This article was downloaded by: [130.241.30.

106] On: 08 March 2022, At: 08:32

Publisher: Institute for Operations Research and the Management Sciences (INFORMS)
INFORMS is located in Maryland, USA

Information Systems Research

Publication details, including instructions for authors and subscription information:
http://pubsonline.informs.org

What’s in a “Username”? The Effect of Perceived

Anonymity on Herding in Crowdfunding
Yang Jiang, Yi-Chun (Chad) Ho, Xiangbin Yan, Yong Tan

To cite this article:

Yang Jiang, Yi-Chun (Chad) Ho, Xiangbin Yan, Yong Tan (2021) What’s in a “Username”? The Effect of Perceived Anonymity on
Herding in Crowdfunding. Information Systems Research

Published online in Articles in Advance 29 Nov 2021

. https://doi.org/10.1287/isre.2021.1049

Full terms and conditions of use: https://pubsonline.informs.org/Publications/Librarians-Portal/PubsOnLine-Terms-and-

Conditions

This article may be used only for the purposes of research, teaching, and/or private study. Commercial use
or systematic downloading (by robots or other automatic processes) is prohibited without explicit Publisher
approval, unless otherwise noted. For more information, contact permissions@informs.org.

The Publisher does not warrant or guarantee the article’s accuracy, completeness, merchantability, fitness
for a particular purpose, or non-infringement. Descriptions of, or references to, products or publications, or
inclusion of an advertisement in this article, neither constitutes nor implies a guarantee, endorsement, or
support of claims made of that product, publication, or service.

Copyright © 2021, INFORMS

Please scroll down for article—it is on subsequent pages

With 12,500 members from nearly 90 countries, INFORMS is the largest international association of operations research (O.R.)
and analytics professionals and students. INFORMS provides unique networking and learning opportunities for individual
professionals, and organizations of all types and sizes, to better understand and use O.R. and analytics tools and methods to
transform strategic visions and achieve better outcomes.
For more information on INFORMS, its publications, membership, or meetings visit http://www.informs.org
 INFORMATION SYSTEMS RESEARCH
Articles in Advance, pp. 1–17
http://pubsonline.informs.org/journal/isre ISSN 1047-7047 (print), ISSN 1526-5536 (online)

What’s in a “Username”? The Effect of Perceived Anonymity on

Herding in Crowdfunding
Yang Jiang,a Yi-Chun (Chad) Ho,b Xiangbin Yan,c,* Yong Tand
a
School of Business, Nanjing University, Nanjing, 210093 Jiangsu, China; b School of Business, George Washington University, Washington,
District of Columbia 20052; c School of Economics and Management, University of Science and Technology Beijing, 100083 Beijing, China;
d
Michael G. Foster School of Business, University of Washington, Seattle, Washington 98195-3226
*Corresponding author
Contact: yangjiang@nju.edu.cn, https://orcid.org/0000-0003-1368-7967 (YJ); chadho@gwu.edu, https://orcid.org/0000-0003-0383-1216
(Y-C(C)H); xbyan@ustb.edu.cn, https://orcid.org/0000-0001-7442-1244 (XY); ytan@uw.edu, https://orcid.org/0000-0001-8087-3423
(YT)

Received: March 7, 2018 Abstract. This research examines the role of perceived anonymity in shaping herding be-
Revised: April 3, 2019; February 2, 2020; havior in online crowdfunding markets. Drawing on theories from social psychology liter-
March 8, 2021; June 1, 2021 ature, we argue that a lender forms different credibility perceptions toward preceding
Accepted: July 1, 2021 peers based on their perceived anonymity state; the lender then uses such perceptions to
Published Online in Articles in Advance: adjust the lender’s herding momentum toward them. Using data collected from a leading
November 29, 2021
debt-based crowdfunding platform, we classify an individual’s username as either anony-
mous or real-seeming, with the latter referring to as a user identification that seems to re-
https://doi.org/10.1287/isre.2021.1049 veal one’s real name. The results show that successors demonstrate weaker herding
momentum toward predecessors who are presented with real-seeming usernames than
Copyright: © 2021 INFORMS
anonymous ones. This finding, which we attribute to a lower extent of perceived credibility
derived from a nonconforming behavior, challenges the conventional wisdom that consid-
ers anonymity a negative factor for source credibility. We further show that the uncovered
positive effect of perceived anonymity on herding is accentuated in the early stage of the
fundraising period; nevertheless, we find no such discrepancies between listings that are
assigned with high-risk and low-risk credit grades. Our study contributes to the literatures
dealing with anonymity and herding in online environments.

History: Xiaoquan (Michael) Zhang, Senior Editor; Khim Yong Goh, Associate Editor.
Funding: This work was supported in part by the National Natural Science Foundation of China [Grants
72025101, 72001103, and 71729001].
Supplemental Material: The online appendix is available at https://doi.org/10.1287/isre.2021.1049.

Keywords: crowdfunding • herding • observational learning • anonymity • privacy • source credibility

1. Introduction observational learning, a mechanism under which pred-

Debt-based crowdfunding is a collective process
where the crowd pool together and lend money to un-
known borrowers for interest earnings (Lin and Vis-
wanathan 2016). Information asymmetry inherent in
this market poses a tough challenge to individual
lenders; they must evaluate the potential risk of each
loan listing with very little information about the bor-
rower (Burtch et al. 2016). To facilitate decision mak-
ing, they resort to observe and follow the decisions
made by preceding peers, giving rise to herding behav-
ior (Zhang and Liu 2012, Xiao et al. 2021). The ratio-
nale is that the crowd is uncertain about the potential
risk of a listing; nonetheless, each individual may
hold some private knowledge about it, perhaps
through the individual’s previous experience with
similar listings or through personal connections
with the borrower (Agrawal et al. 2015). Consequent-
ly, lenders can infer a listing’s potential risk through
ecessors’ lending decisions are treated as signals that
carry useful information in successors’ eyes.
Then, are these individual signals perceived as
equally credible? Prior research suggests that indi-
viduals tend to attach uneven weights to informa-
tion sources that have different attributes, such as
network ties and expertise (Lee et al. 2015, Kim and
Viswanathan 2019). The cognitive process underly-
ing this differentiation behavior is that, upon receiv-
ing signals, individuals first assess source credibility
based on some observable personal cues (Petty et al.
1983). They then use such perceived source credibili-
ty to moderate the value of the information, which,
in turn, impacts their subsequent decision making.
Following this notion, we anticipate that lenders are
likely to incorporate a similar differentiation process
into their observational learning. If they indeed dis-
tinguish the informational value of previous lending

1

2 Information Systems Research, Articles in Advance, pp. 1–17, © 2021 INFORMS
transactions (i.e., signals), we should observe succes-
sors demonstrating different herding momentum to-
ward distinct types of predecessors (i.e., information
sources).
In online environments, a username is a unique and
persistent personal cue that is presented to and per-
ceived by others (e.g., Ma and Agarwal 2007, Forman
et al. 2008, Burtch et al. 2016). This role of usernames
could be particularly salient in contexts where other
personal information is hidden for privacy preserva-
tion (as in debt-based crowdfunding markets). Indeed,
empirical research has documented that usernames
can carry informational value that affects a beholder’s
perception in e-commerce settings (Silva et al. 2017,
Silva and Topolinski 2018). Along this line, conceal-
ment of usernames is further shown to discourage
subsequent crowdfunding activities, because an ab-
sence of identity information about others triggers a
sense of unease from a socially normative perspective
(Burtch et al. 2016).
Although it has been recognized that usernames,
albeit evaluated in various dimensions, can affect a
beholder’s perception, little is known about their im-
plication in shaping observational learning behavior,
which manifests as herding. To fill this gap, we inves-
tigate whether and how predecessors’ usernames—as
evaluated from a perspective of perceived anonymity—
affect successors’ herding momentum through the
varying extent of perceived source credibility. More-
over, as the main purpose of lenders engaging in ob-
servational learning is to infer investment risk (Zhang
and Liu 2012), a natural follow-up question to ask is
whether lenders would adjust their reliance on obser-
vational learning when dealing with different levels
of perceived risk. Thus, we are also interested in
examining whether and how the effect of perceived
anonymity on herding is moderated by listing charac-
teristics that could signal a listing’s potential risk.
We conceptualize perceived anonymity as the be-
holder perception that arises from observing prede-
cessors who present themselves with an anonymous
username. Conventional wisdom suggests that ano-
nymity provides no personal cue about information
sources, which lowers the perceived source credibility
and reduces the informational value of the signal
(Hass 1981, Kruglanski et al. 2006, Wagenknecht et al.
2016). Intuitively, a prospective lender is likely to per-
ceive predecessors presented with an anonymous
username as less credible, and accordingly exhibit a
weaker herding momentum toward them. Recent
work on online anonymity, however, implies that an
opposite effect may emerge as well. Debt-based
crowdfunding is an online community (Burtch et al.
2016) wherein members have little incentive to reveal
identity cues (e.g., real names) to others. As remaining
anonymous to “lie low” emerges as an agreed
Jiang et al.: The Effect of Perceived Anonymity on Herding
behavioral standard in such an investment-oriented
context (Pritchard 2020), the behavior of keeping a
high profile by voluntarily revealing one’s identity in
a username is likely to arouse disapproval and skepti-
cism from the beholders (Van Kleef et al. 2015), atten-
uating the magnitude of the herding effect. Bearing
the competing arguments presented earlier, we leave
whether perceived anonymity has a positive or nega-
tive effect on lender herding as an empirical question
worth academic examination.
To study our research questions, we collect data
from a leading debt-based crowdfunding platform in
Asia. Our data record crowdfunding activities on this
platform, including characteristics at the listing, bor-
rower, and transaction level. The platform provides
an ideal setting for answering the posed question for
two reasons. First, the platform conceals all personal
information about all lenders for privacy preservation
purposes, leaving usernames the only salient identity
cue that can affect a beholder’s perception of source
credibility. This contextual feature allows us to isolate
the role of usernames in impacting herding behavior,
after controlling for other listing attributes observable
to subsequent lenders. Second, unlike some platforms
where transactional information is partially disclosed
(e.g., the context studied by Burtch et al. 2016),
prospective lenders in our context observe detailed
information about all previous funding activities in a
salient fashion. With this complete transaction log, we
can recover the cumulative lending amount a listing
has received at a given point of time. With this key
measure for the herd volume, we can quantify the
herding effect, allowing us to empirically investigate
whether successors exhibit different herding magni-
tudes toward predecessors presented with different
types of usernames.
Our preliminary result confirms the existence of
herding, as reflected by a positive correlation between
lending amounts funded by precedent and subse-
quent lenders. To investigate the role of perceived an-
onymity, we first classify each individual’s username
as either anonymous or real-seeming, with the latter
referring to as a user identification that seems to re-
veal one’s real name. We then construct our herding
measure (i.e., cumulative lending amount a listing has
received at a given time) and decompose it into two
based on the classification result. The empirical results
suggest that successors demonstrate weaker herding
momentum toward predecessors who are presented
with real-seeming usernames than anonymous ones.
This finding, which we attribute to a lower extent of
perceived credibility derives from a nonconforming
behavior (i.e., deviating from others by presenting
oneself with a real-seeming username), challenges the
conventional wisdom that considers anonymity a neg-
ative factor for source credibility. Further, in
Jiang et al.: The Effect of Perceived Anonymity on Herding
Information Systems Research, Articles in Advance, pp. 1–17, © 2021 INFORMS
examining the moderating role of risk-related attrib-
utes, we find that the effect of perceived anonymity
on herding is more pronounced in the early stage of
the fundraising period. Nonetheless, no such systemic
difference is found between listings that differ in cred-
it grades assigned by the platform.
The current research makes several contributions to
the literature. First, our study extends prior work
dealing with anonymity in the information systems
literature. Unlike past work that takes a socially nor-
mative perspective and therefore considers perceived
anonymity a factor that undermines source credibility
(e.g., Forman et al. 2008, Jiang et al. 2013, Burtch et al.
2016), we argue observational learning as a distinct
mechanism under which the effect of perceived ano-
nymity can manifest in either direction. Our empirical
analyses demonstrate that perceived anonymity inten-
sifies herding in the debt-based crowdfunding market.
Though seemingly surprising, our main finding is not
uncommon. In this regard, we add to the emerging re-
search stream that underscores the bright side of ano-
nymity in a variety of online settings, such as product
reviews (Hong and Li 2017) and user-generated con-
tent (Lee et al. 2014). Second, prior credibility litera-
ture (e.g., Sternthal et al. 1978, Harmon and Coney
1982, Wagenknecht et al. 2016) has largely focused on
the effect of source credibility on persuasive commu-
nication (direct, initiated by information sources). Our
work adds to this literature by showing the possibility
that source credibility can also play a role in observa-
tional learning (indirect, information receivers are
proactive). Furthermore, our work also contributes to
the literature stream focusing on the identification of
herding, specifically in online crowdfunding markets
(e.g., Zhang and Liu 2012, Jiang et al. 2018). Our find-
ings provide a nuanced yet important understanding
of the mechanism that underlies herding behavior. In-
stead of mimicking precedent transactions in a ho-
mogenous manner, successors first form different
credibility perceptions toward predecessors based on
observable personal cues (De Vries and Pruyn 2007);
they then use such perceptions to adjust the informa-
tional value derived from observational learning ac-
cordingly. Results from our extended analysis help
better understand that the observed behavior of dif-
ferentiated herding can vary with some risk-related
characteristics that can potentially convey the invest-
ment risk associated with a listing.
Our study is closely related to Burtch et al. (2016),
who also analyze backers’ behavior on a reward-
based crowdfunding site. In their study context, indi-
vidual backers can decide whether to conceal their
usernames or not at the transaction level. Despite
some similarities, our research differs from theirs in
three major aspects. First, the two studies explore dis-
tinct social dynamics. Burtch et al. (2016) observe a
3
positive association between backers’ decision to con-
ceal identity cues (i.e., usernames), citing social norms
as the driving mechanism. However, we investigate
the sequential correlation in lenders’ lending amount,
arguing herding arises as a result of observational
learning. Our extended analysis suggests that listing
characteristics that may signal a listing’s risk level can
moderate herding, confirming observational learning,
rather than simply mimicking or socially normative
behavior, being the mechanism that underlies lenders’
decision making. Second, the main findings are oppo-
site. Their work finds that early peers’ decision to re-
main anonymous has a negative influence on later
backers’ contributions, which pertains to the “I’ve got
nothing to hide” argument from the law literature (So-
love 2007). By contrast, our results suggest that prede-
cessors presented with an anonymous username in-
duce a stronger herding effect, in that such privacy-
preservation behavior is subject to fewer disapprovals
in an online investment community. Finally, their
work leverages backers’ choice of information control
at the transaction level as a unique contextual feature
and helps understand how the effect of identity hid-
ing manifests within a crowdfunding project. Differ-
ent from that, our research explores the variation in
perceived anonymity at the lender level, allowing us
to uncover the broader impact of usernames as a per-
sistent personal cue, in a more general, online envi-
ronment. Taken together, these conceptual and con-
textual differentiations help explain why the
uncovered effects of perceived anonymity are quite
different between the two studies.
2. Theoretical Background and Hypotheses
2.1. Herding
Herding, defined as the tendency for an individual to
follow the actions of preceding peers (Banerjee 1992;
Bikhchandani et al. 1992, 1998), has drawn growing
research attention in online environments. One of the
most prominent behavioral mechanisms behind herd-
ing is observational learning which describes situa-
tions in which people draw inference from mere ob-
servations of others’ actions (Zhang 2010). For
example, Cai et al. (2009) examine the effect of obser-
vational learning on diners’ dish choices in a restau-
rant setting. They find that when diners are presented
with information about popular dishes ordered by
others, the demand for these dishes increases. Using
data from a field experiment, Tucker and Zhang
(2011) study how the popularity information affects
consumers’ choices on a website that lists wedding
service vendors. They show that narrow-appeal ven-
dors attract more visits than their broad-appeal coun-
terparts after the introduction of popularity informa-
tion. Zhang and Liu (2012) investigate whether the

4 Information Systems Research, Articles in Advance, pp. 1–17, © 2021 INFORMS
observed herding behavior is rational or irrational in
microloan markets. Their empirical results reveal
that the magnitude of herding is accentuated by unfa-
vorable listing characteristics but attenuated by favor-
able ones, providing evidence of rational observation-
al learning.
Drawing on observational learning theory, we
argue that lenders are observational learners who re-
gard predecessors’ lending transactions as informa-
tive signals reflecting the collective assessment of a
listing. Having little to no information about a bor-
rower, each individual lender observes the cumulative
amount collectively funded by preceding peers, and
uses this information to draw an inference of the list-
ing’s potential risk. As such, we expect that lenders
would exhibit a tendency to make a larger lending
amount to a listing that has received more funds.
There is a stream of literature that examines the role
of early decision makers’ attributes in influencing
follow-up user activities in various online communi-
ties. Research in this stream conveys a same message:
predecessors (and their actions) are not perceived as
equal by successors. For example, Lee et al. (2015) ex-
amine users’ reviewing behavior in response to previ-
ously posted ratings on a social movie site. They find
that moviegoers value ratings posted by their online
friends more than those by the crowd. Kim and
Viswanathan (2019) find a similar differentiation
behavior on a crowdfunding platform for mobile ap-
plications. For apps in the prerelease stage, investors
value preceding peers with app development experi-
ence more than those with investment experience;
however, they behave conversely for apps that are
already being sold in the market. An important
cognitive process that rationalizes such differentiation
behavior is that individuals tend to gauge first the
credibility of an information source based on some ob-
servable personal and identity cues (Petty et al. 1983);
they then use the perceived credibility to adjust the
value of the information obtained from the source
(Hovland et al. 1953, Ohanian 1990).
2.2. The Role of Perceived Anonymity
Prior literature has documented that the anonymity
state of information sources may influence a behold-
er’s perception of source credibility (Rains 2007a, b).
In most online settings, individuals can freely choose
a username, which serves as a unique and persistent
personal cue presented to others (e.g., Ma and Agar-
wal 2007, Forman et al. 2008, Burtch et al. 2016). This
role of a username would be even more salient when
other identity-related information is unobservable to
beholders. Such freedom of username choice grants
individuals a means to present themselves, which, in
turn, would determine how they are perceived by
others in the same community. In our framework,
Jiang et al.: The Effect of Perceived Anonymity on Herding
we do not use anonymity as a term to describe a
status of whether one’s offline identity (i.e., real name)
being known (or unknown) to others. Instead, we con-
ceptualize anonymity, or perceived anonymity to be
more exact, as the beholder perception that arises
upon observing other individuals presenting them-
selves with a username that seems (or doesn’t seem)
to reflect their real name. This conceptualization is
consistent with much prior work that considers per-
ceived anonymity dichotomous (e.g., Postmes and
Spears 1998, Hite et al. 2014). Specifically, we consider
a username either anonymous (associated with a high
state of perceived anonymity) or real-seeming (associ-
ated with a low state of perceived anonymity). In
what follows, we first propose a link between the per-
ceived anonymity state of a username and the per-
ceived credibility of the information source. We then
use the proposed link to rationalize the role of per-
ceived anonymity in shaping herding behavior in
debt-based crowdfunding markets.
Marx (1999) argues that we live in a culture where
the norm is that one must be identified, and names an-
chor us in the world as a “little detail in which big so-
cial meanings may reside.” Refusing to disclose such
information disregards our culture and makes it diffi-
cult for others to assess the credibility of the source
(Hovland et al. 1953, Kruglanski et al. 2006). Early
work has shown that a high anonymity state under-
mines the perceived trustworthiness, lowering the
perceived credibility of information sources (Hass
1981, Kruglanski et al. 2006). More recent research
suggests that the same cognitive process exists in on-
line environments as well. For example, information
contributed by anonymous (Forman et al. 2008) or un-
identified (Racherla and Friske 2012, Wagenknecht
et al. 2016) sources is perceived as less trustful by
shoppers in e-commerce settings. Individuals are also
found more hesitant to transact when the information-
al cues about preceding peers are missing, as an
absence of personal information results in negative at-
titudes, such as skepticism (Gneezy et al. 2012) and
uncertainty (Burtch et al. 2016). Applying this cogni-
tive process to our context, we argue that a high state
of perceived anonymity is likely to result in a lower
extent of perceived credibility, which in turn attenu-
ates the magnitude of the herding effect.
Perhaps more interestingly, sources presented
with a low state of perceived anonymity may not nec-
essarily be viewed as more credible either. Before we
proceed to discuss the reasoning, it is useful to under-
score the feature that distinguishes our study context
from others. In many online settings, such as social
media and knowledge-sharing sites, users are moti-
vated, either intrinsically or extrinsically, to present
and establish themselves by revealing some identity
cues (e.g., using their real names as part of the
Jiang et al.: The Effect of Perceived Anonymity on Herding
Information Systems Research, Articles in Advance, pp. 1–17, © 2021 INFORMS
usernames). On a debt-based crowdfunding platform,
by contrast, individual lenders have relatively less in-
centive to do so. To increase the transparency of the
crowdfunding process, most platforms publicize all
funding activities (Burtch et al. 2015) at the expense of
lender privacy. The high visibility and traceability of a
lender’s past transactions may give rise to a worry of
being watched by an unauthorized third-party entity,
or even a fear of receiving unwanted solicitations
(from either friends or strangers) should one’s identity
is exposed. Such concerns attribute to a real-world ob-
servation that most lenders have a strong desire for
privacy preservation and accordingly prefer to pre-
sent themselves with a high state of anonymity (i.e.,
through using an anonymous username).1 As remain-
ing anonymous to lie low emerges as an accepted
standard in an investment-oriented community (Xu
et al. 2008, Bansal et al. 2016, Pritchard 2020), the be-
havior of keeping a high profile—by voluntarily re-
vealing one’s real name in a username2—is likely to
be perceived as nonconforming, arousing disapprov-
als and skepticism from other community members
even when no actual harm has occurred (Helweg-
Larsen and LoMonaco 2008, Van Kleef et al. 2015).
These unfavorable attitudes undermine the perceived
credibility of predecessors who are presented with a
low state of perceived anonymity, dampening succes-
sors’ herding momentum toward them.
The previous discussion proposes that the perceived
anonymity of predecessors could pose two opposite ef-
fects on source credibility perceived by successors, and
which effect prevails may depend on the nature of the
context. In determining source credibility, we argue
that the negative effect of perceived anonymity plays a
predominant role in a general, consumption-oriented
context, whereas the positive effect of perceived ano-
nymity manifests as a critical determinant in a more
specialized, investment-oriented context. The reason-
ing is twofold. First, in a consumption-oriented context
such as an e-commerce site, users voluntarily contrib-
ute to the community (e.g., writing reviews to share
product experiences) with motives like altruism and
helping the company (Hennig-Thurau et al. 2004). As
individuals have an innate need to be appreciated or
recognized for their contribution (Ma and Agarwal
2007), they may have an incentive to present them-
selves with a low state of anonymity such that their
identity can be associated with the good reputation
they have established. However, this is less unlikely
to occur in an investment-oriented context wherein in-
dividuals’ core task is to make profits. Since online in-
vestors have strong privacy concerns (Xu et al. 2008,
Bansal et al. 2016, Pritchard 2020), they are in general
more inclined to present themselves with a high state
of anonymity. Second, in a consumption-oriented con-
text, the concern of content manipulations (e.g., fake
5
reviews) casts serious doubt on the perceived credibility
of user-generated information (Ho et al. 2017), resulting
in distrust toward information sources presented with
a high state of anonymity. In an investment-oriented
context, however, the contents of peer-generated signals
(i.e., actual lending amount) are more objective (as op-
posed to online reviews that tend to be more subjective)
and less likely subject to manipulations. This contextual
nature diminishes lenders’ reliance on perceived ano-
nymity as a means of assessing source credibility. Bear-
ing this reasoning this mind, we anticipate that the posi-
tive effect of perceived anonymity on herding is likely
to outweigh its negative effect. Formally, we propose
the following hypothesis.
Hypothesis 1. The positive relationship between predeces-
sors’ cumulative lending amount and successors’ lending
amount toward a listing is stronger for predecessors pre-
sented with an anonymous username than those with a
real-seeming one.
2.3. The Moderating Role of Risk-Related Factors
To facilitate decision making with little information,
lenders observe and learn from the decisions made by
preceding peers. As we previously argued, during
this process, lenders tend to demonstrate dissimilar
herding momentum toward predecessors who are
presented with different states of perceived anonymi-
ty. Given the main purpose of engaging in observa-
tional learning is to better evaluate a listing’s potential
risk, it is anticipated that lenders may further adjust
such a differentiation behavior as the perceived risk
level varies across listings. In this subsection, we
hypothesize that the predicted effect of perceived ano-
nymity on herding (Hypothesis 1) is likely to be mod-
erated by two listing characteristics that could signal
the investment risk to potential lenders: fundraising
stage and listing credit grade.
In our context, a listing can materialize into a loan
only when the requested amount is reached. Under
this mechanism, a listing situated in the early stage
of its fundraising period is likely to be regarded
as less certain with respect to loan materialization
(Burtch et al. 2018, Kuppuswamy and Bayus 2018),
compared with the other otherwise identical listing
situated in its late fundraising stage. When dealing
with such a listing, a prospective lender may tend to
rely more on observational learning, making the pre-
dicted positive effect of perceived anonymity (Hy-
pothesis 1) more pronounced. Consistent with this
logic, prior literature on online reviews holds that
the uncertainty of a product is related to its life cycle,
which then affects the value of the credibility signal
conveyed by reviewers (e.g., Yazdani et al. 2018). In
addition, a listing in its early fundraising stage re-
ceives fewer transactions such that a prospective

6 Information Systems Research, Articles in Advance, pp. 1–17, © 2021 INFORMS
lender can differentiate predecessors per their ano-
nymity state in a relatively manageable manner.
However, as the listing progresses into its late fund-
raising stage, the uncertainty of loan materialization di-
minishes and the number of total transactions increases,
together weakening lenders’ motivation of engaging in
observational learning. Therefore, we expect that the
differentiated herding behavior toward distinct prede-
cessor groups would diminish in the late stage of a list-
ing’s fundraising process.
Hypothesis 2 (a). The predicted positive effect of predeces-
sors’ perceived anonymity on herding is weaker in the late
fundraising stage than in the early fundraising stage.
To help mitigate information asymmetry, the plat-
form derives and assigns a credit grade to each list-
ing based on various listing attributes as well as the
financial status of the borrower (Iyer et al. 2015).
Such an aggregate measure is found to have the abil-
ity to predict loan default (Herzenstein et al. 2008,
Zhang and Liu 2012) and thus can signal the poten-
tial risk of a listing. As such, a prospective lender
may regard a listing that has a low credit grade as
riskier with respect to the likelihood of default, cete-
ris paribus. Given that anonymous predecessors are
perceived as more credible in assessing a listing’s
potential risk under our theoretical framework (Hy-
pothesis 1), lenders (as observational learners) are
likely to pay more attention to their lending deci-
sions when dealing with a listing that might ulti-
mately turn into delinquency or default. By contrast,
for a listing that has a high credit grade, prospective
lenders would reduce their dependence on the sig-
nal originated from different predecessor groups
due to reduced concern on the investment risk.
Therefore, we anticipate that the predicted positive
relationship between predecessor anonymity and
herding magnitudes would be stronger on listings
that receive low credit grades.
Hypothesis 2 (b). The predicted positive effect of predeces-
sors’ perceived anonymity on herding is stronger on list-
ings with high-risk credit grades than those with low-risk
credit grades.
3. Methods and Data
3.1. Empirical Context
Our research context is one of the leading debt-based
crowdfunding platforms in Asia. As of August 2019,
the platform has been in operation for nine years, fa-
cilitating nearly 1.67 million transactions. In what fol-
lows, we provide the details about the lending process
and the relevant information from a prospective lend-
er’s perspective.
Jiang et al.: The Effect of Perceived Anonymity on Herding
To participate in any activity on the platform, an in-
dividual first needs to register for an account and
choose a username. The platform imposes nearly no
naming restriction such that a new user can freely use
a combination of Chinese characters, English letters,
and numbers as a username. The lending process
starts with a borrower submitting a loan application,
in which the borrower is required to provide informa-
tion such as the requested amount, the loan duration,
the fixed interest rate the borrower would like to offer,
the borrower’s age, as well as the listing title. The
borrower could also voluntarily disclose additional
personal information such as education, marriage,
employment, and income. Loan applications ap-
proved by the platform would turn into listings for
prospective lenders’ browsing. For an interested list-
ing, a prospective lender can observe the listing- and
borrower-specific characteristics in the top section of
the listing’s main page. In the bottom section of the
same page, the lender can view detailed information
about all previous lending transactions, including the
preceding lender’s username, the lending amount,
and the timestamp of each transaction.3 Unlike other
crowdfunding sites on which the transactional infor-
mation is partially available (e.g., the context studied
in Burtch et al. 2016), the entire transaction log is dis-
played on a single page for easy browsing. This con-
textual feature provides us an ideal setting to quantify
how a successor’s lending amount is associated with
the cumulative lending amount collectively made by
predecessors.
3.2. Data Description
We collect data from all listings that were posted on
the platform from March to August in 2015.4 To make
our sample comparable to those posted on other ma-
jor sites (e.g., LendingClub and Prosper), we restrict
our attention to listings that were requested online
and were not guaranteed by offline institutions.5 We
also restrict our sample to listings that received at
least two transactions. The resulting data set for our
main analysis contains 3,504 listings, which attracted
a total of 64,925 transactions from 17,816 distinct
lenders. We construct our panel data at the transaction
level. For each transaction, we observe characteristics
at the listing, the borrower, and the transaction level.
Listing characteristics include the requested amount,
the interest rate, the loan duration, and the credit
grade. Borrower characteristics include each bor-
rower’s username, demographic information, and the
lending history on the platform. Transaction charac-
teristics include the lender’s username, the lending
amount, the sequential index of the transaction associ-
ated with a given loan, and the channel through
which the transaction is made (desktop vs. mobile).
Jiang et al.: The Effect of Perceived Anonymity on Herding
Information Systems Research, Articles in Advance, pp. 1–17, © 2021 INFORMS
Next, we discuss how perceived anonymity is
operationalized and how it affects the perception of
beholders. In the spirit of prior work on online
anonymity (Nissenbaum 1999, Qian and Scott 2007,
Tsikerdekis 2012), we ask research assistants to
label all usernames appearing in our data as either
real-seeming or anonymous.6 Based on the theoretical
background provided in Section 2, we argue that dis-
tinct types of usernames may carry dissimilar infor-
mational values, which may lead to heterogeneous
perceptions held by beholders (Hite et al. 2014)—the
primary focus of this research.
It is imperative for us to emphasize that the objec-
tive of this research is to explore the informational
role of perceived anonymity in shaping the perceived
source credibility from a subsequent lender’s perspec-
tive. Although it is likely for a user to choose a real-
seeming username that doesn’t match the user’s real
name, such a scenario is less of a concern in our re-
search framework. Since the platform discloses no
identity cue about a lender, there is no way for other
users to discern whether a real-seeming username in-
deed reflects the focal lender’s real name. As such, a
subsequent lender is likely to regard predecessors
who are presented with real-seeming usernames as
the same kind, and consequently forms the same atti-
tude toward them. To see if lenders demonstrate such
differentiation behavior in our data, we calculate the
partial correlation between the lending amount of a
transaction and the cumulative lending amount made
by different predecessor groups given the percentage
of the unfunded amount. The correlation is positive
and significant for the all-predecessor group (0.052,
p < 0.001). If we split all predecessors based on their
username types, we see correlations deviating to some
extent, with 0.043 for the anonymous group (p < 0.001)
and 0.023 for the real-seeming group (p < 0.001).
We report the definition and summary statistics of
our main variables in Table 1. An average listing re-
quests $3,521 with an interest rate of 11.8% for a peri-
od of 14 months and receives an amount of $161 per
transaction. The dummy variable for credit risk takes
a value of 1 if a listing is assigned with a low credit
grade by the platform (i.e., HR), and 0 otherwise (i.e.,
AA, A, B, C, D, or E).7 The age of borrowers ranges
from 22 to 56, with an average value of 31.8. About
78.2% of the borrowers hold a bachelor’s degree or
higher, 58.5% are married, 45.8% have worked for
more than three years, and 29.3% have a monthly in-
come greater than $3,072. Among all borrowers,
12.3% of them present themselves with a real-seeming
username, whereas the number is 9.6% for lenders.
3.3. Model Specification and Estimation
A straightforward test of herding is to look for the se-
quential correlation in the lending amount. If herding
7
exists, we should expect to see a positive correlation
between (1) the lending amount of a given transaction
and (2) the cumulative lending amount the listing re-
ceived prior to that transaction. Such examination of
transaction-level dynamics based on the sequential or-
der is commonly used in studying herding (Banerjee
1992, Bikhchandani et al. 1992, Zhang 2010). Leverag-
ing the detailed transaction log, we can observe the se-
quential order of all transactions and construct our
herding measure—the cumulative lending amount
made to the same listing up to any given point of
time.
We next specify our amount model. Let i denote the
lender and j denote the listing. The logarithm of the
lending amount of the tth transaction made to listing j
by lender i, yijt, is modeled as:
yijt  α + βYj,t−1 + Xjt γ1 + Zj γ2 + Wi γ3 + λw + εijt , (1)
where t  2, . . . , T and T is the total number of trans-
actions for listing j, and Yj, t−1 denotes the logarithm of
the cumulative lending amount listing j has received
prior to the arrival of its tth transaction. The parame-
ter of interest is β, which captures the magnitude of
herding. A positive and significant estimate of β will
indicate that subsequent lenders’ lending amount is
increasing in the cumulative lending amount made to
the same listing, confirming the existence of herding.
The time-variant listing characteristics Xjt include
the cumulative number of transactions prior to the
current transaction (Lag Num of Transactions, prior to
t), the time gap between two consecutive transactions
(Time Interval), and whether the transaction is made
through a desktop (Lender Device). To control for the
payoff externality effect at the listing level, we include
the percentage of the unfunded amount prior to the
current transaction (Lag Percent Needed) in Xjt. A
dummy variable is also included to indicate whether
a listing is situated in its late fundraising stage if the
percentage of the unfunded amount is less than 50%
(Late Fundraising Stage  1). Time-invariant listing and
borrower characteristics Zj includes Amount Requested,
Interest Rate, Loan Duration, Credit Risky, Paid Loans,
Overdue Loans, Age, Education, Marriage, Employment
Length, Monthly Income, and a dummy variable indi-
cating the borrower perceived anonymity (Real Sm
Borrower). We include in Wi a dummy variable that in-
dicates the lender perceived anonymity (Real Sm Lend-
er). We also insert week fixed effects λw, which helps
capture potential unobserved temporal dynamics be-
yond the listing level. The error term is εijt.
To further differentiate the heterogeneous herding
effects induced by different groups of predecessors,
a r
we decompose Yj,t−1 and use Yj,t−1 and Yj,t−1 to denote
the logarithm of the cumulative lending amount made
by predecessors who present themselves with anony-
mous and real-seeming usernames, respectively.
 Jiang et al.: The Effect of Perceived Anonymity on Herding
8 Information Systems Research, Articles in Advance, pp. 1–17, © 2021 INFORMS

Table 1. Variable Description and Summary

Variable Description Mean Standard deviation Min Max

Listing-level characteristics
Amount Requested Listed amount of a listing ($) 3,521 2,827 461 30,704
Interest Rate Interest rate provided by the borrower (%) 11.801 1.067 7 13
Loan Duration Number of months the borrower makes payments 14.078 7.348 3 36
Title Length Number of characters in the title of a listing 8.218 3.593 2 29
Credit Risky 1 if a listing’s credit grade is HR, 0 otherwise 0.422 0.494 0 1
Paid Loans Number of paid loans made by the borrower 0.774 1.706 0 38
Overdue Loans Number of overdue loans made by the borrower 1.429 2.201 0 18
Num of Transactions Total number of transactions a listing receives 20.926 15.851 2 230
Borrower-level characteristics
Real Sm Borrower 1 if a borrower presented with a real-seeming username, 0 otherwise 0.123 0.329 0 1
Age Age of the borrower at the time of posting the listing 31.820 6.288 22 56
Education Unreported (base) 0.046 0.210 0 1
A borrower has a high school diploma or lower (low) 0.172 0.377 0 1
A borrower has a college diploma or higher (high) 0.782 0.413 0 1
Marriage Unreported (base) 0.010 0.102 0 1
A borrower is married (married) 0.585 0.493 0 1
A borrower is single, divorced, or widowed (others) 0.405 0.491 0 1
Employment Length Unreported (base) 0.131 0.337 0 1
Length of employment is less than three years (short) 0.411 0.492 0 1
Length of employment is more than three years (long) 0.458 0.498 0 1
Monthly Income Unreported (base) 0.124 0.330 0 1
Monthly income is less than ¥ 20,000 (~$3,072) (low) 0.583 0.493 0 1
Monthly income is more than ¥ 20,000 (high) 0.293 0.455 0 1
Transaction-level characteristics
Percent Needed Percentage of unfunded amount (%) 51.860 33.140 0 100
Late Fundraising Stage 1 if percentage of unfunded amount is less than 50%, 0 otherwise 0.519 0.500 0 1
Amount Invested Amount of a transaction ($) 161 383 8 13,817
Real Sm Lender 1 if a lender presented with a real-seeming username, 0 otherwise 0.096 0.295 0 1
Lender Device 1 if a transaction is made via desktop, 0 via mobile 0.429 0.495 0 1
Note. 1 USD  6.51 CNY, www.oanda.com (accessed January 10, 2018).

Accordingly, we can enhance (1) as:
yijt  α + βa Yj,t−1
a
+ βr Yj,t−1
r
+ Xjt γ1 + Zj γ2 + Wi γ3 + λw
+ εijt ,
where βa and βr measure the subsequent lender’s
herding momentum toward preceding peers who
are presented with anonymous and real-seeming
usernames, respectively. The relative magnitude of
these two coefficients will inform us of the role of
perceived anonymity in lenders’ observational learn-
ing process.
Multiple Sources of Unobserved Heterogeneity. Our
amount model so far explores herding at the trans-
action level. However, the variation in lending
amounts could be further attributed to some
higher-level sources of variation where each trans-
action resides. For example, there could be unob-
served heterogeneity beyond the observable listing
characteristics we have controlled for. Similarly,
lenders with different characteristics may demon-
strate different funding behavior. Ignoring these
unobserved heterogeneities could lead to an incon-
sistent estimation and misleading inference. To
deal with this concern, we use a cross-classified
multilevel modeling approach (Gelman and Hill
2006, Rasbash and Browne 2008). In our data set,
(2) we observe transactions made to different listings
from distinct lenders. Thus, transactions can be
viewed as lowest-level units nested within two
higher levels: listings and lenders. Listings and
lenders are not nested within each other (i.e., cross-
classified), because neither all transactions made by a
given lender are made to the same listing, nor do all
transactions made to a given listing come from the same
lender. Accordingly, we enhance (1) and (2) by inserting
lender-specific intercepts (ηi) and listing-specific inter-
cepts (ηj) to better control for unobserved heterogeneity
at the lender and listing level, respectively.
Participation Decisions of Lenders. Another concern is
that our analysis of lending amount does not consider
individuals who do not participate in lending ($0). Po-
tential selection bias may arise if factors that affect a
lender’s choice of which listing to participate also af-
fect the lender’s decision of how much to lend. The
following steps are taken to address this concern.
First, for each lender who made at least one transac-
tion on a given day, we construct the lender’s
Jiang et al.: The Effect of Perceived Anonymity on Herding
Information Systems Research, Articles in Advance, pp. 1–17, © 2021 INFORMS 9

consideration set using all listings that were active on Table 2. Results of Participation Decision
that day.8 Then, we adopt a two-stage estimation pro-
Variable Coefficient (Std. err.)
cedure (Heckman 1979) to correct the potential selec-
tion bias. In the first stage, we model an individual log Amount Requested 0.2981*** (0.0032)
lender’s decision of whether to participate in a listing Interest Rate −0.0569*** (0.0041)
Loan Duration 0.0065*** (0.0006)
(in the lender’s consideration set) using a standard
Title Length 0.0016*** (0.0005)
probit specification. We refer to this specification as Max Title Length −0.0071** (0.0030)
the participation model. The dependent variable here is Credit Risky −0.0284** (0.0138)
a vector of dummy variables indicating whether a Paid Loans −0.0015 (0.0012)
lender participates in a listing, and the independent Overdue Loans 0.0010 (0.0008)
Real Sm Borrower (1  yes) −0.0100* (0.0056)
variables include corresponding listing and lender Age 0.0623* (0.0326)
characteristics. To satisfy the exclusion restriction, we Education (low) 0.0036 (0.0102)
include the maximum title length of other listings, Education (high) −0.0303*** (0.0091)
that is, all except the focal listing, in a lender’s consid- Marriage (married) −0.0009 (0.0198)
eration set (Max Title Length). This variable choice can Marriage (others) 0.0107 (0.0198)
Employment Length (short) 0.0004 (0.0068)
be justified by lenders’ decision-making processes. Employment Length (long) 0.0101 (0.0068)
When a prospective lender chooses which listing to Monthly Income (low) 0.0091 (0.0070)
participate, the lender is likely to evaluate all listings Monthly Income (high) −0.0067 (0.0073)
in the consideration set in an exploratory manner Real Sm Lender (1  yes) 0.0207*** (0.0058)
(akin to online shoppers’ behavior discussed by Janis- Intercept −2.1042*** (0.1492)
Number of observations 1,909,057
zewski 1998). At this stage, useful soft information Log-likelihood −272,726.6
can be inferred from listing titles to help grasp a gen- Akaike Information Criterion (AIC) 545,715.2
eral impression about a specific listing, and to also al- Bayesian Information Criterion (BIC) 547,347.7
low a quick comparison with plausible alternatives of Note. Standard errors are reported in parentheses.
which the listing with the maximum title length is *p < 0.1; **p < 0.05; ***p < 0.01.
likely a good representative. When proceeding to the
next stage to determine a lending amount, the lender parameters that reflect the degree of heterogeneity at
will evaluate a focal listing’s attributes in a more de- the lender and listing level, respectively.
liberate manner, focusing on detailed information Correspondingly, the full model for testing the role
posted on that listing’s main page. The information of perceived anonymity can be expressed as:
about other listings, though not directly displayed, is
not expected to play a significant role in this decision. yijt  α + βa Yj,t−1
a
+ βr Yj,t−1
r
+ Xjt γ1 + Zj γ2 + Wi γ3
Hence, the maximum title length of alternative listings + γ4 λ̂ ij + ηi + ηj + λw + εijt : (4)
only indirectly affects the lending amount through the
participation decision. To further justify our variable To empirically test Hypotheses 2(a) and 2(b), we en-
choice from a statistical perspective, we follow the hance (4) by allowing our herding measures to inter-
strategy of Kumar et al. (2018) and include Max Title act with Late Fundraising Stage and Credit Risky.
Length in our models to test the previous justification.
The results indicate that the exclusion restriction is
4. Results
well satisfied: the maximum title length of alternative
The estimation results from the participation model
listings has good explanatory power in the participation
are presented in Table 2. The results show that lenders
model (reported in Table 2) but turns insignificant in
are more likely to choose listings that request a larger
the amount model (reported in Table A1 of the online
amount and pay back over a longer period, and are
appendix). Finally, we compute the inverse Mill’s ratios
submitted by a senior borrower. We also find that
(λ̂ ij ) at the lender-listing level and include them as an
lenders’ participation decisions are negatively associ-
additional variable in the amount model for the second-
stage estimation. ated with interest rates and credit riskiness, providing
Based on the previous discussion, we can rewrite evidence that lenders draw rational inferences based
the full model for our herding analysis as: on various observable cues. As expected, the effect of
Title Length is positive and significant, whereas Max
yijt  α + βYj,t−1 + Xjt γ1 + Zj γ2 + Wi γ3 + γ4 λ̂ ij + ηi + ηj Title Length of alternative listings negatively affects
the selection of the focal listing.
+ λw + εijt , (3)
Table 3 presents the estimation results from the
where ηi ~ N(0, σ2ηi ) and ηi ~ N(0, σ2ηj ). The means for amount model. We first look at whether herding be-
both parameters are zeros as any nonzero mean will havior exists in our context. Column (1) of Table 3 re-
be absorbed into α. σ2ηi and σ2ηj are the variance ports the parameter estimates specified in (3). The
 Jiang et al.: The Effect of Perceived Anonymity on Herding
10 Information Systems Research, Articles in Advance, pp. 1–17, © 2021 INFORMS

estimated coefficient of Y is positive and significant,
confirming the existence of herding among lenders.
We proceed to test our hypothesis about the role of
perceived anonymity in herding (Hypothesis 1). Col-
umn (2) of Table 3 reports parameter estimates from
(4). The estimated coefficients of Ya and Yr are both
positive and statistically significant. Specifically,
subsequent lenders demonstrate a stronger herding
magnitude (βa  0.0393) toward predecessors who are
presented with an anonymous username, compared
with those presented with a real-seeming one (βr 
0.0138). To verify whether the two estimates are statis-
tically different from each other, we test a null hy-
pothesis: βa  βr. The result rejects the null hypothesis
(p < 0.001), lending empirical support to our predic-
tion that perceived anonymity intensifies herding,
Therefore, Hypothesis 1 is supported.
The results for testing our second set of hypotheses
are presented in column (3) of Table 3. We find that
the estimated coefficients of the terms interacting
with Late Fundraising Stage are both negative and sta-
tistically significant. A quick calculation on estimated
coefficients reveals that the difference in herding
magnitude toward two distinct predecessor groups is
further attenuated in the late stage. These results lend
empirical support to our prediction that observational
learning bears less informational value and hence
turns less impactful in the late stage of the fundraising
process (Hypothesis 2(a) is supported). Nonetheless,
we do not observe lenders demonstrating any system-
atic differences when dealing with listings that differ
in credit grades (Hypothesis 2(b) is not supported).
This result, with the insignificant main effect of Credit
Risky, together imply that lenders tend to pay little

Table 3. Main Results

Variable (1) Herding (2) Perceived anonymity (3) Risk-related factors

Y: All Predecessors 0.0462*** (0.0035)

Ya: Anonymous Predecessors 0.0393*** (0.0033) 0.0373*** (0.0034)
Yr: Real-seeming Predecessors 0.0138*** (0.0021) 0.0188*** (0.0030)
Lag Percent Needed 0.0101*** (0.0004) 0.0103*** (0.0004) 0.0103*** (0.0004)
Lag Num of Transactions 0.0006 (0.0004) 0.0002 (0.0004) 0.0011** (0.0004)
Late Fundraising Stage 0.2422*** (0.0187) 0.2389*** (0.0187) 0.4945*** (0.0726)
Time Interval −0.0065*** (0.0018) −0.0061*** (0.0018) −0.0059*** (0.0018)
Real Sm Lender (1  yes) 0.0546* (0.0331) 0.0599* (0.0331) 0.0583* (0.0331)
Lender Device 0.1183*** (0.0144) 0.1185*** (0.0144) 0.1181*** (0.0144)
log Amount Requested 0.0296** (0.0133) 0.0327** (0.0134) 0.0229 (0.0140)
Interest Rate 0.0474*** (0.0118) 0.0481*** (0.0119) 0.0499*** (0.0121)
Loan Duration −0.0076*** (0.0016) −0.0076*** (0.0017) −0.0077*** (0.0017)
Title Length −0.0015 (0.0014) −0.0016 (0.0014) −0.0017 (0.0014)
Credit Risky −0.0619 (0.0397) −0.0574 (0.0399) −0.0585 (0.0407)
Paid Loans −0.0005 (0.0037) −0.0001 (0.0037) −0.0005 (0.0038)
Overdue Loans −0.0014 (0.0023) −0.0013 (0.0023) −0.00004 (0.0036)
Real Sm Borrower (1  yes) −0.0081 (0.0156) −0.0077 (0.0157) −0.0074 (0.0159)
Age 0.0797 (0.0906) 0.0806 (0.0911) 0.0766 (0.0925)
Education (low) 0.0384 (0.0284) 0.0396 (0.0285) 0.0391 (0.0290)
Education (high) 0.0434* (0.0255) 0.0456* (0.0256) 0.0467* (0.0260)
Marriage (married) 0.0789 (0.0549) 0.0708 (0.0552) 0.0748 (0.0559)
Marriage (others) 0.0677 (0.0549) 0.0599 (0.0551) 0.0633 (0.0558)
Employment Length (short) 0.0041 (0.0187) 0.0052 (0.0188) 0.0048 (0.0190)
Employment Length (long) −0.0072 (0.0184) −0.0063 (0.0185) −0.0071 (0.0188)
Monthly Income (low) −0.0068 (0.0191) −0.0080 (0.0192) −0.0088 (0.0195)
Monthly Income (high) −0.0048 (0.0200) −0.0057 (0.0201) −0.0055 (0.0204)
Ya × Late Fundraising Stage −0.0277*** (0.0094)
Yr × Late Fundraising Stage −0.0127*** (0.0038)
Ya × Credit Risky −0.0004 (0.0038)
Yr × Credit Risky −0.0021 (0.0048)
Intercept 3.2925*** (0.2827) 3.3155*** (0.2835) 3.2175*** (0.2877)
λ̂: Heckman correction term −0.2505*** (0.0344) −0.2549*** (0.0345) −0.2561*** (0.0350)
σ2ηi : Lender-level variation 1.2151 1.2152 1.2143
σ2ηj : Listing-level variation 0.0208 0.0215 0.0237
Week fixed effects Yes Yes Yes
Number of observations 64,925 64,925 64,925
Log-likelihood −101,572.5 −101,571.8 −101,561.1
AIC 203,250.9 203,251.6 203,238.2
BIC 203,732.2 203,742.0 203,764.9
Note. Standard errors are reported in parentheses.
*p < 0.1; **p < 0.05; ***p < 0.01.
Jiang et al.: The Effect of Perceived Anonymity on Herding
Information Systems Research, Articles in Advance, pp. 1–17, © 2021 INFORMS
attention to credit grades assigned by the studied plat-
form when making the decision of how much to lend.
Our finding challenges the conventional wisdom
concerning how anonymity is perceived in the eye of
beholders. A predominant notion that has long been
used to justify the negative effect of anonymity is that
one must be identified (Marx 1999). Extensive litera-
ture has documented that anonymity provides few
cues about the information source and therefore
would undermine the trustworthiness perceived by
others (Postmes et al. 2001, Jiang et al. 2013). In con-
trast to this point of view, our results lend empirical
support to a competing argument in which anonymity
poses a positive impact on source credibility. As we
previously argued, this seemingly counterintuitive
finding is quite reasonable due to lenders’ strong de-
sire for privacy in an investment-oriented context (Xu
et al. 2008, Bansal et al. 2016, Pritchard 2020) like debt-
based crowdfunding. Our finding complements an
emerging stream of research that uncovers the posi-
tive role of perceived anonymity in various online set-
tings, such as the trustworthiness of online reviews
(Hong and Li 2017), quality of online posts (Lee et al.
2014), and online charity giving (Peacey and Sanders
2014). Moreover, we show that the effect of perceived
anonymity on herding is further moderated by risk-
related characteristics of various listings. This finding
supports the established notion that lenders are ratio-
nal observational learners who use observable attrib-
utes to moderate their perception of preceding peers’
credibility (Zhang and Liu 2012), endorsing observa-
tional learning as the behavioral mechanism that
drives herding.
4.1. Robustness Checks
In this section, we perform a series of robustness
checks. Specifically, we use an alternative measure for
herding, include additional listing sample, account for
serial correlation at the listing level, employ general-
ized method of moments (GMM) estimators, and esti-
mate fixed-effects models. The consistent parameter
estimates demonstrate the robustness of our main
findings.
4.1.1. Alternative Herding Measure. In our main anal-
ysis, we use the logarithm of the cumulative lending
amount made to a listing as our herding measure. To
rule out the possibility that the uncovered heteroge-
neous herding effects are attributed to a potential
change in the base of two predecessor groups, we use
the logarithm of the average lending amount of previ-
ous transactions in this robustness check. Following a
similar procedure, we decompose this alternative
herding measure according to the group membership.
As column (1) of Table 4 presents, the new results cor-
roborate our main findings.
11
4.1.2. Alternative Listing Sample. As previously dis-
cussed in Section 3.2, we restrict our sample to listings
that are comparable to those available on other major
platforms. In this robustness check, we re-estimate
our models using the whole sample, which contains
all listings posted during our sample period (i.e., in-
cluding listings that are requested offline and guaran-
teed by offline institutions). The results, reported in
column (2) of Table 4, are largely consistent with our
main analysis.
4.1.3. First-Order Autoregression. So far, we assume
that there is no serial correlation among error terms. If
this assumption is violated, error terms would be cor-
related with lagged independent variables, introduc-
ing bias into the estimation. To address this potential
issue, we fit multilevel models that allow for first-
order autoregressive correlation structure for resid-
uals at the listing level. As reported in column (1) of
Table 5, despite changes in magnitude, the coefficients
of our key independent variables continue to support
our main findings.
4.1.4. Multilevel GMM. If the random intercepts in our
multilevel models are correlated with the herding
measure, our estimation could be biased due to the
endogeneity stemming from such correlations. To ad-
dress this concern, we fit models using the multilevel
generalized method of moments (GMM) estimators
(Kim and Frees 2007). The results are robust, as re-
ported in column (2) of Table 5.
4.1.5. Reflection Problem. The reflection problem
(Manski 1993) describes that the observed social
interactions can arise from three sources: endogenous
effects, exogenous (contextual) effects, and correlated
effects. To identify social influence (such as herding in
our study) that arises from endogenous effects, re-
searchers must account for the other two sources, that
is, exogenous and correlated effects. Failure to control
those two effects could lead to endogenous group for-
mation due to homophily (Kim and Viswanathan 2019).
Put in our research framework, exogenous (contex-
tual) effects could arise when the observed herding is
caused by the shared characteristics at the listing level.
For example, the correlated behavior between a pre-
decessor and a successor may be driven by some un-
observed listing characteristics, such as project type.
In this case, the inclusion of listing fixed effects can
help control for homophily driven by listing-related
factors (Kim and Viswanathan 2019). On the other
hand, correlated effects could also arise when the ob-
served correlated behavior is caused by similarity
among lenders. For example, the correlated behavior
between a predecessor and a successor may be driven
by some unobserved lender characteristics, such as
 Jiang et al.: The Effect of Perceived Anonymity on Herding
12 Information Systems Research, Articles in Advance, pp. 1–17, © 2021 INFORMS

Table 4. Alternative Herding Measure and Listing Sample

(1) Alternative herding measure (2) Alternative listing sample

Variable Perceived anonymity Risk-related factors Perceived anonymity Risk-related factors

Ya: Anonymous Predecessors 0.0462*** (0.0040) 0.0439*** (0.0041) 0.0534*** (0.0030) 0.0471*** (0.0031)
Yr: Real-seeming Predecessors 0.0170*** (0.0023) 0.0224*** (0.0030) 0.0222*** (0.0020) 0.0321*** (0.0027)
Lag Percent Needed 0.0097*** (0.0003) 0.0098*** (0.0003) 0.0145*** (0.0003) 0.0145*** (0.0003)
Lag Num of Transactions 0.0013*** (0.0004) 0.0023*** (0.0004) 0.0046*** (0.0004) 0.0063*** (0.0004)
Late Fundraising Stage 0.2308*** (0.0186) 0.4896*** (0.0724) 0.2772*** (0.0161) 0.7447*** (0.0708)
Time Interval −0.0066*** (0.0018) −0.0065*** (0.0018) −0.0003 (0.0017) −0.0007 (0.0017)
Real Sm Lender (1  yes) 0.0608* (0.0331) 0.0588* (0.0331) 0.0403 (0.0328) 0.0369 (0.0328)
Lender Device 0.1192*** (0.0144) 0.1186*** (0.0144) 0.1142*** (0.0136) 0.1136*** (0.0136)
log Amount Requested 0.0235* (0.0130) 0.0142 (0.0136) 0.0259** (0.0127) 0.0212 (0.0132)
Interest Rate 0.0461*** (0.0116) 0.0478*** (0.0118) 0.0807*** (0.0135) 0.0830*** (0.0138)
Loan Duration −0.0073*** (0.0016) −0.0075*** (0.0016) −0.0099*** (0.0019) −0.0101*** (0.0019)
Title Length −0.0016 (0.0014) −0.0017 (0.0014) −0.0011 (0.0016) −0.0011 (0.0016)
Credit Risky −0.0596 (0.0386) −0.0613 (0.0394) −0.0100 (0.0452) −0.0076 (0.0460)
Paid Loans −0.0004 (0.0036) −0.0007 (0.0037) 0.0014 (0.0038) 0.0006 (0.0039)
Overdue Loans −0.0015 (0.0023) −0.0007 (0.0035) −0.0022 (0.0026) −0.0009 (0.0039)
Real Sm Borrower (1  yes) −0.0079 (0.0153) −0.0075 (0.0155) −0.0055 (0.0174) −0.0061 (0.0177)
Age 0.0836 (0.0886) 0.0804 (0.0901) 0.0662 (0.0980) 0.0766 (0.0996)
Education (low) 0.0396 (0.0278) 0.0394 (0.0282) 0.0484* (0.0285) 0.0465 (0.0289)
Education (high) 0.0459* (0.0250) 0.0472* (0.0254) 0.0624*** (0.0231) 0.0638*** (0.0234)
Marriage (married) 0.0722 (0.0539) 0.0769 (0.0546) 0.0330 (0.0241) 0.0352 (0.0244)
Marriage (others) 0.0619 (0.0539) 0.0660 (0.0546) 0.0213 (0.0264) 0.0220 (0.0268)
Employment Length (short) 0.0064 (0.0183) 0.0059 (0.0186) 0.0021 (0.0214) 0.0001 (0.0217)
Employment Length (long) −0.0049 (0.0181) −0.0059 (0.0183) −0.0272* (0.0157) −0.0297* (0.0159)
Monthly Income (low) −0.0104 (0.0187) −0.0112 (0.0190) −0.0173 (0.0227) −0.0182 (0.0230)
Monthly Income (high) −0.0065 (0.0195) −0.0064 (0.0198) −0.0286 (0.0232) −0.0293 (0.0235)
Ya × Late Fundraising Stage −0.0276*** (0.0093) −0.0247*** (0.0092)
Yr × Late Fundraising Stage −0.0135*** (0.0036) −0.0115*** (0.0033)
Ya × Credit Risky 0.0004 (0.0037) 0.0003 (0.0038)
Yr × Credit Risky −0.0027 (0.0048) −0.0022 (0.0044)
Intercept 3.3082*** (0.2789) 3.2103*** (0.2832) 2.6309*** (0.3032) 2.4457*** (0.3085)
λ̂: Heckman correction term −0.2506*** (0.0337) −0.2519*** (0.0342) −0.1917*** (0.0464) −0.1909*** (0.0471)
σ2ηi : Lender-level variation 1.2159 1.2150 1.2068 1.2053
σ2ηj : Listing-level variation 0.0177 0.0199 0.0517 0.0548
Week fixed effects Yes Yes Yes Yes
Number of observations 64,925 64,925 86,662 86,662
Log-likelihood −101,571.4 −101,559.3 −134,934.1 −134,886.0
AIC 203,250.8 203,234.6 269,976.2 269,888.0
BIC 203,741.1 203,761.3 270,483.1 270,432.5
Note. Standard errors are reported in parentheses.
*p < 0.1; **p < 0.05; ***p < 0.01.

disposable income. In this case, adding lender fixed
effects can help control for homophily driven by
lender-related factors (Burtch et al. 2016). Following
the strategy employed by Kim and Viswanathan
(2019) and Burtch et al. (2016), we conducted a robust-
ness check by including fixed effects at both the listing
and lender level. As shown in Table 6, the observed
differentiated herding behavior toward distinct prede-
cessor groups continues to prevail.
4.2. Additional Analyses
We conduct additional analyses that may help re-
searchers further understand or expand our study.9
Specifically, we differentiate the herding effect from
earlier or recent predecessors, account for the sentiment
contained in usernames, and consider a subsequent
lender’s investment history during our observation
period.
4.2.1. Primacy or Recency Effects. Prior literature
suggests that individuals tend to remember informa-
tion that is presented at the beginning (primacy) and
end (recency) of a list (e.g., Haugtvedt and Wegener
1994, Murphy et al. 2017). In our context, if the prima-
cy effect or the recency effect exists, lenders may rely
more on the easiest or the most recent transactions ap-
pearing in the transaction log. To account for this pos-
sibility, we re-estimate (4) by investigating the role of
the (five or 10) earliest transactions (reported in Table
A3 of the online appendix) and the (five or 10) most
recent preceding transactions prior to a given lender
funding activity (reported in Table A4 of the online
Jiang et al.: The Effect of Perceived Anonymity on Herding
Information Systems Research, Articles in Advance, pp. 1–17, © 2021 INFORMS 13

Table 5. Alternative Model Specifications

(1) First-order autoregression (2) GMM

Variable Perceived anonymity Risk-related factors Perceived anonymity Risk-related factors

Ya: Anonymous Predecessors 0.0795*** (0.0045) 0.0761*** (0.0047) 0.0761*** (0.0045) 0.0731*** (0.0047)
Yr: Real-seeming Predecessors 0.0252*** (0.0029) 0.0322*** (0.0040) 0.0240*** (0.0029) 0.0301*** (0.0040)
Lag Percent Needed 0.0150*** (0.0005) 0.0152*** (0.0005) 0.0142*** (0.0005) 0.0143*** (0.0005)
Lag Num of Transactions 0.0029*** (0.0005) 0.0047*** (0.0005) 0.0014*** (0.0005) 0.0024*** (0.0005)
Late Fundraising Stage 0.3024*** (0.0251) 0.7488*** (0.0983) 0.2975*** (0.0255) 0.5921*** (0.0980)
Time Interval −0.0036 (0.0024) −0.0034 (0.0024) −0.0044* (0.0024) −0.0043* (0.0024)
Real Sm Lender (1  yes) 0.0898*** (0.0183) 0.0880*** (0.0183) 0.0914*** (0.0186) 0.0901*** (0.0186)
Lender Device 0.1160*** (0.0117) 0.1156*** (0.0117) 0.1146*** (0.0119) 0.1139*** (0.0119)
log Amount Requested 0.0094 (0.0180) 0.0074 (0.0191) 0.0038 (0.0172) 0.0078 (0.0180)
Interest Rate 0.0759*** (0.0158) 0.0782*** (0.0163) 0.0861*** (0.0148) 0.0878*** (0.0150)
Loan Duration −0.0142*** (0.0022) −0.0144*** (0.0022) −0.0154*** (0.0020) −0.0155*** (0.0021)
Title Length −0.0029 (0.0019) −0.0030 (0.0020) −0.0028 (0.0018) −0.0028 (0.0018)
Credit Risky −0.0353 (0.0536) −0.0353 (0.0555) −0.0528 (0.0478) −0.0525 (0.0485)
Paid Loans −0.0044 (0.0050) −0.0045 (0.0051) −0.0049 (0.0043) −0.0052 (0.0044)
Overdue Loans −0.0046 (0.0032) −0.0073 (0.0050) −0.0030 (0.0030) −0.0044 (0.0046)
Real Sm Borrower (1  yes) −0.0107 (0.0214) −0.0111 (0.0219) −0.0111 (0.0201) −0.0115 (0.0203)
Age 0.0278 (0.1244) 0.0203 (0.1277) 0.0100 (0.1171) 0.0055 (0.1183)
Education (low) 0.0394 (0.0389) 0.0403 (0.0399) 0.0321 (0.0367) 0.0328 (0.0371)
Education (high) 0.0533 (0.0349) 0.0570 (0.0359) 0.0484 (0.0330) 0.0513 (0.0333)
Marriage (married) 0.0580 (0.0751) 0.0630 (0.0768) 0.0372 (0.0715) 0.0417 (0.0721)
Marriage (others) 0.0481 (0.0751) 0.0520 (0.0768) 0.0284 (0.0714) 0.0322 (0.0721)
Employment Length (short) 0.0093 (0.0256) 0.0074 (0.0262) 0.0074 (0.0242) 0.0065 (0.0244)
Employment Length (long) −0.0118 (0.0252) −0.0140 (0.0259) −0.0087 (0.0239) −0.0099 (0.0241)
Monthly Income (low) −0.0108 (0.0262) −0.0128 (0.0269) −0.0106 (0.0248) −0.0124 (0.0250)
Monthly Income (high) −0.0158 (0.0273) −0.0164 (0.0280) −0.0136 (0.0256) −0.0139 (0.0259)
Ya × Late Fundraising Stage −0.0491*** (0.0127) −0.0308** (0.0126)
Yr × Late Fundraising Stage −0.0207*** (0.0051) −0.0169*** (0.0051)
Ya × Credit Risky 0.0011 (0.0052) 0.0004 (0.0049)
Yr × Credit Risky −0.0035 (0.0066) −0.0022 (0.0065)
Intercept 1.6428*** (0.3802) 1.4729*** (0.3885) 1.6597*** (0.2563) 1.5379*** (0.2625)
λ̂: Heckman correction term −0.1181*** (0.0457) −0.1193** (0.0469) −0.1298** (0.0426) −0.1312*** (0.0430)
Week fixed effects Yes Yes Yes Yes
Number of observations 64,925 64,925 64,925 64,925
Log-likelihood −115,663.6 −115,663.9 −115,600.7 −115,614.0
AIC 231,435.1 231,443.7 231,483.4 231,550.0
BIC 231,925.5 231,970.3 232,763.8 233,012.1
Note. Standard errors are reported in parentheses.
*p < 0.1; **p < 0.05; ***p < 0.01.

appendix). The results support our main findings that
successors exhibit a stronger herding behavior toward
predecessors presented with an anonymous username
than those with a real-seeming one.
4.2.2. Sentiment in Usernames. A username takes a
textual form that may contain positive, neutral, or
negative meanings, and the sentiment embedded in a
username is likely to affect lenders’ herding behavior.
To capture the overall sentiment in predecessors’
usernames, we calculate the fraction of preceding
peers who are presented with usernames that convey
positive (Frac Positive), neutral (Frac Neutral), and neg-
ative (Frac Negative) sentiment as recognized by re-
search assistants. We then include Frac Positive and
Frac Negative as moderators in (4). Frac Neutral is
dropped because of multicollinearity. The results are
reported in column (1) of Table A5 of the online
appendix. The estimated coefficient of Frac Positive is
positive and significant, suggesting that subsequent
lenders tend to invest a larger amount when a large
proportion of predecessors are presented with user-
names that convey positive meanings. However, we
find no evidence that these sentiment measures signif-
icantly moderate the effect of perceived anonymity
on herding.
4.2.3. Previous Investment Records. We also consider
the potential impact of a lender’s short-term invest-
ment history. We calculate the logarithm of a lender’s
cumulative number of transactions prior to the lend-
er’s current transaction during our observation period
(Lender Records) and include it as a moderator. Col-
umn (2) of Table A5 in the online appendix provides
the results. We find that one’s short-term investment
history is negatively associated with one’s lending
 Jiang et al.: The Effect of Perceived Anonymity on Herding
14 Information Systems Research, Articles in Advance, pp. 1–17, © 2021 INFORMS

Table 6. Results from the Fixed-Effect Models

Variable (1) Lender fixed effects (2) Listing fixed effects (3) Lender and listing fixed effects

Ya: Anonymous Predecessors 0.0288*** (0.0034) 0.1117*** (0.0047) 0.0431*** (0.0036)

Yr: Real-seeming Predecessors 0.0091*** (0.0021) 0.0520*** (0.0041) 0.0204*** (0.0032)
Lag Percent Needed 0.0079*** (0.0004) 0.0222*** (0.0005) 0.0126*** (0.0004)
Lag Num of Transactions 0.0026*** (0.0004) 0.0159*** (0.0006) 0.0066*** (0.0005)
Late Fundraising Stage 0.2055*** (0.0196) 0.3285*** (0.0261) 0.2307*** (0.0208)
Time Interval −0.0082*** (0.0019) −0.0022 (0.0026) −0.0006 (0.0020)
Real Sm Lender (1  yes) 0.0454 (0.7383) 0.4123 (0.6761) 0.0938 (0.9171)
Lender Device 0.1235*** (0.0196) 0.1261*** (0.0122) 0.1416*** (0.0203)
log Amount Requested 0.0255** (0.0121)
Interest Rate 0.0407*** (0.0103)
Loan Duration −0.0047*** (0.0014)
Title Length −0.0011 (0.0012)
Credit Risky −0.0626* (0.0327)
Paid Loans 0.0025 (0.0033)
Overdue Loans 0.00004 (0.0020)
Real Sm Borrower (1  yes) −0.0149 (0.0138)
Age 0.0918 (0.0783)
Education (low) 0.0396 (0.0248)
Education (high) 0.0459** (0.0222)
Marriage (married) 0.1022** (0.0495)
Marriage (others) 0.0876* (0.0495)
Employment Length (short) 0.0042 (0.0166)
Employment Length (long) −0.0041 (0.0163)
Monthly Income (low) −0.0113 (0.0169)
Monthly Income (high) −0.0069 (0.0174)
λ̂: Heckman correction term −0.2825*** (0.0304) −65.7518 (89.3102) 30.5200 (71.6562)
Week fixed effects Yes Yes Yes
Number of observations 64,925 64,925 64,925
Log-likelihood −78,513.3 −112,799.8 −74,935.8
AIC 192,758.6 232,673.5 192,573.6
BIC 354,996.8 264,792.5 386,458.5
Note. Standard errors are reported in parentheses.
*p < 0.1; **p < 0.05; ***p < 0.01.

amount, whereas it does not significantly moderate the
observed effect of perceived anonymity on herding.
5. Conclusion
Online crowdfunding provides researchers an oppor-
tunity to gain new insights into the role of perceived
anonymity in herding. This research pursues such an
inquiry based on the data collected from a leading
debt-based crowdfunding platform. We find that
subsequent lenders demonstrate stronger herding mo-
mentum toward preceding peers who present them-
selves with an anonymous username than those with
a real-seeming one. This finding, which we attribute
to a lower extent of perceived credibility resulting
from a nonconforming behavior, challenges the con-
ventional wisdom that considers anonymity a nega-
tive factor for source credibility. An extended analysis
further reveals that the positive effect of perceived an-
onymity on herding is accentuated in the early fund-
raising stage. Nonetheless, we find no differentiated
herding effect between listings that have high-risk
and low-risk grades.
Our findings complement prior research on herding.
Recent work finds empirical evidence of herding in
various contexts, such as dish choices from diners, kid-
ney refusals from patients, and lending activities from
lenders (Cai et al. 2009, Zhang 2010, Herzenstein et al.
2011). Some research further identifies factors that may
affect the magnitude of herding, such as network ties
(Lee et al. 2015), listing attributes (Zhang and Liu
2012), platform attributes, and government-imposed
regulations (Jiang et al. 2018). We show that subse-
quent lenders may exhibit dissimilar extents of herd-
ing behavior toward predecessors who are presented
with different states of perceived anonymity. Such a
difference is more prominent under high-risk circum-
stances, such as the early stage of the fundraising peri-
od. Besides, our findings also complement the emerg-
ing stream of literature that identifies the positive role
of perceived anonymity in a variety of settings, such as
trustworthiness of online reviews (Hong and Li 2017),
quality of online posts (Lee et al. 2014), and so on.
Our main findings lead to actionable recommenda-
tions for managers of debt-based crowdfunding plat-
forms. For example, the role of the uncovered user
Jiang et al.: The Effect of Perceived Anonymity on Herding
Information Systems Research, Articles in Advance, pp. 1–17, © 2021 INFORMS
heterogeneity with respect to perceived credibility
(manifested in observable personal cues such as user-
names) in herding should be better utilized to facili-
tate lenders’ decision making. To illustrate, we show
that predecessors who are presented with anonymous
usernames in our case are perceived as more credible
and hence induce a stronger herding magnitude from
their successors. Accordingly, to encourage subse-
quent lending activities (in amount), web designers
can consider displaying previous transactions made
by those lenders in a more salient fashion (e.g., listing
them on the top of the transaction log), rather than list
all prior transactions in chronological order as is cur-
rently done. This new design feature is also likely to
be useful in mitigating information overload on the
lenders’ side, particularly when they face listings that
have received a large number of transactions. More-
over, we find that the previous heterogeneity is
contingent on the fundraising stage. During the fund-
raising period, a prospective lender may face varying
degrees of risk, and hence may have different needs
for peer-generated lending information (find signals)
and peer differentiation (find credible signals).
Crowdfunding platforms should be cognizant of such
needs and develop information-disclosure strategies
that provide useful risk-related information to facili-
tate lenders’ decision making. For example, though
the measure of fundraising progress is usually con-
structed as the percentage of the requested amount
that is funded, platforms could develop an alternative
measure by incorporating additional information such
as the number of days left. Meanwhile, in designing
the fundraising progress bar, platforms could use dif-
ferent colors to highlight different fundraising stages
(e.g., a dark color for early and late fundraising
stages). Such a design may enable lenders to quickly
get the fundraising progress information.
Empirical results from this work also have important
implications for borrowers who desire to seek money
from the crowd. For example, we find that lenders are
less prone to participate in listings that offer high inter-
est rates, perhaps because they rationally associate a
high return to a high level of risk. As such, borrowers
who are in urgent need of money should refrain from
overcommitting to interest rates that are much higher
than market prices as this strategic move is likely to
backfire in the end. To encourage participation from the
crowd, borrowers instead should exert effort in provid-
ing more information about their listings, at least in the
length of listing title as our empirical results suggest.
This work has some limitations that can point direc-
tions for future research. As with most empirical stud-
ies of crowdfunding, we have limited information
about individual lenders. For example, in controlling
for lenders’ choice of listings, we construct lenders’
consideration sets based on their transaction records
15
within a fixed time period (e.g., one day). Future re-
search may leverage data on lenders’ browsing histo-
ry to accurately recover their consideration set at the
time of choice, accounting for lenders’ selection of list-
ings in a more precise manner. To mitigate the con-
cern about the reflection problem, we carry out fixed-
effects models as an alternative specification (similar
to Kim and Viswanathan 2019 and Burtch et al. 2016)
and confirm that our results are robust. This strategy
can largely account for the correlated behavior in
crowdfunding settings like ours, where investors re-
lease little personal information and have little room
for direct communication or information sharing (as
discussed by Kim and Viswanathan 2019). When de-
mographic and geographic data are accessible in other
settings, researchers may also account for the potential
correlation among lenders according to other model
specifications (e.g., a spatial autoregressive form).10
Further, in addition to the perceived anonymity that
is present in our setting, perceived expertise may also
play a role in forming the perceptions about source
credibility (Hovland et al. 1953, Ohanian 1990) in oth-
er online environments (e.g., Kim and Viswanathan
2019). Researchers may explore whether and to what
extent observation learning is induced by source cred-
ibility stemming from peer expertise based on de-
tailed information on peers’ knowledge and skills
(self-reported or platform-generated).
To conclude, our work is among the first empirical
studies to investigate the effect of online anonymity
on subsequent lenders’ herding behavior in a debt-
based crowdfunding platform. Our finding regarding
online anonymity challenges the common wisdom
that considers anonymity a negative factor for source
credibility. This finding would seem to be particularly
relevant in similar investment-oriented contexts, such
as social lending. Because individuals face great risks
in such contexts, they may actively engage in sophisti-
cated cognitive processing of peer anonymity. Per-
haps more broadly, however our work highlights the
significance of understanding online anonymity
through the lens of observational learning. Observa-
tional learning has become increasingly important in
individual decision making in online environments
due to improved visibility and traceability. Consider-
ing this trend, our work may inspire further interest
in understanding the role of online anonymity and
uncovering interesting behavioral mechanisms in oth-
er settings where a similar observational learning
mechanism has a strong presence, such as online re-
views, online health, or online education.
Acknowledgments
The authors thank the senior editor, the associate editor,
and the anonymous reviewers for constructive comments
throughout the review process.

16
Endnotes
1 of peer recommendations: The mediating role of social trust.
Our study context observes the same pattern: around 90% of the
lenders in our sample use an anonymous username.
2
It is imperative to note that, from a beholder’s perspective, it is the
motivation of choosing a real-seeming username that triggers the
sense of unease and disapproval. Whether a real-seeming username
trustfully presents one’s real name (or whether a beholder can veri-
fy this association) plays a least important role in this cognitive pro-
cess and is thus not our focus in this study.
3
In early 2017, the platform started masking the usernames of all
previous lenders on a listing page. Under this new policy, only the
first and the last character of a given username are visible, with the
rest being masked by asterisks (e.g., “whgqtu” was masked as
“w****u”).
4
Our data scraping task complied to the website’s data collection
rules (as specified in robots.txt). We thank the associate editor for
pointing this out to demonstrate the legitimacy of our data usage.
5
In addition to listings that were comparable to those available on
major debt-based crowdfunding sites (e.g., LendingClub and Pros-
per), the platform displayed listings that were collected offline and
guaranteed by third-party institutions. In Section 4.1, we assess the
robustness of our results by including these two types of listings.
6
We observed a 96.76% agreement between two research assistants
(Cohen’s Kappa  0.91, perfect agreement). The remaining 3.24%
were further labeled by a third research assistant. A majority rule
was then used to determine the classification result.
7
The platform derives a credit grade for each listing based on bor-
rower demographics (e.g., age), verification documents (e.g., in-
come, education, and assets), and repayment records (e.g., paid,
overdue, and severe overdue repayments).
8 Herzenstein M, Dholakia UM, Andrews RL (2011) Strategic herding
We thank the associate editor for this valuable suggestion. We
also construct alternative consideration sets using different time
windows (three days and seven days) and find robust results (re-
ported in Table A2 of the online appendix).
9
We thank the associate editor for this suggestion.
10
We thank the associate editor for pointing this out.
References
Agrawal A, Catalini C, Goldfarb A (2015) Crowdfunding: Geogra-
phy, social networks, and the timing of investment decisions. J.
Econom. Management Strategy 24(2):253–274.
Banerjee AV (1992) A simple model of herding behavior. Quart. J.
Econom. 107(3):797–817.
Bansal G, Zahedi FM, Gefen D (2016) Do context and personality
matter? Trust and privacy concerns in disclosing private infor-
mation online. Inform. Management 53(1):1–21.
Bikhchandani S, Hirshleifer D, Welch I (1992) A theory of fads, fash-
ion, custom, and cultural change as informational cascades. J.
Political Econom. 100(5):992–1026.
Bikhchandani S, Hirshleifer D, Welch I (1998) Learning from the be-
havior of others: Conformity, fads, and informational cascades.
J. Econom. Perspect. 12(3):151–170.
Burtch G, Ghose A, Wattal S (2015) The hidden cost of accommodat-
ing crowdfunder privacy preferences: A randomized field ex-
periment. Management Sci. 61(5):949–962.
Burtch G, Ghose A, Wattal S (2016) Secret admirers: An empirical
examination of information hiding and contribution dynamics
in online crowdfunding. Inform. Systems Res. 27(3):478–496.
Burtch G, Hong Y, Liu D (2018) The role of provision points in on-
line crowdfunding. J. Management Inform. Systems 35(1):117–144.
Cai H, Chen Y, Fang H (2009) Observational learning: Evidence
from a randomized natural field experiment. Amer. Econom.
Rev. 99(3):864–882.
Jiang et al.: The Effect of Perceived Anonymity on Herding
Information Systems Research, Articles in Advance, pp. 1–17, © 2021 INFORMS
De Vries P, Pruyn A (2007) Source salience and the persuasiveness
de Kort Y, IJsselsteijn W, Midden C, Eggen B, Fogg BJ, eds. In-
ternat. Conf. Persuasive Tech. (Springer, Berlin, Heidelberg), 164–
175.
Forman C, Ghose A, Wiesenfeld B (2008) Examining the relationship
between reviews and sales: The role of reviewer identity disclo-
sure in electronic markets. Inform. Systems Res. 19(3):291–313.
Gelman A, Hill J (2006) Data Analysis Using Regression and Multilevel/
Hierarchical Models (Cambridge University Press, New York).
Gneezy A, Gneezy U, Riener G, Nelson LD (2012) Pay-what-you-
want, identity, and self-signaling in markets. Proc. Natl. Acad.
Sci. USA 109(19):7236–7240.
Harmon RR, Coney KA (1982) The persuasive effects of source
credibility in buy and lease situations. J. Marketing Res. 19(2):
255–260.
Hass RG (1981) Effects of source characteristics on cognitive re-
sponses in persuasion. Petty R, Ostrom TM, Brock TC, eds. Cog-
nitive Responses in Persuasion (Psychology Press, New York),
141–172.
Haugtvedt CP, Wegener DT (1994) Message order effects in persua-
sion: An attitude strength perspective. J. Consumer Res. 21(1):
205–218.
Heckman JJ (1979) Sample selection bias as a specification error.
Econometrica 47(1):153–161.
Helweg-Larsen M, LoMonaco BL (2008) Queuing among U2 fans:
Reactions to social norm violations. J. Appl. Soc. Psych. 38(9):
2378–2393.
Hennig-Thurau T, Gwinner KP, Walsh G, Gremler DD (2004) Elec-
tronic word-of-mouth via consumer-opinion platforms: What
motivates consumers to articulate themselves on the Internet? J.
Interactive Marketing 18(1):38–52.
behavior in peer-to-peer loan auctions. J. Interactive Marketing
25(1):27–36.
Herzenstein M, Andrews RL, Dholakia UM, Lyandres E (2008) The
democratization of personal consumer loans? Determinants of
success in online peer-to-peer lending communities. Working
paper, University of Delaware and Rice University.
Hite DM, Voelker T, Robertson A (2014) Measuring perceived ano-
nymity: The development of a context independent instrument.
J. Methods Measurement Soc. Sci. 5(1):22–39.
Ho Y-C, Wu J, Tan Y (2017) Disconfirmation effect on online rating
behavior: A structural model. Inform. Systems Res. 28(3):626–642.
Hong C, Li C (2017) The effect of “anonymous reviewer”: A study
of anonymity, affect intensity, and message valence in the cy-
berspace. J. Language Soc. Psych. 36(5):504–524.
Hovland CI, Janis IL, Kelley HH (1953) Communication and Persua-
sion: Psychological Studies of Opinion Change (Yale University
Press, New Haven).
Iyer R, Khwaja AI, Luttmer EF, Shue K (2015) Screening peers softly:
Inferring the quality of small borrowers. Management Sci. 62(6):
1554–1577.
Janiszewski C (1998) The influence of display characteristics on
visual exploratory search behavior. J. Consumer Res. 25(3):
290–301.
Jiang Z, Heng CS, Choi BC (2013) Privacy concerns and privacy-
protective behavior in synchronous online social interactions.
Inform. Systems Res. 24(3):579–595.
Jiang Y, Ho Y-CC, Yan X, Tan Y (2018) Investor platform choice:
Herding, platform attributes, and regulations. J. Management In-
form. Systems 35(1):86–116.
Kim J-S, Frees EW (2007) Multilevel modeling with correlated ef-
fects. Psychometrika 72(4):505–533.
Kim K, Viswanathan S (2019) The experts in the crowd: The role of
experienced investors in a crowdfunding market. Management
Inform. Systems Quart. 43(2):347–372.
Jiang et al.: The Effect of Perceived Anonymity on Herding
Information Systems Research, Articles in Advance, pp. 1–17, © 2021 INFORMS
Kruglanski AW, Pierro A, Mannetti L, De Grada E (2006) Groups as
epistemic providers: Need for closure and the unfolding of
group-centrism. Psych. Rev. 113(1):84–100.
Kumar N, Qiu L, Kumar S (2018) Exit, voice, and response on digi-
tal platforms: An empirical investigation of online management
response strategies. Inform. Systems Res. 29(4):849–870.
Kuppuswamy V, Bayus BL (2018) Crowdfunding creative ideas:
The dynamics of project backers. Cumming D, Hornuf L, eds.
The Economics of Crowdfunding: Startups, Portals and Investor Be-
havior (Palgrave Macmillan, Cham, Switzerland), 151–182.
Lee Y-J, Hosanagar K, Tan Y (2015) Do I follow my friends or the
crowd? Information cascades in online movie ratings. Manage-
ment Sci. 61(9):2241–2258.
Lee H, Choi J, Kim KK, Lee AR (2014) Impact of anonymity on in-
formation sharing through internal psychological processes: A
case of South Korean online communities. J. Global Inform. Man-
agement 22(3):57–77.
Lin M, Viswanathan S (2016) Home bias in online investments: An
empirical study of an online crowdfunding market. Management
Sci. 62(5):1393–1414.
Ma M, Agarwal R (2007) Through a glass darkly: Information tech-
nology design, identity verification, and knowledge contribu-
tion in online communities. Inform. Systems Res. 18(1):42–67.
Manski CF (1993) Identification of endogenous social effects: The re-
flection problem. Rev. Econom. Stud. 60(3):531–542.
Marx GT (1999) What’s in a name? Some reflections on the sociolo-
gy of anonymity. Inform. Soc. 15(2):99–112.
Murphy J, Hofacker C, Mizerski R (2017) Primacy and recency effects
on clicking behavior. J. Comput. Mediated Comm. 11(2):522–535.
Nissenbaum H (1999) The meaning of anonymity in an information
age. Inform. Soc. 15(2):141–144.
Ohanian R (1990) Construction and validation of a scale to measure
celebrity endorsers’ perceived expertise, trustworthiness, and
attractiveness. J. Advertising 19(3):39–52.
Peacey MW, Sanders M (2014) Masked heroes: Endogenous ano-
nymity in charitable giving. Working paper, University of Bris-
tol, Bristol, UK.
Petty RE, Cacioppo JT, Schumann D (1983) Central and peripheral
routes to advertising effectiveness: The moderating role of in-
volvement. J. Consumer Res. 10(2):135–146.
Postmes T, Spears R (1998) Deindividuation and antinormative be-
havior: A meta-analysis. Psych. Bull. 123(3):238–259.
Postmes T, Spears R, Sakhel K, De Groot D (2001) Social influence
in computer-mediated communication: The effects of anonymity
on group behavior. Personality Soc. Psych. Bull. 27(10):1243–1254.
Pritchard J (2020) How to make financial transactions as anonymous
as possible. The Balance. Accessed September 17, 2020, https://
www.thebalance.com/how-to-make-payments-and-bank-
anonymously-315816.
Qian H, Scott CR (2007) Anonymity and self-disclosure on weblogs.
J. Comput. Mediated Comm. 12(4):1428–1451.
17
Racherla P, Friske W (2012) Perceived ‘usefulness’ of online con-
sumer reviews: An exploratory investigation across three serv-
ices categories. Electronic Commerce Res. Appl. 11(6):548–559.
Rains SA (2007a) The anonymity effect: The influence of anonymity
on perceptions of sources and information on health websites.
J. Appl. Comm. Res. 35(2):197–214.
Rains SA (2007b) The impact of anonymity on perceptions of source
credibility and influence in computer-mediated group commu-
nication: A test of two competing hypotheses. Comm. Res. 34(1):
100–125.
Rasbash J, Browne WJ (2008) Non-hierarchical multilevel models.
Leeuw Jd, Meijer E, eds. Handbook of Multilevel Analysis (Spring-
er, New York), 301–334.
Silva RR, Topolinski S (2018) My username is IN! The influence of
inward vs. outward wandering usernames on judgments of on-
line seller trustworthiness. Psych. Marketing 35(4):307–319.
Silva RR, Chrobot N, Newman E, Schwarz N, Topolinski S (2017)
Make it short and easy: Username complexity determines trust-
worthiness above and beyond objective reputation. Frontiers
Psych. 8:2200.
Solove DJ (2007) I’ve got nothing to hide and other misunderstand-
ings of privacy. San Diego Law Rev. 44:745.
Sternthal B, Dholakia R, Leavitt C (1978) The persuasive effect of
source credibility: Tests of cognitive response. J. Consumer Res.
4(4):252–260.
Tsikerdekis M (2012) The choice of complete anonymity vs. pseudo-
nymity for aggression online. eMinds Internat. J. Human Com-
put. Interaction 2(8):35–57.
Tucker C, Zhang J (2011) How does popularity information affect
choices? A field experiment. Management Sci. 57(5):828–842.
Van Kleef GA, Wanders F, Stamkou E, Homan AC (2015) The social
dynamics of breaking the rules: Antecedents and consequences
of norm-violating behavior. Current Opinion Psych. 6:25–31.
Wagenknecht T, Teubner T, Weinhardt C (2016) The impact of ano-
nymity on communication persuasiveness in online participa-
tion. Internat. Conf. Inform. Systems (ICIS), https://aisel.aisnet.
org/icis2016/HumanBehavior/Presentations/1/.
Xiao S, Ho Y-C, Che H (2021) Building the momentum: Information
disclosure and herding in online crowdfunding. Production
Oper. Management ePub ahead of print April 3, https://doi.org/
10.1111/poms.13425.
Xu H, Dinev T, Smith HJ, Hart P (2008) Examining the formation of in-
dividual's privacy concerns: Toward an integrative view. Internat.
Conf. Inform. Systems (ICIS), https://aisel.aisnet.org/icis2008/6/.
Yazdani E, Gopinath S, Carson S (2018) Preaching to the choir: The
chasm between top-ranked reviewers, mainstream customers,
and product sales. Marketing Sci. 37(5):838–851.
Zhang J (2010) The sound of silence: Observational learning in the
U.S. kidney market. Marketing Sci. 29(2):315–335.
Zhang J, Liu P (2012) Rational herding in microloan markets. Man-
agement Sci. 58(5):892–912.