ArticlePDF Available

Social Capital Accumulation Through Social Media Networks: Evidence from a Randomized Field Experiment and Individual-Level Panel Data

Authors:

Abstract and Figures

Work-related social media networks (SMNs) like LinkedIn introduce novel networking opportunities and features that promise to help individuals establish, extend, and maintain social capital (SC). Typically, work-related SMNs offer access to advanced networking features only to premium users in order to encourage basic users to become paying members. Yet, little is known about whether access to these advanced networking features has a causal impact on the accumulation of SC. To close this research gap, we conducted a randomized field experiment and recruited 215 freelancers on a freemium, work-related SMN. Out of these recruited participants, more than 70 received a randomly assigned, free, 12-month premium membership voucher. We observe that individuals do not necessarily accumulate more SC from the ability to access advanced networking features, as the treated freelancers did not automatically change their digitized networking engagement. Those features will only unfold their full utility if the individuals are motivated to proactively engage in networking: Freelancers who have access to advanced networking features increase their SC by 4.609% for each unit increase on the strategic networking behavior scale. We confirm this finding in another study utilizing a second, individual-level panel data covering 52,392 freelancers; in tandem, we investigate the dynamics that active vs. passive features play in SC accumulation. Based on the findings, we introduce the “theory of purposeful feature utilization”: essentially, individuals must not only possess an efficacious “networking weapon”—they also need the intent to “shoot” it.
Content may be subject to copyright.
Social Capital Accumulation Through Social Media Networks: Evidence from a Randomized
Field Experiment and Individual-Level Panel Data
Michael Weiler
1
Simon Stolz
2
Andreas Lanz3
Christian Schlereth2
Oliver Hinz1
Forthcoming: Management Information Systems Quarterly
Keywords: Social Capital, Field Experiment, Instrument Variable Approach, Complier Average
Causal Effect, Social Media Network, Agency for Networking
Acknowledgements: This work has been funded by the DFG as part of project B.3 within the RTG
2050 “Privacy and Trust for Mobile Users”. We thank the senior editor, the area editor, and the re-
viewer team for their constructive and helpful feedback. They substantially helped to improve the
paper. Moreover, we would also like to show our gratitude to Michael Kosfeld and Bernd Skiera for
their valuable suggestions and comments on this work as well as Christina Hofmann-Stoelting for
her support. We also highly thank the Head of Business Analytics and the Director Marketing of
“WorkSMN”.
1
Michael Weiler, Oliver Hinz, Goethe University Frankfurt, Chair of Information Systems and
Information Management, Theodor-W.-Adorno Platz 4, 60323 Frankfurt am Main, Germany,
Phone: +49-69-798-34656, me@michaelweiler.de, ohinz@wiwi.uni-frankfurt.de.
2
Simon Stolz, Christian Schlereth, WHU Otto Beisheim School of Management, Chair of Digital
Marketing, Burgplatz 2, 56179 Vallendar, Germany, Phone: +49-261-6509-455,
simon.stolz@whu.edu, christian.schlereth@whu.edu.
3 Andreas Lanz, HEC Paris, Assistant Professor of Marketing, 1 Rue de la Libération, 78350 Jouy-
en-Josas, lanz@hec.fr.
2
Michael Weiler has an academic background in sociology and was a doctoral student at the chair of
Professor Oliver Hinz. During his doctorate, his research focused on the areas of social capital and
social media networks. His research has been published in journals such as Review of Managerial
Science and in the proceedings of the International Conference on Information Systems.
Simon Stolz is a doctoral student at WHU - Otto Beisheim School of Management at the chair of
Professor Christian Schlereth. His research explores user data in online networks via predictive and
econometric methods. His work has been published in the Journal of Interactive Marketing and in
conference proceedings of the European Marketing Academy Conference (EMAC) and AMA Mar-
keting Science. Simon has a background in management consulting in the field of data science and
data strategy.
Andreas Lanz is Assistant Professor of Marketing at HEC Paris. His research projects lie primarily
in the intersection between marketing, economics, and information systems, and are geared towards
helping managers make better-informed decisions utilizing granular data from online platforms. An-
dreas' work has been published in the Journal of Marketing Research (JMR) and Management Infor-
mation Systems Quarterly (MISQ). He is the recipient of the Rigour & Relevance Research Award
2020 from the Swiss Academy of Marketing Science and was the runner-up in the 2018 Doctoral
Dissertation Competition of the European Marketing Academy (EMAC) / Sheth Foundation.
Christian Schlereth is Professor of Digital Marketing at WHU Otto Beisheim School of Manage-
ment. His research primarily addresses questions about the use of digitization, whether its benefits
can be quantitatively determined, and what implications can be derived for practice. His research
has been published in journals like Management Science, Journal of the Academy of Marketing Sci-
ence, International Journal of Research in Marketing, International Journal of Electronic Commerce,
Journal of Interactive Marketing, Business & Information Systems Engineering (BISE) and in a
number of proceedings (e.g. ICIS, ECIS, EMAC).
Oliver Hinz is Professor of Information Systems and Information Management at Goethe Univer-
sity Frankfurt. He is interested in research at the intersection of technology and markets. His re-
search has been published in journals like Information Systems Research, Management Information
Systems Quarterly, Journal of Marketing, Journal of Management Information Systems, Decision
Support Systems, Business & Information Systems Engineering (BISE) and in a number of proceed-
ings (e.g. ICIS, ECIS, PACIS).
3
Social Capital Accumulation Through Social Media Networks: Evidence from a Randomized
Field Experiment and Individual-Level Panel Data
Abstract: Work-related social media networks (SMNs) like LinkedIn introduce novel networking
opportunities and features that promise to help individuals establish, extend, and maintain social
capital (SC). Typically, work-related SMNs offer access to advanced networking features only to
premium users in order to encourage basic users to become paying members. Yet, little is known
about whether access to these advanced networking features has a causal impact on the accumula-
tion of SC. To close this research gap, we conducted a randomized field experiment and recruited
215 freelancers on a freemium, work-related SMN. Out of these recruited participants, more than 70
received a randomly assigned, free, 12-month premium membership voucher. We observe that indi-
viduals do not necessarily accumulate more SC from the ability to access advanced networking fea-
tures, as the treated freelancers did not automatically change their digitized networking engagement.
Those features will only unfold their full utility if the individuals are motivated to proactively en-
gage in networking: Freelancers who have access to advanced networking features increase their SC
by 4.609% for each unit increase on the strategic networking behavior scale. We confirm this find-
ing in another study utilizing a second, individual-level panel data covering 52,392 freelancers; in
tandem, we investigate the dynamics that active vs. passive features play in SC accumulation. Based
on the findings, we introduce the “theory of purposeful feature utilization”: essentially, individuals
must not only possess an efficacious “networking weapon”—they also need the intent to “shoot” it.
Keywords: Social Capital, Field Experiment, Instrument Variable Approach, Complier Average
Causal Effect, Social Media Network, Agency for Networking
4
Social Capital Accumulation Through Social Media Networks: Evidence from a Randomized
Field Experiment and Individual-Level Panel Data
INTRODUCTION
Social capital (SC) is linked to all kinds of social phenomena (e.g., Portes 1998; Weiler and Hinz
2019 for reviews). While SC has received varying interpretations, it essentially captures the idea
that individuals can access tangible and intangible resources through their connections (Adler and
Kwon 2002; Portes 1998). Among the most widely studied relationships is the one between SC and
work-related outcomes. They have been a focal point in early works of social network theory
(Granovetter 1973), with a variety of research finding that SC positively relates to hiring outcomes
(e.g., Gee et al. 2017), salaries, and career paths (e.g., Seibert et al. 2001).
Nowadays, work-related social media networks (SMNs) such as LinkedIn (LI) promise to facili-
tate the accumulation and maintenance of one’s SC online. They intend to support users in manag-
ing their professional network and finding new jobs or other job-related opportunities. Work-related
SMNs typically operate under a so-called freemium business model (Bapna and Umyarov 2015;
Voigt and Hinz 2016), i.e., they offer access to the platform free of charge but require a fee-based
premium membership to unlock advanced networking features. Despite their popularity for the job-
search process, there is a dearth of empirical evidence on their effectiveness (Forret 2018; Garg and
Telang 2018). Specifically, the literature lacks an understanding of how the advanced networking
features available under premium membership increase SC.
Prior research in information systems (IS) points to largely different mechanisms that could ex-
plain how premium membership could support SC accumulation: First, premium badges are com-
mon in work-related SMN. According to social resource theory, prestigious items confer status,
which is closely associated with SC (Lin 1999). Empirical findings in virtual communities highlight
that ownership of such prestigious items increases SC (Hinz et al. 2015). Second, premium users
5
can identify their profile visitors. In dating SMNs, where profile browsing visibility likewise de-
pends on the membership type, identifiable profile visits provide “weak signals” of interest to a fo-
cal user, which significantly drives matching success (Bapna et al. 2016). These two mechanisms
depend on the action of the surrounding network, i.e., they are passive. Third, network scholars
(e.g., Borgatti and Halgin 2011) argue that the individual’s agency is pivotal in SC accumulation.
As findings in a music freemium service show, premium converters disproportionately increase
their activity and SC (Bapna et al. 2018). This explanation, in contrast, emphasizes the activity of
the focal users themselves. In the work-related SMN, advanced networking features enhance both
sides, active and passive, but it is unclear, which of the two can drive the SC accumulation.
With this research, we seek to examine the impact of having access to the advanced networking
features provided by work-related SMNs on SC gains over time (RQ1)––and to identify which spe-
cific types of features (active vs. passive) drive these gains (RQ2). As pointed out in Sundararajan et
al. (2013), despite their inherent relevance for the IS discipline, these questions have received little
scholarly attention so far. This is remarkable, given that work-related SMNs enable easier SC accu-
mulation in comparison to offline contexts. Among other things, SMNs eradicate spatial and tem-
poral boundaries in contact formations, make contact lists explicitly visible (Kane et al. 2014), pro-
vide algorithmic contact suggestions (Liben‐Nowell and Kleinberg 2007), and stimulate contact for-
mation by pointing to similarities between users (Sun and Taylor 2020). Hence, SMNs have “poten-
tially altered the processes by which social networks evolve” (Sundararajan et al. 2013, p. 895) and
given rise to different network formations, depending on their features (Kane et al. 2014).
To date, the empirical findings in the context of work-related SMNs and SC mostly come from
cross-sectional data (see Appendix Section A) only a few exceptions use longitudinal data (e.g.,
Utz and Breuer 2016). These findings largely agree that one’s digital presence on work-related
6
SMNs is advisable (Nikitkov and Sainty 2014), especially due to the platform’s professional infor-
mational benefits (Utz 2016; Utz and Breuer 2016). Moreover, SC maintained within SMNs can
play an important role in deriving job benefits (e.g., Aten et al. 2017; Garg and Telang 2018). These
studies implicitly assume that users exogenously receive SC, which is responsible for the corre-
sponding job-related outcomes. The scarcity of studies that address online SC accumulation in a
causal fashion can be explained through the difficulty of altering SC via experimental stimuli. Prior
studies in IS have found that stimuli that alter the user’s status (Hinz et al. 2015) and reduce social
boundaries (Bapna and Funk 2020) can result in SC increases in specific settings. With our study,
we seek to build a broader understanding of the dynamicsof how individuals accumulate digital-
ized SC in general (i.e., the role of their agency) (Ahuja et al. 2012; Stuart and Sorenson 2007)
3
, and
the role of networking features in particular (Bapna et al. 2016; Kane et al. 2014).
To investigate the causal evidence of SC accumulation, we conducted a randomized field experi-
ment on one of the largest European work-related SMNs. We issued a free, 12-month premium
membership to a special segment of users in the SMN, i.e., freelancers. This segment is of interest
for our study, because they rely on their social network to succeed economically (Van den Born and
Van Witteloostuijn 2013; Wu 2013). We randomly assigned potential freelancers to the treatment
group, which granted them access to advanced networking features, whereas the control group did
not receive such a membership. The premium conversion serves as our treatment variable (e.g.,
Bapna and Umyarov 2015). With this experimental set-up, we can inherently address the endogene-
ity problems that often plague SC research, such as omitted variable bias, measurement error, and
simultaneity bias (c.f., Mouw 2006). Without addressing these issues, the estimated SC effect can be
biased and lead to misinterpretations of the findings.
3
Even in the in the context of real-world SC, only a few studies focused on the role of the individuals’ agency in net-
work theorizing (e.g., Lee 2010; Tröster et al. 2019).
7
To provide insights into which type of features drive SC accumulation (RQ2), we acquired a
second dataset consisting of individual-level panel data that allows us to analyze individual behavior
over time. This panel data covers 52,392 freelancers in discount mailing campaigns. This dataset
contains users’ digital footprint data in two directions: on the active side, it captures the outgoing
activity of users; on the passive side, it captures the incoming activity of the network. It is comple-
mentary to the field experiment and provides insights into the types of features that support SC ac-
cumulation. It also allows us to investigate the dynamic evolution of SC and activities.
Our first study’s findings, derived from the field experiment using the complier average causal
effect (CACE), suggest that freelancers do not automatically change their digitalized networking en-
gagement just because they have access to advanced networking features. Those premium features
can only prove their full value if the freelancers are also motivated to proactively and purposely uti-
lize the given resources as part of strategic networking behavior. To be specific, scoring one point
higher on the strategic networking behavior scale will increase freelancers’ SC by 4.609% when
treated with a premium membership. The individual-level panel data yields converging results:
Every doubling in the number of contact-invites sent before the discount mailing resulted in SC in-
creases of 4.148 additional contacts among freelancers who converted to premium. Passive features
(e.g., prestigious premium badges), which make users more salient, are also positively linked to SC
accumulation. Yet, their impact is substantially lower compared to active features (e.g., personal
messages to non-contacts). Thus, the possession of an efficacious “networking weapon” is not
enough by itself; it must also be accompanied by the intent to “shoot” it. We summarize this finding
through the “theory of purposeful feature utilization”: namely that individuals need to accompany
their access to advanced networking features with the motivation to proactively exploit them.
Our work contributes to the IS literature in multiple ways. Among other things, we improve the
understanding of the formation of digitalized SC, which is an essential part of prior theorizing on
8
SMN (Kane et al. 2014; Karahanna et al. 2018). Further, we provide new insights to the topic of
freemium business models, where previous IS studies have analyzed the interplay of premium mem-
berships and behavioral patterns (Bapna et al. 2018; Oestreicher-Singer and Zalmanson 2013).
Our findings are relevant for work-related SMNs as they provide insights on which advanced
networking features foster SC accumulation. Essentially, it is in the best interest of SMNs to encour-
age networking to gain a competitive advantage (Hinz et al. 2020; Shapiro and Varian 1999). Our
results show that giveaway trials of premium memberships will do little to encourage users to accu-
mulate SC if they have a low agency for networking. Moreover, for the product managers of work-
related SMN, our study points to the role of active features in SC accumulation, which is an im-
portant consideration in managerial decisions concerning pricing, versioning, and feature develop-
ment (Shapiro and Varian 1999). Finally, platform users benefit from our insights: A premium
membership alone is insufficient to increase SC. Only premium users willing to actively approach
others (e.g., through personal messages) will be able to derive additional SC.
The remainder of the paper is structured as follows: The next section reviews the related re-
search on SC and its association with networking. The third section depicts the causal evidence of
SC accumulation in a field experiment (RQ1). Section four complements the experimental insights
into which types of features drive SC accumulation (RQ2): firstly, by providing evidence for the ro-
bustness of these findings; secondly, by developing further insights from the individual-level panel
data. In the final section, we discuss our results and outline potential future work.
THEORETICAL BACKGROUND
Social Capital
Since the concept’s mainstream scientific inception almost 40 years ago, a plethora of different SC
definitions have permeated the disciplines of social science (i.e., sociology, economics, and political
science––see Adler and Kwon (2002) for an overview). These myriad definitions generally align
9
with two camps: individual-level and collective-level SC (Adler and Kwon 2002; Borgatti et al.
1998). In broad terms, representatives of the individual-level perspective explicitly focus on the in-
dividual and his/her relationships with fellow humans. This perspective perceives SC as a private
good of one individual (i.e., a node in a network). Thus, researchers commonly use network analyti-
cal measures, like the node degree, to operationalize it (Borgatti et al. 1998). Maintaining relation-
ships with others provides individuals with exclusive access to all kinds of tangible and intangible
resources, which will ultimately help them attain their specific goals. In the words of Nahapiet and
Ghoshal (1998, p. 243), SC is “[…] the sum of the actual and potential resources embedded within,
available through, and derived from the network of relationships possessed by an individual or so-
cial unit. Social capital thus comprises both the network and the assets that may be mobilized
through that network”. Thus, “[…] social capital is a ‘metaphor about advantage’” (Takac et al.
2011, p. 189) that enables individuals to be more successful than others.
Scholars who look at the collective-level perspective of the concept (e.g., Putnam 2000) focus
on larger social structures, such as groups and communities. They look at the members’ ties, which
exist among each other within these collectives, and assess whether these groups eventually achieve
cohesion (Adler and Kwon 2002). This perspective emphasizes that every individual in the collec-
tive will benefit from the created SC, regardless of whether he/she put effort into the creation pro-
cess (Borgatti et al. 1998). Thus, according to this perspective, “[…] social capital is a collective
good and […] it is non-exclusive in consumption […]” (Rostila 2011, p. 311).
In this study, we follow Nahapiet and Ghoshal’s (1998) understanding of individual-level SC.
We do so because it is the individual user who eventually benefits from his/her resources on the
SMN platform and thus reaps assets, such as information about potential working opportunities.
10
Networking
Networking refers to the process of cultivating and establishing SC (Forret and Dougherty 2001).
The term “networking” is generally understood to mean a “[…] strategic (i.e., rationally motived)
behavioral effort that involves the dyadic exchange of interpersonal resources, which are directed
toward building and maintaining network relationships with specific network contacts and moti-
vated by whether they have access to specific interpersonal resources” (Porter and Woo 2015, p.
1481). For instance, individuals who engage in such behavioral efforts of networking may achieve a
higher salary (e.g., Wolff and Moser 2009) and more job offers (e.g., Van Hoye et al. 2009).
Although individuals are aware of the importance of networking, they often do not, or only re-
luctantly, participate in such behavior. This self-effacement may be due to difficulties in socializing
or a belief that users are insincere by establishing ties solely for instrumental purposes (Casciaro et
al. 2014; Kuwabara et al. 2018). By contrast, others enjoy creating instrumental ties and actively
aim to establishing them (Bensaou et al. 2014). To put it differently, the motivation to proactively
adjust one’s career trajectory largely seems to depend on an individual’s beliefs and attitudes toward
such instrumentally oriented behaviors (Kuwabara et al. 2018). Individuals who perceive network-
ing as a skill that can be learnedand thus changed through continuous effortwill embrace such
beneficial behavior. In contrast, individuals who deeply believe that it is a more or less fixed skill
and thus part of their stable individual traitswill rather refrain from participating in it (Kuwabara
et al. 2018). Next to attitudes of networking behavior, experimental evidence indicates that specific
external interventions can promote and encourage networking. For instance, Bapna and Funk (2020)
highlight that female IT conference participants who receive a nonreciprocal list of contact recom-
mendations significantly extended their digitalized SC on LI. This observation raises the question of
whether online stimuli can likewise positively affect SC.
11
The process of networking has greatly changed due to the introduction of SMNs. The online set-
ting eliminates frictions in an offline setting: SMNs provide explicit contact lists (Kane et al. 2014)
and support networking through a range of features (Karahanna et al. 2018). Individuals not only
receive recommendations on potential contacts directly from SMNs, but they can also leverage the
IS-inherent search functions to identify potential contacts (Kane et al. 2014). Therefore, SMNs mir-
ror a huge repository of diverse social contacts, which the user can tap into and access other individ-
uals who may otherwise be (spatially) out of reach. SMN users can also browse through their per-
sonal network members’ profiles, allowing them to inform themselves about their contacts’ friends,
whom they can easily befriend by simply sending a contact request.
Moreover, SMNs commonly provide users with contact suggestions that are algorithmically de-
rived based on network structures (Li et al. 2017; Liben‐Nowell and Kleinberg 2007) and common-
alities between users (Sun and Taylor 2020). Facilitating network formation among users is a pri-
mary strategic interest to SMN platform providers. A better-connected network reflects a lock-in of
users because of network effects (Shapiro and Varian 1999). In other words, users are unwilling to
undertake the effort to recreate their network on a competing SMN platform. As contact lists evolve
over several years of usage, SMNs form an ensemble of actively maintained and inactive (dormant)
ties, both of which can be valuable in terms of SC (McCarthy and Levin 2019).
This rich set of features enables users of SMNs to not only maintain their offline contacts, but
also reach out to users who they do not know well offline or even at all. Manago et al. (2012, p. 6)
found that on Facebook, networks grow “primarily through relatively more distant kinds of relation-
ships”. In work-related SMNs, such weak connections are particularly commonplace (Garg and
Telang 2018). Freelancersthe target group of our studyespecially rely heavily on SC to acquire
project-based labor arrangements and achieve job security through a diversified set of contacts (Van
den Born and Van Witteloostuijn 2013; Wu 2013). As in the offline world, looking at each other
12
and initiating a conversation are typically the first steps to forming a connection. The equivalents to
such offline behaviors in an SMN context are profile visiting and messaging (Bapna et al. 2016).
In summary, while the intrinsic agency for networking differs across users, SMNs strongly sup-
port SC accumulation through a wide range of features, eliminating frictions in offline contexts
(Kane et al. 2014; Karahanna et al. 2018). We expect in work-related SMNs, and for freelancers in
particular, that this desire for SC accumulation through more distant connections is strong so that
advanced networking features will be particularly helpful for accumulating SC.
ACCUMULATING SOCIAL CAPITAL: CAUSAL EVIDENCE (RQ1)
Experimental Set-Up
Our experiment aims to causally test whether freelancers with access to advanced networking fea-
tures accumulate more SC than those without (RQ1). It builds on the idea that such features in a
work-related SMN can bolster freelancers’ networking opportunities, and thus their ability to in-
crease and nurture their SC. This argument broadly resembles the ideas associated with the af-
fordances concept (e.g., Leonardi 2013), which considers SMN features as the foundation for influ-
encing the users’ action possibilities (Bucher and Helmond 2017). Thus, users approach the same
SMN features with different motivations and needs, and apply them in diverse ways: “[…] [U]sers
will appropriate certain features of a technology only when they perceive that those features offer
them affordances for action, but if those features are not appropriated, their material qualities cannot
afford social structural change” (Leonardi 2013, p. 752).
We set up an experiment in one of the largest European work-related SMNs, which features ap-
proximately 17 million users; we refer to the platform by the pseudonym WorkSMN. Like its Amer-
ican counterpart, LI, WorkSMN enables its users to build and maintain their professional social net-
work to find new jobs or projects––either free of charge or by signing up for a premium member-
13
ship. Examples of such advanced features of the premium membership are functionalities that ena-
ble users to see which other users visited their profile, enhanced search capabilities, or the possibil-
ity to send messages to non-contacts (c.f., Appendix Section B). Essentially, users’ networking ca-
pabilities are either more constrained or enabled based on their type of membership.
Our study focuses on freelancers, a population that is only marginally represented in the current
research landscape (Lo Presti et al. 2018) because they are quite difficult to access (Kuhn 2016).
Nonetheless, freelancers epitomize a promising research population for our endeavor due to their
heavy reliance on networking and the resulting SC to gain projects and succeed in their careers (e.g.,
Van den Born and Van Witteloostuijn 2013). Qualitative studies highlight that independent musi-
cians value the networking opportunities that SMNs provide them (Haynes and Marshall 2018) and
that freelancers see an active SMN profile as imperative to their success (Gandini 2016). So far,
only relatively few studies have used quantitative data to examine how freelancers engage in a digi-
tal environment. Those few pay more attention to how freelancers behave in online labor markets
(e.g., Leung 2014; Shevchuk and Strebkov 2018) or enterprise SMNs (Wu 2013) rather than in
work-related SMNs. Hence, there is a need for deeper insights into this research population.
As we could not rely on a pre-defined list of freelancers to sample this hard-to-reach study popu-
lation (Kuhn 2016), we recruited participants for our experiment through a pre-study by directly
contacting them on WorkSMN and asking them to confirm our contact request. The contact request
also included the study's access link, intended solely for recruitment purposes. We used the pre-
mium features of WorkSMN to search for users who self-reported themselves as freelancers in their
digital user profile and asked them to fill out a questionnaire about “The perceived usefulness of
work-related SMNs in generating job offers”. The end of the questionnaire contained two questions
to qualify for the experiment: (1) whether they are willing to use a donated premium membership
for six months and (2) whether they are willing to support further studies from the researchers. If the
14
participants responded affirmatively to both questions, we redirected them to another separate ques-
tionnairea way to ensure their anonymityand asked them to enter their name, e-mail address,
and WorkSMN profile ID.
In total, it took us roughly ten months (January to October 2017), and approximately 6,350 man-
ually sent contact invites to collect a sample of 243 eligible freelancer responses. Out of these 243
responses, we excluded 26 cases because they either purchased a premium membership on their
own or quit working as a freelancer. Among the remaining 217 freelancers, we used random number
seeds to select 75 participants who received a voucher code to upgrade their basic account to a pre-
mium membership. Henceforth, we refer to this group as the treatment group. The platform service
provider graciously donated the vouchers. The control group consisted of the remaining 142 non-
gifted (i.e., non-treated) freelancers.
We did not want the treatment and control group members to perceive that they were participat-
ing in an experiment. Thus, we communicated to the treatment group that they won a free, 12-month
premium membership as a thank-you gift for participating in our pre-study. Our message included
the premium voucher code, the redemption instructions, and a notice about the code’s automatic ex-
piration outside of a special period to discourage them from postponing their redemption. After two
weeks, treated freelancers who did not redeem their vouchers received a personal reminder note via
their WorkSMN profile and a final reminder after another two weeks. This effort was worthwhile,
as 65 gifted freelancers redeemed their voucher.
Data
Before the experiment, we manually harvested digital footprint data by visiting each freelancer’s
profile page. Specifically, we browsed through each recruited participant’s profile page, collected
the publically visible variables of interest (mentioned below), and subsequently stored the corre-
sponding information in our database. For instance, in line with Burbano (2016), we looked at the
15
freelancer’s profile picture and assessed his/her “gender” based on the corresponding profile name
and displayed portrait. However, in a handful of cases, some information, such as the number of
contacts (i.e., SC), was not directly visible to us due to the specific profile settings. In these in-
stances, we utilized WorkSMN’s API, with their permission, to request and retrieve the relevant
data. Aside from “gender”, the data we collected included “freelancers’ tenure on WorkSMN”, pro-
file details, such as freelancers’ self-reported “number of haves” and “wants”, as well as the “num-
ber of subscribed groups”. While the “have” section allows the freelancers to promote their skills,
products, and services, the “wants” section enables them to specify what they are looking for, e.g.,
finding new project partners. We also harvested the “number of direct contacts” shortly before the
experiment and after the experiment had run for about six months. Across many disciplines, includ-
ing IS (e.g., Hinz et al. 2015), this metric is a well-established way to operationalize SC. In short,
well-connected individuals maintain more heterogeneous networks and have access to a wider set of
diverse resources (Borgatti et al. 1998). We calculated the average percentage change in the number
of direct contacts over six months as our measure “SocialCapitalAccumulation”.
While harvesting the node degree from the freelancers’ profiles for the second time, we had to
exclude two participants from our analysis: one assigned to the treatment (did not redeem the
voucher) and the other to the control condition. We faced technical issues in the data collection for
one of them, and the other deleted the SMN account. After their exclusion, we had a final sample
size of 215 cases (see Appendix Section C.1 for a flow chart of the experiment).
Table 1 summarizes the covariate balance check to test for distributional differences between the
treatment and control groups. Due to unequal sample sizes, we report the Welch’s t-test. Treatment
(n = 74) and control (n = 141) group do not statistically differ before the experiment for any of the
pre-treatment variables (p > 0.05). We conclude that both groups are comparable and well-balanced.
16
Table 1. Covariate balance check before manipulation
Covariate
Treatment
(assigned)
Mean (SD) or %
Difference
p-Value
Gender (Female = 1;
Male = 2)
0 (n = 141)
56.738%
-6.776%
0.337
1 (n = 74)
63.514%
Social Capital (Number
of direct contacts)
0
153.348 (153.882)
-16.693
0.506
1
170.041 (183.990)
Number of groups
0
6.482 (7.858)
0.360
0.717
1
6.122 (6.372)
SMN tenure (in years)
0
7.028 (3.969)
-0.310
0.578
1
7.338 (3.815)
Number of ‘haves’
0
11.879 (10.430)
-1.229
0.394
1
13.108 (9.776)
Number of ‘wants’
0
4.333 (5.618)
0.211
0.754
1
4.122 (4.151)
Notes: Test statistic reflect Welch’s t-test for continuous variables and Pearson’s Chi-squared test for discrete variables
About six months after the manipulation, we sent the participants a link to our follow-up ques-
tionnaire. We implemented two mechanisms to build trust and galvanize them to participate (e.g.,
Laurie and Lynn 2009): First, we sent all freelancers who qualified for the experiment an e-mail
containing a summary of the pre-study research findings. Second, a couple of days later, we sent
them a pre-paid 5-Euro Amazon voucher together with a notification that we would greatly appreci-
ate their participation in our follow-up questionnaire.
The follow-up questionnaire’s survey link contained a unique identifier featuring a randomly
generated six-digit token (e.g., “Xv5P32”), such that we can match participants’ digital footprint
data with their self-reported survey data. We collected several covariates: socio-demographic varia-
bles like “age” (M = 44.94 years; SD = 10.98) and “level of education” (high: 83.08%; low: 0%;
middle: 16.92%); freelancers’ personality trait “openness” (M = 5.49; SD = 1.13, Cronbachs α =
0.65) using a seven-point Likert scale on the validated version of the Big Five Inventory (BFI)
(Hahn et al. 2012); freelancers’ “SMN usage intensity” (several times a month: 34.62%; less than
once a month: 6.15%; once a month: 17.69%; once a week: 11.54%; several times a week: 14.62%;
once a day: 10.77%; and several times a day: 4.62%), and the “number of weekly hours” working as
17
a freelancer (M = 38.33; SD = 16.29) (Forret and Dougherty 2001). Further, we surveyed the free-
lancers about the “field” they work in (e.g., Shevchuk and Strebkov 2018) using a pre-chosen list of
response categories (e.g., graphic design (15.38%) or journalists/PR (15.38%)). Moreover, we asked
whether they had “other earnings” apart from their freelance work (Van den Born and Van
Witteloostuijn 2013), which was not the case for the majority (66.92%). We also asked freelancers
to assess their “strategic networking behavior” by adapting the four-item scale by Utz and Breuer
(2016). An example item reads: “I send contact requests to a great number of people, in order to get
a large network.” The participants rated their strategic networking behavior on a five-point Likert
scale ranging from “strongly disagree” (1) to “totally agree” (5). The Cronbach’s alpha measure
suggests adequate internal consistency of the scale (α = 0.66), which we constructed by averaging
the four items (M = 2.35; SD = 0.91). In the Appendix, Section C.4, we detail the definitions as well
as the used measures of our key variables and provide summary statistics.
Challenges and Estimation
The first challenge we encountered is sample attrition, i.e., that not every freelancer reacted to the
follow-up questionnaire. To alleviate this issue, we used two different channels (i.e., e-mail and the
messaging function of WorkSMN) to motivate our freelancers to participate, with up to three re-
minders. We offered freelancers who were still unresponsive a small, delayed cash incentive (5
Euro). While several freelancers responded to the third reminder, only five participants requested
their promised cash incentive. Eventually, 65.12% (i.e., 140 of our 215 participants) took part in our
follow-up questionnaire six months after the experiment. Specifically, 130 of these respondents re-
ported that they are still working as a freelancer, while ten stated they quit pursuing freelance work
and that they transitioned into traditional forms of employment. Thus, in the upcoming analysis, we
focus on those who were still active as freelancers.
18
To rule out a potential violation of the random assignment assumption, we repeated the covari-
ate balance check for the 130 respondents. We find no statistically significant differences between
the two groups (p > 0.05) (see Appendix Section C.4). The results suggest that attrition is not related
to any pre-treatment covariate, and the initial internal validity (see Table 1) remains unchanged
4
.
We encountered a second challenge: two-sided noncompliance, which emerged because we
could not guarantee that all of the gifted participants redeemed their vouchers and because some
freelancers in the control group voluntarily bought a premium membership on their own. Two-sided
noncompliance is a common threat to experiments that utilize some sort of encouragement design
(Gerber and Green 2012). For instance, this issue often appears in clinical trials, where some pa-
tients are encouraged to take a medical treatment while others are not. Eventually, the decision
about whether to redeem the randomly assigned voucher or not (i.e., whether to comply) relates to
observed or unobserved individual characteristics, which is troublesome because it may result in the
perilous problem of self-selection (Angrist and Pischke 2008).
Thankfully, some approaches enable researchers to handle two-sided noncompliance, which
“[…] allow the recovery of a causally interpretable estimate of the treatment effect, even though
they alter the interpretation and the generalizability of the experimental results” (Lonati et al. 2018,
p. 26). These are the intention-to-treat (ITT) analysis, the as-treated (AT) analysis, the per-protocol
(PP) analysis, and the complier average causal effect (CACE, also referred to as local average treat-
ment effect or LATE) (Imbens and Rubin 2015; Sagarin et al. 2014). Among these approaches,
4
We also estimated a logistic regression to check whether the freelancers who took part in the post-treatment follow-up
questionnaire differed from their non-responding counterparts, using the binary variable “Participation in Follow-up
Questionnaire” (i.e., questionnaire non-participant = 0 versus participant = 1) as our dependent variable. The results
show that we can mitigate potential concerns of a self-selection bias because none of the pre-treatment variables or in-
teraction terms (i.e., assigned treatment condition x pre-treatment variables) included in the logistic regression signifi-
cantly predicted the probability of filling out our post-treatment follow-up questionnaire.
19
CACE is the only one that recovers an unbiased causal estimate of treatment received on the out-
come (Imbens and Rubin 2015). As such, we focus on the CACE for the analysis, as it captures the
average treatment effect for a subset of participantsnamely the so-called “compliers”. Therefore,
we use an instrumental variable (IV) approach (c.f., Appendix Section C.2), which is an established
approach for deriving causal estimates (Angrist and Pischke 2008). We define freelancers as com-
pliers when a) they redeem a premium membership voucher after being assigned to receive it (treat-
ment group), and b) they remain untreated after being assigned to the control group, i.e., they do not
voluntarily buy the premium membership (Gerber and Green 2012; Imbens and Rubin 2015).
As stressed by several scholars (e.g., Bollen 2012), randomly generated variables inherently pos-
sess the ideal conditions to pass as a valid instrument. Thanks to our encouragement design, we al-
ready have such a promising IV. Namely, our randomly formed binary variable “assignment to the
premium membership subscription voucher (Treatment assigned).” Using this established setup, we
follow recent articles from the IS domain, which also utilized the CACE approach with the random
assignment of the treatment as an IV (Sun et al. 2019a; Sun et al. 2019b). Specifically, our choice of
this IV gives us the perfect conditions to tackle the non-compliance issue, as we are subsequently
able to uncover the causal effect for those freelancers who complied with their initial treatment as-
signment (Angrist and Pischke 2008). As a result of utilizing our IV, the identified CACE is inde-
pendent of unobserved confounding factors (e.g., believes about the utility of premium membership)
that give rise to self-selection. To put this into perspective: By only looking at the compliers, and
thus remedying self-selected movement between treatment conditions, we ensure that the treatment
and control groups remain comparable. Thus, the CACE estimate is “[…] undiluted by non-compli-
ance and unaffected by selection bias” (Angrist 2006, p. 35).
20
A valid instrument should meet two requirements: (1) instrument relevance, i.e., it is highly cor-
related with the independent endogenous variable of receiving the treatment, and (2) instrument ex-
ogeneity, i.e, it is uncorrelated with the unobservable error term (Bollen 2012; Gerber and Green
2012). While we can assess instrument relevance with statistical means such as the first-stage F-sta-
tistic, no such statistical tests exist for instrument exogeneity. Instead, they must provide theoretical
reasons why a specific variable fulfills this assumption (Hinz et al. 2016).
We will report and comment on the corresponding statistical tests for the instrument relevance
requirement in the upcoming subsection; however, we can already conclude that our IV fulfills the
instrument exogeneity requirement. Specifically, the fact that our IV resulted from our randomiza-
tion procedure is an inherent assurance that it “is independent of other preexisting causes of Yi,
which makes Zi a potentially valid instrument” (Sovey and Green 2011, p. 189).
Results
Model-Free Evidence
We examine whether freelancers who have access to an SMN’s advanced networking features accu-
mulate more SC: In our experiment, freelancers receiving the treatment accumulate on average
11.41% (SD = 16.72) more WorkSMN contacts. In comparison, their counterparts in the control
group gain an average of 9.41% (SD = 15.37) more WorkSMN contacts. In other words, the model-
free analysis provides the first empirical evidence that treated freelancers, with their access to ad-
vanced networking features, do not accumulate significantly more SC (t(104.79) = -0.690, p =
0.492). The absence of a significant effect indicates that the access to advanced networking features
does not automatically lead treated freelancers to accumulate more SC.
CACE Estimation and IV Selection
We next derive the CACE estimate, while including other covariates to increase its precision
(Gerber and Green 2012; Imbens and Rubin 2015).
5
We summarize our two-stage least squares
5
We display the correlation among these variables in Appendix Section C.3.
21
(2SLS) IV model in Equations (1) and (2). Equation (1) displays the first stage, where we regress
“Treatment received” on our IV “Treatment assigned” and the aforementioned covariates.
(1)        
       
         
Equation (2) describes the second stage. We regress the freelancers’ level of SC increase against
the predicted values of “Treatment received” from Equation (1) and the set of same covariates.
(2)        
       
         
Heterogeneity within the experimental groups may introduce too much variance to arrive at sig-
nificant main effects. In some situations, accounting for some of this heterogeneity through an inter-
action variable can reveal the underlying effects. As individuals differ in their aspirations to partici-
pate in networking behaviors, there could be a mechanism at play that obfuscates the relationship
between our treatment and dependent variable (“SocialCapitalAccumulation”). Specifically, it is
not enough that the premium membership offers more visibility on WorkSMN; the freelancers must
also proactively use these new networking opportunities to accumulate SC. In other words, they
need to have the corresponding attitudes, values, and beliefs necessary to motivate their pursuit of
instrumental ties (Ahuja et al. 2012; Bensaou et al. 2014; Kuwabara et al. 2018; Porter and Woo
2015). For instance, Pollack et al. (2015) demonstrate that entrepreneurs who are motivated by
growth and advancement cultivate more relevant contacts for their business during the week. Simi-
larly, entrepreneurs shape their network by actively reaching out to potential ties (Vissa 2012),
which highlights the role of individual agency in enacting SC accumulation (Ahuja et al. 2012).
We use “strategic networking behavior” as an agency variable, which captures whether freelanc-
ers are rationally motivated to proactively and purposely befriend other users who might, in some
way, be or become relevant for their future career (Utz 2016). Thus, the key variable of interest in
22
our model is the interaction term between “treatment received” and “strategic networking behavior”.
Individuals who are strategic networkers “contrive to get into advantageous network positions”
(Vissa 2011, p. 154) to promote their economic success (Ahuja et al. 2012; Stuart and Sorenson
2007). However, with the inclusion of the interaction term into our model, we need a second IV to
identify this interaction (Ebbes et al. 2016). We follow the lead of others (e.g., Hoisl and Mariani
2017) by using the interaction term between our first instrument (i.e., treatment assigned) and strate-
gic network behavior to construct the required second instrument. Table 2 shows the second-stage
estimates of our IV regression analysis (for the first-stage, see Appendix Section C.5). While model
E1 pertains to the CACE estimation without the outlined interaction term, model E2 adds this key
variable into the model. In the upcoming analysis, we will focus on the interpretation of model E2.
Table 2. Complier average causal effect estimation of the effect of having access to an IS with advanced
networking features on SC accumulation
Model E1.
2nd stage estimates: “So-
cial capital accumulation”
Model E2.
2nd stage estimates: “Social capital ac-
cumulation” (including interaction term)
Treatment received
0.300 (2.472)
-10.867 (6.224)
Strategic networking behavior
3.306 (1.247)
**
1.020 (1.698)
Treatment received * Strategic net-
working behavior
4.609 (2.338)
*
Male (Ref.: female)
-4.104 (2.397)
-3.874 (2.380)
High level of education (Ref.: middle)
-4.723 (3.053)
-3.804 (3.063)
Age (in years)
-0.219 (0.112)
-0.233 (0.112)
*
Openness
-1.150 (1.022)
-1.432(1.023)
SMN tenure
-1.195 (0.287)
***
-1.243 (0.285)
***
Work hours per week
0.031 (0.067)
0.038 (0.066)
Other income (Ref.: no other income)
0.807 (2.242)
0.317 (2.237)
Usage frequency (once a month)(Ref.:
less often than once a month)
-1.496 (4.771)
-2.427 (4.756)
Usage frequency (several times a month)
1.012(4.639)
1.044 (4.601)
Usage frequency (once a week)
0.591 (5.029)
0.550 (4.988)
Usage frequency (several times a week)
-1.082 (5.139)
-0.623 (5.102)
Usage frequency (once a day)
2.498 (5.185)
2.075 (5.146)
Usage frequency (several times a day)
9.438 (6.201)
9.035 (6.153)
F
3.452
3.435
R2
0.351
0.362
RMSE
10.338
10.252
Underidentification test (Anderson
canon. corr. LM statistic)
108.070
***
104.394
***
Weak identification test (Cragg-Donald
Wald rk F-statistic)
473.076
193.656
n
130
130
Notes: 2SLS instrumental variable regression providing the complier average causal effect (CACE). Standard errors in
23
parentheses; unstandardized regression estimates. Constant not reported. The paper of Stock and Yogo (2005) lists the
critical values for the Cragg-Donald Wald F statistic. The Stock-Yogo weak ID test critical value for a 10% maximal IV
size is 16.38 for model E1 and 7.03 for model E2, respectively. In our model specification, we partialled out the variable
freelance field from all the other variables in our estimation using the Frisch-Waugh-Lovell theorem. Significance lev-
els: * p < 0.05, ** p < 0.01, *** p < 0.001
We test whether our instruments are working correctly and provide unbiased estimates (Column
2 in Table 2). Through the suitability tests, we can verify that both of our IVs are not weak instru-
ments: To assess the relevance of both IVs (i.e., treatment assigned and the instrumental interaction
term are each correlated with the endogenous regressor), we run an under-identification test using
the Anderson canonical correlation LM statistic. We reject the null hypothesis that the model is
overall under-identified, at p < 0.001. To evaluate whether we have weak instruments, we use the
Cragg-Donald Wald F-statistic measure (193.66), which by far exceeds the critical values proposed
by Stock and Yogo (2005) for a 5% maximal IV estimator as well as 10% maximal size bias. In
sum, we are confident that our IVs are sufficiently strong and valid instruments.
6
Interaction between Strategic Networking Behavior and Treatment Received
The result of the second-stage regression of model E2 indicates that the treatment has no direct sig-
nificant causal effect on accumulating SC. Instead, we find a significant positive interaction effect
between our treatment and strategic networking behavior (B = 4.609; p < 0.05). This means that the
freelancers’ level of strategic networking behavior moderates this relationship. Complying freelanc-
ers, who score higher on the strategic networking behavior scale can significantly increase their SC.
To put it another way, if a freelancer exhibits a strong motivation for strategic networking behavior
and has access to an IS with suitable networking features, he/she will substantially increase his/her
SC relative to that of his/her counterparts in the control group. Scoring higher on the strategic net-
working behavior scale by one point will increase a freelancer’s SC by 4.609% relative to that of
his/her non-treated equivalents. In Appendix Section C.6, we also display the ITT parameters,
6
In Appendix Section C.5, we report the Sanderson and Windmeijer (2016) first-stage conditional under-identification
and weak identification tests.
24
which closely mirror those in the CACE estimation (see Table 2) and suggest that the identified ef-
fect persists: Being assigned to the treatment and scoring high on our strategic network behavior
variable positively relates to the SC a freelancer accumulates.
In Table 3, we depict the identified interaction effect in more detail using a two-by-two matrix.
We classify freelancers with a value smaller than the median (i.e., 2.25) of our variable strategic net-
working behavior into the “low” group and freelancers with a score equal or greater than the median
into the “high” group. On average, treated freelancers in the latter group accumulate 13.24% more
SC, while their counterparts in the former group achieve, on average, only 8.93% more SC. Like-
wise, freelancers accumulate more SC on average if they have access to an IS that equips them with
suitable networking features.
Table 3. Descriptive mean comparisons of strategic networking behavior (median split) and treatment
received groupings
Treatment received
No (0)
Yes (1)
Strategic networking behavior
Low
8.33%
8.93%
High
10.39%
13.24%
For gaining SC, the advanced networking features that freelancers’ need to utilize actively are of
particular importance (see Appendix Section B). In Appendix Section C.7, we test whether the
agency variable “strategic networking behavior” potentially suffers from post-treatment bias be-
cause we elicitated this information in the follow-up questionnaire. We provide theoretical argu-
ments why this is not the case and empirically rule out group differences between the treatment and
control group. Ultimately, we conclude that a post-treatment bias is quite unlikely.
To sum up, the field experiment results suggest a causal link between having access to advanced
networking features and SC accumulation, but this is conditional on the users’ agency for network-
ing. In the next section, we explore which types of features drive this accumulation.
25
FEATURE UTILIZATION: SUPPORTIVE EVIDENCE (RQ2)
The field experiment suggests that the mere increase in status portrayed by a premium membership
badge, an increase in visibility or the sheer availability of advanced active features is not sufficient
to attract SC by itself. Otherwise, we would have observed a significant main effect for all treated
users. Instead, only users with a high agency for networking were able to significantly increase their
SC. This begs the question: How did these users achieve this positive outcome? That is, we want to
investigate which types of advanced networking features drive SC accumulation (RQ2). This ques-
tion is of paramount interest as work-related SMNs typically advertise their premium membership
with claims along the lines of expand your network,” while also promising to “turn profile views
into new opportunities.
7
Hence, they are unclear about the mechanism and what the contribution of
the platform’s advanced premium features may be. We resolve this ambiguity, providing insights to
users who want to benefit most from the premium membership. Yet, also for platform providers, we
provide insights into the magnitude of the value that their premium features offer, which is an im-
portant consideration in designing their products.
To investigate RQ2, we follow the needs-affordances-features framework, which states that af-
fordances and features are closely related: features enable affordances, i.e., “action possibilities”
(Karahanna et al. 2018). Hence, we now focus on how users make use of these affordances follow-
ing a premium conversion.
Framework
We distinguish two types of advanced networking features: Active features operate via activity, i.e.,
targeted actions of the premium user, whereas passive features operate via saliency, i.e., incoming
attention that the premium user receives. In Appendix Section B, we list these features in detail. The
7
https://premium.linkedin.com/, Retrieved on 01/21/2021
26
arguably most important active features are the following: Premium users have access to an exclu-
sive advanced search filter. While basic users can only search by name, premium users can search
through WorkSMN’s user directory by industry, firm, and role, or even a combination of criteria.
For example, by these means, premium users can identify profiles of targets in the buying center of
a specific company––without knowing their names––and get in contact with them. This advanced
search filter is especially important to freelancers because they rely heavily on acquiring project-
based labor arrangements (Wu 2013). Besides, only premium users can send personal messages to
such non-contacts and thereby reach out to them. One can think of profile visiting and messaging as
steps on a funnel of SC accumulation: users first identify potential targets and then actively “make a
move” (Bapna et al. 2016) on targets outside of their network.
Passive features take effect via the premium user’s saliency. This incoming attention results
from better and more highlighted positioning in search results and a prestigious premium badge that
accompanies the user’s profile. Among others, this premium badge also highlights the profile in
contact suggestions, groups, and contact lists. For this reason, passive features unfold their effec-
tiveness irrespective of the user’s activity and generate attention in the form of traffic onto the pre-
mium users’ profiles––beyond the reciprocal traffic as a result of activity. The traffic due to saliency
may, for example, originate from companies looking for project-based labor arrangements. Along
these lines, all premium users are likely to experience a lasting increase in SC accumulation due to
passive features. Moreover, the passive features enable premium users to evaluate the traffic onto
their profiles and access user statistics on the visitors. Previous research in dating SMN noted that
the ability to follow up on profile visits is a pivotal feature in achieving matches (Bapna et al. 2016).
Likewise, we assume that this passive feature may improve SC accumulation.
Given that advanced networking features primarily support targeted actions, we operationalize
profile visits and messages sent as activity, and profile visitors and messages received as saliency.
27
As illustrated in Figure 1, activity relates to active features, while saliency relates to passive fea-
tures. The seeding literature on SMNs suggests that saliency is a result of activity (Chen et al. 2017;
Lanz et al. 2019); however, it could also be the other way around, i.e., that a user’s saliency triggers
activity. Correspondingly, we expect a reciprocate association between activity and saliency (see
arrow 1). The seeding literature also finds that activity is effective in accumulating SC (Hinz et al.
2011; Lanz et al. 2019). Hence, we expect that profile visiting and personal messaging drive SC ac-
cumulation (arrow 2). Finally, regarding the association between saliency and SC accumulation (see
arrow 3), research shows that prestigious items, like a premium badge, can result in more contacts
on SMNs (Hinz et al. 2015).
Figure 1. Hypothesized framework
Data
WorkSMN kindly shared data with us on three discount mailing campaigns conducted in May 2015,
May 2016, and November 2016. A combined set of 52,392 freelancers from a larger sample re-
ceived a 50% discount coupon on their first 12-month subscription. Of them, 626 freelancers con-
verted to premium. Even though it is impossible to warrant unbiased causal effects from such obser-
vational data, it allows us further to investigate the process of SC accumulation under premium
membership and accounts for the dynamic interplay of activity and saliency with SC. The restricted
discount mailing campaign allows us to carry out our analysis in a uniform environment; in most
cases involving SMNs, the discounts and communication tend to be highly personalized and vary
over time. This approach also boosted the number of conversions, which might otherwise be too low
28
due to the rare-event nature of premium membership conversions (Bapna et al. 2018; Oestreicher-
Singer and Zalmanson 2013). In line with the experimental setting in the first study, the campaigns
were broadly targeting active freelancers in terms of total logins ever (M = 949.80, SD = 1,476.26).
They resulted in a heterogeneous sample with a wide range of user characteristics (e.g., from 18 to
70+ years old, various career levels, academic degrees, and diverse backgrounds like engineers, de-
signers, and alternative practitioners).
The dataset contains digital footprint data, which we summarize in Table 4. Tracking user activ-
ity on the platform is multifaceted; thus, WorkSMN only stores a selected set of variables that it
considers the most relevant over a longer period. These include the monthly individual-level panel
data on users’ number of profile visits and messages sent, as well as their number of contacts four
months before and six months after the mailing campaigns. While some scholars have studied simi-
lar activity information (e.g., Bapna et al. 2018; Oestreicher-Singer and Zalmanson 2013), only a
few have succeeded in capturing the incoming activities (for an exception in the context of dating
platforms, see Bapna et al. 2016). In our case, we have access to the monthly numbers of profile vis-
its and messages received from all focal users. In addition to the individual-level panel data,
WorkSMN granted us access to a granular set of snapshot data collected before each mailing cam-
paign to select users. This snapshot data contains the number of contact invites sent before the mail-
ing campaign––so that we can use it to remedy post-treatment concerns of the experiment with re-
spect to strategic networking behavior. To overcome the high skewness of the digital footprint vari-
ables, and to implement the notion that excessive networking behavior is rewarded with diminishing
returns to scale, we took natural logarithms of “+1”-transformed values for our independent varia-
bles (Stock and Watson 2015, pp. 269-276).
As reflected in Table 4, non-converters appear similar to converters with regard to SMN tenure,
age, gender, education, and career level. Nevertheless, converting freelancers sent out more contact
29
invites, had more logins one month before conversion, and were more job-seeking. We find strong
behavioral changes (Δ) for premium converters across all captured dimensions of SC, activity, and
saliency. Also, a correlation analysis (see Appendix D.2) reveals a general association between ac-
tivity and saliency measures.
30
Table 4. Covariate summary
Covariate
Premium conversion
Mean (SD) or %
Difference
Test Statistics
p-Value
Measure
SC accumulation (Δ)
0 (n = 51,766)
11.719 (25.205)
-18.535
6.933
<0.001
Monthly
1 (n = 626)
30.254 (66.854)
Profile visits
(Δ Log)
0
0.003 (0.998)
-0.806
16.212
<0.001
Monthly
1
0.809 (1.242)
Messages sent
(Δ Log)
0
0.038 (0.659)
-0.463
13.529
<0.001
Monthly
1
0.501 (0.854)
Profile visitors
(Δ Log)
0
-0.079 (0.751)
-0.364
10.925
<0.001
Monthly
1
0.285 (0.830)
Messages received (Δ
Log)
0
0.062 (0.700)
-0.325
9.808
<0.001
Monthly
1
0.387 (0.826)
Contact invites sent
before (Log)
0
2.842 (1.471)
-0.331
5.766
<0.001
Before
campaign
1
3.173 (1.427)
Logins before (Log)
0
1.746 (1.201)
-0.329
6.679
<0.001
Before
campaign
1
2.075 (1.223)
SMN tenure (in years)
0
8.052 (2.439)
0.113
1.051
0.294
Before
campaign
1
7.939 (2.686)
Age
0
43.040 (9.684)
0.664
1.697
0.090
Before
campaign
1
42.376 (9.726)
Gender (Female = 1;
Male = 2)
0
74.971%
-0.709%
0.130
0.719
Before
campaign
1
75.680%
Education title (Yes =
1, No = 0)
0
50.995%
-2.926%
2.000
0.157
Before
campaign
1
53.920%
Level: Entry & expe-
rienced
0
58.024%
2.984%
2.938
0.230
Before
campaign
1
55.040%
Level: Management
0
13.335%
0.055%
1
13.280%
Level: Senior
0
28.641%
-3.039%
1
31.680%
Job-seeking (active)
0
36.840%
-11.480%
34.436
<0.001
Before
campaign
1
48.320%
Notes: Test statistic reflect Welch’s t-test for continuous variables and Pearson’s Chi-squared test for discrete variables
Challenges and Estimation
One challenge of observational data is that self-selection is inherent. Unlike in the field experiment,
in our observational data freelancers had to pay for the premium membership by themselves. Thus,
if a variable is associated with the self-selection effect and the dependent variable at the same time,
it will skew the estimated effect size (i.e., a confounder). We acknowledge that there are countless
hypothetical confounding factors. In our setting, the latent job situation of a freelancer is the most
obvious candidate. A freelancer who lost an important client is more likely to convert to premium
and increase networking activity. We include variables to capture this confounder in our analyses.
31
Later on, we provide an extensive set of robustness tests using propensity score matching methods,
enabled by the rich dataset.
Hence, even though we cannot categorically rule out all potential confounding factors, the ob-
servational data provides us with a valuable complementary perspective: It allows us to assess the
previous findings' convergent validity while also resolving some of the previously mentioned limita-
tions of the experiment. The overall idea is that if the results of the two approaches converge, we
can gain more confidence in the findings. One example is that it was not possible to rule out the
post-treatment bias in the experiment completely. With the individual-level panel data, we can be
sure it does not exist, because we utilize in our analysis the number of contact requests sent before
the mailing campaign.
Another challenge is the assumed directionality of activity and saliency on the SC accumulation
(i.e., arrows 2 and 3 in Figure 1). Studies that seek to quantify content contribution provide evidence
for an audience effect in some contexts (e.g., blogging networks in Shriver et al. 2013): A bigger
audience can incentivize activity, which is why researchers commonly refer to this reverse associa-
tion as an audience effect. Other experimental findings suggest that this effect is rather nuanced
(Toubia and Stephen 2013). Because our study primarily focuses on directed networking actions, we
expect this effect to be insubstantial. Nevertheless, we will explicitly test for reverse associations.
We continue as follows: First, we replicate the analysis of RQ1 using the observational data.
While we acknowledge that this additional analysis builds only on observational data, it contains an
objective and pre-treatment variable for strategic networking behavior, alleviating concerns of the
prior analysis. Second, we turn our attention to the outgoing and incoming activities (i.e., activity
and saliency) after the premium conversion. We use model-free evidence to investigate the develop-
ment of activity, saliency, and SC before and after the premium conversion. Third, we seek empiri-
cal support on the directionality of activity and saliency on SC accumulation. For this purpose, we
32
follow the examples of Dewan and Ramaprasad (2014) and Chen et al. (2015) and estimate a panel
vector autoregression model. Finally, we conduct a difference-in-differences analysis to assess the
influence of behavioral changes over time, before alleviating self-selection concerns through pro-
pensity score matching.
Results
Accumulating Social Capital: Robustness Check (RQ1)
Consistent with the experiment, we examine SC before the discount mailing and six months after.
Since the campaigns target a large and heterogeneous user base of freelancers, the number of con-
tacts (M = 210.12, SD = 336.87) is sometimes low, even up to no contacts at all. For this reason, we
measure changes in SC accumulation as the absolute number of direct contacts.
8
To objectively
measure strategic networking behavior, we use the individual number of contact invites sent before
the mailing campaign (M = 47.56, SD = 194.69). Replacing the self-reported agency variable “stra-
tegic networking behavior” of the 2SLS-IV regression presented in Equation 2 with the operational-
ization “contact invites sent before”, we derive the OLS regression estimates presented in Table 5.
While our approach cannot guarantee the absence of additional unobserved confounders, we seek to
capture the confounding factors where possible. Hence, we control for self-reported job-seeking ac-
tivity and the number of logins before (the mailing). To provide conservative estimates, we report
heteroscedasticity-robust standard errors.
As reported in Table 5, freelancers who exhibit a 1% higher number of contact invites sent be-
fore converting are associated with an additional increase in SC accumulation of 0.059 contacts (=
log(1.01)·5.985; p < 0.01) if they convert. In other words, as a premium user, every doubling of the
number of contact invites sent is associated with an increase in SC accumulation of approximately
8
We show in Appendix D.1 that our findings are robust for the percentage change, when following Bapna and Umyarov
(2015) and removing the freelancers with the lowest 15-percentile in SC prior to the mailings.
33
4.148 contacts (log(2) · 5.985). Table 5 suggests a significant main effect of increased SC accumu-
lation (p < 0.001) for “premium conversion”, which was not significant in the field experiment. We
assume that this difference is due to the non-experimental setting where freelancers pay for pre-
mium membership (Bapna et al. 2018).
Table 5. OLS regression of the effect of having access to an IS with advanced networking features on
SC accumulation
Model P1:
„ΔSC accumulation”
Model P2: „ΔSC accumulation”
(including interaction with contact
invites sent before)
Premium conversion
16.027 (2.604)
***
-2.918 (4.489)
Contact invites sent before (Log)
4.351 (0.162)
***
4.284 (0.163)
***
Premium conversion *
Contact invites sent before (Log)
5.985 (1.993)
**
Male (Ref.: female)
1.009 (0.219)
***
0.991 (0.220)
***
Level of education (Ref.: no acad. title)
1.006 (0.224)
***
1.010 (0.224)
***
Age
-0.068 (0.015)
***
-0.068 (0.015)
***
SMN tenure
-0.134 (0.055)
*
-0.134 (0.055)
*
Career segment: Manager (Ref.: Entry)
1.795 (0.351)
***
1.809 (0.351)
***
Career segment: Senior executive
0.231 (0.282)
0.222 (0.280)
Job-seeking: active (Ref: inactive)
0.820 (0.238)
***
0.808 (0.237)
***
Logins before (Log)
2.683 (0.142)
***
2.684 (0.142)
***
F
560.364
523.712
R2
0.114
0.115
RMSE
24.643
24.626
n
52,392
52,392
Notes: Constant and fixed effects of the three mailing campaigns are not reported. Robust heteroscedasticity-consistent
standard errors in parentheses. Significance levels: * p < 0.05, ** p < 0.01, *** p < 0.001
Upon the inclusion of the interaction term, the premium conversion's main effect is no longer
significant, suggesting that freelancers with an agency for networking are accumulating SC when
they have access to advanced networking features. These results support the field experiment’s main
conclusion.
Model-Free Evidence
We now visually explore which specific types of features (active or passive) drive these gains
(RQ2). Figure 2 contrasts the percentage changes in SC of premium users vs. non-premium users
and other digital footprint variables. For ease of interpretation, we index all variables to four months
before the mailing campaign and grey out the mailing campaign month because we cannot decom-
pose the numbers into the ones before and after upgrading to premium.
34
Figure 2. Monthly co-evolution of average SC accumulation, activity, and saliency before and after the
premium conversion
The indexed SC of premium subscribers is similar to the non-subscribers before the mailing
campaign. After the campaign, we observe a peak for premium users that declines quickly yet stays
on a significantly higher index level (p < 0.001) compared to non-premium users. Concerning activ-
ity (i.e., profile visits and messages sent), its trajectory coincides with SC accumulation in magni-
tude and duration. The peak is particularly pronounced for profile visits (404% versus 253% for
messages sent). A possible reason is that, in contrast to message sending, profile visiting is a lower-
involvement activity. When looking at the graphs on the right-hand side, saliency does not peak at
the same magnitude but followed a similar pattern (172% vs. 205%).
The peak in the first month after premium conversion and the quick decline suggest that SC ac-
cumulation has a stronger relationship with activity than saliency. Bapna et al. (2018) observe a
similar peak in the context of music freemium services. They attribute this change to users’ desire to
extract value from the service they now pay for. Regarding saliency as an explanation of SC gains,
we would have expected a more durable increase in SC after the premium conversion; after all, the
features that help a user stand out do not change in the course of the premium membership. Yet, be-
cause saliency also peaks in the first month, the graphs suggest that activity may be the underlying
35
driver of the saliency increase. Next, we turn our attention to a difference-in-differences model that
allows us to control for user characteristics. However, as discussed above, before we can do this, we
first need to rule out a reverse association between activity and SC accumulation as well as saliency
and SC accumulation.
Panel Vector Autoregression
We follow the examples of studies in IS (Chen et al. 2015; Dewan and Ramaprasad 2014) and use
panel vector autoregression (PVAR) to model the dynamic interdependencies between our main var-
iables via Generalized Method of Moments (GMM). GMM uses transformed observations as instru-
ments for the lagged dependent variables (Hansen 1982). As detailed in Appendix Section D.4, we
first downsample non-converters due to the computational intensity of PVAR calculations via
GMM. We test the log-transformed individual values for stationarity (Harris and Tzavalis 1999) and
find an optimal lag length of three months (Andrews and Lu 2001). The resulting estimates of a sys-
tem PVAR model, which controls for time fixed effects, are presented in Table 6.
Both saliency and activity are positively related to SC accumulation. In Table 6, all lagged ob-
servations of activity and saliency are associated with each other and SC accumulation. Converting
to premium is positively associated with all digital footprint variables (all p < 0.001). However, the
reverse associationnamely that a higher number of contacts might lead to an increase in saliency
and a higher activity level (e.g., Shriver et al. 2013)was not supported. SC accumulation has no
explanatory value for consecutive activity nor saliency; thus, we can rule out the presence of a re-
verse effect. Moreover, the results show that activity and saliency are interrelated, but they do not
allow us to conclude that one precedes the other.
36
Table 6. Panel vector autoregression model with three lags
Profile visits
Profile visitors rec.
Messages sent
Messages rec.
Δ SC acc.
Profile visitst-1
0.209 (0.015)***
0.059 (0.009)***
0.041 (0.011)***
0.046 (0.010)***
0.048 (0.012)***
Profile visitors rec.t-1
0.126 (0.019)***
0.207 (0.013)***
0.057 (0.013)***
0.052 (0.014)***
0.089 (0.016)***
Messages sentt-1
-0.015 (0.02)
0.015 (0.013)
0.117 (0.016)***
0.039 (0.016)*
-0.005 (0.019)
Messages rec.t-1
0.021 (0.019)
-0.019 (0.012)
0.045 (0.015)**
0.111 (0.014)***
0.016 (0.018)
Δ SC acc. t-1
0.020 (0.015)
0.000 (0.010)
0.009 (0.011)
0.016 (0.011)
0.029 (0.015)
Profile visitst-2
0.043 (0.014)**
-0.01 (0.008)
0.003 (0.009)
0.008 (0.009)
-0.023 (0.012)*
Profile visitors rec.t-2
0.073 (0.018)***
0.091 (0.013)***
0.016 (0.013)
0.016 (0.013)
0.063 (0.016)***
Messages sentt-2
-0.021 (0.019)*
0.011 (0.012)
0.037 (0.015)*
-0.022 (0.014)
-0.018 (0.018)
Messages rec.t-2
-0.024 (0.018)***
-0.019 (0.011)
-0.004 (0.014)
0.052 (0.013)***
-0.015 (0.017)
Δ SC acc. t-2
0.023 (0.013)
0.013 (0.009)
0.013 (0.010)
0.010 (0.010)
0.035 (0.015)*
Profile visitst-3
0.029 (0.012)*
0.011 (0.008)
0.000 (0.008)
0.011 (0.009)
-0.009 (0.01)
Profile visitors rec.t -3
0.080 (0.016)***
0.135 (0.011)***
0.012 (0.011)
0.000 (0.011)
0.012 (0.013)
Messages sentt-3
-0.022 (0.018)
-0.015 (0.011)
0.016 (0.014)
-0.025 (0.014)
-0.017 (0.016)
Messages rec.t-3
-0.022 (0.017)*
-0.019 (0.011)
0.001 (0.013)
0.030 (0.013)*
-0.005 (0.017)
Δ SC acc. t-3
-0.001 (0.012)
-0.014 (0.008)
-0.005 (0.010)
0.008 (0.009)
0.022 (0.013)
Premium con.t
0.476 (0.034)***
0.229 (0.023)***
0.287 (0.026)***
0.189 (0.026)***
0.264 (0.027)***
Notes: Constant and seasonality estimates not reported; Significance levels: * p < 0.05, ** p < 0.01, *** p < 0.001
Active Versus Passive Features
To explore which specific features drive SC accumulation, we extend the set of control variables in
the difference-in-differences based models (P1 and P2 in Table 5) of the robustness check as fol-
lows: For each freelancer, we model the change in the activity variables (i.e., profile visits and mes-
sages sent) and saliency variables (i.e., profile visitors and messages received) by differencing the
log-transformed value before with the average six months after the campaign. Moreover, for model
P2, we substitute the interaction term between “premium conversion” and “contact invites sent be-
fore” with two new terms: namely, “premium conversion” and the increases in activity and saliency,
resulting in equation (3):
(3)      
        
Next to the direct effect estimates of activity βA and saliency βS, the interaction estimates βIA (in-
teraction activity) and βIS (interaction saliency) reflect the ability of active and passive advanced
37
networking features to accumulate SC. If access to advanced networking features plays no role in
SC accumulation, we expect these interaction terms to turn out insignificant.
When testing this model for concerns of multicollinearity, we find that all variance inflation fac-
tors are in a moderate range (smaller 3), suggesting that we can distinguish activity from saliency.
Because the PVAR results do not indicate whether saliency follows activity or vice versa, we per-
form the subsequent analysis in a stepwise manner: We report the results with either activity or sali-
ency alone and then include their interaction term with premium conversion.
As shown in Table 7, model P3 supports that changes in activityi.e., profile visits (B = 3.777;
p < 0.001) and messages sent (B = 3.830; p < 0.001)are associated with SC accumulation. In
model P4, we find similar effects for the saliency variablesi.e., profile visitors received (B =
3.706; p < 0.001) and messages received (B = 3.898; p < 0.001)albeit with lower R².
When including the interaction terms of activity (model P5), the significant main effect of the
premium conversion disappears (B = -0.003; p > 0.05). At the same time, both interaction terms
namely between premium conversion and profile visits (B = 5.669; p < 0.05) as well as messages
sent (B = 13.500; p < 0.05)turn out to be positive and significant. The interaction terms of the sa-
liency variables (profile visitors received and messages received) with premium conversion in
model P6 and P7 yield no significant parameters (p > 0.05) and only partially explain the premium
conversion main effect in model P6 (B = 6.671; p > 0.01).
38
Table 7. OLS regression of the effect of changes in activity, saliency, and their interactions with pre-
mium conversion on SC accumulation
Δ SC accumulation explained by…
Model P3:
Δ Activity
Model P4:
Δ Saliency
Model P5:
Δ Activity and
interactions
Model P6:
Δ Saliency and
interactions
Model P7:
Δ Activity, Δ Sali-
ency, and interac-
tions
Premium conversion
11.133 (2.455)***
13.496 (2.519)***
-0.003 (2.738)
6.671 (2.212)**
-0.074 (2.767)
Δ Profile visits (Log)
3.777 (0.168)***
3.666 (0.160)***
3.255 (0.150)***
Δ Messages sent (Log)
3.830 (0.277)***
3.553(0.257)***
2.108 (0.362)***
Δ Profile visitors (Log)
3.706 (0.241)***
3.584 (0.211)***
2.161 (0.186)***
Δ Messages rec. (Log)
3.898 (0.322)***
3.693 (0.296)***
1.458 (0.391)***
Premium conversion *
Δ Profile visits (Log)
5.669 (2.740)*
4.239 (2.060)*
Premium conversion *
Δ Message sent (Log)
13.500 (5.412)*
12.695 (5.886)*
Premium conversion *
Δ Profile visitors rec. (Log)
8.317 (6.925)
3.814 (6.012)
Premium conversion *
Δ Messages rec. (Log)
11.817 (8.040)
0.956 (9.058)
F
674.598***
606.619***
610.314***
545.072***
503.201***
R2
0.153
0.140
0.157
0.143
0.161
GVIF < 3
Yes
Yes
Yes
Yes
Yes
RMSE
24.095
24.283
24.034
24.238
23.977
n
52,392
Notes: We include all control variables as in model P1 (Table 5). Robust heteroscedasticity-consistent standard errors in
parentheses. Significance levels: * p < 0.05, ** p < 0.01, *** p < 0.001
Overall, increases in activity and saliency are associated with increases in SC. Moreover, look-
ing at the estimated interaction coefficients suggests that, in the setting of work-related SMN, activ-
ity does not uniformly drive SC for basic and premium users. Instead, we observe a significant inter-
action between premium conversion and changes in those activities that are supported by active fea-
tures. The interaction effects of saliency with premium conversion are not significant, indicating
that passive features are less effective than active features for premium users. In other words, while
we observe in the model-free evidence (Figure 2) that saliency and activity both increase following
the premium conversion, the OLS regression shows that advanced networking features only enable
the SC accumulation through activity.
These patterns complement our conclusions from the field experiment, which revealed that only
freelancers who convert to premium and exhibit a strategic networking behavior accumulate more
39
SC. The individual-level panel data further suggests that among the advanced networking features,
particularly active features support the accumulation of SC. The passive features, such as prestig-
ious premium badges and the ability to follow up on profile visitors, provide limited value in terms
of SC accumulation.
Propensity Score Matching
A limitation of the previous analysis is that self-selection among premium converters may exist. Ta-
ble 4 shows that premium converters and non-converters differ substantially, so that they may not be
directly comparable. We follow previous IS research (e.g., Bapna et al. 2018; Hinz et al. 2015;
Oestreicher-Singer and Zalmanson 2013) and address this potential limitation by conducting pro-
pensity score matching (PSM). We can test for the robustness of the results by selecting only those
non-converting users who closely resemble premium converters.
The objective of this approach can be thought of as “finding an artificial twin” that closely re-
sembles a premium converter (Rosenbaum 2020). We first have to form propensity scores based on
pre-treatment variables. Following model P1, we include all basic variables such as age, gender, ed-
ucation, and self-indicated job-seeking activity. However, the strength of our observational data is
that it contains a wide range of variables beyond these basic ones. As a large scale comparison of
experimentation and matching shows, granular behavioral matching can yield results indistinguisha-
ble from experimental results (Eckles and Bakshy 2020). We build upon this idea and try to account
for a potential self-selection process: In addition to the basic variables of model P1, we further in-
clude pre-treatment behavioral variables, namely in terms of activity (i.e., profile visits and mes-
sages sent) and saliency (i.e., profile visits received and messages received).
Behavioral variables are especially promising because they can uncover latent processes, such as
the users’ current job situation (Ebbes and Netzer 2018). For example, we find in Figure 2 that pro-
file visiting already increases before a premium conversion, which supports the assumption. In line
40
with this phenomenon, profile visits before treatment have a significant (p < 0.01) positive effect on
the decision to convert to premium, as estimated via the underlying propensity model (reported in
Appendix D.5).
The propensity scores are the basis for matching, where the most common form is 1:1 matching.
Since 626 freelancers converted to premium, the corresponding “artificial twins” also amount to 626
among the abundant control observations (i.e., 1.209% from n = 51,766). To reduce sampling varia-
bility, one can further apply an elbow criterion (Rosenbaum 2020), which allows us to adapt the
matching ratio to 1:4 and increase the selection to 1,841 most identical observations (i.e., 3.556%
from n = 51,766). Based on these two matching criteria and the two sets of variables, Table 8 shows
the modeling results of four PSM specifications. For direct comparison, we also include the un-
matched sample (model P7 from Table 7).
Table 8. Propensity score matching overview
Propensity Score Matching
Unmatched
(comparison)
Matching variables
Controls
Controls,
Activity before,
Saliency before
Controls
Controls,
Activity before,
Saliency before
(unmatched)
Matching ratio
1:1
1:1
1:4
1:4
-
Matching caliper
-
-
0.002
0.002
-
Treatment obs. (dropped)
626 (0)
626 (0)
623 (3)
617 (9)
626
(Matched) controls
(626)
(626)
(1,841)
(1,841)
51,766
Interaction Coefficients of Model P7
Premium conversion *
Δ Profile visits (Log)
5.468
(2.632)*
3.286
(2.424)
4.053
(2.343)+
4.292
(2.321)+
4.239
(2.060)*
Premium conversion *
Δ Message sent (Log)
13.073
(5.994)*
12.822
(5.732)*
14.886
(6.180)*
13.345
(5.822)*
12.695
(5.886)*
Premium conversion *
Δ Profile visitors (Log)
1.189
(6.097)
-0.982
(9.150)
3.541
(5.951)
3.792
(5.747)
3.814
(6.012)
Premium conversion *
Δ Messages received (Log)
2.286
(9.221)
2.429
(5.732)
0.752
(9.368)
0.292
(9.111)
0.956
(9.058)
Notes: Robust heteroscedasticity-consistent standard errors in parentheses. Significance levels: + p < 0.1, * p < 0.05, **
p < 0.01, *** p < 0.001. Matching caliper: to be interpreted in standard deviations of treatment propensities, where
0.002 reflects an inflection point to dropping treatment observations.
Table 8 shows that for all four PSM specifications, the interaction of message sending and pre-
mium conversion remains positive and significant. As with the unmatched sample, no significant
41
interaction with saliency occurs. The interaction of profile visiting and premium conversion is sig-
nificant for the most common form of PSM––and at a higher level (p < 0.1) when increasing the ra-
tio to 1:4 (potentially caused by the smaller dataset). These results further underpin that the ad-
vanced networking features, especially messaging, are valuable for SC accumulation.
DISCUSSION
Prior research focuses on the benefits of the network itself rather than the proactive and purposive
actions that structure the network, resulting in numerous calls to take the latter issue more seriously
(Ahuja et al. 2012; Bensaou et al. 2014; Borgatti and Halgin 2011; Stuart and Sorenson 2007). Our
research aim is to assess whether individuals’ access to advanced networking features on work-re-
lated SMNs helps them accumulate more SC and which type of features support this accumulation.
To this end, we conducted a field experiment with freelancers on a large European, work-related
SMN and complement the insights with individual-level panel data. Freelancers are an interesting
population because they rely on their network to succeed economically (e.g., Van den Born and Van
Witteloostuijn 2013).
In our first study (RQ1), we focus on the CACE estimate via an IV approach. We do this to ad-
dress the two-sided noncompliance that arose from our randomized experiment. This estimate repre-
sents the mean causal effect for the subpopulation of the compliers (Gerber and Green 2012; Imbens
and Rubin 2015). Thus, in terms of those compliers, we find a significant interaction effect between
our treatment (i.e., receiving a premium membership) and the self-reported agency variable “strate-
gic networking behavior”. We are the first to highlight this interaction effect, which suggests that
those who proactively and purposefully shape their social network enjoy a larger positive effect on
their SC accumulation.
We also test our findings’ robustness: In the individual-level panel dataset, we find converging
results of a significant interaction effect between premium conversion and SC accumulation when
42
using the objectively measured number of contact invites sent before a mailing campaign. We fur-
ther utilize this dataset for our second study (RQ2). We find a positive relationship between SC ac-
cumulation and the observed changes in user behavior concerning profile visiting and message
sending. Along these lines, a freelancer joining a work-related SMN cannot rely on passive features
alone to accumulate SC. Prestigious premium badges, common in freemium settings (e.g., Bapna et
al. 2018), may encourage other users to obtain premium and impact saliency. Yet, for SC accumula-
tion, these badges play a minor role.
Theoretical Implications
Our study makes several contributions to prevalent theories in the IS literature. First, our study ad-
vances the stream of experimental research on SC accumulation in SMN. Specifically, we offer a
better understanding of which and how experimental stimuli drive the formation of online SC. Our
multilayered stimuli, the provision of advanced networking features, does not automatically increase
SC among individuals but is conditional on the heterogeneity in the recipient’s agency for network-
ing. This is interesting to researchers facing the tricky objective to manipulate SC in SMN experi-
mentally. For example, artificial manipulations of connections via fake profiles (Toubia and Stephen
2013) would not properly reflect the SC accumulation process. Hence, to assess the effect of SC on
further outcomes, such as job search (Bapna and Funk 2020), job security (Wu 2013), or user be-
havior (Wasko and Faraj 2005), SC needs to be organically manipulated. As we highlight, such or-
ganic manipulations in the form of advanced networking features in work-related SMN are only
successful in light of users’ agency for networking. This insight complements previous findings in
IS that seek to identify the mechanisms of SC accumulation experimentally. Such studies show that
nonreciprocal online stimuli (i.e., e-mails with contact suggestions), known only to the recipient, in-
43
crease SC (Bapna and Funk 2020). Meanwhile, in non-work-related contexts, experimental dona-
tions of prestigious items that are visible beyond the recipient can increase SC by making a user
more salient (Hinz et al. 2015).
Second, our insights about individual motivation and agency advance the insights of prior IS
studies, which assessed measures to enable network change in the wake of additional digital af-
fordances (Bapna et al. 2018; Leonardi 2013; Wu 2013). Our experiment agrees with these studies
that the provision of online networking capabilities “was, by itself, incapable of bringing […] any
network change” to them (Leonardi 2013, p. 772) and that gifting access to premium features does
not automatically change user activity (Bapna et al. 2018). Beyond this finding, we show that the
premium membership’s advanced networking features will only unfold their effectiveness if free-
lancers are motivated to proactively and purposely utilize the resource and shape their surrounding
social network. The users’ agency (i.e., their networking behavior) particularly determines the bene-
fit of having access to advanced networking features. This finding extends a burgeoning stream of
research on the individual agency for networking (e.g., Tröster et al. 2019; Vissa 2012) and con-
firms theorizing that individual networking agency is the key driver of SC accumulation (Borgatti
and Halgin 2011). Thus, the affordances (i.e., additional action possibilities for networking) pro-
vided by work-related SMNs do not automatically prompt an action; instead, they just invite the
freelancer to potentially seek out valuable ties. Building upon our findings, we can weave together
the SMN affordances lens and the role of the individual’s agency. Specifically, we introduce the
theory of purposeful feature utilization” to capture our study’s key findings, as outlined above.
Third, scholars have hypothesized that individual features of an SMN define the way users net-
work and build SC online (Kane et al. 2014; Karahanna et al. 2018; Sundararajan et al. 2013). We
contribute a theoretical framework that describes how advanced networking features drive SC accu-
44
mulation by enabling two pivotal forces: activity and saliency. Using this framework, we further in-
vestigated those mechanisms among self-selected individuals who pay for the premium membership
and generally exhibit a high baseline agency for networking. As our field experiment demonstrates
in a work-related SMN, a prestigious appearance in the form of a premium badge is, by itself, insuf-
ficient to accumulate significantly more SC. Our individual-level panel data further uncovered the
mechanisms of increased SC accumulation among premium users: They increase their SC by send-
ing more messages and outgoing profile visits. The ability to follow up on “weak signals” of visible
profile visits, a crucial mechanism in dating SMNs (Bapna et al. 2016). For work-related SMNs, it
appears to have no measurable effect: increasing profile visits did not lead to a higher SC accumula-
tion if the focal user had access to a premium membership. Overall, we observe that premium con-
verters generally become more active, which is positively associated with SC.
In sum, our findings contradict a deterministic view of SC accumulation in SMNs: Digitalized
networks do not emerge exogenously, as implicitly assumed by cross-sectional studies (see Appen-
dix Section A). Researchers must be cautious when interpreting the effect of an individual’s social
network position on a certain outcome if they do not take the endogenous nature of the networks
into account. For example, Lee (2010) strikingly demonstrates that the positive relationship between
individuals’ network positions and their performance vanishes once he controlled for their past per-
formance. We concur with those findings that one cannot fully understand the effect of SC on out-
comes without taking individual agency into account (Ahuja et al. 2012; Stuart and Sorenson
2007)a reality that should assume a more central position in SC research and theorizing (Bensaou
et al. 2014).
Practical Implications
Our findings have several practical implications for (work-related) SMNs that want to better design
45
and promote their platforms, on the one hand, and for users who want to capitalize on premium fea-
tures, on the other. From a platform’s perspective, the chief focus should be encouraging users to
network and accumulate SC. Users who have established a large network on one SMN are unlikely
to repeat the effort on another SMN that fulfills the same purpose. Giveaways of trial periods of pre-
mium membership are a common marketing instrument in freemium businesses (Koch and Benlian
2017), and they can be used to foster network effects, among others, in work-related SMNs. How-
ever, such a free giveaway only leads to the formation of new contacts if premium users are moti-
vated to network. Hence, free giveaways are less suitable for “waking up” passive networkers. In-
stead, SMNs should target those users with a high agency for networking.
The strong association between premium users’ messaging and SC accumulation highlights the
role of a particular premium feature: The improved ability to reach out via personal messages. This
finding provides two strategic directions for platform operators: Because of the value of personal
messages to their users, SMNs should consider differentiating their offer to the actual user needs
through the manifold possibilities of nonlinear pricing (e.g., Schlereth and Skiera 2012). Instead of
offering this important feature in an unlimited way, WorkSMN could differentiate its premium
memberships according to the number of messages that one can send to non-contacts (e.g., ten mes-
sages for 5€ per month; 30 messages for 10€ and unlimited for 99€). A different strategic direction
could be that SMNs incentivize users to actively network: Because SMNs generate a high value
from dense networks with active users, the platform could remunerate the most active users through
free access to advanced networking features.
Moreover, our findings imply that SMNs should provide templates once a user intends to send a
personal message to a non-contact. An algorithm could generate such a template, including a target
user, easing the contact initiation based on similarities between profiles and common contacts. Our
study also highlights that the opportunities provided by advanced networking features are complex
46
and may be challenging to comprehend for users. As a solution, we suggest providing video tutori-
als that display and summarize the advanced networking features, alongside special use cases. Plat-
forms can further improve clarity by embedding instructions and encouragements in their design
(Oestreicher-Singer and Zalmanson 2013).
From a user’s perspective, our insights are beneficial for understanding the value of converting
to a premium membership. Work-related SMNs are often unspecific about the mechanisms of ad-
vanced networking features that a premium membership provides. Hence, attempts to estimate the
contribution of individual premium features in freemium networks have garnered interest (Bapna et
al. 2016). Findings from our field experiment and individual-level panel data agree that passive fea-
tures supply little benefit to SC accumulation itself. Premium users in work-related SMNs, should
not expect that the visual highlighting via prestigious premium badges suffices for successful SC
accumulation. Likewise, the “weak signals” provided through profile visiting are not a strong mech-
anism in work-related SMNs. Instead, the agency for networking is the very basis for successful SC
accumulation. Only those who actively use targeted actions––i.e., profile visiting and personal mes-
saging––that are enabled by active premium features can accomplish increased SC accumulation
through their premium membership.
LIMITATIONS AND FUTURE RESEARCH
First, concerning RQ1 and the randomized field experiment, we collected the variable strategic net-
working behavior in a post-treatment survey. Hence, it potentially has been affected by the experi-
mental treatment of premium membership. To alleviate this concern, we provide statistical tests and
point to theoretical arguments that our analyses do not suffer from a post-treatment bias. Neverthe-
less, we acknowledge that fully addressing this potential contamination requires ruling it out in the
first place, namely by measuring the variables of interest before administering the treatment. Hence,
47
to assess the robustness of our findings from the randomized field experiment, we utilize comple-
mentary individual-level panel data that contain, among others, a pre-treatment variable for strategic
networking behavior. Given the observational nature of the individual-level panel data, we
acknowledge that our additional analyses may leave room for alternative explanations through un-
observed confounders. However, together with the analyses from the randomized field experiment,
they allow us to provide a high level of convergent validity with respect to our findings on active
freelancers on WorkSMN.
Second, our conclusions on RQ2 are based solely on observational data: We generate supportive
evidence––without claiming causality, in contrast to RQ1––by exploiting the panel structure of the
data. Although we gain robustness by applying PSM, a limitation remains since it only matches
based on observables. Hence, unobserved confounders can theoretically still bias the results. Such
confounders may relate to job satisfaction and the risk of losing a job because users who consider
switching jobs may be more active on WorkSMN and simultaneously more likely to convert to pre-
mium. We assume that such confounders are sufficiently captured by the self-indicated job-seeking
status and the additional behavioral variables––activity and saliency––that can explain the latent sit-
uation of a user (Ebbes and Netzer 2018). Yet, PSM does not allow us to rule out the existence of
additional unobserved confounders. Hence, future research should rely on randomized field experi-
ments to further understand the role of advanced networking features in SC accumulationideally
by directly manipulating platform features (Bapna et al. 2016; Gee 2019).
Third, our findings’ generalizability might be limited because our studies focused on a certain
population. More specifically, utilizing a randomized field experiment and individual-level panel
data, we gain convergent validity about our findings on freelancers who are active on WorkSMN.
Besides, the study population stems from one country. Therefore, we cannot generalize the findings
to other countries, other professional categories, nor inactive accounts. Furthermore, we conducted
48
our analyses on WorkSMN data; hence, our findings may not be generalizable to other platforms.
Note that the generalizability of the field experiment’s CACE estimate might be limited as well
(Bisbee et al. 2017; Mogstad and Torgovitsky 2018), as it only indicates the causal effect for the
subpopulation of compliers, reflecting 83.59% of our sample, and does not provide the average
treatment effect for the non-compliers. While we would not expect a strong response from non-re-
deemers (12.16%), who may not even have noticed a change in membership type, we can only hy-
pothesize how the treatment would have affected them and the always-takers (4.25%). To examine
external validity, future studies should apply our analyses to other work-related SMNs, ensuring that
the study population differs in terms of professional category and geographical location. Concerning
the field experiment, future studies could impose premium memberships to avoid non-compliance.
Finally, all our work focuses on the structural dimension of SC, which we operationalized in line
with prior research as the number of contacts, i.e., node degree (e.g., Hinz et al. 2015; Lanz et al.
2019). While this measure is easy to harvest manually by visiting the member’s profile, it is also a
very crude measure of the concept: It does not differentiate how many contacts the new connection
led to and whether the new connection ended up as a meaningless entry in the user's contact list. Be-
sides, by only looking at SC in terms of the plain acceptance of a sent digital contact request, our
work does not consider that contacts might possess an unequal interest in their potential instrumen-
tality. Thus, future research should explore how and why individuals cultivate some of those accu-
mulated instrumental ties while allowing others to remain a digital artifact.
REFERENCES
Adler, P. S., and Kwon, S.-W. 2002. "Social Capital: Prospects for a New Concept," Academy of
Management Review (27:1), pp. 17-40.
Ahuja, G., Soda, G., and Zaheer, A. 2012. "The Genesis and Dynamics of Organizational
Networks," Organization Science (23:2), pp. 434-448.
49
Andrews, D. W., and Lu, B. 2001. "Consistent Model and Moment Selection Procedures for Gmm
Estimation with Application to Dynamic Panel Data Models," Journal of Econometrics
(101:1), pp. 123-164.
Angrist, J. D. 2006. "Instrumental Variables Methods in Experimental Criminological Research:
What, Why and How," Journal of Experimental Criminology (2:1), pp. 23-44.
Angrist, J. D., and Pischke, J.-S. 2008. Mostly Harmless Econometrics: An Empiricist's Companion.
Princeton university press.
Aten, K., DiRenzo, M., and Shatnawi, D. 2017. "Gender and Professional E-Networks: Implications
of Gender Heterophily on Job Search Facilitation and Outcomes," Computers in Human
Behavior (72), pp. 470-478.
Bapna, R., Ramaprasad, J., Shmueli, G., and Umyarov, A. 2016. "One-Way Mirrors in Online
Dating: A Randomized Field Experiment," Management Science (62:11), pp. 3100-3122.
Bapna, R., Ramaprasad, J., and Umyarov, A. 2018. "Monetizing Freemium Communities: Does
Paying for Premium Increase Social Engagement?," MIS Quarterly (42:3), pp. 719-735.
Bapna, R., and Umyarov, A. 2015. "Do Your Online Friends Make You Pay? A Randomized Field
Experiment on Peer Influence in Online Social Networks," Management Science (61:8), pp.
1902-1920.
Bapna, S., and Funk, R. 2020. "Interventions for Improving Professional Networking for Women:
Experimental Evidence from the It Sector," MIS Quarterly (forthcoming).
Bensaou, B. M., Galunic, C., and Jonczyk-Sédès, C. 2014. "Players and Purists: Networking
Strategies and Agency of Service Professionals," Organization Science (25:1), pp. 29-56.
Bisbee, J., Dehejia, R., Pop-Eleches, C., and Samii, C. 2017. "Local Instruments, Global
Extrapolation: External Validity of the Labor SupplyFertility Local Average Treatment
Effect," Journal of Labor Economics (35:S1), pp. S99-S147.
Bollen, K. A. 2012. "Instrumental Variables in Sociology and the Social Sciences," Annual Review
of Sociology (38), pp. 37-72.
Borgatti, S. P., and Halgin, D. S. 2011. "On Network Theory," Organization Science (22:5), pp.
1168-1181.
Borgatti, S. P., Jones, C., and Everett, M. G. 1998. "Network Measures of Social Capital,"
Connections (21:2), pp. 27-36.
Bucher, T., and Helmond, A. 2017. "The Affordances of Social Media Platforms," The SAGE
Handbook of Social Media), pp. 233-253.
Burbano, V. C. 2016. "Social Responsibility Messages and Worker Wage Requirements: Field
Experimental Evidence from Online Labor Marketplaces," Organization Science (27:4), pp.
1010-1028.
Casciaro, T., Gino, F., and Kouchaki, M. 2014. "The Contaminating Effects of Building
Instrumental Ties: How Networking Can Make Us Feel Dirty," Administrative Science
Quarterly (59:4), pp. 705-735.
Chen, H., De, P., and Hu, Y. J. 2015. "It-Enabled Broadcasting in Social Media: An Empirical
Study of Artists’ Activities and Music Sales," Information Systems Research (26:3), pp. 513-
531.
Chen, X., Van Der Lans, R., and Phan, T. Q. 2017. "Uncovering the Importance of Relationship
Characteristics in Social Networks: Implications for Seeding Strategies," Journal of
Marketing Research (54:2), pp. 187-201.
Dewan, S., and Ramaprasad, J. 2014. "Social Media, Traditional Media, and Music Sales," MIS
Quarterly (38:1), pp. 101-122.
50
Ebbes, P., and Netzer, O. 2018. "Using Social Network Activity Data to Identify and Target Job
Seekers." SSRN (3200214).
Ebbes, P., Papies, D., and van Heerde, H. J. 2016. "Dealing with Endogeneity: A Nontechnical
Guide for Marketing Researchers," in Handbook of Market Research, C. Homburg, M.
Klarmann and A. Vomberg (eds.). Springer International Publishing AG.
Eckles, D., and Bakshy, E. 2020. "Bias and High-Dimensional Adjustment in Observational Studies
of Peer Effects," Journal of the American Statistical Association), pp. 1-11.
Forret, M. L. 2018. "Networking as a Job-Search Behavior and Career Management Strategy," The
Oxford Handbook of Job Loss and Job Search), p. 275.
Forret, M. L., and Dougherty, T. W. 2001. "Correlates of Networking Behavior for Managerial and
Professional Employees," Group & Organization Management (26:3), pp. 283-311.
Gandini, A. 2016. "Digital Work:Self-Branding and Social Capital in the Freelance Knowledge
Economy," Marketing Theory (16:1), pp. 123-141.
Garg, R., and Telang, R. 2018. "To Be or Not to Be Linked: Online Social Networks and Job Search
by Unemployed Workforce," Management Science (64:8), pp. 3926-3941.
Gee, L. K. 2019. "The More You Know: Information Effects on Job Application Rates in a Large
Field Experiment," Management Science (65:5), pp. 2077-2094.
Gee, L. K., Jones, J., and Burke, M. 2017. "Social Networks and Labor Markets: How Strong Ties
Relate to Job Finding on Facebook’s Social Network," Journal of Labor Economics (35:2),
pp. 485-518.
Gerber, A. S., and Green, D. P. 2012. Field Experiments: Design, Analysis, and Interpretation. W.
W. Norton & Company.
Granovetter, M. S. 1973. "The Strength of Weak Ties," American Journal of Sociology (78:6), pp.
1360-1380.
Hahn, E., Gottschling, J., and Spinath, F. M. 2012. "Short Measurements of PersonalityValidity
and Reliability of the Gsoep Big Five Inventory (Bfi-S)," Journal of Research in Personality
(46:3), pp. 355-359.
Hansen, L. P. 1982. "Large Sample Properties of Generalized Method of Moments Estimators,"
Econometrica (50:4), pp. 1029-1054.
Harris, R. D., and Tzavalis, E. 1999. "Inference for Unit Roots in Dynamic Panels Where the Time
Dimension Is Fixed," Journal of Econometrics (91:2), pp. 201-226.
Haynes, J., and Marshall, L. 2018. "Beats and Tweets: Social Media in the Careers of Independent
Musicians," New Media & Society (20:5), pp. 1973-1993.
Hinz, O., Hill, S., and Kim, J.-Y. 2016. "Tv's Dirty Little Secret: The Negative Effect of Popular Tv
on Online Auction Sales," MIS Quarterly (40:3), pp. 623-644.
Hinz, O., Otter, T., and Skiera, B. 2020. "Estimating Network Effects in Two-Sided Markets,"
Journal of Management Information Systems (37:1), pp. 12-38.
Hinz, O., Skiera, B., Barrot, C., and Becker, J. U. 2011. "Seeding Strategies for Viral Marketing: An
Empirical Comparison," Journal of Marketing (75:6), pp. 55-71.
Hinz, O., Spann, M., and Hann, I.-H. 2015. "Research Note—Can’t Buy Me Love… or Can I?
Social Capital Attainment through Conspicuous Consumption in Virtual Environments,"
Information Systems Research (26:4), pp. 859-870.
Hoisl, K., and Mariani, M. 2017. "It’sa Man’s Job: Income and the Gender Gap in Industrial
Research," Management Science (63:3), pp. 766-790.
Imbens, G. W., and Rubin, D. B. 2015. Causal Inference in Statistics, Social, and Biomedical
Sciences. Cambridge University Press.
51
Kane, G. C., Alavi, M., Labianca, G., and Borgatti, S. P. 2014. "What’s Different About Social
Media Networks? A Framework and Research Agenda," MIS Quarterly (38:1), pp. 275-304.
Karahanna, E., Xu, S. X., Xu, Y., and Zhang, N. A. 2018. "The NeedsAffordancesFeatures
Perspective for the Use of Social Media," MIS Quarterly (42:3), pp. 737-756.
Koch, O. F., and Benlian, A. 2017. "The Effect of Free Sampling Strategies on Freemium
Conversion Rates," Electronic Markets (27:1), pp. 67-76.
Kuhn, K. M. 2016. "The Rise of the "Gig Economy" and Implications for Understanding Work and
Workers," Industrial and Organizational Psychology (9:1), p. 157.
Kuwabara, K., Hildebrand, C. A., and Zou, X. 2018. "Lay Theories of Networking: How
Laypeople’s Beliefs About Networks Affect Their Attitudes toward and Engagement in
Instrumental Networking," Academy of Management Review (43:1), pp. 50-64.
Lanz, A., Goldenberg, J., Shapira, D., and Stahl, F. 2019. "Climb or Jump: Status-Based Seeding in
User-Generated Content Networks," Journal of Marketing Research (56:3), pp. 361-378.
Laurie, H., and Lynn, P. 2009. The Use of Respondent Incentives on Longitudinal Surveys. Wiley
Online Library.
Lee, J. J. 2010. "Heterogeneity, Brokerage, and Innovative Performance: Endogenous Formation of
Collaborative Inventor Networks," Organization Science (21:4), pp. 804-822.
Leonardi, P. M. 2013. "When Does Technology Use Enable Network Change in Organizations? A
Comparative Study of Feature Use and Shared Affordances," MIS Quarterly), pp. 749-775.
Leung, M. D. 2014. "Dilettante or Renaissance Person? How the Order of Job Experiences Affects
Hiring in an External Labor Market," American Sociological Review (79:1), pp. 136-158.
Li, Z., Fang, X., Bai, X., and Sheng, O. R. L. 2017. "Utility-Based Link Recommendation for
Online Social Networks," Management Science (63:6), pp. 1938-1952.
Liben‐Nowell, D., and Kleinberg, J. 2007. "The Link‐Prediction Problem for Social Networks,"
Journal of the American Society for Information Science and Technology (58:7), pp. 1019-
1031.
Lin, N. 1999. "Social Networks and Status Attainment," Annual review of sociology (25:1), pp. 467-
487.
Lo Presti, A., Pluviano, S., and Briscoe, J. P. 2018. "Are Freelancers a Breed Apart? The Role of
Protean and Boundaryless Career Attitudes in Employability and Career Success," Human
Resource Management Journal (28:3), pp. 427-442.
Lonati, S., Quiroga, B. F., Zehnder, C., and Antonakis, J. 2018. "On Doing Relevant and Rigorous
Experiments: Review and Recommendations," Journal of Operations Management (64), pp.
19-40.
Manago, A. M., Taylor, T., and Greenfield, P. M. 2012. "Me and My 400 Friends: The Anatomy of
College Students' Facebook Networks, Their Communication Patterns, and Well-Being,"
Developmental Psychology (48:2), p. 369.
McCarthy, J. E., and Levin, D. Z. 2019. "Network Residues: The Enduring Impact of Intra-
Organizational Dormant Ties," Journal of Applied Psychology).
Mogstad, M., and Torgovitsky, A. 2018. "Identification and Extrapolation of Causal Effects with
Instrumental Variables," Annual Review of Economics (10:1), pp. 577-613.
Mouw, T. 2006. "Estimating the Causal Effect of Social Capital: A Review of Recent Research,"
Annu. Rev. Sociol. (32), pp. 79-102.
Nahapiet, J., and Ghoshal, S. 1998. "Social Capital, Intellectual Capital, and the Organizational
Advantage," Academy of management review (23:2), pp. 242-266.
Nikitkov, A., and Sainty, B. 2014. "The Role of Social Media in Influencing Career Success,"
International Journal of Accounting & Information Management (22:4), pp. 273-294.
52
Oestreicher-Singer, G., and Zalmanson, L. 2013. "Content or Community? A Digital Business
Strategy for Content Providers in the Social Age," MIS Quarterly (37:2), pp. 591-616.
Pollack, J. M., Forster, W. R., Johnson, P. D., Coy, A., and Molden, D. C. 2015. "Promotion-and
Prevention-Focused Networking and Its Consequences for Entrepreneurial Success," Social
Psychological and Personality Science (6:1), pp. 3-12.
Porter, C. M., and Woo, S. E. 2015. "Untangling the Networking Phenomenon: A Dynamic
Psychological Perspective on How and Why People Network," Journal of Management
(41:5), pp. 1477-1500.
Portes, A. 1998. "Social Capital: Its Origins and Applications in Modern Sociology," Annual review
of sociology (24:1), pp. 1-24.
Putnam, R. D. 2000. "Bowling Alone: America’s Declining Social Capital," in Culture and Politics.
Springer, pp. 223-234.
Rosenbaum, P. R. 2020. "Modern Algorithms for Matching in Observational Studies," Annual
Review of Statistics and Its Application (7), pp. 143-176.
Rostila, M. 2011. "The Facets of Social Capital," Journal for the Theory of Social Behaviour (41:3),
pp. 308-326.
Sagarin, B. J., West, S. G., Ratnikov, A., Homan, W. K., Ritchie, T. D., and Hansen, E. J. 2014.
"Treatment Noncompliance in Randomized Experiments: Statistical Approaches and Design
Issues," Psychological Methods (19:3), pp. 317-333.
Sanderson, E., and Windmeijer, F. 2016. "A Weak Instrument F-Test in Linear Iv Models with
Multiple Endogenous Variables," Journal of Econometrics (190:2), pp. 212-221.
Schlereth, C., and Skiera, B. 2012. "Measurement of Consumer Preferences for Bucket Pricing
Plans with Different Service Attributes," International Journal of Research in Marketing
(29:2), pp. 167-180.
Seibert, S. E., Kraimer, M. L., and Liden, R. C. 2001. "A Social Capital Theory of Career Success,"
Academy of Management Journal (44:2), pp. 219-237.
Shapiro, C., and Varian, H. R. 1999. Information Rules: A Strategic Guide to the Network Economy.
Harvard Business Press.
Shevchuk, A., and Strebkov, D. 2018. "Safeguards against Opportunism in Freelance Contracting
on the Internet," British Journal of Industrial Relations (56:2), pp. 342-369.
Shriver, S. K., Nair, H. S., and Hofstetter, R. 2013. "Social Ties and User-Generated Content:
Evidence from an Online Social Network," Management Science (59:6), pp. 1425-1443.
Sovey, A. J., and Green, D. P. 2011. "Instrumental Variables Estimation in Political Science: A
Readers’ Guide," American Journal of Political Science (55:1), pp. 188-200.
Stock, J. H., and Watson, M. W. 2015. Introduction to Econometrics, (Third (Updated) ed.).
Pearson Education.
Stock, J. H., and Yogo, M. 2005. "Testing for Weak Instruments in Linear Iv Regression," 0898-
2937, National Bureau of Economic Research, New York, pp. 80-108.
Stuart, T. E., and Sorenson, O. 2007. "Strategic Networks and Entrepreneurial Ventures," Strategic
Entrepreneurship Journal (1:3‐4), pp. 211-227.
Sun, C., Shi, Z. J., Liu, X., Ghose, A., Li, X., and Xiong, F. 2019a. "The Effect of Voice Ai on
Consumer Purchase and Search Behavior," SSRN (3480877) (Working Paper).
Sun, T., Shi, L., Viswanathan, S., and Zheleva, E. 2019b. "Motivating Effective Mobile App
Adoptions: Evidence from a Large-Scale Randomized Field Experiment," Information
Systems Research (30:2), pp. 523-539.
53
Sun, T., and Taylor, S. J. 2020. "Displaying Things in Common to Encourage Friendship
Formation: A Large Randomized Field Experiment," Quantitative Marketing and Economics
(Not assigned to an issue), pp. 1-35.
Sundararajan, A., Provost, F., Oestreicher-Singer, G., and Aral, S. 2013. "Research Commentary
Information in Digital, Economic, and Social Networks," Information Systems Research
(24:4), pp. 883-905.
Takac, C., Hinz, O., and Spann, M. 2011. "The Social Embeddedness of Decision Making:
Opportunities and Challenges," Electronic Markets (21:3), p. 185.
Toubia, O., and Stephen, A. T. 2013. "Intrinsic Vs. Image-Related Utility in Social Media: Why Do
People Contribute Content to Twitter?," Marketing Science (32:3), pp. 368-392.
Tröster, C., Parker, A., Van Knippenberg, D., and Sahlmüller, B. 2019. "The Coevolution of Social
Networks and Thoughts of Quitting," Academy of Management Journal (62:1), pp. 22-43.
Utz, S. 2016. "Is Linkedin Making You More Successful? The Informational Benefits Derived from
Public Social Media," New Media & Society (18:11), pp. 2685-2702.
Utz, S., and Breuer, J. 2016. "Informational Benefits from Social Media Use for Professional
Purposes: Results from a Longitudinal Study," Cyberpsychology: Journal of Psychosocial
Research on Cyberspace (10:4).
Van den Born, A., and Van Witteloostuijn, A. 2013. "Drivers of Freelance Career Success," Journal
of Organizational Behavior (34:1), pp. 24-46.
Van Hoye, G., Van Hooft, E. A., and Lievens, F. 2009. "Networking as a Job Search Behaviour: A
Social Network Perspective," Journal of Occupational and Organizational Psychology
(82:3), pp. 661-682.
Vissa, B. 2011. "A Matching Theory of Entrepreneurs' Tie Formation Intentions and Initiation of
Economic Exchange," Academy of Management Journal (54:1), pp. 137-158.
Vissa, B. 2012. "Agency in Action: Entrepreneurs' Networking Style and Initiation of Economic
Exchange," Organization Science (23:2), pp. 492-510.
Voigt, S., and Hinz, O. 2016. "Making Digital Freemium Business Models a Success: Predicting
Customers’ Lifetime Value Via Initial Purchase Information," Business & Information
Systems Engineering (58:2), pp. 107-118.
Wasko, M. M., and Faraj, S. 2005. "Why Should I Share? Examining Social Capital and Knowledge
Contribution in Electronic Networks of Practice," MIS Quarterly (29:1), pp. 35-57.
Weiler, M., and Hinz, O. 2019. "Without Each Other, We Have Nothing: A State-of-the-Art
Analysis on How to Operationalize Social Capital," Review of Managerial Science (13:5),
pp. 1003-1035.
Wolff, H.-G., and Moser, K. 2009. "Effects of Networking on Career Success: A Longitudinal
Study," Journal of Applied Psychology (94:1), p. 196.
Wu, L. 2013. "Social Network Effects on Productivity and Job Security: Evidence from the
Adoption of a Social Networking Tool," Information Systems Research (24:1), pp. 30-51.
54
Appendix:
Social Capital Accumulation Through Social Media Networks: Evidence from a Randomized
Field Experiment and Individual-Level Panel Data
55
A. RELATED WORK ON SOCIAL CAPITAL (SC) AND SOCIAL MEDIA NETWORKS (SMN)
Table 9. Overview of related work: SC and work-related SMNs
Study
Data/Method
Measure(s) of SC
Outcome(s)
Main Finding
Mayer
(2012)
Multiple, linked data sources:
(1) administrative records, (2) self-
reported survey data, (3) observed
FB data. n = 3,303 students from an
American university.
Network structure (i.e.,
number of FB friends,
number of friends of
friends, cluster coeffi-
cient)
Salary
Employment status
No strong empirical support for Granovetter’s (1973) strength-
of-weak-tie hypothesis.
Rienties et
al. (2012)
Online survey; cross-sectional data;
n = 386 encompassing students,
non-managers, and managers.
Total number of overall
digitized network contacts
and how many of those
are strong or weak ties
Structural holes
Receiving of job infor-
mation
LI users with larger networks receive more job-related infor-
mation. A larger number of weak ties within one’s LI network
correlates positively with the obtainment of job-related infor-
mation (see also Granovetter 1973). However, no empirical
support for Burt’s (2009) structural holes hypothesis is pro-
vided.
Burke and
Kraut (2013)
Three-wave panel survey of FB us-
ers. Survey responses were linked
to the participant’s digital footprint
FB data. n = 3,358.
Tie strength (i.e., strong or
weak ties) (self-reported)
Finding a job
Communicating with strong FB ties correlates positively with
obtaining a new job. However, no empirical support for Gran-
ovetter’s (1973) strength-of-weak-tie hypothesis is provided.
Nikitkov and
Sainty
(2014)
Archival data (academic records)
enriched with manually retrieved
information from SMNs and other
data sources. n = 1,182 accounting
alumni from a Canadian university.
Number of contacts on
different SMNs, such as
FB and LI
Being a member on
SMNs
Twelve career success
indicators, such as being
employed at a large ac-
counting firm
Compared to non-work-related SMNs, the presence and num-
ber of contacts on the work-related SMN LI is positively asso-
ciated with indicators of career success.
Zhitomirsky-
Geffet and
Bratspiess
(2015)
Online survey of Israeli FB and LI
users; n = 222.
Network utilization
Perceived and actual ef-
fectiveness of FB and LI
in job finding processes
Participants perceive LI (in comparison to FB) to be more ef-
fective in helping them find a job. However, in terms of actual
helpfulness, neither SMN seems to differ. Individuals who re-
veal more information about themselves are more effective us-
ing SMNs to search for employment than those who reveal
less information.
Utz (2016)
Representative online survey of
Dutch Internet users; n = 1,959.
Network structure (i.e., to-
tal number of SMN con-
tacts; their tie strength;
strategic networking be-
havior)
Posting professional con-
tent
Professional informa-
tional benefits
Compared to non-users, LI users report greater professional
informational benefits (e.g., timely access to information and
referrals). Notably, strong and weak ties provide these benefits
in equal measures. Yet, informational benefits obtained
through FB are primarily linked to strong ties. Engaging in
strategic networking behavior also leads to informational ben-
efits.
56
Utz and
Breuer
(2016)
Longitudinal data (four waves);
representative online survey of
Dutch Internet users.
Wave 1: n = 3,367
Wave 2: n = 2,678
Wave 3: n = 2,273
Wave 4: n = 1,953
Network structure (i.e., to-
tal number of SMN con-
tacts and how many of
those are strong or weak
ties; strategic networking
behavior)
Posting professional con-
tent
Professional informa-
tional benefits
Participants who experience more informational benefits from
using LI are more likely to use the platform. Networking pro-
cesses take some time for the corresponding effort to unfold,
as strategic networking at one-point leads to informational
benefits at a later point in time.
The benefit of receiving informational benefits stimulates fur-
ther strategic networking behavior. The importance of strong
and weak ties seems to favor strong ties over the course of
time. The number of LI contacts predicts informational bene-
fits at a later point in time.
Aten et al.
(2017)
Online survey of LI users; n = 366.
Network composition
(i.e., number of LI con-
tacts)
Network facilitation (i.e.,
assistance of LI contact in
getting the job)
Job search duration
Salary attained
LI users whose networks consist of many connections with
male contacts are not only likely to find a new job more
quickly than those without as many male contacts but also to
attain a higher income.
Buettner
(2017)
Online survey of work-related SMN
(i.e., XING and an enterprise solu-
tion) users; n = 523.
Number of contacts (i.e.,
node degree)
Receiving job offers
The relationship between the number of digitized contacts and
the obtainment of job offers runs in a nonlinear way. SMN us-
ers receive an increasing number of job offers when they have
up to approximately 150 SMN contacts, which reflects the so-
called Dunbar number. However, beyond that threshold num-
ber, more work-related contacts do not seem to translate into
more job offers, which indicates that the arbitrary befriending
of other users “[…] does not make sense in terms of getting
job offers” (Buettner 2017, p. 380).
Baruffaldi et
al. (2017)
Multiple, linked data sources, en-
compassing (1) LI digital footprint
data, (2) university records, (3)
publication records, and (4) infor-
mation from curricula vitae; n = 422
PhD graduates.
Total number of LI con-
nections (i.e., node de-
gree)
Total number of LI con-
nections
Having an LI account
PhD holders who leave academia to work in the industry sec-
tor cultivate larger LI networks. The same result (i.e., having
larger LI networks) emerges for PhD holders who work in an-
other country after finishing their theses. Compared to
younger PhD graduates, older ones cultivate smaller LI net-
works.
Gee et al.
(2017)
FB digital footprint data partially
enriched with survey data; n =
12,263 FB users and more than 1.4
million dyads.
Tie strength (i.e., strong or
weak ties)
Working with a friend
for the same employer
(reflects a proxy meas-
ure “[…] for being
helped by a connection
to find a job” (Gee et al.
2017, p. 491))
Weak ties collectively unfold their potential in helping indi-
viduals find a job due to their sheer number in FB users’ net-
works. However, when looking closer at a single dyad, strong
ties seems to be more helpful than their weak counterparts.
57
Garg and
Telang
(2018)
Survey of unemployed job seekers;
n = 424.
Total number of SMN
contacts (i.e., node de-
gree)
Tie strength (i.e., strong
or weak ties)
Job outcomes (i.e.,
leads, interviews, and
offers)
For unemployed individuals, maintaining a large LI network is
related to receiving more job interviews and offers. While
strong LI ties play a central role in obtaining job leads, inter-
views, and offers, weak ties are largely unrelated to such bene-
ficial outcomes.
Gloor et al.
(2018)
Publicly accessible digital footprint
data from the work-related SMN
XING, combined with university
data, company websites, and a busi-
ness database. n = 981385 crawled
user profiles. Analysis on small
subset of n = 960 entrepreneurs and
n = 960 non-entrepreneurs.
Degree centrality (i.e.,
network size)
Betweenness centrality
Entrepreneurial success
Neither maintaining a large digitized network, nor being cen-
tral in terms of a high betweenness centrality is positively cor-
related with entrepreneurial success.
Mashayekhi
and Head
(2018)
Survey of adult LI users; n = 275.
Network size (i.e., number
of LI contacts)
Perceived social connect-
edness
Perceived networking
value (defined, inter alia,
as access to information,
referrals, or influence)
LI users who reveal more information on their profile page not
only maintain larger digitized networks but also experience
more social connectedness than those who reveal less infor-
mation. Both variables (i.e., network size and social connect-
edness), on the other hand, correlate positively with the net-
working value of LI.
Banerji and
Reimer
(2019)
Digital footprint data from Crunch-
base and LI of startup companies (n
= 129) and their founder(s) (n =
227).
Network size (defined as
the number of LI follow-
ers)
Raised amount of finan-
cial resources
The authors provide correlational evidence that startup compa-
nies raise more money when their founder(s) possess more LI
connections.
Ma and
Leung
(2019)
Online survey of Chinese LI users;
n = 301.
Perceived bridging SC
(based on Williams 2006)
Perceived bridging SC
Using LI to follow and react to professional information and to
participate in strategic networking behavior is correlated with
higher levels of perceived SC.
Utz and
Breuer
(2019)
Representative online survey of
Dutch Internet users; n = 685 work-
ing individuals, 259 of whom are LI
users.
Network structure (i.e., to-
tal number of SMN con-
tacts, and their tie
strength)
Professional content
Professional informa-
tional benefits
Extending their previous works (c.f., Utz 2016; Utz and
Breuer 2016), the authors show that individuals who partici-
pate more in external networking endeavors are more likely to
be LI users. Both networking styles (i.e., internal and external)
correlate with receiving professional informational benefits on
the platform. However, engaging intensively in external net-
working leads to an accrual of larger numbers of strong and la-
tent ties on LI.
Davis et al.
(2020)
Paper and pencil survey data (n =
133) and digital footprint data (n =
56) from American business stu-
dents
Total number of LI con-
tacts (i.e., node degree)
(self-reported as well as
digital footprint)
Career benefits
Although the frequency of LI usage seems to be important to
utilize positive career outcomes, the maintenance of a large
network appears to play no pivotal role.
Note: Facebook = FB; LinkedIn = LI.
58
B. OVERVIEW OF FEATURES ON WORKSMN
Table 10. Overview of features on WorkSMN by membership type
Basic
membership
Premium
membership
Active or
Passive
More detailed profile visitor’s statistics
View profile visitors, including their full
name, position, and company
-
Passive
Find out how profile visitors came across
your profile and the date they visited it
-
Passive
Link to the visitor’s profile page
-
Passive
More professional appearance through cus-
tomization of profile page
No ad banners visible
-
Passive
Possibility to customize business card
with individual cover image
-
(Both)
Possibility to highlight your top skills
-
(Both)
Possibility to add and embed more multi-
media content to your profile page includ-
ing videos from YouTube
-
(Both)
More advanced member search and getting
found more easily
Member search, maximum number of dis-
played search results
10
300
Active
More search options, including special
search filters (e.g., working field, com-
pany, location)
Active
Possibility to create automatic search
alerts
-
Active
Highlighted entry of your profile in search
results
-
Passive
More efficient communication and net-
working
Possibility to send contact requests with a
personalized message plus to browse and
edit pending contact requests
-
Active
Possibility to send messages to non-con-
tacts
-
Active
Updates of network contacts are visible at
a glance (e.g., if they changed a job)
-
Passive
Note: WorkSMN features as they were present at the time of conducting this study.
59
C. FURTHER INFORMATION ON THE FIELD EXPERIMENT (RQ1)
C.1 Flow Diagram
Figure 3. Flow diagram of the freelancers’ recruitment, treatment condition assignment, and attrition
Complete responses of freelancers in pre-
study, n = 531
Freelancers who qualified for the experiment,
n = 243
Freelancers who qualified for the randomi-
zation procedure, n = 217
Data amendments 1: Exclusion of 26 free-
lancers, who quit working as a freelancer
etc.)
Treatment group: n = 74
Control group: n = 141
Treated freelancers who redeemed the
voucher, n = 65
Treated freelancers in follow-up
study, n = 52
Data amendments 2: Exclusion of two
freelancers (technical scrapping prob-
lem and quit using the platform)
Non-treated freelancers in follow-up
study, n = 78
Gifted freelancers who did not re-
deemed voucher, n = 9
Freelancers who are assigned to control
group but received treatment, n = 6
Freelancers who quit working as free-
lancers, n = 10
60
C.2 Elaboration on the Complier Average Causal Effect (CACE)
In this section, we elaborate on why the CACE is the most appropriate strategy for our research pur-
pose and provide more (technical) details on the employed approach.
Issues of the Per-Protocol and As-Treated Approaches to Handle Non-compliance
The per-protocol (PP) and as-treated (AT) approaches are rather inadequate in their way of address-
ing non-compliance, as they take away the salient and unique benefit that is established through the
randomization process, i.e., the introduction of balance of observed and unobserved characteristics
across the treatment and control condition. While the AT approach reassigns non-compliant partici-
pants for the analysis, the PP approach excludes those deviant ones altogether (Sagarin et al. 2014).
To put this into perspective, the AT approach looks at participants as they self-select into the treat-
ment condition and, therefore, pays no attention to the initial group assignment (Sagarin et al. 2014).
For instance, if a freelancer assigned to the control group does buy a premium membership sub-
scription on his/her own, this formerly untreated freelancer will now be analyzed as treated. Like-
wise, a freelancer who is assigned to receive the treatment but does not redeem the corresponding
premium membership voucher will be reassigned to the untreated condition. The PP approach, how-
ever, keeps the initial group assignment intact, but simply eliminates from the analysis all those par-
ticipants who did not stick to their assigned treatment condition (Sagarin et al. 2014). Unsurpris-
ingly, “[…] in the presence of noncompliance, there is no compelling justification for these two […]
approaches [i.e., AT and PP]” (Imbens and Rubin 2015, p. 535; Sagarin et al. 2014).
In comparison with the AT or PP approach, the intention-to-treat (ITT) approach focuses
solely on the groups as they were initially created by the randomization process (i.e., the actual as-
signment). Consequently, this approach is advantageous, as it perpetuates the initial groups and is
therefore able to produce causal estimates. Despite this undeniable benefit, however, the ITT ne-
61
glects the actual receipt of the treatment, and therefore no empirical evidence emerges about this ef-
fect of interest (Sagarin et al. 2014). Taken together, as highlighted, none of these three outlined ap-
proaches can produce unbiased causal estimates regarding the effect of the treatment received
(Imbens and Rubin 2015; Sagarin et al. 2014). Thus, against this backdrop, we chose the CACE ap-
proach on account of the fact that, in the face of non-compliance, it is the only strategy that is able
to reveal a causal estimate of treatment received on the outcome, namely for the subgroup of the
compliers (for a detailed description of the term see the next subsection) (Imbens and Rubin 2015).
Rationale of the CACE approach
In the light of non-compliance, we can classify freelancers into different groups based on the unique
arrangement of their potential outcomes di(z). While z refers to whether the freelancer is assigned to
the control group (z=0) or the treatment group (z=1), di refers to whether the corresponding free-
lancer is actually treated (di =1) or not (di =0) (Gerber and Green 2012). Thus, in the face of two-
sided non-compliance, four different groups exist:
(1) We can classify freelancers as “compliers” when they are a) assigned to receive the premium
membership subscription voucher (treatment group) and they actually redeem it (di (1) = 1), as
well as when they are b) assigned to the control group and they remain untreated (i.e., they do
not voluntarily buy one premium membership subscription on their own) (di (0) = 0). The ra-
tionale of what we just delineated is visualized by the blue crosshatched areas in Figure 2.
(2) We can categorize freelancers as “never-takers” when they are a) assigned to the treatment
group, but do not actually redeem the received premium membership subscription voucher (di
(1) = 0), as well as when they are b) assigned to the control group and do not voluntarily buy a
subscription on their own (di (0) = 0). In Figure 2, never-takers are displayed by the red colored
area.
62
(3) We can group freelancers as “always-takers” when they are a) assigned to receive the premium
membership subscription voucher and they actually redeem it (di (1) = 1), as well as when they
are b) assigned to the control group, but voluntarily buy a premium membership subscription on
their own (di (0) = 1). In Figure 2, this group is visualized by the green colored area. Regardless
of their treatment assignment, always-takers and never-takers do not change their corresponding
treatment condition, respectively.
(4) Finally, we can categorize freelancers as “defiers” when they are a) assigned to receive the pre-
mium membership subscription voucher, but do not actually redeem it (di (1) = 0), as well as
when they are b) assigned to the control group, but voluntarily buy a premium membership sub-
scription (di (0) = 1) (Gerber and Green 2012; Imbens and Rubin 2015). We can subsume groups
two, three, and four under the notion of “non-compliers”. As outlined in the upcoming subsec-
tion, due to the monotonicity assumption, we assume no defiers in our study and therefore,
there is no reason to visualize them in Figure 2.
63
Figure 4. Graphical representation of the rationale of the CACE and the ITT (in the style of Sun et al.
(2019b))
Notes: The semicolon rimmed box in the right corner (in the style of Lousdal (2018)) graphically details how the differ-
ent groups (i.e., always-takers (green), never-takers (red), and compliers (blue)) respond to their assigned treatment
condition (z). For the compliers, the group of interest within the CACE approach, we see that their treatment status mir-
rors exactly one-to-one their randomly assigned treatment condition. However, the other two groups remain unrespon-
sive to the corresponding treatment offer (Angrist and Pischke 2008).
Figure 4 also illustrates the technical construction of the CACE estimate. The CACE can be
manually estimated using two separate OLS regressions. However, this strategy can produce incor-
rect standard errors. A better way is to use the genuine and expedient 2SLS-IV regression routines
available in statistical analysis software tools (Angrist 2006) (see section “Accumulating Social
Capital: Causal Evidence (RQ1)” in the main text, in which we elaborate on the utilized 2SLS-IV
regression model). Moreover, the identification of the CACE rests on a series of assumptions, which
have to be fully met to arrive at unbiased estimates (see the upcoming subsection “Underlying As-
sumptions for the Identification of the CACE”).
Specifically, from a statistical point of view, we can regard the CACE estimate as the ratio of
two components (i.e., ITT and share of compliers), whereby each one is the output of a specific
64
OLS regression. To put this into perspective, first, we regress “treatment assigned (zi)” on “treat-
ment received (di)”, which ultimately reveals the share of compliers in the study
9
. Second, we re-
gress “treatment assigned (zi)” on the outcome of interest “∆SocalCapitalAccumulation”, which
then estimates the so-called ITT parameter. Finally, we divide both of the obtained coefficients to
identify the   
 (Gerber and Green 2012).
Underlying Assumptions for the Identification of the CACE
Beyond needing a successful random assignment of participants into the treatment and control
groups (see subsection Accumulating Social Capital: Causal Evidence (RQ1) - Data in the main
text), the CACE depends on three key assumptions for an unbiased estimate: First is the non-inter-
ference assumption (also referred to as the stable unit treatment value assumption or SUTVA),
which indicates that the potential outcome of each freelancer depends only on their condition as-
signment and is not affected by other freelancers‘ condition assignment (Gerber and Green 2012).
Second is the excludability presupposition, which assumes that the assignment to the treatment con-
dition zi has no direct effect on the potential outcome Yi. Instead, the assignment only affects the
value of the actual treatment di, which is the corresponding factor that will influence Yi (Gerber and
Green 2012). In other words, being assigned to receive the premium membership subscription has
no direct effect on the accumulation of the freelancers’ SC, except if the receipt changes the free-
lancer to redeem the voucher. Third is the monotonicity assumption, which states that none of the
freelancers in our sample are actually “defiers(Gerber and Green 2012). We will give more details
on this below and elaborate on why these three assumptions likely hold in the present study and,
consequently, that our CACE estimation led to unbiased results.
9
In the language of the 2SLS regression analysis, this step resembles the procedures “first-stageregression.
65
Stable unit treatment value assumption. Although our experiment was performed in a network
setting, where at least theoretically, the treatment of one individual could influence other (direct)
network members (Bapna and Umyarov 2015; Eckles et al. 2017), our design vastly rules out poten-
tial spillover concerns (Walker and Muchnik 2014). The reasons are the following: (a) Generally,
the likelihood of interference between freelancers decreases when the network distance between
them increases (Walker and Muchnik 2014). When we conducted the experiment, WorkSMN had
approximately 14 million platform users, out of whom 153,111 identify themselves as freelancers.
From this latter subset of network members, 215 freelancers comprised the final sample size for our
experiment (see Figure 1 in Appendix C.1). They were recruited from places all over Germany and
a myriad of freelancing working fields. Thus, we believe that the odds would be very small that two
freelancers from the same geographical area (which would also increase the chance of their prior
history of working together), who knew each other on the SMN, would found their way into our
sample. Besides, the number of freelancers assigned to the treatment group was relatively small
compared with the total number of freelancers on the platform, further alleviating spillover issues
(Gerber and Green 2012).
(b) Theoretically, even in the unlikely case that the freelancers in our sample were directly con-
nected to each other, the probability is very low that they interacted or that a single direct connec-
tion would be sufficient to exert influence on a control group member (Walker and Muchnik 2014).
An interaction between two freelancing parties would not necessarily imply a violation of the
SUTVA, as they could have only communicated before the treatment assignment and ceased inter-
acting afterwards (Eckles et al. 2017). Even if two connected freelancers communicated, the chance
that both discussed the study and/or the treatment would seem marginal. As mentioned in the main
text, we informed the part-taking freelancers that we raffled the premium membership subscription
as a thank-you gift for their participation in our pre-study. This implied that they should be unaware
66
that they received the voucher as part of an experimental study and that they were actually partici-
pating in one, respectively (see also Bapna and Umyarov 2015). Hence, the freelancers had only
marginal to no incentive to talk about the study and/or the treatment. For instance, we explicitly
asked our participating freelancers to invite their freelancing colleagues to join our pre-study by for-
warding them the provided online survey link. None of them snowballed our pre-study survey to
their freelancing colleagues. This anecdote provides evidence of the marginal role played by our
study in the lives of the recruited freelancers.
Additionally, they might not attach much importance to their communication about the treat-
ment due to the following theoretical reasons. In broad terms, having access to advanced networking
features is not comparable with something like a novel technique, which perhaps has also restricted
access; rather, it is “nothing really fancy.” A premium membership subscription is similar to a non-
excludable product, as all freelancers can purchase one on their own, as long as they would like to
afford it. The platform service provider promotes its premium membership subscription all over its
service, indicating that every freelancer is probably aware of it. Usually, every time they log in to
the platform, basic members are greeted with a pop-up message, which advertises the premium
membership subscription's benefits. Against the background information that in our sample, the
freelancers’ average tenure on the SMN was approximately seven years, it would seem implausible
that they would be unaware of the existence of such a premium membership subscription (Bapna
and Umyarov 2015). Instead, it might be the case that some of those tenured freelancers had already
gained experience as premium members, having tried the fee-based premium membership in the
past before returning to the free basic membership. In general, not every member was interested in
or needed access to an information system (IS) with advanced networking features, as indicated by
the gifted freelancers who did not redeem the premium membership subscription vouchers given to
67
them. In particular, previous research on freemium business models demonstrated that the willing-
ness to pay for a premium membership subscription was closely linked to the users’ participation
level on the platform, whereby active users were more willing to pay for the product (Oestreicher-
Singer and Zalmanson 2013). Related to this finding, the participants would probably not pay much
attention to the fact that one of their contacts had purchased a premium membership subscription
because, as mentioned, the contact’s purchase would not affect their own access. Finally, the plat-
form members might not even have noticed that one of their contacts now owns a premium mem-
bership subscription, especially if they maintained a large digitized network (Bapna and Umyarov
2015), as they probably rarely visited the profile pages of their contacts (Skeels and Grudin 2009).
Taken together, these reasons give us confidence that the SUTVA holds true in our study.
Excludability assumption. We also assume that this conjecture holds in our study. We expect
that no other actor, such as the SMN service provider itself, could intervene in any way in our study
because none except one of the researchers knew which participants qualified for the experiment
and what their corresponding treatment condition was. Thus, they could not interfere and persuade
the freelancers who did not belong to the treatment group to be more active in their networking be-
havior. As further support, it would be highly unlikely that our set of participants was involved in a
similar experiment that was conducted concurrently to ours and therefore received a treatment from
these researchers. In general, our work-related SMN and our research population under scrutiny had
received only marginal scientific attention so far. Moreover, as indicated by our experimental de-
sign, the participants were neither aware of their treatment condition status nor of the condition of
the others in our participant pool. Being assigned to the control group and thus not receiving the pre-
mium membership voucher would highly unlikely have demoralizing effects on them to make them
reduce their platform activity in general, or their digitized networking behavior in particular. Be-
sides, metaphorically speaking, the premium membership subscription voucher consists only of a
68
random sequence of eight numbers and letters (e.g., nfo7g2ck). Thus, the voucher has to be re-
deemed to reveal its full potential and enable access to the premium features. Otherwise, it would
render itself useless and could not impact the outcome on a direct path (Sun et al. 2019a).
We can also rule out that measurement asymmetry emerged as a threat to the excludability as-
sumption (Gerber and Green 2012). In particular, we ensured that the freelancers were handled in a
uniform way during the experiment with the help of the following steps: (a) We used the same
standardized web-based questionnaire for all freelancers regardless of their treatment condition. (b)
We concurrently gathered the data from both groups (i.e., treatment and control), suggesting no time
lag in the data collection between them. (c) Only one of the researchers was aware of the treatment
condition of each participant. (d) Finally, the same researcher was in charge of handling the commu-
nication with the freelancers while conducting the experiment, which was also achieved by applying
the same standardized procedure (Gerber and Green 2012).
Monotonicity assumption. We suppose that this assumption holds in our study, too, as our treat-
ment assignment incentivized the freelancers to take the treatment rather than disincentivized them
(Baiocchi et al. 2014). Moreover, we expect that in our sample, no freelancer would decline the pre-
mium membership subscription if offered as part of the treatment condition and yet become a pre-
mium subscription member if assigned to the control condition.
69
C.3 Descriptive Statistics and Bivariate Correlations
Table 11. Correlations and descriptive statistics
Note: Significance level p<0.05, as indicated by bold numbers
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[1] SC Accumulation
[2] Treatment Received
0.027
[3] Treatment Assigned
-0.020
0.920
[4] Gender
-0.118
0.076
0.158
[5] Level of Education
0.028
-0.008
-0.058
0.022
[6] Age
-0.212
0.062
0.081
0.236
0.005
[7] Openness
-0.054
0.136
0.067
-0.193
0.012
-0.278
[8] SMN tenure
-0.455
0.058
0.072
0.187
0.103
0.294
-0.019
[9] Strategic networking behavior
0.301
0.038
0.052
-0.014
0.044
-0.150
0.090
-0.162
[10] Other Income
-0.083
0.093
0.071
0.083
0.056
0.134
-0.104
-0.007
-0.073
[11] Work hours per week
0.014
-0.124
-0.148
0.130
0.094
0.005
0.070
0.041
-0.038
-0.227
[12] WorkSMN Usage Intensity
0.275
0.209
0.153
0.077
0.163
0.164
-0.021
0.043
0.236
0.033
0.037
Mean
10.209
0.400
0.392
1.569
2.831
44.946
5.494
7.184
2.358
0.331
38.331
3.615
SD
15.890
0.492
0.490
0.497
0.376
10.989
1.126
3.847
0.911
0.472
16.296
1.572
70
C.4 Covariate Definitions, Summary, and Balance Check After Attrition
Table 12. Overview of variables and their definitions
Variable
Definition
Measure
Digital footprint data (observable)
Gender
Freelancer’s gender
Female = 1; Male = 2
Social capital
Contacts the freelancer maintains on WorkSMN; symbolizes ac-
cess to valuable resources
Number of direct WorkSMN con-
tacts, i.e., node degree (Hinz et al.
2015), visible on profile page
SMN tenure
Amount of time the freelancer is registered as a member on
WorkSMN
Number of years
Number of
haves
Details on skills and prior work experiences the freelancer pos-
sesses, which can serve for self-marketing purposes
Number of ‘haves’ listed on pro-
file page
Number of
wants
(Career) Goals and wishes the freelancer seeks to achieve and to
fulfill while using WorkSMN
Number of ‘wants’ listed on pro-
file page
Number of
groups
Virtual spaces in which the freelancer can engage in discussions
regarding various topics; membership to specific groups can be
interpreted as an individual freelancer’s interests and expertise
Number of memberships to
groups, visible on profile page
Follow-up questionnaire (self-reported)
Age
Freelancer’s age
Number of years alive
Level of edu-
cation
Freelancer’s level of education according to the International
Standard Classification of Education for Germany (Statistisches
Bundesamt 2016)
Low = 1; Middle = 2; High = 3
Openness
Characterizes whether an individual is curious and enjoys em-
bracing new experiences, activities, and/or ideas in his/her life
(Costa Jr and McCrae 2014)
Three-items, seven-point Likert
scale, e.g., “I see myself as some-
one who is original, comes up
with new ideas” (Hahn et al.
2012)
Usage fre-
quency
How often the freelancer uses WorkSMN on average
10 answer categories ranging
from “less than once a month (1)”
to “all the time (10)”
Other income
Refers to whether the freelancer has earnings apart from his/her
freelance work, such as a part-time employment
No other income = 0; Other in-
come = 1
Weekly hours
Weekly hours working as a freelancer
Number of weekly working hours
Freelance area
Area/field in which the freelancer primarily works
Several categories of answers re-
flecting different professional
fields (e.g., graphic design; art-
ists/photographers)
Strategic net-
working be-
havior
Assesses whether freelancers are rationally motivated to proac-
tively and purposely befriend other users who might, in some
way, be or become relevant for their future career (Utz 2016)
Four-items, five-point Likert
scale, e.g., “I send contact re-
quests to a great number of peo-
ple, in order to get a large net-
work” (Utz and Breuer 2016)
71
Table 13. Covariate summary and balance check after attrition
Covariate
Treatment
(assigned)
Mean (SD) or %
Difference
Test
Statistic
p-Value
Cronbach’s
α
Digital footprint data
Gender (Female = 1;
Male = 2)
0 (n = 79)
50.633%
-16.034%
3.2493
0.071
1 (n = 51)
66.667%
Social Capital (Number
of direct contacts)
0
156.696
(150.184)
-20.970
-0.643
0.522
1
177.667
(199.313)
Number of Groups
0
7.418 (9.401)
1.222
0.868
0.387
1
6.196 (6.627)
SMN Tenure (in years)
0
6.975 (3.876)
-0.535
-0.776
0.440
1
7.510 (3.818)
Number of Haves
0
13.519 (12.047)
0.754
0.395
0.694
1
12.765 (9.622)
Number of Wants
0
5.127 (6.587)
1.264
1.347
0.180
1
3.863 (4.109)
Follow-up questionnaire
Strategic Networking
Behavior
0
2.316 (0.882)
-0.105
-0.629
0.531
0.66
1
2.421 (0.959)
Age
0
44.329 (11.299)
-1.573
-0.808
0.421
1
45.902 (10.530)
Level of Education:
Middle
0
15.189%
-4.419%
0.430
0.512
1
19.608%
High
0
84.810%
4.418%
1
80.392%
Openness
0
5.418 (1.183)
-0.196
-1.001
0.319
0.65
1
5.614 (1.031)
Other Income (0 = No;
1 = Yes)
0
30.380%
-6.875%
0.662
0.416
1
37.255%
Weekly Hours
0
40.241 (16.548)
4.868
1.696
0.093
1
35.373 (15.597)
Usage Frequency: Less
often than once a month
0
6.329%
0.447%
8.890
0.180
1
5.882%
Once a month
0
24.051%
16.207%
1
7.843%
Several times a month
0
30.380%
-10.797%
1
41.176%
Once a week
0
13.924%
6.081%
1
7.843%
Several times a week
0
13.924%
-1.762%
1
15.686%
Once a day
0
7.595%
-8.091%
1
15.686%
Several times a day
0
3.797%
-2.085%
1
5.882%
Notes: Test statistic reflect Welch’s t-test for continuous variables and Pearson’s Chi-squared test for discrete variables
72
C.5 First-Stage Conditional Underidentification and Weak Identification Tests
The measures presented (Table 2 in the main text) allow us to verify the overall strength of our in-
strumental variables (IVs). Nonetheless, we can also separately evaluate each of our instruments in
Model E2 using the Sanderson and Windmeijer (2016) first-stage under-identification and weak iden-
tification tests. We assess this statistic over the standard first-stage F-test of excluded instruments
because the Sanderson and Windmeijer (2016) metric is required in cases where the model has more
than one endogenous regressor, which is the case in Model E2. Both conditional tests confirm that
each of our IVs is relevant and strong, also if we examine them separately (see Table 14 below).
Table 14. First-stage estimates (second-stage regression displayed in Table 2 in the main text)
Model E1.
First-stage re-
gression esti-
mates: Treatment
received
Model E2.
First-stage regression
estimates: Treatment
received
Model E2.
First-stage regression:
Treatment received *
Strategic network be-
havior
Treatment assigned (IV)
0.891 (0.041)***
0.841 (0.107)***
-0.214 (0.235)
Treatment assigned * Strategic networking be-
havior (IV)
0.021 (0.041)
0.990 (0.090)***
Male (Ref.: female)
-0.054 (0.045)
-0.053 (0.045)
-0.133 (0.099)
High level of education (Ref.: middle)
0.044 (0.057)
0.047 (0.057)
0.069 (0.125)
Age (in years)
-0.000 (0.002)
-0.000 (0.002)
-0.000 (0.005)
Openness
0.022 (0.019)
0.020 (0.019)
0.048 (0.042)
SMN tenure (in years)
-0.000 (0.005)
-0.000 (0.005)
0.006 (0.012)
Strategic networking behavior
-0.021 (0.023)
-0.031 (0.030)
-0.027 (0.067)
Work hours per week
0.001 (0.001)
0.001 (0.001)
0.002 (0.003)
Other income (Ref.: no other income)
0.037 (0.042)
0.035 (0.042)
0.123 (0.092)
Usage frequency (once a month) (Ref.: less than
once a month)
-0.029 (0.089)
-0.033 (0.089)
-0.033 (0.195)
Usage frequency (several times a month)
-0.005 (0.086)
-0.004 (0.087)
0.031 (0.189)
Usage frequency (once a week)
-0.085 (0.093)
-0.085 (0.093)
-0.206 (0.204)
Usage frequency (several times a week)
0.066 (0.096)
0.067 (0.096)
0.115 (0.210)
Usage frequency (once a day)
0.170 (0.096)
0.168 (0.096)
0.423 (0.210)*
Usage frequency (several times a day)
0.011 (0.115)
0.009 (0.116)
0.034 (0.253)
F-test of excluded instruments
473.08***
234.84***
347.99***
Underidentification test (Sanderson and Wind-
meijer 2016, first-stage chi-squared test)
640.62***
698.56***
1080.36***
Weak identification test (Sanderson and Wind-
meijer 2016, multivariate F-test of excluded in-
struments)
473.08***
510.49***
789.49***
Notes: Standard errors in parentheses; unstandardized regression estimates. Constant not reported. In our model specifi-
cation, we partialled out the variable freelance field from all the other variables in our estimation using the Frisch-
Waugh-Lovell theorem. Significance levels: * p < 0.05, ** p < 0.01, *** p < 0.001
73
C.6 Intention-to-Treat (ITT) Parameter
Table 15. ITT estimates of the effect of treatment assigned on SC accumulation
Model E1.
Social capital
accumulation”
Model E2.
Social capital
accumulation”
(including inter-
action term)
Treatment assigned
0.267 (2.205)
-10.123 (5.664)
Strategic networking behavior
3.300 (1.242)**
1.229 (1.608)
Treatment assigned * Strategic networking behavior
4.337 (2.183)*
Male (Ref.: female)
-4.120 (2.420)
-3.911 (2.387)
High level of education (Ref.: middle)
-4.710 (3.051)
-3.999 (3.027)
Age (in years)
-0.219 (0.112)
-0.232 (0.111)*
Openness
-1.144 (1.016)
-1.435 (1.012)
SMN tenure (in years)
-1.195 (0.287)***
-1.213 (0.283)***
Work hours per week
0.031 (0.067)
0.040 (0.066)
Other income (Ref.: no other income)
0.818 (2.241)
0.502 (2.213)
Usage frequency (once a month)(Ref.: less often than once a
month)
-1.505 (4.763)
-2.227 (4.706)
Usage frequency (several times a month)
1.010 (4.642)
1.234 (4.575)
Usage frequency (once a week)
0.565 (5.010)
0.527 (4.935)
Usage frequency (several times a week)
-1.062 (5.148)
-0.823 (5.073)
Usage frequency (once a day)
2.549 (5.154)
2.195 (5.081)
Usage frequency (several times a day)
9.441 (6.207)
9.091 (6.117)
F
3.447
3.475
R2
0.350
0.369
RMSE
10.345
10.192
n
130
130
Notes: Standard errors in parentheses; unstandardized regression estimates. Constant not reported. In our model specifi-
cation, we partialled out the variable freelance field from all the other variables in our estimation using the Frisch-
Waugh-Lovell Theorem. Significance levels: * p < 0.05, ** p < 0.01, *** p < 0.001
C.7 Discussion of Post-Treatment Bias for Variable Strategic Networking Behavior
Given the central role of the agency variable “strategic networking behavior” in our study, we need
to evaluate and discuss whether this variable suffers from a potential post-treatment bias. As ex-
plained in the main text subsection “Accumulating Social Capital: Causal Evidence (RQ1) - Data”,
we measured this key variable as part of our follow-up questionnaire. Thus, a potential concern is
that the treatment might have influenced the strategic networking behavior of freelancers, and the
inclusion of this post-treatment variable in our statistical model could have caused biased causal es-
timates (Montgomery et al. 2018). However, as outlined in the following paragraphs via statistical
and theoretical arguments, we are confident that our self-reported variable “strategic networking be-
havior” is not contingent upon the treatment.
74
As a statistical check regarding the presence of this issue, we ran a Welch’s t-test. As shown in
Table 13 in section C.4, there are no statistically significant differences between the treatment and
control groups regarding our agency variable “strategic networking behavior”. The well-balanced
groups suggest that the treatment did not cause the treated freelancers to exhibit greater strategic
networking behavior. We also find no statistically significant differences between the two groups on
the other post-treatment covariates that were collected (p > 0.05) (see also Table 13 in section C.4).
After ruling out group-differences, theoretical arguments provided further support that the moti-
vation to engage in a strategic networking behavior was not a consequence of the treatment. Like-
wise, Gerber and Green (2012, p. 97) highlight that “[g]ranted, characteristics such as sex or race
may be so stable that we may safely assume that they are unaffected by the intervention even if they
are measured afterwards. But this is a substantive assumption that will need to be evaluated on a
case-by-case-basis.” While we can equate the behavior of strategic networking with a personality
trait (see Utz and Breuer 2016), it is also plausible to assume that such intrinsic characteristics are
more or less fixed over time (Gertler et al. 2016). For instance, several studies (e.g., Cobb-Clark and
Schurer 2012; Gustavsson et al. 1997) demonstrate the stability of personality traits over the years.
Anecdotal evidence obtained from our study’s follow-up questionnaire free-text answers further
corroborates this conjecture
10
. For instance, one freelancer mentioned: “I think that nowadays net-
working is an important part of the career management, yet for me personally it is difficult”. In a
similar fashion and more explicitly, another participant stated: “Eventually I should/could use the
platform more offensive, but it is a question of character und I tend less to force my person on oth-
ers, because I do not like it the other way around either”. Against this backdrop, we can reasonably
assume that, within the six months’ timeframe in which we conducted our experiment (i.e., between
10
The original statements are in German. The authors translated them into English.
75
the manipulation and the follow-up questionnaire), it is quite unlikely that the motivation to engage
in a strategic networking behavior was susceptible to change due to our issued treatment.
D. FURTHER INFORMATION ON THE ANALYSES OF THE INDIVIDUAL-LEVEL
PANAL DATA (RQ2)
D.1 Robustness of Percentage Change in Social Capital (SC) Accumulation
We expect that freelancers who pay for the premium membership will generally have a high agency
for networking and affordances provided by WorkSMN. Hence, in percentage terms, the magnitude
in SC accumulation is disproportionally higher for freelancers with just a few contacts on
WorkSMN. To address this potential distortion (in the form of outliers), we conduct a sensitivity
analysis when repeating the OLS regression with percentage values for “Δ Social capital accumula-
tion” as the dependent variable, i.e., “Δ % Social capital accumulation”. More specifically, we fol-
low the example of Bapna and Umyarov (2015) and apply a stepwise exclusion approach by percen-
tile regarding the freelancers’ SC before the discount mailing. For each step, we assess the estimated
effect size of the interaction term of “premium conversion * contact invites sent before”. The results
in Figure 5 show that only after removing freelancers in the lowest 10-percentile, the estimates sta-
bilize around a positive interaction of 0.012 - 0.014.
76
Figure 5. Interaction effect size estimates on „Δ % Social capital accumulation“
D.2 Descriptive Statistics and Bivariate Correlations
Table 16. Correlations and descriptive statistics
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[1] SC accumulationit
[2] Profile visits (Log)it
0.273
[3] Messages sent (Log)it
0.266
0.508
[4] Profile visitors rec. (Log)it
0.241
0.404
0.409
[5] Messages rec. (Log)it
0.254
0.423
0.691
0.468
[6] Genderi
-0.019
-0.089
-0.036
-0.031
-0.042
[7] Educationi
0.011
0.042
0.064
0.082
0.067
0.022
[8] Agei
-0.046
-0.097
-0.020
-0.157
-0.081
-0.071
-0.021
[9] SMN tenurei
-0.033
0.011
0.012
0.039
0.010
-0.021
0.186
0.133
[10] Job seekingi
0.022
0.037
0.039
0.060
0.043
0.029
0.015
-0.085
-0.148
Mean
1.800
1.005
0.441
2.061
0.537
1.250
0.510
43.032
8.051
0.370
SD
8.086
1.095
0.702
0.930
0.738
0.433
0.500
9.684
2.442
0.483
Notes: Significance level p<0.05, as indicated by bold numbers. Observations over freelancer i and time t, as in Figure 2
in the main text.
77
D.3 Covariate Definitions
Table 17. Overview of variables and their definitions
Variable
Definition
Measure
Individual-level panel data
Profile visits
Number of times another profile of a WorkSMN user was opened
by the freelancer (i.e., activity)
Number of profile visits in a
given month
Messages sent
Number of messages sent by the freelancer via the messaging
function of WorkSMN (i.e., activity)
Number of messages sent in a
given month
Profile visi-
tors rec.
Number of times another WorkSMN user opened the profile of
the freelancer (i.e., saliency)
Number of profile visitors re-
ceived in a given month
Messages re-
ceived
Number of messages received by the freelancer via the messaging
function of WorkSMN (i.e., saliency)
Number of messages received in a
given month
Contact in-
vites sent be-
fore
Number of contact invites sent in the month before the campaign,
which serves as an objective operationalization of “strategic net-
working behavior”, in particular: “I send contact requests to a
great number of people, in order to get a large network.(Utz and
Breuer 2016)
Number of contact invites in the
month before the campaign
Logins before
Number of logins in the month before the campaign, where each
visit counts as a login
Number of logins in the month
before the campaign
Education
title
Whether freelancer had any title indicated in his profile in a free-
text field
Yes = 1; No = 2
Level
Freelancer’s career level, as identified by WorkSMN (derived via
years of experience and job title)
Entry-Level = 1, Management =
2, Senior = 3
Job-seeking
Whether the person has indicated to be job-seeking in a drop-
down field, which can be seen on the profile and by recruiters
1 = active, 0 = not active
Notes: Social Capital, Age, Gender, and SMN tenure as described in Section C.4
D.4 Panel Vector Autoregressive Model (PVAR)
Vector autoregressive (VAR) modeling allows us to assess the coevolution of multiple variables as
interdependent system. In other words, VAR models do not impose a priori assumptions about the
directionality of effects among the variables entered. Panel VAR, or PVAR, essentially is a general-
ization of this approach to panel data sets that contain time-series information on sets of heterogene-
ous units, i.e., freelancers in our setting. The nascent modeling approach has found recent applica-
tion in IS literature (e.g., Chen et al. 2015; Dewan and Ramaprasad 2014).
Because regular OLS estimation of PVAR models yields biased coefficients, most applications
are estimated via the generalized method of moments (GMM). GMM uses transformed observations
78
as instruments for the lagged dependent variables (Hansen 1982). An advantage of PVAR is that it
captures unobserved differences of the microunits.
We assess the variables associated with SC accumulation, activity, and saliency. We take
monthly differences in SC accumulation and take natural logarithms to eradicate the positive trend
and skewness. Monthly observations surrounding the discounting campaigns amount to 16 observa-
tions. In the first part of our data, we report Harris-Tzavalis tests to assess for the stationarity of the
data, reported in Table 18. Harris-Tzavalis is particularly suitable for a short period, large cross-unit
panel datasets, as in our case. All variables in the panel data set are stationary.
Table 18. Harris-Tzavalis test
Variable
Statistic
p-Value
Profile visits
0.191
p < 0.001
Messages sent
0.136
p < 0.001
Profile visitors rec.
0.139
p < 0.001
Messages received
0.123
p < 0.001
SC accumulation
0.028
p < 0.001
Specification
We specify a system GMM model with forward orthogonal deviations (FOD). FOD minimizes data
loss, as it takes averaged future observations as instruments. For the sake of computational effi-
ciency, we down-sample the 47,341 non-converters in equal proportion to converters and arrive at
1,430 observations. The presented results have been confirmed with multiple random seeds. We
look at the Andrews and Lu (2001) statistics to assess the optimal lag length and choose 3 lags, re-
flecting a full quarter––a reasonable duration for interdependencies between the variables to occur.
As shown in Table 19, MAIC, MBIC, and MQIC values concur that a 3-lag specification is optimal.
Table 19. MAIC, MBIC, and MQIC (Andrews and Lu 2001)
Lags
MBIC
MAIC
MQIC
1
-3,603
326
-1,074
2
-3,673
179
-1,198
3
-3,952
-115
-1,492
79
D.5 Propensity Score Matching
Table 20. Logistic regression of propensity to convert to premium
Premium conversion
Contact invites sent before (Log)
0.045 (0.033)
Male (Ref.: female)
-0.036 (0.097)
Level of education (Ref.: No acad. title)
0.087 (0.083)
Age
0.001 (0.004)
SMN tenure
-0.004 (0.017)
Career segment: Manager (Ref.: Entry)
0.004 (0.124)
Career segment: Senior Executive
0.168 (0.092)
Job seeking: Active
0.434 (0.082)
***
Logins before (Log)
0.098 (0.042)
*
Messages sent before (Log)
0.013 (0.074)
Profile visits before (Log)
0.130 (0.045)
**
Messages rec. before (Log)
-0.037 (0.071)
Profile visitors rec. before (Log)
0.101 (0.058)
AIC
6,709.561
Nagelkerke R2
0.017
n
52,392
Notes: Constant not reported. Significance levels: * p < 0.05, ** p < 0.01, *** p < 0.001
Following the propensity score matching (PSM), we evaluate the match via standardized mean dif-
ferences (SMD). Given the specification with ratio 1, the focal variables SMD reduce from 0.232 to
0.013 for Job seeking, from 0.180 to 0.041 for Logins before (Log), and from 0.275 to 0.005 for
Profile Visits before (Log). All matching variables in the matched groups appear below the 0.1
threshold to conclude that matching was successful. The distributions of the propensity scores after
matching appear close to identical. We obtain similar results for matching with ratio 4.
Figure 6. Distribution of propensity scores
80
REFERENCES
Andrews, D. W., and Lu, B. 2001. "Consistent Model and Moment Selection Procedures for Gmm
Estimation with Application to Dynamic Panel Data Models," Journal of Econometrics
(101:1), pp. 123-164.
Angrist, J. D., and Pischke, J.-S. 2008. Mostly Harmless Econometrics: An Empiricist's Companion.
Princeton university press.
Aten, K., DiRenzo, M., and Shatnawi, D. 2017. "Gender and Professional E-Networks: Implications
of Gender Heterophily on Job Search Facilitation and Outcomes," Computers in Human
Behavior (72), pp. 470-478.
Baiocchi, M., Cheng, J., and Small, D. S. 2014. "Instrumental Variable Methods for Causal
Inference," Statistics in medicine (33:13), pp. 2297-2340.
Banerji, D., and Reimer, T. 2019. "Startup Founders and Their Linkedin Connections: Are Well-
Connected Entrepreneurs More Successful?," Computers in Human Behavior (90), pp. 46-
52.
Bapna, R., and Umyarov, A. 2015. "Do Your Online Friends Make You Pay? A Randomized Field
Experiment on Peer Influence in Online Social Networks," Management Science (61:8), pp.
1902-1920.
Baruffaldi, S. H., Di Maio, G., and Landoni, P. 2017. "Determinants of Phd Holders’ Use of Social
Networking Sites: An Analysis Based on Linkedin," Research policy (46:4), pp. 740-750.
Buettner, R. 2017. "Getting a Job Via Career-Oriented Social Networking Markets," Electronic
Markets (27:4), pp. 371-385.
Burke, M., and Kraut, R. 2013. "Using Facebook after Losing a Job: Differential Benefits of Strong
and Weak Ties," Proceedings of the 2013 conference on Computer supported cooperative
work: Association for Computing Machinery, pp. 1419-1430.
Burt, R. S. 2009. Structural Holes: The Social Structure of Competition. Harvard university press.
Chen, H., De, P., and Hu, Y. J. 2015. "It-Enabled Broadcasting in Social Media: An Empirical Study
of Artists’ Activities and Music Sales," Information Systems Research (26:3), pp. 513-531.
Cobb-Clark, D. A., and Schurer, S. 2012. "The Stability of Big-Five Personality Traits," Economics
Letters (115:1), pp. 11-15.
Costa Jr, P. T., and McCrae, R. R. 2014. "The Neo Inventories," in Personality Assessment. 2nd Ed.
New York, NY, US: Routledge/Taylor & Francis Group, pp. 229-260.
Davis, J., Wolff, H.-G., Forret, M. L., and Sullivan, S. E. 2020. "Networking Via Linkedin: An
Examination of Usage and Career Benefits," Journal of Vocational Behavior (118), p.
103396.
Dewan, S., and Ramaprasad, J. 2014. "Social Media, Traditional Media, and Music Sales," MIS
Quarterly (38:1), pp. 101-122.
Eckles, D., Karrer, B., and Ugander, J. 2017. "Design and Analysis of Experiments in Networks:
Reducing Bias from Interference," Journal of Causal Inference (5:1), pp. 1-23.
Garg, R., and Telang, R. 2018. "To Be or Not to Be Linked: Online Social Networks and Job Search
by Unemployed Workforce," Management Science (64:8), pp. 3926-3941.
Gee, L. K., Jones, J., and Burke, M. 2017. "Social Networks and Labor Markets: How Strong Ties
Relate to Job Finding on Facebook’s Social Network," Journal of Labor Economics (35:2),
pp. 485-518.
Gerber, A. S., and Green, D. P. 2012. Field Experiments: Design, Analysis, and Interpretation. W.
W. Norton & Company.
81
Gertler, P. J., Martinez, S., Premand, P., Rawlings, L. B., and Vermeersch, C. M. 2016. Impact
Evaluation in Practice, (Second Edition ed.). The World Bank.
Gloor, P. A., Woerner, S., Schoder, D., Fischbach, K., and Colladon, A. F. 2018. "Size Does Not
Matter-in the Virtual World. Comparing Online Social Networking Behaviour with Business
Success of Entrepreneurs," International Journal of Entrepreneurial Venturing (10:4), pp.
435-455.
Granovetter, M. S. 1973. "The Strength of Weak Ties," American Journal of Sociology (78:6), pp.
1360-1380.
Gustavsson, J. P., Weinryb, R. M., Göransson, S., Pedersen, N. L., and Åsberg, M. 1997. "Stability
and Predictive Ability of Personality Traits across 9 Years," Personality and individual
differences (22:6), pp. 783-791.
Hahn, E., Gottschling, J., and Spinath, F. M. 2012. "Short Measurements of Personality–Validity
and Reliability of the Gsoep Big Five Inventory (Bfi-S)," Journal of Research in Personality
(46:3), pp. 355-359.
Hansen, L. P. 1982. "Large Sample Properties of Generalized Method of Moments Estimators,"
Econometrica (50:4), pp. 1029-1054.
Hinz, O., Spann, M., and Hann, I.-H. 2015. "Research Note—Can’t Buy Me Love… or Can I?
Social Capital Attainment through Conspicuous Consumption in Virtual Environments,"
Information Systems Research (26:4), pp. 859-870.
Imbens, G. W., and Rubin, D. B. 2015. Causal Inference in Statistics, Social, and Biomedical
Sciences. Cambridge University Press.
Lousdal, M. L. 2018. "An Introduction to Instrumental Variable Assumptions, Validation and
Estimation," Emerging themes in epidemiology (15:1), pp. 1-7.
Ma, S. Q., and Leung, L. 2019. "The Impacts of Personality Traits, Use Intensity and Features Use
of Linkedin on Bridging Social Capital," Applied Research in Quality of Life (14:4), pp.
1059-1078.
Mashayekhi, M., and Head, M. 2018. "Taking Advantage of Using Professionally-Oriented Social
Network Sites: The Role of Users' Actions and Profiles," in: AMCIS 2018. New Orleans,
LA, USA.
Mayer, A. 2012. "The Structure of Social Networks and Labour Market Success," Applied
Economics Letters (19:13), pp. 1271-1274.
Montgomery, J. M., Nyhan, B., and Torres, M. 2018. "How Conditioning on Posttreatment Variables
Can Ruin Your Experiment and What to Do About It," American Journal of Political Science
(62:3), pp. 760-775.
Nikitkov, A., and Sainty, B. 2014. "The Role of Social Media in Influencing Career Success,"
International Journal of Accounting & Information Management (22:4), pp. 273-294.
Oestreicher-Singer, G., and Zalmanson, L. 2013. "Content or Community? A Digital Business
Strategy for Content Providers in the Social Age," MIS Quarterly (37:2), pp. 591-616.
Rienties, B., Tempelaar, D., Pinckaers, M., Giesbers, B., and Lichel, L. 2012. "The Diverging
Effects of Social Network Sites on Receiving Job Information for Students and
Professionals," in Technological Change and Societal Growth: Analyzing the Future.
Hershey, PA, USA: IGI Global, pp. 202-217.
Sagarin, B. J., West, S. G., Ratnikov, A., Homan, W. K., Ritchie, T. D., and Hansen, E. J. 2014.
"Treatment Noncompliance in Randomized Experiments: Statistical Approaches and Design
Issues," Psychological Methods (19:3), pp. 317-333.
Sanderson, E., and Windmeijer, F. 2016. "A Weak Instrument F-Test in Linear Iv Models with
Multiple Endogenous Variables," Journal of Econometrics (190:2), pp. 212-221.
82
Skeels, M. M., and Grudin, J. 2009. "When Social Networks Cross Boundaries: A Case Study of
Workplace Use of Facebook and Linkedin," Proceedings of the ACM 2009 international
conference on Supporting group work, pp. 95-104.
Statistisches Bundesamt. 2016. "Statistik Und Wissenschaft. Demographische Standards 2016,"
Statistisches Bundesamt, Wiesbaden, Germany.
Sun, C., Shi, Z. J., Liu, X., Ghose, A., Li, X., and Xiong, F. 2019a. "The Effect of Voice Ai on
Consumer Purchase and Search Behavior," SSRN (3480877) (Working Paper).
Sun, T., Shi, L., Viswanathan, S., and Zheleva, E. 2019b. "Motivating Effective Mobile App
Adoptions: Evidence from a Large-Scale Randomized Field Experiment," Information
Systems Research (30:2), pp. 523-539.
Utz, S. 2016. "Is Linkedin Making You More Successful? The Informational Benefits Derived from
Public Social Media," New Media & Society (18:11), pp. 2685-2702.
Utz, S., and Breuer, J. 2016. "Informational Benefits from Social Media Use for Professional
Purposes: Results from a Longitudinal Study," Cyberpsychology: Journal of Psychosocial
Research on Cyberspace (10:4).
Utz, S., and Breuer, J. 2019. "The Relationship between Networking, Linkedin Use, and Retrieving
Informational Benefits," Cyberpsychology, Behavior, and Social Networking (22:3), pp. 180-
185.
Walker, D., and Muchnik, L. 2014. "Design of Randomized Experiments in Networks," Proceedings
of the IEEE (102:12), pp. 1940-1951.
Williams, D. 2006. "On and Off The’net: Scales for Social Capital in an Online Era," Journal of
computer-mediated communication (11:2), pp. 593-628.
Zhitomirsky-Geffet, M., and Bratspiess, Y. 2015. "Perceived Effectiveness of Social Networks for
Job Search," Libri (65:2), pp. 105-118.
ResearchGate has not been able to resolve any citations for this publication.
Article
Full-text available
Friendship formation is of central importance to online social network sites and to society, but can suffer from significant and unequal frictions. In this study, we demonstrate that social networks and policy makers may use an IT-facilitated intervention – displaying things in common (TIC) between users (mutual hometown, interest, education, work, city) – to encourage friendship formation, especially among people who are different from each other. Displaying TIC may update an individual’s belief about the shared similarity with another and reduce information friction that may be hard to overcome in offline communication. In collaboration with an online social network, we design and implement a randomized field experiment, which randomly varies the prominence of different types of things in common information when a user (viewer) is browsing a non-friend’s profile. The dyad-level exogenous variation, orthogonal to any (un)observed structural factors in viewer-profile’s network, allows us to cleanly isolate the role of individuals’ preference for TIC in driving network formation and homophily. We find that displaying TIC to viewers may significantly increase their probability of sending a friend request and forming a friendship, and is especially effective for pairs of people who have little in common. Such findings suggest that information intervention is a very effective and zero-cost approach to encourage the formation of weak ties, and also provide the first experimental evidence on the crucial role of individuals’ preference (versus structural embeddedness) in network formation. We further demonstrate that displaying TIC could improve friendship formation for a wide range of viewers with different demographics and friendship status, and is more effective when the TIC information is more surprising to the viewer. Our study offers actionable insights to social networks and policy makers on the design of information intervention to encourage friendship formation and improve the diversity of the friendship, at both an aggregate and an individual level.
Article
Full-text available
Even though networking has been identified as an important career self-management behavior, research to date has been limited to traditional channels of communication. With the advent of social media, networking opportunities have expanded to a new set of technology-based methods. However, very little is known about the benefits of using such social networking platforms, especially LinkedIn, which was designed for professional purposes. Integrating research on networking and careers with research on social networking sites, we introduce a model relating extraversion and protean career orientation via networking ability to social networking site usage (i.e., number of contacts and frequency of usage) and a variety of career benefits (e.g., career sponsorship). Using a mixed-mode survey, data were collected from 322 working professionals enrolled in business-related graduate programs, of which 133 used LinkedIn as their predominant professional networking tool. In addition, a subset of the sample using LinkedIn granted us permission to access their LinkedIn profile, which we coded for nine characteristics. In general, data from the survey responses as well as the LinkedIn profiles supported the proposed model. Results also showed that it was not the number of contacts one had, but the frequency of usage, that mattered for the receipt of career benefits.
Article
Full-text available
Previous research has shown that users of social network sites designed for professional purposes, such as LinkedIn, report higher professional informational benefits than nonusers. However, this effect could only be partly explained by social media use as there was also a selection effect, such that people who have more informational benefits were more likely to use LinkedIn. The goal of this study was to explore whether differences in networking, defined as a set of behaviors with the aim of building, maintaining, and using internal and external contacts for instrumental purposes, can explain this selection effect. We used data from a panel study with a representative sample of Dutch Internet users (n = 685; 259 LinkedIn users) to examine the relationships between networking and LinkedIn use as well as professional informational benefits, that is, timely access to relevant information. The results showed that people scoring high on external networking (but not internal networking within their organization) are also more likely to use LinkedIn. External networking was also positively correlated with active and passive use as well as the number of strong and latent ties on LinkedIn. However, in a mediation model the direct effect of networking on informational benefits was not mediated by actual social media use and network composition; instead, the number of weak ties had a direct effect on informational benefits. The results thus indicate that networking is a major driver of informational benefits from LinkedIn use. © Sonja Utz and Johannes Breuer 2018; Published by Mary Ann Liebert, Inc. 2018.
Article
Full-text available
Although experiments are the gold standard for establishing causality, several threats can undermine the internal validity of experimental findings. In this article, we first discuss these threats, which include the lack of consequential decisions and outcomes, deception, demand effects and unfair comparisons, as well as issues concerning statistical validity (e.g., minimum sample size per cell, estimating variance correctly). We expose each problem, show potential solutions, and bring to the fore issues of relevance of the findings (i.e., external and ecological validity). Thereafter, we take stock of the state-of-the-science regarding validity threats using a representative sample of 468 recent experiments from 258 articles published in top-tier journals. We compare research practices in three fields of study-management, social psychology, and economics, which regularly use experimental research-to operations management, which has recently begun to use the experimental paradigm. Our results underscore the importance for journals and authors to follow what we identify to be best-practice methodological suggestions (i.e., the "ten commandments" of experimental research). We show that-on average-markers of methodological rigor and generalizability positively and significantly predict the citations received by published articles. Finally, given that experiments are infeasible in some settings, we conclude with a brief review of often overlooked quasi-experimental designs, which are useful for generating strong counterfactuals and hence allow making causal claims in the field.
Article
Professional networks are vital for individuals’ career advancement. Research demonstrates, however, that women are often disadvantaged in their access to such networks. Using a randomized field experiment at an IT conference, we found that women had worse networking outcomes than men. Relative to men, women met 42% fewer new contacts, spent 48% less time talking to them, and added 25% fewer LinkedIn connections. We theorize that in fields where women are underrepresented (e.g., IT) two networking barriers—search and social—differentially affect men and women. We designed and experimentally tested interventions for reducing these barriers. The search intervention was designed to facilitate locating diverse contacts and information. The social intervention was designed to facilitate helping behavior and connecting across social boundaries. Our findings indicate that the search intervention increased the number of new contacts women met by 57%, the time they spent talking with them by 90%, the number of LinkedIn connections they added by 29%, and their odds of changing jobs by a factor of 1.6. The social intervention also increased the time women spent talking to new contacts by 66%. However, the interventions did not improve men’s outcomes. Our results show that simple interventions can help women expand their networks and find jobs.
Article
The proliferation of the Internet has enabled platform intermediaries to create two-sided markets in many industries. Time-series data on the number of customers on both sides of the markets allow platform intermediaries for estimating the direction and magnitude of network effects, which can then support growth predictions and subsequent information technology (IT) or marketing investment decisions. This article investigates the conditions under which this estimation of same-side and cross-side network effects should distinguish between its impact on the number of new customers (i.e., acquisition) and existing customers (i.e., their activity). The authors propose an influx-outflow model for doing so and conduct a simulation study to benchmark the new model against the traditional model. Further they compare the models in an illustrative empirical study in which they study the growth of an Internet auction platform. The results show that this separation of effects is beneficial because the existing customers on both sides of the market can influence the acquisition and dropout of other customers asymmetrically. The paper thus makes an important contribution that should impact the way how researchers and business practitioners measure network effects in two-sided markets.
Article
Using a small example as an illustration, this article reviews multivariate matching from the perspective of a working scientist who wishes to make effective use of available methods. The several goals of multivariate matching are discussed. Matching tools are reviewed, including propensity scores, covariate distances, fine balance, and related methods such as near-fine and refined balance, exact and near-exact matching, tactics addressing missing covariate values, the entire number, and checks of covariate balance. Matching structures are described, such as matching with a variable number of controls, full matching, subset matching and risk-set matching. Software packages in R are described. A brief review is given of the theory underlying propensity scores and the associated sensitivity analysis concerning an unobserved covariate omitted from the propensity score.
Article
It is often seen as axiomatic in the social capital literature that relationships require ongoing maintenance to remain valuable. As a result, nearly all social network research has only considered relationships that are active (or recently restored to activity after a period of dormancy). Seldom considered is the impact of still-dormant, unmaintained ties that remain dormant. Using two bounded-network studies (one in a school district, one in a private company), we find consistently that still-dormant ties increase employees' organizational commitment. However, we also find that active ties can "dilute" this effect, rendering dormant ties less relevant and meaningful to individuals, or act as "preservatives" keeping dormant ties relevant and valued. In a third, vignette-based study, we find two mechanisms for the effect of dormant ties on organizational commitment: perceptions of the past (fond recollections) and future (information access). Thus, we offer a more time-oriented, layered theory of social network ties that can be activated in people's minds even when not active in practice.
Article
Making sustainable profits from a baseline zero price and motivating free consumers to convert to premium subscribers is a continuing challenge for all freemium communities. Prior research has causally established that social engagement (Oestreicher-Singer and Zalmanson 2013) and peer influence (Bapna and Umyarov 2015) are two important drivers of users converting to premium subscribers in such communities. In this paper, we flip the perspective of prior research and ask whether the decision to pay for a premium subscription causes users to become more socially engaged. In the context of the Last.fm music listening freemium social community, we establish, using a novel 41-month-long panel dataset, a look-ahead propensity score matching (LA-PSM) procedure coupled with a difference-in-difference estimator of the treatment effect, that payment for premium leads to more social engagement. Specifically, we find that paying for premium leads to an increase in both content-related and community-related social engagement. Free users who convert to premium listen to 287.2% more songs, create 1.92% more playlists, exhibit a 2.01% increase in the number of forum posts made, and gain 15.77% more friends. Thus, premium subscribers create value not only for themselves by consuming more content, but also for the community and site by organizing more content and adding more friends, who are subsequently engaged by the social diffusion emerging from the focal user's activities.