Three essays on opinion leadership and social networks
TABLE OF CONTENTS
Title Page i
Authorization Page ii
Signature Page iii
Table of Contents vi
List of Figures vii
List of Tables viii
Essay 1: The Influence of Opinion Leaders 12
Essay 2: Opinion Leader-induced Cascades: An Experimental Study 54
Essay 3: The Emergence of Opinion Leaders in Social Networks 84
List of Figures
1.1 A network with M = 15 consumers and T = 6 49
1.2 Timeline of the model 50
1.3 A “chain” network in the queue model 51
1.4 A network with a “chain” followed by a “bulge” 51
The “chain” network used in the present experiment 73
2.1B The decision task 73
Display at decision-making point for the player in position 3 74
2.2B Feedback screen at the end of a typical “game” (round) in the experiment 74
2.3 Occurrence of cascades as predicted by theory 75
2.4 Benchmark response rates under different response conditions and across half-games 76
2.5 Comparison of average realized profits, expected profit of realized choices, and expected profits of different strategies 77
3.1 Timeline of the model 121
3.2 A example of a set L of two consumers that satisfy description (b) in Proposition 3.4(2), together with their (four) neighbors 122
3.3 A dyad (K 2 ) 123
3.4 Complete graphs K M with M = 3, 4, and 5 123
List of Tables
Consumer 2’s possible posterior means for q at the end of period 1 in the queue model given that consumer 1 buys the new product 52
1.1B Consumer 3’s possible posterior means for q at the end of period 2 in the queue model given that consumer 2 buys the new product 52
1.2 Comparison between the queue model and mainstream herd behavior model 53
Position 2 player’s possible posterior means for q after observing a right-arm lottery outcome of the player in position 1 78
2.1B Position 3 (Position 4) player’s possible posterior means for q after observing the lottery choice and outcome of the player in position 2 (position 3) 78
Statistics of theoretically predicted cascades 79
2.2B Proportional occurrence of cascades within each half-game 79
2.3 Occurrence of “runs” 80
2.4 Definition of information type as observed by a player in positions 2- 4 80
2.5 Definitions of strategies 81
2.6 Conditional responses of different strategies 82
2.7 Comparison between the setups of the experiment and the typical herd behavior model 83
3.1 The follower’s possible posterior means for q in the dyad model 124
3.2 Payoff table in the dyad model when T = 2 124
Three Essays on Opinion Leadership and Social Networks
by MAK, Wah Sung Vincent Department of Marketing The Hong Kong University of Science and Technology
This is a collection of essays related to opinion leaders, consumers who exert disproportionate influence on the purchase decisions of other consumers. In Essay 1, I model opinion leaders’ influence by an economic model in which each consumer in a social network makes once-in-a-lifetime choice between a new product and an outside option. Through this model, I find that opinion leaders potentially wield huge influence. In fact, if the consumers believe a priori that the new product is better than the outside option but only moderately so, a bad recommendation of the new product from the opinion leader is sufficient to stop further new-product adoption, resulting in consumers imitating each others’ outside-option purchases in a cascade of behavior. But the reverse phenomenon of imitative purchase of the new product occurs under more restricted conditions, suggesting a reason why negative word of mouth (WOM) often has more impact than positive WOM. Following these conclusions, in Essay 2, I describe an experiment designed to find empirical support for purchase cascades. I create theoretically predicted cascades successfully under all experimental conditions, and find evidence of increasing occurrence of cascades as the game proceeds. A major challenge in many WOM marketing campaigns is the cost-effective identification of opinion leaders, but empirical studies can agree on few consumer characteristics that are strong predictors of opinion leadership. In Essay 3, I attempt to explain these null or weak findings through a game-theoretic perspective on opinion leader-follower relationships. I find that opinion leaders who purchase with certainty at the beginning of the game emerge whenever certain general conditions regarding network structure and consumer time preferences are met; moreover, counter-intuitively, these opinion leaders might not be the consumers with the lowest time discount factors, suggesting that opinion leaders are not “born” but “made”.
The strategic use of word-of-mouth (WOM) communication as a marketing tool has gained increasing prominence in recent years. 1 An industry report (PQ Media 2007) estimates that US companies’ spending on WOM marketing jumped 35.9% in 2006 to US$981.0 million, compared with an overall growth of 7.7% in marketing expenditure, and would reach US$1 billion in 2007. Novel catchphrases such as “buzz advertising” and “viral marketing” are now well known to marketers, while industry players from established conglomerate Procter and Gamble to enterprising startup BzzAgent (with clients like Coca-Cola and Kellogg) continuously press ahead efforts in utilizing WOM as a medium for product promotion. Yet, amidst “the buzz on buzz” (Dye 2000), much remains to be clarified as to how product recommendations are spread or shared among consumers, how they affect consumer purchase timing and decisions, and how they may help or hurt profits. A key issue regards opinion leaders (alternatively called “influentials”), consumers who have disproportionate influence on other consumers’ purchase decisions. Opinion leaders, as uncovered in numerous studies from the 1940s to the present, are often ordinary consumers with direct influence limited to only a few fellow consumers who are described as followers or opinion seekers (Flynn et al. 1996; Rogers 2003). Their importance in new-product diffusion, as early adopters from whom other consumers obtain product recommendations through WOM communication, cannot be overstated. Firms carrying out WOM marketing often hope to identify opinion leaders among consumers and focus their resources on them, making them deserving topics of investigation in marketing and economics alike; yet, there has been relatively scant research on opinion leadership in either discipline. Moreover, whatever research there is on opinion leaders in marketing, economics, and sociology (in which the topic receives a good deal of attention under communication studies), it is usually empirical and survey-based. While such an approach is definitely worthwhile, the essays in this thesis represents another perspective on opinion leadership based on economic modeling and experimentation,
1 Exact definitions of WOM are somewhat difficult to construct (see Godes et al. 2005), and I here simply operationalize it as one-to-one communication of information among consumers; this definition excludes settings that lead to herd behavior in the traditional sense in the economics literature (Banerjee 1992; Bikhchandani, Hirshleifer, & Welch 1992), with which an agent may only observe other agents’ choice behavior without direct communication.
uncovering, in the process, hitherto unnoticed phenomena that are demonstrated by theorizing and also in the laboratory. Essays 1 and 3, which are theoretical in approach, are moreover distinguished by incorporating social networks into their models and seeking to establish results that are generally true with a large class of networks. Again, there has been little research in marketing as well as economics that is conducted in such a spirit.
OPINION LEADERSHIP Opinion leadership as an idea originated in communication studies in the 1940s and 1950s that led to Katz & Lazarsfeld’s (1955) seminal work. These authors suggest, in a “two-step flow” framework, that an opinion leader, under the influence of the mass media, formed her opinion, which was then passed on to her followers or opinion seekers. In addition, they offered the following explication:
“What we shall call opinion leadership, if we may call it leadership at all, is leadership at its simplest: it is casually exercised, sometimes unwitting and unbeknown, within the smallest grouping of friends, family members, and neighbors. It is not leadership on the high level of a Churchill, or of a local politico, nor even a social elite. It is at quite the opposite extreme: it is the almost invisible, certainly inconspicuous, form of leadership at the person-to-person level of ordinary, intimate, informal, everyday contact.” (p. 138)
Subsequent studies on opinion leadership are consistent with this notion (examples in marketing include King & Summers 1970, Childers 1986, Flynn et al. 1996, Nair et al. 2006, and Godes & Mayzlin forthcoming), which means that an “opinion leader” in those studies can be, in many ways, the polar opposite of a more intuitive view of an opinion leader as “a Churchill”, “a local politico”, “a social elite”, or media celebrity in general. While media celebrities certainly “lead opinions” in many cases, they can often be classified as “innovators” in the typical paradigm of diffusion of innovations, 2 because they often learn about new products and are invited to try them much earlier than all others. Their influence can also be
2 Researchers on diffusion of innovations classify adopters into five categories: innovators, early adopters, early majority, late majority, and laggards (Rogers 2003). Note that what Bass (1969) originally called “innovators” is different from Rogers’ (2003) sense of the word; see the discussion in Mahajan, Muller, and Srisvastava (1990), who adapt Rogers’ classification to the Bass model.
considered part of the influence of the mass media, because of the high exposure of their opinions in the media. However, the opinion leaders that are studied in the literature as well as here are rather in line with Katz and Lazarsfeld’s conceptualization. In fact, an opinion leader in this sense may not even be aware (“unwitting”) that she is an opinion leader at all, nor may a follower be aware that her purchase decision is influenced by a fellow consumer (“unbeknown”). Researchers have also now recognized that flow of opinion may consist of more than two steps, so that an opinion leader in one opinion leader/follower relationship can be a follower in another such relationship. Thus, in this thesis, I shall follow the literature in understanding opinion leaders as consumers “who exert an unequal amount of influence on the decisions on others” (Rogers & Cartano 1962; Flynn et al. 1996) at a person-to-person level. No other a prior notions about opinion leadership is assumed, in accordance with previous studies (in fact, one of my goals in Essay 3 is to show that no other a priori notions about opinion leadership can be assumed). As Flynn et al. point out, opinion leadership must also be accompanied by opinion seeking: an opinion leader must have someone to “lead”, and form an opinion leader/follower relationship with that other consumer. Thus what I shall focus on in my research (and establish formally in the essays) are consumer relationships in which there is a systematic, unidirectional influence of one consumer’s purchase decision and recommendation on another’s purchase decision; this influence can be direct (that is, between “neighbors” in a social network), as well as indirect. I shall be especially interested in “global” opinion leaders, who influence other consumers but are themselves not influenced by any other opinion leaders. Unless otherwise stated, the term “opinion leader” in this thesis is intended to denote these consumers.
CONCEPTUAL FRAMEWORK In this study, I conceptualize opinion leaders as consumers who influence other consumers but are themselves not influenced by any consumers, and model their influence and emergence under the following framework:
1. Consumers are connected through a social network consisting of WOM communication links;
2. There is a product category for which selling takes place over a finite number of time periods; 3. Every consumer demands at most one unit of the category during the selling horizon, and in addition, in the model in Essay 3, has preference over early rather than late purchase and consumption, all else being equal; 4. Over the selling horizon, the category consists of a new product and also a well- known outside option, both of which are available at fixed, constant prices; 5. Consumers are ex ante homogeneous but ex post heterogeneous regarding the consumption utility of every product; that is, the expected utility of a product before consumption is the same for every consumer, but its realized consumption (experience) utility can be different from one consumer to another; 6. All consumers hold the same prior information regarding the category; 7. An opinion leader purchases and consumes the new product at the beginning of the selling horizon; 8. After consumption, the opinion leader communicates with her immediate network neighbors about her consumption experience truthfully; 9. Consumption experience regarding the outside option has no informational value by itself, since the outside option is well known; consumption experience regarding the new product has informational value by itself and can be used to reduce uncertainty over the new product (in a Bayesian manner to be specified in the formal models); 10. There is a positive ex ante probability that the follower’s purchase decision will differ according to the opinion leader’s consumption experience.
It needs be re-emphasized that, in my framework, an opinion leader needs not be directly communicating with her follower: they may be separated by one or more consumers in a chain of network links. But an opinion leader can still have an indirect influence (in the sense of feature 10 above) on the follower through a series of purchase decisions/consumption experience along the chain that links them up, while a follower may be influenced by more than one opinion leader. This second possibility justifies the clause of positive ex ante probability in feature 10: suppose, for example, a follower takes recommendation from r opinion leaders, and her belief structure is such that she purchases the new product if and only if at least m < r opinion leaders offer positive recommendation for the product. Then, if it turns out
that more than m opinion leaders offer positive recommendation, any single opinion leader’s recommendation is ex post irrelevant given the all other opinion leaders’ recommendations. But, ex ante, there is a positive probability that any single opinion leader’s recommendation is pivotal. I acknowledge that not all opinion leader/follower relationships in reality possess all the above features. For example, some opinion leaders merely give recommendations of products without having personally consumed them, or may give dishonest recommendations. Moreover, opinion leaders sometimes (though less so than innovators) introduce new product information into the population, while here I only examine the function of an opinion leader in reducing consumption risk for the follower by sharing her consumption experience of the new product with her network neighbors. Lastly, in reality, new information about the product over and above fellow consumers’ consumption experience may appear after it becomes available. But I believe that my simplification, summarized above, do possess sufficient realism in capturing many opinion leader-follower relationships while allowing me to illustrate clearly the ideas about opinion leadership formation behind this thesis. To help understand this conceptual framework, consider a stylized example of two new theatre shows being put up at the same time over the same duration or “run”. One of them (the “new product”) is created by new talents, while the other (the “outside option”) is created by a well-known production/acting team. Previews for media representatives were completed before the shows were open to the public, and thus media information about both shows is already widely circulated at the beginning of the run. 3 Assume that every member of the theatre-going public would choose to see at most one show during the run. In this context, the opinion leaders would be theatre goers who see the show created by the new talents at the beginning of the run and then share their feelings towards it with their social contacts. 4 My first major research question is: how could these opinion leaders’ recommendations, despite being communicated directly to only small circles of social contacts,
3 I may also assume that avid theatre enthusiasts (the “innovators” in this context) see both shows as soon as the runs begin and quickly spread information about and opinions on the shows in blogs and forums, which then become absorbed by the rest of the population at the start of the run. The rest of the population is the target of my model. 4 A recent example of a “new-talent” theatre show being marketed by an active WOM campaign is described in Cox (2006).
ultimately influence the theatre-going public at large, perhaps even causing one show to flop while the other become a running success (Essays 1 and 2)? Now, suppose every member of the theatre-going public prefers to enjoy a show earlier rather than later. A dilemma then arises: while the show by the well-known creative team can offer a predictable expected quality, the show by the new talents may be refreshingly excellent or the opposite, and poses more uncertainty than the former. A theatre goer may thus like to wait for an acquaintance to go to the show created by the new talents first and give her recommendation (in addition to the media information that she already knows), before making a decision on how to spend her money and time on theatre-going during the run period. My second major research question is: how might opinion leaders emerge out of the theatre-going public in the very first place to see a show always earlier than others in situations like the one described, and influence others through “ordinary, intimate, informal” contact (Essay 3)?
OUTLINE OF THE ESSAYS Essay 1 is a theoretical attempt in modeling the influence of opinion leaders. Many opinion leaders are ordinary consumers who directly communicate with only a limited number of other consumers. How, then, may they influence consumer choices beyond their immediate contacts? Based on an economic model that combines features of WOM communication, consumer heterogeneity, and social network in a setting in which each consumer makes once-in-a-lifetime choice between a new product and an outside option, I find that opinion leaders do potentially wield huge influence in the society. In fact, whatever the network, if the consumers believe a priori that the new product is better than the outside option but only moderately so, a bad recommendation of the new product from the opinion leader is sufficient to stop further new-product adoption, resulting in consumers imitating each others’ outside-option purchases. But the reverse phenomenon, namely that a good recommendation of the new product from the opinion leader deterministically leads to complete adoption, is impossible in many networks unless the consumers are very optimistic about the new product. I consider that this asymmetry helps explain why negative WOM has more impact than positive WOM. Essay 2 describes an experiment designed to find empirical laboratory support for a major result in Essay 1. According to Essay 1, an opinion leader’s
recommendation regarding a new product can sometimes lead to the rest of the social network adopting or not adopting that product through a cascade of imitative purchase. The research in Essay 2 is an attempt to create such cascades in the laboratory using a simple realization of the paradigm in Essay 1 in the form of a “chain” network of four subjects. I successfully create theoretically predicted cascades under all experimental conditions, although they occur less frequently than the expected 100% occurrence. I also find evidence of increasing occurrence of cascades and increasing strategic sophistication as the game proceeds. Coupled with the fact that inefficient cascades appear consistently throughout the experiment, I conclude that the kind of imitative purchase behavior put forward in Essay 1 is indeed possible in reality, and can lead to (a) failure of new-product diffusion simply because of a chance failure of the new product at the opinion leader, and (b) the consumer population making the inefficient adoption decision merely because of a single trial with the opinion leader. Essay 3 is, like Essay 1, a theoretical analysis, but at a deeper level concerning the formation of opinion leader/follower relationships. The essay is motivated by the fact that a major challenge in many WOM marketing campaigns is the cost-effective identification of opinion leaders, while empirical studies can agree on few consumer characteristics that are strong predictors of opinion leadership. In Essay 3, I attempt to explain these null or weak findings through a game-theoretic perspective on opinion leader-follower relationships. I seek to demonstrate how opinion leaders may emerge endogenously in a population, and how opinion leaders with counter- intuitive individual characteristics may appear, through a game-theoretic model that is developed from that presented in Essay 1. The model combines not only WOM communication, consumer heterogeneity, and social network, but also time preference, in a setting in which each consumer makes once-in-a-lifetime choice between a new product and an outside option. I find that opinion leaders who purchase with certainty at the beginning of the game emerge whenever certain general conditions are met and the consumers’ time discount factors are all sufficiently low or all sufficiently high; in the latter case, counter-intuitively, it is possible that the most “patient” consumer becomes an opinion leader. It must be said that economic models, as with other attempts at understanding human behavior using mathematical means, unavoidably make idealizing assumptions about humans – mainly in terms of human rationality – that are, as a
rule, not completely truthful. And yet, they can be seen as benchmark scenarios, points of reference, approximations, or thought experiments, by which I can understand reality or make predictions about behavior. In the theoretical models in Essays 1 and 3, Bayesian rationality is assumed to be true and common knowledge among the consumers – a rather strong assumption, but similarly strong assumptions are not hard to find in the literature either. Once the model is set up and the assumptions incorporated, the theoretical results are just mathematical certainties. But the rationality assumptions, being strong as they are, are subjected to questions as to their empirical validity. This motivates the experiment in Essay 2, which is a check on one of the major results of Essay 1. (Bayesian) rationality is apparently not adopted by every subject in every round, and yet, the sizeable presence of purchase cascades that cannot be formed if subjects use simple decision heuristics rather than perform high-order inference in a Bayesian sense, rather shows the power of theorizing. It is my hope that these essays convey both the importance and limit of theorizing in marketing and economics, and also the value of economic experimentation in the study of marketing problems.
POSITION IN THE LITERATURE The research in this thesis stands at the confluence of several streams of literature. To start with, a key feature of my framework is that consumers share their consumption experience through WOM communication with their network neighbors. WOM has long been recognized as an important influence on consumer decision (Katz & Lazarsfeld 1955); indeed, it has been demonstrated that face-to-face WOM has a greater impact on consumer choice over printed information because of its vividness and credibility (Borgida & Nisbett 1977, Herr, Kardes, & Kim 1991). Much research has since been conducted on disentangling the variables that affect the psychological impact of WOM. For example, Grewal, Cline, & Davies (2003) study how brand similarity between an early and later entrant to the market influences the effect of WOM regarding the later entrant. In recent years, the rise of the Internet makes written WOM (“Word of Mouse”) through online channels a subject of both theoretical and empirical research including Godes & Mayzlin (2004), Chevalier & Mayzlin (2006), and Chen & Xie (forthcoming). Despite such progress, a closer look at how social networks influence WOM communication is relatively lacking in the marketing literature. Brown & Reingen
(1987) is an early exception with a detailed empirical study of the effects of social networks on WOM communication; they notably find evidence for Granovetter (1973)’s “strength-of-weak-ties” theory, which states that rarely activated weak social ties are important bridges that bring information across tightly knit social groups. But otherwise, as has been remarked in Godes et al.’s (2005) overview:
“Another important individual-level dimension is the structure of one’s network and the role the actor plays in it. While the popular press has discussed this – Gladwell’s (2000) ‘connectors,’ Rosen’s (2000) ‘network hubs’ – the marketing literature has been relatively silent about these factors.” (p. 419)
I aim to help fill this gap in the marketing literature with the present research, which takes as its basic objective the economic modeling of WOM communication in social networks. Recent development along this line includes Mayzlin (2002), who considers simple chain and circular networks in her model of “buzz” advertising. But otherwise, social network studies are more common in economics and sociology than in marketing. In economics, whenever social interactions are studied, researchers traditionally either assume that agents with private information make publicly observable decisions in sequence (Banerjee 1992, Bikhchandani, Hirshleifer, & Welch 1992), or agents randomly sample other agents in the population for their opinion (Banerjee 1993, Ellison & Fudenberg 1995, Banerjee & Fudenberg 2004). Çelen & Kariv (2004b) also take on the Banerjee (1992) and Bikhchandani et al. (1992) type of social learning problem but limits each agent’s information about the history of play to only the decision of the agent immediately in front of her in the queue (rather than the whole history, as in the earlier papers), and find a gradual convergence to herd- like behavior, which is commonly observed in this stream of literature. There is also an accumulating literature on social networks in economics. For example, there are studies on network formation (Jackson & Wolinsky 1996, Bala & Goyal 2000), or agents playing cooperative (Myerson 1977, Bolton, Chatterjee, & McGinn 2003) or dyadic non-cooperative games in exogenous networks (Corominas- Bosch 2003, Cassar 2007). Bala & Goyal (1998) and Gale & Kariv (2003) investigate how information flows in general networks and find conditions for long- run convergence of payoffs and behavior. Both papers investigate how information
flows in general networks and find conditions for long-run convergence of payoffs and behavior. However, agents in both models choose action simultaneously at regular intervals of time, while here I look at “once-and-for-all” decisions between two products (at least within some finite duration). Gale & Kariv (2003), moreover, adopts the traditional approach of assuming agents to have received independent private signals at the outset, an approach that is not taken up here. Bala & Goyal (1998) assume that agents in their model do not make inferences concerning the experience of unobserved agents (for example, the neighbors of their neighbors) through the observed actions of direct neighbors; I do not make this bounded rationality assumption in the present research. Social networks have also been studied for many years in sociology; random network is often taken as a benchmark by theorists in that field, though it has long been acknowledged that this needs be adjusted to accommodate real data (Rapoport 1951a, 1951b). But network research in the past ten years highlights even more sharply the complexity of real-life networks. Empirical evidence has been found for Watts & Strogatz’s (1998) and Watts’ (1999) “small-world” model (a concept that can be dated back to Milgram 1967) as well as Barabási & Albert’s (1999) model, which is most prominently characterized by a scale-free distribution of the number of neighbors to which an agent is connected. These models have been finding their way slowly into the marketing literature. For example, Garber et al. (2004), building from Goldenberg, Libai, & Muller (2001, 2002), is a recent, pioneering attempt to apply the small-world model in marketing problems through both simulation and analysis of real life data. But it must be said that, despite the success of both network models in many examples, they also fail to characterize network structure satisfactorily in a lot of cases too (Watts 2003); but at least they suggest that studies of information flow among consumers should better provide conclusions that are robust across a wide variety of social networks. It is in the spirit of finding non- trivial conclusions for WOM communication with marketing implications in general networks that this research is conducted. WOM is a major factor in new-product diffusion (Mahajan, Muller, & Bass 1990, Rogers 2003), and I focus the present investigation on this aspect of WOM. I shall maintain the assumptions implicit in models stemming from Bass (1969), that consumers are simultaneously influenced by external mass media as well as interpersonal communications in their social networks (Mahajan et al. 1990).