Estimating Trading Risks in the Presence of Reporting Bias: Identifiability Problems in Online Feedback Mechanisms: Chrysanthos Dellarocas Charles A. Wood 3rd May 2006 Abstract Online feedback mechanisms have become an important component of electronic business, helping to elicit good behavior and cooperation among loosely connected and geographically dispersed economic agents (Del- larocas 2003). For example, eBay’s feedback mechanism is the primary means through which eBay elicits honest behavior and, thus, facilitates transactions among strangers over the Internet (Resnick and Zeckhauser 2002). Since most details of commercial transactions are unobservable to the market authority, the majority of online feedback mechanisms rely on voluntary self-reporting of transaction outcomes by one or both parties involved. As a consequence, not every transaction receives feedback. More importantly, self-reporting opens the door to several forms of reporting bias : traders may strategically misreport some outcomes, or may selectively choose to report certain types of outcomes and not others. If reporting bias is severe enough, posted feedback provides a distorted view of the risks that are associated with trading in a given market. Its usefulness, both in deterring fraud and in informing buyers, then becomes severely diminished. There are important indications that reporting bias is present in online feedback: Posted feedback in most systems is overwhelmingly positive. For example, more than 99% of all feedback posted on eBay is positive (Resnick and Zeckhauser, 2002). A naïve reading of this empirical fact may lead one to conclude that more than 99% of eBay transactions result in satisfactory outcomes. Such a conclusion runs against wide-spread reports of consumer fraud in online auctions. Internet Auctions accounted for 16% of all consumer fraud complaints received by the Federal Trade Commission in 2004, the highest level of fraud of any Internet transaction type (see http://www.consumer.gov/sentinel/). One possible explanation for the discrepancy between the overwhelmingly positive online feedback and the large number of consumer fraud complaints is that, whereas satisfied traders generally express their satisfaction online, dissatisfied traders often prefer to remain silent rather than to report their dissatisfaction to the system. The reciprocal nature of auction feedback is considered the main reason behind such reporting bias. Specifically, it is widely believed (though, so far, not rigorously proven) that many traders choose to remain silent because they are afraid that, if they report their negative experience, their partner will retaliate with negative feedback. The presence of reporting bias on eBay has been discussed by several authors (Resnick and Zeckhauser 2002; Reichling 2004; Klein et al. 2005). However, so far there has not been an attempt to quantify the degree to which it is present or an assessment of the extent to which it distorts the distribution of published feedback relative to the underlying distribution of transaction outcomes that this feedback is supposed to track. We fill this gap by offering what we believe is the first quantitative method for assessing and “repairing” the impact of reporting bias on feedback mechanisms. Our method draws heavily on ideas drawn from the statistical theory of identifiability of finite mixtures (see, for example, Prakasa Rao 1992) and derives general possibility and impossibility results regarding the ability to infer the distributions of latent transaction 1