ANALYZING COMPETITION AND COLLUSION STRATEGIES IN ELECTRONIC MARKETPLACES WITH INFORMATION ASYMMETRY Robert J. Kauffman Associate Professor of Information and Decision Sciences Charles A. Wood (contact author) Doctoral Program in Information and Decision Sciences Carlson School of Management University of Minnesota, Minneapolis, MN 55455 Phone: 612-624-8030; Fax: 612-626-1316 Last Revised: November 26, 2000 ______________________________________________________________________________ ABSTRACT Electronic commerce (EC) researchers are divided between those who show how EC can cause intense Bertrand competition and those who show an absence of this competition. This research takes a more central stance to show when intense competition occurs and when collusion occurs in an EC environment. We present a multi-industry investigation of competition and collusion related to firm pricing behavior using a customized data-collecting Internet agent, the Time Series Agent Retriever (TSAR). Using an analytical model, we show how EC technology increases the ability of firms to tacitly collude to keep prices higher than expected, even amid intense competition. We then develop a probit regression model to test the predictions of our theory. The results show that EC technology allows either intense competition or tacit collusion, and that EC vendors tend to choose one or the other based on industry and firm attributes. ______________________________________________________________________________ KEYWORDS: Econometric analysis, electronic commerce, information asymmetry, market power, pricing, probit, tacit collusion. INTRODUCTION Researchers have predicted intense competition among electronic commerce (EC) firms (e.g., Bakos 1997). In certain industries, such as the airline parts industry (Choudhury, Hartzel and Konsynski 1998), this competition certainly seems to be occurring, yet in other well-developed EC industries, such as the book industry, such intense competition is rare (Varian 2000; Dillard 1999). Just as Bakos (1997) predicted, EC technology drastically reduces search costs. Indeed, Byrnjolfsson and Smith (1999) point out how "shopbots" are "the great equalizer" and allow the EC consumer to drastically reduce search costs by searching for a product from several sellers simultaneously. However, we wonder if there is more to the story of pricing goods on the Internet than has been understood by academic researchers up to this time. In fact, up till now, even though the majority of research and trade journals point out how the Internet can increase competition, there are several anecdotal indications that tacit collusion may be occurring in several market areas, instead of the intense competition that low friction search might predict. For this research, we define collusion as the matching of a competitor’s price above average cost, whereas we define competition as beating a competitor’s price when that price is above average cost. These definitions follow the suggestions of Varian (2000) and Dillard (1999) in their empirical research. Tacit collusion, in contrast to these other definitions that we have offered, is collusion without any formal agreement. For example, Carvajal (1999) reports how, when Amazon (www.amazon.com) reduced prices, other top bookselling vendors exactly matched the reduction. Bailey (1998), Dillard (1999), Kauffman and Wood (2000a), and Varian (2000) all show evidence of how booksellers and CD vendors collude and may exactly match each other in both price increases and decreases. Developments such as these prompt BusinessWeek to ask: “So where are all the bargains?” (as discussed in Sager and Green 1998) Since it is more convenient for shoppers to shop online for goods with a high level of electronic transactability (e.g., commodity products, information goods, items with low complexity specifications, etc.), Internet sellers are often more interested in matching rather than beating their competitors’ prices. Varian (2000) comments on how EC vendors can use “shopbot” technology to tacitly collude with each other as easily as customers use shopbots to compare prices. Thus, EC technology allows a decrease in information asymmetry among sellers, thereby allowing sellers to match prices, either up or down, nearly instantaneously. Such capabilities allow strategies that are not feasible in traditional markets. As information systems (IS) researchers, it is imperative that we understand how EC technology impacts pricing strategy. New evidence gathered in the EC environment can help researchers to shed light on the choice Internet sellers make between tacit collusion and Bertrand competition. For example, Brynjolfsson and Smith (1999) show that enough market friction exists so that EC consumers still must rely on reputation and brand awareness, even when choosing between firms selling identical products, while Clemons, Hann, and Hitt (1998) discuss how price dispersion exists in the online travel agent industry, which argues against the one price rule resulting from Bertrand competition. This market friction is a form of a decrease in information asymmetry between the EC seller and the EC buyer in which EC technology decreases the consumer’s knowledge of product and seller, forcing the consumer to rely on trust and reputation. Such a change in dynamics gives rise to different pricing strategies employed by Internet sellers. In this research, we address the following research questions: • How can researchers empirically evaluate the level of competition or collusion in an EC marketplace? • What are the effects of reduced information asymmetry among Internet competitors on the pricing strategy employed by the Internet seller? • What factors cause an Internet seller to collude, and what factors cause an Internet seller to compete intensely? We will argue that different factors can determine the pricing strategy of a firm. These factors can be industry-specific, product-specific, or firm-specific. In this research, we explore the economic, marketing and IS perspectives on pricing, collusion, and competition. We develop conceptual and analytical models that describe why EC technology enhances the feasibility of tacit collusion. Then, using a data-collecting Internet agent we call the Time Series Agent Retriever (TSAR), we collect pricing data for best selling items in multiple industries sold by multiple vendors. We then extract price changes from this pricing data, and develop and test a limited dependent variable econometric model that describes the motivation for these price changes. On the basis of this analysis, we are able to empirically demonstrate that pricing strategy depends on both the relative market power between a firm and its competitor, and the previous actions of that competitor. Internet sellers respond either competitively or collusively based on a strategy that mirrors the established strategy of competitors with more market power (i.e., the market leaders). Furthermore, we will show that a single market leader that refuses to collude will place an entire industry in an intensely competitive mode, while if the market leaders collude, the rest of the industry will follow their lead and collude as well. LITERATURE REVIEW Marketing research tells us that firms compete asymmetrically where the firms with greater market power, such as Amazon, Barnes and Noble (i.e., www.bn.com, hereafter referred to as BN.com), and CDNow (www.cdnow.com), are not affected by promotions from smaller firms, such as Books-aMillion (www.booksamillion.com) or Rock.com (www.rock.com) (e.g., Carpenter et al. 1989; Blattberg and Wisniewski 1989). Asymmetric competition suggests that firms with a large market share and significant brand recognition compete differently than smaller firms. By contrast, IS research has concentrated on how the EC environment increases the tendency toward intense Bertrand competition (e.g., Bakos 1997; Choudhury, Hartzel and Konsynski 1998; Bailey 1998). Bertrand competition occurs when two or more firms sell identical products (e.g., books or CDs), and compete so fiercely that their price is driven down to average costs (Bertrand 1883). In this research, we seek to integrate these two research streams by adding a discussion of tacit collusion borrowed from the economics literature (Chamberlin 1929). This will enable us to more fully understand the role of pricing in the technologyintensive EC environment. Tirole (1997) describes how tacit collusion is difficult to implement in traditional non-technological selling environments. He argues that this is because “cheating” can occur, where one party secretly lowers its price, thereby stealing market share from the other firms. In this research, we show how EC technology decreases the likelihood of “cheating” by decreasing the information asymmetry among Internet sellers, and therefore makes tacit collusion a more viable strategy. EC Product Pricing Bakos (1991) was the first to associate Bertrand (1883) competition with the sale of products on the Internet. He predicted that EC technology decreases customer search costs and therefore allows customers to easily compare prices, thus leading to a reduction of switching costs between vendors. The result will be intense Bertrand price competition between identical products. Choudhury, Hartzel and Konsynski (1998) show this to be true in the online airline parts industry. Alba et al. (1997) describe how vendors avoided embracing the Web because of fears of intense Bertrand competition. Reluctance to transact through the EC environment allowed entry of new players, for instance, Amazon in the bookselling business and e*Trade (www.etrade.com) in the brokerage industry. Lynch and Ariely (2000) show how consumers are more price-sensitive when they purchase wine online. However, there have been numerous observations of pricing behavior that argue against Bertrand competition. Bertrand competition theory argues that prices will converge to a single price, where demand intersects marginal costs (Tirole 1998), resulting in what economists term the Bertrand's one-price rule. Many IS researchers have shown how the one-price rule does not seem to apply in the EC environment. Clemons, Hann, and Hitt (1998) describe how online travel agents charge different prices when given the same customer request. Brynjolfsson and Smith (1999) review customer actions when using a shopbot and demonstrate how friction still exists in online markets, and how several consumers do not necessarily purchase the cheapest item, but rather still gravitate toward the brand name retailers. Lal and Sarvary (1999) also show a preference for specific vendors in an EC environment, especially if the product being sold has relevant non-digital attributes. In this research, we attribute this preference to an increase in information asymmetry between EC sellers and buyers. While buyers can search for the lowest costs, they do not actually know if or when they will get a product, and whether or not the firm they are dealing with is reputable. To exacerbate this problem, EC sellers can remain anonymous and may resist detection and punishment of opportunistic behavior (Kauffman and Wood 2000a). This increase in information asymmetry can reverse the effect of reduced search costs, and actually add to the switching costs when an EC buyer switches from a known and trusted seller to an unknown seller. Tirole (1998) describes that economists call Bertrand’s one price rule the Bertrand Paradox because vendors of identical products often still manage to sell their commodity at a higher price. Bertrand’s intense competition works well in a single-period game where every player is trying to clear all inventory. However, in a multi-period game, sellers know that they may be hurting their profits in the long term by drastic price cuts, and will instead avoid this intense Bertrand competition, concentrating instead on competitor reaction, reputation, and service to gain market share rather than low prices. Asymmetric Competition Carpenter et al. (1989) show how market leaders compete asymmetrically in that price promotions of the market leaders can steal customers from market followers, but smaller firms’ price promotions have little effect on market leaders. Blattberg and Wisniewski (1989) carry this thread further and show how industries form price tiers where market leaders compete with each other at a higher price level, and market followers compete with each other at a lower price level. We argue that information asymmetries are at the heart of asymmetric competition: the reason a consumer might prefer one Internet seller over another for an identical product, such as a book or CD, is because there are information asymmetries present, such as reliability to ship the product in a timely fashion or confidentiality with personal information. Such information asymmetry coerces a consumer to include trust as a factor in addition to price when considering one vendor over another. As EC technology causes information asymmetry to increase, Internet buyers rely on past experience and reputation, as well as price, when buying products from an Internet seller. Thus, an empirical study on pricing in the Internet selling context should consider the effect of asymmetric competition. For this research we will employ a comparative ratio of firm market power to competitor market power to absorb the competitive asymmetry effect that causes smaller firms to compete with bigger firms, but allows bigger firms to somewhat ignore the actions of smaller firms. Tacit Collusion Chamberlin (1929) introduces tacit collusion and shows how competitors will tacitly cooperate with each other to avoid intense competition. Rather than the average cost price of Bertrand competition, tacit collusion brings the market price closer to the monopoly price. One limitation of tacit collusion is that colluding players will “cheat,” if possible, and charge a little less to capture a larger market share and profit at the expense of their colluding partners (Tirole 1998). Varian (2000) and Dillard (1999) provide evidence by tracking a single book’s price and show how collusion appears to be occurring in the bookselling industry. Campbell, Ray, and Muhanna (1999) develop an analytical model that shows that the fast response times allowed by EC technology can cause price increases by facilitating tacit collusion. Tacit collusion can result from a commitment to raise prices. Many researchers discuss how firms that promise to match prices of their competitors result in higher customer prices. For example, Dixit and Nalebuff (1991) discuss an example of stereo price wars in New York, where Crazy Eddie and Newmark & Lewis both had price guarantees on consumer electronics products that not only guaranteed the lowest price, but also refunded the difference of the prices plus a percentage. The end result was that any price promotion from one vendor would actually make customers buy from the other vendor, to get the low price plus a percentage of the difference. Thus, by guaranteeing the lowest price, vendors were actually committing to a tacitly collusive strategy by punishing the price promotions of their competitors. In this research, we show how EC technology allows the same commitment. By matching the competitors' prices, firms minimize any competitor benefit from a price promotion. We also show how competitors cannot increase their market share by price promotions and are forced into a tacitly collusive strategy. Thus, EC technology allows the same commitment to tacit collusion by punishing any price promotion in the form of reduced competitor revenues and limited competitor benefits. Tacit collusion is a form of Stackelberg competition, as opposed to Bertrand competition. Stackelberg (1934) argues that, when possible, competitors react to each other rather than try to anticipate each other's actions, leading to rich and varied strategies such as reactive competition or tacit collusion. Kauffman and Wood (2000b) empirically show how the EC environment reduces information asymmetry among competitors and allows immediate evaluation of competitors’ pricing through browsers and Internet agents, thus facilitating Stackelberg competition. They argue that because EC technology allows sellers to respond to each other before the market, Stackelberg competition, where sellers set a price and then react to competitors' prices, is more descriptive of Internet seller pricing strategy than Bertrand competition, where prices are set simultaneously based on expectations of competitors' behavior. In this research, we add to this previous work by showing when tacit collusion is optimal, and when reactive pricing is optimal. We test to determine what factors affect an Internet seller’s desire to compete or collude in the marketplace. THEORETICAL MODEL Our model begins with a game theoretic prisoner's dilemma duopoly that describes why collusion is difficult to achieve. A prisoner’s dilemma duopoly occurs when two “players” who represent firms perform sub-optimally as a “group” because each player optimizes her own situation as an individual. The end result is that the players are each worse off than before. Collusion may lead to better profits for each firm, but each firm can be made better off by “cheating” and secretly lowering prices, thereby gaining market share. As a result, both firms will feel compelled to better their own situation at the expense of their competitor. The technologies that are used in Internet selling aid in immediate detection of “cheating,” and have the potential to resolve the dilemma and facilitate tacit collusion among competitors. We will introduce the concept of asymmetric competition between firms into our model, and show why some firms may not collude if there is a high probability of immediately preventing entry of future competitors or eliminating current competitors. Basic Setup Firms selling a single identical product on the Internet have three pricing strategies. First, they can compete with other firms directly in intense Bertrand competition, thus driving prices down to marginal costs. Second, firms can match other firms’ actions in a collusive manner, thus driving prices upward toward the monopoly price. Third, firms can be unresponsive to each other’s actions and follow independent strategies of their own. For simplicity, we first assume a duopoly and a friction-free market where customers are equally disposed to purchase from either vendor, and therefore the lowest price will always prevail. Later in this analysis, we relax these assumptions. We also assume that marginal cost for the vendors is the same, and that there is enough demand to absorb the lowest-priced items. In a friction-free market, it is unlikely that a firm will be unresponsive to a competitor’s actions. Doing so allows a responsive competitor to completely absorb all the industry revenues by charging marginally less than the unresponsive firm. Figure 1 shows a firm’s prices in a purely competitive versus a collusive market with respect to fixed costs (FC), marginal costs (MC) and average costs (AC). Figure 1. Prices in a Competitive and Collusive Industry P FC MC P tc AC P bc Q bc Q tc Q In Figure 1, the price and quantity of the product sold in Bertrand competition (Pbc and Qbc, respectively) are compared to the price and quantity of the product sold in the presence of tacit collusion (Ptc and Qtc, respectively). In a competitive market, firms are willing to produce up to the point where marginal cost equals average cost. The increased profit from charging a higher price enabled by collusion gives firms an incentive to produce more products. In Figure 1, both firms, either explicitly or tacitly, agree to charge a higher price, thereby increasing their profits. As a result, both firms are better off by colluding. When collusion is observed instead of intense Bertrand competition in an industry, the firm’s profit will result from the aggregate of its excess marginal revenue from each sale of a product, Marginal Revenue = Price - Marginal Costs, as shown in Equation 1 below. A Firm’s Profit from Collusion: π= Qtc ∫ (P tc 0 − MC ) dq − Qbc ∫ (P bc − MC ) dq (1) 0 Equation 1 shows that profit from collusion is equal to the marginal revenue enabled by the higher price less the marginal revenue that would be earned in a competitive market. The typical outcome in a market in which collusion occurs involves both firms splitting the market exactly in half. As a result, both will enjoy higher marginal revenues from collusion. However, collusion causes a prisoner’s dilemma: each firm would be better off by not cooperating and slightly reducing prices, thereby capturing the entire market share (Tirole 1998). The firm knows that, in any period, its competitor can capture all industry revenues, and cause the firm to have a loss equal to fixed costs. Therefore, the firms will opt for competition, and each will make less profit than it otherwise would, rather than losing all revenues to the competitor. Furthermore, since firms can continue be uncooperative in every period, the prisoners’ dilemma will be stable across a multi-period game as well. Firms will continue to compete unless there is quick and immediate detection of their non-cooperation so colluding firms can respond immediately to competitor price changes. Modeling Information Asymmetry In prior research, several authors (Bakos et al. 1999; Choudhury, Hartzel and Konsynski 1998) point out how the EC environment reduces search costs for consumers who are seeking to purchase various kinds of goods on the Internet, thereby resulting in increased competition. However, EC also reduces information asymmetry between sellers, resulting in faster and less costly detection of competitor actions, such as price changes. While sellers are able to effectively and efficiently compare prices between vendors, vendors can use this same technology to monitor each other’s prices (Varian 2000). Shapiro (1983) introduces a variable, λ, to indicate the probability of detection of opportunistic behavior. For this study, we expand λ to be defined as the probability of being able to detect and respond to a colluding partner’s cooperation (e.g., whether the competitor matches prices or lowers prices) before the market responds. By reducing the information asymmetry between sellers, EC technology allows firms to more easily accurately detect collusive behavior (i.e., a high λ). Therefore, EC technology leads to pricing strategies that tend to involve more collusion. EC technology allows firms to make quick responses to competitor price changes, as shown in Equation 2: A Firm’s Expected Excess Marginal Revenue from Collusion: π =λ Qbc Qtc (Ptc − MC ) dq − (Pbc − MC ) dq − ∫0 ∫ 0 (1 − λ )FC (2) In Equation 2, the chance for higher profits from collusion is compared to the potential lost revenue to cover fixed costs. With a high λ, the chance for excess profits from collusion is great, while the chance to lose all sales from normal competition because of non-cooperation is slight. A high λ also presumes that competitors are willing to collude with each other. If a non-cooperative competitor has no wish to collude (e.g., to establish barriers to entry or to try to drive competitors out of the market), then the probability of detecting collusive cooperation is zero. On the other hand, if competitors show a willingness to collude with each other and firms have enhanced their capability to detect the noncooperative actions of their competitors, as becomes possible with the technologies that support Internetbased selling, then firms will be more likely to collude with a competitor, either tacitly or explicitly. Similarly, a previously-cooperative competitor will be less likely to lower prices just to capture market share if the chance of detection is great, because such an action will be prompt an immediate response to and will hurt both firms. It will not grant the benefit desired by the uncooperative firm. If πi is positive, the firms will collude; if not, the firms will compete. To make our model more realistic, we next relax two assumptions. First, the assumption of a duopoly needs to be relaxed to include multiple firms. Second, the assumption of a friction-free market may also be somewhat unrealistic for the contexts in Internet selling that we will study. EC markets show friction in that EC customers feel more comfortable buying from certain vendors, and are willing to pay a little more for products from firms with a reputation for perceived reliability (Brynjolfsson and Smith 1999). With many firms in a market, asymmetric competition will begin; price promotions of established market leaders will affect more competitors than price promotions of smaller, newer vendors for the same products (Carpenter et al. 1988; Blattberg and Wisniewski 1989). With multiple firms in a market containing some market friction, firms are now more interested in the actions of specific competitors. A firm’s collusive and competitive responses to its competitors’ actions depend on the likelihood that those actions will affect the firm’s profit. In a multi-firm competitive industry that contains market friction, firms now have possible motivation to ignore the actions of some of their competitors. Firms will focus only on those competitors that can impact their profitability. By the same token, firms will not respond to the actions of smaller competitors in an environment where response and detection is likely because such an action will result in lower profits for the firm, since smaller firms have little or no effect on the sales of the market leaders. With this in mind, we introduce in our model the parameter γ to indicate the perceived magnitude of impact (in the continuous interval [0,1]) that a competitor’s actions will have on a firm’s profits, as shown in Equation 3. A Firm’s Expected Profit from Collusion in a Market with Friction: Qbc Qtc π = γ λ ∫ (Ptc − MC ) dq − ∫ (Pbc − MC ) dq − ( 1 − λ ) FC 0 0 (3) Equation 3 shows how competitor actions will have differing effects on firm profits. The larger the effect, the more likely the firm will collude or compete with that competitor. With an extremely low γ, the competitor’s actions will have a minimal effect on the firm’s profits, and the firm is likely to ignore the competitor’s actions in lieu of more profitable strategies. With a high γ, the competitor’s actions can significantly affect the firm’s profits, and the firm will respond either in a collusive or competitive manner to competitor actions. Examining Equation 3 in light of asymmetric competition theory, we know that a smaller competitor is likely to be ignored by a market leader. Equation 3 is still valid in a multi-period game because the profit from operations does not contain a time element, and will be evaluated identically every period. However, a firm will tend to compete more intensely if future competitors may cause future profits to fall. A firm will compete if there is a perceived high probability in eliminating future competition, or if there is a perceived probability that intense competition will eliminate current competitors, leaving the firm with a higher market share. Equation 4 represents this scenario. A Firm’s Expected Profit from Collusion Less Future Sales in a Market with Friction: Qbc Qtc T 1 π = γ λ ∫ (Ptc − MC ) dq − ∫ (Pbc − MC ) dq − ( 1 − λ ) FC − ∑ (θ tπ t ) (4) 0 t =1 rt 0 In Equation 4, the profit in the current period, π, is equal to the profit from collusion less the discounted increased future profits of driving out current competitors, πt, and the impacts of eliminating new competitors who now will avoid this intensely competitive market. θ t is the probability of eliminating the entry of a new firm or eliminating a current competitor in time period t, and π t is the increase in profit because of that elimination. These future profits are discounted by rate r t , which reduces the effects of profits in the future on current operations. If a firm believes that it can drive out competitors and prevent new entrants, then that firm will compete more fiercely. However, this is a dangerous strategy since the longer the competitor continues competing, the less will be the beneficial impact on present profits. A firm that underestimates the strength of its competitors can find itself losing profits from collusion while not realizing the profits from a larger market share. HYPOTHESES Our theoretical model indicates that without at least some cooperation from every market leader in an industry, the entire industry is forced to compete at an extreme level, resulting in few, if any profits, as prices dive toward average costs (Bakos 1997). Based on our theoretical analysis, we form three hypotheses and one corollary, and discuss them below. THE ‘MODIFIED TIT-FOR-TAT’ HYPOTHESIS (H1). Internet-based sellers will tend to react collusively if the competitors that they are responding to tend to act collusively. Conversely, Internet-based sellers will tend to act competitively (i.e., non-collusively) if competing Internet-based sellers that they are responding to refuse to tacitly collude. INDUSTRY COROLLARY (HC1): Levels of collusion and competition will vary among different industries for which Internet-based selling is observed. The Modified Tit-For-Tat Hypothesis (H1) illustrates how collusion will take place only in the presence of cooperation among competitors. "Tit-for-tat" refers to the act of mirroring your competitor's actions. Thus, if a competitor shows strong competition with you, you will respond with strong competition. Dixit and Nalebuff (1991, p. 113-114) describe a modified tit-for-tat strategy that involves reviewing the past recent actions of a competitor in aggregate, and responding with a similar strategy. For example, if an Internet seller's competitor is competitive on one product, but colludes with all others, the Internet seller will also be, for the most part, collusive. On the other hand, if a competitor has shown a reluctance to collude, overall, in past periods, there is no reason to believe that competitor will collude in current periods. In terms of Equations 2, 3, and 4, a firm has no chance of detecting collusive behavior (λ = 0) if the competitor refuses to collude. Collusion requires cooperation among firms, so if a firm is responding competitively to a price offered by a competitor, then that action is likely to cause the competitor to also respond competitively. In terms of Equation 3, market leaders will have a high value for the perceived magnitude of impact of a competitor’s price change actions, γ. If the market leaders can collude with a high probability of detecting whether or not other firms are cooperating (e.g., both λ and γ are close to one from Equations 2 and 3 for all firms), the marginal revenue from each product sold increases throughout the industry. Conversely, if a market leader firm refuses to cooperate and does not collude (e.g., the probability of detection of opportunistic behavior λ is near to 0 and γ is close to one from Equations 2 and 3 for all firms), the entire industry will be forced into intense Bertrand competition at least with that firm and at worst with every firm in the industry. This is because with an intensely competitive market leader, there is no benefit in colluding with other competitors at a higher price, since the market leader will command a larger market share not only because of price but also because of brand recognition. The Industry Corollary (HC1) describes how competitive and collusive decisions are likely to vary between different firms in each industry, and therefore affect competition at the industry level. ASYMMETRIC COMPETITION HYPOTHESIS (H2): The more market power on Internet-based seller has in relation to another Internet-based seller, the more likely that the former will respond with a competitive (e.g., lower) price. The Asymmetric Competition Hypothesis (H2) describes how competitive responses are more likely when firms that sell goods on the Internet respond to other Internet-based sellers with higher market power. In terms of Equations 3 and 4, a firm with high market power has a greater γ, and therefore its actions have a greater chance of impacting a competitor. Market leaders have the greatest impact on smaller firms because market leaders' price promotions can affect smaller firms’ profits, while smaller firms’ price promotions tend to not affect the market leaders’ profits (Carpenter, et. al. 1989). Indeed, as a general rule, smaller firms need to price themselves lower than the market leaders because the brand recognition of the market leaders allows them to charge more than the smaller firms. Brynjolfsson and Smith’s (1999) results are indicative of this; the authors showed that Amazon, BN.com, and Borders.com could charge around 8% higher than other firms because of their reputation. COMPETITOR RESPONSE TIME HYPOTHESIS (H3a): The less time, on average, that a firm takes to respond to its competitors, the more likely its competitors will tend to act collusively toward that firm. Conversely, the more time, on average, that a competitor takes to respond to its competitors, the more likely its competitors will tend to act competitively (i.e., non-collusively) toward that firm. FIRM RESPONSE TIME HYPOTHESIS (H3b): The more quickly a firm responds to its competitors, the more likely that firm will tend to act competitively (i.e., non-collusively). The Competitor Response Time Hypothesis (H3a) describes how firms will find no profit in competing with a competitor that will respond immediately to a price promotion. Kauffman and Wood (2000a) argue that if a competitor responds immediately to a price promotion, then the firm that makes the price promotion will not benefit from any increased market share. Conversely, a firm may be more likely to respond competitively if its actions typically do not result in an immediate response, allowing the firm's competitive behavior to be rewarded by market demand at the lower price. The Firm Response Time Hypothesis (H3b) describes how competitive behavior is more likely to be observed for a firm with immediate response time. If a firm has the ability to quickly respond, we hypothesize that this competitor will be more likely to exhibit competitive behavior to take advantage of its quick response capabilities by reducing prices just below those of the competition and taking advantage of increased market share before they can respond. DATA We examine Internet data from the bookselling and music CD industries with the assistance of a data collection software agent. Haltiwanger and Jarmin (2000) note that it is difficult to collect electronic commerce data and that traditional data collecting techniques are often inadequate when measuring the digital economy. Data-collecting agents, we feel, address this problem by allowing massive data collection while minimizing cost and effort. Such tools as the one we will now describe make it possible to carry out research designs that require data that would be impossible to collect by any other means. Our Internet data-collecting agent is called Time Series Agent Retriever (TSAR), and it collects data that are appropriate for our investigation of price collusion. Figure 2 shows TSAR’s data collection functionality. Figure 2. TSAR Data Collection Functionality TSAR retrieves vendor-neutral lists of the top-selling items from the USA Today site for books (http://www.usatoday.com/life/enter/books/leb1.htm) and the Billboard Magazine site for the top CDs (http://www.billboard.com/charts/bb200.asp). TSAR then uses these lists and retrieve data from “shopbots” (e.g., www.mysimon.com) by querying on the bestsellers books and music CDs, and then storing the resulting data into a database. For the purposes of this research, TSAR ran at 4:00 a.m. every day from February 21 to March 29, 2000. The resulting dataset contains 70,552 daily prices, including 1793 price changes from 169 products and 53 firms. PRELIMINARY MODELING ISSUES AND ANALYSIS For this research, we examine how Internet sellers respond to their competitors. We next present and discuss some of the preliminary evidence we obtained and how we use this evidence to derive our operationalization of collusion and competition. Empirical Operationalization of Collusion and Competition In the data set that we will use in our empirical analysis, there were 871 price increases and 922 price decreases. The Bertrand competition literature disallows price increases unless there is an aggregate change in marginal costs across competitors in an industry. Since there are almost as many price increases as price decreases, and it is unlikely that marginal costs changed this drastically in the 38 days of this study, it intuitively follows that intense Bertrand competition is not the sole explanation for competitive pricing strategy in an EC environment, although it may possibly be used to explain a portion of some firms’ pricing strategies. We operationalize collusion as a firm’s matching of any price currently charged by a competitor for the same product. We operationalize competition as a firm’s beating of any price currently charged by competitors for the same product. In the case of matching one competitor’s price and beating another competitor’s price, a collusive response is assumed. We recognize that it is also possible that a price change is neither competitive nor collusive, e.g., when a firm raises its prices above its competitors. We will not investigate price changes that are determined to be neither competitive nor collusive. Of the 1793 price changes, we detected 494 price changes that can be viewed as either competitive responses or collusive responses based on our operationalization. Thus, we considered these 494 price changes for our study. Preliminary Evidence of Collusion and Competition To illustrate the appropriateness of the operationalization of competition and collusion in this research, we introduce specific products to which our empirical data will apply. In Figure 3, we illustrate the price trajectories for The Hours book by Michael Cunningham and for the Mary music CD by Mary J. Blige during February and March 2000. (See Figure 3.) Figure 3. Competition and Collusion Examples THE HOURS by Michael Cunningham $11 Price $10 $9 $8 $7 37 35 33 31 29 27 25 23 21 19 17 15 13 11 9 7 5 3 1 $6 Day Number Am azon BN Books A Million Borders MARY by Mary J. Blige $15 Price $14 $13 37 35 33 31 29 27 25 23 21 19 17 15 13 11 9 7 5 3 1 $12 Day Number Am azon BN Borders CDNow Evidence of Collusion Around Book Prices. In Figure 3, prices for The Hours book appear to show a great deal of collusion. At first, all vendors charged $6.50 for the book. Then, when Amazon decided to raise its price to $10.40 (neither a competitive nor a collusive price response), BN.com and Borders.com eventually raised their price to the exact same amount. Books A Million (www.booksamillion.com) refused to collude, however, forcing Amazon to change its price back to $6.50 (thus colluding with Books A Million). Note that this is a collusive move with Amazon colluding with Books A Million, not a competitive move with Amazon competing with Borders.com and BN.com. BN.com and Borders.com quickly followed Amazon’s lead and reduce prices back to $6.50. Finally, Books A Million, the holdout whose non-cooperation, it could be argued, forced its competitors to reduce their prices back to the $6.50 level in the first place, raised its own price to $10.40 (neither a competitive nor a collusive price response). The other vendors quickly matched Books A Million’s price (with collusive responses), and by the end of the study, all were charging $10.40 for The Hours. We interpret the booksellers’ actions as follows. Amazon first raises prices as a signal to the rest of the market to raise their price for The Hours, thereby allowing every vendor to make more money on sales of The Hours. We do not classify this signal as either a collusive or competitive move. After some time, Amazon decides that Books A Million will not cooperate and reduces the price back down to the previous $6.50 level. We view this as a collusive move: all competitors did not respond to Amazon’s signal, forcing Amazon to collude at a lower level, and BN.Com and Borders.com to follow Amazon with their own collusive responses. This scenario describes why we view price changes that match one competitor but beat with other competitors as collusive responses, not competitive responses. Amazon’s change in price back to the previous level is clearly collusive with Books A Million, and not a competitive response to BN.com or Borders.com. Once Books A Million decides to “play nice” and raise its price, all three other vendors quickly follow with collusive responses. Thus, we see a picture emerging of how price collusion impacts the EC bookselling environment that is hard to deny, even with the very simple data that we present in this illustration. Evidence of Collusion and Competition Around CD Prices. Let us return to Figure 3 again for some additional inspection. Mary, a music CD, exhibited a wide variety of prices during the study period. Throughout this period, Amazon charged $13.99 and made no price changes on this product (and therefore made no collusive nor competitive moves). BN.com was charging $12.99 at the start of this study, and raised its price to $14.98 for just two days (again, neither a competitive nor a collusive price response.) No firm matched BN.com’s price increase, and two days later, BN.com reduced its price to $12.99 (considered a competitive move against Amazon) where it stayed for the rest of the study period. Meanwhile, Borders.com was charging $13.59 for Mary at the start of this study. When BN.com raised its price to $14.98, Borders responded by raising its price to $14.39 (a competitive response, still undercutting BN.com). But then, when BN.com moved its price back to $12.99, Borders.com reduced its price to $13.59 (a competitive move, since Borders.com now undercut Amazon by 40 cents). CDNow started with a $12.99 price just like BN.com. When BN.com increased its price to $14.98, CDNow decreased its price to $12.58 (a competitive move, undercutting everyone else). And, when BN.com decreased its price to $12.99, CDNow raised its own price to $12.99, matching BN.com (a collusive response). Later in the study, CDNow decreased its price to $12.49 (a competitive response) and ended by raising its price again to match BN.Com at $12.99 (also a collusive response). While there are aspects of collusion by different firms selling the Mary CD, we can see a much more competitive environment with the Mary CD then we do with The Hours book even though many of the same vendors are present. This indicates to us that the same vendor may choose to compete or collude based on some different factors. This prompts us to investigate what factors lead to intense competition and what factors lead to collusion in an EC environment. AN EMPIRICAL MODEL FOR EXPLAINING INTERNET-BASED PRICE COLLUSION We compare an Internet seller’s propensity to compete with its propensity to collude. Because of this motivation for this research, we limited our study to include only responses that were determined to be either competitive or collusive. Price changes that were neither competitive nor collusive (e.g., price increases) were eliminated. Our econometric analysis, which we now explain, uses the behavior of competitors with large market power to predict the level of collusion in Internet selling. Empirical Modeling Issues Our model choice for this research is the probit model of binary choice (Greene 1999). A probit model is one of a number of choice models that enables the dependent variable to be a binary variable. However, before we implement this model with our data to test our hypotheses, we first need to discuss a number of empirical modeling issues that arise in the presence of a binary dependent variable, and why they led us to select the probit model. Nonlinearity of Binary Choice. Because of the binary nature of the dependent variable (e.g., companies competitively respond either with tacit collusion or Bertrand competition in mind), a linear model is inappropriate. Using a linear model often results in estimates of the dependent variable that fall outside the appropriate range limits. Specifically for our case, the probability of a response cannot exceed 1 and cannot be less than 0. But simple regression can lead to estimated values for the dependent variable that exceed the maximum and minimum values for a probability. A probit transformation or logit transformation are able to resolve this problem by forcing predictions about the dependent value to lie within the [0, 1] interval. Thus, it is generally appropriate to use a logit model or a probit model to estimate the probability of a specific binary choice. Independence of Irrelevant Alternatives. Logit transformations are computationally simpler than probit transformations, making logit transformations the model of choice for most applied econometric analysis, according to Geweke et al. (1994). However, logit transformations also assume independence of irrelevant alternatives. Specifically, if we test for the probability of collusion or competition, but then add another choice (e.g., price increases), a logit model would assume that the new choice (e.g., price increases) will draw equally from the first two alternatives. This is appropriate in situations in which alternatives are vastly different from each other (Greene 1999; Kennedy 1998). We question whether it is appropriate in our context though. For example, suppose a firm is twice as likely to beat a price (i.e., competing) as to match a price (i.e., colluding). In this example, the probability of beating versus matching would be 2/3 and 1/3, respectively. Now suppose an extra choice, increasing prices, were added to the model. One would expect that in a competitive environment, raising prices would not occur that often, because of the competitive nature of firm interactions. Conversely, in a collusive environment, raising prices might act as a signal to a firm’s competitors that prices should be raised. (This is often the case with test price increases in the airline industry, for example.) Now, assuming that competing firms do not find raising prices that appealing of a choice, and colluding firms would tend to opt for raising prices -- and for this illustration that increasing prices would hold as much appeal as matching -- one would expect the probability of increasing, matching, and beating to be 1/6 (increasing), 1/6 (matching), and 2/3 (beating) for all three choices (1/3 for increasing (1/6) or matching (1/6), and 2/3 for beating). Instead, the logit model produces probabilities 1/4, 1/4, and 2/4 for increasing, matching, and beating, respectively, to preserve the relative probabilities between matching and beating. (See Kennedy 1998 or Greene 1999 for a more in depth discussion.) As shown by this example, the independence of irrelevant alternatives assumption insists that that there are no alternatives that might be added that will change the basic odds ratio, or the ratio between the existing choices. We have no reason to assume that this assumption is true, however. In fact, as shown by this example, there are reasons to believe this is not the case. And, thus, the logit transformation is inappropriate. The probit model relaxes the independence of irrelevant alternatives assumption, making the probit model generally preferred to the logit model, especially if the number of different alternatives is small. As a result, it offers a less restrictive test that can lead to more reliable coefficient estimates (McFadden 1984). The Probit Model General Form Equation 5 and Table 1 show the form of the general form of the probit model that we develop to use to test our hypotheses. Prob[ycompetition=0, ycollusion=1] = f(Competitor Behavior, Competitor Relative Market Power, Competitor Response Time, Firm Response Time, Product Descriptors,Industry Effects) (5) Table 1. Variables in the Probit Model VARIABLES DESCRIPTION Prob[ycompetition=0, ycollusion=1] Indicates whether the firm's price changes are tacitly collusive or competitive, related to competitors’ price changes. Competitor Behavior Indicates whether a competitor’s typical price change responses are collusive or competitive. Competitor Relative Market Power Measures market power for the firm relative to its competitors. Firm Response Time Measures how long it takes a firm to respond to a competitor’s price change. Competitor Response Time Measures how long, on average, it takes a competitor to respond to other competitors' price changes. Product Descriptors Act as control variables that absorb variance due to the specific product differences (e.g., price and sale item popularity). Industry Effects Controls for industry-specific aspects of price competition in Internet selling (e.g., for the music CD and bookselling industries). Specification of the Empirical Model Variables We next discuss a number of issues that arise in the definition of our model’s variables, and in the details of how we test our model with them. Competitor Behavior. One of our hypotheses, the Modified Tit-For-Tat Hypothesis, states that firms will respond in a like manner to the competitive actions of their competitors. This is more than just a petty “tit for tat” strategy, but rather is indicative of a larger issue. When a competitor tries to use price competition to drive a firm out of business rather than collude, a viable strategy is to respond by competing as fiercely as possible, especially if there is a good chance of a loss of market share. Hence, collusion will only tend to take place when all competitors collude. If a single strong competitor (e.g., CDNow) decides not to collude, the industry environment may be changed from one of general collusion to one of intense competition. One problem we encountered when operationalizing the Competitive Behavior variable is that there is often a competitive or collusive response to a firm that had very little price changes and no price changes that were considered either competitive or collusive. For example, Best Buy’s online sales unit typically charges more than other vendors in several product areas. Best Buy does make price changes, but in this study, Best Buy did not match or beat any competitor’s price for any product during any period that we observed. As a result, since Best Buy made no competitive or collusive price changes, we are not able to discern Best Buy’s competitive or collusive behavior; in effect, Best Buy was missing data for Competitive Behavior. Our approach, with this problem in mind, was to remove from our study any price change that was a response to a competitor who had no competitive or collusive price changes. This reduced our data from 494 records to 475 records. All 19 records that were removed were from the music CD market, and exhibited competitive responses to Best Buy. This is explainable in that some CD vendors (e.g., Best Buy) rely on brand name recognition or niche positions, and do not wish to participate in the competitive pricing that seems to be occurring in the online CD market. Overall, the omitted data accounted for 1.8% of the total observations, and since competitive responses make up 68% of our data, this minimizes the likelihood that we have introduced unacceptable biases into our data set. Competitor Relative Market Power. Our model uses the number of unique users on a Web site as a proxy for market power of a firm. For this study, we use a simple ratio for Competitor Relative Market Power: market power of the competitor divided by the market power of the firm (measured in terms of unique users). However, this ratio is not intended to have a linear effect; market power then can be infinite in a single direction. But a logarithmic transformation can resolve this problem. For instance, PC Data online reported that eCampus (www.ecampus.com) had 110 unique users during the month when this study took place. Amazon, on the other hand, had 14,812 unique users. If eCampus responds to Amazon’s price, the (untransformed) market power ratio would be 14,812/110 = 134.65. On the other hand, if Amazon responds to eCampus, the market power ratio would be 110/14,812 = .0074. Taking the natural logarithm of these numbers, eCampus’ response to Amazon is ln(14,812 /110) = 4.90, and Amazon’s response to eCampus is ln(110/14,812) = - 4.90. This logarithmic transformation not only generates a sign for the ratio describing if the competitor has more (+) or less (-) market power, but also preserves the magnitude of difference between firms and reduces the corrupting effects of strong or weak outliers. Response Time. We include two variables to examine price time dynamics based on our theory. Competitor Response Time is a variable to measure the average time it takes a competitor to respond to price changes. One of our hypotheses, the Competitor Response Hypothesis, suggests that firms will not want to intensely compete with a competitor that has the capability to respond quickly to their price promotions. If a competitor responds quickly, the result will be reduced revenue for both firms. Profitmaximizing firms will realize that any competition with a competitor that tends to respond quickly will result in reduced profits with no increase in market share, and, therefore, will tend to act collusively with that firm. Hence, we think we will see more collusive behavior when firms are responding to competitors who are able to quickly respond reciprocally, and more competitive behavior from firms who are responding to competitors that have not shown the ability to quickly respond reciprocally. On the other hand, our Firm Response Time hypothesis states that, from a firm's perspective, it seems that if they can respond more quickly, they would tend toward competitive behavior. By having the ability to quickly respond to its competitors, a firm may feel that it can profit from competitive behavior, at least more than a firm that responds more slowly. Hence, we think we will see more collusive behavior when firms are able to respond quickly to competitors’ price changes, and more competitive behavior from firms that are not able or willing to respond quickly to competitors’ price changes. The Firm Response Time variable measures how quickly that particular response is to a competitor price change. To determine response time when collusion occurs, we assume that the firm is responding to the competitor who last made a price change with the same amount as the firm. For competition, we assume the firm was responding to the competitor with the lowest price listed for the same product. In the case of a tie with either the competitive or collusive responses, we assume the firm was responding to the competitor with the largest market power. Endogeneity. One of the major potential pitfalls encountered with this data set is the presence of endogeneity. For example, Amazon reacts to BN.com’s price changes, and BN.com reacts to Amazon’s price changes. If such a reaction occurs in a time frame that is shorter than the data collection intervals that we employ, then endogeneity will result. We will not be able to distinguish its impact. In this example, if BN.com and Amazon both change their price on a book on the same day, we do not know if BN.com responded to Amazon, or Amazon responded to BN.com. Since our data were collected once per day (instead of on a multiple collection intra-day basis), same day reactions to price will result in endogeneity and will reduce the predictive power of the estimates that we can obtain from our model, since simultaneous predictors are omitted. Thus, if we interpret endogenous reactions incorrectly, we can corrupt the parameter estimates produced by our model. For instance, if we say that BN.Com responded to Amazon when, in fact, Amazon responded to BN.Com, then our estimates of the Competitor Relative Market Power and Competitor Behavior coefficients will be biased. For this research, we resolve the same-day pricing action endogeneity by eliminating same-day reactions. This resulted in the elimination of 32 price changes that were responses to competitors who made price changes on the same day out of our 475 remaining competitive or collusive price changes. These 32 price changes included 15 competitive price changes and 17 collusive price changes. This resulted in 443 price changes that were used for our study. Because of the relatively small number of observations, and because of the balance of collusive and competitive price changes, we do not that this significantly biases the results we will report. By removing concurrent responses and only considering reactions to previous period competitors, we have safely removed the possibility of endogeneity from our study, since endogenous reactions are only possible in concurrent situations. Product Descriptors. We added two product control variables to help explain variance at the product level. Product Rank is the rank shown by the listing agent, and goes from 1 to 100. Our datacollection methodology involved downloading data from best-selling music CD and book lists from Billboard Magazine and USA Today. Each of these lists tells what item is selling the best, what is second, etc., up until the 100th best-seller in each industry at the time the competitive or collusive response was made. We include Product Rank to absorb information about how item popularity can affect competitive actions. Product Rank may intensify a firm’s propensity to either act collusively or competitively. We also include List Price to represent information about how much an item costs. Higher-priced items tend to generate larger returns, and may affect a company’s decision to compete or collude. Since collusion is typically more profitable and intense competition is typically more risky, high priced items should show a tendency toward collusion, not competition. The Final Probit Model The full form of the probit model that we will use for the empirical analysis of our data is shown in Equation 6 and Table 2. Prob[ycompetition=0, ycollusion=1] = α + β1 CompetitorBehavior + β 2 ln(CompetitorMarketPowerRatio) + β 3 FirmResponseTime + β 4 CompetitorResponseTime + β 5 IndustryEffects + β 6 ListPrice + β 7 ProductRank + ε (6) Table 2. Variables in the Probit Model VARIABLES DESCRIPTION α Intercept. β 1, ..., β 7 Coefficients of the model variables. Prob[ycompetition=0, ycollusion=1] A binary variable coded to indicate a firm’s specific strategic response to a competitor price change: competitive = 0; collusive = 1. CompetitorBehavior Ratio of collusive actions to competitive actions of the competitor. CompetitorMarketPower Ratio The ratio of the number of unique visitors to a competitor’s Web site to the number of unique visitors to the firm’s Web site. FirmResponseTime The number of days since last competitor price change that was: • the exact same price, if collusion is occurring; or, • the closest higher price, if competition is occurring. CompetitorResponseTime The number of days on average that it takes for a competitor to respond to price changes. IndustryEffects A binary variable that captures industry effects, with bookselling industry = 0, and music CD industry = 1. ListPrice The list price of the product. ProductRank The current rank of the product among the best-selling 100 items ranked by USA Today (books) or by Billboard Magazine (music CDs). ε An error term for the probit regression. Collinearity and Multicollinearity. Strong pairwise correlation between independent variables can lead to severe estimation problems of coefficients, as can perfect collinearity (or multicollinearity) among multiple variables (Kennedy 1998). We examined the correlation matrix and found that there were no unacceptable pairwise correlations (in excess of 65%); instead, the highest single pairwise correlation was only 38%. A simple pairwise correlation test only tests linear relationships between two variables, and not relationships between an independent variable and a linear combination of two or more other independent variables. Thus, multicollinearity can still exist even though the correlation matrix shows no high correlation between predictor variables. Moreover, multicollinearity cannot be readily observed by eyeballing the data, it is valuable to test for its presence. To test for close-to-perfect collinearity between multiple variables, a condition number test is useful. Greene (1999) describes the condition number of a square matrix as the square root of the largest and smallest matrix roots of the normalized X’X matrix that is formed in regression analysis. A condition number of 20 or greater may indicate the presence of collinearity between one independent variable and a linear combination of other independent variables. We performed this test with our data, and derived a condition number of 7.3. Thus, with such a low condition number, we are assured that multicollinearity does not exist in our data set. RESULTS AND DISCUSSION The results of our study are promising. We estimated the probit model in Equation 4 with 443 observations, using LIMDEP 7.0 (www.limdep.com), an econometrics and statistics package described in Greene (1995). The results converged in five iterations to yield parameter estimates. Empirical Results Model Parameter Estimates. Probability tests for the model yielded χ2=190 with 7 degrees of freedom, which is well below the 1% probability needed to reject the null hypothesis and accept the validity of the model. Table 3 shows our parameter estimates for each variable in the probit model. (See Table 3.) Table 3. Parameter Estimates for the Probit Model VARIABLE α (Intercept) COEFFICIENT STANMARGINAL STANt-STAT DARD EFFECTS DARD ERROR ERROR -0.117 0.295 -0.047 0.118 0.397 CompetitorBehavior 0.696 0.267 0.277 0.107 2.605*** ln(CompMarketPowerRatio) 0.108 0.028 0.043 0.011 3.844*** FirmResponseTime 0.001 0.006 0.000 0.002 CompetitorResponseTime 0.040 0.014 0.016 0.006 2.853*** -1.787 0.193 -0.713 0.077 -9.262*** ListPrice 0.022 0.012 0.009 0.005 1.840* ProductRank 0.001 0.003 0.000 0.001 0.306 IndustryEffects 0.199 Note: *=significant at the p ≤ 10% level; ***=significant at the p ≤ 1% level. The reader also should note that the CompetitorMarketPowerRatio coefficient cannot be interpreted as the marginal effect of the related variable in the usual sense, since the data that are used to construct the variable undergoes a transformation. Instead, the marginal effects apply to the transformed data. In our probit model, we used a dummy variable for type of response, either collusive or competitive. We coded collusive responses as 1 and competitive responses as 0. Hence, positive coefficients indicate that an increase in the variable causes a tendency toward more collusive reactions, while negative coefficients indicate that an increase in the variable creates a tendency toward more competitive reactions. Marginal Effects. Liao (1994) and Greene (1996) describe how the magnitudes of coefficients in a binary choice model like probit can be misleading. The dependent variable is a probability. Unlike linear models, a change in a coefficient of an independent variable should not be used to predict a correlated response of the dependent variable in a probit model. Greene (1996) develops a marginal effects measure of coefficients that can be used for probit models. He shows that, in order to determine the marginal effect of a variable, the independent variables must be scaled using a variable, z, derived from the Z distribution, on which the probit model is based. For instance, in our binary probit model, the effect of a change in CompetitorBehavior on propensity to compete, when all other independent variables are held constant, can be represented (in simple terms) by measuring an increase that a one unit change in an independent variable has on the dependent variable is to calculate over continuous variable z the change in one unit, as shown by Equation 7. Effect on E [y1|CompetitorBehavior1] = E [y1| z CompetitorBehavior1 = 1] – E[y1 | z CompetitorBehavior1 = 0] (7) The actual derivation for all variables to determine marginal effects is quite complicated and its complexity increases geometrically as more variables are considered. Fortunately, LIMDEP 7.0 (Greene 1995) has added the ability to calculate marginal effects of probit coefficients, following Greene's (1996) research. The reader should note that the marginal effect coefficients allow for clearer understanding of the impact a variable has on the probability of an observed behavior. We believe that the marginal effect standard errors offer similarly useful interpretative power. So both columns are included in our results. Model Fit. Kennedy (1998) notes that there is no universally-accepted goodness of fit measure, such as R2 for linear regression, for binary choice models such as probit and logit. It is also considered poor form to simply take a measure of the percentage of correctly predicted observations, because in a binary choice model, predictions may predict well overall because they are strong predictors of one choice but weak predictors of other choices. For example, with our data, if we predict that competition always occurs, we would be right about 51% of the time (224 / 443). However, we would never successfully predict a collusive outcome, and thus would be mislead into thinking that competitive responses are a natural assumption. Table 4 shows the actual versus predicted values of competitive and collusive price changes from the online CD and book industries. (See Table 4.) Table 4. Frequencies of Actual and Predicted Outcomes for the Probit Model PREDICTED ACTUAL Collusion Competition Total Collusion Competition Total 162 57 219 30 194 224 192 251 443 As Table 4 shows, our probit model correctly predicted collusive price changes 74% of the time (162 / 219) and correctly predicted competitive price changes 87% of the time (194 / 224). McIntosh and Dorfman (1992) suggest a technique that involves adding the sum of the percentage of correct predictions in each category. They claim that the measure is of value if this sum is greater than 100%. The sum of the percentage of correctly predicted competitive responses (87%) plus the sum of the percentage correctly predicted collusive responses (74%) is 161%, which is well above their threshold. While the traditional R2 measure of fit used in linear models is not valid for binary choice models, such as logit and probit, several psuedo-R2 measures have been developed to measure goodness of fit of binary choice models. Many of these measures are evaluated by Veall and Zimmermann (1996), and they recommend a test, developed by McKelvey and Zavoina (1975), that estimates R2 from the underlying latent linear regression. In the LIMDEP 7.0 documentation, Greene (1995) calculates the McKelvey and Zavoina test, R2MZ_LIMDEP, using the latent regression Yi* = x i* β + ω,, where ω is the inverse Mill’s ratio.1 The LIMDEP implementation of the McKelvey and Zavoina psuedo-R2 test yielded a value of R2MZ_LIMDEP = 58.4%, which indicates a very good degree of predictability for the estimated model. Veall and Zimmermann (1996) argue that Greene’s use of the inverse of the Mill’s ratio (ω ) calculates R2MZ conditionally upon the discrete outcomes, and therefore seriously overestimates R2MZ as defined by McKelvey and Zavoina. Instead, Veall and Zimmermann suggest that R2MZ_VZ should be derived using Yi* = x i* β . For our data, the Veall and Zimmermann implementation of the McKelvey and Zavoina psuedo-R2 test yielded a value of R2MZ_VZ = 48.4%, which, while less than the R2MZ calculated using the methodology suggested by Greene, still indicates a good predictive capability for our probit model. 2 Results of Hypotheses Tests Our model verifies three of our four hypotheses and reveals some interesting relationships. As we mentioned earlier, the dynamics of competition in different industries are somewhat different, and may cause the same firm to act differently when competing in different industries. In addition, we described how firms would be more likely to collude when sale items are more expensive, since more revenue is at stake. This is weakly shown by our results in the ListPrice variable, which was significant 1 Mill's ratio is defined as the ratio between the density and distribution of a set of numbers. Mill’s ratio is calculated using R(x) = (1-Φ(x)) /φ (x) where Φ(x) and φ(x) represent, respectively, the standardized normal density and distribution functions (Ruben 1961). Historically, it has been approximated with numerical methods. Ruben (1961) claims that Laplace (1812) is the first known recorded use of an approximation of the Mill's ratio. 2 A weaker linear test, which is known to be less applicable when probit models are used, resulted in an R2 = 37.8% with the same signs on all predictor variables. The linear model showed the intercept to be significant and the list price to be significant with 13%. All other variables showed the same level of significance. We again caution the reader to recognize that fit of econometric models of binary choice and categorical choice are underestimated with traditional R2 values using a linear test. See Kennedy (1998, p.233) for further explanation. within the 10% level and has a positive marginal effect (Marginal Effect = 0.009). A positive marginal effect for a variable in our probit model is consistent with increases in the firm’s propensity to make collusive price reactions. Product popularity, proxied by ProductRank, was not significant, so product popularity does not appear to have an impact on the decision of whether to collude or compete. However, the lack of a statistically significant coefficient does not lead us directly to a conclusion that the related variable is unimportant as an explanatory factor. In this study, we only considered the top-100 products in both industries, when there are many, many more products that firms are selling via the Internet. We feel more comfortable suggesting (though not confirming) that firms do not distinguish between the popularity of the top-100 best-sellers. Comparisons between best-sellers and weaker-selling products (those that sell far below the top-100 books and music CDs) are beyond the scope of this study. More tests should be conducted to examine the affect of a best-seller with a weaker-selling product in these areas. We next discuss each of our hypotheses, and what we learn about their applicability based on our empirical results. Modified Tit-for-Tat Hypothesis (H1) (supported at the 1% significance level). We test HI by examining the CompetitiveBehavior variable. The marginal effect estimate for this variable is positive and significant (Marginal Effect = 0.277), indicating that an increase in a competitor’s competitive responses causes a firm to act more competitively. Thus, we show that firms echo their competitor’s typical behavior. Firms tend to collude with a colluder and compete with a competitor. This makes sense in light of our analytical model, which shows that collusion is more likely when there is a high probability (in terms of λ in Equations 2, 3, and 4) to accurately detect collusive behavior from a competitor. Industry Corollary (HC1) (also supported at the 1% significance level). HC1, the Industry Corollary, tests whether different industries exhibit different competitive or collusive responses, even when some of the same firms occupy both industries. Our results clearly show that the industry is likely to affect the behavior of the firms within that industry, based on the marginal effect estimate for the IndustryEffect variable (Marginal Effect = -0.712). We speculate that CDNow, the market leader in the music CD industry, led the Internet-based music CD industry towards greater competition towards an intensely competitive environment by its own competitive behavior. Conversely, Amazon, the market leader in the Internet-based bookselling industry, led the market towards collusive behavior based on its own collusive actions. These market leaders’ actions led to a “ripple effect” which caused the rest of the industry to imitate their strategy. Asymmetric Competition Hypothesis (H2) (also supported at the 1% significance level). We tested H2 by examining the ln(CompetitorMarketPowerRatio) variable. The marginal effect of this variable was significant and had a positive coefficient (Marginal Effect = 0.043), indicating that the more market power you had when compared to your competitor, the more likely you were to act collusively. In terms of our analytical model, actions of the low smaller firms have little impact (e.g., low γ in Equations 3 and 4) on the sales and revenue of market leaders. This empirical test shows the existence of asymmetric competition within EC industries. Thus, intense competition predicted by Bertrand (1883) for identical items does not appear to be occurring. Competitor Response Time Hypothesis (H3a) (supported at the 1% significance level). H3a tests for whether a high average price change response time for a competitor leads to greater propensity for collusive behavior in a firm, via the estimated coefficient of the CompetitorResponseTime variable. Implicit in this hypothesis is the assumption that, if a competitor possesses the ability to respond more quickly to a competitor, it would do so, and that there are not other factors that would cause competitors to delay this response. The results indicate support for the hypothesis, with a positive estimated marginal effect (Marginal Effect = 0.016). In addition, the marginal effect is positive, as we would expect with this hypothesis. Thus, we accept this hypothesis that firms who are able to respond quickly coerce competitors into less competitive behavior and more collusive behavior. Firm Response Time Hypothesis (H3b) (not supported within the 10% significance level). H3b tests if increased unit price change response time of a firm leads to greater propensity for the firm's competitive behavior, via the estimated coefficient of the FirmResponseTime variable. Again implicit in this hypothesis is the assumption that, if a firm possesses the ability to respond more quickly to a competitor, it would do so, and that there are not other factors that would cause a firm to delay this response. Although we considered a quick response to allow competitive behavior, our results were not significant (p = 84%). Thus, we reject this hypothesis that ability to respond more quickly leads to more competitive behavior and believe that rejection of the hypothesis suggests (but does not prove) that firms do not consider their own response time, but rather only their competitors’ response time when making price strategy decisions. Further research in this area will shed more light on how a firm’s (rather than competitors’) ability to quickly respond affects its own pricing strategy. DISCUSSION Both our theoretical and empirical models show how EC technology increases the possibility of collusion because EC technology can be used to more quickly respond to competitors' actions, thus reducing any benefit from price cuts. In this section, we show that industry-wide affects competition and collusion as predicted by our models, and we initiate a discussion about CDNow's competitive behavior, and the Internet CD and Bookselling industries in general. This section ends with implications for other industries brought to light by this study. Evidence of Collusion and Competition at the Aggregate Industry Level Figure 4 shows price change behavior among the top three firms (e.g., the market leaders) in each industry we examined. In the bookselling industry, we considered the actions of Amazon, BN.Com, and Borders.com, which sell to about 95% of the books based on both unique user data obtained by PC Data Online. Brynjolfsson and Smith (1999) agree with this assessment, indicating that the “Big Three” online bookstores represent an estimated 88% of the total market share. There were 44 collusive responses (61%) and only 28 competitive responses (39%) among these market leaders. The situation is exactly reversed in the CD industry. Using the same technique, we identify that the top three CD Internet sellers are CDNow, Amazon, and BN.Com with 83% of the unique hits in the CD industry. The EC CD market leaders had only 19 (43%) collusive responses and 25 (57%) competitive responses. These preliminary results show how industry structure and dynamics help define how much collusion and how much competition takes place. Figure 4. Market Leaders’ Responses to Other Market Leaders’ Prices We also examine the aggregate of competitive or collusive responses of all the Internet sellers in our study. Figure 5 shows, in aggregate, competitive or collusive responses from all competitors in our study. The low level of competitive responses in the bookselling industry indicates that most participants practice collusion. In addition, the extremely low number of 30 competitive responses (15%) and the high number of 169 collusive responses (85%) indicate that market leaders are willing to “go down” to collude with smaller competitors. Such actions by the market leaders make price competition of smaller companies infeasible in the EC book industry since competitive price changes will be matched. Figure 5. Industry-Wide Competitive and Collusive Responses to Competitors Conversely, the CD industry has only 81 collusive responses (27%) and 214 competitive responses (73%) overall. This preliminary indicates that the music CD industry is intensely competitive. Rather than adopting a definitive stance on competition or collusion as the appropriate EC pricing strategy, this research tests what competitive models are used and when they are appropriate. Qualitative Industry Analysis On the surface, one expects that the CD industry and the bookselling industry should be quite similar in their tendency to exhibit collusive and competitive pricing on the Internet. Both industries sell identical commodity-like products to end users through similar Internet-based interfaces. In addition, both industries have Amazon and BN.com as two of their top three competitors. However, the largest competitor in the CD industry during the time frame that our data were collected was CDNow (Brady 2000), and CDNow, as the market leader, may be the reason for the change in marketplace dynamics. Taylor and Jerome (1999) documented intense price competition between CDNow and N2K's Music Boulevard, and also noted how CD companies are forced to compete against free MP3 music downloads from Web sites such as Napster. Hill and O'Brien (1999) describe how Amazon's entry into CD music sales forces CDNow to intensely compete on price because of worries that Amazon's large Web presence would erode CDNow's customer base. English (1999) echoes this sentiment and describes how CDNow may be forced to lower prices because a large customer base is initially considered more valuable than current period profits. Thus, CDNow competed fiercely with its competitors despite the benefits of collusion. When we examined high level of competition in the CD industry in the time frame of our data set, we found that CDNow made competitive responses about 67% of the time and collusive responses only 33% of the time. Compare this with Amazon, a market leader in both industries. Amazon made competitive responses to other market leaders only 27% of the time and collusive responses 73% of the time. If our interpretation is correct, Amazon tacitly cooperates with other Internet-based booksellers, and these firms appear to be willing to allow (or perhaps are even forced to allow) such collusion to occur in the marketplace. Conversely, CDNow does not appear to be cooperating with the other firms, thus making the music CD industry extremely competitive, new entry difficult, and probably hurting all competitors, especially smaller firms. Thus, the evidence that we have points to CDNow’s propensity to use price competition to capture market share while driving its competitors out of the CD business and discouraging new entrants. In terms of Equation 4, CDNow can be thought of as having a high value of θ, the probability of eliminating the entry of a new firm or eliminating a current competitor. As a result, CDNow’s competitive actions would result in less competition in the future because of a large brand-name preference. CDNow, therefore, avoided collusion and embraced intense competition. Clearly, as defined by Equations 2 and 3, without tacit cooperation, tacit collusion is impossible. CDNow's strategy did not appear to drive out new sellers, however, as Amazon, BN.Com, and Borders.com all began selling CDs within the last two years. Furthermore, CDNow appears to have hurt its own performance. After the data in this study was collected, CDNow was forced sell out to German media conglomerate, Bertelsmann, for a mere $117 million, despite having the largest at home purchasing market of any Internet-based CD seller (Farmer 2000). This amount is extremely low, especially when considering that CDNow purchased N2K for $522 million just a year earlier (Peterson 1999). Implications for Other Industries In this study, we show how EC technology gives Internet-based sellers the ability to tacitly collude. We show that the bookselling industry is collusive, while the music CD industry is competitive. The industries are similar, selling copyrighted work developed by others, produced and duplicated by a publisher, and distributed by a distributor. Both industries share several top-tier vendors, including Amazon, BN.com, and Borders.com. Yet our research shows that different pricing strategies are employed in each industry. We attribute this difference to the presence of CDNow, which has shown itself to be intensely competitive. CDNow, perhaps as a result of the approach it took to its marketplace, recently became financially troubled and was forced to sell out to Bertelsmann for what many would call a low price. Although we do not examine other industries in this study, our model shows how companies should tacitly collude unless there is a chance for market dominance by driving out less-efficient competitors, or by erecting barriers to entry to prevent future competitors, thus increasing the potential profit from the competitive action. In EC industries, pricing strategy is vital for success, and, as shown by CDNow, the wrong strategy by a market leader can place the leader in a competitive environment of its own making that reduces profits to minimal levels and makes survival difficult. EC companies of all types can examine these results and apply our findings to their own situations. By examining the competitive actions of the market leaders, new entrants will understand the competitive demands of the industry they are entering. Existing companies can use this discussion to clarify their existing strategy, and to decide the best way to compete in the future. Finally, as this research shows, a firm’s ability to effectively respond reduces competitive behavior from other firms in an industry. Since fast response time can lead to a less competitive environment, firm revenue and survivability can be enhanced by employing technology to rapidly respond to competitors. CONCLUSION Because of recent developments in the new electronic markets of the Internet, IS researchers know relatively little about whether Internet-based sellers intensely compete or tacitly collude. In this paper, we have made the case for the possibility that collusion is a part of their repertoire of potentially effective strategy. Our model suggests how EC technology enables firms to be extremely responsive to pricing changes for products. This responsiveness aids in strategizing with respect to price changes. It ultimately enables firms to tacitly collude with each other to help ensure higher profits for all. Our perspective is a sharp contrast to the primary findings of existing IS literature, which maintains that technology will force Internet-based sellers to compete with intense Bertrand competition. The results of the in-depth examination of the online bookselling and CD industries in this paper initially surprised us. We guessed that the CD and bookselling industries would be quite similar. They have similar competitors, especially in the top tier, and the products that they sell are similar. Both booksellers and music CD sellers sell products that are developed by an author or artist, produced by a publisher or label, and sold to a distributor. However, we show in this paper that firms within these industries act quite differently. We attribute this to the intensely competitive actions of CDNow, a market leader in the CD industry that does not compete in the bookselling industry. We conclude that an “unfriendly competitor” who chooses not to tacitly collude will cause other firms to compete more intensely, especially with that unfriendly competitor. This paper also shows that the same firms can act differently in different environments, and that the competitive choices of the market leaders in an industry can determine the nature of competitive interactions throughout the industry. Such interaction can lead to an environment that is either collusive or highly competitive. Our econometric estimates of collusion and competition confirm our hypotheses describing how a firm’s choice of strategy is influenced. It turns out, according to our findings, that that they are influenced by the actions of the competitor to whom the firm is responding, by the market power of that firm relative to the market power of the competitor, and by competitors’ ability to immediately respond. We also show that a firm may choose to intensely compete rather than collude to drive current and future competitors from the market, but this strategy is risky. Perhaps the best real world example that we can offer that parallels our finding is the failure of CDNow, which failed to succeed in the competitive online CD environment that it created through its pricing strategy. This research makes several contributions. First, it is one of the first multi-industry empirical studies of Internet sellers’ competitive pricing behavior. Therefore, this research gives EC researchers better insight into the effects that technology has on Internet seller pricing strategy, especially in the bookselling and CD industries. The results of this study are surprising, and our econometric model should be useful for IS and EC researchers who wish to study other contexts. Second, this study incorporates tacit collusion theory and demonstrates that collusion is a viable option to intense Bertrand competition in Internet-based selling. This is particularly interesting since the inclusion of tacit collusion theory gives Internet-based sellers the opportunity to make larger profits. This is in sharp contrast to Bertrand competition, which is typically used in IS literature to describe competitive interaction between Internet-based sellers. Third, this study makes a contribution by showing how large firms in an EC environment set the tone for actions in the entire industry because EC technology. This observation can allow more in-depth study of the competitive interaction within industries and also will aid businesses develop their own strategy. Fourth, this study shows the strategic importance of developing EC technology in a firm that can respond to a competitor's price promotions. Such investment in technology can allow more profitable strategies for the firm. For academic researchers, it is important to understand the effect that Internet technology has on an industry, whether that effect is collusive or competitive. By allowing for more than an intensely competitive response, researchers can examine firms with greater dimensions and therefore develop greater insight. For managers, this research not only adds to general understanding of EC product pricing dynamics, but also can be used to develop an optimum pricing strategy depending on the competitive environment of a firm. By evaluating the actions of the market leaders, managers can develop strategies that fit better with their industry. By understanding the impact of their own actions, including increasing their response time, Internet sellers can help define the very environment in which they compete. This study has two limitations. First, the data collection technique that we used ignored roughly 75% of price changes that we considered to be neither competitive nor collusive (e.g., price increases as a means for price discovery). This was because we wanted to simplify the choice structure of our empirical model to clarify the choice between competition and collusion. Because we used a probit model, our coefficient estimates should not be affected by the addition of more pricing alternatives if a more robust model were to be examined. These price changes need to be investigated in future research to gain a greater understanding of pricing in Internet-based selling. Second, we assumed that any undetected price changes were zero. Several times, especially with small companies, Web sites were unreachable or the required price information was not available. Assuming price stability in these cases, in our view, is valid and conservative. REFERENCES Alba, J., Lynch, J., Weitz, B., Janiszewski, C., Lutz, R., Sawyer, A., and Wood, S. 1997. “Interactive Home Shopping: Consumer, Retailer, and Manufacturer Incentives to Participate in Electronic Marketplaces,” Journal of Marketing 61: 38-53. Bailey, J. P. May 1998. "Intermediation and Electronic Markets: Aggregation and Pricing in Internet Commerce," Ph.D. Dissertation, Program in Technology, Management and Policy, Massachusetts Institute of Technology. Bakos, Y. 1991. "A Strategic Analysis of Electronic Marketplaces," MIS Quarterly, 15(3): 295-310. Bakos, J.Y. 1997. "Reducing Buyer Search Costs: Implications for Electronic Marketplaces," Management Science 43 (12): 1676-1692. Bakos, J.Y., Lucas, H. C., Jr., Oh, W. S., Viswanathan, S., Simon. G, and Weber, B. November 1999. "Electronic Commerce in the Retail Brokerage Industry: Trading Costs of Internet Versus Full Service Firms," working paper, Stern School of Business, New York University, New York, NY. Bertrand, J. 1883. "Review of Theorie Mathematique de la Richesse Sociale and Researches sur les Principes Mathematicque de la Theories des Richesse," Journal des Savants: 499-508. Blattberg, R. C. and Wisniewski, K. J. 1989. “Price-Induced Patterns of Competition,” Marketing Science 8 (4): 291-299. Brady, M. July 13, 2000. "When Giants Stumble," E-Commerce Times, http://www.ecommercetimes.com/news/viewpoint2000/view-000713-1.shtml. Brynjolfsson, E. and Smith, M. December 1999. "The Great Equalizer? The Role of Price Intermediaries in Electronic Markets," Workshop on Information Systems and Economics (WISE-99), Charlotte, NC. Brynjolfsson, E. and Smith, M. 2000. "Frictionless Commerce? A Comparison of Internet and Conventional Retailers," Management Science 46 (4) 563-585. Campbell, C., Ray, G., and Muhanna, W. A. December 1999. "Search and Collusion in Electronic Markets," Workshop on Information Systems and Economics (WISE-99), Charlotte, NC. Carpenter, G.S., Cooper, L.G., Hanssens, D. M., and Midgley, D. F. 1988. "Modeling Asymmetric Competition," Marketing Science 7 (4): 393-412. Carvajal, D. May 18, 1999. “3 On-Line Book Retailers Cut Prices on Best Sellers.” New York Times, New York, NY Late Edition (East Coast); pg. C.10 Chamberlain, E. 1929. “Duopoly: Value Where Sellers Are Few,” Quarterly Journal of Economics 43: 63-100. Choudhury, V., Hartzel, K. S., and Konsynski, B. R. 1998. "Uses and Consequences of Electronic Markets: An Empirical Investigation in the Aircraft Parts Industry," MIS Quarterly 22(4): 471-507. Clemons, E. K., Hann, I., Hitt., L. M. 1998. “The Nature of Competition in Electronic Markets: An Empirical Investigation of Online Travel Agent Offerings,” Working Paper, The Wharton School, University of Pennsylvania, Philadelphia, PA. Dillard, M. 1999. The Economics of Electronic Commerce: A Study of Online and Physical Bookstores, Bachelor’s Honors Thesis, University of California, Berkeley, CA. Dixit, A. K., and Nalebuff, B, J. 1991. Thinking Strategically: The Competitive Edge in Business, Politics, and Everyday Life, W. W. Norton & Company, New York, NY. English, D. "How Low Can You Go? Finding Maximum Value on the Web" Computer Shopper, July 1999, available at http://www.zdnet.com/computershopper/edit/cshopper/content/9907/406116.html. Farmer, M. A. July 20, 2000. "Million Dollar Merger Helps CDNow Play On," CNET News.com, http://news.cnet.com/news/0-1007-200-2298875.html?tag=st.ne.1002.srchres.ni. Geweke, J., Keane, M., Runkle, D. 1994. Alternative Computational Approaches to Inference in the Multinomial Probit Model," The Review of Economics and Statistics 76 (4): 609-632. Greene, W. H. 1995. LIMDEP Version 7.0 User's Manual. Econometric Software, Inc. Bellport, NY. Greene, W. H. 1999. Econometric Analysis, 4th Edition, Prentice Hall, Englewood Cliffs, NJ. Greene, W. H. 1996. "Marginal effects in the Bivariate Probit Model," working paper, Stern School of Business, New York University, New York, NY. Haltiwanger J. and Jarmin, R. S. 2000. “Measuring the Digital Economy,” in in Brynjolfsson, E. and Kahin, B. (editors), Understanding the Digital Economy, MIT Press, Cambridge, MA. Hill, A and O'Brien, B. February 12, 1999. "The Hard Edge: AOL Takes Over; Bill Fights the Net; and Alice's Shocking News About the Y2K Bug," Computer Shopper. Available at http://www.zdnet.com/computershopper/stories/reviews/0,7171,387610,00.html. Kauffman, R. J., March, S. T., Wood, C. A. December 2000. "Mapping Out Design Aspects for DataCollecting Agents." International Journal of Intelligent Systems in Accounting, Finance, and Management (in press). Kauffman, R. J. and Wood, C. A. August 2000a. "Running Up the Bid: Modeling Seller Opportunism in Internet Auctions," in Proceedings of the 2000 Americas Conference on Information Systems (AIS/AMCIS 2000), Chung, H. M. (ed.), Long Beach, CA. Kauffman, R. J. and Wood, C. A. December 2000b. "Follow the Leader? Strategic Pricing in ECommerce," in The 21st Annual International Conference on Information Systems (ICIS-2000), Ang, S., Krcmar, H., Orlikowski, W., Weill, P., and Degross, J. (eds.), Brisbane, Australia. Kennedy, P. 1998. A Guide to Econometrics (4th ed.). MIT Press, Cambridge, MA. Lal, R. and Sarvary, M. 1999. “When and How Is the Internet Likely to Decrease Price Competition,” Marketing Science 18(4): 485-503. Liao, T. F. 1994. “Interpreting Probability Models: Logit, Probit and Other Generalized Linear Models,” Sage University Series on Quantitative Applications in the Social Sciences, Series No. 07-101, Sage Publishers, Thousand Oaks, CA. Lynch, J. G., Ariely, D. 2000. “Wine Online: Search Costs and Competition on Price, Quality, and Distribution,” Marketing Science 19(1). McFadden, D. 1994. “Qualitative Response Models,” in Griliches, Z., and Intriligator, M. (editors), Handbook of Econometrics, Volume 2, North-Holland, Amsterdam, Netherlands, 1395-1458. McIntosh, C. S. and Dorfman, J. H. 1992. "Qualitative Forecast Evaluation: A Comparison of Two Performance Measures." American Journal of Agricultural Economics, 74: 209-214. McKelvey, R. D. and Zavoina W. 1975. “A Statistical Model for the Analysis of Ordinal Level Dependent Variables.” Journal of Mathematical Sociology 4: 103-120. Peterson, A. "CDnow and N2K Make Music Together: The $522 Million Merger Is an Aggressive Attempt to Corner the Online Music Market," WSJ Interactive Edition, March 17, 1999, available at http://www.zdnet.com/zdnn/stories/news/0,4586,2227403,00.html. Ruben, H., "A New Asymptotic Expansion for the Normal Probability Integral and Mill's Ratio," Journal of the Royal Statistical Society, Series B (Methodological), 24 (1), 1962, pp. 177-179. Smith, M. D., Bailey, J., Brynjolfsson, E. 2000. “Understanding Digital Markets: Review and Assessment,” in Brynjolfsson, E. and Kahin, B. (editors), Understanding the Digital Economy, MIT Press, Cambridge, MA. Taylor and Jerome/Smart Business. "The Net Is a Casino for Imbeciles," Ziff Davis News Network, January 21, 1999, available at http://www.zdnet.com/zdnn/stories/comment/0,5859,2190580,00.html Tirole, J. 1998. The Theory of Industrial Organization, The MIT Press, Cambridge, MA. Varian, H. R. 2000. “Market Structure in the Network Age,” in Brynjolfsson, E. and Kahin, B. (editors), Understanding the Digital Economy, MIT Press, Cambridge, MA. Also based on a presentation in the Understanding the Digital Economy Conference, Department of Commerce, Washington, DC, May 2526, 1999. Veall, M.R. and Zimmermann, K.F. 1996. “Pseudo-R2 Measures for Some Common Limited Dependent Variable Models.” Journal of Economic Surveys 10: 241-259. Von Stackelberg, H., 1934. Marktform und Gleichgewicht, Springer, Vienna.