AUGUST 2014 VOLUME 61 NUMBER 3 IEEMA4 (ISSN 0018

advertisement

AUGUST 2014 VOLUME 61 NUMBER 3 IEEMA4 (ISSN 0018-9391)

EDITORIAL

Editorial for August 2014 Issue . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . R. Sabherwal 390

RESEARCH ARTICLES

Spotting Lemons in Platform Markets: A Conjoint Experiment on Signaling . . . . . . . . . . . . . . . . . A. Tiwana and A. A. Bush 393

Energy Technological Change and Capacity Under Uncertainty in Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . E. Shittu 406

Strategic Management of Cloud Computing Services: Focusing on Consumer Adoption Behavior . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . J. Shin, M. Jo, J. Lee, and D. Lee 419

Exploration, Exploitation, and Growth Through New Product Development: The Moderating Effects of Firm Age and

Environmental Adversity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Y. R. Choi and P. H. Phan 428

Contracting a Development Supplier in the Face of a Cost-Competitive Second Source of Supply . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A. Mensendiek and R. Mitri 438

Two-Phase Differential Evolution for the Multiobjective Optimization of Time–Cost Tradeoffs in Resource-Constrained

Construction Projects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . M.-Y. Cheng and D.-H. Tran 450

The Relationship Between Innovation and Diversification in the Case of New Ventures: Unidirectional or Bidirectional? .

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . I. Deligianni, I. Voudouris, and S. Lioukas 462

Energy Efficiency Benefits: Is Technophilic Optimism Justified? . . . . . . . . . . . . . . . . . . R. Nishant, T. S. H. Teo, and M. Goh 476

Openness and Appropriation: Empirical Evidence From Australian Businesses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . F. Huang, J. Rice, P. Galvin, and N. Martin 488

Innovation and Performance: The Role of Environmental Dynamism on the Success of Innovation Choices . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A. P´erez-Lu˜no, S. Gopalakrishnan, and R. V. Cabrera 499

Is an Efficacious Operation a Safe Operation: The Role of Operational Practices in Worker Safety Outcomes . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . M. Pagell, C. Dibrell, A. Veltri, and E. Maxwell 511

(Contents Continued on Page 389)

IEEE TRANSACTIONS ON ENGINEERING MANAGEMENT, VOL. 61, NO. 3, AUGUST 2014

Spotting Lemons in Platform Markets: A Conjoint

Experiment on Signaling

Amrit Tiwana and Ashley A. Bush

393

Abstract—This study addresses the understudied question of how content integrity for digital goods is signaled ex-ante in the absence of centralized oversight in self-organizing platforms. We build on signaling theory to theorize three classes of signals to explain how and why they influence platform user behavior. Experimental data from 380 users show that in the absence of centralized oversight in platforms a portfolio of signaling mechanisms is used to assess content integrity. Platform users differentially weigh platform, content, and contributor signals but simultaneously triangulate them to form holistic inferences about risk vis-`a-vis benefit to spot potential “lemons” in a platform market. Implications for practice, especially for platform design, are also discussed.

Index Terms—Conjoint experiment, digital platforms, digital platform design, self-organizing platforms, signals, signal classes, signaling theory.

I. I NTRODUCTION

S ELF-ORGANIZING platforms have emerged as a growing mechanism for distributing digitized content such as music, video, software, and IT services (hereafter digital goods). Unlike centralized approaches to content distribution, self-organizing platforms do not rely on a central intermediary to supply content, monitor, oversee transactions, or coordinate activities [13].

Users of such platforms are often both providers and consumers of the content [2]. Organizations use such platforms as a low cost, scalable mechanism to distribute content that would otherwise be potentially costly to distribute in a centralized manner.

Their use is growing in diverse industries such as publishing, film distribution, smartphones, gaming, and software services.

For example, BBC distributes unbroadcast footage of television programs [38]; Mozilla distributes software patches to users of its products. Such platforms are referred to as content delivery networks [17], peer-to-peer (P2P) content networks [2], or subsumed under the broader umbrella of digital platforms [41]. For simplicity, we refer to self-organizing content sharing platforms as self-organizing platforms.

A key problem that users of such platforms face is in assessing ex-ante the quality and integrity of content [2]. Digitized products are consumption goods or what economists describe as

Manuscript received March 8, 2012; revised January 14, 2013 and January

25, 2014; accepted March 5, 2014. Date of publication April 4, 2014; date of current version July 15, 2014. Review of this manuscript was arranged by Department Editor B. C. Y. Tan.

A. Tiwana is with the University of Georgia, Athens, GA 30606 USA (e-mail: tiwana@uga.edu).

A. A. Bush is with the College of Business, Florida State University,

Tallahassee, FL 32306-1110 USA (e-mail: abush@fsu.edu)

Digital Object Identifier 10.1109/TEM.2014.2311074

experience goods or credence goods [11] which means that their quality can only be assessed ex-post after they have been consumed but not ex-ante. Users of such platforms often find it difficult to differentiate high-integrity digital goods from malicious low-integrity digital goods masquerading as “good” ones (trojan/decoy content, spyware, “torrent poisoning,” or poor quality content). As Akerlof [1] would put it, the challenge for end-users is one of spotting “lemons” in such platform markets. Recent studies show that as many as 80% of users of such platforms experience content integrity problems [8]. Self-organizing platforms therefore face a challenge in ensuring content integrity [2], which can undermine their survival.

How do users of such platforms distinguish ex-ante between high integrity and rogue digital goods? This question has largely been neglected in information systems research on platforms, which has largely focused on the supply-side under two overarching themes: 1) their business models and economic impacts and 2) their design. The first set of studies have evaluated viable business and pricing models for platforms [5], their economic impact on music industry sales [22], and the strategic implications of product digitization [3]. The second set of studies has focused on their design, both from a technical architecture perspective and from a governance perspective. For example, how their architecture affects performance [19], [21], their governance [13], [45], and the interplay between architecture and governance [41]. This focus on the supply side of platforms has left open the important issue of how users on the consumption-

side of such platforms discern content integrity.

Research on the economics of information does focus on the consumption-side of platforms. Unfortunately this body of work assumes away the content integrity problem, focusing instead on a theoretically downstream notion of its value to a consumer.

Unlike Varian [44] who focuses on the value of an information good to a consumer, our emphasis is on the integrity of the digital good. For example, Varian discusses different mechanisms through which the seller of a digital good can credibly convince consumers of its value using strategies such as branding, crippled free versions, and sampling previews. However, they assume the digital good is what it claims to be, and in the process assume away the potential upstream problem of ensuring its integrity. Nevertheless, platform sustainability and vibrant functioning increasingly hinges on assuring end-users of this integrity. Failure to ensure integrity can be detrimental to how platform markets function, as malware proliferation on Google’s self-organizing Android marketplace illustrates [6]. Thus, unlike Shapiro and Varian [32] who are concerned with the utility that a consumer will derive from watching Finding Nemo, we are concerned if the digital copy offered to the consumer is

0018-9391 © 2014 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.

See http://www.ieee.org/publications standards/publications/rights/index.html for more information.

394 IEEE TRANSACTIONS ON ENGINEERING MANAGEMENT, VOL. 61, NO. 3, AUGUST 2014 indeed Finding Nemo. In summary, prior studies have not addressed the problem of how users of self-organizing platforms, in the absence of centralized control, differentiate low from high integrity digital goods. This study focuses on this gap, guided by the following research question: How is integrity of a digital good signaled among participants in self-organizing platforms?

We build on signaling theory to develop two ideas. First, users rely on a variety of signals to assess the integrity of a digital good. We introduce the notion of signal classes—contributor signals, content signals, and platform signals—that correspond to who, what, and where facets of the digital good. We show how users integrate these signals to formulate a holistic perception of the differential between risks and benefits, which in turn influences their transaction likelihood involving a digital good.

Second, properties of the platform, i.e., where signaling occurs alter whether who is signaling or what object the signal is about is more influential on signal recipients. We test the proposed ideas through a field study of 380 users of peer-to-peer, selforganizing platforms.

The study makes several novel contributions. Our overarching contribution is a theoretical explanation for how content integrity is communicated ex-ante through various classes of signals in self-organizing platforms. We show that users simultaneously rely on all three classes of signals conceptualized here—signals about what the digital good is, who contributed it, and its host platform—to formulate an integrative assessment of risk-benefit differential, which provides the explanatory link between signaling and their decision to transact. We also show that the signaling environment impacts what signals platform users attend to. Signals exhibit signal class substitution effects:

Contributor and content signals swap the dominant role depending on the signaling by the platform. The remainder of the paper progresses as follows. Next, we theoretically develop the hypotheses (see Section II), followed by the methodology and data collection (see Section III), analyses and results (see Section IV); and conclude with the contributions and implications of the study (see Section V).

II. T

A. Signaling Theory

HEORY AND H YPOTHESES

Signaling theory, the theoretical foundation for our study, is about how potential transactors in a competitive market overcome challenges of market exchange due to an information asymmetry between the transacting parties [10], [20], [29], [36].

Such information asymmetry can upset market exchange of goods or services whose quality cannot be ascertained ex-ante, and even lead to market demise [1], [11]. There are three elements to signaling theory—signalees, signalers, and signals.

Signalee is the recipient of a signal and signaler is the entity

(object, person, or organization) that the signal is about [10].

Spence [35, p. 2] defines signals as activities or attributes of individuals in a market which, by design or accident, convey information to alter the signalee’s subjective probabilistic beliefs about the signaler’s unobservable ability to fulfill her needs or demands. A signal is therefore an observable trait that can potentially convey information about an ex-ante unobservable trait. In Spence’s example, a signalee (a prospective employer) confronts a plethora of potential signals about a job applicant in the form of observable attributes of the signaler (e.g., education, employment record, gender), which the signalee must use to discern unobservable quality as a future employee. Spence’s signals are therefore verifiable at varying costs, as we also assume.

The crux of signaling theory is that the signaler transmits information (a signal) to the signalee to influence the signalee’s perception of the signaler’s quality (e.g., reliable versus unreliable) [35, p. 10]. The role of a signal is to help resolve the signalee’s classification problem in the presence of potential deception by the signaler. The challenge is that low-quality signalers might attempt to imitate signals that make them appear to be high quality [35, p. 38], implying that the signalee must consider the signal to be credible for it to influence her assessment of the signaler [10]. For a signal to be reliable in distinguishing the signaler’s type, it must be 1) observable and 2) costly for the low type to send but low cost for the high type to send [7], [36].

Signalees therefore will discount signals that can be produced at low cost since they do not reliably separate high from lowquality signalers.

In making sense of weak and ambiguous signals, humans rely on information about both the signal and its source to interpret them [30]. More broadly, signaling theory emphasizes the need for multiple signals for signal recipients to have greater confidence in the judgments that they make using them. Spence’s [34] early work also emphasized the need for a “sufficient number of signals,” arguing that a single signal is often insufficient to judge quality. Although the idea that individuals often use multiple signals is pervasive in contemporary signaling theory [10], [20], [27, p. 22], Spence did not distinguish between different types—or classes—of signals. Although signaling often involves use of many signals, few studies have considered more than one signal [12]. Signals can broadly be classified into three broad classes of signals—signals about who, what, and where. This typology of signals is theoretically grounded in signaling theory, which emphasizes the signaling information communicated by characteristics of the object of signaling, the observable characteristics of the sender of the signal, and the characteristics of the environment in which signals are sent and received [28]. Building on Spence’s [34] notion of multidimensional signals, each class of signals can have multiple, imperfectly correlated dimensions. Each signal within a class is therefore merely correlated with the unobservable desirable property of a digital good’s integrity, and signalees might combine multiple types of signals from different classes [20]. The greater the consistency among multiple signals, the more the signalee is likely to trust that the overall signal is a credible one [10].

Although Spence [35, p. 4] developed signaling theory in the hiring context, he explicitly intended it to be a broader theory with widespread utility [36]. Consider how signaling is manifested in self-organizing platforms, where the ex-ante unobservable property desired by a user is integrity of the digital good (e.g., a file). The fundamental problem here is adverse selection, which refers to an information asymmetry where one

TIWANA AND BUSH: SPOTTING LEMONS IN PLATFORM MARKETS: A CONJOINT EXPERIMENT ON SIGNALING 395

Fig. 1.

Research model.

party (the signalee) in the transaction lacks information that the other party (the signaler) possesses [1], [20]. A user searching for specific content (the signalee) who encounters a plethora of options might not know which content file has acceptable integrity, i.e., is indeed what it purports to be. In this setting, the signalee cannot reliably know ex-ante but can tell ex-post if the downloaded file was the intended one. Therefore, she must rely on informational cues—or signals—as potential indicators of unobservable digital good integrity, and use them to infer if each file is rouge or high integrity. Various signals therefore influence the signalee’s likelihood of transacting a digital good by shaping her overall perception, as discussed next. Fig. 1 summarizes the forthcoming research model. Our dependent variable is transaction likelihood, which we define as the likelihood that the signalee will download a digital good (a file) from a self-organizing content sharing platform. This criterion variable—the decision of whether or not to trade with another party—has extensively been used in the literature on information asymmetry in markets (e.g., [11], [18].

1) Mediating Role of Risk-Benefit Differential: The decision to download depends on the user’s perception of risk relative to perceived benefits. Risk comes from the possibility that the file is not what it appears to be. For example, the file might be a malicious program disguised as a legitimate file, a rouge software application, or a poor quality copy of the original.

Prior research suggests that the risks of downloading such a file might range from a harmless expenditure of resources such as bandwidth, lost opportunity costs, to outright harmful [2]. (Note our field experiment controls for the first two risks by ensuring that the decision to download is not costless and nontrivial by specifying the file size as substantial.) The potential benefit of deciding to download the file is that the user will get the digital good that she was seeking. Faced with such ambiguity about the integrity of the file, the user is likely to engage in riskbenefit calculus, i.e., estimate the differential between perceived risks vis-`a-vis perceived benefits. We build on the notion of perceived risk-benefit differential in Gundlach et al.’s [15] social information processing framework.

We define perceived risk-benefit differential as the net difference between the risks and expected benefits as perceived by the signalee in transacting for a digital good (i.e., downloading a file). When perceived risks exceed expected benefits, a user is less likely to download the file. Thus, the higher the perceived risk-benefit differential, the lower the transaction likelihood. Risk-benefit differential in turn is influenced by various signals as perceived by the signalee. However, the signaling literature says nothing about the underlying processes through which signalees make holistic inferences from signals [7], [10].

Consumers often combine signaling information from different sources to form an overall perception of a product [18], [27], p. 72], which suggests that a signalee is likely to combine multiple signals from different signal classes to arrive at an integrative, holistic assessment of risk-benefit differential. The signalee must make inferences about each signal, i.e., either discount it or weigh it based on their existing mental model. Therefore, signal processing requires integration of diverse potential signals to arrive at a holistic assessment [35, p. 69]. Signalees’ assessments are based on their interpretation of various signals, which is the process of translating the signals into perceived meaning [10].

Thus, signalees first form beliefs about signalers and then use these beliefs in making decisions [28]. As the signalee’s perception of the integrity of a file becomes more certain through an integrative interpretation of signals, the likelihood of the signalee downloading it increases. We therefore posit that the three

396 IEEE TRANSACTIONS ON ENGINEERING MANAGEMENT, VOL. 61, NO. 3, AUGUST 2014 classes of signals in our model influence transaction likelihood by altering the signalee’s holistic perception of risk-benefit differential, which serves as the central explanatory concept to elucidate how signals influence transaction likelihood.

B. Three Classes of Signals in Self-Organizing Platforms

The three classes of signals pertaining to who, what, and where respectively translate into three classes of signals in platform settings: Contributor, content, and platform signals.

1

Contributor signals pertain to who the contributor of a digital good is. Content signals represent signals about what the digital good is. Platform signals relate to where signaling about the digital good is taking place. Note that these three classes of signals have not been distinguished in prior studies and are theoretically developed in this study. Next, we theoretically develop each class of signals as well an explanation for how and why they influence transaction likelihood in such platforms.

1) Contributor Signals: Signaling effectiveness is contingent on who is signaling [10], which is the first class of signals that we consider. Building on signaling theory, we focus on two types of contributor signals, which focus on the digital good’s contributor, independent of file or platform characteristics. First, contributor market reputation 2 is the accumulated favorable ratings ascribed to the file’s contributor by other users of the platform. Such reputation therefore indicates the degree of the contributor’s cumulative history of signal fulfillment in prior interactions (not limited to the focal file) with other participants in that platform market. This notion also builds on the idea that reputations are socially constructed [10]. Behavior in prior transactions with other users might be difficult for a signalee to observe, but can plausibly be inferred from the contributor’s reputation in the platform marketplace [see [43]. Users in the self-organizing platform themselves might be in a good position to monitor each other [43]. The presence of such mutual peer monitoring therefore imposes a social control wherein individual users of the platform notice and respond to their peers’ behavior [23], [43]. This deters inappropriate behavior by contributors (e.g., supplying malicious content) by increasing the likelihood that it will be detected and penalized [2]. Accumulated contributor reputations on the platform therefore span time and discrete individual transactions, as suggested by Spence [36] for Internet-based market environments. This signal, which requires investment of time and effort to create, is therefore costly for a contributor of low integrity digital goods to mimic [34].

The peer group of other platform users thus functions as an institution to ensure mutual accountability and impose communal penalties [43]. Such peer monitoring is especially effective in the absence of centralized supervisory monitoring [23], as is the case in self-organizing platforms. Thus, all else being equal, higher contributor market reputation will lower the signalee’s perception of risk relative to benefit in transacting with that

1

Our three signal classes map 1:1 to Prabhu and Stewart’s [28] three categories: product signals (content signals in our study), sender related contextual signals (contributor signals), and environmental contextuals (platform signals).

2

Reputation is closely related to the concept of valence, which focuses on the proportion of positive to negative ratings.

contributor, and in turn increase the likelihood of a transaction between the signaler and the signalee.

Hypothesis 1a: Higher market reputation of a file’s contributor increases transaction likelihood because it decreases the signalee’s perceived risk-benefit differential.

Second, contributor rating volume, is whether a large or small number of other users have rated the file’s contributor, independent of how they rated her. This attribute builds on the observation that signals are only credible if they are costly to fake [20], [29]. For a signal to allow its recipient to differentiate high quality from low-quality contributors, it should be costly enough to dissuade a low-quality signaler from creating it but not too costly for a high-quality signaler to create. Signals are therefore valuable only when they are observable and costly to imitate (i.e., easy to produce truthfully but prohibitively costly to produce falsely) [36]. A larger volume of peer contributor ratings are likely to be a credible differentiating signal since the history of prior signaling cannot readily be mimicked at a low cost [10]. Therefore, volume of contributor ratings, independent of whether they are positive or negative, might affect the end user’s perceptions of risk-benefit differential in addition to the positive ratings alone described in the preceding hypothesis.

Thus, a larger volume of a file’s contributor ratings will decrease the signalee’s perception of risk relative to benefit in transacting with that contributor, and in turn increase transaction likelihood.

Hypothesis 1b: A larger rating volume of a file’s contributor ratings increases transaction likelihood because it decreases the signalee’s perceived risk-benefit differential.

2) Content Signals: The second class, content signals, are signals about the digital good (file that is prospectively transacted) itself. Building on signaling theory, we focus on two types of content signals, both about the digital good, independent of who its contributor is or the platform where it is offered. File

market reputation is the accumulated favorable ratings ascribed to the file by other users of the platform who have previously downloaded the file. Such reputation therefore indicates the cumulative history of signal consistency experienced by other users (see [10], or what Dulleck et al. [11]) characterize as accumulated market reputation. This signal, which requires investment of time and effort to create, is therefore costly for low integrity digital goods to mimic [34]. Thus, higher file market reputation will decrease the signalee’s perception of risk relative to benefit in downloading the file, and in turn increase the likelihood of a transaction involving the file.

Hypothesis 2a: Higher market reputation of a file increases its transaction likelihood because it decreases the signalee’s perceived riskbenefit differential.

Second, file rating volume is whether a large or small number of other users have rated the file contributor, independent of how they rated her. This attribute builds on Spence’s critical observation that signals are credible only if they are costly to fake. For a signal to allow its recipient to differentiate a high integrity file from a rogue file, it should be costly enough for a rogue file to mimic yet not too costly for a high integrity file to send [34]. A larger volume of platform user ratings of a file is

TIWANA AND BUSH: SPOTTING LEMONS IN PLATFORM MARKETS: A CONJOINT EXPERIMENT ON SIGNALING 397 a credible signal differentiating files of differing integrity. The sheer volume of file ratings, independent of whether they are positive or negative, may affect the end user’s perceptions of risk-benefit differential in addition to the positive file reputation described in the preceding hypothesis. Thus, a larger volume of file ratings will decrease the signalee’s perception of risk relative to benefit of downloading the file, and in turn increase the likelihood of a transaction involving the file.

Hypothesis 2b: A larger rating volume of a file increases its transaction likelihood because it decreases the signalee’s perceived riskbenefit differential.

3) Platform Signals: The third class, platform signals, pertains to where the signaling occurs, which in our context is the self-organizing platform where a digital good is hosted.

Signaling theory refers to this as the signaling environment, which plays a key role in determining the credibility of other signals [10], [12]. The context in which signaling occurs is as much a source of information as the signals about the digital good itself, and this context may change the way a signal is interpreted [28]. Although signaling theory recognizes the importance of the signaling environment, it is among the most understudied aspects of signaling theory [10], [28]. Our conceptualization of platform signals is therefore theoretically grounded in signaling theory, but anchored in its patterns of manifestation in our preliminary observations in 100 different peer-to-peer platforms. We focus on five platform signals, which pertain to where a file is hosted and are independent of file or contributor signals.

Platform reputation: The first platform signal that we consider is platform reputation, which we define as the level of confidence that the signalee has that the host platform is not inundated with low integrity digital goods. In his study of signaling in job markets, Spence [35, p. 56] suggested that third parties such as employment agencies sometimes take over the signaling role for job applicants. Ackerlof [1] further argued that the economies of scale available to such agencies make this a viable signaling approach, but requires that the certifying establishment be credible. Spence asserts that passing a screening by the employment agency itself constitutes a signal. In our study, a parallel intermediary is the platform itself. To stay viable, such a platform, just like employment agencies, will have an incentive to nurture a reputation of being turstworthy. They can do so by placing “institutional restrictions” or establishing norms for acceptable behavior [11]. Therefore, the stronger the reputation of a platform, the more confidence its users will likely have that digital goods hosted on the platform are not of low integrity, rogue, or malicious, thus reducing the signalee’s perception of risk relative to benefits from transacting on that platform. Thus, the choice of platform itself has signaling value. Therefore, we expect that higher reputation of the file’s host platform will decrease—independent of file or contributor characteristics— the signalee’s perception of risk relative to benefits in downloading a file hosted on the platform, thus increasing transaction likelihood. This leads to our next hypothesis.

Hypothesis 3a: Higher reputation of the file’s host platform increases transaction likelihood because it decreases the signalee’s perceived risk-benefit differential.

Platform participant legitimacy: The second platform signal is platform participant legitimacy, which we define as the degree to which a file contributor’s identity can be verified by a given platform. For example, some platforms utilize input control that discourages participation by low integrity contributors by requiring disclosure of personal information, registration, confirmable e-mail addresses, etc. This corresponds to access control, authentication, and identity management that the peerto-peer literature has emphasized as being important to their reliable functioning [2]. In contrast, if the platform allows users to join anonymously, it is infeasible to discover the true origin of malicious or low integrity digital goods [2]. Irrespective of the actual effect of such legitimation of digital goods contributors on a platform, we expect that a greater perception of legitimacy of users of a platform will engender greater confidence in a signalee’s belief about the integrity of the digital goods on that platform. We therefore expect that a higher perceived participant legitimacy will decrease the signalee’s perception of risk relative to benefits, thus increasing the transaction likelihood of a file on that platform. This leads to the next hypothesis.

Hypothesis 3b: Higher perceived legitimacy of users of the file’s host platform increases transaction likelihood because it decreases the signalee’s perceived risk-benefit differential.

Contributor rating system usage: Peer surveillance by the platform is one way to increase the likelihood of digital goods’ integrity in content sharing platforms [2]. The underlying rationale is that surveillance increases the perception that low integrity digital goods and its contributors will be communally penalized. Peer surveillance is usually implemented through content and contributor rating systems, thus signaling users of that platform. The next two platform signals pertain to this platform characteristic. First, we define contributor rating system

usage as the degree to which a contributor rating system is used in the platform. It is an attribute of the file’s host platform rather than of the file or its contributor. Reputation management systems are an integral part of self-organizing content sharing systems, and considerable variance has been observed in their use among competing platforms [2]. The more extensively such reputation tracking and management is used to monitor contributors, the more frequently users are likely to find ratings and comments from other users of the platform. The presence and widespread use of the associated disciplinary mechanisms are critical to credible signaling about the digital goods hosted by the platform. They create an incentive mechanism to stimulate acceptable behaviors among users and for them to communally discourage participation by low integrity digital goods providers [2]. This platform characteristic is costly for a low integrity platform to mimic, thus increasing the likelihood that greater use of such a contributor rating system will lower the signalee’s perception of risk in downloading a file. We therefore expect that greater use of a contributor rating system on a platform will decrease the signalee’s perception of risk relative to benefits from downloading a file, thus increasing transaction

398 IEEE TRANSACTIONS ON ENGINEERING MANAGEMENT, VOL. 61, NO. 3, AUGUST 2014 likelihood involving a file on that platform. This leads to our next hypothesis.

Hypothesis 3c: Greater usage of a contributor rating system on the file’s host platform increases transaction likelihood because it decreases the signalee’s perceived risk-benefit differential.

Content rating system usage, which we define as the degree to which a content rating system is used in the platform, serves a similar purpose for digital goods hosted on the platform. Such centralized rating systems for digital goods can reduce the number of rogue or malicious files on the platform [2], signaling to potential consumers that a communal penalty is likely to apply to low integrity digital goods. Their greater use is likely to reduce the signalee’s uncertainty about the digital goods hosted on that platform, thus lowering the signalee’s perception of risk relative to benefits from downloading a file hosted on that platform. This in turn should increase the transaction likelihood involving a file hosted on that platform, independent of file and contributor characteristics. This leads to our next hypothesis.

Hypothesis 3d: Greater usage of a content rating system on the file’s host platform increases transaction likelihood because it decreases the signalee’s perceived risk-benefit differential.

Decisional facilitation: Platforms, just like other types of markets, can be rife with signals [28]. The abundance of signals can result in informational overload. Therefore, a platform that provides its users mechanisms to readily use the variety of signals to make transaction decisions can help mitigate this problem. We define the ease of use of a platform’s rating system to help a signalee make informed download decisions as deci-

sional facilitation. For example, a platform might allow users to filter searches using personalized criteria (e.g., only show four or more star rated files or files uploaded in the last week) and aggregate various types of collected ratings. The extent to which a platform has such functionality and makes it readily accessible to signalees itself signals the integrity of the digital goods on the platform and the platform community’s stance toward low integrity digital goods. Greater decisional facilitation provided by a platform is therefore likely to lower the signalee’s perception of risk relative to benefits from downloading a file hosted on that platform, increasing transaction likelihood involving a file hosted on that platform. This leads to our final hypothesis.

Hypothesis 3e: Greater levels of decision-making facilitation functionality provided by the file’s host platform increases transaction likelihood because it decreases the signalee’s perceived risk-benefit differential.

III. R ESEARCH M ETHODOLOGY

A. Research Setting

We used peer-to-peer file sharing networks as the experimental context for testing our hypotheses, which was motivated by four theoretical advantages. First, it meets the two necessary conditions for signaling prevalence identified by Spence [35, p.

107]: a large number of participants and relative infrequency with which individuals transact with each other. Signalers and signalees are relatively numerous and interact dyadically sufficiently infrequently in peer-to-peer networks. This context also satisfies the necessary condition of market competition [11], as a user can potentially choose among several suppliers of the same file. Second, it meets the criteria for being self-organizing emphasized by Androutsellis-Theotokis [2]: Direct exchange of resources rather than intermediation by a centralized server or authority. Third, it exhibits sufficient variance in signaling mechanisms, particularly interesting in light of Spence’s [36] observation that the Internet potentially alters the signal-to-noise ratio in ways that could both impede and facilitate signaling. Finally, it circumvents the nontrivial confounding effect of pricing policies as no money is exchanged. The unit of analysis is the digital good, operationalized as a file.

B. Overview of the Conjoint Research Design

We used a conjoint design, a multiattribute judgment analysis technique that involves ex-post decomposition of the respondent’s decision-making process [24]. The conjoint methodology has been used in a variety of prior studies in strategy [42], management [33], operations [40], and MIS [39]. Two considerations motivated our choice of conjoint over a survey or an experiment: mitigation of social desirability bias and concurrent consideration of all signals.

3 A conjoint design combines the control of an experiment and the external validity of a survey, allowing us to isolate the focal theoretical constructs that neither a survey nor an experiment can. A conjoint design consists of three elements: 1) an attribute refers to a decision criterion (our nine signals) that respondents might use to evaluate the criterion variable; 2) overall utility is the overall value assigned by the respondent to a dependent variable and the contribution of each conjoint attribute its part-worth utility; and 3) conjoint profiles are different combinations of attribute levels.

Each respondent is sequentially presented the conjoint profiles with different combinations of attribute levels, and provides an assessment of the criterion variables for each profile. Each profile describes a file in terms of the levels of each of the nine signals and requires the respondent to assess risk-benefit differential and likelihood of downloading a file with that profile.

The underlying structure of the respondents’ cognitive models can then be statistically inferred by analyzing the responses at the individual scenario level and aggregate respondent levels using multiple regression [24]. The analyses are therefore comparable to repeated measures experimental design and account for nonindependence among the multiple assessments by a single respondent. To identify attributes that are statistically significant at the aggregate level, the regression coefficient for each attribute is averaged across individuals. A Z-statistic aggregates the T-statistics derived from the individual-level analysis for each conjoint attribute to assess the significance of each

3

First, it lowered social desirability bias, as the decision to download files on

P2P networks might be perceived as legally and socially controversial. Since conjoint profiles are hypothetical and do not require a recall of prior downloads, the approach is less susceptible to social desirability bias, retrospection bias, and confidentiality concerns [37]. Second, this design allows us to understand how users concurrently consider all nine signals and the three classes of signals to assess whether to download a file, allowing us to infer the relative importance they ascribe to each signal.

TIWANA AND BUSH: SPOTTING LEMONS IN PLATFORM MARKETS: A CONJOINT EXPERIMENT ON SIGNALING

TABLE I

D ESCRIPTIVE S TATISTICS FOR THE 12 C ONJOINT P ROFILES

399 predictor [26]. Hays’ ω 2 (a measure of the variance explained by each predictor) indicates its relative importance [24].

C. Operationalization of the Conjoint Profiles and Pilot Tests

We used two-level predictors with values of “high” and “low,” consistent with the majority of prior signaling studies, which use two-level treatments of signaled properties (e.g., high/low quality, good/bad, reliable/unreliable) [10], [29].

4

The number of possible combinations of nine signals each with two possible values (high or low) is 512 (2 9 ). To mitigate the infeasibility of having each respondent evaluate 512 conjoint profiles, an informationally more efficient fractional-factorial, orthogonal design was used [24]. The conjoint algorithm in SPSS 11.5 generated the fewest profiles with nine attributes that yields the most information. The tradeoff in this design is that it does not allow testing of interaction effects among the predictors, which is acceptable given the objectives of our study. This yielded 12 profiles as shown in Table I, each with varying combinations of the nine signal attribute levels (high or low), which were then sequentially evaluated by each respondent (each respondent evaluates all profiles). We piloted and refined the instrument with a convenience sample of 11 undergraduate students representative of the broader population of peer-to-peer platform users and five academic experts.

In the first class of signals, contributor signals, contributor market reputation was operationalized as the number of positive ratings by other users of that peer-to-peer platform of the

4

For theoretical tractability, the signaling literature follows Spence [35] in using two categories, high and low, which we also use for each signaled attribute.

This reduced the length of the conjoint survey to a realistic 12 conjoint profiles per respondent. In contrast, a three level manipulation (high, medium, and low) would have resulted in 3

9 possible profiles, which would result in respondent fatigue even with a fractional-factorial design. Two-level manipulations reduce the cognitive burden of simultaneously evaluating multiple signals.

contributor of that file. Contributor rating volume was operationalized as the total number of other users that have rated the contributor of the file, both positively and negatively. The second class, content signals, involved file market reputation and file rating volume. File market reputation was operationalized as the number of positive ratings by previous downloaders of that specific file, and file rating volume as the total number of previous downloaders that have rated that file. The third class, platform signals, was defined as characteristics of the peer-to-peer platform on which the user obtained the search results. Platform reputation was operationalized as the degree of confidence that the specific peer-to-peer platform on which the user obtained the search results was not flooded with fake or malicious files

(e.g., containing spyware or viruses). Platform participant legitimacy was operationalized as the extent to which the file contributors’ identities are verified in the specific peer-to-peer platform on which the user obtained the search results (e.g., by requiring registration, valid e-mail addresses, etc.). Contributor rating system usage was operationalized as how extensively a contributor rating system is used in the peer-to-peer platform on which the user obtained the search results (e.g., by users regularly leaving comments and ratings on each other). Content rating system usage was operationalized as how extensively a file rating system is used in the peer-to-peer platform on which the user obtained the search results (e.g., detailed comments left by previous downloaders). Decisional facilitation was operationalized as how easy it is to use the host platform’s file and contributor rating system to make file download decisions.

The two criterion variables for each conjoint profile were measured using two nine-point semantic-differential scales with bipolar anchors. The risk-benefit differential mediator was measured as the respondent’s perception of the risks versus benefits associated with downloading the particular file described by the specific conjoint profile. Transaction likelihood variable was measured as the likelihood that the respondent would download the file described in each conjoint profile.

400 IEEE TRANSACTIONS ON ENGINEERING MANAGEMENT, VOL. 61, NO. 3, AUGUST 2014

Fig. 2.

Data collection using the conjoint instrument.

D. Data Collection

We collected data from 380 undergraduate students enrolled in two sections of a core MIS course with 598 enrolled students

(63% response rate) at a major public university. Such students are representative of the intended target sampling frame, typical peer-to-peer platform users. The historical push by RIAA in targeting U.S. university undergraduates in its legal action against music piracy provides additional evidence to support this assertion. We used five iPods and fifteen $15 iTunes gift cards as incentives. The data collection progressed in three steps as summarized in Fig. 2. Each respondent was provided instructions and a printed reference card with definitions of the nine signal attributes. We instructed the respondents to assume they are searching Internet peer-to-peer networks such as Napster,

Kazaa, Limewire, or Bit Torrent (popularly recognized illustrative networks at the time of the study) to find a copy of a video game or movie DVD of interest to them. Each respondent was sequentially presented 12 search result conjoint profiles and asked to provide assessments for perceived risk-benefit differential and transaction likelihood on nine-point semanticdifferential scales for each of the 12 profiles, based on the information provided in the profile and their own expertise. Only the levels (high/low) of the nine signals were manipulated across the 12 profiles. Individual demographics and control variables were collected after the respondents evaluated all 12 profiles.

Anonymity, dollar value of the digital good, file size, download time, and peer-to-peer platform age was explicitly held constant to mitigate confounds. This provided 4,560 assessments (380 respondents × 12 conjoint profiles) for analyses, which explicitly accounts for nonindependence in the dataset. No significant differences existed between early (first 100) and late (last 100) respondents, mitigating the threat of nonresponse bias.

E. Rival Explanations: Control Variables and Confounding

Variables

We used eight control variables to account for rival explanations of transaction likelihood. The asserted utilities of each signal are largely derived from the past experience of the signalee (see [18], [35, p. 8]), here user characteristics. First, we controlled for Internet usage intensity (measured as the number of hours per day the respondent used the Internet), peer-topeer platform usage intensity (measured as whether the respondent regularly, occasionally, or almost never used peer-to-peer platforms), and peer-to-peer platform experience (measured in years). The second set was demographic characteristics that could influence the respondent’s evaluations. We controlled for the respondent’s technical expertise using a dummy for MIS major. We accounted for their educational level by constructing an index based on whether the respondent was a freshman, sophomore, junior, or senior (all were undergraduate students). The third set of demographic controls was age and gender. Finally, following prior conjoint studies, respondents’ self-reported level of confidence used an 11-point semantic-differential scale with bipolar anchors “little confidence” and “high confidence.”

Other variables that are neither included in the model nor as controls were explicitly held constant to prevent them from confounding the results. These variables include the downloader’s perception of anonymity in using the peer-to-peer platform, dollar value of the digital good, file size, file download time, and platform’s age. The respondents were explicitly instructed to assume that: 1) they were completely anonymous; 2) the retail value of the digital good was $30; 3) the file was 500 Megabytes in size; 4) it would take 10 minutes to download; and 5) the peerto-peer network has been in existence for five years.

F. Descriptive Statistics and Respondent Demographics

The majority (60.5%) of the participants in the study used the

Internet 2 to 4 hours a day, and the majority (57.9%) were occasional users of peer-to-peer platforms with about an equal mix

(

21%) of heavy and infrequent users. These demographics and variance in respondent experience with peer-to-peer platforms suggests that the sampling frame is appropriate for testing our hypotheses. The majority (64.7%) had three or more years of experience using peer-to-peer platforms, non-MIS backgrounds

(86.2%), largely male (63.5%), and were either sophomores

(47.6%) or juniors (34.9%). Means and standard deviations for each of the 12 conjoint profiles are summarized in Table I. Recall from Section III.C that the use of a fractional-factorial design reduced the number of needed profiles to 12, making the field data collection feasible.

TIWANA AND BUSH: SPOTTING LEMONS IN PLATFORM MARKETS: A CONJOINT EXPERIMENT ON SIGNALING

TABLE II

R ESULTS

401

IV. A

NALYSIS AND

R

ESULTS

Following procedures native to conjoint analysis, both individual respondent and aggregate search profile data were analyzed using hierarchical regression [14], [24].

β values indicate direction, Z-statistics statistical significance, and Hays’ ω

2 indicate the relative importance of each predictor in the model.

A. Hypothesis Tests

Each of our nine hypotheses were mediation hypotheses. We conducted two sets of analyses, one assessing the relationship between the mediator on the dependent variable (step 1), and the other between the nine independent variables on the mediator

(step 2). We used these results to conduct mediation tests using the updated version of Baron and Kenny’s [4] Sobel mediation test [25]. The results are summarized in Table II.

For Step 1, we used a three step hierarchical regression model in which we first introduced the control variables (step 1.1), followed by the main effects (step 1.2), and then the mediator

(step 1.3). Six of eight significant control variables explained

3.6% of the variance. The mediator, risk-benefit differential, explained 37.7% of the variance in transaction likelihood. As shown in step 1.3, risk-benefit differential (the mediator) had a significant positive relationship with transaction likelihood (the dependent variable). In the analyses for Step 2, the predictors for each of the three classes of signals were entered incrementally to compare the explanatory contribution of each class of signals in predicting risk-benefit differential. These steps correspond to contributor signals (step 2.1), content signals (step

2.2), and platform signals (step 2.3). As shown, the aggregated

Z-statistics were significant and β coefficients were in the hypothesized directions for all nine signaling predictors. Mediation test statistics are also shown in Table II along with the measure of relative importance ω 2 for each of the nine predictors. The relationship between the mediator and the dependent variable was negative and significant, suggesting that when perceived risk exceeds the benefits of downloading a file, transaction likelihood is lowered. Hypotheses 1a and 1b pertain to contributor signals. Contributor market reputation had a negative and significant relationship with risk-benefit differential ( β = − .25, z -stat =

19.63, p < .001), and its effect on transaction likelihood was significantly mediated by risk-benefit differential

( T

Sob el

= 18.84, p < 0 .

001 ). This suggests that greater contributor market reputation lowers risk-benefit differential, which in turn increases transaction likelihood, supporting Hypothesis 1a.

Contributor rating volume had a negative, significant relationship with risk-benefit differential ( β =

.14, z -stat =

10.94, p < .001), which significantly mediated its effect on transaction likelihood ( T

Sob el

= 10.8, p < .001). This suggests that contributor rating volume increases transaction likelihood by lowering risk-benefit differential, supporting Hypothesis 1b.

Hypotheses 2a and 2b pertain to content signals. File market reputation had a negative and significant relationship with risk-benefit differential ( β = − .29, z -stat = − 22.5, p < .001), and its effect on transaction likelihood was significantly mediated by risk-benefit differential ( T

Sob el

= 21.33, p < .001), supporting Hypothesis 2a. File rating volume had a negative, significant relationship with risk-benefit differential ( β =

.13, z -stat =

9.94, p < .001), which significantly mediated its effect on transaction likelihood ( T

Sob el

= 9.83, p < .001), supporting

Hypothesis 2b. Hypotheses 3a–3e pertain to platform signals.

Platform reputation had a negative, significant relationship with risk-benefit differential ( β =

.23, z -stat =

18.16, p < .001), which significantly mediated its effect on transaction likelihood

( T

Sob el

= 17.53, p < .001), supporting Hypothesis 3a. Platform participant legitimacy had a negative, significant relationship with risk-benefit differential ( β = − .15, z -stat = − 11.52, p

< .001), which significantly mediated its effect on transaction likelihood ( T

Sob el

= 11.35, p < .001), supporting Hypothesis 3b.

Contributor rating system had a negative, significant relationship with risk-benefit differential ( β =

.08, z -stat =

6.61, p < .001), which significantly mediated its effect on transaction

402 IEEE TRANSACTIONS ON ENGINEERING MANAGEMENT, VOL. 61, NO. 3, AUGUST 2014 likelihood ( T

Sob el

= 6.58, p < .001), supporting Hypothesis 3c.

Content rating system usage had a negative, significant relationship with risk-benefit differential ( β =

.12, z -stat =

9.24, p < .001), which significantly mediated its effect on transaction likelihood ( T

Sob el

= 9.15, p < .001), supporting Hypothesis 3d.

Finally, decisional facilitation lowers risk-benefit differential ( β

= − .09, z -stat = − 6.74, p < .001), which significantly mediated its effect on transaction likelihood ( T

Sob el

= 6.71, p < .001). This suggests that the greater the extent to which a self-organizing platform provides mechanisms to facilitate download decision making, the greater transaction likelihood will be because this lowers perceived risk-benefit differential, supporting Hypothesis 3e. The relationships in hypotheses 1a, 3b, and 3c were fully mediated and the other six were partially mediated. In all cases, mediation tests reveal that risk-benefit differential significantly mediates the effect of all nine predictors on transaction likelihood. All nine hypotheses were therefore fully supported.

TABLE III

R ELATIVE I MPORTANCE OF THE T HREE S IGNAL C LASSES

Therefore, the platform on which the digital good exists is the primary determinant of its subsequent transaction, followed by the signals about the digital good, and the relatively least important is who contributes that digital good. Thus, where is more important than what, and what is more important than who in self-organizing platforms.

B. Relative Importance of Individual Signals

The magnitude of ω 2 in Table II provides insights into each signal’s relative importance.

5 The results suggest that users of self-organizing platforms ascribe greatest importance (in that order) to content reputation, contributor market reputation, and platform reputation. The top three signals encompass all three classes of signals—contributor, content, and platform. This highlights that users triangulate different classes and sources of diverse signals to make a holistic assessment about content integrity. The relatively least importance is ascribed to content rating system usage, decisional facilitation, and contributor rating system usage.

D. Limitations

Four limitations merit consideration. First, a fractionalfactorial conjoint design might result in spurious, unrealistic scenarios; but our pretests did not suggest that this was the case.

The orthogonal design precludes assessing interactions among signals. Second, our design limited us to nine signals; this does not trivialize the contribution since this is among the first studies to theoretically develop a signaling perspective. Further, we did not control respondent attitudes toward piracy and their risk propensity. Third, we do not differentiate content quality problems (e.g., truncated files or poor quality files) from content integrity (e.g., malicious files). While the former is likely to decrease benefits, the latter is likely to increase risk; this distinction awaits future work. Finally, our student subjects are highly representative of peer-to-peer platform users [9], but caution is warranted when generalizing our results to other types of users.

C. Relative Importance of the Three Classes of Signals

The relative importance of the three classes of signals is indicated by the Δ R 2 for each block of class-specific predictors in steps 2.1, 2.2, and 2.3 in Table II, and by complementing superattribute analysis. The largest proportion of variance was explained by platform signals ( R 2 = 10.2%, F ( Δ R 2 ) = 127.4, p < .001), followed by content signals ( R 2 = 9.7%,

= 265.1, p < .001), and finally contributor signals ( R 2

F

=

( Δ R 2 )

8.1%,

F ( Δ R 2 ) = 197.9, p < .001). However, since the study included more platform signals than in the other two classes, there is a plausible risk that this relative importance could be an artifact of this difference. To mitigate this risk, we followed the approach recommended by Hair et al. [16, p. 576], and created three superattributes (one for each signal class) by computing separate composite scores for signal variables within each class. We then reestimated the model using these superattributes instead of individual signal variables (see Table III).

The magnitude of the β s in Table III shows that ordering of signal class importance is consistent with Table II. The model explained 24.8% of the variance in risk-benefit differential and the model was significant at the 0.1% level (Model F = 492.6).

5

We tested our model without the risk-benefit differential to examine the direct effects of the signal classes on the dependent variable. The model while less parsimonious in light of our theorizing remained robust with a consistent pattern of significant results.

V. D ISCUSSION AND I MPLICATIONS

This study was motivated by the inattention to how users discern ex-ante the integrity of digital goods on self-organizing platforms using multiple types of signals. On one hand, such platforms are increasingly pervasive in diverse industries (e.g., publishing, entertainment, mobile computing, and software services). Simultaneously, the Internet is changing the informational structure of conventional markets by lowering both signaling and signal reception costs, making it even more difficult for users to discern high- and low-quality participants in markets [36]. Our study departs from economists’ debatable assumption that reputation involving transactions of experience goods is not visible to others over time ([e.g., [11]), as well as the emphasis of prior studies on the supply-side at the neglect of the consumption-side of platforms.

We used Spence’s [34], [36] signaling theory as a starting point for our theoretical development. Signaling theory acknowledges the existence of multiple signals yet prior research has not identified how signals may be interrelated and form classes of signals. We introduced the notion of three classes of signals that pertain to the “who, what, where” facets of a digital good. We proposed that end-users simultaneously glean signaling information from multiple signals in each class, which in turn influences their decision to transact for a digital good on

TIWANA AND BUSH: SPOTTING LEMONS IN PLATFORM MARKETS: A CONJOINT EXPERIMENT ON SIGNALING 403 self-organizing platforms. Further, we developed a theoretical explanation for how these signals influence transaction likelihood: Multiple signals shape users’ holistic, integrative riskbenefit differential regarding a potential transaction, which in turn influences digital good transaction likelihood. Tests using data collected from 380 users support the proposed ideas. Our results contribute four new theoretical insights to research on signaling theory and three to research on platforms.

TABLE IV

E FFECTS OF C ONTRIBUTOR AND C ONTENT S IGNALS IN L OW AND H IGH

R EPUTATION P LATFORMS

A. Contributions and Theoretical Implications for Signaling

Theory

Our contributions to signaling theory are four fold: 1) introduction of theoretically generalizable classes of signals; 2) signaling portfolios; 3) delineation of the intervening mechanisms between signaling and behavior; and 4) the role of the signaling environment in shaping which class of signal dominates. Our first contribution to signaling theory is in theoretically developing the underdeveloped notion of classes of signals. Although the idea that individuals often use multiple signals is pervasive in contemporary signaling theory [10], [20], [27, p. 22], Spence did not distinguish between different types—or classes—of signals.

Signaling theory has long recognized the existence of multiple types of signals [10], but with little empirical research on multiple signals [20], sheds little light on what these types are. The theoretically generalizable three classes of signals introduced here—who, what, and where—addresses this gap. We theoretically developed these classes of signals building on signaling theory’s emphasis on signaling information communicated by characteristics of the object of signaling, the sender of the signal, and the environment in which signals are sent and received [28].

The three classes of signals corresponded to contributor, content, and platform signals in our model. Our results indicate that all three classes of signals influence platform users’ perceptions.

This is precisely why there is theoretical value in distinguishing among signal classes: a platform owner might otherwise overlook an entire class of signaling mechanisms in the implementation of a platform. This distinction adds a nuanced new level of granularity to the broader signaling perspective that has previously not theoretically developed such distinctions.

Our second contribution to signaling theory is in showing that platform users differentially weigh and synthesize different types of signals into a portfolio of signals. In a recent review of the signaling literature, Connelly [10] highlighted that portfolios of signals might be used to maximize their collective effectiveness, yet no prior study has conceptualized such signal portfolios. We also do not know the relative emphasis signal recipients place on different types of signals within a signal portfolio [28]. Our findings imply that recipients, in their mental calculus, differentially weigh but triangulate signaling information from all three classes to form holistic inferences.

Our findings therefore directly address the gap of how receivers’ interpretation of multiple signals influences their beliefs about signalers and their responses [28].

Our third contribution to signaling theory is developing and testing an explanation for why signals—by altering the signal recipient’s perception of risk-benefit differential—influence signalees’ behavior. This directly addresses a persistent gap in signaling theory, where the underlying mechanisms that explain how signaling mechanisms produce judgments have been absent [7], [27, p. 59]. The mediating role of risk-benefit differential in explaining how signals affect transaction likelihood therefore theoretically fleshes out Pentland’s [27, p. 72] intuition about the role of information integration in understanding how signalees aggregate multiple signals into “meaningful wholes” [10].

The fourth insight for signaling theory—that the signaling environment alters what class of signal dominates—is offered by additional post-hoc analyses. Spence [34] suggested that one category of signals might take on increasing importance in the absence of others. We conducted additional analyses to assess this theoretical point, which has received scarce attention in prior studies. Building on this insight, we used post-hoc analyses to explore whether content or contributor signals matter more in platforms with low or ambiguous integrity (after confirming that platform reputation differences indeed lead to significant differences in predicted variables.

6 ) A comparison of aggregated contributor and content signals for high and low reputation platforms suggests that the signaling environment alters whether a signal about who (contributor) or what (content) matters more to signal recipients (see Table IV). Contributor signals are more influential in platforms with low reputation, while content signals are more influential in platforms with high reputation (as illustrated in Fig. 3). This implies that where the signaling occurs alters whether who is signaling or what object the signal is about matters more to signal recipients. In nascent platforms, contributor signals are therefore more important than content signals but in platforms with established positive reputations, content signals are more important than contributor signals. This has implications for platform design.

7

6

A one-way ANOVA showed that risk-benefit differential was significantly lower in platforms with high reputation compared to those with low reputation

(4.40 versus 4.99; F-value 86.7; p < .001) and transaction likelihood significantly higher in platforms with high reputation vis-`a-vis low reputation (5.29

versus 4.69; F-value 68.7; p < .001). (We aggregated the five platform signals to create the two subgroups.)

7

Important systems design implication can be drawn. Realistically only a small subset of signaling-related software features can be implemented under time/ budget constraints. Knowing which of the features are most effective and when helps. Various integrity assurance functionality in platforms can broadly be classified into three classes of mechanisms—contributor, content, and platformrelated mechanisms. Our findings suggest that systems designers should ideally implement technical features that target assurance of the integrity of content, its contributors, as well as the platform itself. These are interrelated in that a system feature primarily in one category might simultaneously enhance system capabilities in the other two classes. But as Table III shows, content integrity is the most important and thus deserves primary attention in the design of the sys-

404 IEEE TRANSACTIONS ON ENGINEERING MANAGEMENT, VOL. 61, NO. 3, AUGUST 2014

Fig. 3.

Flipping importance of contributor and content signals based on platform reputation.

B. Contributions and Theoretical Implications for Platforms Research

Our results make three novel contributions to the emerging literature on platform-based markets, which unlike this study, focuses largely on the supply side of platforms. We show that self-organizing platforms—in the absence of centralized oversight—require signaling mechanisms at multiple levels to assure users about content integrity. Second, they imply that platform designers must incorporate signaling functionality in a platform such as through persistent contributor and content ratings management systems, making this data easy to use, and exercising input control and authentication to regulate new user selection. This aspect of platform design, by highlighting the consumer-facing considerations, complements prior studies that have focused on how their architecture and design affects supply-side challenges such as performance optimization and externality generation (e.g., [2], [19], [21]). Our study thus is a first step to answer Androutsellis-Theotokis’ [2] call for research on design attributes that are critical to the functioning of such platforms. Third, we extend the economics literature on information goods by broadening the focus from ex-ante uncertainty about consumer-perceived value to their upstream tem, followed by contributor integrity, and then platform integrity. For systems designers, this implies that features pertaining to peer user evaluations of content should be easy to use, difficult to game, and cumulative and persistent over time. Further, as the number of content ratings for each unit of content grows over time, system designers should present aggregated patterns in a manner that minimizes information overload for users. This can be accomplished through features that average content ratings across all raters, show distribution patterns of ratings, and allow slicing and dicing of large volumes of ratings using automated mechanisms such as tag clouds and keyword extraction. Other features that also may be valuable to end users are file checksums, user comments, and review usefulness ratings by users. The next most influential feature pertains to contributor reputation. Systems designers should therefore provide mechanisms that ensure that contributors are reliably screened and vetted as legitimate

(e.g., by validating non-free email addresses), and that contributor ratings are difficult to game or fake. As contributor rating volumes grow, systems designs can increase the utility of the system by making the volume of this information easier to digest (e.g., by creating contributor scorecards that summarize lifetime patterns of contributor ratings and recent ratings). Such data can also be used to categorize contributors into easy-to-interpret classes (e.g., novice or veteran) and to automatically assign badges/ icons that succinctly communicate their level of trustworthiness based on their rating histories. Finally, system designers can tweak platform level features and functionality over its lifecycle.

integrity (e.g., [31], [44]). Communicating content integrity, the focus of this work has largely been overlooked in economics.

C. Directions for Future Research

Our findings lead to six promising questions for future research on signaling in platform markets. First, how do signals interact with a platform’s licensing arrangements (e.g., Creative Commons and open source licensing)? Second, we circumvented pricing issues by focusing on platforms that did not involve the exchange of money, but those issues are increasingly important in commercial platforms for app distribution, cloud services, and digital media sales. Third, how do platform governance (e.g., distribution of decision rights among end-users, the platform owner, and content contributors) and signal verifiability make some classes of signals less crucial in self-organizing platforms? Fourth, in crowded platform markets where attention is scarce, how can digital goods producers design signaling strategies to differentiate and garner attention to their offerings?

Fifth, although our research design precluded testing interactions, studies on the interplay among signal classes can offer deeper insight into potential synergies among signals. Finally, future work can explore signaling mechanisms that span multiple, independent platforms wherein content integrity and contributor ratings reside elsewhere but are summoned by the host platform.

In conclusion, we focused on how content integrity is signaled among participants in self-organizing platforms that are now widespread for distributing digital goods. The overarching insight from the study is about how and why signals of different types—by helping spot “lemons” [1] in the market—influence the consumption of digital goods in self-organizing platforms.

The unusual unit of analysis of the digital good allowed us to uncover nuanced insights into how signaling behavior combines three elements at the transaction-specific level—the transactable digital good (what), its contributor (who), and the context in which the signaling occurs (where). In addition to showing the relative importance that users ascribe to each signal, we also showed that the relative salience of content and contributor signal classes is swapped depending on the reputation of the

TIWANA AND BUSH: SPOTTING LEMONS IN PLATFORM MARKETS: A CONJOINT EXPERIMENT ON SIGNALING 405 platform where signaling occurs. This study provides a first step for using a signaling lens to understand how platforms function, but richer theory development opportunities abound as such platforms become pervasive in contemporary markets.

R EFERENCES

[1] G. Akerlof, “The market for lemons: Quality uncertainty and the market mechanism,” Quarterly J. Econ., vol. 84, no. 3, pp. 488–500, 1970.

[2] S. Androutsellis-Theotokis and D. Spinellis, “A survey of peer-to-peer content distribution technologies,” ACM Comput. Surveys, vol. 36, no. 4, pp. 335–371, 2004.

[3] R. Aron and E. K. Clemons, “Achieving the optimal balance between investment in quality and investment in self-promotion for information products,” J. Manage. Inform. Syst., vol. 18, no. 2, pp. 65–88, 2001.

[4] R. Baron and D. Kenny, “The moderator-mediator variable distinction in social psychological research: Conceptual, strategic, and statistical considerations,” J. Personality Soc. Psychol., vol. 51, no. 6, pp. 1173–1182,

1986.

[5] S. Bhattacharjee, R. Gopal, K. Lertwachara, and J. Marsden, “Consumer search and retailer strategies in the presence of online music sharing,” J.

Manage. Inform. Syst., vol. 23, no. 1, pp. 129–159, 2006.

[6] N. Bilton. (2010). Malware infection hits Russian android phones

New York Times, August 10, [Online]. Available: bits.blogs.nytimes.

com/2010/08/10/malware-infection-hits-russian-android-phones/

[7] W. Boulding and A. Kirmani, “A consumer-side experimental examination of signalling theory,” J. Consumer Res., vol. 20, no. 1, pp. 111–123, 1993.

[8] BPI. (2010). British recorded music industry digital music na-

tion study [Online]. Available: www.bpi.co.uk/assets/files/Digital%

20Music%20Nation%202010.pdf (accessed August 19, 2011)

[9] E. Cohn. (2008). RIAA targets college students The Cornell Daily Sun,

[Online]. Available: cornellsun.com/node/28168

[10] B. Connelly, S. Certo, D. Ireland, and C. Reutzel, “Signaling theory: A review and assessment,” J. Manage., vol. 37, no. 1, pp. 39–67, 2010.

[11] U. Dulleck, R. Kerschbamer, and M. Sutter, “The economics of credence goods: An experiment on the role of liability, verifiability, reputation, and competition,” Amer. Econ. Rev., vol. 101, no. 2, pp. 526–555, 2011.

[12] M. Engers, “Signalling with many signals,” Econometrica, vol. 55, no. 3, pp. 663–674, 1987.

[13] A. Forte, V. Larco, and A. Bruckman, “Decentralization in wikipedia governance,” J. Manage. Inform. Syst., vol. 26, no. 1, pp. 49–72, 2009.

[14] P.E. Green, K. Helsen, and B. Shandler, “Conjoint internal validity under alternative profile presentations,” J. Consumer Res., vol. 15, no. 3, pp. 392–

397, 1988.

[15] M. Gundlach, S. Douglas, and M. Martinko, “The decision to blow the whistle: A social information processing framework,” Acad. Manage. Rev., vol. 28, no. 1, pp. 107–123, 2003.

[16] J. F. Hair, Jr., R. E. Anderson, R. L. Tatham, and W. C. Black, Multivariate

Data Analysis, 4th ed.

Englewood Cliffs, NJ, USA: Prentice-Hall, 1995.

[17] K. Hosanagar, J. Chuang, R. Krishnan, and M. Smith, “Service adoption and pricing of content delivery network services,” Manage. Sci., vol. 54, no. 9, pp. 1579–1593, 2008.

[18] P. Huang, N. Lurie, and S. Mitra, “Searching for experience on the web:

An empirical examination of consumer behavior for search and experience goods,” J. Marketing, vol. 73, no. 2, pp. 55–69, 2009.

[19] M. Johar, S. Menon, and V. Mookerjee, “Analyzing sharing in peer-to-peer networks under various congestion measures,” Inform. Syst. Res., vol. 22, no. 2, pp. 325–345, 2011.

[20] A. Kirmani and A. Rao, “No pain, no gain: A critical review of the literature on signaling unobservable product quality,” J. Marketing, vol. 64, no. 2, pp. 66–79, 2000.

[21] K. Lang and R. Vragov, “A pricing mechanism for digital content distribution over computer networks,” J. Manage. Inform. Syst., vol. 22, no. 2, pp. 121–139, 2005.

[22] S. Liebowitz, “Testing file sharing’s impact on music album sales in cities,”

Manage. Sci., vol. 54, no. 4, pp. 852–859, 2008.

[23] M. Loughry and H. Tosi, “Performance implications of peer monitoring,”

Organization Sci., vol. 19, no. 6, pp. 876–890, 2008.

[24] J. Louviere, Analyzing Decision Making: Metric Conjoint Analysis.

Beverly Hills, CA, USA: Sage, 1988.

[25] D. MacKinnon, C. Lockwood, J. Hoffman, S. West, and V. Sheets, “A comparison of methods to test mediation and other intervening variable effects,” Psychol. Methods, vol. 7, no. 1, pp. 83–104, 2002.

[26] W. Moore, “Levels of aggregation in conjoint analysis: An empirical comparison,” J. Marketing Res., vol. 17, no. 4, pp. 516–523, 1980.

[27] S. Pentland, Honest Signals.

Cambridge, MA, USA: MIT Press, 2008.

[28] J. Prabhu and D. Stewart, “Signaling strategies in competitive interaction:

Building reputations and hiding the truth,” J. Marketing Res., vol. 38, no. 1, pp. 62–73, 2001.

[29] J. Riley, “Silver signals: Twenty-Five years of screening and signaling,”

J. Econ. Literature, vol. 39, no. 2, pp. 432–478, 2001.

[30] P. Schoemaker and G. Day, “How to make sense of weak signals,” Sloan

Manage. Rev., vol. 50, no. 9, pp. 81–89, 2009.

[31] C. Shapiro and H. Varian in Versioning: The Smart Way to Sell Informa-

tion., Harvard Business Rev., Nov./Dec. 1998, pp. 16–114.

[32] C. Shapiro and H. Varian, Information Rules: A Strategic Guide to the

Network Economy.

Boston, MA,, USA: Harvard Business School Press,

1999.

[33] D. A. Shepherd, “Venture capitalists’ assessment of new venture survival,”

Manage. Sci., vol. 45, no. 5, pp. 621–630, 1999.

[34] A. Spence, “Job market signaling,” Quart. J. Econ., vol. 87, no. 3, pp. 355–

374, 1973.

[35] A. Spence, Market Signaling: Informational Transfer in Hiring and Re-

lated Screening Processes.

Cambridge, MA, USA: Harvard Univ. Press,

1974.

[36] A. Spence, “Signaling in retrospect and the informational structure of markets,” Amer. Econ. Rev., vol. 92, no. 3, pp. 434–459, 2002.

[37] S. Stremersch, A. Weiss, B. Dellaert, and R. Frambach, “Buying modular systems in technology-intensive markets,” J. Marketing Res., vol. 40, no. 3, pp. 335–350, 2003.

[38] The Economist. (2010). The bigger picture [Online]. Available: www.economist.com/node/15582215

[39] A. Tiwana and A. Bush, “A comparison of transaction cost, agency, and knowledge-based predictors of IT outsourcing decisions,” J. Manage. In-

form. Syst., vol. 24, no. 1, pp. 263–305, 2007.

[40] A. Tiwana, M. Keil, and R. Fichman, “IS project continuation in escalation situations: A real options model,” Decision Sci., vol. 37, no. 3, pp. 357–

391, 2006.

[41] A. Tiwana, B. Konsynski, and A. Bush, “Platform evolution: Coevolution of architecture, governance, and environmental dynamics,” Inform. Syst.

Res., vol. 21, no. 4, pp. 675–687, 2010.

[42] B. Tyler and H. Steensma, “Evaluating technological collaborative opportunities: A cognitive modeling perspective,” Strategic Manage. J., vol. 16, pp. 43–70, 1995.

[43] H. Varian, “Monitoring agents with other agents,” J. Institutional Theo-

retical Econ., vol. 146, pp. 153–174, 1990.

[44] H. Varian, “Buying, sharing and renting information goods,” J. Ind. Econ., vol. 48, no. 4, pp. 473–488, 2000.

[45] M. Xia, Y. Huang, W. Duan, and A. Whinston, “To continue sharing or not to continue sharing? An empirical analysis of user decision in peer-to-peer sharing networks,” Inform. Syst. Res., vol. 23, no. 1, pp. 244–259, 2012.

Amrit Tiwana is a Professor at the University of Georgia. He serves on the editorial boards of Strategic Management Journal, Information Systems Research,

Journal of Management Information Systems, and IEEE Transactions on Engi-

neering Management.

Ashley A. Bush is currently an Associate Professor at Florida State University,

FL, USA. She received the Ph.D. degree from Georgia State University, GA,

USA. Her research has appeared in Information Systems Research, Journal of

Management Information Systems, IEEE Transactions on Engineering Man-

agement, and others.

Download