網路市集賣家影響下的評價分數與評論取向 - 政大學術集成

全文

(1)國立政治大學資訊管理學系. 碩士學位論文指導教授:管郁君博士、林淑瓊博士政治. 立. 大. ‧. ‧ 國. 學. io. sit. y. Nat. 網路市集賣家影響下的評價分數與評論取向. al. er. Rating Score and Comment Orientation under the Influence of. n. v i n C hMarketplace USellers Online engchi. 研究生：謝宜廷中華民國 102 年 7 月.

(2) 謝辭轉瞬間兩年的碩士班生活就以此篇論文畫下句點，而這篇論文受到許多人的幫助才得以順利完成，其中特別感謝兩位指導老師管郁君博士與林淑瓊博士的教導與照顧，從文獻的閱讀、題目的發想、再到研究架構的建立以及文章的撰寫，兩位老師都給予了許多的指導，也受到了管老師的鼓勵以英文撰寫論文，過程中當然遇到了許多的問題，兩位老師總是願意撥出時間一同討論並解決問題，逐字逐句的修正，耐心地指導，著實令我受惠良多。兩年的時間，我有了滿滿的收穫，不只是在學業上更包含. 政治大話，但每句話所包含的意義我都將永遠記在腦海中，非常榮幸能成為兩位老師的學生，立也非常感激兩位老師給予的指導與照顧。此外，感謝兩位口試委員不辭辛勞參與口試，並給予了許多寶貴的意見，讓我的論文能夠更加精進。. 學. ‧ 國. 了待人處事方面讓我有所成長，或許我無法完整引用管老師或是淑瓊學姐曾經說過的. ‧. 感謝系辦的兩位助教，詩晴助教和雨儒助教，在各項行政事務上總是很有耐心詳細的解釋流程，大力幫忙，在系辦工讀期間也深受兩位助教的照顧，由衷感謝兩位助. sit. y. Nat. 教。另外，感謝與我一同奮鬥的實驗室夥伴們，五個人相輔相成，一同學習、一同遊玩、相互打氣，這段期間留下了許多回憶，也感謝我們能夠彼此認識，並珍惜此段友. io. n. al. er. 誼。也感謝實驗室的學長姐所給予的經驗分享以及學弟妹的協助幫忙，特別感謝勝為. i n U. v. 學長、國華學長以及家榮學長在每次討論過後總是不吝嗇的給我許多意見與建議，讓. Ch. engchi. 我能從更多不同的面向去思考。去年，我們規畫了第一次的實驗室之旅，蔚藍的天空，碧藍的海，留下了許多蘭陽之旅的回憶，今年藉著參加研討會，我們也有了韓國之旅，連續兩年與大家一起旅遊，是一種特別的經驗，大家共同參與使得旅程更加的豐富精彩，並留下深刻且美好的回憶，期待下次的相聚。最後，感謝我最重要的家人，爸爸、媽媽以及姐姐，因為有你們的支持我才能無後顧之憂的一路念到碩士，我很幸福，因為我有最堅強的後盾，我很感激，因為我有最珍惜我的人，我也同樣珍惜我最重要的家人。漫長的學生生活從幼稚園延伸至此，雖以這篇論文畫下句點，新的段落也從此展開，期待自己的蛻變，並給與未來的自己多一點勇氣前往冒險，再次感謝一路上給予幫助的所有人，我都將永遠銘記。 i.

(3) Abstract Online evaluation systems play an important role in creating trust between sellers and buyers in online marketplaces. Consumers often rely on these evaluations when making purchase decisions. As a result, sellers need to take buyers’ evaluations seriously. However, evaluation manipulation occurs in many online marketplaces. Various types of manipulation. 治政 sellers engage in systematic evaluation manipulation 大 by constantly posting reminders to 立 influence buyers’ evaluations. The reminders generally aim to prompt a positive rating or. are apparent. Although most appear to be spurious and disappear rather quickly, many. ‧ 國. 學. prohibit a negative rating. This study examines the influence of seller manipulation on rating scores and comment orientation in an evaluation system. The research framework. ‧. categorizes three types of seller manipulation: allurement, rejection and none. Taobao.com, the largest online marketplace in China, is used to evaluate the proposed model.. y. Nat. io. sit. The results show that seller manipulation only influences rating scores. However, the. er. interaction between seller manipulation and buyer experience influences both the. al. n. v i n C of seller manipulation. differences among the three types h e n g c h i U Buyer experience mediates the. orientation of comments and the percentage of positive ratings, which leads to significant. relationship between seller manipulation and buyer evaluation. Although rating scores are. influenced by seller manipulation, comments tend to contain much more valuable information. It is thus suggested that consumers rely on rating scores and comments when making purchase decisions. Keywords: Online evaluation system, Rating score, Comment orientation, Manipulation.. ii.

(4) 摘要. 網路市集中，網路評價系統扮演著重要角色在於買、賣雙方之間的信任建立，消費者往往仰賴網路評價系統協助購買決策的制訂，因此網路賣家必需認真看待買家給予的評價。然而，我們也不意外各式各樣的評價操弄行為開始出現於許多的網路市集中。僅管大多數操弄方式以隨機方式出現並且迅速的消失，其中一種具系統化且時常出現的操弄方式為賣家透過公告提醒事項試圖影響買家的評價，透過這些提醒事項去. 治政誘導買家留下正面評價或是禁止負面評價的產生。此研究提出一個研究設計用來檢視大立賣家的操弄行為對於評價系統之評價分數與評論取向的影響，而所提出研究架構將賣 ‧ 國. 學. 家的操弄行為分成三類分別為利誘、拒絕與無操弄。本研究選擇中國最大的網路市集. ‧. ─ 淘寶網用來衡量研究架構。. 研究結果顯示在賣家的操弄下，僅有評價分數會受到影響進而在三組間造成顯著. y. Nat. io. sit. 差異，然而，賣家的操弄與買家經驗的交互作用會影響評論取向與好評率，在賣家的. n. al. er. 操弄的三種類型中產生顯著差異。由此可知，買家經驗確實在賣家的操弄與買家評價. Ch. i n U. v. 之間有調節的作用。僅管研究結果顯示賣家的操弄確實影響了評價分數，但也更突顯. engchi. 出評論內容可能包含了更具價值的資訊，因此消費者不僅可以依據評價分數更可參考評論內容選擇賣家進行購買。關鍵字：網路評價系統、評價分數、評論取向、評價操弄. iii.

(5) Content Index 謝辭 .............................................................................................................................................I ABSTRACT ............................................................................................................................. II 摘要 ......................................................................................................................................... III CONTENT INDEX................................................................................................................. IV FIGURES ................................................................................................................................ VI. 治政大 TABLES ................................................................................................................................. VII 立 CHAPTER 1 INTRODUCTION............................................................................................. 1 ‧ 國. 學. CHAPTER 2 LITERATURE REVIEW ................................................................................. 6. ‧. 2.1 Online Marketplaces ..................................................................................................... 6. sit. y. Nat. 2.2 Evaluations and Evaluation Systems ............................................................................ 8. io. er. 2.3 Evaluation Manipulation ............................................................................................ 11. al. 2.4 Percentage of Positive Ratings and Comment Orientation ........................................ 17. n. v i n Ch 2.5 Buyer Experience ....................................................................................................... 18 engchi U. CHAPTER 3 RESEARCH METHOD ................................................................................. 20 3.1 Hypotheses and Research Model ................................................................................ 20 3.2 Data Collection and Sampling .................................................................................... 24 3.3 Measurement .............................................................................................................. 26 3.3.1 Research Variables and Definition .................................................................. 27 3.3.2 Measurement of Buyer Experience ................................................................. 27 3.3.3 Constructing a Preliminary Scale .................................................................... 28 iv.

(6) 3.4 Data Analysis .............................................................................................................. 29 3.5 Pretesting and initial item reduction ........................................................................... 31 CHAPTER 4 RESULTS ......................................................................................................... 35 4.1 Sampling Data ............................................................................................................ 35 4.1.1 Seller Evaluation Data ..................................................................................... 35 4.1.2 Coder Recruitment ........................................................................................... 37 4.2 Measurement Model ................................................................................................... 42. 政治大. 4.3 Hypothesis Testing...................................................................................................... 44. 立. 4.3.1 MANOVA and ANOVA analysis ..................................................................... 45. ‧ 國. 學. 4.3.2 ANOVA analysis results .................................................................................. 47. ‧. 4.3.3 PERCENTAGE OF POSITIVE RATINGS ..................................................... 47 4.3.4 COMMENT ORIENTATION ......................................................................... 51. y. Nat. io. sit. 4.4 Discussion................................................................................................................... 55. n. al. er. CHAPTER 5 CONCLUSION ............................................................................................... 59. Ch. i n U. v. 5.1 Conclusion and Management Implications ................................................................ 59. engchi. 5.2 Limitations and Future Research ................................................................................ 61 APPENDIX A .......................................................................................................................... 64 APPENDIX B .......................................................................................................................... 67 APPENDIX C .......................................................................................................................... 70 REFERENCES ....................................................................................................................... 72. v.

(7) Figures Figure 1. A seller’s evaluations on Taobao.com ....................................................................... 10 Figure 2. Seller manipulation behavior on Taobao.com ........................................................... 16 Figure 3. Research model ......................................................................................................... 23 Figure 4. Effect of the interaction on the percentage of positive ratings .................................. 50 Figure 5. Effect of the interaction on comment orientation ..................................................... 54. 政治大. Figure 6. Hypotheses testing results ......................................................................................... 55. 立. ‧. ‧ 國. 學. n. er. io. sit. y. Nat. al. Ch. engchi. vi. i n U. v.

(8) Tables Table 1. Common forms of reputation manipulation (Dini & Spagnolo, 2009) ...................... 14 Table 2. Convergent validity analysis results (initial results) ................................................... 32 Table 3. Convergent validity analysis results ........................................................................... 33 Table 4. Discriminate validity analysis results ......................................................................... 34 Table 5. Sellers’ categories on Taobao.com .............................................................................. 37. 政治大 Table 7. Demographic descriptive 立 statistics ............................................................................. 39 Table 6. Coder recruitment results ............................................................................................ 38. ‧ 國. 學. Table 8. Online shopping experience ....................................................................................... 41 Table 9. Item analysis ............................................................................................................... 43. ‧. Table 10. Analysis results for discriminate validity ................................................................. 44. Nat. sit. y. Table 11. Correlation analysis result ........................................................................................ 44. n. al. er. io. Table 12. Levene’s Test of Equality of Error Variancea........................................................... 46. i n U. v. Table 13. Levene’s Test of Equality of Error Variancea........................................................... 48. Ch. engchi. Table 14. Descriptive Statistics ................................................................................................ 48 Table 15. Two-way ANOVA: percentage of positive ratings ................................................... 49 Table 16. Multiple Comparisons: percentage of positive ratings ............................................. 50 Table 17. Levene’s Test of Equality of Error Variancea........................................................... 51 Table 18. Descriptive Statistics ................................................................................................ 52 Table 19. Two-way ANOVA: comment orientation ................................................................. 53 Table 20. Multiple Comparisons: comment orientation ........................................................... 54 vii.

(9) CHAPTER 1 INTRODUCTION. With the rapid growth of e-commerce, consumers are gradually becoming accustomed to online shopping, and the number of online shoppers is increasing year by year. According to a global online shopping survey across 10 countries, about 93% of consumers have. 政治大. purchased a product on the Internet (Pitney Bowes Inc., 2011). Furthermore, about 69% of. 立. consumers are reported to be highly capable of making online purchases (Maxwell &. ‧ 國. 學. Hudson, 2011). Consumers can now buy almost anything on the Internet simply by making. ‧. a few clicks, and online shopping has the convenience of not being limited by location or. Nat. io. sit. y. time. However, the lack of trust between sellers and buyers still interferes with consumers’. er. online transaction decisions (Ba & Pavlou, 2002; Utz, Matzat, & Snijders, 2009). Trust. al. n. v i n C htransactions, andUconsumers continue to be concerned remains a critical factor in electronic engchi about the credibility of sellers. Therefore, most online marketplaces use some form of feedback mechanism, such as an evaluation system, to promote mutual trust between sellers and buyers (Ba & Pavlou, 2002; Bolton, Katok, & Ockenfels, 2004; Resnick, Zeckhauser, Friedman, & Kuwabara, 2000).. Because of the information asymmetry between sellers and buyers, potential online 1.

(10) buyers tend to rely on evaluation systems and others’ feedback when making decisions (Ba & Pavlou, 2002; Gregg & Scott, 2006). Moreover, information asymmetry may bring about transaction disputes or even result in fraud, such as the non-delivery of purchased items, misrepresentation, and the sale of black market goods or triangulation (Gregg & Scott, 2006). Triangulation occurs when the seller buys a product from another seller or merchant. 政治大. using a stolen credit card or account and then resells it online (Gregg & Scott, 2006). In. 立. addition, the existing online evaluation systems have limitations, as they cannot resolve. ‧ 國. 學. problems such as anonymity, online identity and online fraud (Gregg & Scott, 2006;. ‧. Resnick et al., 2000). Sometimes, sellers offer cheap goods to attract buyers and accumulate. y. Nat. er. io. sit. positive ratings or negotiate with buyers so that they will not leave negative ratings. Moreover, some sellers retaliate against buyers who give negative ratings (Gregg & Scott,. n. al. Ch. engchi. i n U. v. 2006). Steiner (2003) confirms that about 19% of eBay users have received retaliatory feedback and about 16% have been a victim of feedback extortion. Online shoppers also face other retaliatory behavior, such as telephone harassment, the sending of threatening packages and the use of threats to induce the purchase of online goods or services. To avoid receiving threats from sellers, consumers can obtain relevant information from news media or online shopping forums and avoid leaving negative feedback when retaliation is likely. 2.

(11) Other sellers engage in a more systematic form of manipulation by constantly posting reminders to leave feedback in an attempt to influence buyer evaluation and to sustain good standing. Seller manipulation is also observed on sites such Taobao.com but also on well-known online marketplaces, such as eBay and Yahoo!. Today, the world’s most famous online marketplaces (e.g., eBay and Amazon.com). 政治大. have the best choice of consumers. As a result of the rapid growth of online shopping in. 立. ‧ 國. 學. China, the C2C online marketplace Taobao.com has become one of the most well-known online shopping websites in Asia. Taobao.com was launched in 2003 by the Alibaba Internet. ‧. business group, which was founded in 1999 (iResearch, 2012; Oberholzer-Gee & Wulf,. sit. y. Nat. io. n. al. er. 2009). In 2011, there were approximately 500 million registered users of Taobao.com and. i n U. v. the site received over 60 million unique visitors per day. The largest C2C online. Ch. engchi. marketplace in China, Taobao.com accounts for over 80% of the e-commerce transactions in China (iResearch, 2012). The website provides a platform for individual sellers and small businesses to sell items via auction or as a buy-it-now (BIN) fixed-price offer. Most sellers choose the BIN model by setting a fixed price and waiting for buyers to purchase the products directly (Ye, Li, Kiang, & Wu, 2009). Taobao.com also provides an online evaluation system to maintain trust between sellers and buyers. The evaluation system 3.

(12) enables consumers to check sellers’ rating scores, which are evaluated by buyers, and buyers’ comments. This information is critical for consumers to decide whether or not to purchase from a seller. Accordingly, sellers on Taobao.com tend to care about their rating scores, which are an accumulated value, as they provide clear evidence of their credibility.. A number of previous studies examine the issues relating to online reputation systems.. 政治大. For example, Ba and Pavlou (2002) investigate the effects of trust building technology.. 立. ‧ 國. 學. Bolton et al. (2004) explore the effectiveness of electronic reputation mechanisms, while Resnick et al. (2000) point out the relationship between reputation systems and trust. Gregg. ‧. and Scott (2006) emphasize the role of reputation systems in reducing online auction fraud.. sit. y. Nat. io. n. al. er. Zhou and Windle (2008) suggest the design of an effective reputation system for online. i n U. v. marketplaces. The extant research indicates that the potential negative characteristics of. Ch. engchi. sellers are hard to present in reputation systems (Resnick et al., 2000). Moreover, although online marketplace providers try to enhance the reliability of their evaluation systems by adopting credit card and real name authorization, evaluation manipulations still occur. As mentioned, most evaluation systems have difficulty deterring and disclosing these manipulations. Therefore, the evaluation system approach still has a number of limitations. In particular, manipulation of evaluation by sellers tends to create unreliable transaction 4.

(13) environments in which there are large amounts of positive feedback and few negative ratings (Gregg & Scott, 2006; Resnick & Zeckhauser, 2000).. The purpose of this study is to examine the effects of seller manipulation on buyer evaluations, namely, rating scores and text comments. Although rating scores can be easily influenced, buyers’ comments are not so easy to manipulate. Because ratings are cumulative,. 政治大. a higher score means a more reputable seller. Therefore, sellers tend to care more about their. 立. ‧ 國. 學. rating scores than comments. Rating scores are also used to reflect the reliability of the sellers on Taobao.com. However, Pavlou and Dimoka (2006) find that 97% of consumers. ‧. read buyers’ comments and 81% read at least 25 comments, because buyers’ comments. sit. y. Nat. io. n. al. er. contain more information concerning the quality of products than rating scores. The purpose. i n U. v. of this study is to examine the effects of seller manipulation on buyer evaluations, which. Ch. engchi. include rating scores and comments, and to examine the consistency between rating scores and comments.. 5.

(14) CHAPTER 2 LITERATURE REVIEW. 2.1 Online Marketplaces. Online marketplaces have become popular channels for buyers and sellers to transact anytime and anywhere (Post, Shah, & Mislove, 2011). In addition to the traditional B2B and. 政治大 B2C models, the advent of the Internet has enabled the development of the C2C transaction 立. ‧ 國. 學. model, which has grown to an unprecedented scale. However, sellers in the C2C. ‧. marketplace are typically unknown and are unrecognizable for potential buyers. Moreover,. sit. y. Nat. the items that they sell may not carry a brand name. Conversely, sellers have limited. n. al. er. io. assurance that buyers will comply with their payment obligations. Trust between buyers and. Ch. i n U. v. sellers is thus critical in C2C e-commerce. Online evaluation systems have been shown to. engchi. be relatively effective in inducing trust and promoting the growth of C2C e-commerce, as they provide reference information on sellers, which can help consumers to make purchase decisions (Ba & Pavlou, 2002; Bolton et al., 2004; Resnick et al., 2000).. The well-known online marketplaces, such as Amazon.com for B2C e-commerce and eBay and Taobao.com for C2C e-commerce, have enabled numerous people to establish 6.

(15) lucrative online businesses. The C2C e-commerce platform in particular provides an accessible transactional environment in which sellers can easily sell services or products that buyers can bid for or purchase directly. With the development of C2C e-commerce and the increasing number of online consumer purchases, it has become increasingly easy for businesses to engage in C2C e-commerce in online marketplaces. In this case, while the. 政治大. platform is C2C, the real business model is B2C. Recently, however, many individual. 立. sellers and small businesses have begun to trade on C2C type online marketplaces, such as. ‧ 國. 學. online auction sites, because of the lower cost of setting up a store front. More than simply. ‧. arenas for exchanging second-hand goods or online flea markets, online C2C marketplaces. y. Nat. er. io. sit. have become places where it is normal for businesses to sell items through online auctions. More and more sellers are now becoming professional fulltime online sellers. Due to the. n. al. Ch. engchi. i n U. v. unique business model provided by the C2C platform, sellers now have more incentive to manipulate online evaluation systems, because reputable sellers attract more buyers (Huang, Chen, & Lu, 2011; MacInnes et al., 2005). This study thus focuses on the evaluation system in an online marketplace that uses the C2C platform to investigate the problem of evaluation manipulation.. 7.

(16) 2.2 Evaluations and Evaluation Systems. Online purchasing differs from traditional retail transactions in that it occurs without face to face communication between sellers and buyers. In online marketplaces, buyers are given incomplete merchandise information through the seller’s description or a few. 政治大. photographs. Information asymmetry is an obvious problem in this case, as the buyer does. 立. not know exactly who the seller or what the commodity quality is. Similarly, the seller is. ‧ 國. 學. unable to clearly identify who the buyer is. In this situation, a lack of trust between the. ‧. transacting parties can affect whether a transaction will be completed. The trust between. Nat. io. sit. y. sellers and buyers is an important issue to bridge in C2C ecommerce because the brands and. er. the traders’ credibility tend to be unknown. Most online marketplaces provide public. al. n. v i n Ctrust evaluation systems to help build sellers and buyers. Resnick et al. (2000) h ebetween ngchi U define online evaluation systems as systems that collect, distribute and aggregate feedback about participants’ past behavior. The evaluations relate to consumers’ past interactions with producers (sellers) who provide services (goods) and are generated by the evaluation system for future reference by system users. Evaluations usually constitute positive and negative assessments of the quality of the product or service as determined by an individual customer 8.

(17) (Standifird, Weinstein, & Meyer, 1999; Simonson, 2001). A typical online marketplace evaluation method consists of buyer’s ratings and text comments, such as the evaluations presented on Taobao.com shown in Figure 1. Here, the website provides three types of rating for buyers to judge the transaction or seller: positive, neutral and negative. In Figure 1, the red flower represents a positive rating, the yellow represents a neutral rating and the. 政治大. black flower represents a negative rating. The evaluations are provided by everyone who. 立. has transacted with the seller, and the rating score is accumulated to represent the seller’s. ‧ 國. 學. credibility. Because the evaluation system collects evaluations about the sellers and directly. ‧. displays the information on the website, the system enhances the trading transparency, as. y. Nat. n. er. io. al. sit. anyone can review a seller’s credibility in a few clicks.. Ch. engchi. 9. i n U. v.

(18) 立. 政治大. ‧. ‧ 國. 學. n. er. io. sit. y. Nat. al. Ch. engchi. i n U. v. Figure 1. A seller’s evaluations on Taobao.com 10.

(19) Various studies have confirmed that evaluation systems can be of value when they are properly designed and managed (Dellarocas, 2003; Resnick et al., 2000; Pavlou & Dimoka, 2006). Sellers in online marketplaces accumulate credibility through evaluation systems and consumers put their faith in other buyers’ assessments of a seller’s credibility when making purchase decisions (Pavlou & Dimoka, 2006). Higher positive ratings increase sellers’ sales. 政治大. opportunities (Ba & Pavlou, 2002; Gregg & Scott, 2006). Evaluation systems can also. 立. disclose past transaction information through the evaluations of both buyers and sellers. In. ‧ 國. 學. addition, such systems help marketplaces to automatically eliminate untrustworthy sellers. ‧. by continuously accumulating rating scores. Overall, these systems enable buyers to avoid. y. Nat. er. io. sit. sellers who sell products of uncertain quality, thereby increasing the overall quality of sellers in online marketplaces (Akerlof, 1970). Therefore, effective evaluation systems are. n. al. Ch. engchi. essential for the sustainability of online marketplaces.. i n U. v. 2.3 Evaluation Manipulation. Because evaluation systems provide critical information for buyers’ purchase decisions, various forms of seller manipulation have emerged to encourage buyers to leave positive ratings (Huang et al., 2011; Melnik & Alm, 2002). Evaluation manipulation occurs 11.

(20) when sellers attempt to influence buyer evaluation and encourage or urge buyers to write favorable comments. Although buyer evaluations are not easily manipulated, opportunistic sellers still take various manipulation schemes. Moreover, the evidence from online marketplaces suggests that evaluation manipulation continues to occur (Dini & Spagnolo, 2009).. 政治大. In addition to acting as a kind of word-of-mouth testimony, online evaluations also. 立. ‧ 國. 學. help sellers to obtain the trust of buyers. Higher rating scores and favorable comments appeal to buyers. McDonald and Slawson (2002) point out that sellers’ evaluations are. ‧. important indicators for buyers in dynamic markets, as most buyers use them to identify. sit. y. Nat. io. n. al. er. trustworthy sellers to reduce their transaction risk (Ba & Pavlou, 2002; Gregg & Scott,. i n U. v. 2006). Li (2010) suggests that a seller’s credibility as indicated by an evaluation system. Ch. engchi. results in more buy-it-now transactions. Sellers’ evaluations also have positive effects on the final price and the probability of a successful transaction in online auctions (Melnik & Alm, 2002; Huang et al., 2011). Evaluations also affect sellers’ net revenue (Song and Baker, 2007). On Taobao.com and many other online marketplaces, sellers’ rating scores are cumulative, with each positive rating leading to an incremental increase and each negative rating a decrease. Although neutral ratings do not increase or reduce the score, they increase 12.

(21) the denominator in the fraction calculation, and thus dilute the positive ratings. Furthermore, negative evaluations always attract consumers’ attention, just as defects are easily spotted, which greatly influences buyers’ purchase intentions (Simonson, 2001). Sellers thus adopt different methods to avoid negative evaluations to maintain a good reputation.. Previous studies have found that online transaction disputes, customer complaints and. 政治大. manipulation behavior challenge the viability of evaluation systems (Dini & Spagnolo, 2009;. 立. ‧ 國. 學. Gregg & Scott, 2008; MacInnes et al., 2005). To avoid negative evaluations, some sellers disguise themselves as buyers and leave positive comments, while others urge buyers to. ‧. alter their evaluations by using intimidating language (Hu, Bose, Koh, & Liu, 2012, 2011;. sit. y. Nat. io. n. al. er. Hu, Liu, & Sambamurthy, 2011). Recently, online marketplaces have begun to attract more. i n U. v. and more manipulation behavior designed to interfere with buyer evaluations. As shown in. Ch. engchi. Table 1, common evaluation manipulations include unfair ratings, identity changes, multiple accounts, feedback theft, and purchase reputations (Dini & Spagnolo, 2009). Newcomers and unscrupulous sellers frequently sell commodities at a low price as a tactic to attract buyers and accumulate positive evaluations (Dellarocas, 2003; Resnick et al., 2000). However, some opportunistic buyers also use evaluation manipulation as a means to threaten sellers to get benefits. They may buy the cheapest thing and then intentionally leave 13.

(22) a negative rating to extort money from the seller. On Taobao.com this is called “malicious poor evaluators (差評師)” (VentureData.org, 2012). “Deleting evaluators (刪評師)” also exist who use methods to help sellers to delete their negative evaluations to obtain benefits (bbs.taobao.com, 2012). In this case, the evaluator may harass buyers until they delete their negative evaluations. Unscrupulous sellers manipulate buyers not only to garner higher. 政治大. evaluation scores, but also to conduct fraud.. 立. Table 1. Common forms of reputation manipulation (Dini & Spagnolo, 2009). 學. ‧ 國. Description. Form of manipulation. ‧. benefits. This tactic can be used to inflate the reputation of partners or destroy competitors.. y. Example:. sit. Sellers can sign up to secondary accounts or hire partners to make. io. transactions with the primary account (competitors) then leave positive. al. er. Nat. Unfair rating. A seller colludes with a group of buyers to be given ratings to gain. v. n. (negative) ratings to improve their reputation (destroy competitors).. Ch. i n U. Change the identity on the online marketplace.. Identity change. Example:. engchi. A seller abandons an identity with a bad reputation then creates a new “clean” identity and continues to sell products. Sellers use more than one account to inflate the reputation of the primary account or destroy competitors. Example:. Multiple accounts. A seller has more than one account and uses the secondary account to purchase from the primary account and leave positive ratings to improve their reputation on the primary account. The seller can also purchase from a competitor and leave a negative rating. 14.

(23) Description. Form of manipulation. A user takes control of another user’s account to gain benefits. Example:. Feedback theft. A user may hack into another person’s account and leave positive feedback to inflate their own reputation. Dishonest users can sell feedback in ‘fake’ or ‘shill’ auctions by exchanging money for a positive rating by making a fake transaction.. Purchase reputation (Shill auction). Example: The user sells a product named ‘feedback exchange’ or ‘new recipe gets positive feedback from all’ and then gives the buyer a positive. 政治大 These manipulation methods tend to involve spurious transactions that are quickly 立 rating.. ‧ 國. 學. completed. A systematic form of manipulation enacted by sellers is to constantly post. ‧. reminders in an attempt to influence buyer evaluation (as shown in Figure 2). These. sit. y. Nat. reminders include statements such as “give us full stars and you will get a discount,” “do. n. al. er. io. not leave negative or neutral evaluations without trying to resolve the problem with us,” and. Ch. i n U. v. “if you intend to leave a negative evaluation without contacting us first, please do not. engchi. purchase here.” The tone of the reminders ranges from prompting a positive rating to prohibiting a negative rating. Although the manipulations do not appear to be illegal, many online marketplaces restrict buyers’ comments and ratings. Therefore, this study focuses on evaluation manipulations that attempt to influence buyer evaluations by posting reminders. This type of manipulation can be categorized as either allurement or rejection. 15.

(24) 立. 政治大. ‧. ‧ 國. 學. n. er. io. sit. y. Nat. al. Ch. engchi. i n U. v. Figure 2. Seller manipulation behavior on Taobao.com 16.

(25) 2.4 Percentage of Positive Ratings and Comment Orientation. As discussed, sellers care about buyer evaluations, as most consumers base their purchase decisions on those evaluations (Ba & Pavlou, 2002; Huang et al., 2011; Li, 2010; Melnik & Alm, 2002; Song & Baker, 2007). The two must common evaluation methods are rating. 政治大. scores and text comments on subjective perceptions of transactions. Ratings include positive,. 立. neutral and negative, with the buyer choosing one rating to judge the seller or transaction.. ‧ 國. 學. Sellers usually pay more attention to rating scores than buyers’ comments, because higher. ‧. scores imply better credibility and are easily understood by buyers. Rating scores also serve. io. sit. y. Nat. as an assessment of a seller’s “level” on Taobao.com. Sellers desire to advance through. er. levels to achieve higher credibility. In this case, the accumulated rating score is tied to how. al. n. v i n C hbusiness in the marketplace. long the seller has been conducting As time passes, the rating engchi U score can accumulate to a high number, and thus the rating does not accurately reflect the seller’s current situation. Thus, most online marketplaces also display the percentage of positive ratings, as shown in Eq. 1, which is determined by dividing the number of positive ratings by the total number of ratings in a certain period, say six months. Neutral ratings obviously play an important role in this percentage calculation, because they are factored 17.

(26) into the total number of ratings.. Percentage of positive ratings= # of positive ratings ⁄ total # of ratings. (Eq. 1). Although text comments are not as direct or obvious as rating scores, many consumers read text comments. Pavlou and Dimoka (2006) found that 97% of consumers read buyers’ comments and 81% of consumers read at least 25 comments. Consumers read buyers’. 政治大 comments to gain more information about the merchandize and to better assess the 立. ‧ 國. 學. credibility of the seller. For instance, in many cases, buyers who give positive ratings write. ‧. comments to express their dissatisfaction with the transaction process. Inconsistencies. sit. y. Nat. between the orientation of a comment and the rating score can challenge buyers’. n. al. er. io. discernment. Although a few studies have examined comment orientation (Kusumasondjaja,. Ch. i n U. v. Shanka, & Marchegiani, 2012; Park & Lee, 2009), none have considered comment. engchi. orientation under the influence of seller manipulation.. 2.5 Buyer Experience. Consumers depend on their online experience to make all sorts of judgments in online marketplaces. Although consumers are familiar with purchasing online, different online 18.

(27) marketplaces have different purchase processes and consumers have to learn how to complete transactions on different websites. This is also the case with evaluations, as buyers rely on their past experience when making ratings and comments. Therefore, buyers have to understand the operating process and that the atmosphere of evaluation relies on their past experience. When buyers realize the importance of evaluations to sellers, they tend to take. 政治大. more caution with their evaluations. MacInnes et al. (2005) have indicated that experienced. 立. buyers tend to be involved in fewer disputes.. ‧ 國. 學. Because buyers’ online conduct is closely related to their online experience, how. ‧. buyers evaluate sellers can also be affected by buyers’ online shopping experiences. Gregg. sit. y. Nat. io. n. al. er. and Scott (2006) also found that experienced buyers are more capable of effectively. i n U. v. utilizing evaluation systems to evaluate sellers and avoid online fraud. Although these. Ch. engchi. studies do not directly show the relationship between buyer experience and buyer evaluation, they do confirm that buyer experience is an important moderator to consider in this study.. 19.

(28) CHAPTER 3 RESEARCH METHOD This study examines the effects of seller manipulation on buyers’ positive ratings and comments in online evaluation systems, using buyer experience as a moderator. It also investigates the consistency between the percentage of positive ratings and the orientation. 政治大. of the comments received by sellers.. 立. ‧ 國. 學. 3.1 Hypotheses and Research Model. ‧. This study seeks to examine the influence of seller manipulation on buyer evaluation.. Nat. io. sit. y. First, seller manipulation can affect a buyer’s evaluation of a seller so that it is in line with. er. what the seller expects. According to the literature, sellers care a great deal about their. al. n. v i n rating scores because it can affectCtransactions h e n g candhtheir i Urevenue (Huang et al., 2011; Melnik & Alm, 2002; Song & Baker, 2007). Therefore, if they want to survive in the online marketplace, sellers must get more positive ratings and maintain a good reputation. However, because evaluations are accumulated with each individual trade, it takes time and a large number of transactions to build a reputation. Hence, sellers take every evaluation seriously and may even try to manipulate their reputation. Past studies have also indicated 20.

(29) that sellers use various tricks to accumulate more positive ratings (Dellarocas, 2003; Dini & Spagnolo, 2009; Melnik & Alm, 2002). It can therefore be inferred that sellers will seek to manipulate buyers’ evaluations to get more positive ratings. In this study, manipulation refers to sellers posting reminders in an attempt to influence buyers’ evaluations. The manipulations are classified as allurement or rejection. Allurement manipulation means that. 政治大. a seller uses a reminder to offer a discount if a buyer leaves a positive rating. Rejection. 立. means that a seller seeks to reject a transaction if the buyer does not agree to leave positive. ‧ 國. 學. feedback. A contrast setting is also introduced whereby the seller does not set any. ‧. restrictions on manipulating its reputation to compare with the two types of manipulation. y. Nat. er. io. sit. behavior. Furthermore, the percentage of positive ratings is considered instead of the accumulated rating score, because the former is a more accurate indicator of a seller’s. n. al. Ch. engchi. i n U. v. current credibility. This leads to the following null hypothesis: H1: There is no significant relationship between seller manipulation and the percentage of positive ratings. Second, sellers’ behavior suggests that they are more concerned about positive ratings than positive comments. Hence, seller manipulation usually focuses on rating scores instead of text comments. However, an evaluation consists of a rating and a comment, and most 21.

(30) consumers will read buyers’ comments to help to make their decisions (Pavlou & Dimoka, 2006). For this reason, it is worth examining whether or not seller manipulation influences comments. Comment orientation is examined rather than comment quality, because consumers tend to quickly assess the orientation of comments rather than their quality, and a positive/negative orientation is likely to influence their purchase decision. The correlation. 政治大. between the comment orientation and the percentage of positive ratings under the influence. 立. of seller manipulation is also examined. It is assumed that buyers express what they really. ‧ 國. 學. perceive when writing comments, even under the influence of seller manipulation. This. ‧. leads to the following null hypothesis:. y. Nat. sit. H2: There is no significant relationship between seller manipulation and the orientation of. er. io. buyers’ comments.. al. n. v i n When making evaluations,Cbuyers rely on their past experience of online shopping. hengchi U Past studies indicate that experienced buyers are better able to use evaluation systems and avoid transaction disputes and fraud (Gregg & Scott, 2006; MacInnes et al., 2005). Because these buyers are familiar with the evaluating process, they tend to make careful evaluations. Therefore, buyer experience is likely to moderate the relationship between seller manipulation and buyer evaluation. 22.

(31) H3: Buyer experience does not moderate the relationship between buyer evaluation and seller manipulation. H3a: Buyer experience does not moderate the relationship between seller manipulation and the percentage of positive ratings. H3b: Buyer experience does not moderate the relationship between seller manipulation and the orientation of buyers’ comments. The research model for this study is depicted in Figure 3. The conceptual framework. 政治大. of the study centers on examining the effect of seller manipulation on the percentage of. 立. ‧ 國. 學. positive ratings and the orientation of buyers’ comments in the evaluation system, with buyer experience as the moderator. Seller manipulation is divided into the three categories. ‧. of allurement, rejection and none.. n. er. io. sit. y. Nat. al. Ch. engchi. i n U. Figure 3. Research model. 23. v.

(32) 3.2 Data Collection and Sampling. The proposed model is evaluated using data from Taobao.com, the largest online marketplace in China. Owing to the huge number of sellers on Taobao.com, a keyword search was used to prepare a list of about 200 sellers who had been operating on. 政治大. Taobao.com for more than six months. For the comment orientation evaluation, each seller’s. 立. most recent 25 buyer comments were captured from the evaluation system of Taobao.com. ‧ 國. 學. using the screen print function. Only the most recent comments are considered, as these are. ‧. what consumers are most likely to read. The number of comments captured corresponds to. al. er. io. sit. y. Nat. Pavlou and Dimoka’s (2006) finding that 81% of consumers read at least 25 comments.. n. The seller sampling choice process is as follows. First, Chinese keywords were drawn. Ch. engchi. i n U. v. from a Chinese character frequency survey (Ho, 1998) to search for the names and merchandise of sellers. The search was then progressively refined by adding more Chinese keywords and advanced search functions. This resulted in about 210 sellers from Taobao.com, which is close to what was expected. Finally, the samples were classified into three groups: allurement, rejection, and none, based on the tone of the reminders. Allurement is defined as a seller posting a reminder that offers a discount to encourage a 24.

(33) buyer to leave a positive rating. Rejection is defined as a seller posting a reminder that claims to reject or not provide a sale to a buyer who does not agree to provide a positive evaluation. None means that the seller does not display any significant manipulation behavior on the merchandise page. Twenty-five buyer comments were collected on each seller for the evaluation data. The comments were also directly coded as buyer experience. 政治大. according to the buyer level symbol displayed in the Taobao.com evaluation system. The. 立. average of the 25 buyer comments was used to represent buyers’ experiences of each seller.. ‧ 國. 學. For each seller, buyers’ comment orientation was coded by five coders who were. ‧. randomly sampled online. One thousand and fifty participants from an online shopping. sit. y. Nat. io. n. al. er. forum were invited to act as coders by responding to online web questionnaires. All of the. i n U. v. participants had online purchase experience and were familiar with the operation of online. Ch. engchi. marketplaces. Online participants rather than experts were invited to judge the comment orientation, because evaluation systems are an important indicator for consumers and function to induce trust between sellers and buyers. Consumers are the main reviewers of evaluation systems and what they perceive from the comments will directly influence their purchase decisions. Accordingly, online participants can be expected to easily judge the orientation of comments as they often make purchases online. A seven-point Likert scale 25.

(34) ranging from 1 ‘strongly disagree’ to 7 ‘strongly agree’ is used to measure the comment orientation.. Process of Evaluation in Taobao.com. After making a purchase, a buyer will wait for the package to be delivered. When the buyer receives the package he or she will need to respond with “received” and then proceed. 政治大 to evaluate the product. The buyer can then evaluate the shop and the trustworthiness of the 立. ‧ 國. 學. seller. The shop evaluation uses 5 stars to evaluate the degree to which the description. ‧. matches the product, the seller’s service attitude, the speed of the shipping and the speed of. sit. y. Nat. the delivery of the transaction. Buyers make their evaluations anonymously. The buyer. n. al. er. io. evaluates the trustworthiness of the seller by giving a positive, neutral, or negative rating. Ch. i n U. v. with text comments on each product. The buyer can choose to evaluate anonymously. In this. engchi. study, the data on product evaluation was extracted because the evaluation consists of both a rating score and text comments.. 3.3 Measurement. The measurement scale is based on the literature review and related research and then 26.

(35) modified to suit this study. The scale was developed to measure comment orientation because the measures used in prior studies did not suit the needs of this study. Buyer experience was measured by directly coding the symbols displayed in the evaluation system. The measure was then used as a surrogate for buyer experience, where a higher number means more experience.. 政治大 3.3.1 Research Variables and Definition 立. ‧ 國. 學. The operational definitions of the variables are described as follows. Sellers’ manipulation behavior: the type of seller evaluation manipulation, namely,. ‧. allurement, rejection and none.. y. sit. months.. Nat. Percentage of positive ratings: the percentage of a seller’s positive ratings in the last six. er. io. Comment orientation: the positive/negative orientation of comments (Kusumasondjaja et. al. n. v i n Cratings Buyer experience: The number of by each buyer is coded directly from the h e nreceived gchi U al., 2012).. corresponding symbols displayed in the evaluation system and then used as a surrogate for buyer experience. 3.3.2 Measurement of Buyer Experience. Because buyers also receive evaluations from sellers, a buyer’s experience was estimated according to the symbols displayed in the evaluation system, which represent 27.

(36) each buyer’s total number of ratings. This approach is used in existing research (MacInnes et al., 2005). Buyer evaluation ratings on Taobao.com can be considered as transaction records, with more transaction records indicating more experience. The 25 buyers’ evaluations were averaged and the mediating effect of this average experience examined.. 3.3.3 Constructing a Preliminary Scale. 立. 政治大. The web questionnaire for coders was divided to three parts: demographic data, online. ‧ 國. 學. purchase experience and comment orientation. In addition, a filtering question was used to. ‧. make sure that the coders had adequate online shopping experience to fill in the. sit. n. al. er. io Comment Orientation. y. Nat. questionnaire.. Ch. engchi. i n U. v. Buyers’ comments express what buyers feel about the transactions and the quality of the products. Various factors can affect the orientation of a comment. First, a seller’s service quality in the transaction process can directly affect a buyer. Taobao.com uses the TradeManager (AliWangwang) instant messenger application to facilitate direct communication between sellers and buyers in the form of “questions and answers.” The 28.

(37) communication process enables the buyer to perceive the seller’s attitude. Second, information quality is especially critical in online shopping due to information asymmetry. The only avenue for buyers to know about product quality is through the sellers’ descriptions. Numerous disputes in online marketplaces are caused by products not matching their description (MacInnes et al., 2005). Information quality also includes the. 政治大. promises that sellers put on the webpage. When a buyer receives a product, the most. 立. important issue to them is the quality of product. The final factor is whether buyers express. ‧ 國. 學. their satisfaction in text comments, which can be easily ascertained from the orientation of. ‧. the comments. A scale was developed to measure the orientation of comments based on four. y. Nat. er. io. sit. dimensions: service quality, information quality, product quality, and buyer satisfaction. For the service quality, information quality and buyer’s satisfaction measures, the measurement. n. al. Ch. engchi. i n U. v. items in Chong (2004) were adopted. Buyer satisfaction was identified through emotional expression (Kim & Gupta, 2012). The measurement of product quality used the items developed by Purohit and Srivastava (2001) and Zaichkowsky (1994).. 3.4 Data Analysis. To assure the quality of the scales, Hinkin’s scale development procedure was 29.

(38) followed. The SPSS17 and SamrtPLS2.0 (Ringle, Wende, and Will, 2005) software packages were used for the descriptive statistics, measurement model assessment, and model fit evaluation. The measurement model examines reliability and convergent and discriminant validity. Multivariate analysis of variance (MANOVA) was used to test the hypotheses. The consistency between the percentage of positive ratings and the comment orientation was also examined.. 立. 政治大. ‧ 國. 學. Descriptive statistical analysis was used to summarize the characteristics of the respondents. Cronbach’s α was used to measure the reliability. An alpha coefficient above. ‧. 0.7 indicates strong item covariance (Cortina, 1993). The validity measurement includes. sit. y. Nat. io. n. al. er. convergent and discriminate validity. The convergent validity of the items was tested using. i n U. v. average variance extracted (AVE), factor loading, and composite reliability. The value of the. Ch. engchi. AVE is expected to exceed 0.5 (Bagozzi & Yi, 1988), whereas the value of the composite reliability should exceed 0.7 (Fornell & Larcker, 1981). The set of observed variables is then reduced to a small set of variables according to the factor loading threshold; loadings of ±0.4 are considered more important and loadings greater than ±0.5 are considered to be of practical significance. Previous research suggests that the loading must exceed ±0.7 for a factor to account for 50 percent of the variance (Hair, Anderson, Tatham, and Black, 1998). 30.

(39) Therefore, the loadings were expected to exceed 0.7 and to be at least above 0.5 to make the variables representative of the factors. The discriminate validity was then tested. The Fornell-Larcker criterion requires a latent variable to share more variance with its assigned indicators than with any other latent variable. Based on this criterion, the square root of the AVE on the diagonal should be larger than the corresponding inter-construct correlations.. 政治大 3.5 Pretesting and initial item reduction 立. ‧ 國. 學. Participants were recruited using an Internet web questionnaire. All of the participants. ‧. had online shopping experience. The valid sample size was 34. Data were collected on each. y. Nat. al. er. io. sit. type of seller manipulation, namely, allurement, rejection and none. These data were coded. n. by at least ten coders to evaluate the comment orientation. The collected questionnaires were then analyzed.. Ch. engchi. i n U. v. First, the items’ convergent validity was tested using factor loading, AVE, and construct validity. The recommended value for factor loading is 0.5, and values above 0.7 indicate better loading (Hair et al., 1998), so the items below 0.7 were progressively deleted until the values of all the reserved items were above 0.7. The recommended values for AVE and composite reliability are 0.5 and 0.7, respectively (Bagozzi & Yi, 1988; Fornell & 31.

(40) Larcker, 1981). Cronbach’s α was used to test the reliability. The recommended value for Cronbach’s α is 0.7 (Cortina, 1993). The initial analysis results are shown as Table 2. Table 2. Convergent validity analysis results (initial results) Factor Construct/ Item loading Service quality (SQ) SQ1 SQ2 SQ3. 0.1. 0.33 -0.11 0.15 0.55 -0.00 -0.49 -0.04. ‧ 國. 立. 0.58. PQ 4 PQ 5. 0.53 0.84. Satisfaction (SAT) SAT 1 SAT 2 SAT 3 SAT 4. y. sit 0.76. 0.43. n. 0.21 0.66 0.82. 0.80. er. io. 0.40. PQ 1 PQ 2 PQ 3. delete delete. ‧. 0.93 0.77 0.83. al. delete. 0.84. Nat. Product quality (PQ). delete. 政治大. Information quality (IQ) IQ 1 IQ 2 IQ 3 IQ 4. Composite Cronbach’s Keep/delete reliability α item 0.87 0.03. 學. SQ4 SQ5 SQ6 SQ7. AVE. Ch. engchi U. v ni. delete 0.87 delete. delete 0.54. 0.30. 0.65. 0.93 0.21 0.52. delete. 0.13. delete. 32.

(41) Finally, nine items were deleted at the pretesting stage and the remaining items revised according to the feedback of the participants. After deleting the eight items, all of the items showed convergence and reliability (as shown in Table 3). Table 3. Convergent validity analysis results Construct/ Item Average Service quality (SQ). 立. 0.78 政治 0.79 大. 1.15 1.03 1.28. 0.72. ‧ 國. 1.28. 0.93. 4.24 4.97. 1.26 0.90. 0.77 0.84. Nat. io. 4.88 4.35 4.74. SAT 1 SAT 3. al. n. Satisfaction (SAT) 4.15 3.94. 1.07 1.10 0.90. Ch. 0.84 0.85 0.97. e n g c1.00 hi. 1.50 1.43. 0.88. 0.81. 0.92. 0.88. 0.86. 0.82. y. 0.79. Product quality (PQ) PQ 2 PQ 3 PQ 5. ‧. 4.76. 學. IQ 2 IQ 3. Composite Cronbach’ reliability sα 0.85 0.77. 0.88. Information quality (IQ) IQ 1. AVE 0.66. sit. 5.32 4.82 4.76. Factor loading. er. SQ 1 SQ 3 SQ 4. Standard error. i v0.76 n U. 0.72. The discriminate validity was then tested. Based on the Fornell-Larcker criterion (Fornell & Larcker, 1981), the square root of the AVE on the diagonal should be larger than the corresponding inter-construct correlations. Table 4 lists the correlations among the 33.

(42) constructs, with the square root of AVE on the diagonal. All of the diagonal values are larger than the corresponding inter-constructor correlations. Table 4. Discriminate validity analysis results Construct. Service quality. Service quality Information quality Product quality Satisfaction. 0.814 0.583 0.356 0.228. 立. Information quality. Product quality. Satisfaction. 0.888 0.300. 0.872. 0.848 0.628 0.238. 政治大. ‧. ‧ 國. 學. n. er. io. sit. y. Nat. al. Ch. engchi. 34. i n U. v.

(43) CHAPTER 4 RESULTS. 4.1 Sampling Data. In this study, seller evaluation data was collected and coders recruited to evaluate the orientation of the comments. This section describes the process for collecting the two kinds. 立. 學. ‧ 國. of data.. 政治大. 4.1.1 Seller Evaluation Data. ‧ sit. y. Nat. A keyword search on Taobao.com was used to collect the sellers’ evaluation data. The. n. al. er. io. keywords were chosen randomly from the website of a Chinese character frequency survey. Ch. i n U. v. (Ho, 1998). The following steps were then used to collect the seller evaluation data. First,. engchi. the randomly chosen keywords were used to search the website and keywords were added until the results matched expectations. The advanced search tools, “combine same seller” and “search shop identity”, provided by Taobao.com were then applied to narrow the results and to avoid collecting the same seller repeatedly. Second, only online shops that had operated for at least six months were included in the seller list. Third, the seller 35.

(44) manipulation was classified as either allurement, rejection, or none based on the reminders the sellers put on the product pages. Pavlou and Dimoka (2006) found that buyers usually read up to 25 comments, therefore the 25 most recent comments were collected for each seller sampled. Buyer experience was coded directly from the symbols displayed in the evaluation system, which show the amount of ratings a buyer has accumulated. The symbols. 政治大. represent from zero to twenty levels of buyer experience. The buyers’ 25 comments were. 立. then averaged to represent buyers’ experience of each seller. The averaged buyer experience. ‧ 國. 學. was classified as either high or low based on the median of the buyer experience. The. ‧. collected sellers’ evaluation data comprised the percentage of positive ratings in the last six. y. Nat. er. io. sit. months and the 25 most recent comments. The seller evaluation data were coded by five coders to evaluate the comment orientation.. n. al. Ch. engchi. i n U. v. Table 5 presents the categories of the items sold on Taobao.com. The main categories are clothing, shoes, and bags (37.7%). The categories are decided by the sellers from a list of 15. The details of each category are given in Appendix C, which is a summary from Taobao.com.. 36.

(45) Table 5. Sellers’ categories on Taobao.com Category. Frequency. Ratio (%). Jewelry and Accessories. 4. 2.0%. Sports and Outdoors. 4. 2.0%. Books, Music and Videos. 5. 2.5%. Games and Phone Charges. 5. 2.5%. Automotive. 7. 3.4%. Collectibles and Entertainment. 7. 3.4%. Home Improvement. 7. 3.4%. 8 治政 Household Products 11 大 Beauty and 13 立Skin Care. Food and Health Care. Clothing, Shoes and Bags. 77. Other. 0. TOTAL. 204. Nat. io. n. al. 4.1.2 Coder Recruitment. 8.8% 10.3% 37.7% 0.0%. y. 21. sit. Mother and Baby. 8.3%. 100.0%. er. 18. ‧ 國. Living Services. 6.4%. ‧. 17. 5.4%. 學. Electronics. 3.9%. Ch. engchi. i n U. v. The seller data were coded by five coders. More than 1,000 coders were recruited from a population that had previously purchased products or services online. The coders were invited to read a list of 25 comments and then make a judgment on the orientation of the comments. 37.

(46) The coders were recruited through a web questionnaire and links posted on the PTT Bulletin Board System (BBS) and Facebook.com. Details of the questionnaires are given in Appendix B. Because the aim was to recruit coders with online shopping experience, the first question of the questionnaire screened the coders to ensure that they were all experienced online shoppers. The questionnaire link was posted on the online shopping. 政治大. board on PTT BBS and on the author’s personal Facebook.com page, in the hope that. 立. Facebook.com friends would share the link, because over 1,000 coders were needed to. ‧ 國. 學. participate in the study. Gifts were also offered to attract coders. It took about one month to. ‧. gather the data from 1,090 coders. Any questionnaires that were incomplete or contained the. y. Nat. er. io. sit. wrong answer to a verified question that examined whether the coder read the comments seriously were rejected. This process left 1,052 coders to participate in the study, a valid. n. al. Ch. engchi. i n U. v. response rate of 96.5%. All of the coders had online shopping experience. For each question, the coders were asked to choose the response that best described their level of agreement. Table 6. Coder recruitment results Number. Ratio (%). Overall coders Invalid questionnaires. 1090 38. 100 3.5. Valid questionnaires. 1052. 96.5. 38.

(47) The sample comprised 788 (74.9%) females and 264 (35.1%) males. Most of the coders were Taiwanese, because the questionnaire link was posted on PTT BBS, which is mainly used by people from Taiwan. The majority of the coders were aged between 15 and 30 years and had a bachelor or Master’s degree. The demographic statistics of the sample are summarized in Table 7.. 政治大 Option Frequency. Table 7. Demographic descriptive statistics. ‧ 國 Nat. 0.4. Others. 3. 0.3. Male. 264. 25.1. Female. 788. 74.9. 7. 0.7. 256. 24.3. sit. 15~20. 687. 65.3. 84. 8. 14. 1.3. 51~55. 2. 0.2. Upon 56. 2. 0.2. Junior high school or below. 10. 1. Senior high and vocational school. 74. 7. Bachelor. 745. 70.8. Master’s. 211. 20.1. PhD. 12. 1.1. n Education. 4. er. io Age. China. Under 15. al. 99.3. ‧. Gender. 1045. 學. Nationality. Taiwan. Ratio (%). y. 立. Measure. 21~30. C31~40 hengchi U 41~50. 39. v ni.

(48) Table 8 summarizes the online shopping experience of the coders. The first question of the questionnaire was designed to ensure that all of the coders were experienced online shoppers. Sixty-eight percent of the coders were somewhat familiar with Taobao.com and had browsed the website’s evaluation system. Some had purchased items from Taobao.com. Moreover, a few of the coders (2.2%) were also sellers on Taobao.com. Most of the coders. 政治大. chose the Yahoo! auction website when making online purchases (57.7%) and nearly all. 立. indicated that they read evaluations before purchasing (98%). In particular, 87.3% of the. ‧ 國. 學. coders usually read the rating scores and text comments and 92% read the text comments.. ‧. Most searched for information before purchasing (94.3%), and the majority often referred to. y. Nat. n. er. io. al. sit. the e-shopping board of BBS PTT (57.1%).. Ch. engchi. 40. i n U. v.

(49) Table 8. Online shopping experience Measure Option Frequency Internet experience 0~3years 38 4~7years 169 8~11years 427 12~15years 275 more than 15 years 143 0~3years 334 4~7years 488 Experience of online 8~11years 189 shopping 12~15years 34 more than 15 years 7 Browsed the Taobao.com No 339 evaluation page Yes 713 No 449 Under 1 year 328 2 years 179 3 years 59 Experience of shopping on 4 years 15 Taobao.com 5 years 15 6 years 2 7 years 5 8 years 0 More than 9 years 0 No 1029 A seller on Taobao.com Yes 23 Taobao.com 192 Yahoo! Auction 607 Website most often used for PC Home 111 purchases Tmall.com 0 360buy.com 0 Others 142 No 25 Only scores 60 Browsing evaluation habit Only comments 49 Both scores and comments 918 No 275 PTT e-shopping 601 Online shopping forum meilishuo.com(美丽说) 0 often browsed mogujie.com(蘑菇街) 7 Mobile01 145 Others 24 Search information before No 60 online shopping Yes 992. 立. 政治大. ‧. ‧ 國. 學. n. er. io. sit. y. Nat. al. Ch. engchi. 41. i n U. v. Ratio (%) 3.6 16.1 40.6 26.1 13.6 31.7 46.4 18 3.2 0.7 32.2 67.8 42.7 31.2 17 5.6 1.4 1.4 0.2 0.5 0 0 97.8 2.2 18.3 57.7 10.6 0 0 13.5 2.4 5.7 4.7 87.3 26.1 57.1 0 0.7 13.8 2.3 5.7 94.3.

(50) 4.2 Measurement Model. Item analysis was conducted using the SmartPLS2.0 (Ringle, Wende, & Will, 2005) and SPSS 17 software packages to assess the reliability and validity of the items/constructs. Cronbach’s α was used to measure reliability. As shown in Table 9, the Cronbach’s α values. 政治大. were all greater than 0.7 (Cortina, 1993). Thus, the constructs used in this study had high. 立. levels of reliability and consistency. The convergent validity was then tested using the AVE,. ‧ 國. 學. factor loading, and composite reliability. The value of the AVE exceeded 0.5 (Bagozzi & Yi,. ‧. 1988) and the composite reliability exceeded 0.7 (Fornell & Larcker, 1981). One item’s. Nat. io. sit. y. factor loading value was 0.68 and lower than the threshold (greater than 0.7) (Hair et al.,. er. 1998). However, because it was close to 0.7 and greater than 0.5, the loading is considered. al. n. v i n C h Besides, there are partially significant (Hair et al., 1998). e n g c h i U only two items in the construct for buyer satisfaction, and thus the item was retained to make the construct more comprehensive and to ensure that its inclusion would not affect the validity of the construct. Consequently, all of the variables in the study had convergent validity.. The discriminate validity was then tested based on the Fornell-Larcker criterion (Fornell & Larcker, 1981) that the square root of the AVE on the diagonal should be larger 42.

(51) than the corresponding inter-construct correlations. Table 10 lists the correlations among the constructs, with the square root of the AVE on the diagonal. All of the diagonal values are larger than the corresponding inter-constructor correlations, and this condition is satisfied in all cases. Table 9. Item analysis Standard Factor Construct/ Item Average error loading AVE 0.847 Service quality (SQ). 立 1.214 1.203 1.233. ‧ 國. 5.00 5.00 4.91. 0.911 0.902 0.948. 0.913. 0.909. sit. al. 1.287 1.275 1.241. Ch. 0.803 0.873 0.965. engchi. 0.825. 0.900. 5.48. 1.244. 0.978. 5.52. 1.207. 0.680. 43. er. 0.779. n. 5.09 5.05 4.56. 0.868 0.916 0.861. y. 1.299 1.234 1.119. Buyer satisfaction (SAT) SAT 1 SAT 2. 0.861. ‧. io. PQ 1 PQ 2 PQ 3. 4.82 4.63 5.09. Nat. Product quality (PQ). 0.913. 0.778. Information quality (IQ) IQ 1 IQ 2 IQ 3. Cronbach’s α 0.914. 學. SQ 1 SQ 2 SQ 3. 政治大. Composite reliability 0.943. i n U. v. 0.709.

(52) Table 10. Analysis results for discriminate validity Construct Service quality Information Product quality quality Service quality 0.921 0.795 Information quality 0.882 0.669 0.762 Product quality 0.883 0.558 0.565 0.603 Satisfaction. Satisfaction. 0.842. A correlation test was conducted before the hypothesis testing to examine the correlation between the percentage of positive ratings and the orientation of the comments. 政治大. under the influence of seller manipulation. As presented in Table 11, there was a low. 立. ‧ 國. 學. correlation (p<0.01) between the percentage of positive ratings and comment orientation under the influence of seller manipulation. According to this result, a buyer might give a. ‧. io. sit. Table 11. Correlation analysis result 1. n. al. Ch. n U i e n g c h.169**. 1. Percentage of positive ratings 2. Comment orientation. er. Nat. y. positive rating, but not necessarily write a positive comment.. iv. 2 .169**. **P<.01. 4.3 Hypothesis Testing. Multivariate analysis of variance (MANOVA) was initially used to test the hypotheses. Because Levene’s test of homogeneity was significant, the MANOVA was replaced with ANOVA, as explained in detail in the next section. 44.

(53) 4.3.1 MANOVA and ANOVA analysis. A 3 X 2 (three level of seller manipulation and two level of buyer experience) multivariate analysis of variance (MANOVA) was used to test the effect of seller manipulation behavior on buyer evaluations. Because MANOVA can compare several. 政治大. criteria variables simultaneously across groups to assess the effect of a set of independent. 立. variables, it is an appropriate analytical technique for this study. However, Levene’s test of. ‧ 國. 學. homogeneity in the percentage of positive ratings was significant, which violates the. ‧. assumption of MANOVA. A check of the data revealed that there was an outlier in the seller. Nat. io. sit. y. manipulation data that had a much lower positive rating than the others. According to Hair. er. et al. (1998), MANOVA is especially sensitive to outliers, which have a disproportionate. al. n. v i n C h researchers are effect on the overall results. Therefore, e n g c h i U strongly encouraged to eliminate outliers from their analysis. After deleting the outlier, a total of 203 seller evaluation data and 1046 valid questionnaires remained. However, Levene’s test of homogeneity was still significant (as shown in Table 12) even after deleting the outlier. Therefore, MANOVA could not be used to examine the hypotheses, and the analysis method was changed such that the hypotheses had to be tested separately. 45.

(54) a. Table 12. Levene’s Test of Equality of Error Variance Dependent variable. F. df1. df2. Sig.. Percentage of positive ratings. 30.099. 5. 1040. .000. Comment orientation. 1.010. 5. 1040. .410. Tests the null hypothesis that the error variance of the dependent variable is equal across groups. a. Design: Intercept +Seller Manipulation +Buyer Experience +Seller Manipulation*Buyer Experience. 政治大. A two-way analysis of variance (ANOVA) was used in place of the MANOVA test,. 立. even though the problem of homogeneity still exists and violates the assumption of ANOVA.. ‧ 國. 學. However, past research indicates that “F tests in ANOVA are robust with regard to these. ‧. assumptions except in extreme cases” (Hair et al., 1998, p. 347). Two-way ANOVA, which. y. Nat. al. er. io. sit. is an extension of one-way ANOVA, is used when there is more than one independent. n. variable and there are multiple observations for each independent variable. In addition to. Ch. engchi. i n U. v. determining the main effects of the contributions of each independent variable, the two-way ANOVA can determine if there is a significant interaction effect between the independent variables. Therefore, two-way ANOVA was used to test the hypotheses in two separate stages: (1) the influence of seller manipulation on the percentage of positive ratings as mediated by buyer experience and (2) the influence of seller manipulation on comment orientation as mediated by buyer experience. 46.

(55) 4.3.2 ANOVA analysis results. Each recruited coder was randomly assigned one of the three types of seller manipulation and was asked to judge the comment orientation. A 3 X 2 (three level of seller manipulation and two level of buyer experience) MANOVA was used to test for the effect. 政治大. of seller manipulation on buyer evaluations. However, Levene’s test of homogeneity was. 立. significant with regards to the percentage of positive ratings (as shown in Table 12).. ‧ 國. 學. Therefore, the MANOVA test could not be used to examine whether the interaction of seller. ‧. manipulation and buyer experience affected buyer evaluation. A two-way ANOVA test was. Nat. io. sit. y. then conducted in place of the MANOVA to examine the effect of the interaction of seller. n. al. er. manipulation and buyer experience on the percentage of positive ratings and comment orientation separately.. Ch. engchi. i n U. v. 4.3.3 PERCENTAGE OF POSITIVE RATINGS. First, the effect of seller manipulation on the percentage of positive ratings as mediated by buyer experience was examined. Although Levene’s test of homogeneity was significant (as shown in Table 13), past research indicates that ANOVA is robust even when 47.

(56) the equal variation assumption is violated, as long as the group sizes are equal (Hair et al., 1998). a. Table 13. Levene’s Test of Equality of Error Variance Dependent variable. F. df1. df2. Sig.. Percentage of positive ratings. 30.099. 5. 1040. .000. Tests the null hypothesis that the error variance of the dependent variable is equal across groups. a. Design: Intercept +Seller Manipulation +Buyer Experience +Seller Manipulation*Buyer Experience. 治政 Table 14 shows the descriptive statistics of the data, 大which comprise 358 allurement, 立 ‧ 國. 學. 341 none, and 347 rejection seller evaluation manipulations. The mean percentage of the positive ratings in the none group was higher than in the other groups.. n. al. Allurement. None. Rejection. Total. Mean. y. sit. io. Seller manipulation Buyer experience. Std. Deviation. er. Nat. Dependent Variable: Percentage of positive ratings. v i n CHigh U h e n g c h99.3202 i Total 98.8405 Low. 98.5466. ‧. Table 14. Descriptive Statistics. N. 1.07439. 222. .72754. 136 358 141. Low. 99.4556. 1.02767 .58559. High Total Low High Total. 99.6172 99.5504 99.0562 99.5878 99.3151. .44821 .51490 .92927 .46897 .78697. 200 341 178 169 347. Low High Total. 98.9512 99.5273 99.2294. .99082 .55715 .86022. 541 505 1046. 48.