using Online Conversa-tions to Study Word-of-Mouth Communication, MarketingScience, 23 (4), 54560.

, , Yubo Chen, Sanjiv Das, Chrysanthos Dellarocas,Bruce Pfeiffer, et al. (2005), The Firms Management ofSocial Interactions,Marketing Letters, 16 (34), 41528.

Granger, Clive W.J. (1969), Investigating Causal Relations byEconometric Models and Cross-Spectral Methods,Economet-rica, 37 (3), 42438.



19/24

Hanssens, Dominique M. (1980), Market Response, CompetitiveBehavior, and Time Series Analysis, Journal of Marketing

Research, 17 (November), 47085., Leonard J. Parsons, and Randall L. Schultz (2001), Mar-

ket Response Models: Econometric and Time-Series Research,2d ed. Boston: Kluwer Academic Publishers.

Harvey, Andrew C. (1991), Forecasting, Structural Time SeriesModels and the Kalman Filter. Cambridge, UK: CambridgeUniversity Press.

Heeler, Roger M. and Thomas P. Hustad (1980), Problems in Pre-dicting New Product Growth for Consumer Durables, Man-

agement Science, 26 (October), 10071020.Joshi, Amit and Dominique M. Hanssens (2006), AdvertisingSpending and Market Capitalization, Marketing Science Insti-tute Report No. 04-110.

Kao, Ling-Jing and Greg M. Allenby (2005), Estimating State-Space Models of Economic Behavior: A Hierarchical BayesApproach, working paper, Max M. Fisher College of Busi-ness, Ohio State University.

Katz, Elihu and Paul F. Lazarsfeld (1955), Personal Influence; ThePart Played by People in the Flow of Mass Communications.Glencoe, IL: The Free Press.

Kozinets, Robert V. (2002), The Field Behind the Screen: UsingNetnography for Marketing Research in Online Communities,

Journal of Marketing Research, 39 (February), 6172.Kwiatkowski, Denis, Peter C.B. Phillips, Peter Schmidt, and

Yongcheol Shin (1992), Testing the Null Hypothesis of Sta-tionarity Against the Alternative of a Unit Root: How Sure AreWe that Economic Time Series Have a Unit Root? Journal of

Econometrics, 54 (13), 15978.Liu, Yong (2006), Word of Mouth for Movies: Its Dynamics and

Impact on Box Office Revenue, Journal of Marketing, 70(July), 7489.

Misner, Ivan R. (1999), The Worlds Best Known MarketingSecret: Building Your Business with Word-of-Mouth Marketing,2d ed. Austin: Bard Press.

Money, Bruce R., Mary C. Gilly, and John L. Graham (1998),Explorations of National Culture and Word-of-Mouth ReferralBehavior in the Purchase of Industrial Services in the UnitedStates and Japan,Journal of Marketing, 62 (October), 7687.

Nail, Jim (2005), Whats the Buzz on Word-of-Mouth Market-ing? Social Computing and Consumer Control Put Momentum

into Viral Marketing, (May 3), (accessed May 11, 2009),[available at http://www.forrester.com/Research/Document/Excerpt/0,7211,36916,00.html].

Nielsen//NetRatings (2005), A Community Uprising, (February6), (accessed May 11, 2009), [available at http://www.nielsen-online.com/pr/pr_060206_uk.pdf].

Nijs, Vincent, Marnik Dekimpe, Jan-Benedict E.M. Steenkamp,and Dominique Hanssens (2001), The Category-DemandEffects of Price Promotions,Marketing Science, 20 (1), 122.

Pauwels, Koen (2004), How Dynamic Consumer Response,Competitor Response, Company Support, and Company InertiaShape Long-Term Marketing Effectiveness, Marketing Sci-ence, 23 (4), 596610.

and Enrique Dans (2001), Internet Marketing the News:Leveraging Brand Equity from Marketplace to Marketspace,

Brand Management, 8 (4), 303314., Dominique Hanssens, and S. Siddarth (2002), TheLong-Term Effects of Price Promotions on Category Incidence,Brand Choice, and Purchase Quantity, Journal of Marketing

Research, 39 (November), 42139., Jorge Silva-Risso, Shuba Srinivasan, and Dominique

Hanssens (2004), New Products, Sales Promotions, and FirmValue: The Case of the Automobile Industry,Journal of Mar-keting, 68 (October), 14256.

Pesaran, M. Hashem, Richard G. Pierse, and Kevin C. Lee (1993),Persistence, Cointegration, and Aggregation: A Disaggregated



20/24

Analysis of Output Fluctuations in the U.S. Economy,Journalof Econometrics, 56 (March), 5788.

and Yongcheol Shin (1998), Generalized ImpulseResponse Analysis in Linear Multivariate Models,Economics

Letters, 58 (1), 1729.Reibstein, David J. and Paul W. Farris (1995), Market Share and

Distribution: A Generalization, a Speculation, and Some Impli-cations,Marketing Science, 14, (3, Part 2), 190202.

Rosen, Emmanuel (2000), The Anatomy of Buzz: How to CreateWord-of-Mouth Marketing. New York: Doubleday.

Sims, Christopher A. (1980), Macroeconomics and Reality,

Econometrica, 48 (1), 148. and Tao Zha (1999), Error Bands for ImpulseResponses,Econometrica, 67 (5), 111355.

Slotegraaf, Rebecca and Koen Pauwels (2008), The Impact ofBrand Equity and Innovation on the Long-Term Effectivenessof Promotions, Journal of Marketing Research, 45 (June),293306.

Srinivasan, Shuba, Koen Pauwels, Dominique M. Hanssens, andMarnik G. Dekimpe (2004), Do Promotions Benefit Manufac-turers, Retailers, or Both? Management Science, 50 (May),61729.

Srinivasan, V. and Charlotte H. Mason (1986), Nonlinear LeastSquares Estimation of New Product Diffusion Models, Mar-keting Science, 5, (2), 16978.

Stephen, Andrew T. and Donald R. Lehmann (2008), RecipientCharacteristics and Product-Related Word-of-Mouth Transmis-sion: The Role of Social Capital, working paper, ColumbiaBusiness School, Columbia University.

Trusov, Michael, Anand Bodapati, and Randolph E. Bucklin(2008), Determining Influential Users in Internet Social Net-works, working paper, Robert H. Smith School of Business,University of Maryland.

Van den Bulte, Christophe and Gary Lilien (2001), Medical Inno-vation Revisited: Social Contagion Versus Marketing Effort,

American Journal of Sociology, 106 (5), 14091435.Van Heerde, Harald J., Marnik. G. Dekimpe, and William P. Putsis

Jr. (2005), Marketing Models and the Lucas Critique, Jour-nal of Marketing Research, 42 (February), 1521.

Villanueva, Julian, Shijin Yoo, and Dominique M. Hanssens(2008), The Impact of Marketing-Induced Versus Word-of-Mouth Customer Acquisition on Customer Equity Growth,

Journal of Marketing Research, 45 (February), 4859.Whitman, Janet (2006), Lessons Learned: New Dot-Coms Seek

Attention Cheaply, The Wall Street Journal, (February 13),B3A.

Wikipedia (2008), List of Social Networking Websites,(accessed August 28, 2008), [available at http://en.wikipedia.org/wiki/List_of_social_networking_websites].



21/24

1We thank an anonymous reviewer for this insight.2A wrong choice for the number of lags in the test may erro-

neously conclude the absence of Granger causality (e.g., Hanssens1980). Because we are applying these tests to investigate the needfor modeling a full dynamic system, we are not interested inwhether variable X causes variable Y at a specific lag but rather inwhether we can rule out that X Granger-causes Y at any lag.Therefore, we run the causality tests for lags up to 20 and reportthe results for the lag that has the highest significance for Grangercausality.

3We thank the guest editor and reviewers for suggesting alterna-

tive models and the comparison tests.4Van den Bulte and Lilien (2001) raise the question whetherWOM effects are significant after marketing actions are accountedfor in a contagion model. To investigate this further, we estimatedtwo additional VAR models. In the first, we dropped the traditionalmarketing variables, media and events, leaving all else unchanged.In the second, we dropped WOM. In both cases, we found nomaterial change in the estimated elasticity values for WOM ormedia and events. Although the Granger-causality tests indicatedthat indirect effects were present, at least for our data, the ultimateelasticities we obtained did not substantially differ. Part of thismay also be due to a greater robustness of VAR model estimates toomitted variables.

5Our use of the term dynamic refers to the carryover effectsthat a change in one variable has on other variables over time. Thisdiffers from a model in which parameters themselves change overtime, as we detail subsequently.

6In our data set, the ratio of daily average number of sign-ups tothe daily average number of WOM referrals is close to one (Table2). Accordingly, the estimation of long-term marginal effect ofWOM (.52) appears to be not significantly different from the esti-mation of WOM elasticity (.53). Additional details on the simula-tion and estimation of alternative models are available on request.



22/24


WOM InferenceEffect of WOM on

Customer AcquisitionComparison with

Traditional Marketing

Indirect Effectsof WOM and/or

Traditional Marketing

Initial (Katz andLazarsfeld 1955) andmost studies

Self-report fromsurveys

Inferred from self-reports on relative

influence

WOM two times moreeffective than radio

advertisements, fourtimes more than

personal selling,seven times morethan print

advertisements

Not analyzed

Customer lifetime value(Villanueva, Yoo, andHanssens 2008)

Inferred from newcustomers self-report

Not analyzed Customers acquiredthrough WOM add

two times the lifetimevalue of customersacquired through

traditional marketing

Customers acquiredthrough WOM spreadmore WOM and bringin twice as many new

customers

Social contagion(Coleman, Katz, andMenzel 1966; Vanden Bulte and Lilien2001)

Inferred from adoption Difficult to attributeobserved contagion

to WOM versustraditional marketing

Contagion effectsdisappear when

traditional marketingeffects are included in

the model

Not analyzed

Determinants of WOMtransmissions(Stephen andLehmann 2008)

Directly: personsdecision to transmit

WOM in experimentalsetting

Expectedresponsiveness of

WOM recipient drivestransmitters decision

to pass WOM

None Not analyzed

Valence of online WOM(Chevalier andMayzlin 2006; Liu2006)

Inferred from Website posts and

reviews

Higher number ofreviews leads to

higher relative sales

None Not analyzed

Impact of social ties onWOM effect (De

Bruyn and Lilien2008)

Directly: thoughe-mails sent

Social tie effectsdepend on stage of

decision making

None Not analyzed

This article Directly: throughreferrals sent

Quantify direct andindirect effects of

WOM and marketing

Compare immediateand carryover effects

of WOM andtraditional marketing

Demonstrate indirecteffects and quantify

total monetary value ofWOM activity

TABLE 1

Comparison of Empirical Studies on the Effectiveness of WOM

TABLE 2Descriptive Statistics (Daily Data)

Maxi- Mini-M Mdn mum mum SD

Sign-ups 11.36 11.30 11.89 10.86 .29WOM referralsa 11.37 11.42 12.09 10.53 .38Media .92 0 8 0 1.34Events .39 0 4 0 .64

aThe figures reflect a linear transformation of the original daily datato preserve the anonymity of the collaborating site. Actual datawere used in the econometric analyses.


23/24


TABLE 5

In-Sample and Out-of-Sample Performance for Alternative Models

Base Bass with VaryingVAR AR ARDL Bass Covariates Coefficients

In-Sample RMSE .687 .731 .714 .784 .747 .692MAD .457 .525 .521 .544 .529 .472

Out-of-Sample RMSE 1.302 1.630 1.454 1.563 1.525 1.681

MAD 1.051 1.384 1.182 1.217 1.207 1.427

TABLE 3

Results of the Granger Causality Tests (Minimump-Values Across 20 Lags)

DependentVariable IsGranger- Sign- WOMCaused by Ups Referrals Media Events

Sign-ups .02a .00 .00WOM referrals .00 .22 .08

Media .00 .58 .02Events .02 .00 .01

aWOM referrals are Granger-caused by sign-ups at the .02 signifi-cance level.

TABLE 4Short-Term Versus Long-Term Elasticity of Sign-

Ups with Respect to WOM Referrals andMarketing Activities

One Three Seven Long-Day Days Days Term

WOM referrals .068 .171 .330 .532Media .008 .017 .017 .017Events .008 .022 .026 .026

FIGURE 1Referrals Process at Friendster.com


24/24

FIGURE 4

IRFs: Response of Sign-Ups to Shock inReferrals, Media, and Promotional Events

FIGURE 2

Modeling Framework

WOMReferrals

TraditionalMarketing

NewSign-Ups

FIGURE 3Model Fit:Tracking Plot

.00

2.00

4.00

6.00

8.00

10.00

12.00

14.00

16.00

2/3/20

05

ActualVAR

10/13/20

05

9/15/2

005

8/18/2

005

7/21

/2005

6/23

/2005

5/26

/2005

4/28

/2005

3/31

/2005

3/3/20

05

Ne

w

Sign-Ups

(Linear

Transformation)

Notes The y-axis values reflect a linear transformation used to dis-guise the identity of the source.

.5

0

.5

1

1.5

2

2.5

3

1 3 5 7 9 11 13 15 17 19 21 23 25 27 29

1 3 5 7 9 11 13 15 17 19 21 23 25 27 29

1 3 5 7 9 11 13 15 17 19 21 23 25 27 29

Response to shock in referrals

Days

New

Sign-Ups

New

Sign-Ups

New

Sign

-Ups

.5

0

.5

1

1.5

2Response to shock in media

Days

.5

0

.5

1

1.5

2Response to shock in events

Days

effects of wom vs traditional marketing

Documents