MANAGEMENT SCIENCE

Vol. 59, No. 1, January 2013, pp. 17–35

ISSN 0025-1909 (print) — ISSN 1526-5501 (online) http://dx.doi.org/10.1287/mnsc.1120.1560
© 2013 INFORMS

Judging Borrowers by the Company They Keep:

Friendship Networks and Information Asymmetry in
Online Peer-to-Peer Lending

Mingfeng Lin
Eller College of Management, University of Arizona, Tucson, Arizona 85721, [email protected]

Nagpurnanand R. Prabhala, Siva Viswanathan

Robert H. Smith School of Business, University of Maryland, College Park, Maryland 20742
{[email protected], [email protected]}

W e study the online market for peer-to-peer (P2P) lending, in which individuals bid on unsecured
microloans sought by other individual borrowers. Using a large sample of consummated and failed listings
from the largest online P2P lending marketplace, Prosper.com, we find that the online friendships of borrowers
act as signals of credit quality. Friendships increase the probability of successful funding, lower interest rates
on funded loans, and are associated with lower ex post default rates. The economic effects of friendships show
a striking gradation based on the roles and identities of the friends. We discuss the implications of our findings
for the disintermediation of financial markets and the design of decentralized electronic markets.
Key words: peer-to-peer (P2P) lending; value of social networks; signaling; information asymmetry;
credit markets
History: Received August 31, 2010; accepted March 3, 2012, by Sandra Slaughter, information systems.
Published online in Articles in Advance September 4, 2012.

1. Introduction results show a striking gradation along friend type,

The ability of online markets to efficiently bring with greater effects when friends have roles and
together buyers and sellers has transformed busi- identities that signal better credit quality. Our evi-
nesses, spawned success stories, and redefined the dence suggests that friendships serve as informational
roles of traditional intermediaries. In this paper, we cues of a borrower’s credit quality. More generally,
study the online market for peer-to-peer (P2P) lend- they support the central premise of signaling models
ing, where individual lenders make unsecured loans that agents facing asymmetric information adapt by
to other individual borrowers. This market was vir- using signals to mitigate adverse selection. The results
tually nonexistent in 2005 but has since grown. also highlight how technology aids this process by
The biggest market, Prosper.com, has logged over facilitating the generation and transmission of new
200,000 listings seeking $1 billion in funding since its sources of information.
inception. To motivate our empirical strategy, we briefly con-
We study the relation between the online friend- sider the literature on adverse selection and signal-
ships and transactional outcomes on the P2P mar- ing. As Spence (2002) notes in his Nobel prize lecture,
ket. We find that borrowers with friends are more this strand of research is pioneered by Akerlof (1970),
likely to have their loan requests funded and that who studies the used car market. Akerlof points out
these loans have lower interest rates. We then exam- that when sellers of used cars know more about car
ine why friendships matter. Lenders may use friends quality than buyers, cars are treated as “lemons” by
as a signal because friendships serve as an informa- buyers. High-quality car sellers unable to communi-
tional cue of a borrower’s credit quality. Alternatively, cate quality cannot charge for it and withdraw from
friendships could have no pecuniary implications and the market, leading to a market failure. Spence (1973)
mislead lenders to irrational lending choices. To test argues that the Akerlof adverse selection problem can
these possibilities, we examine ex post outcomes of be mitigated if high-quality types use “signals” to
loans on the P2P market. We find that friendships communicate quality. Low-quality types do not mimic
lower default probability. Both the ex ante and ex post this behavior due to the costs of acquiring signals.
17
 Lin, Prabhala, and Viswanathan: Friendship Networks and Information Asymmetry
18 Management Science 59(1), pp. 17–35, © 2013 INFORMS

While Spence focuses on using education to signal friendships. This occurs because borrowers with no
worker productivity, the framework has led to a rich friends on Prosper.com are known only by user IDs
literature in economics.1 with identities protected for privacy concerns. How-
The signaling framework makes both ex ante and ever, the borrower identity is unmasked to friends in
ex post predictions. For instance, consider the Spence the email invitation establishing the friendship. Thus,
(1973) model in which workers signal quality through borrowers more prone to defaulting should avoid
education. Ex ante, educated workers should be forming friendships to avoid stigma costs. This makes
more likely to find employment and get paid more, friendships a credible signal of default, particularly
given that they are signaling high quality. Ex post, for friends active as lenders who may be more atten-
these workers should be more productive, given their tive to defaults on Prosper.com.
higher quality. Likewise, in the used car market, war- Figure 1 illustrates the empirical implementation
ranties could signal used car quality (Grossman 1981). of the tests based on friend types. At the top level
If so, cars with warranties should command better in Figure 1 are friends who have registered on
prices ex ante and should have fewer quality prob- Prosper.com. These friends amount to little more
lems ex post. These tests readily map to the P2P lend- than an email address. Friends who have roles pass
ing context. If friendship signals better credit quality, Prosper.com identity screens for a social security
borrowers with friends should be more likely to number, bank accounts, and driver’s license. These
attract funding at lower interest rates. Ex post, these friends can be further differentiated by roles as
borrowers should default less given their higher- lenders or borrowers. Lenders must pass minimum
credit quality. These are the baseline empirical predic- income and wealth screens. Level 3 differentiates
tions that we test. lenders between those who are merely registered and
To gain more insights into the role played by friend- those who actually have a lending history. Friendship
ships, we consider the type of friends. These tests with lenders who actually bid is a more credible sig-
can be motivated by the observation in Spence (2002) nal of quality. Levels 4 and 5 indicate borrowers with
that a signal’s effectiveness is more when the cost of lender-friends who bid on the borrower’s listing, and
acquiring it is greater. Some “friends” on Prosper.com bid and win on the listing, respectively.
are little more than (potentially fake) email addresses. The numerically higher levels of friendship in Fig-
These friends should have little signaling value. It is ure 1 should convey progressively stronger informa-
harder for borrowers to establish friendships in which tional cues of borrower quality. For instance, whereas
friends undergo screening to verify identity, credit it is costless for lower-quality borrowers to create
history, and meet the income and wealth screens ties with nonverified friends, it is progressively more
required of Prosper.com lenders. The bar is higher difficult for these borrowers to find lender-friends,
for friends with established histories of bidding and lender-friends with bidding histories, and lender-
especially successful bids that require outlay of cap- friends who are willing to personally take the risk and
ital. These friend types should have especially high invest in their loans. The stigma costs also increase
signaling value.
along this hierarchy, because defaults are likely to be
Multiplicative effects on friend type are produced
more visible to lender-friends who participate actively
by the social stigma costs of default. The personal
in the P2P market. Such costs are especially elevated
finance literature points out that a default not only
when a borrower’s friends participate in the bor-
lowers credit scores, but also imposes an additional
rower’s listing because the participating friends are
stigma cost on a borrower, which is disutility suffered
instantaneously informed when a borrower defaults.
by a defaulter when a friend learns about the default
The bottom line prediction is that as we go down
(see e.g., Thorne and Anderson 2006, Cohen-Cole
the friendship hierarchy, the signal of credit quality
and Duygan-Bump 2008).2 If stigma costs matter, bor-
conveyed by friends should be stronger and result in
rowers who are more likely to default should avoid
greater economic effects.
1
Our first empirical tests focus on two ex ante out-
See the Nobel prize address of Spence (2002), as well as
comes, the probability of successful funding and the
Weiss (1995) and Bedard (2001), on education signaling. Other
applications include warranties (Grossman 1981), capital structure interest rate of loans. Consistent with the signaling
(Ross 1977), initial public offerings (Leland and Pyle 1977), and hypothesis, friendships increase the probability of a
dividends (John and Williams 1985). successful listing and lower loan interest rates with
2
In Thorne and Anderson (2006, p. 83), survey respondents say more pronounced effects as we go down the friend-
that “0 0 0 [bankruptcy] is a mark against my name 0 0 0 it was too ship hierarchy in Figure 1. Thus, lenders in the P2P
embarrassing 0 0 0 that’s a sign of failure.” Other work in economics market do appear to judge borrowers by the (qual-
establishes stigma as an important force in individual behavior.
For instance, influential work by Moffitt (1983) and Bertrand et al. ity of) company they keep. For the ex post tests,
(2000) explains low welfare participation as a consequence of we follow the literature on consumer defaults (Gross
stigma. and Souleles 2002) and estimate Cox survival models.
Lin, Prabhala, and Viswanathan: Friendship Networks and Information Asymmetry
Management Science 59(1), pp. 17–35, © 2013 INFORMS 19

Figure 1 Hierarchy of Friends

# Friends

# Friends with # Friends with

Level 1
no roles roles

# Pure
Level 2 # Lender-friends
borrower-friends

# Potential # Real
Level 3 lender-friends
lender-friends

# Potential lender-friends # Potential # Real lender-friends # Real lender-friends

Level 4
who do not bid lender-friends who bid who bid who do not bid

# Real lender-friends # Real lender-friends

Level 5
who bid and win who bid but lose

We find that friendships lower the default hazard especially interesting because of the unusual research
along the friendship hierarchy. We also show that the context. The “information gap” between buyers and
weight placed on the friendship signal is economi- sellers that drives adverse selection (Spence 2002)
cally sensible relative to its effects on default. We con- is particularly elevated in P2P loans. This is an
duct extensive robustness tests. We estimate panel anonymous credit market with no face-to-face contact
models, model the fraction of a loan funded rather between agents. Agents lend as little as $50 apiece,
than the funding probability, funding from nonfriend less than 1% of the typical loan request. Moreover, the
lenders, subsamples of first time listings, and controls signal i.e., friendship, is subtle. Inferring its relation
for text and images. Our results are robust. to credit quality requires sophisticated economic rea-
The rest of this paper is organized as follows. Sec- soning. Yet, agents adapt remarkably in line with the
tion 2 reviews the literature and theory motivating predictions of economic theories of signaling.
our work. Section 3 summarizes our research context Our findings also add to the work on adverse
and describes the data used in the study. Sections 4 selection specifically related to credit markets. The
and 5 describe our data set and empirical methodol- finance literature (see, e.g., Gorton and Winton 2003
ogy, respectively. Section 6 contains the results of the for a review) argues that information asymmetry is
study. Section 7 discusses the robustness of the results an important feature of credit markets. In such envi-
to alternative specifications of controls, images, and ronments, gathering “soft” information about credit
text. This section also discusses a method to quantify quality beyond credit scores and standard ratios is
the value of the friendship signal. Section 8 concludes critical to successful lending outcomes (Petersen and
this paper. Rajan 2002). Almost all work in this area views finan-
cial intermediaries as repositories of such information
2. Theoretical Motivation because of their incentives, expertise, and economies
Our study draws on and contributes to research of scale and scope (Fama 1985, Granovetter 1985,
in multiple disciplines including economics, finance, Petersen and Rajan 1994, Uzzi 1999, Agarwal and
information systems, and sociology. To place our con- Hauswald 2007).
tributions in perspective, we review the relevant liter- We add to this literature in two ways. First, we
ature and discuss how our findings add to the work show that soft information can be produced and used
in these areas. without financial intermediaries. In our study, friend-
ship is borrower-generated and is used by individ-
2.1. Adverse Selection and Signaling uals putting small sums of money to work. Second,
We add to the signaling literature pioneered by we identify a new source of soft information, viz.,
Akerlof (1970) and Spence (1973). Our contribution to friends and friendship types. We show that they add
this body of work is to offer new empirical evidence to the standard hard credit variables such as Fair
on the importance of signaling in markets where Isaac Credit Organization (FICO) scores or debt-to-
agents face information asymmetry. Our evidence is income ratios used in the literature (Rajan et al. 2010,
 Lin, Prabhala, and Viswanathan: Friendship Networks and Information Asymmetry
20 Management Science 59(1), pp. 17–35, © 2013 INFORMS

Agarwal et al. 2006). We illustrate how information We also offer substantive evidence on the channels
technology hardens this soft information into usable through which social capital matters. As Granovetter
form for lenders. Our evidence also establishes a (1972) writes, social capital is conventionally concep-
counterpoint to the policy concern that decentralized tualized as an individual attribute that generates an
electronic markets lead to loss of soft information economic benefit, or as a group attribute of a col-
(Hauswald and Marquez 2003). Although it is cer- lection of individuals that enhances the transactional
tainly correct that information technology could sub- efficacy for economic gain (Coleman 1988, Mizruchi
tract some forms of soft information, our point is that 1992, Putnam 1993). Our results indicate that social
it could also facilitate production and transmission of capital between individuals plays another role: It
new sources of soft information. facilitates transactions with third parties outside the
The literature on eBay auctions also considers dyad creating the social capital. In our study, social
adverse selection, and emphasizes the role of seller capital reflected in friendships generates additional
reputation through buyer feedback in mitigating it credit information that is harvested to facilitate trans-
(see, for instance, Dellarocas 2003, Bolton et al. 2004, actions with outsiders such as nonfriend lenders in
Houser and Wooders 2006, Ghose and Ipeirotis 2011, financial markets. In the framework of Podolny (1993,
Hill et al. 2006, Resnick et al. 2006). There are several 2001) or Granovetter (1972), social connections can be
differences between the eBay literature and our study. beneficial not only because they are pipes that con-
Buyer feedback on eBay establishes seller credibility. vey resource flows to individuals, but also because
In P2P lending, standards for verifying roles and iden- the ties act as prisms, or informational cues that out-
tities establish the friend’s credibility. In the eBay con- siders can use to infer the quality of an agent.4 Our
text, the scores establishing seller reputation come results also show that such effects are enhanced when
from post-purchase feedback from buyers. The infor- the ties are credible and when outsiders have gran-
mational cue here, friendship, is not derived from ular information on the individuals forming ties. For
post-purchase feedback by the buyer (lender). Rather, instance, friendships help when the skepticism from
friendship is a characteristic of the seller known ahead lenders is mitigated by verified, credible data on the
of and formed outside the product purchase context roles and identities of the friends.
of the loan listing. There are also differences in the
product. Here, the product is not a consumable, but is 2.3. Online Networks
like a durable good whose utility (repayment) flows Our study is of separate interest because it exam-
over 36 months as the loan is repaid. Finally, our ines the economic value of online networks, an area
study focuses on the gradation along the roles and in which there is little prior research. We show
identities of the friends rather than the star rating of that online networks have economic value, espe-
the seller, i.e., the borrower. cially when data on the friendship network is made
credible and granular. Our study also addresses a
2.2. Social Capital and Economic Outcomes major limitation of the received work on online net-
A growing strand of research examines how social works, i.e., the difficulties in measuring objective out-
capital facilitates economic exchange. The literature comes, and in identifying the ties relevant to the
originates in sociology, but has attracted considerable economic decision, which necessitate costly meth-
attention in economics of labor, prices setting, produc-
ods such as surveys or interviews (e.g., Karlan 2007,
tion, innovation, and entrepreneurship (Granovetter
Moran 2005, Uzzi 1999), or accepting subjective mea-
2005, Guiso et al. 2004, Sapienza et al. 2007). We offer
sures of outcomes (e.g., Bagozzi and Dholakia 2006,
methodological and substantive contributions to this
Uzzi and Lancaster 2003). More recently, informa-
literature.
tion systems researchers are increasingly interested in
Our methodological contribution is on the identifi-
cation of social capital. Granovetter (2005) writes that
social capital is best thought of as being generated contacts 0 0 0” (p. 9). In a careful survey, Durlauf and Fafchamps
(2005) suggest that the most compelling empirical work is likely
by actions, patterns, or processes of people outside
to be based on friendships of individuals (formed outside the eco-
the economic setting being studied. The challenge is nomic context under study).
to find such metrics that reflect outside interactions. 4
A related point is made by education signaling models. As Spence
We offer such a setting. The social capital of borrowers (1973, 2002) points out, education does not need to cause produc-
i.e., their friendships, are formed outside Prosper.com. tivity to increase to serve as a useful signal. Acquiring education
We study their effect on economic outcomes, precisely can separate high and low-quality workers. Likewise, friendships
the approach favored by economists and sociologists on Prosper.com do not need to increase the physical resources of a
borrower. Friendships, graded by the roles and identities of friends,
(see Granovetter 1985).3 can serve as an informational cue that helps borrowers signal their
creditworthiness to outside lenders. We later show that friendships
3
Burt (1992) also states that a promising avenue to identify attract funding from strangers, i.e., bidders outside a borrower’s
social capital is using “friends, colleagues, and more general friendship network.
Lin, Prabhala, and Viswanathan: Friendship Networks and Information Asymmetry
Management Science 59(1), pp. 17–35, © 2013 INFORMS 21

online social networks. Studies draw from network- credit reporting agency. Prosper.com lenders are also
ing websites where network ties are explicit (e.g., Aral subject to verification of the social security number,
and Walker 2011, Susarla et al. 2012, Gnyawali et al. driver’s license number, and bank account number.
2010) or copurchase and recommendation networks To protect privacy, the true identity of borrowers and
(e.g., Oestreicher-Singer and Sundararajan 2012). In lenders is never publicly revealed on the website. All
our study, the network itself and the economic out- users are identified with user-names that are chosen
comes are quantifiable using relatively objective mea- when signing up.
sures such as funding probability, interest, or default To seek funding, a borrower makes an online list-
rates. ing, which indicates the loan amount, the maximum
Finally, we add to the literature on P2P lending. interest rate, and optionally, free format text descrip-
A small but but growing body of research in this area tion and images that are not verified by the website.
exploits data on personal characteristics of borrow- Listings can be funded as closed auctions that end
ers to test theories of taste-based discrimination. Pope as soon as the total amount bid reaches the amount
and Sydnor (2011) examine loan listings between June sought at the borrower’s asking rate. Alternatively, in
2006 and May 2007. Ravina (2008) examines listings the open format, the auction remains open for up to
for a one-month period between March 12, 2007, and seven days even if amount and rate criteria are met
April 16, 2007. Both papers focus on facial attributes to let lenders bid down the interest rate of the loan.
such as race and beauty of the borrowers, addressing The listing includes the borrower-supplied informa-
the literature on racial bias and the beauty premium tion such as amount, rates, or loan purpose as well as
(Hamermesh and Biddle 1994, Mobius and Rosenblat other hard credit data such as the number of credit
2006). We control for race and beauty in our tests, but inquiries in the last six months and a letter credit
our focus is different. We examine the role of friend- grade from AA (high quality) to HR (low quality),
ships. Our specific focus is on the use of the roles which is a coarse version of the borrower’s FICO
and identities of friendship as economic signals that score.6 The listing shows friendship data but excludes
mitigate adverse selection and information asymme- (prohibited) personal information such as phone or
try. The flavor of our findings is similar to that of Iyer address.
et al. (2009), who find that loans in the P2P market-
place seem to reflect default information beyond the 3.2. Bidding, Funding, and Repayment
traditional hard credit variables. Before bidding, lenders transfer sufficient funds
to their noninterest bearing Prosper.com account.
3. Institutional Background An individual lender can can bid an amount of $50
Our data come from the online P2P lending web- or more and specify the minimum interest rate she
site, Prosper.com, which opened on February 5, 2006. desires for the bid. The actual bidding process uses
By the end of 2008, it had 830,000 members and a proxy bidding mechanism. If the loan has not yet
over $178 million in funded loans. Borrowers are lim- been funded 100%, the ongoing interest rate is the
ited to a maximum of two concurrent loans with borrower’s asking rate, even if the lenders’ minimum
total amount less than $25,000. Loans amortize over rate is lower. After 100% of the requested funding has
a 36-month period. Loan proceeds are credited to the been reached, the auction closes if it is of closed for-
bank account from which repayments are automati- mat, but remains open for lenders to lower rates if
cally withdrawn. The rest of the section describes the it is an open format. All bids are firm commitments
lending process and information provided by borrow- with no withdrawals allowed. From a lender’s view-
ers on this network.5 point, a bid could win or be outbid, in which case
the lender can place a second bid to rejoin the auc-
3.1. Verification and Listing tion. If the loan is not fully funded by the end of the
Users join Prosper.com by providing an email auction, the request is deemed to have failed and no
address, which is verified by the website. To engage funds are transferred: No partial funding is allowed.
in a transaction, users must go through additional Successful auctions go to Prosper.com staff for
verification. Borrowers must reside in the United further review. If documentation is in order, funds
States, have a valid social security number, a valid are collected from the winning bidders’ Prosper.com
bank account number, a minimum FICO credit score accounts and transferred to the borrower’s account,
of 520, and a valid driver’s license and address. after deducting fees of up to 2% of the loan amount.
The details are verified by Prosper.com, which also Loans on Prosper.com have a fixed maturity of
extracts a credit report from Experian, a major U.S. 36 months with repayments in equated monthly

5 6
The descriptions are accurate for the time period that we study. Specifically, AA = (FICO score ≥ 760); A = 720–759; B = 680–719;
Some features have changed since then. C = 640–679; D = 600–639; E = 560–599; HR = 520–559.
 Lin, Prabhala, and Viswanathan: Friendship Networks and Information Asymmetry
22 Management Science 59(1), pp. 17–35, © 2013 INFORMS

installments that are automatically deducted from a We obtain information on borrowers’ credit histories,
borrower’s bank account and distributed to lenders’ unique Prosper IDs, friendships, and outcome of their
Prosper.com accounts. If the monthly payment is loan listings using an API provided by Prosper.com.
made on time, the loan status for that month is con- Importantly, we ensure that the descriptive fields in
sidered current. If a monthly bill is not paid, the loan our analysis are in the information set of potential
status will be changed to “late,” “one month late,” lenders. We gather information on loan requests in
“two months late,” etc. If a loan is late for two months real time so that information about borrowers’ friends
or more, it is sent to a collection agency. Lenders is current at the time of the loan requests. We describe
on Prosper.com must agree that the proceeds of the the variables used in our analysis and discuss some
collection represent the full settlement of loans. Delin- descriptive statistics.
quencies are reported to credit report agencies and
can affect borrowers’ credit scores. Borrowers who 4.1. Friendships
default on their loans are not allowed to borrow using In our data set, 56,584 listings report friends. A key
Prosper.com again. focus of our analysis is friend type. Figure 1 describes
the hierarchical levels of the roles and identities of
3.3. Friendships and the Company That friends underlying our analysis. Level 1 distinguishes
Borrowers Keep friends according to whether their identities are veri-
Any Prosper.com member with a verified email fied on Prosper.com or they are mere email addresses.
account can create or join a friendship network. Level 2 categorizes the verified friends based on their
To form friendships, the inviting member fills out specific roles as borrowers or lenders. Lenders are
the friend’s email address and a short message on individuals with extra financial capital. Level 3 fur-
Prosper.com. Prosper.com then generates an email ther differentiates lender-friends by whether they are
message with a link that the recipient can click to real lender-friends who have lent to other borrowers
establish a friendship. Thus, individuals who are before the current listing. Level 4 differentiates real
friends on Prosper.com have at least some offline, lender-friends according to whether they bid on the
nonpublic information about each other, such as an specific borrower’s listing. Level 5, the finest classifi-
email address. Although Prosper.com users are nor- cation, distinguishes between lender-friends who bid
mally identified by user IDs (e.g., “banker234”), this on the borrower’s listing and won and those who bid
situation changes with friends. Members on either but did not win. As we progress from level 1 to level 5,
end of a friendship tie know the real person behind the relationship between the borrower and lender
the ID. Thus, friends can link user IDs of a borrower strengthens. The difficulty in mimicking or “faking”
who defaults to the actual identities of the defaulter, a friend type, i.e., setting one up where one does not
potentially imposing social stigma costs on borrowers exist, becomes greater as we go from level 1 to level 5.
with friends.
From an empirical viewpoint, the important point 4.2. Control Variables
is that a member’s friend information is highly visible Table 1 lists the control variables included in our
on the members’ profile pages. Friendship data are analysis. Among the hard credit variables is the
prominently displayed in a listing. Indeed, it is one Prosper.com letter grade for each borrower, which
of the most prominent pieces of information outside ranges from AA to HR as described previously. Rather
the credit information and listing data about the bor- than a numerical score (e.g., AA = 1, A = 2, etc.), we
rower. Friends who bid on a listing are also tagged include a full set of dummy variables for each letter
very clearly by a special icon in the list of bids, so they grade. We also include other hard credit information
are readily visible to other potential bidders. Data on on the listing such as a borrowers’ debt-to-income
friend types are also accessible in a straightforward ratio and the number of credit inquires in the six
way by clicking on links to see the profile of friends. months before the listing.8 These variables allow for
the possibility that the letter grade itself is not a suf-
ficient statistic for credit risk.
4. Data Set We include an extensive set of controls in our anal-
Our sample comprises all listings that seek funding
ysis. One of them is auction type because closed
on Prosper.com between January 2007 and May 2008.7
auctions are likely to indicate that the borrower has
more urgent financial needs. Some states have usury
7
An advantage of this time frame is that website features
remain largely consistent throughout this period. In October 2008,
Prosper.com shut down and started a registration process with friendship information remained unchanged, but there is too much
the U.S. Securities and Exchange Commission. When they fully noise to combine data generated before 2008 and that generated
reopened over a year later, some features were introduced or after 2009.
8
modified, partly to comply with SEC regulations. Interestingly, This does not include requesting loans on Prosper.com.
Lin, Prabhala, and Viswanathan: Friendship Networks and Information Asymmetry
Management Science 59(1), pp. 17–35, © 2013 INFORMS 23

Table 1 Main Variables and Their Descriptions

Hard credit information

Hard credit information variables include the following information about the borrower at the time of listing:
• A series of dummy variables indicating borrower’s credit grade (CreditGradeAA, CreditGradeA, CreditGradeB, CreditGradeC, CreditGradeD, CreditGradeE,
CreditGradeHR).
• Borrower’s debt-to-income ratio (dti).
• Borrower’s bank card utilization (bankcardutilization), or the percentage of credit line that the borrower has used; quadratic term of bank card utilization.
• Number of credit inquiries on borrower’s credit report in the six months before listing (InquiriesLast6month).
• Borrower’s credit history length (YearsSinceFirstCredit), or the number of years between borrower’s first credit line and the time of listing.
Auction characteristics
Auction characteristics variables include the following:
• Auction format (auctionformat), or whether the auction is “immediate funding” or “open for duration”.
• The maximum interest rate specified by the borrower (borrowermaxrate) and its quadratic term.
• Loan amount requested by the borrower (amountrequested).
• Logarithm of the total length of text on borrower’s listing (logttltext).
• A series of dummy variables indicating the loan category that the borrower specified (listingcatX), where X = 0 (no category specified—this is the baseline
and not included in regressions), 1 (debt consolidations), 2 (home improvements), 3 (personal loans), 4 (business loans), 5 (student loans), 6 (automobile
loans), and 7 (other loans).
• A series of dummies for the duration of auctions (durationY), where Y = 31 51 7, or 10; Duration3 is the baseline dummy and not included in the regressions.
• Borrower’s closing fees charged by Prosper.com at time of listing (BorrowerFee).
• Lender’s closing fees charged by Prosper.com at time of listing (LenderFee).
Variables about groups
• Logarithm of the number of members in the group that the borrower belongs to at time of listing (loggroupsize).
• Dummy variable for whether the borrower’s group leader is awarded for successful borrowing (1 if yes, 0 otherwise) (groupleaderrewarded).
• A series of dummy variables indicating the nature of the group that the borrower belongs to: _Alumni for alumni groups related to a school or an employer;
_Geography for geographically-oriented groups; _Military for groups for military members and their families; _Medical for groups related to medical reasons
for borrowing or lending; _Demographic for groups highlighting particular demographics, such as single parents or Hispanics; _Hobbies for groups targeting
people with particular hobbies; _Religion for religiously oriented groups; and _Business for groups highlighting small businesses, or business development.
Variables about borrowers’ friendship ties
• Number of friends of the borrower (ttlFriends).
• Number of borrower’s friends whose identities have been verified (either as borrower or lender) (ttlRole).
• Number of borrower’s friends whose identities have not yet been verified (neither borrower nor lender) (ttlNorole); note that ttlFriends = ttlRole + ttlNorole.
• Number of borrower’s friends who are borrowers and not lenders (ttlpureborrow).
• Number of borrower’s friends who are lenders (ttllend); note that ttlRole = ttlpureborrower + ttllend.
• Number of borrower’s friends who are real lenders, i.e., friends who are not only lenders but also had lent to other borrowers on Prosper.com before the
current borrower’s auction (ttlreallend).
• Number of borrower’s friends who are potential lenders, i.e., friends who are signed up as lenders but have yet to make a loan at the time of borrower’s
auction (ttlpotentlend). Note that ttllend = ttlreallend + ttlpotentlend.
• Number of borrower’s real lender-friends who place bids on the current listing/auction (ttlrealbid).
• Number of borrower’s real lender-friends who do not bid on the current listing (ttlrealnobid); note that ttlreallend = ttlrealbid + ttlrealnobid.
• Number of borrower’s potential lender-friends who place bids on the current listing (ttlpotentbid).
• Number of borrower’s potential lender-friends who do not place bids on the current listing (ttlpotentnobid).
• Number of borrower’s real lender-friends whose bids are among the final winning bids (i.e., those who actually became one of the funding lenders to the
borrower) (ttlrealbidwin).
• Number of borrowers’ real lender-friends who bid on the listing, but are outbid, and therefore are not contributing to the final loan; note that ttlrealbid =
ttlrealbidwin + ttlrealbidlose.
Additional control variables
• A dummy variable for whether the borrower is constrained by usury laws (i.e., the maximum rate allowable) at the time of listing (usurystate).
• The average interest rate on a 36-month consumer loan from a bank in the same market as the borrower, in the same month as the time of listing, and in
the same credit grade range as the borrower (bankrate).
• A dummy variable for abnormal search activities on Google for Prosper.com (spikedays).
• A dummy variable for whether the borrower also had a lender role on his/her profile (lendrrole).
• A dummy variable for whether the borrower is also a group leader (leaderrole).

laws that enforce a cap on the maximum interest started collaboration with a bank in an effort to cir-
rates on consumer loans. Although usury laws are cumvent that limit. Our sample spans both periods,
intended to protect customers, they could reduce the so we include a control for listings if a usury law is in
chances of successful funding if the supply curve effect. Each borrower indicates a maximum borrow-
intersects the demand curve at a rate above a state’s ing rate that he or she is willing to pay. Whereas low
usury limit. Whether the laws have this bite is an rates indicate less profitable loans, high rates could
empirical issue. After April 15, 2008, Prosper.com also indicate less profitable loans because of greater
 Lin, Prabhala, and Viswanathan: Friendship Networks and Information Asymmetry
24 Management Science 59(1), pp. 17–35, © 2013 INFORMS

likelihood of default for borrowers willing to pay high roles and identities of friends. These tests examine
rates (Stiglitz and Weiss 1981). Thus, we include this if the effects on transactional outcomes are stronger
term in quadratic form. We control for broad lending when borrowers have better quality friends. Affirma-
rates through a proprietary data set from a profes- tive evidence would support the joint hypothesis that
sional company. The data include the average inter- (i) investors rationally adapt to informational asym-
est rate for borrowers in each credit grade in each metry by relying on signals of credit quality; and
regional market for each month on 36-month loans. (ii) a borrower’s friendship ties provide such a signal.
This variable proxies for outside borrowing and lend- While such a role of friendships, and friendship types,
ing options controlling for temporal and regional vari- can be viewed as part of our null hypothesis, the vari-
ations in consumer lending. ables are equally well motivated by economic theories
We further control for the purpose of loans. Bor- of social stigma in the context of defaults (Crocker
rowers indicate several types of needs, including debt et al. 1998, Thorne and Anderson 2006).
consolidation, home improvement, business loans, To gain further insights on why friendships mat-
personal loans for a variety of purposes (including ter, specifically whether they are useful because they
vacations), student loans, or auto loans. The loan pur- harbor information about default, we conduct ex post
pose is self-indicated by borrowers and can thus be performance tests. We test whether friendship ties
thought of as cheap talk. However, potential lenders are associated with lower ex post defaults in funded
may communicate with borrowers during the auc- loans. Furthermore, we test whether this association
tion process and seek more tangible details. On bal- follows a gradation along friend type, i.e., whether
ance, it is likely that there is some information in the better quality friendships that delve deeper into the
loan purpose, so we include this in the regressions. hierarchy in Figure 1 are associated with lower rates
Prosper.com has received significant media exposure of default. Following the consumer finance literature
since its inception. Articles in the media make it (e.g., Gross and Souleles 2002), we model time-to-
more likely for the website to attract new borrowers default using a flexible Cox hazard model. This test
and lenders after their publication. To proxy for this represents an out-of-sample test of the adverse selec-
effect, we download the search volume on Google for tion hypothesis implemented on a disjoint database
Prosper.com and construct a dummy variable based far removed from transactional data on loan funding
on whether there is a significant change in search or interest rates.
volume, which we call spikedays.9 Finally, we include A familiar concern in empirical modeling is endo-
quarterly fixed effects to control for unobserved time geneity. One concern is reverse causality (e.g., Doreian
effects. 2001 in the context of networks). Although this is a
key issue in the traditional peer effects literature (e.g.,
Manski 1993, Kremer and Levy 2008), it is less rel-
5. Empirical Modeling and evant in our context. The key question in the peer
Identification effects literature is whether an individual’s peers have
The theoretical motivation for our study comes from causal effects on behavioral outcomes. The reverse
economic theories such as Spence (1973), which posit causality issue is that an individual may select peers
signaling as a mechanism to alleviate asymmetric based on shared traits or preferences. As Manski
information and adverse selection. These problems (1993) points out, the reflection problem makes it dif-
are especially elevated in P2P lending where lenders ficult to separate peer and contextual effects.10 This
are small and have neither the close relationships is less important in our context where borrowers
nor the sophisticated analytics of institutions such form friendships outside the P2P platform, and the
as banks. The key question is whether a borrower’s actions leading to loan success largely come from
friendship ties help potential lenders adapt by acting individuals outside a borrower’s friendship network.
as informational cues or signals of credit quality. If so, The overwhelming majority of lenders are arm’s-
a testable hypothesis is that friendship ties should be length (stranger) lenders funding small portions of
associated with better ex ante outcomes, or greater a borrower’s request. Moreover, the possibility of
probability of a successful listing and lower interest lending leading to friendships is remote because all
rates on consummated loans.
We begin by testing the relation between ex ante 10
For instance, a regression of a student’s outcomes on peers’
outcomes and the number of friends. To further iden- outcomes commingles two confounding effects: A high-achieving
tify the economic channels through which friendship student may choose peers who are also high achieving, or high
acts, we next focus on the types of friends based on achieving peers may causally improve a student’s own perfor-
mance. The reflection problem (Manski 1993) is not relevant here as
we do not examine the behavior of friends in a peer friendship net-
9
This variable can serve as an exclusion restriction for one of our work. We examine the behavior of small, anonymous, arm’s-length
empirical models. We discuss this in §6.1. lenders who respond to friendship ties of borrowers.
Lin, Prabhala, and Viswanathan: Friendship Networks and Information Asymmetry
Management Science 59(1), pp. 17–35, © 2013 INFORMS 25

operational transactions such as fund transfers and Table 2 Estimated Models

ex post monitoring of repayments are carried out by
Variable set
Prosper.com. There is no room for borrower-lender
interaction after origination, preserving the mutual Model 1 2 3 4 5 6
anonymity of borrowers and lenders. These factors Funding Spec. P1 Spec. P2 Spec. P3 Spec. P4 Spec. P5
mitigate concerns about reverse causality, or the view probability
that loans are made to form friendships.11 Interest rate Spec. H1 Spec. H2 Spec. H3 Spec. H4 Spec. H5 Spec. H6
Unobservables represent a second concern. Do our Loan default Spec. C1 Spec. C2 Spec. C3 Spec. C4 Spec. C5 Spec. C6
estimates pick up the effect of friendship variables or Note: The sets of variables used in each model are described in Table 3.
unobservable variables used by agents that happen to
be correlated with friendships? Our setting mitigates friendship or particularly grades of friendships based
this concern about unobservables. The unobservable on their roles. Nevertheless, we consider them in
argument is most credible when the econometrician multiple ways. We analyze listings with or without
does not have access to as much data as the eco- images. We also examine and encode the descriptive
nomic agents engaged in decision making. In our text in listings using methods derived from com-
study, loans are funded by aggregated contributions putational linguistics. Our friendship results remain
from many small lenders, virtually all of whom are robust.
strangers with no private information about the cred-
itworthiness of the borrowers.12 We have access to the 6. Results
complete vector of information that a potential lender Our sample has 205,132 listings seeking an aver-
sees about a borrower. Thus, if a variable is unobserv- age of $6,973. We find that 56,584 (27.58%) listings
able to us, it is also unobservable to potential lenders. report friends. Listings showing friends are spread
This identification strategy is consistent with Angrist across credit grades. For instance, of the 6,523 AA
(1998, p. 251), who writes that he has access to infor- listings, 1,881 or 28.84% have friends, whereas 9,462
mation about “most of the characteristics used by the of the 33,068 D grade listings (28.61%) show friends.
military to screen applicants,” and thus can elimi- In the HR category, 22,556 (26.39%) of 62,904 list-
nate bias due to unobservables. We have a similar ings show friends. Listings in which borrowers have
or an even stronger setting as we observe (and con- no friends have mean debt-to-income ratios of 58%,
trol) for all information available to potential lenders. whereas those with friends have debt-to-income ratios
As Jackson (2010, p. 519) writes, this strategy is justi- of 57%. Borrowers with no friends have about 4.17
fiable in studies of networks if we “0 0 0 observe all of credit inquiries in the six-month period before the
the relevant factors affecting behavior.” listing date against 4.22 inquiries for borrowers with
Our additional tests further mitigate concerns about friends.13 In our data, 16,500 (8.04%) listings attract
unobservables. Most importantly, our empirical strat- full funding. For the sample of borrowers with no
egy does not rely on friendships alone, but also on friends, 10,410 of 148,548 listings, or 7.04%, are suc-
friendship types, which identify economic effects by cessfully funded. By contrast, the success rate for list-
exploiting the differences among friendships based on ings with friends is 10.76%.
friends’ roles and identities. We also consider several We explore loan funding probability, interest rates
nonstandard controls based on information available and loan default in multivariate specifications. Table 1
to potential lenders, but not included in conventional provides a full list of the explanatory variables.
studies of credit markets. These include images and Tables 2 and 3 describe the different models that we
descriptive texts contained in many Prosper.com list- report in the paper and the set of variables used
ings. As these variables are self-reported and thus in each specification. For instance, specification P1
cheap talk, it is not clear that they could subsume is a model of funding probability that uses variable
set 1 (Table 3). From Table 3 we can see that vari-
11
able set 1 corresponds to the root level of the friend-
Prosper.com also explicitly disallows borrowers to reveal person-
ally identifiable information in loan requests. This includes email ship hierarchy and uses ttlfriends, or the number of
addresses, which are necessary to create friendship ties. In unre- friends, plus the common variables listed in Table 1 as
ported robustness tests, we obtain highly consistent findings on explanatory variables. Section 6.1 reports the ex ante
friendship variables after either excluding borrowers who had outcomes, funding probability, and interest rates. Sec-
loans before our sampling time frame, or adding a variable indicat- tion 6.2 models the probability of default. Each of
ing the number of loans that they obtained before our time frame.
12
these sections focuses on the friendship variables, and
In more than 95% of the listings associated with friends, contri-
§6.3 discusses the coefficients for control variables.
butions from friends account for less than 4.4% of the funds that
borrowers receive. In other words, more than 95% of funds for
13
over 95% of listings come from lenders who are strangers to the These are credit applications to banks or credit card companies,
borrower. and do not include loan requests made on Prosper.com.
 Lin, Prabhala, and Viswanathan: Friendship Networks and Information Asymmetry
26 Management Science 59(1), pp. 17–35, © 2013 INFORMS

Table 3 Variable Sets Used in the Models

Corresponding level of the

Variable sets friendship hierarchy Common variables Additional variables

1 Root level ttlFriends

2 1 Hard credit information ttlNoRole, ttlRole
3 2 ttlNoRole, ttlPureBorrow, ttlLend
4 3 Auction characteristics ttlNoRole, ttlPureBorrow,
ttlPotentLend, ttlRealLend
5 4 Social network information—Groups ttlNoRole, ttlPureBorrow,
ttlPotentLend, ttlRealNoBid,
ttlRealBid
6 5 Additional control variables ttlNoRole, ttlPureBorrow,
ttlPotentLend, ttlRealNoBid,
ttlRealBidWin, ttlRealBidLose

6.1. Funding Probability and Interest Rates with roles as borrowers and those with roles as
Table 4 reports estimates of a Probit model for the lenders, both adding up to the total number of
probability that a listing is successfully funded.14 If x friends with roles. We also include the total num-
is a vector of information about a listing, we estimate ber of friends with no roles. Specification P3 gives
the results. Friends with no verified roles have nega-
Probability4Funded = 1 — x5
tive effects as before. Connections to borrowers have
= ê41 Hard Credit+2 Friendship Ties+3 Controls51 insignificant effects while having friends with roles
as lenders increases the probability of the loan being
(1)
funded.
where ê denotes the standard cumulative normal dis- Specification P4 further differentiates between real
tribution. Specification P1 in Table 4 shows that the lender-friends, who have made loans on Prosper.com
number of friends is positively related to the proba- to other borrowers before the current listing, and
bility of successful funding and is significant at 1%. potential lender-friends who have not yet made loans
However as we discuss below, this relation reflects a on Prosper.com before the start of the current list-
more extensive relation based on the roles and identi- ing. We continue to include other friendship variables,
ties of the friends.15 Specification P2 in Table 4 distin- including all friends with no roles, and friends who
guishes friends according to whether their identities are borrowers but not lenders, as control variables.
are verified on Prosper.com. This process effectively There is a continued gradation of the friendship
decomposes a borrower’s number of friends into two effects. Having lender-friends matters only to the
orthogonal pieces, friends who are verified and those extent that the friends are real lenders who have
who are not. We find that unverified connections,
already lent to other borrowers. The coefficient almost
i.e., connections that merely signify a valid email
doubles relative to that for the total number of lender-
address, represent insignificant cheap talk or even
friends, and remains statistically significant at the 1%
negative signals at the 10% significance level. In con-
trast, ttlrole, which denotes friends with verified roles, level. Holding all other variables at their median, we
has a positive coefficient that is significant at 1%. find that those with one real lender-friend have an
These results constitute the first evidence that roles estimated funding probability of 3% versus 1.4% for
and identities, or the nature of the company that bor- those with no real lender-friends. The funding proba-
rowers keep, matter. bility has more than doubled.
We next categorize friends based on their roles on Specification P5 decomposes real lender-friends
Prosper.com. To this end, we decompose the verified into those who bid on the specific borrower’s listing
friends into orthogonal and additive pieces: friends and those who do not. At this level, it is also pos-
sible that a potential lender who has not lent in the
14
We obtain similar results with a logit model. past now chooses to initiate bidding with the current
15
In the literature on social networks, this simple count of friends is loan. Thus, we decompose both potential lenders and
the degree centrality of a borrower. Other network metrics include real (past) lenders into those who bid on the current
coreness, effective size of network, eigenvector centrality, and effi- listing and those who do not. We find positive and
ciency (Hanneman and Riddle 2005). These notions have limited
significant effects for potential lenders who bid on the
relevance in our context due to the fragmented nature of the net-
work with several disconnected components. We discuss this issue current listing. Interestingly, borrowers whose real or
further in the robustness tests. potential lender-friends do not bid are less likely to
Lin, Prabhala, and Viswanathan: Friendship Networks and Information Asymmetry
Management Science 59(1), pp. 17–35, © 2013 INFORMS 27

Table 4 Probability of Funding

Spec. P1 Spec. P2 Spec. P3 Spec. P4 Spec. P5

∗∗∗
# Friends 00033
4000085
# Friends who are neither borrowers nor lenders −00020∗ −00020∗∗ −00017∗ −00017
4000105 4000105 4000105 4000105
# Friends who are either borrowers or lenders 00106∗∗∗
4000175
# Pure borrower-friends −00006 −00002 00018
4000235 4000215 4000185
# Lender-friends 00170∗∗∗
4000235
# Potential lender-friends 00025
4000225
# Real lender-friends 00312∗∗∗
4000555
# Potential lender-friends who did not bid −00062∗∗
4000285
# Potential lender-friends who bid 00325∗∗∗
4000505
# Real lender-friends who bid 00849∗∗∗
4000445
# Real lender-friends who did not bid −00148∗∗∗
4000225
Credit grade – A (dummy) −00375∗∗∗ −00372∗∗∗ −00372∗∗∗ −00373∗∗∗ −00381∗∗∗
4000375 4000375 4000375 4000375 4000365
Credit grade – B (dummy) −00806∗∗∗ −00805∗∗∗ −00805∗∗∗ −00805∗∗∗ −00814∗∗∗
4000625 4000615 4000615 4000615 4000635
Credit grade – C (dummy) −10457∗∗∗ −10455∗∗∗ −10456∗∗∗ −10457∗∗∗ −10470∗∗∗
4000565 4000565 4000565 4000565 4000575
Credit grade – D (dummy) −20105∗∗∗ −20102∗∗∗ −20107∗∗∗ −20109∗∗∗ −20133∗∗∗
4000945 4000945 4000935 4000945 4000945
Credit grade – E (dummy) −20831∗∗∗ −20825∗∗∗ −20833∗∗∗ −20837∗∗∗ −20867∗∗∗
4001395 4001385 4001385 4001395 4001395
Credit grade – HR (dummy) −30310∗∗∗ −30304∗∗∗ −30312∗∗∗ −30318∗∗∗ −30354∗∗∗
4001325 4001315 4001315 4001315 4001315
Debt-to-income ratio −00102∗∗∗ −00102∗∗∗ −00102∗∗∗ −00103∗∗∗ −00106∗∗∗
4000055 4000055 4000055 4000055 4000055
Group leader is rewarded (dummy) 00141∗∗∗ 00145∗∗∗ 00146∗∗∗ 00145∗∗∗ 00141∗∗∗
4000265 4000255 4000255 4000255 4000255
(Other controls) (Included in estimation)
N 205,131 205,131 205,131 205,131 205,131

Notes. This table reports estimates of a probit specification in which the dependent variable is 1 if a listing on Prosper.com is funded and 0 otherwise. The
explanatory variables include a borrower’s hard credit variables, social network variables, group affiliation, and other characteristics of the loan, the loan
domicile, and the borrower plus quarterly time period fixed effects. For brevity, some covariates are not shown. Table 1 gives the detailed definitions of the
variables. Robust standard errors are in parentheses.
∗
p < 001; ∗∗ p < 0005; ∗∗∗ p < 0001.

get funded.16 In contrast, having real lender-friends into those who win or lose as the bid status is not
bid on a listing elevates the chances of a successful known to lenders, but only observable after the out-
funding. We do not further decompose real bidders come is known.
We next examine the price effects of friendship, i.e.,
16
Bidders seem wary of passive friends with no roles or those who whether it lowers interest rates. We regress interest
do not participate in bidding, consistent with our main point that rates on hard credit variables, controls, and friend-
the types and actions of friends matter more than friends per se. ship ties for loans that are successfully funded using
We remain cautious, however, because of the mixed significance of
Heckman’s (1979) familiar two-step method:
the passive friend variables and their insignificance in the default
models of Table 6. We instead stress the significant and consistent E4r — x1funded = 15 = ƒ1 Hard Credit+ƒ2 Friendship Ties
results for lender-friends, particularly those who bid and win in
the hierarchy of Figure 1. +ƒ3 Controls+‹4x3 51
ˆ (2)
 Lin, Prabhala, and Viswanathan: Friendship Networks and Information Asymmetry
28 Management Science 59(1), pp. 17–35, © 2013 INFORMS

Table 5 Interest Rate on Funded Listings

Spec. H1 Spec. H2 Spec. H3 Spec. H4 Spec. H5 Spec. H6

∗∗∗
# Friends −00002
4000015
# Friends who are neither 00002∗∗∗ 00002∗∗∗ 00002∗∗∗ 00001∗∗∗ 00001∗∗∗
borrowers nor lenders 4000015 4000015 4000005 4000005 4000005
# Friends who are either −00005∗∗∗
borrowers or lenders 4000015
# Pure borrower-friends 00000 −00000 −00001 −00001
4000015 4000015 4000015 4000015
# Lender-friends −00006∗∗∗
4000015
# Potential lender-friends −00001
4000015
# Real lender-friends −00007∗∗∗
4000015
# Potential lender-friends 00002∗∗ 00002∗∗
who did not bid 4000015 4000015
# Potential lender-friends −00008∗∗∗ −00008∗∗∗
who bid 4000015 4000015
# Real lender-friends who −00007∗∗∗
bid 4000015
# Real lender-friends who 00002 00001
did not bid 4000015 4000015
# Real lender-friends who −00006∗∗∗
bid and won 4000015
# Real lender-friends who −00006∗∗∗
bid but lost 4000025
(Other controls) (Included in estimation)
Inverse mills ratio −00081∗∗∗ −00073∗∗∗ −00064∗∗∗ −00047∗∗∗ −00016∗∗∗ −00013∗∗∗
4000075 4000065 4000055 4000035 4000025 4000025
Selection equation: All variables used but not reported for conciseness
Spikedays −00050∗∗ −00053∗∗∗ −00053∗∗∗ −00054∗∗∗ −00052∗∗∗ −00051∗∗
4000205 4000205 4000205 4000205 4000205 4000205
N 205,132 205,132 205,132 205,132 205,132 205,132

Notes. This table reports two-stage estimates of a model in which the dependent variable is the interest rate on Prosper.com listings that are successfully
funded. The probit selection equation models the probability of a listing being successfully funded. The explanatory variables include a borrower’s hard credit
variables, social network variables, group affiliation, and other characteristics of the loan, the loan domicile, and the borrower plus quarterly time period fixed
effects. Because of page limits, we only report results on the main variables and do not report the first-stage probit equation estimates, which are consistent
with the estimates in Table 4. Robust standard errors are in parentheses.
∗
p < 001; ∗∗ p < 0005; ∗∗∗ p < 0001.

where r is the interest rate at which the loan is The interest rate results in Table 5 are remark-
funded. The the inverse Mills ratio ‹4x3 5,ˆ based on ably consistent with those for funding probability.
Probit estimates ˆ from Equation (1), accounts for the The coefficients for the friend types show direction
fact that interest rates are only observed when listings and gradation consistent with the results for fund-
are successfully funded. Because ‹4·5 is a nonlinear ing probability. Connections to friends not verified
function, Equation (2) can be identified without exclu-
sion restrictions (e.g., Overby and Jap 2009, Uzzi
Google Trends for “Prosper.com.” Spikedays is a dummy variable
1999). Alternatively or additionally, one could have an for whether the listing started in a week when the search vol-
exclusion restriction with a variable in the Probit step ume was above the 75th percentile in our sample period (3.55).
not included in the interest rate regression. We esti- When Spikedays is high, Prosper.com traffic volumes increase and
mated the model using both methods and the results borrowing activities increase, but lenders adjust slower because of
are similar.17 delays due to bank verification and funds transfer, and because old
bids cannot be withdrawn as they are firm commitments. We find
that Spikedays has a negative coefficient with F -statistic of 50, well
17
For the exclusion restriction, we consider a variable, Spikedays. above the cutoff of 10 for the Staiger and Stock (1997) criterion for
We download the relative scaling results for the search volume on a strong instrument.
Lin, Prabhala, and Viswanathan: Friendship Networks and Information Asymmetry
Management Science 59(1), pp. 17–35, © 2013 INFORMS 29

Table 6 Time-to-Default of Successful Listings

Spec. C1 Spec. C2 Spec. C3 Spec. C4 Spec. C5 Spec. C6

# Friends 10017
4000215
# Friends who are neither 10048 10048 10047 10047 10047
borrowers nor lenders 4000315 4000315 4000315 4000315 4000315
# Friends who are either 00968
borrowers or lenders 4000275
# Pure borrower-friends 10061 10061 10058 10055
4000555 4000555 4000555 4000535
# Lender-friends 00912∗∗
4000345
# Potential lender-friends 00950
4000615
# Real lender-friends 00877∗∗∗
4000445
# Potential lender-friends 00964 00964
who did not bid 4000735 4000715
# Potential lender-friends 00910 00916
who bid 4001505 4001505
# Real lender-friends who 00856∗∗
bid 4000525
# Real lender-friends who 00938 00938
did not bid 4001135 4001135
# Real lender-friends who 00791∗∗∗
bid and won 4000625
# Real lender-friends who 10086
bid but lost 4001465
(Other controls) (Included in estimation)

Notes. This table reports hazards ratio estimates of a Cox proportional hazards model of the time to default for borrower listings that are successfully funded
on Prosper.com. The explanatory variables include a borrower’s hard credit variables, social network variables, group affiliation, and other characteristics of
the loan, the loan domicile, and the borrower plus quarterly time period fixed effects. Because of page limits, we only report results on the main variables.
Table 1 gives the detailed definitions of the variables. The table reports the exponentiated coefficients (hazards ratio), where values greater than 1 indicate that
a higher value of the explanatory variable increases the risk of default. Robust standard errors are in parentheses.
∗
p < 001; ∗∗ p < 0005; ∗∗∗ p < 0001.

by Prosper.com tend to increase interest rates, as et al. 2007). In this model, the hazard h4t5 is speci-
reflected by the coefficient for the variable ttlNoRole: fied as
On average, an additional unverified friend increases h4t — x5 = h0 4t5exp4x‚5 (3)
the interest rate by 20 basis points (or 0.20%). More
important, friends with lender roles have the opposite where h0 4t5 is a baseline hazard rate, and x denotes
effect: Having a lender-friend decreases the interest a vector of covariates. For each covariate xj in the
rate by about 60 basis points. Interest rates fall the Cox model, we report the exponentiated form of the
most when real lender-friends who have participated coefficient ‚, which is called the hazards ratio, whose
in past loans on Prosper.com also participate in the standard error is obtained using the Delta method.
current listings. These effects are significant regardless A hazards ratio greater than 1.0 for variable xj indi-
of whether they win in the listing. cates that it is positively associated with the probabil-
ity of default, whereas a ratio less than 1.0 indicates
6.2. Loan Defaults that xj is negatively associated with the probability of
Prosper.com records the status of loans in each default.
month, or payment cycle. Loans are current if repay- Table 6 reports the Cox model estimates in odds
ments occur on time. Otherwise, loans can be “late,” ratios. We use coefficients ‚ to construct a base-
“one month late,” “two months late,” and so on. We line hazard function (Kalbfleisch and Prentice 2002),
model a default as occurring if a payment is late by which increases initially, peaks at about 10 months,
at least two months. We follow the consumer finance and then slowly wears off. This pattern is remark-
literature (e.g., Gross and Souleles 2002) and model ably consistent with consumer lending delinquencies
the time-to-default using a Cox proportional hazards analyzed in Gross and Souleles (2002). In specifica-
specification (e.g., Grover et al. 1997, Bhattacharjee tion C1, the total number of friends is insignificant
 Lin, Prabhala, and Viswanathan: Friendship Networks and Information Asymmetry
30 Management Science 59(1), pp. 17–35, © 2013 INFORMS

as a predictor of default. Specification C2 decom- proxy for the borrower’s age and credit experience. It
poses friends into those with verified identities as has small effects on funding probability and interest
lenders or borrowers and friends with no verifica- rate and no effect on default.
tion. Having more unverified friends is associated Auctions that close upon funding can encourage
with an increased odds of default, as indicated by a aggressive early bidding, but also higher interest
hazards ratio of 1.05, and having friends with veri- rates as there are fewer opportunities for lenders to
fied identity is associated with lower odds of default. bid down the interest rate. The results in Tables 4
However, neither variable is statistically significant. and 5 support this view: Closed auctions result in
Specification C3 shows statistically significant effects higher funding probability, but higher interest rates.
for Prosper.com verified friends who are lenders. The We also examine loan purpose. Specifying some pur-
exponentiated hazards ratio of 0.91 shows that bor- pose appears to lend credibility to listings as it
rowers with a lender-friend are about 9% (1−0091) increases overall probability of funding and lowers
less likely to default than those without. interest rates. Business loans appear to be more risky.
Specification C4 includes the number of lender- They are less likely to be funded, attract higher inter-
friends, but controls for whether they actually partic- est rates, and are about 24% more likely to default at
ipated in lending before the borrower’s listing. The a 10% significance level. Debt consolidation loans are
hazards ratio for real lender-friends is 0.88, indicat- more likely to be funded than other loans at lower
ing that having real lender-friends is associated with interest rates. Lenders value the fact that borrowers
lower probabilities of default. Likewise, the hazards are using Prosper.com to shop interest rates or limit
ratio is 0.86 when we consider lender-friends who credit card debt. These loans are about as likely to
bid on the borrower’s listing. Both coefficients are default as other loans.
significant at 1%. The hazards ratio for friends who Borrowers willing to pay low rates may be less
bid on and win a listing is 0.79 and is significant at profitable to lenders and may be less likely to be
1%. Thus, the odds of default are significantly lower funded. However, in the spirit of credit rationing the-
when lender-friends bid and win on the borrower’s ories (Stiglitz and Weiss 1981), high rates may sig-
listing. nal risky borrowers and deter bidding. Our results
The result for the real lenders who bid indicate that support this view. The linear term has a positive
financial stakes taken by friends are strong signals coefficient and the quadratic term has a negative
for outside lenders that a borrower is creditworthy. sign. While rationing theories argue for linear and
An alternative explanation may be that borrowers quadratic terms in the funding probability equation,
have better repayment histories because the loans are it is less obvious that there is a similar implication
essentially funded by friends. The data suggest that for funded loans. The linear term is negative in four
this is not a first-order force because in more than specifications and positive in two others, whereas the
95% of all loan requests associated with friends, funds squared term is consistently positive in all models. In
from friends account for less than 4.4% of the amount unreported results, we estimated models with the lin-
that the borrowers receive. The results are more con- ear term alone, and obtained a positive and significant
sistent with the hypothesis that investors are not mis- coefficient.
led by relying on friends, especially friend types at Finally, we consider group affiliations. Any member
the root level of the hierarchy in Figure 1, as informa- on Prosper.com can propose a group and members are
tional cues about borrower credit quality. admitted based on eligibility criteria. Some groups,
such as employee or university alumni groups require
6.3. Controls verification of the qualifying criteria, whereas oth-
Whereas friendships are the primary focus of our ers have looser joining criteria. An individual can
study, we also include an exhaustive set of controls. be a member of only one group at a time. Whereas
We briefly review and comment on the key controls. borrowers cannot leave or change groups until loan
Hard credit variables have the expected signs in all repayment, entry or exit into other groups is free.
specifications. Lower rated listings, as well as those We manually code all groups that have at least three
with more credit inquiries and higher debt-to-income members and create dummy variables based on mem-
ratios are less likely to be funded, attract higher inter- bership criteria such as religious affiliation. We also
est rates, and are more likely to default. Bank card uti- control for group size.
lization has a positive coefficient, whereas its square There are 4,139 groups in our sample; 59,978 loan
has a negative coefficient. Some card utilization is requests representing 29% of loan requests, indicate
beneficial as it signals creditworthiness, but excessive a group affiliation. Group variables explain fund-
usage is not as it likely signals borrowers who are ing probability; 7.09% of listings not affiliated with
financially stretched. An interesting variable is the a group are successfully funded, whereas 10.36% of
number of years since a borrower’s first credit line, a listings with a group affiliation are funded. Groups
Lin, Prabhala, and Viswanathan: Friendship Networks and Information Asymmetry
Management Science 59(1), pp. 17–35, © 2013 INFORMS 31

with a religious motif enjoy lower interest rates by members post only one listing. We estimate a random
between 70 and 200 basis points, perhaps reflect- effects model and also consider a fixed effect logit
ing taste-based discrimination (Becker 1971) because model; both yield consistent results and support the
these groups do not default less. Geography-based role of friendships. We also consider survival spec-
groups matter in models H1–H4, but not in models ifications that model the number of listings that a
H5 and H6. Groups based on company or univer- borrower must post before getting funded for the
sity alumni, which are verifiable criteria, enjoy lower first time. Our primary results remain robust to this
interest rates (about 120 basis points), but only 451 of specification. We also analyze and find similar results
the over 200,000 requests in our sample are associated for borrowers who only have one loan, mitigating
with such groups. In terms of ex post default, Table 6 the reverse causality concern that friendships may
shows that only two types matter for loan perfor- manifest previous loan outcomes on Prosper.com.20
mance: alumni groups and geography-based groups. We also include binary indicators for all friendship
Members of these groups are less likely to default. variables in the hierarchy in Figure 1 instead of their
We also include a dummy variable for whether a counts, and obtain consistent results. Finally, while
leader of a group is rewarded for group member list- Prosper.com defines success as 100% funding of loans,
ings that are funded. Among all listings affiliated with we also estimated a Tobit model with the extent of
groups, 28,006 (46.63%) are associated with a leader funding as a dependent variable rather than the fund-
incentive structure. We find that the presence of group ing probability. The friendship results are consistent.
leader incentives leads to a greater likelihood of fund- We further consider whether structural measures
ing and lower interest rates, but these listings are in of networks, which deal with node positions and
fact more likely to default. Thus, group rewards create network topologies rather than roles and identities,
incentives similar to the originate-sell model of inter- matter. They have little effect on the results for friend-
mediaries held responsible for the 2008 financial crisis ships. This is unsurprising. We find that the friend-
(Hildebrand et al. 2010). This is partly the reason that ship networks on Prosper.com have many disjoint,
such incentives were discontinued by Prosper.com. star-shaped components. Because Prosper.com friend-
The important finding is, however, that including ship networks are formed offline and then revealed
group variables does not affect the results friendship on the platform—rather formed on the platform due
or friendship type in any of our specifications. to identity-shielding—the network has a low degree
of closure.21 This strengthens the case for looking at
6.4. Alternative Specifications the roles and identities of friends. We also consider
We estimate several alternative specifications to assess the number of endorsements received by borrowers.
the robustness of our friendship results. All results This variable is cheap talk and unsurprisingly, it is
not reported here are available from the authors. insignificant. One variable that does matter is the
One potential concern is that friend-lenders may number of friends’ defaults in a borrower’s neighbor-
be driving the results in the funding probability. hood (the ego network in the network analysis liter-
If the majority of funds for a loan comes from a ature). We find that a higher number of defaults in a
friend-lender, there will be information unobservable neighborhood of a borrower is associated with higher
to researchers, and our models may produce biased risk of the ego’s loan (Cohen-Cole and Duygan-Bump
results even though, for 95% of the loans, friends 2008), though inclusion of this variable does not
provide less than 4.4% of the requested amount. meaningfully alter our main results.
To address this concern, we study the number of bids
from strangers (lenders who are not friends with the
borrower) as a function of the same covariates in the 7. Images and Text
probit models.18 We use a negative binomial model Individuals seeking funding on Prosper.com can
to account for overdispersion. Our results across all upload images and add descriptive text to their list-
levels of the friendship hierarchy remain highly con- ings. A priori, it is not clear why these data should
sistent as well.19 subsume friendship. The image and text data are self-
We re-estimate the funding probability model using reported by borrowers and are not authenticated by
a logit model and find consistent results. We consider
the correlation induced by borrowers who re-list after 20
Borrowers who default are not allowed to borrow again, mitigat-
a failed initial listing by constructing a panel data ing the concern that past defaults influence current outcomes.
set, with each member as a unit and each listing as 21
See Coleman (1988) or Burt (2005) for details. As an example, A,
a time period. The panel is unbalanced as 53% of B, and C are members on Prosper.com and are only identified by
random user names. A and B both have a friendship tie with C, but
A and B are not directly connected. Then, A and B still cannot tell
18
We thank the associate editor for this suggestion. or recognize the real identity of each other, reducing the probability
19
These results are available from the authors. of a closure in C’s ego network.
 Lin, Prabhala, and Viswanathan: Friendship Networks and Information Asymmetry
32 Management Science 59(1), pp. 17–35, © 2013 INFORMS

Prosper.com. However, images or text could act as respectively.22 Blacks pay between 40 and 50 basis
a channel by which friendships matter. For instance, points more in interest rates compared to the point
borrowers with better quality friends may leverage estimate of 60–80 basis points reported in Pope and
the friendships to post more persuasive text or images Sydnor (2011). Our estimate is not significant. As in
that might do a better job at attracting funding. Pope and Sydnor, we find that blacks are significantly
Whether images or text subsume friendships becomes more likely to default with a hazards ratio of 1.20 that
an empirical issue that we turn to next. For concise- is significant at 1%. We find that the key coefficients
ness, we focus only on the coefficients for the friend- for friendships are similar and show similar grada-
ship variables. The full results are available from the tion along friend roles and identities as in the full
authors. sample.

7.1. Images 7.2. Text

We next examine whether descriptive text provided
Many listings on Prosper.com either post no images
by borrowers explains some of the content of the
or simply use the boilerplate images provided by
friendship ties. Following Tetlock (2007), we use a dis-
Prosper.com. As a starting point, we consider the role
ambiguation routine to classify text. We use the pro-
of friendship variables in the no-image subsample.
gram Linguistics Inquiry and Word Count (LIWC),
All key variables—friends, friends who are poten-
which classifies words into five broad categories and
tial lenders, real lenders, or real lenders who bid on
80 further (overlapping) subcategories such as long
listings—remain significant in this subsample.
words or punctuation marks as well as complex psy-
We next consider sample listings that include chological, social, and personal categories. The clas-
images. Automatic image processing software yield sification is based on an extensively validated and
unreliable results due to varying photo quality and updated dictionary of words from psychology and
content. We instead manually code the data. Because linguistics (Slatcher and Pennebaker 2006).
of the high costs of coding the entire sample, we work We experiment with and settle on a set of 12
with a 10% random subsample of listings. To ensure LIWC categories that represent a nontrivial fraction
representativeness, we preserve the proportions of of the word count and that seem most relevant to
successful listings, credit grades, and the degrees of lending outcomes. On average, funded listings have
relations depicted in Figure 1. We also consider a sec- more words per listing, shorter sentences, use more
ond subsample comprising all 16,500 funded listings. numerals, more words in the money subcategory, pos-
In the 10% random sample of 20,513 listings, 15,928 itive emotion words, more words of certainty, and
post images, of which 7,986 contain images of adult fewer tentative words. Most variables, however, do
humans. In the sample of 16,500 funded loans, 10,198 not show consistent results across three outcome vari-
listings have images, of which 8,279 listings contain ables. For instance, money words are more likely to
images of adult humans. We hire assistants to code result in funding and lead to lower interest rates, but
objective aspects of images such as race, age, and gen- they are insignificant in default models. Conversely,
der. We implement extensive screens to ensure output quantifiers such as few or many lower defaults, and
quality. Details are available to readers on request. certainty words matter in defaults but not in the fund-
About 14.55% of the random 10% subsample of all ing equation. No text variable shows the consistency
listings have images of blacks, whereas the propor- of the friendship variables across the funding prob-
tion of blacks in the funded loans is only about 8.79%. ability, interest rate, and default specifications. Thus,
Thus, blacks are less likely to be funded, as in Pope the friendship results remain robust to text. This is not
and Sydnor (2011) and Ravina (2008). The differences intended to be a comprehensive verdict on the role of
for other minority racial groups are less significant; linguistic content in determining economic outcomes.
6.20% of listings are Asian and 4.75% are Hispanic, Although LIWC and General Inquirer are accepted
while these populations represent 6.91% and 4.21% of state-of-the-art text mining programs, they may not
loans funded, respectively. Females form 30% of the capture all of the subtle nuances of text or their com-
listings, but 37% of all funded loans, suggesting that plementarity to the other listing data. Furthermore,
women are more likely to attract funding. Young peo- these data are self-reported and unlikely to be pro-
ple below 25 form 23% of all listings, but 19.33% of all cessed as evenly and rationally as credible signals
funded loans. Older people of age 50+ form 6.65% of with more economic content that are subject to exten-
the listings, but only 6.24% of the funded loans. sive authentication by Prosper.com.
We consider multivariate specifications that incor-
22
porate image data jointly with friendships. We find Despite our extensive set of controls, it remains possible that vari-
ations about race may be due to other demographic factors such
that listings with images of older people of age as education and occupation. Therefore, the results on race should
50+ and those with images of black adults are less be interpreted conservatively (we thank the associate editor for this
likely to be funded at the 5% and the 10% levels, insight).
Lin, Prabhala, and Viswanathan: Friendship Networks and Information Asymmetry
Management Science 59(1), pp. 17–35, © 2013 INFORMS 33

7.3. Quantitative Valuation of Friendship Signal which there is an especially severe problem of adverse
The previous results show that loans to borrowers selection. An interesting question is what mechanisms
with friends are made at lower rates. We assess individual agents use to adapt. The mechanism by
whether the reduction in interest rates due to friend- which agents adapt is interesting given that they lack
ship is reasonable relative to its effect on defaults. the sophisticated risk assessment methodologies or
To motivate our approach, we note that the loan inter- scale economies of banks. Individual agents also lack
est rate equals the risk-free rate plus compensation the soft information in lending from a broader vec-
for a borrower’s default risk. Thus, to assess how tor of financial relationships that is available to tradi-
the friendship signal is valued we compare the effect tional intermediaries such as banks. However, agents
of friendship on default rates to its effect on default adapt precisely as predicted by the economic theories
risk. A higher number would indicate that investors of signaling. The results are particularly interesting
value the friendship signal more, whereas a low num- because of the context, which features a rather sub-
ber suggests that the P2P market offers conserva- tle signal, arm’s-length loans, and individual agents
tive adjustments for friendships relative to its ex post putting small sums of money to work. Yet, the results
effect on defaults. To benchmark the findings, com- line up surprisingly well with the predictions of eco-
pare the friendship results to those for hard credit nomic theory.
variables. We also shed light on the role of soft information
in credit markets. Extensive work in economics and
Empirically, the adjustments in interest rates for
finance argues that credit markets suffer from a prob-
friendship variables are modest ranging from 22 to
lem of adverse selection. The literature traditionally
45 basis points. The most informative comparison is
views financial intermediaries as a solution because
how friendship fares relative to hard credit variables.
they can produce soft information about borrow-
Relative to AA borrowers (the omitted category) in
ers. As financial markets undergo disintermediation
our specifications, the raw credit spreads for A and B
driven by information technology, a natural concern is
borrowers are 80 and 190 basis points, and escalate to
that the soft information produced by intermediaries
660 basis points for E credits. Scaled for their effects is lost, adversely affecting credit flows. Additionally,
on the default hazard, the spreads are 47, 94, and while digital technology widens the reach of credit
132 basis points, respectively. These adjustments are markets, it can exacerbate asymmetric information
higher relative to those for the friendship variables. because borrowers become more anonymous. Our
Thus, investors use the friendship information conser- results highlight how these concerns are at least par-
vatively, offering conservative rate adjustments rela- tially mitigated. Information technology could sup-
tive to the ex post effect of friendship on default rates. plant some sources of soft information, but it could
Consistent with the view of Pope and Sydnor (2011), also increase its supply by developing new sources of
markets do not appear to make sufficient adjustments soft information, hardening it, and permitting its self-
for race effects. organization and availability to lenders. The use of
friendship may be seen as one manifestation of such
an effect.
8. Conclusions and Implications Our study can also be viewed as new evidence on
We study the online market for P2P lending, where
whether social capital facilitates economic exchange,
individual lenders make unsecured microloans to
a question of growing interest in information sys-
individual borrowers. We show that borrowers with tems, economics, and management. The P2P lend-
online friends on the Prosper.com platform have bet- ing marketplace offers microlevel data on this issue
ter ex ante outcomes. The results are consistent with with two significant empirical advantages. It identi-
the joint hypothesis that friendship ties act as a sig- fies social capital, which is a key challenge in the lit-
nal of credit quality, and that individual investors erature. In many studies, it is hard to identify the
understand this relationship and incorporate it into social relationships formed outside the economic con-
their lending decisions. To further understand why text being studied. The P2P loan marketplace offers
friendships matter, we examine whether friendships relatively well-defined, objective measures of social
are related to ex post loan outcomes. We find that capital through friendships that are formed outside
borrowers with friends, especially of the sort that are the lending context. It provides additional granular-
more likely to be credible signals of credit quality, ity based on the roles and identities of the friends.
are less likely to default. Our results are more pro- The marketplace also provides well-defined measures
nounced when friends have roles and identities that of transactional outcomes, viz., funding, interest rates,
are more likely to signal better credit quality. and default.
Our findings are of interest from a number of Finally, our results have specific implications for
viewpoints. One perspective is that our study repre- new businesses based on emerging Web 2.0 tech-
sents data from an emerging online credit market in nologies, such as other electronic commerce sites.
 Lin, Prabhala, and Viswanathan: Friendship Networks and Information Asymmetry
34 Management Science 59(1), pp. 17–35, © 2013 INFORMS

Much of the potential in these new businesses comes Angrist JD (1998) Estimating the labor market impact of voluntary
from the new sources of information about users and military service using social security data on military appli-
cants. Econometrica 66(2):249–299.
the ability to harness this information in new busi- Aral S, Walker D (2011) Creating social contagion through viral
ness models. Our study suggests that such businesses product design: A randomized trial of peer influence in net-
could facilitate the creation of online relationships and works. Management Sci. 57(9):1623–1639.
incorporate them into their business models, which Bagozzi RP, Dholakia UM (2006) Open source software user com-
munities: A study of participation in Linux user groups. Man-
could result in economic benefits for end-users. This agement Sci. 52(7):1099–1115.
is particularly true when such relationships are self- Becker GS (1971) The Economics of Discrimination (University of
organizing and impose little cost to the marketplace,23 Chicago Press, Chicago).
as we document in this study. The realization of Bedard K (2001) Human capital versus signaling models: Uni-
versity access and high school dropouts. J. Political Econom.
such benefits, however, depends on the ability of the 109(4):749–775.
platform to make the new information credible and Bertrand M, Luttmer EFP, Mullainathan S (2000) Network effects
enrich the vector of economic attributes of the peo- and welfare cultures. Quart. J. Econom. 115(3):1019–1055.
ple forming the relationships. For instance, in our Bhattacharjee S, Gopal RD, Lertwachara K, Marsden JR, Telang R
(2007) The effect of digital sharing technologies on music mar-
study, friendships matter. However, this is not a quan- kets: A survival analysis of albums on ranking charts. Manage-
tity effect that scales with the number of friends, but ment Sci. 53(9):1359–1374.
depends on the roles and identities of the friends and Bolton GE, Katok E, Ockenfels A (2004) How effective are elec-
the nature of their social and economic relations. Ties tronic reputation mechanisms? An experimental investigation.
Management Sci. 50(11):1587–1602.
that are more difficult to form and more likely to Burt R (1992) Structural Holes (Westview Press, Cambridge, MA).
lead to social stigma in case of default create stronger Burt R (2005) Brokerage and Closure: An Introduction to Social Capital
incentives to repay, and are also more credible signals (Oxford University Press, New York).
of creditworthiness. Cohen-Cole E, Duygan-Bump B (2008) Household bankruptcy deci-
sion, the role of social stigma vs. information sharing. Working
paper, Federal Reserve Bank of Boston, Boston.
Acknowledgments Coleman JS (1988) Social capital in the creation of human capital.
This paper is based on Essay 1 of Mingfeng Lin’s (2010) Amer. J. Sociol. 94(Supplement):S95–S120.
dissertation at the University of Maryland, College Park. Crocker J, Major B, Steele C (1998) Social stigma. Gilbert DT, Fiske
The authors thank Steve Borgatti, Dan Brass, Ethan Cohen- ST, Lindzey G, eds. The Handbook of Social Psychology, Vols. 1
Cole, Sanjiv Das, Jerry Hoberg, Dalida Kadyrzhanova, and 2 (McGraw-Hill, New York), 504–553.
Dellarocas C (2003) The digitization of word of mouth: Promise
Nikunj Kapadia, De Liu, Vojislav Maksimovic, Joe Peek,
and challenges of online feedback mechanisms. Management
Gordon Phillips, Kislaya Prasad, Galit Shmueli, and Kelly Sci. 49(10):1407–1424.
Shue, as well as seminar participants at Carnegie Mellon Doreian P (2001) Causality in social network analysis. Sociol. Meth-
University, University of Utah, University of Kentucky, ods Res. 30(1):81–114.
the 2008 Summer Doctoral Program of the Oxford Inter- Durlauf SN, Fafchamps M (2005) Social capital. Aghion P,
net Institute, the 2008 INFORMS Annual Conference, the Durlauf SN, eds. Handbook of Economic Growth, Vol. 1, Part B
(Elsevier, Amsterdam), 1639–1699.
2009 Workshop on Information Systems and Economics,
Fama E (1985) What’s different about banks. J. Monetary Econom.
the 2009 Western Finance Association Annual Meeting, the
15(1):29–39.
2009 INFORMS Marketing Science Conference, and the 2010 Ghose A, Ipeirotis P (2011) Estimating the helpfulness and economic
Intra-Organizational Network (ION4) Conference for their impact of product reviews: Mining text and reviewer character-
valuable comments and suggestions. Mingfeng Lin also istics. IEEE Trans. Knowledge Data Engrg. 23(10):1498–1512.
thanks the Ewing Marion Kauffman Foundation for the Gnyawali DR, Fan W, Penner J (2010) Competitive actions and
2009 Dissertation Fellowship Award and the Economic Club dynamics in the digital age: An empirical investigation of
social networking firms. Inform. Systems Res. 21(3):594–613.
of Washington (2008) for their generous financial support.
Gorton G, Winton A (2003) Financial intermediation,
The authors also thank Prosper.com for making the data for Constantinides GM, Harris M, Stulz RM, eds. Handbook of
the study available. The contents of this publication are the the Economics of Finance, Vol. 1, Part 1, Chap. 8 (Elsevier,
sole responsibility of the authors. Amsterdam), 431–552.
Granovetter M (1972) The strength of weak ties. Amer. J. Sociol.
78(6):1360–1380.
Granovetter M (1985) Economic action and social structure: The
References problem of embeddedness. Amer. J. Sociol. 91(3):481–510.
Agarwal S, Hauswald RB (2007) The choice between arm’s- Granovetter M (2005) The impact of social structure on economic
length and relationship debt: Evidence from eloans. Working outcomes. J. Econom. Perspect. 19(1):33–50.
paper, Federal Reserve Bank of Chicago, Chicago. http://ssrn Gross DB, Souleles NS (2002) An empirical analysis of personal
.com/abstract=968010. bankruptcy and delinquency. Rev. Financial Stud. 15(1):319–347.
Agarwal S, Ambrose BW, Liu C (2006) Credit lines and credit uti- Grossman S (1981) The informational role of warranties and private
lization. J. Money Credit and Banking 38(1):1–22. disclosure about product quality. J. Law Econom. 24(3):461–483.
Akerlof GA (1970) The market for “lemons”: Quality uncertainty Grover V, Fielder K, Teng J (1997) Empirical evidence on Swan-
and the market mechanism. Quart. J. Econom. 84(3):488–500. son’s tri-core model of information systems innovation. Inform.
Systems Res. 8(3):273–287.
Guiso G, Sapienza P, Zingales L (2004) Role of social capital in
23
We thank an anonymous referee for this valuable insight. financial development. Amer. Econom. Rev. 94(3):526–556.
Lin, Prabhala, and Viswanathan: Friendship Networks and Information Asymmetry
Management Science 59(1), pp. 17–35, © 2013 INFORMS 35

Hamermesh DS, Biddle JE (1994) Beauty and the labor market. Overby E, Jap S (2009) Electronic and physical market channels:
Amer. Econom. Rev. 84(5):1174–1194. A multiyear investigation in a market for products of uncertain
Hanneman RA, Riddle M (2005) Introduction to Social Network Meth- quality. Management Sci. 55(6):940–957.
ods (University of California, Riverside). Petersen M, Rajan R (1994) The benefits of lending relationships:
Hauswald R, Marquez R (2003) Information technology and finan- Evidence from small business data. J. Finance 49(1):3–37.
cial services competition. Rev. Financial Stud. 16(3):921–948. Petersen M, Rajan R (2002) Does distance still matter? The
Heckman J (1979) Sample selection bias as a specification error. information revolution in small business lending. J. Finance
Econometrica 47(1):153–161. 57(6):2533–2570.
Hildebrand T, Puri M, Rocholl J (2010). Skin in the game: Incentives Podolny JM (1993) A status-based model of market competition.
in crowdsourcing. Working paper, E.CA Economics, Berlin. Amer. J. Sociol. 98(4):829–872.
http://ssrn.com/abstract=1615483. Podolny JM (2001) Networks as the pipes and prisms of the market.
Hill S, Provost F, Volinsky C (2006) Network-based marketing: Amer. J. Sociol. 107(1):33–60.
Identifying likely adopters via consumer networks. Statist. Sci. Pope D, Sydnor J (2011) What’s in a picture? Evidence of discrimi-
21(2):256–276. nation from Prosper.com. J. Human Resources 46(1):53–92.
Houser D, Wooders J (2006) Reputation in auctions: Theory and Putnam R (1993) The prosperous community: Social capital and
evidence from ebay. J. Econom. Management Strategy 15(2): economic growth. Amer. Prospect 4(13):35–42.
353–369. Rajan U, Seru A, Vig V (2010) Statistical default models and incen-
Iyer R, Khwaja AI, Luttmer EFP, Shue K (2009) Screening in new tives. Amer. Econom. Rev. 100(2):506–510.
credit markets: Can individual lenders infer borrower credit- Ravina E (2008) The effect of beauty and personal characteris-
worthiness in peer-to-peer lending? Working paper, National tics in credit markets. Working paper, Columbia University,
Bureau of Economic Research, Cambridge, MA. New York.
Jackson MO (2010) An overview of social networks and economic Resnick R, Zeckhauser R, Swanson J, Lockwood K (2006) The value
applications. Benhabib J, Bisin A, Jackson MO, eds. Handbook of reputation on eBay: A controlled experiment. Experiment.
of Social Economics (Elsevier Press, Amsterdam), 511–585. Econom. 9(2):79–101.
John K, Williams J (1985) Dividends, dilution, and taxes: A sig- Ross SA (1977) The determination of financial structure: The
nalling equilibrium. J. Finance 40(4):1053–1070. incentive-signalling approach. Bell J. Econom. 8(1):23–40.
Kalbfleisch JD, Prentice RL (2002) The Statistical Analysis of Failure Sapienza P, Toldra A, Zingales L (2007) Understanding trust. Work-
Time Data, 2nd ed. (Wiley, New York). ing paper, Northwestern University, Evanston, IL.
Karlan D (2007) Social connections and group banking. Econom. J. Slatcher RB, Pennebaker JW (2006) How do I love thee? Let me
117(517):52–84. count the words: The social effects of expressive writing. Psych.
Kremer M, Levy D (2008) Peer effects and alcohol use among col- Sci. 17(8):660–664.
lege students. J. Econom. Perspect. 22(3):189–206. Spence M (1973) Job Market Signaling. Quart. J. Econom.
Leland HE, Pyle DH (1977) Informational asymmetries, finan- 87(3):355–374.
cial structure, and financial intermediation. J. Finance 32(2): Spence M (2002) Signaling in retrospect and the informational
371–387. structure of markets. Amer. Econom. Rev. 92(3):434–459.
Lin M (2010) Decisions under uncertainty in decentralized online Staiger D, Stock JH (1997) Instrumental variables regression with
markets: Empirical studies of peer-to-peer lending and out- weak instruments. Econometrica 65(3):557–586.
sourcing. Dissertation, University of Maryland, College Park. Stiglitz JE, Weiss A (1981) Credit rationing in markets with imper-
Manski C (1993) Identification of endogenous social effects: The fect information. Amer. Econom. Rev. 71(3):393–410.
reflection problem? Rev. Econom. Stud. 60(3):531–542. Susarla A, Oh J-H, Tan Y (2012) Social networks and the diffusion
Mizruchi MS (1992) The Structure of Corporate Political Action: of user-generated content: Evidence from YouTube. Inform. Sys-
Inter-Firm Relations and Their Consequences (Harvard University tems Res. 23(1):23–41.
Press, Cambridge, MA). Tetlock P (2007) Giving content to investor sentiment: The role of
Mobius MM, Rosenblat TS (2006) Why beauty matters. Amer. media in the stock market. J. Finance 62(3):1139–1168.
Econom. Rev. 96(1):222–235. Thorne D, Anderson L (2006) Managing the stigma of personal
Moffitt R (1983) An economic model of welfare stigma. Amer. bankruptcy. Sociol. Focus 39(2):77–97.
Econom. Rev. 73(5):1023–1035. Uzzi B (1999) Embeddedness in the making of financial capi-
Moran P (2005) Structural vs. relational embeddedness: Social tal: How social relations and networks benefit firms seeking
capital and managerial performance. Strategic Management J. financing. Amer. Sociol. Rev. 64(4):481–505.
26(12):1129–1151. Uzzi B, Lancaster R (2003) Relational embeddedness and learning:
Oestreicher-Singer G, Sundararajan A (2012) The visible hand? The case of bank loan managers and their clients. Management
Demand effects of recommendation networks in electronic Sci. 49(4):383–399.
markets. Management Sci., ePub ahead of print June 15, Weiss A (1995) Human capital vs. signaling explanations of wages.
http://dx.doi.org/10.1287/mnsc.1120.1536. J. Econom. Perspect. 9(4):133–154.