Abstract
We conduct a field experiment using a customized Facebook application that allows us to study the linkage between online social tie measures and an economic measure of trust. We endogenously identify latent heterogeneity in Facebook users’ notion of friendship and partition users as selective or non-selective based on their type. For non-selective pairs of users, each of whom are likely to befriend anybody and everybody, we find that their intersection of common friends is insignificant in its ability to predict trust. This provides the first evidence that traditional measures of dyadic-trust, such as embeddedness, used widely in physical world social networks do not directly carry over to online social networks. In contrast to existing network theory, our field study conducted within an online social network is able to capture revealed preference style measures that capture dyadic interactions between users. We find these measures have strong explanatory power in varying levels of trust, and thus we contribute to existing network theory in this new and expanding domain. Methodologically, our study is part of a nascent body of literature that showcases how large-scale online social networks can serve as platform to take social science and economics experimental research traditionally done in the physical laboratory to real-world subjects in a highly generalizable setting.
Keywords : Investment Game, strength of social ties, trust, switching regression, Facebook field study i
1. Introduction
“Put not your trust in money, but put your money in trust.”
~Oliver Wendell Holmes Sr.
Online social networks such as Facebook and Twitter consume an increasingly significant portion of our time and attention. A 2010 Nielsen study
1
estimated that the amount of time the average user spends on Facebook is about seven hours per month, and more importantly, is growing at the rate of
10% per month. With an active user base of over 955 million people at the time of writing, 80% of which is outside the US
2
, this collective and connected mass of humanity can be viewed as a reservoir of social and economic influence. Online social networks are credited with inspiring political action, driving viral and word-of-mouth spread of products and services (Aral and Walker 2012a, Aral and Walker 2011, Aral
2011, Bapna and Umyarov 2012, Hill et al. 2006, Iyengar et al. 2008, Mayzlin 2006) and playing a significant role in the diffusion of ideas through its influence on media consumption (Dellarocas and
Zervas 2011). A primitive underlying this conjectured “influence” is a notion that people trust the opinions, intentions or actions of individuals they are connected to. The availability of population-level instances of these online social networks provides a unique setting within which one might better understand how trust is related to different measures of social ties. Put metaphorically, it enables one to ask: Which of your Facebook friends would you invite to Tahrir Square? And who would you rely on for product recommendations, blog posts or news source tweets?
In this paper we conduct a field experiment in the context of a large-scale online social network, where little is formally known about the nature or veracity of the friendship formation process. Our research agenda is to determine whether there are linkages between users’ naturally occurring interactions on Facebook, as can be measured through our Facebook API, and an economic measure of trust. We believe these are important questions because while pro-social behavior has been widely documented (Mobius and Szeidl 2007) and its underlying drivers such as directed altruism and reciprocity separated out (Leider et al. 2010, Rosenblat and Mobius 2009), the underlying primitives of the technology enabled size, scale and reach of today’s online social networks is: a) less well understood, and b) represents potential for the industrial-strength deployment of social capital to act as the lubricant that reduces social and economic frictions. Consider as a motivating example the
California based peer-to-peer car rental business called getaround.com, or the peer-to-peer room rental community airbnb.com. Potential users can only sign up with these businesses using their Facebook ID.
1 http://blog.nielsen.com/nielsenwire/online_mobile/facebook-users-average-7-hrs-a-month-in-january-as-digitaluniverse-expands/
2 http://newsroom.fb.com/content/default.aspx?NewsAreaId=22
1
The message just below getaround.com’s signup button states that “ …Getaround is a community built on trust and by connecting with Facebook, you enable safe and open sharing .” Clearly, trust is a key factor for anyone wishing to rent their bedroom or automobile, and it appears that we are on the cusp of new business models that can be built upon measures of trust revealed through their naturally occurring interactions in online social networks. Our research aims to shed new light on socially mediated trust models by using an economic measure constructed from users’ Facebook interactions.
A new aspect of today’s environment is that social interactions often take place in IT platforms that offer interfaces (APIs) to the outside world. Subsequently, we as researchers, have unprecedented access to a data trail of rich micro-level dyadic behavior, something that is very hard to capture in more traditional offline social networks. For instance, we can compute real-time, dynamic and revealed preference style strength of tie measures between pairs of individuals by observing their online interactions, such as posting on each-others walls and jointly appearing in photographs. Prior work has relied on relatively costly declared-preference type surveys of friendship networks to elicit social ties
(e.g, Mobius and Szeidl 2007). These approaches are relatively static and do not capture the variety
3
of micro-level data that we can obtain from the Facebook API.
Our research objective is to determine the association between our various strength of ties measures that we can elicit from Facebook, and the underlying trust in an economic exchange between a pair of socially linked individuals. We measure trust in the context of an economic exchange using the seminal “Investment Game” (e.g. Berg, Dickhaut and McCabe 1995) that has been replicated in numerous studies.
4
One of our innovations in this work is to develop a custom Facebook application to play this game amongst “friends”
5
to look under the hood
6
of participants’ Facebook activity. We utilize three validated measures of social ties, namely: 1) How many common friends each user has, which relates to the traditional measure of embeddedness (Aral and Walker 2012b), 2) the number of times these pairs of friends has interacted with each other on their Facebook walls, and 3) how many times they are tagged together in a photograph (reflecting physical world ties as well as affinity). All three measures correlate positively to perceived social distance as validated through a survey by Gilbert and Karahalious (2010). This innovative study sheds new light on how trust and reciprocity in economic exchange relates to social distance and the strength of social ties by using these metrics to deploy a well vetted economic game in an exogenously formed online network. Specifically, armed with a reliable
3
IBM’s view is that big-data is not just about volume, but as much about variety, velocity and veracity.
4 http://www.fasri.net/index.php/2010/06/john-wilson-dickhaut-jr-1942-2010/
5
We use the terms “friendship” and “friends” within Facebook as they are commonly used in the vernacular.
Simply, it denotes a connection between two individuals on Facebook.
6
Within the bounds of Facebook’s privacy policy of course. Our protocol is approved by our respective IRBs.
2
trust measure from the Investment Game and micro-level strength of ties data from Facebook activity, we fit an empirical model that best represents this data and sheds new light on the linkage of our online strength of ties measures and exhibited trust. We believe this linkage will be a key building block for businesses
7
that rely on Facebook as a trust-fostering community, to, say, make better matching and pricing decisions based on our model’s findings. The data necessary to apply our model (i.e., the various strength of ties measures) is readily available from the Facebook API.
We also formally shed new light on the common notion
8
that friendship and tie formation on
Facebook is different from the physical world. Given the ease of befriending others online, we ask whether there is latent-heterogeneity along the dimension of selectiveness, and if so whether the traditional triadic closure measure of embdeddedness (Aral and Walker 2012b), a function of the number of common friends shared between pairs of individuals, has similar predictive power as in physical world networks. Tynan (2011)
9
documents a phenomenon labeled as Facebook’s fake friend’s epidemic and quotes blogger Henry Copel who details
10
the case of one “Nicole Bally,” who managed to befriend some 697 people, including celebrities such as Sean Parker, Arianna Huffington, Chad Hurley,
Henry Blodget, Jimmy Wales, and Seth Godin. To address these issues, and for society to efficiently tap into the vast reservoir of social and economic influence of large-scale online social networks, we need a newer understanding of the linkage between dynamic and behavioral strength of tie measures that can be captured from such networks and levels of trust between friends. This is the focus of our research.
Our measure of trust is derived from an economic game rooted in the experimental economics literature (e.g., Kagel and Roth 1997). It employs decisions involving monetary transfers to examine whether trust and reciprocity are exhibited in an exchange. In coupling the trust measure with social strength of ties measures, our work builds upon Leider et al. (2009) in that it deploys non-anonymous versions of what were traditionally anonymous games (Berg et al. 1995) to link social distance, trust and reciprocity. The primary focus of Leider et al. (2010) was to shed light on why, and using which mechanisms (directed altruism or reciprocity), people act in a pro-social manner. They derived the underlying social network using an innovative survey and task completion design that induced truth-
7
Such as peer-to-peer room and car rental services, namely AirBNB and GetAround.com respectively. These matching-markets have high frictions ex ante. Who would you let into your bedroom or car? If we can derive predictors of dyadic trust from Facebook activity, this can algorithmically plug into risk adjustment, pricing and matching at the transaction level.
8
9 http://www.huffingtonpost.com/kari-henley/are-facebook-friends-real_b_180204.html
http://www.itworld.com/it-managementstrategy/178261/facebook-s-fake-friends-epidemic -- The author received a friend request from a “cute young blonde thing named Marjorie.” He did not know Marjorie from
Adam, but mentions that she was enthusiastic about friending him, initiating chat sessions before he had a chance to respond.
10 http://blog.web.blogads.com/2011/06/08/are-you-also-exposing-your-private-parts-to-strangers-on-facebook/
3
telling about friendship information. They found that directed altruism increased giving to friends by
52% relative to strangers, and enforced reciprocal behavior lead to a further increase of 24% under conditions of efficiency. Our main objective in this study is to establish a relationship between three different, naturally occurring and observable strength of ties measures and differential trust amongst friends in a general global online social network. In doing so, we make inroads towards developing a model of trust within online social networks by identifying important antecedents that can explain variations in exhibited trust.
Methodologically, our work develops a framework for economic studies on Facebook-like networks, which includes protocols for subject selection, pairing and game play. Our framework facilitates reproducibility, thereby making provenance easier to establish. We showcase how large-scale online social networks can serve as a platform to take economic games that are typically run in the physical laboratory to real-world subjects. We expect this approach to be fertile ground for future economic and social science research.
Briefly, our study is modeled along the lines of Rosenblat and Mobius (2009) and is comprised of three stages as depicted in Figure 1 below. The first stage is used to establish the positive effect of social capital on trust, as is consistent with theoretical expectations. We achieve this by contrasting nonanonymous (treatment) trust levels, where subjects do not know the identity of the person they are paired with, and anonymous (control) trust levels. In doing so, we establish the methodological validity of our Facebook based protocol (Rosenblat and Mobius 2009; Leider et al. 2010) by generating results consistent with prior findings. The second stage focuses on the non-anonymous treatment group. For those pairs where identities are known we have trust and reciprocity measures that we match to the three strength of ties measures gathered via our Facebook API. Specifically these measures include the number of shared wall posts, the number of photos in which the dyad is mutually tagged, and a normalized measure of the number of common friends shared between the dyad, which we call
“ Embeddedness ”.
11
The final stage of the research design is an empirical analysis where we develop a model using the trust measures gathered in Stage 1 and the strength of ties measures gathered in Stage
2 to construct a model of trust in online social networks.
11
Embeddedness s,r
= ( number of common friends ) s,r
/ min ( k s
-1, k j
-1) where k s and k r
are the network degree of the
Sender and Receiver respectively. We discuss this measure in greater detail later in the paper.
4
Figure 1: Research Schema – An Experiment is used in Stage 1 to obtain the Trust Measure, Strength of
Tie Measures are matched in Stage 2, and an Empirical Model is developed in Stage 3 to Link Strength of Ties
Obtained from Facebook to Trust
Our approach represents progress over prior studies relating to networks because it uses an exogenously created social network as our basis for inferring social ties. Thus, we avoid problems of endogeneity and generalizability that arise as a result of artificially created networks. For example, prior experimental work in this area has attempted to build artificial social networks in a laboratory or in similarly restricted online settings (e.g., Suri and Watts 2010), leading to concerns regarding the robustness of the network as well as the nature of what are called “social” ties. In contrast, we utilize a pre-formed network that is exogenous to our study, which allows us to use actual social relationships as the basis for measuring social ties between individuals.
Our initial analysis finds that for the average user wall-post exchanges are positively and significant linked to trust. Similarly, we find that social affinity and physical world ties, as captured by the appearance together in photographs, is positively and significantly linked to trust. Interestingly, we find that that the unconditional correlation between our “number of common friends” metric – a widely used structural equivalence or embeddedness based strength of ties measure in the social networks literature - and trust is not significant. Further, conditional on the cohesion based, revealed preference style, strength of ties measure, the number of common friends measure is weakly significant and negatively linked with trust. This motivates us to examine whether there is some latent heterogeneity in
Facebook users’ normative view of friendship.
5
To account for the latent heterogeneity in Facebook users, we estimate a switching regression model with two regimes, wherein the observations are endogenously partitioned into two sets, such that the likelihood of observing the realized experimental data is maximized. We think of these two sets as representing “non-selective users” and “selective users”, and our results from this analysis finds those individuals with fewer total Facebook friends (selective users) exhibit different outcomes than those with more Facebook friends (non-selective users). Specifically, the association between our common friend measure and trust is positive for the selective user model and insignificant for the nonselective user model. We find that for selective users the online social tie strength metrics, as measured by the number of shared wall posts and the number of mutually tagged photos, are significantly and positively linked to trust. For non-selective users, we find no significant relationship between wall-posts or our common friends measure and trust, however, we do find a significant positive linkage between appearing in a photograph together and trust. This suggests that for non-selective users their online interactions are of a seemingly facile nature and their offline connections, exemplified by co-locating in photos, are a signal of social affinity and physical world ties. This is an important finding for researchers and practitioners looking to develop business models based on the Facebook social network.
To the best of our knowledge, ours is the first study to rigorously establish the nuanced nature of friendship in online social networks, and its linkage to trust in an economic setting in large-scale online social networks, and thus, our findings have theoretical and practical implications. Using a rigorous field experiment in the world’s largest online social network, we are able to link different social mechanisms to endogenously determined segments of Facebook users. Our offline strength of ties metric, number of photos tagged, is particularly useful for non-selective Facebook users who may understand the spurious nature of their online friendships. In contrast, structural equivalence embeddedness measures, such as the number of common friends (e.g., Burt 1987), continue to hold sway for those selective Facebook users who are discriminating about whom they befriend. Practitioners developing social contagion and peer-influence business models based on communities such as
Facebook will be served by appreciating this heterogeneity in the user base.
The following section discusses the background for this work, drawing on related literature. We then provide a brief description of the Investment Game and the specific measurements that proxy for the strength of social ties within pairs. Next we report our results, their implications, and conclusions.
2. Background
Our work is set against the backdrop of two related observations. On one hand, despite the rapid growth of online social networking, the resultant social capital has not yet been shown to directly
6
influence online economic activity, and the monetization of social platforms continues to be a challenge
(Sundararajan 2012). This represents a significant research opportunity for several reasons. First, there is a wealth of social networking data available online, and one would expect such information to have vast potential in improving efficiency and reducing frictions in online markets. Second, a variety of exchanges in the offline (“physical”) world are often mediated by pre-existing social ties, and one of the promises of Internet commerce seems to be that it can scale these aspects of “small community” trustbased exchange to a larger global village. We elaborate on the underlying theoretical foundations of these ideas in the next two subsections.
2.1 Trust and Trustworthiness
Some notion of trust as an important determinant of outcomes permeates a variety of academic disciplines, including economics (e.g., Dasgupta 1988), social psychology (e.g. Leqwicki and Bunker 1995;
Lindskold 1978), and marketing (e.g. Anderson and Weitx 1989; Dwyer et al. 1987). Universally, trust can be defined as “the willingness of a party to be vulnerable to the actions of another party based on the expectation that the other will perform a particular action important to the trustor” (Mayer et al.
1995). Trust has also been referred to as “an important lubricant of a social system” (Arrow 1974) and has been shown theoretically to be important in economic exchanges. In some cases trust has also been shown to reduce transaction costs by mitigating opportunistic behavior (Bromiley and Cummings 1995).
There is precedent to viewing trust as something which ensues when an individual calculates the costs and/or rewards of cheating, and upon determining that it would not be in the best interest of one party to cheat, assumes that party can be trusted (Lindskold 1978; Akerlof 1970). Related work examines trust building within anonymous exchanges (Ho and Weigelt 2005) and the relationship between social ties and trust in knowledge transfer outcomes (Levin and Cross 2004). Work looking at the influence of technologically mediated exchange on trust and deception within distributed teams shows that in some cases mediation can lead to greater trust when compared to face-to-face communication (Burgoon
2003). Evolutionary models suggest that trust maximizes genetic fitness and therefore, is likely to eventually emerge in spite of self-interested motives. This suggests that trust can be viewed as a primitive that guides behavior in new situations (Berg et al. 1995).
The dynamics of trust and trustworthiness have a wide array of applications, including the facilitation of social and political action, as highlighted by recent events in Egypt where much of the uprising was enabled by social media outlets such as Facebook and Twitter. Because the expression of alternative political views could have been potentially dangerous under the current regime, it was important for individuals to trust those with whom they were interacting in each respective online
7
medium. Trust and trustworthiness also play an important role in facilitating online economic exchange, as demonstrated when considering the dynamics within buyer and seller dyads (Rice 2012). For example, upon deciding to engage in an online economic transaction a buyer must move first and pay for the good before it is shipped. In other words, the buyer must initially trust that the seller will deliver the good as promised, and in doing so allocate payment to that seller before receipt of the good. If a seller’s trustworthiness is deemed to be low it is less likely the buyer will trust that seller and the exchange may not occur.
Our conceptualization of the connection between trust and social ties starts from Burt (1992) who shows that when two friends share a friend (or in the language of social networks, there is triadic closure in the friendship) this is likely to lead to a greater level of trust between the two. Burt examines two competing hypotheses for this assertion: the bandwidth hypothesis, which suggests that the presence of the shared friend adds a channel of information exchange which facilitates greater trust, and the echo hypothesis, which theorizes that the shared friend acts as a credible “threat” which prevents reneging and thus increases trust. We do not aim to disentangle these two hypotheses in our study, although it is possible that upon further development our platform might be used for this purpose. However, we do account for a greater propensity for trust in triadic relationships by designating a metric related to the number of shared friends as one of the three measures of social tie strength. Importantly, prior literature has not pitted this measure against other behavioral and dynamic measures such as the intensity of interactions and the measure of affinity such as the appearance in a common photograph. Together with the somewhat noisy nature of what constitutes a friend in online social network settings, we are motivated to generally examine the linkage between the various types of strength of social ties measures and trust.
Prior empirical research examining differential trust
12
(Leider et al. 2010) has been restricted to highly specialized networks, or has been limited to revealing gender or ethnicity (Bohnet and Zeckhauser
2004; Eckel and Grossman 2006; Eckel and Wilson 2003; Fershtman and Gneezy 2001). Experimental work incorporating game theoretic design offers insights regarding factors that can impact individual levels of trust in economic exchanges (Bolton et al. 2001; Rice 2012). See Camerer 2003 for a more comprehensive review of experimental studies with other-regarding preferences. Other work looks at the importance of trust formation (McKnight 1998), showing that early development of trust is critical to establishing functional organizational relationships. Our approach differs from this prior literature in that we do not focus on the formation of trust, rather we look at trust within established social ties,
12
as opposed to measuring absolute trust between strangers.
8
using the general population of the largest online social network, the dynamics and fundamental constructs of which are not yet fully understood.
2.2 The Strength of Social Ties
The concept of tie strength can be defined as the following: “The strength of a tie is a combination of the amount of time, the emotional intensity, the intimacy, and the reciprocal services which characterize the tie” (Granovetter 1973). Two types of ties are commonly referred to in the literature, weak ties and strong ties. A weak social tie exists between two individuals who are not closely connected, such as casual acquaintances or co-workers who do not interact regularly, while strong ties exist between close friends who communicate frequently.
Studies on the effects of tie strength show both strong and weak ties have an impact on information dissemination (Granovetter 1973), job seeking (Granovetter 1974; Bridges and Villemez
1986), and even income levels (Simon and Warner 1992; Gorcoran et al. 1980). Experimental work that investigates the formation of social ties using a public good experiment shows that tie formation is contingent upon the success of the game. The aggregation of individual ties comprises a social network, and the importance of social network structure on the formation of trust has been touted in the social psychology literature. Additionally, work by Karlan et al. (2009) models a setting where social structures are used as collateral in procuring loans, showing that networks can build trust when agents use their connections as social collateral.
While the importance of social ties and network structure on social and economic outcomes is clearly of interest to the academic community, to date there are few empirical contributions to this body of literature. Recent studies on the effects of social structure on other regarding behavior include work by Leider et al. (2010) and Rosenblat and Mobius (2009), which maps a friendship network using a combination of surveys and truth-telling inducing tasks to evaluate strength of ties between individuals.
Trust is measured as a function of self-reported social distance and the authors find that friends exhibit higher levels of trust than non-friend pairings. Other experimental work investigating the effects of social ties on trust and altruism does so by first creating a network in the laboratory, and then testing behavioral effects on those same networks (Di Cagno and Sciubba 2008). Methodologically our work differs from these earlier studies in that we do not try to contrive an artificial network in a laboratory, or rely on self-reported measures regarding friend relations.
3. Methodology
9
3.1 Research Design and Procedures
To address our questions regarding the relationship between social strength of ties and trust we developed a Facebook application to play the well-established Investment Game. Subjects play in pairs where one person is the Sender and the other person is the Receiver. The Sender-Receiver pairs are randomly matched either anonymously, in which case they are not told the identity of the individual they are paired with, or non-anonymously where identities are known and individuals are verified
Facebook “friends”. While our primary focus is on the non-anonymous pairings where the strength of ties measures are allowed to vary, we also run the game with anonymous pairings in order to compare the two settings and confirm the veracity of our design. If the results of this comparison were not consistent with prior literature, simply that anonymous pairings generate lower economic returns than non-anonymous pairings, we would have reason to question our design.
The game begins with the Sender (first mover) receiving an endowment, which in our study is
$10. The Sender is then asked how much of this amount he/she would like to send the Receiver (second mover). Any amount sent is tripled and upon receipt of the Sender’s allocation the Receiver decides how much to return. This single shot game concludes after the Sender learns of the Receiver’s return choice and payoffs are reported to both players. Players are paid by depositing money into their designated PayPal account or by mailed check.
13
The game is depicted in Figure 2.
Sender is given $10 and chooses amount to send Reciver
Amount Sent is multiplied by 3
•Amount Sent = The Trust Measure
•Amount Returned = The Reciprocity Measure
Receiver decides how much to return to
Sender
Figure 2: Schematic of the Investment Game
The sub-game perfect equilibrium for a single shot Investment Game is one where Receivers expropriate the entire amount invested by Senders, and so Senders opt not to invest. The purpose of the anonymous version of the game is to validate our design by replicating results consistent with prior studies, serving as a benchmark for us to compare trust levels in the non-anonymous treatment. In this
13
Sender payoffs: $10 – (amount sent) + (amount returned); Receiver payoffs: 3x (amount sent) – (amount returned)
10
game the first mover is the trustor because any investment made can be taken in full by the Receiver.
For the game to proceed, the Sender must trust that the Receiver will return an amount deemed worthy of play and invest accordingly. As such, the trust metric in the Investment Game is the monetary amount sent by the Sender to the Receiver. The other metric in this game is the measure of reciprocity exhibited by the Receiver, which is the amount he or she returns to the Sender. Typically this analysis is secondary and relatively straight forward, as it tends to be overwhelmingly driven by the amount invested by the Sender. As we report later, our results of the Receiver analysis are consistent with prior findings
14
.
Subjects were recruited to play our game via an initial email, in which we asked those who were interested in participating in the study to sign up for our Facebook API. This initial list of email addresses was obtained from undergraduate and graduate classes, as well as from staff and co-workers, across three large Universities. Out of approximately 600 emails initially sent, 190 people signed up for our
Facebook API and became part of the subject pool. Each individual was given $5 as payment for signing up. From this pool of potential players, subjects were randomly assigned to the anonymous or nonanonymous versions of the game via computerized matching. The only requirement was that nonanonymous pairings be known “friends” on Facebook, otherwise there would be no existing social tie to analyze. When subjects were selected to play the game they were sent an email instructing them how to proceed. Prior to the start of each game participants read written instructions, had the option to receive further instruction via a YouTube video
15
, and had to get all answers correct on a quiz that tested comprehension of the game rules. Display panels of the game protocol are available upon request from the authors. The Sender/Receiver exchange occurred via email within the Facebook API. Players were first told the role to which they were assigned, and then the Sender was sent an email containing a link to the decision screen. After Sender decisions were made, the Receiver was sent an email with the
Senders respective choice and was given a link to the screen where the return choice was made. At the point the Receiver submitted his or her return amount the game ended and payoffs were reported to each player in a follow up email.
14
We point out that within the methodology of experimental economics the goal is to explain behavior in the aggregate rather than among individual subsets of the population. Therefore, while studies such as McKnight
(2002) parse trust into multiple psychometric dimensions that could vary by individual, our randomized design allows us to interpret results as mean behavior of the population (Croson 2005).
15 http://www.youtube.com/watch?v=na1Pe39XYO8
11
To minimize the possibility of offline collusion, each Sender was paired with three different, randomly chosen, Receivers, each of whom was a Facebook friend of that Sender
16
. Participants had no control over whom they were paired with. The Sender was given a $10 endowment for each Receiver choice, so Senders were asked to indicate how much of the $10 endowment they would send to each of the three Receivers shown on their screen. Figure 3 depicts a screenshot of what the Sender saw.
Figure 3: Screenshot of Sender’s Move in the Facebook Investment Game
Once allocation decisions were input by the Sender our system randomly chose and implemented only one of the three potential Receivers to actually play the game; however, the Sender did not know who would be chosen at the time of his or her investment decisions. We believe this element of our design has multiple strengths. First, it forces the Sender to cognitively analyze the underlying level of trust as a function of varying social tie strength, which is the effect we want to capture, and second, it is parsimonious in its use of subjects. In addition, and importantly, it eliminated offline collusion for us by increasing the cost of such communication
17
. Because our protocol pairs a randomly selected Sender with three randomly selected and known Facebook “friends”, we obtain three dyadic data points using
16
In an informal post-game questionnaire we asked participants if they attempted to collude with their matched partners and the overwhelming response was that they did not.
17
This feature was added after a pilot study revealed, in a post-experiment survey, that if senders had to make only one decision, they were highly likely to collude.
12
four subjects, whereas a classic between subjects design would require six subjects to gather the same amount of information. This is of significant practical importance in designing and conducting economic experiments in real-world social networks where recruiting subjects can be costly. Finally, with multiple observations per individual Sender, we have the potential to use panel data techniques (i.e., clustering by Sender due to the within subjects nature of the choice) to take care of other unobservable differences in individuals that could possibly confound our estimates.
Following successful completion of our pilot study, we proceeded with the broader study. We recruited our initial pool of subjects via email from three different universities. All who signed up received a $5 payment and were entered in our pool for a chance to play the game. Senders and
Receivers were recruited in the same manner and of these groupings we ended up with a total of 77 completed games to analyze. There were several cases where either the Sender or Receiver did not make a move and so we opted to remove these incomplete games from our dataset.
While getting subjects from which general inferences can be drawn is a challenge in most experimental settings, it becomes particularly so when implementing our study in a real-world social network. While Facebook is a large global network, the effects that we are interested in, namely the heterogeneity in the strength of ties, are local and materialize at the dyadic level. Thus, it is not enough to get a subject pool of a certain size to meet the power requirements of statistical inference. In addition, the subject pool must exhibit sufficient network characteristics, and often these can be conflicting in their impact on the overall research. For instance, while it seems appealing to have participants from a large, even global, geographical swath of users, such a sample might be relatively less clustered than subjects who were members of an interest group or a company or a university. Given that our unit of analysis is a connected friend dyad, a large random sample of unconnected global users, would be useless given our design. Thus, there is a classic generality-clustering trade off in the sampling procedure for social network based economic games of the sort we are interested in. We expect this to be an interesting area for future methodological research in design of network field experiments.
3.3 Measuring the Strength of Social Ties
A critical aspect of our study is quantifying the strength of social ties between individuals in an online social network. We achieve this by incorporating two measures validated by (Gilbert and
Kavahalois 2009) and a third measure from the social networks literature, to estimate tie strength based on readily available information found on individual Facebook pages. The Gilbert and Kavahalois 2009 study finds two primary factors that help explain much of their predictive tie strength model. The first
13
measure they call intensity, and this is the number of common wall posts shared between Sender and
Receiver ( WallPostShared ). The second strength of ties measure they refer to as intimacy, and this is the number of tagged photos ( PhotosTagged ), meaning the number of instances where the Sender was tagged in the Receiver’s photo, or vice versa.
Our third social ties metric comes from the sociology and social networking literature and is a measure of the number of mutual friends shared between two people (e.g., Granovetter 1973, 1974).
While we realize the importance of estimating the effects of the total number of friends shared by the
Sender and Receiver, we also recognize that each Sender/Receiver dyad is limited by the minimum degree of the pairing. Thus, we create an alternative normalized measure of common friends, referred to as Embeddedness , which is defined as:
Embeddedness s,r
= ( number of common friends ) s,r
/ min ( k s
-1, k j
-1) where k s and k r
are the network degree of the Sender and Receiver respectively.
In our empirical analysis, we also control for Sender and Receiver type by including his or her network degree centrality ( ReceiverDegree, SenderDegree ), as well as the overall activity level of the users, captured as the total number of wall posts each has received in the last three months ( ReceiverWall ,
SenderWall ).
18
4. Analysis and Results
Trust in Socially Mediated Exchanges
Our initial objective is to establish the methodological validity of our Facebook based protocol for playing the Investment Game. We achieve this by comparing the anonymous version of the
Investment Game to the non-anonymous version of the game, and find that the levels of trust and reciprocity are not significantly different from the range of values observed by Berg et al. (1995).
Moreover, consistent with Berg et al. (1995) we find that very few subjects among the first movers sent zero
19
and the amount sent was highly variable, resembling a uniform distribution over the different
18
We also initially included demographic controls such as gender and age but these yielded no explanatory power and were subsequently omitted for the sake of brevity.
19
Although the subgame perfect Nash equilibrium is for Senders to send $0 in the anonymous game, this assumes players have preferences solely for economic gain. However, we know that individuals likely have some nonpecuniary preferences, such as a preference for fairness (Fehr and Gachter 2000), which can explain why in most anonymous Investment Games investment levels are not strictly $0.
14
possible transfers. In our setting the 95% confidence interval of the sent amount is [4.75, 6.38] and this maps closely to the amounts cited in Berg et al. (1995).
We also analyze the economic differences between the anonymous benchmark versus the nonanonymous treatment groups. Table1 reports mean differences between the two groups.
Treatment
Anonymous
Non-Anonymous
Amount Sent
$4.75
$6.38
Amount Returned Number of Dyads
$3.81 20
$10.23 57
P value 0.01 0.005 t-test 2.10 3.35
Table 1: Mean Comparisons of Amount Sent and Amount Returned in Anonymous vs. Non-
Anonymous Experimental Treatments
The average amount sent in the non-anonymous treatment is higher than in the anonymous benchmark, and the same case holds for the average amount returned by Receivers. Both comparisons are statistically significant and in the expected direction. These results are consistent with expectations based on extensive prior literature, namely that strangers are less trusting then when identities are known. While the findings are relatively intuitive, conducting this manipulation within our Facebook setting is important to further validate our experimental protocol and allows us to move forward with our analysis of the non-anonymous treatment data.
Empirical Analysis of Trust Predictors
After confirming that significant economic differences exist between the baseline and treatment groups, we drilled down into the non-anonymous data to explore the relationships between our
Facebook strength of ties measures and Senders’ level of trust. We begin by estimating an OLS regression model with clustered errors to account for Sender specific correlated errors across decisions.
20
OLS is justified because our research design eliminates any possible selection issues. Recall that users are first randomly selected to either the anonymous or non-anonymous treatment, and then, in the non-anonymous game users are paired randomly conditional on them being friends. This matching protocol avoids any selection problems that might have arisen from a non-randomized procedure, where users may have selected to play the game with those they trusted. Our dependent variable is the amount sent by the Sender to the Receiver and is a continuous variable [0, 10]. We
20
As a reminder, Senders designate amounts for three different possible Receivers, only one of whom is randomly selected to complete the exchange. So when analyzing the amount sent in order to account for the possibility of correlated errors across these three choices we must cluster the choice by Sender.
15
include our strength of tie measures, including Embeddedness , PhotosTagged and WallPostsShared . We also include, as controls, the network degree centrality of each Sender and Receiver, SenderDegree and
ReceiverDegree, measured as that individual’s total number of Facebook friends, as well as the total number of wallposts of each ( ReceiverWall and SenderWall). Descriptive statistics for the variables used in our OLS analysis are shown in Table 2.
Variable
AmtSent
AmtReturned
PhotosTagged
SharedWallposts
ReceiverDegree
ReceiverWall
ReceiverLO
Mean
6.38
10.23
2.22
1.73
502.29
141.45
0.456
Std Dev
3.59
8.69
15.88
3.48
254.37
163.15
0.502
Median
6
8
0.000
1.00
502
101
0
SenderDegree
SenderWall
Embeddedness
Observations
435.40
215.22
13.182
57
214.32
286.74
22.423
Table 2: Descriptive Statistics of Variables
415
123
6.171
16
Results of the OLS regression with clustered errors (by sender) are reported in Table 3.
Correlation tables are not included for the sake of brevity, but are available from the authors on request.
Variable
Constant
Coefficeint
(standard errors)
6.55
(2.89) p-values
(t-statistic)
0.031
(2.27)
Embeddedness
-0.002
(0.032)
0.952
(-0.06)
PhotosTagged
WallPostShared
ReceiverWall
SenderWall
0.028
(0.013)
0.078
(0.087)
-0.002
(0.004)
-0.001
(0.002)
0.04
(2.15)
0.051
(2.03)
0.677
(-0.42)
0.605
(-0.52)
ReceiverDegree
SenderDegree
-0.000
(0.002)
0.000
(0.004)
2.06
0.997
(0.00)
0.999
(0.00)
VIF
R square .11
Number of 57
Dyads
Table 3: How trust is related to direct and “shared friends” measures of the strength of ties
21
DV = the amount “trusted” by the Sender to their Receiver; errors are clustered by Sender
Embeddedness = ( number of common friends ) s,r
/ min ( k s
-1, k j
-1)
PhotosTagged = total number of photos in which Sender/Receiver are tagged together
WallPostShared = total number of times Sender/Receiver post on the others wall
ReceiverWall/SenderWall = total number of wallposts for each player
ReceiverDegree/SenderDegree = network degree centrality for each player, measured as total number of
Facebook Friends
We find that each additional wall-post made on a friend’s wall is associated with a 7.8 cents increase in trust, which maps to a 6.5% increase given the empirical distribution of the non-anonymous sending level. Similarly, we find that each additional photo in which two friends jointly appear (a signal of social affinity and physical world ties) is associated with a 2.8 cents increase in trust, which maps to a 2.3% increase given the empirical distribution of the non-anonymous sending level. The result we find most
17
striking is that the coefficient of Embeddedness is insignificant. This runs counter to social networks literature, which suggests that triadic closure signals a stronger tie (which in turn should be associated with higher trust). It also counters literature suggesting that trust is directly enhanced by the presence of shared friends, either through an increase in the “bandwidth” of the channel or an increase in the
“echo”, the threat of being ostracized by shared friends for untrustworthy behavior (Burt 1992). This surprising result motivates us to develop a greater understanding as to the true nature of friendship ties in online social networks such as Facebook.
4.1 A Closer Look at the Nature of Friendship Link Formation on Facebook
There could be several possible explanations for embeddedness not being significant in our analysis above. Broadly, as we will see, the possible explanations triangulate on the underlying latent heterogeneity among Facebook users regarding the manner in which they extend or accept friendship requests. First, it may well be the case that after accounting for more direct measures of the strength of a tie (such as shared activity and tagging), the residual information content of shared friendship is minimal. A second explanation might be that what is called “friendship” on Facebook is a different kind of social tie, and one whose strength or bandwidth is not positively associated with triads. A third (and related) explanation might be that people’s social connections are quite heterogeneous across the population, and thus while the presence of mutual friends has a positive impact on trust for those people who form ties which represent “real” social ties, it has little impact on those who form ties which represent more shallow links based on, say fleeting shared activities, business affiliations or simply casual interest (“I like your picture, let’s be friends”). The insignificant coefficient for embeddedness may be consistent with a relative indifference that Facebook friends who form such ties incur. Put differently, failure to find any significant relationship between trust and our mutual friends measure might be driven by those users who are less discriminating and correspondingly, more ambivalent with respect to trusting their vast number of Facebook friends, when compared to those users who are more selective about their friendship choices.
Towards uncovering some of this latent heterogeneity, we next explore potential differences in types of Facebook users. We partition Senders by their total number of Facebook friends and classify them as either “selective users” if their total number of friends is less than the median, otherwise they are classified as “non-selective users”. We are most interested in seeing if conditioning our analysis on user type can provide additional insights into our findings and further our model of trust. In other words, presuming Senders and Receivers with a high number of friends are more likely to form spurious connections on Facebook, and their intersections therefore are likely to be non-informative, might this
18
understanding impact trust and reciprocity in non-selective pairings? Alternatively, are users with fewer friends more likely to exhibit trust as a function of greater discretion employed in choosing Facebook friends? We first explore the mean comparisons between pairings of selective vs. non-selective users.
Table 3 shows some interesting contrasts.
Amount Sent
Mutual Friends
Shared Wallposts
NON-Selective Pairing
$3.84
93.07
1.23
Selective Pairing
$6.16
33.45
1.40
Shared Photos
Sender Total Friends
Receiver Total Friends
Number of Dyads
0
638
678
29
.25
392
291
28
Table 4: Descriptive Mean Comparisons of Selective vs. Non-Selective Pairings
Table 3 shows the amount sent by non-selective pairings is almost half that sent in selective pairings. Not surprisingly, the intersection of two sets of primarily spurious friendships is likely to contain a large number of somewhat facile common friends. When interacting with another user who also has a large number of Facebook friends, our descriptive statistics show the average number of mutual friends is higher (93.07 vs. 33.45), possibly driven by the fact these two non-selective users have more Facebook friendships overall. Moreover, the higher number of mutual friends for non-selective pairings suggests these dyads may be the likely result of contrived Facebook friendships.
To further explore the differences between selective versus non-selective users, and to endogenize the latent class discovery process, we estimate a switching regression model (Maddala
1975, Lokshin and Sajaia 2004) with two regimes. The observations are endogenously partitioned into two sets based on the Senders’ total number of Facebook friends as a function of an underlying latent model linking trust to social tie characteristics. The partition is such that it maximizes the likelihood of observing the realized experimental data and assigns a probability to each Sender as to whether he or she is a selective or non-selective user.
In particular, the model we estimate using maximum likelihood describes the differential behavior of selective and non-selective Facebook users under two latent regimes identified by a criteria function I i
, such that:
19
Note I i
*
, which in general is the utility of participating in a certain manner (in our case its being selective or not selective), as determined by Z i
(a vector of characteristics that influence the participation) is not observed, but when it exceeds a certain (a priori unknown) threshold, the individual decides to participate. This is endogenously estimated. The corresponding observable variable I i
is dichotomous, and takes the value 1 if the individual ‘participates’ (in our case if the person has greater than median number of friends) and is 0 if he/she does not ‘participate.’ The errors u i
, e
1i
, e
2i
, have a trivariate normal distribution with mean zero and a covariance matrix, and the model is identified by construction through non-linearities. Note that the covariates corresponding to Z in the selection equation are
Sender’s wall count, Receiver’s wall count, and the network degree centrality of both the Sender and
Receiver. All are significant at the 1% level in identifying the latent classes. The model includes our three social distance measures, Embeddedness , SharedWallposts and PhotosTagged, along with
ReceiverLO, which is a dummy variable coded as 1 if the Receiver’s total number of Facebook friends is lower than the median. We include this variable in the model to further examine our univariate results, which suggest the Receiver’s type has some impact on the amount sent and that selective pairings differ from non-selective pairings.
Results of the maximum likelihood estimation of the switching regression for selective and non-selective users are shown below in Table 5a and 5b, respectively.
20
Constant
SharedWallposts
PhotosTagged
Coefficient
(standard errors)
-0.788
(0.089)
0.396
(0.006)
0.068
(0.0009) p-value
(t-statistic)
0.000
(-8.86)
0.000
(60.32)
0.000
(69.06)
Embeddedness
ReceiverLO
0.019
(0.002)
9.8
(0.071)
0.000
(7.25)
0.000
(137.21)
R square .9
Table 5a: Switching Regression Results for Selective Facebook Users (errors clustered by Sender)
Constant
SharedWallposts
Coefficient
(standard errors)
6.49
(1.08)
0.244
(0.165) p-value
(t-statistic)
0.000
(6.01)
0.151
(1.47)
PhotosTagged
Embeddedness
0.026
(0.010)
0.003
(0.024)
0.017
(2.54)
0.877
(-0.16)
ReceiverLO -1.02
(0.882)
0.253
(-1.17)
R square .07
Table 5b: Switching Regression Results for NON-Selective Facebook Users (errors clustered by Sender)
Our results show those individuals with fewer total Facebook friends (selective users) exhibit markedly different outcomes than individuals with more Facebook friends (non-selective users).
Specifically, for selective users who presumably do not form spurious friendships on Facebook, our social distance measures all have a positive and significant association with the amount sent. One explanation is that many of the Facebook friends of selective users might also be offline friends, suggesting that their social ties extend beyond an online medium, leading to greater tendencies for exhibiting trusting behavior. Perhaps these online friendships are more authentic, such that wall posts
21
represent meaningful exchanges. Photos tagged suggest an offline connection and the level of shared friendships while fewer, seem more apt to encompass trust. Within this regime we find that for each shared wall post there is a large 39.6% increase in the total amount sent to Receivers. This suggests that interactions are hugely meaningful for selective Facebook users, perhaps because they employ greater discretion in choosing their online friends. At the same time, the coefficient on ReceiverLO is significant and positive, suggesting that selective Senders may be more likely to trust Receivers who are also discerning in choosing their Facebook friends. Perhaps there is a belief that non-selective users are disingenuous and do not form real friendships on Facebook, and therefore, may exhibit less trustworthy behavior in an online exchange.
Turning to our results for non-selective users, we find the only significant factor in trust levels is whether the dyad appeared in photos together. One explanation for this interesting finding is that
Senders with a lot of friends may purposefully delineate between Facebook connections and “real” friends, realizing that the majority of their Facebook friendships are superficial at best. Accordingly, the positive and significant coefficient on PhotosTagged (p-value = 0.017) implies that in non-selective dyads real offline friendships matter most and presumably are closer than online friendships, thereby leading to greater levels of trust. When individuals are tagged in the same photo this implies an offline closeness, which is a different relationship than online linkages, particularly for non-selective users. An offline friendship implies less social distance, which could help explain why the number of mutually tagged photos is positively associated with trust levels in both regimes. Interestingly, we find the total number of wall posts is not significant, indicating that Facebook interactions for non-selective users are probably facile, and have little bearing on trust. Again, this result can be explained by the somewhat artificial friendship formation that is more likely to occur among non-selective users.
The switching regression generates probabilistic weights for each sender, endogenously delineating the likelihood that he/she will be placed in the selective vs. non-selective regime as a function of the selection equation specified. To construct Table 5 below we split Senders above and below the .5 probability of falling into the HI or LO sender regime, as generated by our switching regression.
Sender Friends
Receiver Friends
ReceiverWall
SenderWall
Non-
Selective
459
616
149
272
Selective
392
502
141
109
22
WallShared
PhotosTagged
MutualFriends
Dyads
1.8
3.29
72.2
57
1.6
0.25
33.45
Table 5: Descriptives of Sender HI/LO partition where regime is endogenously determined by the
Switching analysis
One interesting aspect of Table 5 relates to the Dunbar number, a measure which emerged from anthropological studies showing there are cognitive constraints on how many friendship interactions one individual can sustain (Dunbar 1993). Though there is no exact Dunbar number, a common benchmark for the maximum number of friends one individual can manage is approximately 150, and this is much less than both of our Sender and Receiver averages. We purport this difference further strengthens our position that a social tie on Facebook is a unique connection that likely deviates from more established concepts of friendship. In the Dunbar studies it is presumed that friend groups are colocated and interactive, such that the cognitive costs of maintaining a greater number of interactions
(i.e., over 150) would likely be higher than one could feasibly sustain. In contrast, technologically enabled connections are likely relatively lower cost, even for Selective users. Maintaining a group of
“friends” on Facebook does not require constant contact or even regular communication, hence the marginal cost of one more friend is not the same as the marginal cost of maintaining one additional offline friendship. Moreover, Facebook interactions are often asynchronous, further reducing the burden of communication and enabling greater numbers of social connections than the Dunbar number would suggest is possible
22
.
While there seems to be anecdotal evidence of the noisy nature of friendship on Facebook, extant literature has not yet addressed its boundaries. In our data, the median number of Sender friends is 415, and it is evident that above this threshold, triadic closure style metrics don’t offer much with respect to trust in an economic exchange; however, evidence of offline friendships do hold some explanatory power. As such, cohesion style interaction based strength of ties metrics such as appearances in photographs should be used in developing social contagion and peer-influence business models. In contrast, structural equivalence measures, such as the number of common friends (Burt
22
The mean number of friends on Facebook is between 245 and 359, which is quite close to the means of both
Sender and Reciever friend counts http://www.washingtonpost.com/business/technology/your-facebook-friendshave-more-friends-than-you/2012/02/03/gIQAuNUlmQ_story.html
23
1987), tend to continue to hold for those selective Facebook users who are discriminating about whom they befriend.
4.2 Receiver Reciprocity
Historically the primary objective of the Investment Game has been to provide a quantifiable metric of trust as a function of the first mover’s choice, while a secondary aspect of the game is to provide a measure of Reveivers’ reciprocal behavior. For completeness we analyze the return choice of
Receivers and find their return decisions are less sensitive to individual type differences. Specifically, comparative tests show no difference in the amount returned by Recievers when partitioned on mean friendship levels (p-value = .265). Moreover, overwhelmingly the primary driver of reciprocal behavior in the Investment Game is the amount of trust exhibited by the first mover and we have no theoretical explanation for why a selective Receiver might be more or less willing to reciprocate trust than a nonselective Receiver. Therefore, we pool our data and employ an OLS regression to model the return choice.
23
Results are shown in Table 7 below.
23
As a test we also ran our return analysis with the switching regression, mirroring the Sender analysis specifications only reversed for Receivers. We found that all subjects were endogenously placed in a single regime, further supporting the case that the same distinction (selective vs. non-selective types) does not factor into the reciprocity of Receivers and pooling the data into a single regime analysis is appropriate.
24
Variable
Constant
Embeddedness
PhotosTagged
SharedWallposts
ReceiverWall
SenderWall
ReceiverDegree
SenderDegree
AmountSent
VIF
R square
Number of Dyads*
Coefficient
(std errors)
-3.33
(4.28)
-0.003
(0.038)
-0.023
(0.020)
0.592
(0.135)
0.0003
(0.003)
0.0002
(0.004)
0.005
(0.004)
-0.006
(0.007)
1.824
(0.227)
1.56
.90
18 p-value
(t-statistic)
0.455
(-0.79)
0.944
(-0.07)
0.27
(-1.14)
0.002
(4.37)
0.921
(0.10)
0.965
(0.05)
0.164
(1.53)
0.363
(-0.96)
0.000
(8.03)
DV= the total amount returned to the Sender by the Receiver
* By design we randomly select one Receiver out of each Sender’s three investment choices, therefore we have fewer Receiver decision points.
Table 7: OLS Regression Results of Receiver’s Move
We find, consistent with prior Investment Game analyses that the most influential factor in
Receivers return decisions is the amount invested by the Sender. This suggests positive reciprocity plays a significant role in the return amount choice, and the t-statistic on the AmountSent variable (t-test
= 8.03) indicates that a large proportion of the variance in our model is explained by this measure.
Positive reciprocity can be defined as the impulse or the desire to be kind to those who have been kind to us (Fehr et al. 1997), and has been observed in second movers in a variety of trust games and giftexchange games (Fehr et al. 1997, Berg et al. 1995). Specifically, in our setting, when Senders invested larger amounts Receivers were likely to positively reciprocate by allocating greater returns than when lower amounts were invested. Finding that Receivers’ return behaviors are driven in large part by the desire to reciprocate is consistent with prior studies that show when first movers send greater amounts
25
they are rewarded with larger proportional returns from second movers.
24
We also find that
SharedWallposts is positively and significantly related to Receiver levels of reciprocity. While it appears
Receivers reciprocate higher amounts with those Senders whom they have exchanged more wall posts, the specific mechanism behind this is unclear to us. It is a limitation of our design (which, recall, is not focused on reciprocity) and offers a useful direction for future research more explicitly designed to study reciprocal behavior in online social networks.
5. Conclusion
The advent of a social layer over what has been multiple decades of exponential growth in computing, communications and data has opened up an array of possibilities in how we define our socio-economic and socio-political reality. The global online social graph (Facebook, Twitter and other niche social networks) has flexed its muscles in contexts as diverse as the Arab Spring, to the recent negation of SOPA and PIPA, to numerous consumer uprisings against corporate incompetence, neglect and abuse. Alternatively, the building reservoir of social influence is a gold-mine waiting to be tapped for purposes such as shaping our consumption of media and entertainment, rethinking delivery of education, deriving new pricing, marketing and even business models for digital media, designing the next generation of social recommendation engines for books, movies, dating, formation of teams, location of expertise within traditionally organized institutions, and the crowd-sourcing of innovation and talent.
As large-scale online social networks consume an ever-increasing portion of our time and attention, and as trust serves as a primitive for the spread of social and economic influence in such networks, this study informs us on the relative importance of key strength of ties measures that can be computed from observing users’ activities. As such, we explore important dynamics in online social networks where the notion of friendship formation is less understood relative to the physical world.
Further, we measure the marginal effects of three different measures of strength of ties on trust, while shedding new light on the nuanced nature of friendship formation on Facebook. For the average user, we find that each additional wall-post made on a friend’s wall is associated with a 6.5% increase in trust.
Similarly, we find that each additional photo in which two friends jointly appear is associated with a
2.3% increase in trust. We differentiate between selective versus non-selective users based on their total number of Facebook friends to disentangle the often spurious nature of friendship in this online
24
We also interact the amount sent with each of our social ties measures to test whether these had a moderating effect on return amounts. We found none of these interaction terms to be significant.
26
social network. We estimate a maximum likelihood based endogenous switching regression model with two regimes, wherein the observations are endogenously partitioned into “non-selective users” and
“selective users”.
We find that individuals with fewer total Facebook friends (selective users) exhibit different outcomes than those with more Facebook friends (non-selective users). Specifically, our social distance measures are all positive and significantly associated with trust for selective users, perhaps due to their propensity for greater discretion in acquiring Facebook friends. For those users with more Facebook friends, we find only the number of mutually tagged photos has a positive and significant association with trust. This suggests non-selective users may understand the relatively facile nature of their
Facebook linkages and place little credence on their online interactions, as measured by the number of shared wall posts. Our metric for the number of mutual friends ( Embeddedness ) is not significant, as perhaps non-selective users know all too well the motives for acquiring a vast number of Facebook friends do not include the intent to develop meaningful friendships. As such, our results suggest that non-selective users may be more apt to only trust those friends with whom they have an offline friendship, and do not necessarily trust those loosely connected Facebook friends, who may not really be considered friends at all.
We expect our research will contribute to the academic literature by furthering work on the deployment of economic games in real-world online social networks. Our work establishes a framework for economic experiments on Facebook-type networks that includes protocols for subject selection, pairing and game play. As we develop this application into an open platform that leverages Facebook’s shared computational resources and subject pool, we expect this methodology to be fertile ground for future economic and social science research. To date the literature is silent with respect to the potential sociological and economic value of large-scale global online social networks. Thus, developing a framework for discussing these constructs, and testing their effects on behavior and decision-making, is an important aspect of our study. While there are likely subsets of user types within the population that we do not specifically explore, in general we believe our work can serve to further lower frictions and improve the efficiency of Internet based economic activity. Our results also speak to practical applications in the areas of on-line product marketing and development and suggest care be taken when identifying influential users in online social networks. Though conventional wisdom might conclude heavy users with a lot of social connections should be the most lucrative marketing targets on Facebook, our results are not consistent with this conjecture as we show greater levels of trust are more likely among smaller cohorts of selective users.
27
Our research leverages a hitherto under-appreciated aspect of large-scale online social networks. Networks such as Facebook and Twitter are based on a ‘platform’ architecture, which allows outside entities, including the research community, to build applications, interface and ultimately conduct scientific experiments using randomized trials with hundreds of millions of real-world online users in a natural environment. In a business context, we expect such methods to improve our understanding of consumer preferences, as well as pinpoint causal mechanisms underlying peer-to-peer influence and targeted marketing (Aral 2011; Aral, Muchnik and Sundararajan 2009; Bapna and Umyarov
2011). Overall, we showcase how large-scale online social networks can serve as platform to take social science and experimental economics research from the physical laboratory to a highly generalizable real-world context.
28
1.
Akerlof, George A. 1970. "The Market for 'Lemons": Quality Under Uncertainty and the
Market Mechanism."
August, 84: 488-500.
2.
Anderson, Erin, Leonard M. Lodish. and Barton A. Weitz. 1987. "Resource Allocation
Behavior Conventional Channels."
February 23: 254-62.
3.
Aral, Sinan. 2011, “Identifying Social Influence: A Comment on Opinion Leadership and
Social Contagion in New Product Diffusion.”
. 30(2): 217-223.
4.
Aral, Sinan, Walker, Dylan. 2011 “Creating Social Contagion through Viral Product
Design: A Randomized Trial of Peer Influence in Networks.
,
September, 57(9): 1623-1639.
5.
Aral, Sinan, Walker Dylan. 2012a, “Identifying Influential and Susceptible Members of
Social Networks.”
, Articles in Advance, June, 22: 1-5.
6.
Aral, S., Walker D., 2012b. “The Impact of Strength of Tie and Social Embeddedness on
Influence in Social Networks: Evidence from a Randomized Experiment” NYU working paper.
7.
Arrow, K. J. 1974.
W. W. Norton and Company.
8.
Bapna, Ravi, Umyarov, Akhmed. 2011. “Are Paid Subscriptions on Music Social Networks
Contagious? A Randomized Field Experiment,” Presented at the Workshop on
Information Systems Economics, 2011, Shanghai, available at http://users.nber.org/~confer/2012/SI2012/PRIT/Bapna_Umyarov.pdf
9.
Berg, Joyce, John Dickhaut and Kevin McCabe. 1995. "Trust, Reciprocity and Social
History.”
, Vol. 10: 122-42.
10.
Bohnet, Iris and Zeckhauser, Richard. 2004.“Trust, Risk and Betrayal.”
55(4): 467-484.
11.
Bolton, G., E. Katok and A. Ockenfels. 2005. Cooperation Among Strangers With Limited
Information About Reputation.
. 89(8) 1457.
12.
Bromiley, P. and Cummings, L.L. 1995. “Transactions costs in organizations with trust.”
In R. Bies, B. Sheppard and R. Lewicki (eds),
,
Vol. 5, pp. 219–247. Greenwich, CT: JAI Press.
13.
14.
Burt, Ron, S. 1987. “Social Contagion and Innovation: Cohesion versus Structural
Equivalence.”
, 92(6): 1287-1335.
29
15.
Burt, Ron. S. 1992.
; Cambridge,
Mass.: Harvard University Press.
16.
Camerer, C. 2003.
.
Princeton University Press, Princeton.
17.
Dwyer. F Robert, Paul H. Schurr and Sejo, Oh. 1987. "Developing Buyer-Seller
Relationships."
April 51: 11-27.
Dasgupta, P. (1988), "Trust as a Commodity", in D. Gambetta, ed.,
(Oxford: Basil Blackwell).
19.
Eckel, C. and Grossman, P. 2006. “Differences in the Economic Decisions of Men and
Women: Experimental Evidence” in
, edited by C. Plott and V. Smith. Elsevier, New York.
20.
Eckel, C. and Wilson, R. 2003. “The Human Face of Game Theory: Trust and Reciprocity in Sequential Games”, in
, edited by
Elinor Ostrom and Jimmy Walker. Russell Sage Foundation.
21.
Fershtman, C. and Gneezy, U. 2001. “Discrimination in a Segmented Society: An
Experimental Approach.”
, 115(3): 811-846.
22.
Gilbert, E. and Karahalios, K. 2009, “Predicting Tie Strength with Social Media.”
23.
Granovetter, M. 1973. “The Strength of Weak Ties”.
,
78(6).
24.
Granovetter, M. 1974. “
.” University of
Chicago Press.
25.
Ho, Teck-Hua. and Weigelt, Keith, 2005. “Trust Building Among Strangers”.
. 51(4): 519-530.
26.
Karlan, Dean. Mobius, Markus., Rosenblat, T., Szeidl, Adam. 2009. “Trust and Social
Collateral.”
124(3): 1307-1361.
27.
Leider, Stephen; Mobius, Markus; Rosenblat, Tanya; Do, Quoc-Ahn. 2010. “What Do We
Expect from Our Friends?”
March 8(1):
120-138.
28.
29.
Levin, D. and Cross, R. “The Strength of Weak Ties You Can Trust: The Mediating Role of
Trust in Effective Knowledge Transfer”.
50(11): 1477-1490.
30.
Lokshin, Michael, Sajaia, Zurab. 2004. "Maximum likelihood estimation of endogenous switching regression models," The Stata Journal, 4(3): 282-289.
31.
Maddala, GS. 1975. “Switching regression models with exogenous and endogenous switching,”
22: 423
32.
Mayer, Roger C., Davis, J. and Schoorman, F.D. An Integrative Model of Organizational
Trust.
, Vol. 20, No. 3 (Jul., 1995), pp. 709-734
33.
34.
Rice, S. 2012, “Reputation and Uncertainty in On-line Markets: An Experimental Study.”
30
35.
Rosenblat, Tanya. Mobius, Markus. 2009. “Directed Altruism and Enforced Reciprocity in
Social Networks,
, 124(4): 1815-185
36.
Sundararajan, Arun. 2012, “How Facebook Can Still Rule the Internet Economy,”
, available at http://www.bloomberg.com/news/2012-09-
04/how-facebook-can-still-rule-the-internet-economy.html
37.
Suri, S., Watts, D. J. 2010 “Cooperation and Contagion in Web-Based, Networked Public
Goods Experiments,” available at http://arxiv.org/pdf/1008.1276v2
38.
Tynan, Dan. 2011. “Facebook’s fake friends epidemic.”
, June 28, http://www.itworld.com/it-managementstrategy/178261/facebook-s-fake-friends-epidemic
31