Opinion Leadership and Social Contagion in New Product Diffusion
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
CELEBRATING 30 YEARS
Vol. 30, No. 2, March–April 2011, pp. 195–212 doi 10.1287/mksc.1100.0566
issn 0732-2399 eissn 1526-548X 11 3002 0195 © 2011 INFORMS
Opinion Leadership and Social Contagion in
New Product Diffusion
Raghuram Iyengar, Christophe Van den Bulte
The Wharton School of the University of Pennsylvania, Philadelphia, Pennsylvania 19104
{riyengar@wharton.upenn.edu, vdbulte@wharton.upenn.edu}
Thomas W. Valente
Keck School of Medicine, University of Southern California, Los Angeles, California 90089,
tvalente@usc.edu
W e study how opinion leadership and social contagion within social networks affect the adoption of a new
product. In contrast to earlier studies, we find evidence of contagion operating over network ties, even
after controlling for marketing effort and arbitrary systemwide changes. More importantly, we also find that the
amount of contagion is moderated by both the recipients’ perception of their opinion leadership and the sources’
volume of product usage. The other key finding is that sociometric and self-reported measures of leadership are
weakly correlated and associated with different kinds of adoption-related behaviors, which suggests that they
probably capture different constructs. We discuss the implications of these novel findings for diffusion theory
and research and for marketing practice.
Key words: diffusion of innovations; opinion leadership; social contagion; social networks
History: Received: March 6, 2009; accepted: January 25, 2010; Eitan Muller served as the guest editor-in-chief
and Arvind Rangaswamy served as associate editor for this article. Published online in Articles in Advance
December 10, 2010.
1. Introduction users are more influential than light users, an issue of
Marketers are increasingly experimenting with vari- obvious relevance to the identification and targeting
ous forms of network marketing. In the area of new of likely influentials.
product marketing, the rationale for such strategies The present study addresses each of the three
rests on three key assumptions: (1) social contagion assumptions fundamental to many network market-
among customers is at work, (2) some customers’ ing practices. Specifically, we empirically assess three
adoptions and opinions have a disproportionate influ- questions on how social contagion and opinion lead-
ence on others’ adoptions, and (3) firms are able ership affect new product diffusion. First, to what
to identify and target those influentials or opinion extent do sociometric and self-reported opinion lead-
leaders. These assumptions are quite reasonable, as ership go hand in hand and have the same influ-
the first two are consistent with several sociologi- ence on the time of adoption? Second, is there social
cal and marketing theories, and all three have been contagion operating over social ties such that bet-
supported in at least some studies (e.g., Godes and ter connected adopters exert more influence than less
Mayzlin 2009, Goldenberg et al. 2006, Rogers 2003, connected ones, over and above the effect of mar-
Tucker 2008, Valente et al. 2003, Weimann 1994). keting efforts and systemwide influences that vary
However, managers would be remiss to simply take over time? Third, is contagion emanating from prior
those three assumptions for granted. For instance, adopters a function of how much they use the prod-
Van den Bulte and Lilien (1997, 2001) have shown uct rather than simply whether they have adopted it?
that contagion need not be as important as reported We investigate these questions by studying the
in prior studies, Becker (1970) and Watts and Dodds adoption of a new prescription drug by physicians.
(2007) have raised doubts on the importance of opin- Our study combines individual-level adoption data,
ion leaders in speeding up the acceptance of new demographic data, a measure of self-reported opinion
products, and Rogers and Cartano (1962) note dis- leadership, network data on discussion and patient
agreement on whether to identify opinion leaders referral ties among physicians, and individual-level
based on their self-reports or their centrality in social sales call data. Hence we are able to investigate
networks. More recent research by Coulter et al. the presence of contagion dynamics over social net-
(2002) and Godes and Mayzlin (2009) provides con- works in a real market in which traditional marketing
flicting answers to the question of whether heavy efforts are also being deployed, the kind of setting
195Iyengar et al.: Opinion Leadership and Social Contagion in New Product Diffusion
196 Marketing Science 30(2), pp. 195–212, © 2011 INFORMS
that is of greatest relevance to both practitioners and We proceed as follows. We first develop the three
researchers (Van den Bulte and Lilien 2001, Watts and research questions. We then describe our research set-
Peretti 2007). ting and research design. Next, we specify the vari-
The results are of both theoretical and manage- ables created for analysis and present the results. We
rial interest. Not only do we document the existence conclude with a discussion of implications for theory
of contagion in new product adoption after control- and research and for marketing practice.
ling for many potential confounds, including mar-
keting effort, but we also show that the amount of
contagion is moderated by both the recipients’ per-
2. Research Questions
ception of their opinion leadership and the sources’ 2.1. Social Contagion
volume of product usage. The second key finding is The most fundamental assumption of network mar-
that sociometric and self-reported measures of lead- keting is that social influence or social contagion
ership are weakly correlated and associated with dif- among customers is at work. Although this assump-
ferent effects, indicating that they capture different tion is often taken for granted, it need not always
constructs. In other words, we document important be warranted. For instance, several studies have doc-
contingencies in the social contagion process as well umented inflated evidence of contagion as a result
as important differences between the two main opera- of estimation problems and the use of theoreti-
tionalizations of opinion leadership. Although earlier cally overdetermined models. Unless one analyzes
studies have assessed the existence of contagion in individual-level adoption data and network data, con-
new product adoption after controlling for marketing tagion can easily be confounded with other mecha-
effort using actual network data (Hill et al. 2006, Van nisms generating temporal changes in adoption speed
den Bulte and Lilien 2001) or trying to proxy for net- (Van den Bulte and Stremersch 2004). Another source
work ties by geographical propinquity (Bell and Song of upward bias in prior evidence of contagion has
2007, Grinblatt et al. 2008, Manchanda et al. 2008) or been the failure to appropriately control for marketing
by group membership (e.g., Duflo and Saez 2003, Sac- effort and other changes in the market environment,
erdote 2001), or they have assessed the existence of as shown most compellingly in reanalyses of the clas-
social influence in the usage intensity or joint con- sic Medical Innovation study (Marsden and Podolny
sumption of established products and services within 1990, Van den Bulte and Lilien 2001).
dyads (e.g., Hartmann 2009, Nair et al. 2006), none of The second assumption underlying network mar-
those studies has documented how sociometric lead- keting strategies, that some customers’ adoptions and
ership, self-perceived leadership, and volume of prod- opinions have a disproportionate influence on oth-
uct usage interrelate in affecting contagion effects in ers’ adoptions, should not be taken for granted either.
new product diffusion. It is likely to hold when some customers have a
The findings reported here are of interest to much more central position in the social network
researchers seeking to better understand the relations than do others or when potential adopters look for
between opinion leadership, sensitivity to contagion, advice from experts (e.g., Goldenberg et al. 2006).
and time of adoption, a set of issues that recent In contrast, when the social network structure is not
research has shown to be more complex than was very centralized and when what spreads is simply
previously thought (Van den Bulte and Joshi 2007, information about the product’s existence rather than
Watts and Dodds 2007). The findings are also of inter- information that mitigates perceived risk—conditions
est to practitioners seeking to identify effective opin- typically associated with so-called buzz marketing
ion leaders. That contagion is at work, and sociomet- campaigns—then there is not much variation among
ric leadership is more strongly associated with early customers’ relative influence (e.g., Van den Bulte and
adoption than self-reported leadership, for instance, Wuyts 2007, Watts and Peretti 2007).
implies that the former metric is more effective in Thus, we assess whether new product adoption is
identifying early seeding points to jump-start the dif- subject to social contagion operating through network
fusion process. Another key implication is that the ties such that better connected adopters exert more
practice of targeting heavy users, common in the influence than do less connected ones and whether
pharmaceutical industry and elsewhere, is justifiable such contagion operates over and above the effect of
based not only on their higher “stand-alone” cus- targeted marketing efforts and systemwide influences
tomer lifetime value but also on their higher “net- that vary over time. To our knowledge, the network
work value” because they exert more social contagion. contagion effect has never survived such a stringent
However, because the correlation between prescrip- test because prior studies did not use social network
tion volume and sociometric leadership is only mod- data (e.g., Bell and Song 2007), did not account for
erate, just focusing on heavy users will fail to leverage degrees of connectivity (e.g., Hill et al. 2006), did not
all potential influential seeding points. control for marketing effort or time-varying shocksIyengar et al.: Opinion Leadership and Social Contagion in New Product Diffusion
Marketing Science 30(2), pp. 195–212, © 2011 INFORMS 197
(e.g., Coleman et al. 1966, Strang and Tuma 1993), identify the same influentials (convergent validity), as
or did control for marketing effort or time-varying indicated by a recent review of the literature (Valente
shocks but found no evidence of contagion (Marsden and Pumpuang 2007), we first assess to what extent
and Podolny 1990, Van den Bulte and Lilien 2001). their leadership scores are correlated. Apart from
three studies (Jacoby 1974, Kratzer and Lettl 2009,
2.2. Sociometric vs. Self-Reported Rogers and Svenning 1969), none of which used the
Opinion Leadership validated (Childers 1986) scale of self-reported lead-
The third key assumption underlying many network ership standard in marketing nowadays, we are not
marketing strategies is that firms are able to identify aware of any evidence on this fundamental issue.
and target influentials or opinion leaders. Rogers and Even less is known about whether the two leader-
Cartano (1962) discuss three ways to identify such ship constructs have the same association with adop-
people: (1) self-designation, i.e., asking survey respon- tion behavior (nomological validity). People who are
dents to report to what extent they perceive them- often nominated by their peers as someone they
selves to be influential; (2) sociometric techniques, turn to for expertise or discussion are likely to be
i.e., computing network centrality scores after asking true sources of influence. People who perceive them-
survey respondents to whom they turn for informa- selves to be influential, in contrast, may simply have
tion or advice, or after observing interactions through an inflated sense of self-importance. To the extent
other means (e.g., citations among scientists); and that true expertise drives early adoption, sociomet-
(3) the key informant technique, where selected peo- ric leadership may be more strongly associated with
ple are asked to report their opinion about who the early adoption than self-reported leadership is. On the
influentials are. Whereas self-designation is the most other hand, early adoption may be affected more by
popular technique among marketing academics, the how one perceives oneself than by one’s true status.
sociometric technique has been more popular among These arguments suggest that sociometric leaders and
social network analysts. The latter technique is also self-reported leaders need not adopt equally early but
gaining popularity among marketing practitioners to leave open the question which type of leader adopts
identify influential scientists, physicians, and engi- before the other.
neers (e.g., Dorfman and Maynor 2006) and among Whereas many studies have reported evidence that
some consumer network marketing firms, such as one of these two measures of opinion leadership is
Procter & Gamble’s Vocalpoint, that target people associated with early adoption, others have found no
with demographic characteristics associated with hav- effect (e.g., Goldenberg et al. 2009, Van den Bulte
ing a central network position. and Lilien 2001) or even negative effects (Becker 1970,
Doubts exist about the value of both self-reports Leonard-Barton 1985). More importantly, there is no
and sociometric measures. It is likely that self- evidence to date of the effect of one after controlling
reported opinion leadership is biased upward and for the other. That is, there is no evidence that both
that it also reflects self-confidence rather than actual have an independent effect on the speed of adoption.
influence. Conversely, doubts about marketers’ ability Such evidence is critical to the claim that both mea-
to effectively identify influentials using sociometric sures capture different constructs.
methods have arisen recently following a simulation The distinction between sociometric and self-
study by Watts and Dodds (2007) showing that the reported leadership may also affect how sensitive one
customers critical in generating a sudden burst in is to input from one’s peers. Following the original
the speed of diffusion need not necessarily be the two-step flow hypothesis, several studies have doc-
best connected. Although this possibility was already umented that the information flow between opin-
long known to network and diffusion researchers ion leaders and followers is not unidirectional. True
(e.g., Becker 1970, Locock et al. 2001), the recent sim- experts rarely ignore whatever user experience or
ulation results have created a heated debate among other information less prestigious actors have to share
marketing practitioners (Thompson 2008). Much of (e.g., Strang and Tuma 1993, Weimann 1994). This
that debate seems to ignore that the study by Watts suggests that sociometric leaders may be as respon-
and Dodds (2007) was only a simulation demonstrat- sive to contagion as non-leaders are (assuming lead-
ing a possibility, not an empirical study providing ers do not adopt too early to ever experience peer
actual evidence in support of that possibility. Still, the influence, of course). Self-reported leaders, in con-
simulation results bring to the fore potential difficul- trast, may be more or less sensitive to contagion than
ties marketers may face in identifying key influentials their peers. On the one hand, several theories of social
using sociometric methods. identity and status imply that people with a high
To gain a deeper understanding of the issues at sense of self-importance may deem it below their
hand, we test several hypotheses. Because little is dignity to take into consideration, let alone imitate,
known about whether different methods actually the behavior of lower-status actors (Berger and HeathIyengar et al.: Opinion Leadership and Social Contagion in New Product Diffusion
198 Marketing Science 30(2), pp. 195–212, © 2011 INFORMS
2007, Van den Bulte and Joshi 2007). On the other adoption process (e.g., Lin and Burt 1975). For prod-
hand, status competition implies that people who ucts that do not benefit from marketing commu-
think of themselves as having above-average status nication and that present little perceived risk or
might be driven to adopt quickly once they see others ambiguity such that little additional information is
of lower status adopting, out of fear that being out- required in the evaluation stage, Godes and Mayzlin’s
paced will lead their own status advantage to erode (2009) argument implies that light users will be very
(Burt 1987). Taken together, these arguments imply effective sources of influence. For products that are
that self-reported leaders are differentially (either less supported by a fair amount of standard marketing
or more) sensitive to social contagion compared with communication but pose significant perceived risk or
non-leaders, whereas sociometric leaders are not more ambiguity to potential adopters, in contrast, conta-
or less sensitive than non-leaders. gion fosters adoption by operating at the evaluation
Empirical support for the clear distinction between stage rather than at the awareness stage, so heavy
sociometric and self-reported leadership would be of users are likely to be more effective sources of influ-
theoretical importance because it would imply that ence. Testing the heavy-user hypothesis for a product
they are not different measures of the same construct, with significant perceived risk and ambiguity comple-
as advanced by Rogers and Cartano (1962) and Jacoby ments the study by Godes and Mayzlin (2009) of a
(1974), but are distinct theoretical constructs. Investi- low-risk product enjoying very little marketing sup-
port and for which contagion operated most likely by
gating the distinction is also of obvious value to mar-
boosting awareness rather than evaluation.
keters seeking to identify whom to target as seeding
Whether usage volume moderates the amount
points in their campaigns.
of social contagion exerted might conceivably also
depend on the stage of the diffusion cycle. For a firm-
2.3. Social Contagion Through Central
created word-of-mouth campaign started well after
Actors and Heavy Users the product’s introduction, such as the one studied
A key assumption underlying many network mar- by Godes and Mayzlin (2009), it is conceivable that
keting strategies is that some customers are not only all heavy users have already engaged their network
better connected but also more influential than are members (either successfully influencing them or not),
others. Who these customers are is critical for select- making heavy users less effective seeding points than
ing initial targets or “seeding points” in network mar- light users, who may still have many opportunities left
keting campaigns. We investigate whether customers to convert network members. This alternative expla-
who are heavy rather than light users, a characteristic nation implies not a reversal but a corroboration of the
that is easier and cheaper to determine than opinion light-users-are-better finding by Godes and Mayzlin
leadership, are disproportionately influential among for high-risk versus low-risk products.
those to whom they are connected. Whereas prior
research indicates that centrality in the network and
usage volume tend to be associated with early adop-
3. Research Setting
To provide valid answers to our research ques-
tion (Coulter et al. 2002, Taylor 1977, Weimann 1994),
tions and an informative assessment of the assump-
we investigate whether centrality and usage volume
tions underlying many network marketing efforts, the
also affect how effective one is as a source of influence
research setting should ideally satisfy several con-
after one has adopted. ditions. First, the newly launched product should
The answer to that question is far from obvious. have characteristics making it theoretically justified to
Standard theoretical arguments based on (i) the link expect contagion to be at work. Second, one must be
between repeat buying behavior and satisfaction or able to collect data on self-reported leadership and
(ii) the link between experience and source credibil- sociometric leadership for each person whose behav-
ity imply that someone who adopted a while ago ior is analyzed. Third, one must have data on who can
but is not using the product anymore is likely to influence whom. Fourth, one must have data not only
be less enthusiastic and less credible than someone on the adoption of each person whose behavior is
who is still using the product. Conversely, Godes and analyzed but also on the adoption and postadoption
Mayzlin (2009) note that heavy users may tend to be usage of others in their network. Fifth, key market-
connected mostly to people already predisposed to ing efforts deployed must be observed or otherwise
be early adopters. The latter idea implies that heavy controlled for.
users are less likely to generate new adoptions. We secured the cooperation of a pharmaceutical
Whether usage volume enhances or depresses the company to meet those stringent conditions. For rea-
amount of social contagion exerted is likely to depend sons of anonymity agreed upon with the company, we
on whether contagion operates by boosting either do not report its identity nor the drug’s name, treat-
awareness or evaluation, the two key stages in the ment category, or launch date. Like many other firmsIyengar et al.: Opinion Leadership and Social Contagion in New Product Diffusion
Marketing Science 30(2), pp. 195–212, © 2011 INFORMS 199
in its industry, the company was keen on identifying 3.2. The Physicians and Their Local Network
the physicians with the most central and influential Given the specific medical condition the new drug is
positions and on using that information in its medical treating, the company defined the relevant popula-
education and detailing programs. Managers realized, tion as those physicians who had prescribed at least
however, that their premises were in doubt; they were one of the earlier two drugs in the two years prior
therefore keen on facilitating a study about the impor- to the focal drug’s launch. Based on American Medi-
tance of social networks, opinion leadership, and mar- cal Association membership records, IMS prescription
keting effort. data, and the company’s internal records, the com-
pany supplied us with a list of such physicians prac-
3.1. The Product ticing in three large U.S. cities: San Francisco (SF),
The product is a newly launched prescription drug Los Angeles (LA), and New York City (NYC). Hence,
used to treat a specific type of viral infection. the relevant networks were bounded based on both
There are both short-term (acute) and long-term a positional criterion, being a physician practicing in
(chronic) forms of the disease. The chronic form can one of a specific set of five-digit zip codes, and an
cause severe damage to internal organs and—if left event criterion, having prescribed at least one of two
untreated—sometimes can even lead to patient death. drugs in the past (Laumann et al. 1989).
The product we study is the third entry in the cate- When studying opinion leadership and social con-
gory of drugs for treating the chronic condition. No tagion among physicians, it is important to take into
account their localized character. The importance of
later entries occurred during the observation window.
local as opposed to national opinion leaders is well
Because the condition is chronic, physicians cannot
documented in the modern medical literature (e.g.,
observe drug efficacy quickly and adjust a patient’s
Doumit et al. 2007, Keating et al. 2007, Kuo et al.
therapy if necessary. There is uncertainty in the med-
1998). Whereas nationally reputed “expert opinion
ical community regarding the best treatment because
leaders” may be respected for their research, to most
there exists little comparative information about the
physicians they are much less representative than
three drugs’ long-term efficacy. In an issue of a pres-
local “peer opinion leaders” who are members of their
tigious medical journal featuring two separate stud- own community and face similar patients and work-
ies documenting the focal drug’s effectiveness, an ing conditions (Locock et al. 2001). The pharmaceu-
editorial by a director of one of the National Insti- tical industry is keenly aware of the importance of
tutes of Health warned that even though the new such social dynamics at the local level. Better under-
drug seemed like an excellent treatment option given standing of local opinion leadership dynamics was
its low rate of resistance and outstanding potency, the main motivation of the pharmaceutical company
the drug’s use should—for the time being—be tem- to make our study possible.
pered because the medical condition requires long- Because the three cities we study are major
term therapy. metropolitan areas, the local networks also contain
In short, the drug treats a potentially lethal con- several national opinion leaders. That the physicians
dition, but there is considerable ambiguity and risk who the company considered to be national opinion
in making the decision to adopt. In such situations leaders also emerged as opinion leaders within their
characterized by high risk, high complexity, and low city made the network data fully credible to the man-
observability of results, both theory and research agers, who were also quite interested in the identity of
suggest that contagion is likely to be a significant prominent local opinion leaders they had overlooked
driver of adoption behavior (e.g., Hahn et al. 1994, so far.
Rogers 2003).
Those product characteristics also determined how
4. Data Sources
the product was marketed. The complex nature of the
Our data sources consist of a survey of physicians,
treatment decision made detailing (personal selling)
a commercial data vendor providing physician pre-
the main marketing instrument. There was only very scription data, and company records on sales calls to
limited medical journal advertising and no direct-to- each physician.
consumer advertising. There was no sampling either.
Given the chronic nature of the therapy, physicians 4.1. Physician Survey
cannot assess effectiveness rapidly. This drastically We used a mail survey to collect data on the physi-
limits the effectiveness of free samples in triggering cians’ social network ties and characteristics such as
the decision whether to use the drug as part of a treat- patient volume and self-reported opinion leadership.
ment plan. Even more important is the concern that The survey was mailed twice during a two-month
patients develop resistance when they take a sample period, separated by a reminder postcard. There was
but do not continue on the drug. also an online link provided for physicians whoIyengar et al.: Opinion Leadership and Social Contagion in New Product Diffusion
200 Marketing Science 30(2), pp. 195–212, © 2011 INFORMS
Table 1 Response Rates Across the Three Cities Network ties. Following Coleman et al. (1966), we
SF LA NYC
collected network data using a sociometric survey. We
asked each physician to name up to eight physicians
Mailing 187 273 372 with whom he or she feels comfortable discussing
Returned to sender 37 76 88
Return to sender (%) 198 278 237
the clinical management and treatment of the disease
Valid addresses 150 197 284 (discussion ties) and up to eight physicians to whom
Surveys completed 67 57 69 he or she typically refers patients with the disease
Response rate (%) 445 289 243 (referral ties). Both lists may but need not overlap.
Within the network boundary, the 67 respondents in
SF generated 37 unique nominees for discussion and
wanted to complete the survey online; about 10% of
24 unique nominees for referral. In LA, the 57 respon-
participants did so. A $75 honorarium was promised
dents generated 38 unique nominees for discussion
for completing the survey within two weeks of receiv-
ing it. In SF, the first mailing took place two months and 24 for referral, and in NYC the 69 respondents
before the U.S. product launch; in LA and NYC, it generated 43 unique nominees for discussion and 22
took place 10 months after the U.S. launch. for referral. Again following Coleman et al. (1966),
The response rate in SF was markedly higher than we excluded physicians who were nominated by sur-
in LA or NYC (Table 1). That may be due to a higher vey respondents but who were not part of the orig-
interest in the treatment options in the SF area, where inal network boundary (e.g., a physician cited by an
several national thought leaders are based and a siz- LA physician but practicing in Irvine, California, or
able population group lives with an above-average a fellow LA physician who had never prescribed in
risk of contracting the medical condition. It may also the category prior to launch).2 Physicians who were
be due to the higher quality of the mailing list. For within the network boundary but did not respond to
instance, there were a number of instances in LA and the survey, in contrast, were included in the network.
NYC where two entries in the list had the same name We then built a “discussion” and a “referral” network
but different addresses. In any case, the response rates matrix for each city, with respondents as rows and
in all three cities (24%–45%) are quite high by both all network members as columns and with the i jth
industry and social science standards and do not gen- cell being one when i cited j and zero otherwise. We
erate problems for our network-based covariates. We also constructed “total” network matrices by adding
discuss this in more detail in §7. the referral and network matrices in each city. The
Physician characteristics. Following Coleman et al. SF matrices were of size 67 × 150, those for LA were
(1966), we collected data on the type of primary prac- 57 × 197, and those for NYC were 69 × 284. Including
tice and physician specialty. We also asked about the all physicians who were part of the network bound-
number of patients seen and the number referred to ary as columns allows us to take into account the
other physicians because physicians treating many contagion emanating from everyone within the net-
patients are more likely to prescribe new drugs. To work boundary even if they did not respond to the
measure self-reported opinion leadership, we adapted survey.
the scale of Childers (1986) to our particular research As has long been known to researchers of conta-
setting. We used six items pertaining to the likelihood gion in social and spatial networks, symmetry of ties
and frequency of a physician to interact with other and extradyadic cycles (e.g., a triad with ties from
physicians on issues related to the chronic disease. All
node a to node b, from b to c, and from c to a) can
items were measured on a scale of 1 to 7.1
create an endogeneity or reflection problem (e.g., Ord
1975). Our data, however, do not exhibit such a struc-
1
In consultation with industry experts, and consistent with findings
ture. Of the 204 discussion ties among survey respon-
by Flynn et al. (1994), we excluded one item from the seven-item
Childers scale as it was not relevant to our research context. The dents, only three are symmetric. Of the 138 referral
six items included in our survey were these: In general, do you talk ties among survey respondents, none is symmetric.
to others doctors about ? (Never/Very often); When you talk Of the 234 “total” ties, only three are symmetric. Also,
to your colleagues about do you (Offer very little informa-
our data include only one instance of an extradyadic
tion/Offer a great deal of information); During the past six months,
how many physicians have you instructed about ways to treat
? (Instructed no one/Instructed multiple physicians); Compared to
your circle of colleagues, how likely are you to be asked about ways of advice/Often used as a source of advice). In these items, “ ”
to treat ? (Not at all likely to be asked/Very likely to be asked); stands for the medical condition treated by the focal drug.
2
In discussions of , which of the following happens more often? Robustness checks reported in the electronic companion to this
(Your colleagues tell you about treatments/You tell your colleagues paper, available as part of the online version that can be found
about treatments); In general, when you think about your profes- at http://mktsci.pubs.informs.org/, provide no evidence that this
sional interactions with colleagues, are you (Not used as a source exclusion affects our results.Iyengar et al.: Opinion Leadership and Social Contagion in New Product Diffusion
Marketing Science 30(2), pp. 195–212, © 2011 INFORMS 201
cycle: the three symmetric ties just mentioned per- no samples were distributed, the first recorded pre-
tain to a triad in the NYC discussion network.3 Thus, scription corresponds to the actual time of adoption.
reflection does not constitute a likely threat to internal For each physician-month, we create a binary adop-
validity of our contagion analyses. tion indicator variable yit that is set to zero if physi-
cian i has not adopted by period t and is set to one if
4.2. Prescription Data he has. The discrete-time hazard of adoption is then
For each physician within the network boundary (not modeled as
only respondents), the time of adoption is measured
using monthly individual-level prescription data from
P yit = 1 yit−1 = 0 = F xit (1)
IMS Health, a data provider whose role and reputa-
tion in the pharmaceutical industry is similar to that
of ACNielsen and IRI in consumer package goods. where xit is a row vector of covariates, is a column
For the focal drug, the data start from the month vector of parameters to be estimated, and F is a cumu-
the drug was introduced. Prescriptions were tracked lative distribution function (e.g., logistic or standard
for the next 17 months—incidentally, the same dura- normal). Because the population of interest consists
tion as that in Medical Innovation (Coleman et al. only of physicians who had prescribed within the cat-
1966). Data on postadoption prescriptions are avail- egory at least once in the two years prior to the focal
able as well. drug’s launch, we consider each and every physician
Of the 193 doctors across the three cities who to be at risk of adopting the new drug. Because we
responded to the survey, 68 adopted within 17 months. observe all physicians from the time of launch, left-
This adoption rate of 35% is markedly lower than the censoring does not exist in our data. Hence, we use
87% rate for tetracycline in Medical Innovation over the the standard log-likelihood function for discrete-time
same length of time, consistent with the notion that hazard processes, which appropriately handles right-
the present drug poses a greater risk to physicians censoring and can be expressed as
than tetracycline did.
We also have prescription data for the two other
N
Ti
drugs in the category for two years prior to the launch LL = yit lnF xit +1−yit ln1−F xit (2)
of the focal drug. This allows us to identify the heavy i=1 t=1
prescribers in the category before the focal drug was
introduced and to avoid problems that might occur if where Ti is the number of monthly observations on
the firm targeted heavy prescribers with a higher level physician i and N is the total number of physicians.
of marketing effort. By including the variable avail-
able to the decision maker in the model, we avoid
an endogeneity bias in the effect of marketing vari- 6. Covariates
ables included in the model (sales calls) and control
6.1. Opinion Leadership
for other targeted marketing efforts excluded from the
Indegree centrality. Indegree the number of other
model (e.g., direct mail).
physicians who nominate or “send network ties to”
4.3. Sales Call Data a particular physician; it is computed for each physi-
From the company’s internal records, we obtained cian separately in the referral, discussion, and total
data on the number of sales calls (detailing efforts) network (in the latter network, Indegree is the sum of
pertaining to the focal drug for each of the physicians the Indegree in the referral and discussion networks).
and in each of the 17 months we track. Free sam- Indegree centrality is the most basic measure of sta-
ples were not distributed, and the price did not vary tus or prestige in a network (e.g., Van den Bulte
over time. and Wuyts 2007). Because we measured the social
ties as pertaining to patient referral and discussion
of the treatment of a medical condition, physicians
5. Data Analysis Approach with high Indegree are, in the parlance of Goldenberg
We use hazard modeling as the main statistical
et al. (2006), both “social connectors” having many
approach to analyze the data and test the hypotheses.
We operationalize the time of adoption as the time of ties and recognized “experts” with expertise and good
first prescription (e.g., Coleman et al. 1966). Because judgment.
Self-reported Leadership. The reliability of the six-item
3
scale was quite high (Cronbach = 088), and factor
Because only ties between physicians who both responded to the
survey can be shown to be symmetric or to form cycles, this spe-
analysis confirmed the metric validity of the scale. We
cific analysis is limited to ties among respondents and excludes ties construct the Self-reported Leadership variable by taking
between respondents and nonrespondents. the average of the six items.Iyengar et al.: Opinion Leadership and Social Contagion in New Product Diffusion
202 Marketing Science 30(2), pp. 195–212, © 2011 INFORMS
6.2. Social Contagion 6.3. Marketing Effort
We operationalize exposure to prior adopters through We use monthly physician-level detailing (sales calls)
social ties using lagged endogenous autocorrelation as our measure of marketing effort. To allow for
terms. The extent to which physician i is exposed effects spanning multiple months, we construct a
at time t to prior adoptions is captured through the depreciation-adjusted stock measure. Let Dit be the
term j wij zjt−1 , where wij captures how relevant each amount of detailing (number of sales calls) received
physician j is to i and zjt−1 is a variable capturing the by physician i in month t. The detailing stock of
behavior of j at time t − 1.4 physician i for month t DSit is then defined as
The social network weight wij can be constructed follows:
in various ways (e.g., Valente 1995). We use a sim-
t
ple weighting scheme of contagion through direct ties: DSit = Dit + DSit−1 = t−
Di (3)
in both the referral and the discussion network, wij =1
equals one if i nominates j, and it equals zero oth-
where is the monthly carryover rate bounded
erwise. In the total network, the weights are sim-
ply the sum of the referral and discussion weights. between zero and one, and detailing stock in month 1
Because Indegreej = i wij , the number of colleagues is the amount of detailing in that month.5 To control
that a physician can influence directly through discus- for a potential confound in the interaction between
sion or referral ties equals his Indegree. The number contagion and the leadership variables, we allow the
he can influence indirectly is of course greater. effect of marketing effort to be moderated by Indegree
We capture the behavior of fellow physicians in and Self-reported Leadership.
three different ways.
6.4. Control Variables
(1) Adoption. In this variant, zjt−1 = yjt−1 , i.e., the
Physician characteristics. We control for several
lagged adoption indicator. The resulting contagion
physician characteristics that industry experts and
variable assumes that people start influencing once
prior research suggest may be associated with early
they have adopted and that they continue doing so.
adoption (e.g., Coleman et al. 1966, Rogers 2003).
This operationalization is the one commonly used
Identifying systematic heterogeneity in adoption time
in models of network contagion in the adoption of
is also of practical interest to managers seeking to
innovations.
(2) Use. In this variant, zjt−1 = sjt−1 , where sjt−1 is identify and target likely early adopters. Univer-
set to 1 if j wrote at least one prescription at time sity/Teaching Hospital is a dummy variable indicat-
t − 1 and is set to 0 if he did not. The resulting con- ing whether the physician works in or is affiliated
tagion variable assumes that only recent prescribers with a university or teaching hospital. Solo Practice
exert peer influence (e.g., for reasons of enthusiasm is a dummy variable capturing whether the doctor
or credibility). is in solo practice or not. Although this variable was
(3) Volume. In this variant, zjt−1 = qjt−1 , i.e., the important in the original Medical Innovation study, it
number of prescriptions written by j at time t − 1. is not clear a priori whether practicing solo is a use-
The resulting contagion variable assumes that one’s ful predictor once one takes into account actual net-
influence is proportional to one’s recent prescrip- work exposure to previous adopters. Early Referral is a
tion volume (e.g., again for reasons of enthusiasm or dummy variable taking the value one if the physician
credibility). reports sometimes referring patients to other doctors
Having operationalized both the social network before initiating any treatment, and zero otherwise.
weights wij and the various kinds of behavior zjt−1 , we A doctor referring patients even before starting any
calculate the extent of social network exposure physi- treatment is less likely to adopt the focal drug early.
cian i is experiencing at time t and create the follow- Primary Care is a dummy variable capturing whether
ing variables: Adoption Contagion, Use Contagion, and the doctor is a primary care physician rather than a
Volume Contagion. To assess to what extent sociomet- specialist (internal medicine, gastroenterology, infec-
ric and self-reported leadership moderates the effect tious disease) who is more likely to focus on the rele-
of these contagion variables, we also create the neces- vant medical condition. Patients Managed is the num-
sary interaction terms. ber of patients with the medical condition that the
physician reported clinically managing in the last six
4
Lagging avoids endogeneity problems, unless (1) people are months. Physicians with many patients may adopt
forward-looking not only about their own behavior but also that of sooner.
others, and (2) social ties over which influence flows are symmetric. Category-level prescription volume. Prior research sug-
The first condition is quite unlikely in large networks, and the sec- gests that early adopters and opinion leaders tend
ond condition does not hold in our data. Of course, if the contagion
in the data-generating process is contemporaneous, lagging creates
5
misspecification bias. We find no such evidence (see the electronic The carryover parameter is estimated jointly with the vector of
companion for details). slope parameters using standard maximum likelihood.Iyengar et al.: Opinion Leadership and Social Contagion in New Product Diffusion
Marketing Science 30(2), pp. 195–212, © 2011 INFORMS 203
to be heavy users (Coulter et al. 2002, Taylor 1977, and van der Klaauw 1995, Meyer 1990). Robustness
Weimann 1994). Thus, to avoid confounds in our checks indeed do not detect any evidence of unac-
hypothesis tests, we control for the physicians’ pre- counted unobserved heterogeneity (see the electronic
scription volume. Because no one uses the product companion).
before it is launched, usage volume should be mea-
sured at the category level prior to the new product’s
launch to be useful in identifying early adopters and 7. Final Data Set
to avoid reverse causality problems. We therefore use Data on past prescription of the two other oral antivi-
the number of prescriptions for each of the other two rals are missing for eight doctors, three of whom
drugs during the 12 months prior to the launch of had adopted the focal drug. We dropped these eight
the focal drug.6 As mentioned above, including these physicians from our data set. Thus, our analyses are
variables, Drug 1 and Drug 2, also avoids a potential based on data from 185 doctors, 65 of whom had
endogeneity problem in the detailing levels. adopted the focal drug after 17 months.
Outdegree is the number of nominations given or
“network ties sent” by a physician to others and 7.1. Descriptive Statistics
is computed for each physician separately for dis- We organize the data set as a panel from which
cussion, referral, and in total. Given the importance all postadoption observations are deleted because
of out-of-town contacts in the study by Coleman they do not contribute to the likelihood function
et al. (1966), Outdegree includes nominations to both of hazard models. Table 2 presents the descriptive
in-town and out-of-town colleagues. Unlike Indegree, statistics for these data. In this unbalanced panel,
Outdegree is not a measure of status or prestige. Sim- physicians’ weight in computing the means and corre-
ply connecting to many people may be related to lations equals the number of months until they adopt
being an opinion leader but it may as well indicate a or are right-censored. Table 3 reports the descriptive
lack of expertise and confidence (Van den Bulte and statistics for the time-invariant covariates using equal
Wuyts 2007). Hence we include Outdegree as a control weighting.
variable allowing a sharper interpretation of the Inde- Figure 1 shows the diffusion curve, i.e., the cumu-
lative proportion of physicians who adopted, and
gree effect. If Indegree is associated with early adoption
Figure 2 shows the empirical hazard rate, i.e., the
but Outdegree is not, one can be more confident in the
number of adopters divided by the number of those
interpretation of the former as a measure of status.
who have not adopted before. The diffusion curve
City dummies. We control for city-specific differ-
does not have a pronounced S-shape, and the hazard
ences in the propensity to adopt early by including
rate does not exhibit an upward trend. This, how-
dummy variables for LA and NYC, treating SF as the
ever, does not imply the absence of contagion because
reference.
heterogeneity across physicians, which generates a
Time dummies. We include a dummy for each
downward bias in the duration dependence, is not
period. This has two advantages. First, it captures the
accounted for and because the detailing efforts tar-
effect of any systemwide time-varying factor, such as
geted toward physicians who have not yet adopted
changes in disease prevalence or the appearance of
(called physicians “at risk” of adopting) have a clear
new clinical evidence. The dummies capture all cross-
downward trend (Figure 3). That the marketing effort
temporal variation in the mean tendency to adopt,
toward at-risk physicians decreases over time yet the
leaving only variance across physicians within partic-
empirical hazard rate does not suggests that, control-
ular months to be explained by contagion. As a result,
ling for marketing effort, the hazard rate may actually
including the dummies provides a stringent test for the
be increasing, which would be consistent with conta-
presence of network contagion. The second advantage
gion being at work.
of including monthly dummies is that it provides a
Figure 4 shows how the three contagion variables
nonparametric control for duration dependence. This,
for the total network evolve over time among the
in turn, absorbs much of the effects of possible unob-
physicians at risk. Both Adoption Contagion and Use
served heterogeneity in hazard models (e.g., Dolton
Contagion increase for the first nine months and tend
to level off afterward.7 Put simply, the exposure to
6
We also use the number of prescriptions over the six months both adopters and prescribers stopped growing after
prior to launch and over two years prior to launch. The fit (in
log-likelihood) of the model using the one-year window was
7
marginally better than those using the shorter and longer windows, Adoption Contagion and Use Contagion level off after nine months
but there were no substantive differences in the results. The same (Figure 4), even though the total number of adopters keeps growing
holds for using a three-month moving window of the number of roughly linearly (Figure 1) because those adopting after month 9
prescriptions for the other two drugs (see the electronic companion tended to have very low indegree and hence not to exert any con-
for details). tagion on colleagues.Iyengar et al.: Opinion Leadership and Social Contagion in New Product Diffusion
204 Marketing Science 30(2), pp. 195–212, © 2011 INFORMS
Table 2 Descriptive Statistics and Correlations Among All Covariates for All At-Risk Physician-Month Observations
Variable Mean Std. dev. Min Max 1 2 3 4 5 6 7 8 9 10 11 12
1. Adoption yit 0025 016 0 1 100
2. Detailing (sales calls) 029 083 0 9 023 100
3. Indegree—Referral 011 059 0 17 023 021 100
4. Indegree—Discussion 025 078 0 19 025 027 082 100
5. Indegree—Total 036 131 0 36 025 026 094 097 100
6. Outdegree—Referral 137 139 0 5 000 −002 000 −004 −002 100
7. Outdegree—Discussion 236 154 0 6 003 003 004 002 003 037 100
8. Outdegree—Total 373 243 0 10 002 000 002 −001 000 081 085 100
9. Self-reported Leadership 425 129 1 7 011 018 019 025 023 −017 019 002 100
10. LA dummy 031 046 0 1 −001 −002 000 −004 −002 −005 000 −002 015 100
11. NYC dummy 037 048 0 1 −002 004 −002 −003 −002 −007 −011 −011 002 −052 100
12. Solo Practice 039 049 0 1 000 013 −006 −014 −011 −003 −019 −014 −012 −001 009 100
13. University/Teaching Hospital 021 041 0 1 000 −009 −006 −004 −005 −013 000 −007 007 −012 003 −041
14. Primary Care 013 034 0 1 −005 −010 −008 −008 −008 011 −005 003 −024 013 −014 −006
15. Patients Managed 3636 10942 0 1200 004 012 008 012 010 003 −007 −002 002 −014 020 014
16. Early Referral 035 048 0 1 −007 −016 −005 −013 −009 021 −002 011 −044 −017 000 −005
17. Past Drug 1 1089 2576 0 265 025 050 038 055 050 −009 −006 −009 024 −007 015 002
18. Past Drug 2 1045 2486 0 510 025 032 032 042 039 −002 −006 −005 005 −011 009 009
19. Contagion—Referral, 057 091 0 5 002 005 001 −003 −001 062 030 055 −008 −014 −011 −008
adoption
20. Contagion—Referral, use 054 090 0 5 003 003 002 −002 −000 061 029 054 −007 −015 −009 −009
21. Contagion—Referral, volume 403 1045 0 104 005 002 002 −000 000 039 021 035 −006 −016 −009 −007
22. Contagion—Discussion, 065 097 0 6 004 007 004 000 002 031 036 041 −006 −004 −021 −011
adoption
23. Contagion—Discussion, use 057 089 0 5 004 003 006 002 004 031 035 040 −007 −008 −019 −012
24. Contagion—Discussion, 385 964 0 89 006 000 005 005 005 020 021 025 −006 −013 −015 −008
volume
25. Contagion—Total, adoption 123 171 0 9 004 006 003 −001 001 051 037 053 −008 −011 −018 −011
26. Contagion—Total, use 111 163 0 9 004 003 004 000 002 051 036 052 −008 −014 −016 −011
27. Contagion—Total, volume 788 1849 0 178 006 001 004 002 002 033 023 033 −007 −016 −014 −008
13 14 15 16 17 18 19 20 21 22 23 24 25 26 27
13. University/Teaching Hospital 100
14. Primary Care −001 100
15. Patients Managed −013 −008 100
16. Early Referral 013 015 000 100
17. Past Drug 1 −007 −009 022 −016 100
18. Past Drug 2 −006 −005 046 −004 053 100
19. Contagion—Referral, −011 002 003 013 −002 −002 100
adoption
20. Contagion—Referral, use −009 000 003 013 −004 −002 098 100
21. Contagion—Referral, volume −011 −003 002 011 −004 −002 072 076 100
22. Contagion—Discussion, −006 002 −003 009 −004 −005 065 062 047 100
adoption
23. Contagion—Discussion, use −006 003 −003 009 −005 −003 063 064 049 096 100
24. Contagion—Discussion, −008 000 −001 009 −005 −003 047 049 069 067 071 100
volume
25. Contagion—Total, adoption −009 001 000 012 −004 −004 090 087 066 091 088 063 100
26. Contagion—Total, use −008 002 000 013 −005 −004 089 090 068 087 091 066 097 100
27. Contagion—Total, volume −010 −002 000 011 −005 −003 065 068 093 066 065 091 071 073 100
Notes. Values computed on all physician-month observations for which physician was at risk of adopting, N = 2575. All correlations equal to or greater than
0.04 in absolute value are significant at p ≤ 005.
nine months. Under such conditions, the firm’s strat- observation period. If contagion based on peers’ pre-
egy to decrease the sales effort over time might have scription volume is more important than that based
been inappropriate to drive late adoptions. Instead, on their adoption or user status, then the firm’s detail-
increasing detailing once the effect of word of mouth ing strategy may have been quite appropriate. This
has stalled (i.e., after nine months) might have been simple analysis shows how the precise nature of the
more suitable. However, consider how Volume Conta- social contagion process may be quite relevant to mar-
gion trends upward throughout the entire 17-month keting policy.Iyengar et al.: Opinion Leadership and Social Contagion in New Product Diffusion
Marketing Science 30(2), pp. 195–212, © 2011 INFORMS 205
Table 3 Descriptive Statistics and Correlations Among Time-Invariant Covariates
Variablea Mean Std. dev. 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17
3. Indegree—Referral 037 157 100
4. Indegree—Discussion 058 190 095 100
5. Indegree—Total 095 342 098 099 100
6. Outdegree—Referral 134 143 −012 −014 −013 100
7. Outdegree—Discussion 240 158 −009 −011 −010 041 100
8. Outdegree—Total 374 254 −013 −015 −014 082 086 100
9. Self-reported Leadership 446 134 029 034 032 −021 013 −004 100
10. LA dummy 031 046 −009 −013 −011 −002 000 000 009 100
11. NYC dummy 036 048 −009 −007 −008 −013 −008 −012 002 −049 100
12. Solo Practice 038 049 −009 −014 −012 003 −012 −006 −017 000 004 100
13. Univ/Teaching Hospital 022 041 −006 −002 −004 −015 −006 −012 012 −009 005 −041 100
14. Primary Care 011 032 −008 −009 −009 009 −006 001 −027 013 −012 −004 −002 100
15. Patients Managed 4467 10982 026 028 028 −002 −009 −007 009 −016 021 008 −011 −009 100
16. Early Referral 03 046 −012 −017 −015 021 −002 011 −048 −016 000 −001 008 017 −006 100
17. Past Drug 1 2136 4711 054 062 059 −021 −021 −025 037 −011 009 −002 000 −013 029 −022 100
18. Past Drug 2 2144 5655 057 062 060 −012 −012 −014 021 −012 −004 003 −005 −009 031 −014 071
Notes. Values computed on a single observation for each physician, N = 185. All correlations equal to or greater than 0.15 in absolute value are significant at
p ≤ 005.
a
The numbers in front of the variables match those in Table 2. Min and Max values are the same as in Table 3.
7.2. Respondents vs. Nonrespondents Figure 3 Average Detailing Effort per At-Risk Physician
As mentioned earlier, the response rate was markedly 0.50
higher in San Francisco than in the other two cities. 0.45
However, the 185 respondents were not significantly 0.40
Average detailing
different (p > 005) from the 411 nonrespondents on 0.35
0.30
0.25
Figure 1 Cumulative Proportion of Physicians Having Adopted
0.20
0.40
Cumulative proportion of adoption
0.15
0.35 0.10
0.30 0.05
0
0.25 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17
Months
0.20
0.15 time-invariant characteristics of focal interest that we
0.10 observe for both groups:8 the amount of prescription
0.05
of two other drugs in the category for 12 months prior
to launch (21.4 versus 15.6 for drug 1; 21.4 versus 20.4
0
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 for drug 2) and sociometric leadership (total Indegree:
Months 0.95 versus 0.49). Moreover, none of those variables
was associated significantly with the probability of
Figure 2 Empirical Hazard Rate of Adoption responding after controlling for city in a multivariate
test (p > 005). Hence, there is no evidence of response
0.040
bias based on usage or sociometric status.
0.035
7.3. Validity of Network Measures
0.030
Nonresponse raises some special concerns in network
Hazard rate
0.025 studies because variables of interest are measured
0.020 not only on the respondents included in the analy-
sis but also on their connections to nonrespondents.
0.015
We discuss to what extent response rates of 24%–45%
0.010 affect our three network variables, (i) Exposure to Prior
0.005 Adopters, (ii) Outdegree, and (iii) Indegree.
0
8
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 Data on prelaunch prescriptions are missing for 27 of the 438 non-
Months respondents.You can also read
