1944 Exploring Decision Rules for Sellers in Business-to-Consumer (B2C) Internet Auctions business-to-consumer (B2C) auctions (Bapna, Goes, & Gupta, 2001). In B2C auctions, large mer- chants such as Dell, Disney, Home Depot, IBM, Motorola, Sears, Sun Microsystems, and Sharper Image have been able to use Internet auctions to VHOOH[FHVVLQYHQWRU\IRUJUHDWHUSUR¿WWKDQWKH\ would receive from using a liquidator (Dholakia, 2005b; Gentry, 2003; Grow, 2002; Vogelstein, Boyle, Lewis, & Kirkpatrick, 2004). As further evidence of the growth of B2C Internet auctions, E\WKH¿UVWTXDUWHURI,QWHUQHWDXFWLRQHHU eBay alone hosted approximately 383,000 eBay stores worldwide, including 171,000 on Web sites other than their U.S. Web site (eBay, 2006). $V¿UPVFRQWLQXHWRPDNHH[WHQVLYHXVHRI,Q- ternet auctions, the interest in developing sound guidelines for businesses as well as developing theory to advance research will likely continue to grow as well. While many studies have examined the factors W K DWGH W HU P L Q HDQD X FWLR QLWHP¶V ¿ Q DOE LGS U LFH W KH number of bids an item receives, whether a sale is completed, or the revenue earned by a seller, the examination of price premiums (above-aver- DJH¿QDOELGSULFHVLVUHODWLYHO\XQGHUVWXGLHG,Q HFRQRPLFVSULFHSUHPLXPVDUHGH¿QHGDVSULFHV WKDW\LHOGDERYHDYHUDJHSUR¿WV.OHLQ/HI ÀHU 1981; Shapiro, 1983). Price premiums within the ,QWHUQHW DXFWLRQ FRQWH[W KDYH EHHQ GH¿QHG DV “the monetary amount above the average price received by multiple sellers for a certain match- ing product” (Ba & Pavlou, 2002, pp. 247-248). Restated, a number of auctions exist where sell- ers have earned above-average prices, or price premiums, on the items they have auctioned. In this study, we compare the group of auctions that have achieved above-average prices with those that KDYHQRWWRREVHUYHVLJQL¿FDQWGLIIHUHQFHV7R our knowledge, only two studies have previously examined price premiums (Ba & Pavlou, 2002; Pavlou, 2002). Since it is only by maximizing UHYHQXHDQGSUR¿WWKDWD¿UPFDQUHPDLQYLDEOH in the marketplace (Seth & Thomas, 1994), an increased focus on how businesses that rely upon Internet auctions can earn price premiums may SURYHEHQH¿FLDO7KHIRFXVRQSULFHSUHPLXPVLV WKH¿UVWFRQWULEXWLRQRIWKLVVWXG\$VZHLQYHV- tigate price premiums, we examine many of the independent variables that have been considered in previous studies to determine if they are also predictive of price premiums. The second con- tribution is the application of CART analysis to Internet auctions as a tool to generate decision rules. CART analysis is a tree-based method of recursive partitioning for explaining or predict- LQJDUHVSRQVHWRRUGHUYDULDEOHVE\VLJQL¿FDQFH (Brieman, Friedman, Olshen, & Stone, 1984). It generates decision trees and decision rules that may be used as guidelines (by sellers in Internet auctions, in this case). While electronic commerce r e s e a r c h h a s d e m o n s t r a t e d t h a t C A R T a n a l y s i s c a n be used to improve one-to-one Internet market- ing (Kim, Lee, Shaw, Chang, & Nelson, 2001), CART has not yet been applied to Internet auc- tions. Thus, our study is, to our knowledge, the ¿UVWWRXVHDVWDWLVWLFDOO\EDVHGGHFLVLRQPDNLQJ technique to demonstrate how sellers can use quantitative data to decide how to sell products LQ %& ,QWHUQHW DXFWLRQV 7KH WKLUG DQG ¿QDO contribution of this study is the examination (by CART analysis) of variables that have been found (generally by multiple-regression analysis) to be determinants of auction outcome in previous VWXGLHV7KLVFRQ¿UPDWLRQRIYDULDEOHVLGHQWL¿HG as critical factors in other types of analysis is the third contribution of this study. The article will be organized as follows. We begin by reviewing literature on auctions, includ- ing relevant research on both traditional auctions as well as Internet auctions. Next, we present literature on machine-learning techniques that enable the induction of decision trees. Following the literature review, we discuss our methods, including our dataset, variables, and our research GHVLJQ6SHFL¿FDOO\ZHGHVFULEH WKHFROOHFWLRQ DQGDQDO\VLVRI¿HOGGDWDIURP,QWHUQHWDXFWLRQHHU eBay. We then present the results of our analysis. Following the presentation of our results, we 1945 Exploring Decision Rules for Sellers in Business-to-Consumer (B2C) Internet Auctions GLVFXVVRXU¿QGLQJVDQGQRWHWKHLPSOLFDWLRQVRI RXUVWXG\)LQDOO\ZHFRQFOXGHE\EULHÀ\QRWLQJ the limitations of our study and directions for future research. LITERATURE REVIEW Literature pertinent to this study will be selectively drawn from two areas of research. Given that one of the objectives of this study is to investigate factors enabling sellers to earn price premiums LQ ,QWHUQHW DXFWLRQV WKH ¿UVW DUHD IURP ZKLFK we draw theory is that of auction literature. An additional objective—namely, describing a tech- nique for developing decision rules for sellers in Internet auctions—leads us to the second area of research that is pertinent to the present study: decision-tree induction techniques. Auctions $XFWLRQVKDYHEHHQGHVFULEHGDV³a market in- stitution with an explicit set of rules determining resource allocation and prices on the basis of bids from the market participants” (McAfee & McMillan, 1987, p. 701). A vast amount of research addresses the topic of auctions. Numerous surveys of auction literature can be found (Engelbrecht- Wiggans, 1980; Klemperer, 1999, 2000; Krishna, 2002; McAfee & McMillan, 1987; Milgrom, 1985, 1986; Rothkopf & Harstad, 1994; Wilson, 1987), including a bibliography of earlier literature (Stark & Rothkopf, 1979) and a review of experimental auction literature (Kagel, 1995). Auction Mechanisms and Auction Theory Auction mechanisms are generally categorized as: (1) English or ascending-price auctions; (2) 'XWFKRUGHVFHQGLQJSULFHDXFWLRQV¿UVWSULFH sealed-bid auctions; or (4) second-price sealed bid or Vickrey auctions (McAfee & McMillan, 1987). A thorough description of these mechanisms can be found in the recent work of Lucking-Reiley (2000a). Internet auctions on eBay, the point of data collection for this study, have been described by scholars as a hybrid of the English and second- price auctions (Lucking-Reiley, 2000a, 2000b; Ward & Clark, 2002; Wilcox, 2000). Researchers assert that eBay uses a hybrid auction type on the grounds that the presence of a proxy-bidding mechanism ensures that a winning bidder will pay only one increment more than the second-highest bidder’s price. Since this study examines only auctions of the hybrid eBay type, a discussion of how various types of auction mechanisms impact auction outcome is beyond the scope of the present study. Auction theory is often centered around or developed in response to the seminal work of William Vickrey (1961), who described the In- dependent Private Values Model (IPV). In this model, each bidder formulates a valuation for the item being auctioned without an awareness of competing bidders’ valuations. Even if valuations were shared among all bidders, each individual bidder’s valuation would be unaffected by the additional information that competing bidders’ v a l u a t i o n s w o u l d p r o v i d e . I n t h i s w a y, t h e b i d d e r ’s YDOXHLVLQGHSHQGHQWRIWKHLQÀXHQFHRIFRPSHWLQJ bidders and is privately determined. In contrast, the Common Values Model (CV) posits that the value of the item being auctioned is common to all bidders, but incomplete information causes each bidder to formulate a valuation for the item that falls either above or below the common value (Rothkopf, 1969; Wilson, 1969). If it is assumed that bidders’ valuations are normally distributed about the common value, the winner of the auc- tion is the bidder with the valuation that is farthest above the common value. This person incurs the ³ZLQQHU¶V FXUVH´EHFDXVH KH RU VKHKDVOLNHO\ overpaid for the item. An integrative approach, UHIHUUHGWRDVWKH$I¿OLDWHG9DOXHV0RGHO$9 explains that bidder valuations depend upon the bidder’s personal preferences, the preferences of 1946 Exploring Decision Rules for Sellers in Business-to-Consumer (B2C) Internet Auctions others, and the intrinsic qualities of the item being sold (Milgrom & Weber, 1982). Bidders’ valu- DWLRQVDUHGHVFULEHGDVDI¿OLDWHGEHFDXVHDKLJK valuation by one bidder makes a high valuation by other bidders more likely (Milgrom & Weber, 1982). The AV model is a more general concep- tualization of the valuation of items in auctions than the IPV or CV models; both the IPV and CV models can be understood as special cases of the more general AV model (McAfee & McMillan, 1987). Recent studies of Internet auctions rely upon and explicitly mention the merits of the AV model (Dholakia & Soltysinski, 2001; Gilkeson & Reynolds, 2003; Segev, Beam, & Shanthikumar, 2001; Wilcox, 2000). These studies empirically validate the AV model in Internet auctions by GHPRQVWUDWLQJ WKDW ELGGHUV PD\ EH LQÀXHQFHG not only by their own valuation of the item, but also by the behavior of other bidders. Internet Auctions Internet auctions have a relatively brief history. Among the earliest electronic auctions were the auctioning of pigs in Singapore (Neo, 1992) and ÀRZHUV LQ +ROODQG YDQ +HFN YDQ 'DPPH 1997) conducted over a LAN. Auctions on the Internet, conducted via newsgroups and e-mail discussion lists, were the next major development in the Internet auction timeline (Lucking-Reiley, 1999, 2000a). The explosion in popularity of In- ternet auctions, however, did not begin until the 1995 launches of U.S. Web sites Onsale and eBay (Lucking-Reiley, 2000a). By 1999, there were an estimated 200 auction sites on the Internet (Crockett, 1999). The continued growth of Internet auctions is demonstrated by the performance of international industry leader eBay, a company that operates auction Web sites in 24 countries, includes over 180 million registered users, and generated US$ 4.552 billion in sales in 2005 (eBay, ,QWHUQDWLRQDOFRPSHWLWLRQLQFOXGHV¿UPV s u c h a s QX L . c o m i n E u ro p e , Ta o b a o .c o m i n A s i a , and MercadoLibre in Latin America. Following 0|OOHQEHUJSSZHZLOOGH¿QH Internet auctions to mean virtual marketplaces relying on Internet services (such as the World Wide Web) and Internet protocols to conduct auctions. In spite of the relatively short history of Internet auctions, they have begun to draw interest not only from economists, but also from researchers in marketing, information systems, and computer science (see Appendix A for a selective listing of recent studies in each of these disciplines). The general questions that many of these studies seek WRDQVZHUDUH³:KDWLVWKHRSWLPDOZD\WRDXFWLRQ DQLWHP"´RU³+RZLVWKHPDUNHWSODFHFKDQJLQJ DVDUHVXOWRI,QWHUQHWDXFWLRQV"´RU³:KDWIDF- tors should be considered when buying or selling in an Internet auction?” We will generally limit our discussion of Internet auctions to empirical studies that deal with variables that are under the control of the seller (rather than variables under the control of the other two parties to the auction transaction, the auctioneer and the bidder). Since this study focuses on developing decision rules for sellers in single-item B2C Internet auctions, we will reserve exploration of multi-unit auctions and buyer behavior for other researchers. To organize the list of variables that have been investigated in previous studies, we introduce the categories of: (1) selling information, (2) seller information, (3) product information, and (4) delivery informa- WLRQ:HZLOOGH¿QHDQGGLVFXVVHDFKRIWKHVH categories in turn. Selling information includes general infor- mation about the auction and the terms of an item’s sale. The initial bid price, the availability of a buy-now option, the auction duration, and the auction’s ending time are included as selling information variables. Table 1 contains a list of W K HVH Y D U LDE OH V W KHL U G H¿QLW L RQV D QGD O L VWRIV W X G - ies in which they have been investigated. There KDYHEHHQDQXPEHURILPSRUWDQW¿QGLQJVLQWKLV DUHD,WKDVEHHQREVHUYHGWKDWDQLWHP¶V¿QDOELG S U LF HF D QE H VLJQL ¿FD QWO\D I IH FW H GE \L WVL Q LW L DOELG price (Brint, 2003). Bidders have been found to 1947 Exploring Decision Rules for Sellers in Business-to-Consumer (B2C) Internet Auctions sometimes ignore a buy-now option even when buy-now prices are set below prevailing market SULFHV 6WDQGL¿UG 5RHORIV 'XUKDP Setting a buy-now price may, however, enhance revenue for sellers (Budish & Takeyama, 2001) in some situations. The time of day or week that an auction ends, and the duration of an auction are frequently used as either control variables or dependent variables (Bruce, Haruvy, & Rao, 2004; Dholakia & Soltysinski, 2001; Gilkeson & Reynolds, 2003; McDonald & Slawson, 2002; 6W DQGL ¿UG 6W DQGL¿UG5RHORIV'X UKDP 2004; Subramaniam, Mittal, & Inman, 2004), but have not, to our knowledge, been conclusively linked to higher closing prices. Seller informationLVGH¿QHGDVWKHYDULRXV facets of the seller’s feedback rating. The ease with which buyers are able to provide feedback has made a seller’s feedback rating one of the most VLJQL¿FDQWSUHGLFWRUVRIDXFWLRQFORVLQJSULFH Feedback mechanisms can help sellers earn higher prices (Bruce, Haruvy, & Rao, 2004; McDonald & Slawson, 2002; Ottaway, Bruneau, & Evans, 2003) and have been shown in one previous study to play a role in generating price premiums for reputable sellers (Ba & Pavlou, 2002). The number of positive feedback ratings and the number of negative feedback ratings are included as seller information variables in this study (see Table 1). We investigate both positive as well as negative feedback, because it has been found that positive and negative feedback have an asymmetrical HIIHFWXSRQWKH¿QDOELGSULFH6SHFL¿FDOO\SRVL- WLYHIHHGEDFNLVPLOGO\LQÀXHQWLDOLQGHWHUPLQLQJ ¿QDOELGSULFHZKLOHQHJDWLYHIHHGEDFNLVKLJKO\ LQÀXHQWLDO6WDQGL¿UG7KXVLWKDVEHHQ clearly demonstrated that seller information is also an important subset of variables to examine when researching Internet auctions. Product information r e f e r s t o t h e i n f o r m a t i o n provided by the seller or by other bidders about the item being auctioned. Frequently, product information is measured by recording the num- ber of pictures of an item and the number of bids which an item receives (see Table 1). One study has explained that pictures of an item being auctioned on the Internet may affect information SURFHVVLQJDQGXOWLPDWHO\WKHLWHP¶V¿QDOFORVLQJ price (Ottaway, Bruneau, & Evans, 2003). An- other found that detailed descriptions of the item ZHUHVLJQL¿FDQWSUHGLFWRUVRIDFRPSOHWHGVDOH (Gilkeson & Reynolds, 2003) 1 . Other researchers have included product description as a control variable in their studies (Bruce, Haruvy, & Rao, 2004; Dholakia & Soltysinski, 2001; Gilkeson & 5H\QROGV6WDQGL¿UG5RHORIV'XUKDP 2004), giving at least informal credence to the notion that product information, such as pictures RIDQLWHPFDQLQÀXHQFHDQLWHP¶V¿QDOFORVLQJ price. Finally, the number of bids and the number of bidders has been shown to be factors leading to higher closing prices (Dholakia & Soltysinski, 2001; Gilkeson & Reynolds, 2003; Wilcox, 2000). Following the lead of these scholars, and in order WRUHDFKDPRUHGH¿QLWLYHFRQFOXVLRQUHJDUGLQJWKH possible impact of product description on auction prices, we also include product information in our analysis of Internet auctions. Finally, delivery information simply refers to the cost of shipping and to the available delivery options. The availability of expedited delivery, international delivery, and the item’s shipping cost are included here as variables (see Table 1). Relatively few researchers have included this subset of variables within their models. However, one study argues that high seller reputation and GHOLYHU\ HI¿FLHQF\ PD\ FRYDU\ 0F'RQDOG Slawson, 2002), while another includes shipping cost as a control variable (Gilkeson & Reynolds, 2003). We introduce the examination of interna- tional delivery because we believe that, with the increasing level of international activity in Inter- net retailing and Internet auctions, international shipping will become more important to sellers wishing to ensure the largest possible set of poten- tial bidders. To gain a more complete perspective on all factors impacting auction prices, we will include each of the aforementioned delivery at- 1948 Exploring Decision Rules for Sellers in Business-to-Consumer (B2C) Internet Auctions Variable Description Source Initial Bid Price Starting bid price (Gilkeson & Reynolds, 2003; McDonald & Slawson, 6WDQGL¿UG6WDQGL¿UG5RHORIV'XU- ham, 2004) Buy-Now Option Presence or absence of option for bidder to end auction early by purchas- ing at a seller-determined ¿[HGSULFHH%D\¶V%X\ it-Now option) 6WDQGL¿UG5RHORIV'XUKDP Auction Duration Length of auction in days (Dholakia & Soltysinski, 2001; Gilkeson & Reynolds, 2003; McDonald & Slawson, 2002; Mehta, 2002; 6WDQGL¿UG6WDQGL¿UG5RHORIV'XUKDP 2004; Subramaniam, Mittal, & Inman, 2004) Auction Ending Time Time of day auction ends (Dholakia & Soltysinski, 2001; Gilkeson & Reynolds, 2003; McDonald & Slawson, 2002; Mehta, 2002; 6WDQGL¿UG Table 1a. Previous empirical studies measuring selling information variables Variable Description Source Number of Posi- tive Feedback Ratings Total number of eBay positive feedback ratings (Ba & Pavlou, 2002; McDonald & Slawson, 2002; 6WDQGL¿UG Number of Nega- tive Feedback Ratings Total number of eBay negative feedback ratings (Ba & Pavlou, 2002; McDonald & Slawson, 2002; 6WDQGL¿UG Product Information Variables Number of Pic- tures Number of pictures (Ottaway, Bruneau, & Evans, 2003) Number of Bids Total number of bids sub- mitted for item (Dholakia, 2005b; Dholakia & Soltysinski, 2001; Gilkeson & Reynolds, 2003; McDonald & Slawson, 6WDQGL¿UG6XEUDPDQLDP0LWWDO,Q- man, 2004; Wilcox, 2000) Delivery Information Variables Availability of Expedited De- livery Availability of express delivery Availability of International Delivery Possibility to Deliver Internationally Shipping Cost Amount of shipping and handling charges (Gilkeson & Reynolds, 2003; McDonald & Slawson, 2002) 1949 Exploring Decision Rules for Sellers in Business-to-Consumer (B2C) Internet Auctions tributes in our analysis. 5HFHQWVFKRODUO\FRPPHQWDU\LGHQWL¿HVWKUHH approaches that researchers have taken in their studies of Internet auctions: (1) concept discov- ery, which explores new phenomena; (2) process explanation, which seeks to provide an economic, psychological, or social explanation for behavior; and (3) theory deepening, which uses electronic markets to develop and test theories (Dholakia, 2005a). It has been noted that concept discovery and process explanation have received the majority of researchers’ attention, while theory-deepening approaches are relatively few in number (Dhola- kia, 2005a). In the absence of established theory, continued exploratory work such as this study seems warranted. :KLOH WKH IRUHJRLQJ ¿QGLQJV IURP ,QWHUQHW auction research are noteworthy in their own right, they have a limited usefulness even when taken in sum. Without being able to ascertain ZKLFKYDULDEOHVZLOOSURYLGHWKHJUHDWHVWEHQH¿W relative to other variables, businesses are left without guidance for generating price premiums in Internet auctions. In light of this need, we will capitalize upon previous work in a novel way. Rather than simply searching among the myriad DWWULEXWHVRIDQ,QWHUQHWDXFWLRQWR¿QGWKRVHWKDW DUHSUHGLFWLYHRIWKH¿QDOFORVLQJSULFHZHSURSRVH a descriptive model based upon empirical data which ranks the attributes of Internet auctions E\WKHLULPSRUWDQFH$FODVVL¿FDWLRQDQGUHJUHV- sion tree will be produced which can be used to guide businesspeople who are making decisions regarding how to auction their products in B2C auctions. At this point, we will turn our attention to decision-tree induction, a technique capable of producing decision rules for sellers. Decision-Tree Induction Techniques 'HFLVLRQUXOHVRUUXOHVRIFODVVL¿FDWLRQFDQEH deduced from data by using various machine- learning techniques (Tsai & Koehler, 1993). Information gained by analyzing data with these inductive learning techniques can be represented in various forms, including mathematical state- ments, logical expressions, formal grammar, decision trees, graphs, and networks (Kim, Lee, Shaw, Chang, & Nelson, 2001). Decision trees are essentially visual presentations of sets of nested if-then statements. One advantage of using decision trees is that they depict rules that can be readily expressed in words, thus facilitating comprehension by decision-makers (Kim, Lee, Shaw, Chang, & Nelson, 2001). Several algorithms for building decision trees H[LVW WKH\ LQFOXGH &$57 &ODVVL¿FDWLRQ DQG Regression Trees), QUEST (Quick, Unbiased DQG(I¿FLHQW6WDWLVWLFDO7UHH6/,46XSHUYLVHG Learning In Quest), CHAID (Chi-squared Auto- matic Interaction Detector), IC (Interval Classi- ¿HU,'DQG&$JDUZDO$UQLQJ%ROOLQJHU Mehta, Shcafer, & Srikant, 1996; Mehta, Agar- wal, & Rissanen, 1996; Quinlan, 1990). While decision-tree induction allows data analysts to deduce decision rules for both continuous and discrete variables, not all algorithms are equally well-suited for use with both types of variables. For instance, CHAID and C5.0 are restricted to the analysis of categorical variables only (Berry & Linoff, 1997; Zanakis & Becerra-Fernandez, 2005). CART, on the other hand, can analyze either categorical or continuous variables. Clas- VL¿FDWLRQWUHHDQDO\VLVFDQEHXVHGIRUFDWHJRULFDO criterion 2 variables; regression-tree analysis is used for continuous criterion variables (Brieman, Friedman, Olshen, & Stone, 1984). Because of this characteristic of the CART algorithm, and because we intend to make binary splits of our dataset into price premium and non-price premium groups at each node, CART is ideally suited to our study. We now turn to a brief description of the CART decision-tree induction process. &ODVVL¿FDWLRQDQG5HJUHVVLRQ7UHH$QDO\VLV (CART) is a nonparametric procedure that deter- mines the optimal decision tree for classifying observations on the basis of a large number of 1950 Exploring Decision Rules for Sellers in Business-to-Consumer (B2C) Internet Auctions predictive variables (Brieman, Friedman, Ol- shen, & Stone, 1984). CART recursively splits a dataset into non-overlapping subgroups based upon the independent variables until splitting is no longer possible (Kim, Lee, Shaw, Chang, & Nelson, 2001). One of the principal advantages of CART is that it tends to be less-biased than other data analysis methods (Lhose, Biolsi, Walker, & Reuter, 1994; Sorensen, Miller, & Ooi, 2000; Zanakis & Becerra-Fernandez, 2005). For instance, multiple discriminant analysis (MDA) and LOGIT methodologies need to satisfy the assumption of multivariate normality for inde- pendent variables; in addition, MDA requires that the groups’ covariance structure be equal. Thus, if the variables follow some distribution other than the multivariate normal distribution, MDA and LOGIT will give biased results. The assumptions of multivariate normality and equal covariance can be easily violated in empirical GDWDVHWVELDVHGFODVVL¿FDWLRQFDQUHVXOW,QVXFK a situation, CART is preferable because it rests upon more realistic, less-frequently violated as- sumptions. CART assumes only that the groups DUH GLVFUHWH QRQRYHUODSSLQJ DQG LGHQWL¿DEOH (Brieman, Friedman, Olshen, & Stone, 1984). Thus, CART is a data analysis technique that may be well-suited to real-world electronic commerce datasets. Now that some of the merits of CART have been described, we turn to an explanation of the process of decision-tree induction with CART. The decision-tree induction technique begins as a dataset is subdivided into N sub-datasets. N-1 subsets are used as training datasets, and the remaining dataset is used to test the model. 7KH¿UVWWUDLQLQJGDWDVHWLVDQDO\]HGWR¿QGWKH single most important independent variable for classifying the observations into two groups. &$57WKXVPDNHVLWVPRVWVLJQL¿FDQWVSOLW¿UVW at the root node (Berry & Linoff, 1997; Zanakis & Becerra-Fernandez, 2005). Each subgroup is WKHQH[DPLQHGDJDLQZLWKWKHDOJRULWKPWR¿QG the next-most important variable for classifying observations. After this partition, the process continues until only inconsequential variables remain (Berry & Linoff, 1997). The possibility of erroneously classifying some observations is computed by summing the predictive error rate at each split (Zanakis & Becerra-Fernandez, $WWKLVSRLQWWKHWUHHLV³SUXQHG´WRUH- PRYHEUDQFKHVWKDWLQÀDWHWKHHUURUUDWHZLWKRXW providing substantial improvements in predictive power (Berry & Linoff, 1997). After the decision WUHHLVJHQHUDWHGIURPWKH¿UVWWUDLQLQJGDWDVHW the subsequent training datasets are analyzed to UH¿QHWKHWUHH7KLVSURFHVVLVNQRZQDVFURVV validation. Analysis of the training datasets thus generates a decision tree—a predictive model for classifying observations. Finally, the test dataset is a n a l y z e d t o ve r i f y t h a t t h e d e c is i o n t r e e g e n e r a t e d XVLQJWKHWUDLQLQJGDWDVHWDFFXUDWHO\FODVVL¿HVWKH remainder of the data as well. To our knowledge, the use of decision-tree induction techniques to analyze Internet auction data and generate decision rules has not been undertaken. The application of the decision-tree analysis technique to Internet auction data may help to unify and bring coherence to the disparate H[WDQW¿QGLQJVLQ,QWHUQHWDXFWLRQUHVHDUFK,WPD\ also provide perspective on the relative importance of the numerous factors that have been proven to VLJQL¿FDQWO\LPSDFWDXFWLRQRXWFRPH METHOD We present the following analysis in order to answer questions about the variables enabling merchants to earn price premiums in Internet auctions and also to describe the decision rules for these variables. Sample Data was collected over a one-month period in 2005 from eBay’s U.S. Web site. Data from international industry leader eBay has been 1951 Exploring Decision Rules for Sellers in Business-to-Consumer (B2C) Internet Auctions frequently used as the point of data collection for studies of Internet auctions (Ba & Pavlou, 2002; Brint, 2003; Bruce, Haruvy, & Rao, 2004; Dholakia, 2005b; Dholakia & Soltysinski, 2001; *LONHVRQ5H\QROGV6WDQGL¿UG5RHORIV & Durham, 2004; Ward & Clark, 2002; Wilcox, 2000). Data from eBay is used for three reasons. First, eBay data is often used because the realism of such data is often preferable to data collected in an experimentally-controlled laboratory set- ting. Field experiments with auctions present an obvious trade-off between experimental control and realism (List & Lucking-Reiley, 2000). Laboratory experiments of auctions have been criticized on the grounds that subjects’ behavior LQDQDUWL¿FLDOODERUDWRU\HQYLURQPHQWPD\QRW be exactly the same as it would be in real-world conditions (Lucking-Reiley, 1999). It has been argued that experimental subjects have no in- centive to develop optimal bidding strategies or apply experience gained from bidding (Ward &ODUN&ROOHFWLRQRIGDWDIURPD¿HOG setting reduces questions regarding its general- izability to the marketplace. For these reasons, our goal of developing a guideline for selling in Internet auctions that is both descriptive and prescriptive leads us to follow the precedent of WKHVHUHVHDUFKHUVLQXVLQJ¿HOGGDWDUDWKHUWKDQ experimental data. The second reason that researchers often use eBay data is simply that eBay continues to be the Internet auctioneer of choice. EBay continues to lead the industry because of the circular effect of high seller volume eliciting high bidder interest, which in turn motivates sellers to continue to uti- OL]HH%D\:LQJ¿HOG7KXVH%D\SURYLGHV substantial numbers of auctions to observe and numerous points of measurement. 7KHWKLUGDQG¿QDOUHDVRQIRUWKHXVHRIH%D\ data is that eBay is the largest and most interna- tional of the Internet auctioneers. Their auction mechanism and terminology are used more widely than any other auctioneer’s. Thus, in an endeavor to provide the most generalizable results, we have selected eBay as the point of data collection for this study. The items examined in this study are a DVD movie (404 auctions) and a popular MP3 player (366 auctions). All DVD auctions were for the same, new, identically-packaged movie title (the SRSXODUDQLPDWHGIHDWXUH³7KH,QFUHGLEOHV´ and all MP3 player auctions were for the same, QHZ¿UVWTXDOLW\LGHQWLFDOO\SDFNDJHGPRGHORI the device (the 4 GB Apple iPod). All items were GHVFULEHGDV³QHZ´³QHYHUXVHG´³QHZLQER[´RU ³EUDQGQHZ´:HLQFOXGHGWKHVHLWHPVWRVDPSOHD reasonably-broad spectrum of items, ranging from inexpensive (DVD) to relatively expensive (MP3 player). We collected data during a three-week window of time to guard against effects due to changes in the market price (due to the release of new versions of the products, or due to a reduc- WLRQLQFRVWLQ¿[HGSULFHPDUNHWV$GGLWLRQDOO\ these items were examined because their value should not change with the fortunes of a team or individual (as sports collectibles or celebrity memorabilia might). Finally, the high sales volume of these items facilitates data collection. Variables The variables for this study are those listed and GH¿QHGHDUOLHULQ7DEOH$VZHQRWHGHDUOLHU variables studied in previous research as predictors RIDXFW LRQRXW FRPH FD QEHFOD VVL ¿HGL QWRIRX UFDW- egories: selling information, seller information, product information, and delivery information. In addition, the dependent variable of interest is ¿QDOELGSULFH:HGH¿QH¿QDOELGSULFHDVWKH highest bid submitted for a given item. Measurement of Variables Table 2 reports our coding scheme for the vari- ables in the Internet auction. Table 3 reports the descriptive statistics of the data for 404 DVD auctions and 366 MP3 player auctions. 1952 Exploring Decision Rules for Sellers in Business-to-Consumer (B2C) Internet Auctions Research Design This study uses CART to determine the most important variables that sellers should consider to earn price premiums. The reader will recall ¿UVWWKDWSULFHSUHPLXPVKDYHEHHQGH¿QHGDV “the monetary amount above the average price received by multiple sellers for a certain matching product” (Ba & Pavlou, 2002, pp. 247-248) and second, that CART is a nonparametric procedure that determines the optimal decision tree for classifying observations on the basis of a large Variables Coding Criterion (Dependent) Variable: Final Bid Price Continuous: dollars and cents Independent Variables: Selling Information Variables (1) Initial Bid Price Continuous: dollars and cents (2) Buy-Now Option Binary: 0—not available, 1—available (3) Auction Duration Continuous: duration of auction in days (4) Auction Ending Time Categorical: 1: Weekday before 4 PM 2: Weekday after 4 PM 3: Weekend before 4 PM 4: Weekend after 4 PM Seller Information Variables (5) Number of Positive Feedback Ratings Continuous: number of positive ratings (6) Number of Negative Feedback Ratings Continuous: number of negative ratings Product Information Variables (7) Number of Pictures Continuous: number of pictures (8) Number of Bids Continuous: total number of bids submitted Delivery Service Information Variables (9) Availability of Expedited Delivery Binary: 0—not available, 1—available (10) Availability of International Delivery Binary: 0—not available, 1—available (11) Shipping Cost Continuous: dollars and cents Table 2. Data coding scheme 1953 Exploring Decision Rules for Sellers in Business-to-Consumer (B2C) Internet Auctions number of predictive variables (Brieman, Fried- man, Olshen, & Stone, 1984). We perform two analyses with CART: classi- ¿FDWLRQWUHHDQDO\VLVDQGUHJUHVVLRQWUHHDQDO\VLV :H¿UVWXVH¿QDOELGSULFHDVWKHFULWHULRQYDULDEOH IRU FO D V VL¿F D W LRQ W U HHD QDO \VL V7 K H FO D V VL ¿ F DW L R Q WUHHDOJRULWKPLGHQWL¿HVWKHSUHGLFWRUVWKDWEHVW separate our data into categories where an auction yields a price premium (denoted in subsequent ¿JXUHVDV33RUIDLOVWR\LHOGDSULFHSUHPLXP (denoted as NPP). Second, we use number of bids as a criterion variable for regression-tree analysis. We use number of bids as criterion variable because the number of bids is highly and GLUHFWO\FRUUHODWHGZLWKWKH¿QDOELGSULFH7KXV the results should be substantially similar to those LQWKHFODVVL¿FDWLRQWUHHDQDO\VLV DVD Movie (N=404) MP3 Player (N=366) Mean Std. Dev. Mean Std. Dev. Criterion (Dependent) Variable: Final Bid Price 9.74 2.94 187.58 19.34 Independent Continuous variables: (1) Initial Bid Price 4.58 3.89 34.98 69.39 (3) Auction Duration 4.47 2.20 2.93 2.05 (5) Number of Positive Feedback Ratings 849 2616 2374 3176 (6) Number of Negative Feedback Ratings 5.70 9.35 20.04 32.38 (7) Number of Pictures 0.58 0.56 2.58 1.71 (8) Number of Bids 6.14 4.23 23.28 12.60 (11) Shipping Cost 4.20 1.07 16.32 5.50 Independent Categorical Variables Frequencies Frequencies (2) Buy-Now option No: 384,Yes: 20 No: 357, Yes: 9 (4) Auction Ending Time Weekday Morning 98 164 Weekday Afternoon 153 48 Weekend Morning 50 20 Weekend Afternoon 103 134 (9) Availability of Expedited Delivery No: 354, Yes: 50 No: 256, Yes: 110 (10) Availability of International Delivery No: 187, Yes: 217 No: 130, Yes: 236 Table 3. Descriptive statistics . explanation for behavior; and (3) theory deepening, which uses electronic markets to develop and test theories (Dholakia, 2005a). It has been noted that concept discovery and process explanation. bibliography of earlier literature (Stark & Rothkopf, 1979) and a review of experimental auction literature (Kagel, 1995). Auction Mechanisms and Auction Theory Auction mechanisms are generally categorized. models; both the IPV and CV models can be understood as special cases of the more general AV model (McAfee & McMillan, 1987). Recent studies of Internet auctions rely upon and explicitly mention