Universiteit van Amsterdam
The Price Dispersion Peculiarity : Seller
Heterogeneity and Pricing Strategies on
Amazon.co.uk
1
Abstract: Despite the existence of price comparison technologies in electronic markets, price dispersion online remains both systematic and persistent. Despite favourable market conditions, including almost free entry and exit, negligible search costs and information transparency, price dispersion remains prevalent. Moreover, evidence suggests that price dispersion is not a disequilibrium phenomenen that is being corrected over time, as empirical evidence shows that it does persist in the long-run suggesting that it might be an equilibrium phenomenen. This paper adds to the burdgeoning literature on the drivers of price dispersion in online markets. Seller heteregoneity has often been cited as a potential driver of price dispersion and this study adds to existing literature trying to understand to what extent seller characteristics can explain observed price dispersion. This paper is a systematic study on the morphology of price dispersion i.e. a study of the shape and structure of the distribution of prices at which an identical good is sold in a given market during a given period of time. Several potential drivers are investigated to determine how they contribute towards the observed distribution. Results suggest that reputation and price levels are negatively correlated whearas the number of ratings a seller hold is not correlated with price.
2 Table of Contents
1. Introduction ... 3
2. Overview of the Relevant Literature ... 6
2.1 Price Dispersion on the Internet ... 6
2.2 Drivers of Price Dispersion in E-Markets ... 6
2.3 Conclusion of Literature Review ... 15
3. Research Questions and Hypotheses Building ... 16
4. Data ... 19
4.1 Data Collection ... 19
4.2 Descriptive Statistics ... 21
5. Results ... 23
5.1 Price Distribution and Dispersion ... 23
5.2 Price Stability ... 24
5.3 Preliminary Results ... 24
5.4 Regression Analysis ... 25
6. Discussion... 29
6.1 Multiproduct Competition and Price Signaling ... 29
6.2 Regression Analysis Results ... 30
7. Limitations and Directions for Future Research ... 31
8. Concluding Remarks ... 33
References ... 35
3
1. Introduction
‘The explosive growth of the Internet promises a new age of perfectly competitive markets. With perfect information about prices and products at their fingertips, consumers can quickly and easily find the best deals. In this brave new world, retailers’ profit margins will be competed away, as they are all forced to price at cost.’ The Economist, November 20, 1999,
p. 112.
At the turn of the 20th century it was believed that as computers grew in personal use and as markets became accessible online, there would be a step change both in the way consumers shop and retailers operate. Whereas shopping in traditional bricks-and-mortar retailers can be time-consuming and arduous with retailers dispersed geographically, e-markets are able to offer consumers an array of retailers all accessible at the click of a button. Theoretically, online markets characterized by conditions similar to Bertrand competition e.g. minimal search costs, free entry and exit, and near perfect information were hypothesized to lead to the elusive ‘law of one price’. This is based upon the famous Bertrand competition model which suggests that if a market is served by two or more competing firms who sell a homogeneous good, any firm that sells a positive quantity in equilibrium will charge the same price (Azar, 2013). It was believed that e-markets could destabilize profitable oligopolistic pricing in homogenous goods markets thereby leading to price competition that would eliminate firms’ excess profits (Bakos and Yannis, 1997). Such market characteristics represent the true economic ideal of competition, which are “as close to Adam Smith’s notion of perfect competition as economies come” (Petrescu, 2011).
Whilst e-markets offer advantages to buyers in terms of lower search costs and hypothesized lower prices, e-markets can also be advantageous to sellers. Firstly, menu costs are minimal thus overall transaction costs should theoretically decrease. Secondly, since a retailer will only change a price if the expected marginal benefit exceeds the marginal cost of a price change, low menu costs offer the capacity for small and frequent price changes (Bailey, 1998). Moreover, retailers are in a position to quickly screen the offers of their competitors, allowing retailers to quickly react to price alterations. Furthermore, given that sunk costs are negligible in e-markets, the number of sellers would be expected to be increase enabling fierce competition. New entrants are theoretically advantageous in that they can introduce new products and processes, forcing incumbents to become more efficient and innovative and thus reducing the scope for monopoly pricing (Simon, 2005). Given the theoretical
4
predictions outlined above, e-markets were expected to display both price convergence and an increase in overall efficiency.
Early work on e-commerce centered on trying to establish whether, due to perceived efficiency gains, prices were in fact lower online than in traditional bricks-and-mortar retailers. However, early results were often ambiguous and empirical evidence mixed. Bailey (1998) for instance found that prices for books, CDs and software were higher online than in traditional bricks-and-mortar retailers. Conversely, Brynjolfsson and Smith (2000) found that prices were on average between 9 and 16 per cent lower online for books and CDs. Whilst early work on e-commerce did not reach a consensus on whether prices were indeed lower online, it did show that ‘price dispersion’ existed online. Price dispersion can be defined as the distribution of prices (such as range and standard deviation) of an item with the same measured characteristics across sellers of the item at a given point in time (Pan et al, 2004). Despite heady optimism about the potential efficiency gains in e-markets, a litany of subsequent empirical studies to date have reached a unanimous conclusion; namely that prices show no sign of convergence and that ‘price dispersion’ in e-markets remains commonplace.
The persistence of substantial levels of price dispersion for homogeneous products in well-structured e-markets continues to attract much theoretical and empirical attention. As Haynes and Thompson noted (2008, p. 46), the existence of such a wide distribution of prices is ‘one of the most curious empirical findings to emerge from the recent research on e-commerce.’ It represents an important research area as its existence defies economic logic that due to rationality and budget constraints, consumers will buy from the lowest priced seller when the good is homogeneous (Azar, 2013). Even if a degree of price dispersion is observed, a relatively small spread between the highest and lowest price would provide evidence of pricing efficiency (Tang and Xing, 2001). However, when price dispersion is significantly large this can indicate market inefficiency (Pan et al, 2004). This has implications for both consumer search and purchasing behaviour as well as for managerial pricing strategy and public policy (Baye et al, 2003).
It is thus the purpose of this paper to add to the growing body of literature on the drivers of price dispersion in e-markets. In particular, it focuses on the shape and structure of the distribution of prices at which an identical good, a digital camera, is sold at a particular marketplace, Amazon.co.uk, during a given period of time. This study will look in detail at seller characteristics in order to determine to what extent subtle differences between sellers can explain the distribution of prices in the market. Notably, this empirical study will focus on two publicly available measures of seller reputation and experience as well as a range of
5
further indicators which can differentiate sellers. Additionally, this study will test whether or not price dispersion remains stable from day to day or whether there is some evidence of mixed pricing strategy from e-tailers.
The rest of this paper will be organized as follows. Chapter 2 will provide an overview of the relevant literature highlighting the most prominent reasons both theoretically built and empirically tested in studies to date. Chapter 3 will construct the research question and introduce the hypotheses to be tested. Chapter 4 will describe the dataset and provide baseline descriptive statistics. Next, Chapter 5 will provide an overview of the price distribution in the market and subsequently econometric analysis will then test the research hypotheses. Chapter 6 will discuss the findings of this paper and consequently Chapter 7 will discuss limitations of the study as well as grounds for future research. Finally, Chapter 8 will provide a summation of the paper as well as concluding remarks.
6
2. Overview of the Relevant Literature
2.1 Price Dispersion on the Internet
Early research on price dispersion in e-markets, such as Brynjolfsson and Smith (1999), argued that price dispersion may reflect the random “noise” of an immature market and that prices would converge over time (Baylis and Perloff, 2002). Relatedly, Baye et al (2004) hypothesized that price dispersion may be a disequilibrium phenomenon that would be corrected over time. However, subsequent empirical studies to date find little sign of price convergence in homogenous goods markets and price dispersion remains an empirical regularity.
For instance, Brynjolfsson and Smith (1999) in one of the first studies to analyze price dispersion online found that prices for identical books and CDs at different online sellers differed by as much 50 per cent. Moreover, in a study on digital cameras and scanners Baylis and Perloff (2002) found that the price range of digital cameras was 42 per cent of its mean, while the price range of scanners was 29 per cent of its mean. Furthermore, Baye et al (2006) find persistent price dispersion for 36 homogeneous electronic products and found no evidence of price convergence even after an 18 month observation period. A litany of other notable studies (e.g. Ancarani et al, 2004; Baye et al, 2004; Bapna et al, 2006; Petrescu, 2011; Bounie et al, 2012) have consistently proven empirically that there is no evidence of price convergence in e-markets.
In addition, price dispersion has been observed across a variety of different product markets, across countries and has been shown to be persistent over time. The drivers of price dispersion in homogeneous product markets with seemingly low search costs and transparent information has been widely discussed in the literature. Attention will now turn to discussing the most widely cited sources of price dispersion to date.
2.2 Drivers of Price Dispersion in e-markets
The drivers of price dispersion in online markets have and continue to attract a considerable degree of attention in research literature. A number of potential sources of price dispersion have been identified and investigated and it is the purpose of this section to discuss the most prominently cited among them. The most prominent sources of price dispersion commonly
7
cited are: (i) Search cost theory, (ii) Market Characteristics, (iii) Seller Heterogeneity and (iv) Pricing Strategy. Each individual explanation will be analyzed in turn discussing both its theoretical foundation and subsequently the empirical evidence as to their contribution to price dispersion.
Figure 1 - Drivers of Price Dispersion
(i) Search Cost Theory
The existence of search costs is commonly cited as a primary driver of price dispersion in physical markets, where competitors are dispersed geographically. Such an explanation has also been given to price dispersion in online markets. The classic Bertrand competition model assumes that search costs are zero. Therefore, in the event where all buyers are perfectly informed about the available price alternatives, “perfect market conditions’’ would exist and any seller assigning a higher price that his/her competitors will not be able to sell anything
8
(Biswas, 2004). In reality although the internet has greatly reduced search costs, search costs still remain positive and these can provide frictions in the market avoiding the Bertrand competition outcome. There have been several search-theoretic models developed in order to explain price dispersion through search costs. For example, Carlson and McAfee’s model (1983) defined a search as an inspection of one alternative and assumed that buyers search sequentially until the marginal cost of an additional quote is more than the marginal gain. Moreover, Carlson and McAfee showed that if there is a wide distribution of search costs across consumers and if sellers’ costs are heterogeneous, price dispersion will be an equilibrium outcome (Pan et al, 2002). Price dispersion will thus increase if the distribution of consumer search costs becomes more dispersed.
Heterogeneity in search costs across individuals is a logical outcome in e-markets. Buyers may incur different search costs for instance due to their online shopping experience, skills of using search engines, and wealth levels (Feng et al, 2012). Intuitively, consumers willing to undertake longer search will be rewarded with a lower price. However, a proportion of consumers may prefer to buy from a seller that does not offer the lowest price because spending time and effort obtaining information about prices is costly (Azar, 2013). Consequently, consumers can be dichotomized into two distinct groups: informed and uninformed consumers. Informed consumers are willing to undergo extensive search of different products to find the one that offers the highest utility. Conversely, uninformed consumers undertake limited search until finding a product that offers positive utility (Feng et al, 2012). Moreover, consumers are heterogeneous in their degree of brand loyalty, and more brand loyal customers are more likely to limit their searches to specific brands (DiRusso et al, 2011).
Furthermore, if firms are aware that consumers are both imperfectly informed and that information is costly to obtain then firms can set prices to leverage consumer heterogeneity in information and search costs (Brynjolfsson and Smith, 2000). Whilst there are a variety of search theoretic models which try to explain price dispersion through search costs, empirical testing of such models is somewhat more difficult. Firstly, search costs are in general unobservable and what’s more the predicted impact of search costs on levels of dispersion depends on the model and the metric used for measuring price dispersion (Baye et al, 2006). However, in an empirical study on online travel agents, Clemons et al (2002) claim that sellers can overcome price competition by effectively segmenting the market. Clemons et al (2002) claimed that sellers’ opportunistic pricing behavior allowed them to extract extra
9
profits from uninformed consumers that accounts for at least some of the dispersion in prices (Grover et al, 2006).
Additionally, sellers have tools at their disposition which can lead to increased search costs across all consumers. One such avenue open to sellers is price obfuscation. Obfuscation can be thought of as an action that raises search costs and subsequently can lead to higher average markups as the fraction of firms buying from relatively high-price firms increases (Ellison and Fisher Ellison, 2005). Sellers have been shown to set out to obfuscate price offers by subtle variations in supply terms or by bundling products. However, obfuscation can also take many other forms. One such example is that firms can set out to trick boundedly rational consumers. This could occur if firms can trick consumers into paying more for a product than it is worth or altering their utility functions in a way that raises equilibrium profits (Ellison and Fisher Ellison, 2009). Obfuscation in the presence of heterogeneous consumer search costs will lead to at least a subset of consumers to pay above marginal costs, even in a free market (Haynes and Thomson, 2007).
Whilst sellers have tools at their disposal to obfuscate prices, informational problems, such as information asymmetry can often plague markets. For instance, when objective information is given to consumers who would like make a purchase they behave based on rational economic theory (Grover et al, 2006). Consumers engage in a best-value choice strategy, weighing up the marginal cost and benefits of a product given a certain price level. However, at times information is misrepresented or incomplete and this can misguide consumers and in turn allow sellers to make pricing decisions that increase price dispersion in e-markets. Furthermore, there are times when ‘information equivocality’ can arise when there are conflicting views on product information or seller reliability. High equivocality leads to confusion and lack of understanding (Grover et al, 2006). When making a purchase, particularly online, consumers rely on feedback from others in order to have confidence in their purchasing decision. When there is uncertainty and ambiguity in this information, this increases information equivocality.
Additionally, Shapiro and Varian (1999) argued that information overload can occur when the higher number of sellers increases due to low entry costs for sellers. This results in a scenario where more information exists than is required and this creates a cognitive burden on the buyer (Grover et al, 2006). Faced with information overload, consumer’s confidence in their purchasing decision is not high. In a study of 161 different e-markets over a 4-month
10
observation period between December 2004 and March 2005 Grover et al (2006) find positive support for both the Information Equivocality and Information Overload hypotheses. (ii) Market characteristics
This strand of research attempts to find reasons for price dispersion by focusing on market factors such as the number of competing firms. E-markets have evidently made market entry more feasible for potential entrants and economic theory would suggest that lower barriers to entry increase competition. The number of firms is effectively a measure of competitiveness in the market yet there are differing theoretical views on its impact upon price dispersion. Intuitively, a market which is densely populated with sellers would lead to a more competitive market, thereby characterized by lower prices and less price dispersion. However, as discussed above the higher the number of firms, the higher is price dispersion since it becomes more difficult and expensive for consumers to find all the information. Additionally, Azar (2013) suggests that as the number of firms in the market increases so does the average price, as more firms choose to price at the higher end of the price distribution and target only uninformed consumers, since it is less likely that they will end up charging the lowest price. Yet, empirical studies have often failed to reach consensus on the relationship between market structure and price dispersion. In a study using four million price observations in an online market for consumer products for over 1,000 products on Shopper.com spanning a 7 month observation period, Baye et al. (2004) find systematic differences in price dispersion depending on the number of firms listing prices for a given product. They find that the level of price dispersion is greater when smaller numbers of firms list prices than when large numbers do suggesting a strong negative relationship between the number of firms and price dispersion. Haynes and Thompson (2007) also find an important link between price dispersion and market structure but have opposing findings to Baye et al (2004) as they find that price dispersion increases with the number of sellers. In a study on 399 distinct digital camera markets using price data gathered from NexTag.com, they find that there is a strong positive relationship between the number of sellers and price dispersion. However, not all studies have found a significant link between market structure and price dispersion. Lindsey-Mullikin and Grewal (2006) find only partial support for the hypotheses that price dispersion increases in the number of sellers while Petrescu (2011) finds no empirical evidence whatsoever that price dispersion increases in the number of sellers.
11
In addition, several existing studies (e.g. Haynes and Thompson, 2007) have also shown that price dispersion is dependent on various factors, such as the product’s stage in its life cycle. This is of particular importance in markets where technology is continuously advancing such as in consumers electronic product markets. Haynes and Thompson (2007) for example find that whilst the minimum price is observed to modestly decline over time, price dispersion is observed to increase over the model’s life.
Finally, several researchers (e.g. Pan et at, 2001) have asserted that price dispersion is affected by the average price of a product. For example, when the product price is high, this may induce consumers to undertake longer search and thus could lead to price convergence at a lower level if the number of uninformed consumers are low. However, Petrescu (2011) argues that the opposite is true. For example, a saving of £10 on a good worth on average £30 is more important to a consumer than a saving of £10 on a product worth £300 as consumers view savings in relative, not absolute terms. In this case, the result would be that buyers undertake less search and higher price dispersion occurs. Petrescu (2011) does find positive support for this hypothesis although in similar studies the relationship is less clear cut.
(iii) Seller Heterogeneity
One obvious explanation for price dispersion in homogeneous goods markets is seller heterogeneity. One particular reason is that sellers in e-markets are dispersed geographically leading to potential product differentiation and price dispersion. Intuitively, a wide distribution in transportation costs across sellers will lead to more dispersed prices as prices are a function of transportation costs. Moreover, sellers can differ in their cost structure and this could influence price dispersion. While technology has advanced in such a way that the cost of adjusting prices has become negligible, management costs in deciding what prices to be set can still be substantial and these costs vary largely across both sellers and products (Gupta et al, 2006). Furthermore, there is certainly scope for further cost heterogeneity. There are likely differences in supply prices, as larger retailers are able to negotiate larger wholesale discounts and there are presumably differences in handling costs which reflect low rent/low wage costs locations (Haynes and Thompson, 2008).
Whilst firms evidently vary in their cost structure, it is empirically difficult to measure what impact these differences have on price dispersion as data on firm-level costs are difficult to obtain. Most theoretical models on price dispersion tend to use a simplifying assumption that firms’ costs are homogeneous. However, costs only represent one way in which firms are
12
able to differentiate themselves. As online markets have matured, one new avenue in which firms have been able to differentiate themselves is through online ‘word of mouth’ – customer feedback mechanisms. Whilst costs are unobservable to consumers and competitors, reputation through feedback is visible to both buyers and sellers. In contrast to physical storefronts, transactions in online marketplaces are more uncertain as consumers are not able to physically inspect the product before making a purchase. Moreover, the perceived transaction risk is higher online, since the customers need to trust sellers with their personal and financial information (Petrescu, 2011). As a means of overcoming these challenges online, reputation mechanisms have become commonplace. These mechanisms aim to alleviate problems such as adverse selection and moral hazard by providing a publicly available measure of seller reputation (Jolivet et al, 2014). Whilst feedback mechanisms are beneficial to consumers, they are of great importance to researchers alike who wish to study reputation effects. As Jolivet et al (2014) note, “the empirical importance of these reputation indicators is the focus of burgeoning literature in economics, that has arisen from the expansion of e-commerce and increasing data availability.” (p2)
Whilst reputation plays an important role in overcoming risk, it also represents a sign of seller
quality in markets where the good is homogeneous. In e-markets, consumers purchasing a
product online care not only for the product itself, but also for seller specific aspects such as delivery timeframe, post-sale service and return policy. It is commonly believed that buyers are willing to pay price premiums to sellers with a high reputation; therefore high-reputation sellers should charge relatively high prices. Few dispute the notion that reputation acts as a signal for quality in a market setting with information asymmetry and repeated transactions. “Trusted” sellers with a large history of transactions and high ratings evidently reduce uncertainty for risk-averse consumers. Furthermore, sellers with comparatively high ratings indicate to buyers that they are less likely to renege and sacrifice their reputation.
Varian (1999) initially predicted that e-markets would establish two types of firms: those providing poor service with low prices and those firms offering high prices combined with premium service (Bayliss and Perloff, 2002). Given the theoretical underpinnings outlined above, one would expect to see such predictions supported empirically. However, whilst researchers have reached consensus that reputation does have some impact on price dispersion, the magnitude of this impact is often disputed. In one of the earlier studies on the impact of seller reputation on price dispersion, Pan et al (2002) find that only a small proportion of price dispersion in online markets can be explained by differences in service
13
quality and that significant amounts of price dispersion remain after controlling for the influence of e-tailer services. They identified four service factors in their study, notably: reliability, shopping convenience, product information, and shipping and handling. They compiled a data set containing 6,730 price observations for 581 product items in eight different product categories and the Bizrate.com’s service quality ratings for the 105 e-tailers who sold those items. Price dispersion remained persistent after controlling for e-tailer service quality. Additionally, Bayliss and Perloff (2002) show counter intuitively that e-tailers of digital cameras and scanners charge relatively low prices whilst providing superior services, while some internet retailers charged relatively high prices and provided poor services. Moreover, Ba et al (2012) identify an “adverse” price effect which shows a seller may decrease its price when its recognition level increases.
However, while several studies have reported low adjusted r-squared results, suggesting that seller heterogeneity has little explanatory power for price dispersion in e-markets, there are several contradictory findings. Cohen et al (2007) in a study of 542 homogenous products across 13 different product categories report that a large degree of the variation in prices can be explained by seller heterogeneity. Moreover, Ghose et al (2009) find evidence that differences in the level of reputation between sellers have a significant and positive effect on pricing premiums. What has caused difficulty in establishing the relationship between reputation and price dispersion is that there has not often been one singular measure of a seller’s reputation which incorporates sellers’ quality. However, as online feedback mechanisms have matured, it has become simpler to measure their impact and this remains a research area not yet fully explored.
(iv) Seller Pricing Strategy
The reasons outlined above represent the most prominent drivers of price dispersion cited in the relevant literature. However, several theoretical explanations emerged at the turn of the century (e.g. Chen and Hitt, 2003; Smith), 2001) claiming that firms could use random pricing strategy which could account for at least some of the observed price dispersion. Randomized mixed-pricing strategies were first studied by Varian in his seminal paper ‘A Model of Sales’. He found that stores randomly lowered their prices and thus consumers could not learn from experience which stores had the lowest prices (Gupta et al, 2006). This strategy could theoretically be driven by firms’ understanding of the existence of both informed and uninformed consumers in the market. As firms cannot target both
14
simultaneously, sellers’ optimal strategy is to randomize prices in a dynamic and probabilistic setting (Jost, 2012). In equilibrium firms randomize their prices between a lower bound and an upper bound based on their average costs and the consumers reservation level respectively (Jost, 2012). As a result, extremely high prices are more frequently charged whereas the frequency of intermediate prices diminishes. Thus, price dispersion is an equilibrium outcome.
In empirically testing this hypothesis in e-markets, Baye et al (2002) found that there was evidence of mixed pricing strategies on the online shopbot, Shopper.com. They argued that sellers mix between low and high prices resulting in dispersion across retailers and across time (Pan et al, 2004). However, in a study on the digital camera market Bayliss and Perloff (2002) found that the store ranking in the cross-sectional price dispersion is persistent over time. They found that sellers maintained their price rank or changed by at most one position 57% of the time. Only 4% of sellers were found to have large price adjustments which resulted in a large changes in their price rank. Overall, there was little evidence to support the random pricing hypothesis and e-tailers were not observed to unilaterally lower or raise prices, nor were they observed to engage in severe price competition (Pan et al, 2004). Relatedly, there is also evidence that online prices exhibit temporal price dispersion. For instance, Baye, Morgan and Scholten (2004) using a dataset consisting of 36 popular consumer electronics products sold over a 19-month period find considerable evidence for month-to-month changes in the identity of the low-price firms, but more evidence of persistence in the identity of high-priced firms. Similarly, Iyer and Pazgal (2003) collect bi-weekly price data on music CDs, movie videos and books from five price comparison sites and find empirical results suggesting that there is no firm which consistently offers the lowest price (Baye et al, 2006).
In addition, one well-known retail strategy that results in price dispersion is the tactic of cutting prices on one good, known as a “loss leader”, in order to attract more store traffic and increase profits on other goods. Retailers’ incentive to engage in such pricing varies with their selling capabilities. When product demand increases, sellers with high cross-selling capabilities have higher incentives to set loss-leading prices than retailers with low cross-selling capabilities (Li, 2013). This leads to price differentials between the two different types of retailers which thus contribute to the observed higher price dispersion.
15
Finally, Ellison and Fisher Elisson (2005) pointed out that sellers might try to attract buyers by charging low prices on certain items with a view to make higher mark-ups on other items. Ellison and Ellison (2001) for example find evidence of ‘bait-and-switch’ tactics in the market for computer memory modules. They show that sellers engage in strategies of offering inefficiently low quality products at a very low price to score highly on price comparisons: this enables them to attract a base of loyal customers, whom they then try to convince to pay higher for a better quality product (Fisher and Fisher Ellison, 2005). However, this runs the risk of adverse selection. This strategy may attract a disproportionate number of informed consumers who will ‘take the bait but ignore the hooks for higher price ancillary products’ (Haynes and Thompson, 2007).
2.3 Conclusions on Literature Review
As this literature review has shown, there exists several drivers of price dispersion found in existing literature. Whilst the literature has reached common consensus that the law of one price is not a reality in e-markets, there is no common consensus on the drivers of price dispersion online. Empirical results are often ambiguous and a definite proof as to why price dispersion persists in e-markets is lacking. It is the aim of the remainder of this thesis to analyze the drivers of price dispersion, relating primarily to seller heterogeneity and pricing strategy as drivers of price dispersion in e-markets. In particular, the focus is placed on how seller characteristics impact the price distribution in the Amazon.co.uk marketplace.
16
3. Research Questions and Hypotheses Building
This section aims to develop a set of hypotheses based on the research question, drawing from the relevant literature as well as empirical studies similar in nature. Most studies to date focus primarily on price dispersion in homogeneous products within a product category or across different categories and across sellers. This study differs slightly as the focus is on one singular homogeneous product within one specific marketplace. The purpose of this paper is to further investigate the drivers of price dispersion for a seemingly homogenous product in a structured online marketplace, Amazon.co.uk. The aim is to better understand the price distribution in the market and determine why there exists such a variance in prices for an identical product. Whilst similar studies have focused on seller reputation as a potential explanation for price dispersion, this thesis aims to have a more detailed investigation on seller characteristics and to what extent different seller characteristics can explain price dispersion in the online marketplace. Moreover, this paper aims to establish whether or not the price distribution is stable from day to day or is there evidence of individual seller price variability.
In particular, this thesis aims to answer the following research questions:
- Which seller characteristics have the most explanatory power for price dispersion?
- Does seller reputation matter in the price distribution?
- Are price distributions stable from day to day or is there evidence of random strategy pricing from firms?
In order to answer these research questions, it is necessary to construct a clear set of hypotheses. As was discussed extensively in the literature, reputation can play an important role in differentiating sellers as well as affecting consumers purchasing decisions. To a consumer, a high reputation can be a sign of a high level of seller trustworthiness as well as better fulfillment services. Therefore, buyers may intend to pay price premiums on high-reputation sellers for low risks and better services (Feng et al, 2012). Ba et al (2013) argues that the appeal of a lower price diminishes when there is a risk that seller does not fulfill its obligations. Although the empirical evidence is ambiguous on the relationship between seller reputation and prices, it is generally expected that higher seller ratings would be associated with higher prices. Hypothesis 1 is thus formulated as follows:
17
H1: Sellers with a higher numerical rating will have a higher price relative to the mean.
In addition to seller ratings, consumers may also care about seller experience, which can be captured by the number of seller ratings across its lifetime. Generally, a higher number of ratings signify a higher number of transactions and therefore a more experienced seller. Moreover, Schmalensee and Willig (1989) point out that the risks of early market entrants are lower than those of late entrants and that consumers may be willing to pay more for early entrants that are more trustworthy (Ancarani and Sotgiu, 2005). Risk averse buyers will therefore be inclined to pay a higher price to a seller with higher average ratings and more prior transactions (Ghose et al, 2009). Closely related to H2a is H2b. Sellers which have a higher number of products will appear to be more experienced and more established and buyers may be willing to pay a premium to such sellers. Thus H2a and H2b are formulated as follows.
H2a: Sellers with a higher number of ratings will have a higher price relative to the mean. H2b: Sellers which sell a higher number of products will have a higher price relative to the
mean.
In addition to the rating, experience and size of a seller, there are a number of additional ways in which sellers can be subtly different. For instance, it is not only the size of a seller’s product portfolio which may be important, but the products which make up the seller’s portfolio. Potential consumers may be willing to pay a premium to seller’s who specialize in the product category they wish to make a purchase. It is thus hypothesized that the degree of specialization of a seller’s product portfolio will have a relationship with price. Following DiRusso et al (2011) I hypothesise that sellers which have at least 90% electronic products will offer a high price relative to the mean. A specialized electronics vendor may be perceived as more trustworthy than a generalized retailer for an electronics product, resulting in some consumer willingness to pay more (DiRusso et al, 2011). What is interesting to establish additionally is whether sellers which specialize in cameras or camera related products charge a higher price given their specialization. I therefore wish to test whether sellers which have at least 90% camera products hold a higher price.
H3a: Sellers which specialize in electronic products will have a higher price relative to the mean.
18
H3b: Sellers with specialize in camera products will have a higher price relative to the mean.
Additionally, it is important to note that sellers who are active on the Amazon marketplace can also be active in other marketplaces as well as possessing a physical storefront. Empirical evidence from several years of data indicates that multichannel retailers have higher price levels than strictly online sellers (e.g. Ancarani and Shankar 2004, Pan et al. 2003; Ratchford et al, 2003). There are several theoretical reasons for this to be case. For instance, the single distribution channel for a online-only retailer lowers the inventory costs and this could be potentially leveraged to price differentiate by cost leadership (Bapna et al, 2007). Based on the existing empirical studies and theory, I propose Hypothesis 4a and 4b:
H4a: Sellers with a separate sales website will have a higher price relative to the mean. H4b: Sellers who operate a bricks-and-mortar storefront will have a higher price relative to
the mean.
Finally, I wish to establish whether the contact details a seller posts are in any way related to a seller’s posted price. On Amazon.co.uk consumers are able to contact sellers through an online question form provided by Amazon. Although not obligatory, certain sellers also provide a contact telephone number in addition to email contact. Consumers might see such sellers as more trustworthy and therefore may be willing to pay a higher price. H5 is thus written as follows:
H5: Sellers which provide a contact telephone number with have a higher price than the mean.
Table 1 - Summary of Hypotheses
Dependent Variable Abbreviation
Log of Total Price LnPrice
Independent Variables Abbreviation Predicted sign
Rating Rep + No. of Ratings Camera Specialist Electronics Specialist Storefront Website Contact Exp Cam Elec Store Web Cont + + + + + +
19
4. Data
4.1 Data Collection
The dataset is taken from Amazon.co.uk, which has operated as a subsidiary of Amazon.com since 1998. A study of price dispersion on Amazon.co.uk represents a suitable marketplace for a study of this kind and has several advantages over price comparison websites which is the focus of much of the extant literature. Firstly, Amazon is a marketplace in itself with the site brokering the transaction without the requirement to visit the seller’s website to complete the purchase. Secondly, all product information is available on a central landing page; hence there is no heterogeneity in product information across sellers and less tendency for information asymmetry. Thirdly, buyers who visit Amazon.co.uk have almost zero search costs since price listings of each seller are listed automatically from lowest to highest. These factors combined help to mediate the effects of both search costs and information asymmetry which can contribute towards price dispersion.
Similar to previous literature on price dispersion I use publicly available differences in posted prices of a homogenous product. Importantly, prices were only collected for products identified as “new” which mitigates some of the variation caused by product heterogeneity. Because price dispersion, by definition, relates to prices charged by different vendors for the same product, a single product was used as the unit of analysis. The product chosen was a digital camera, namely the Canon EOS 1100D Digital SLR Camera. This is in keeping with previous studies on price dispersion in homogeneous goods markets (e.g. Baye et al, 2004; Haynes and Thomson, 2008). The digital camera is advantageous for pricing research in that it is typically purchased singly, thus avoiding the issue of bulk discounts that apply to products such as books or CDs (Haynes and Thompson, 2012). In addition, Canon provide a 1 year warranty period to UK customers and therefore any manufacturing flaw which occurs post sale is covered by Canon.
A typical page on Amazon’s secondary market contains a list of sellers of both ‘new’ and ‘used’ products, where prices are automatically sequentially listed from lowest to highest (including shipping charge). In addition, each seller’s percent positive feedback from the past twelve months and number of overall ratings are all directly observable and are listed next to the seller’s name. Each seller’s percentage positivity rating is determined by a star rating
20
system out of five. Ratings equal to 2 or less are graded as negative whereas ratings of 4 of more are denoted as positive. Ratings of 3 are denoted as neutral and the combined score is averaged for an overall positivity feedback rating displayed next to the seller. These ratings are directly determined by buyers who have purchased directly from the sellers. Only those who have made a purchase from the seller are permitted to leave feedback and comments, and doing so is not compulsory. In addition, through further search, consumers are able to obtain a full history of buyer feedback, including lifetime ratings and scores from the last 30 days of feedback.1
Data collection began on the 19th May, 2014 until the 24th June 2014, constituting a 38 day observation period. Data on sellers’ posted prices, seller rating and the number of ratings were manually gathered daily with the help of screenshots. The study was limited in scope due to the arduous task of downloading the screenshots and subsequently creating the data set. The data set thus consists of a set of daily price listings, with a total of 1432 observations. However, not all sellers who featured on the first day of observation featured throughout as there was significant entry and exit in the market.
Identifying other distinguishing factors that might contribute to sellers’ pricing dispersion is important and thus further variables must be incorporated into the analysis. In addition to variables outlined above, further binary variables were included. For instance, I verified whether in addition to existence on the Amazon marketplace, sellers also operated an external sales website and/or physical storefront. In addition, information was gathered on whether or not each seller listed a telephone number in addition to email. Furthermore, I also documented for each seller whether or not they were an electronics and/or camera specialist. Lastly, data was gathered on the number of items sold by each seller, which alongside the number of seller ratings is a good indicator of seller experience and size. Intuitively size should matter because the consumer may take market performance as reflecting part of the seller’s quality.
1
For the basis of the study, only data on the past 365 days of feedback was recorded. The reasoning for this is provided in Chapter 6.
21
4.2 Descriptive Statistics
The means and standard deviations of prices and continuous independent variables are displayed in Table 1. It is immediately observable that data on prices shows considerable variance. The mean price across the period was 72% higher than the minimum recorded price. The coefficient of variation - the ratio of standard deviation to the mean price of the product - is 22%.
It is evident from Table 1 that the magnitudes of the standard deviations of the prices are related to the magnitudes of the means of the prices. This suggests that the logarithmic transformation is a suitable choice to normalize the price distribution. Figure 3 displays a histogram of the log of prices throughout the observation period. Additionally, sellers differ greatly in their experience, as measured by the number of ratings in their lifetime. Some sellers are observed to have a handful of ratings whereas others have thousands suggesting a mix of established and non-established sellers in the dataset. Therefore, a logarithmic approach is once more suitable.
Whilst the observed number of ratings are widely dispersed, seller ratings are less so. Analysis of the reputation data shows that firms generally have very high reputation scores. As can be observed in Figure 2, the number of ratings is densely populated at the top end with the mean of positive ratings at 94% suggesting that generally speaking sellers are highly rated on the marketplace. This would suggest that potential buyers may differentiate between small percentages and the difference between a rating of 97 and 95 for instance matters. Like with other continuous variables in the dataset, I take the logged transformation of percentage rating, rendering a linear additive equation.
From the initial sample of 1432 price quotes, 119 observations were dropped from the sample. All sellers which did not have any ratings were removed from the final sample. These were three sellers who were likely individuals who entered the market in order to sell one singular product. Importantly, Amazon were dropped from the analysis as consumers are not able to rate the company. As reputation and the number of ratings were key variables of analysis, this was systemically necessary. In addition, there was one recorded firm which was recorded to have 50% positivity rating based on two reviews, and such an observation has the potential to bias results. Finally, on several days two sellers were observed to have two
22
separate posted prices in the distribution. One posting was systematically removed upon inspection of the real offered price on the individual seller’s page.2
Table 2 - Summary Statistics Continuous Variables Mean Standard Deviation Minimum Maximum Price 389.43 84.31 226.99 908.73 Shipping 3.92 13.37 0 101.42 Total Price 393.49 83.96 232.09 913.83 Rating 94.16 4.69 77 100 No. of Ratings 9816.31 16,262 5 59819 Binary Variables (Proportion) Electronics Camera Website Storefront Contact 0.6 0.14 0.59 0.12 0.51 Total Observations 1313 2
For instance, Big Red Toolbox was observed to post a price of £384.62 on the 19th May as well as a different posted price of £389.24 in the same distribution. Upon verification on the seller’s market page, the price of £384.62 was posted and thus this price was included in the dataset and the price of £384.62 was dropped.
23
5. Results
5.1: Price Distribution and Dispersion
Having now described the dataset and with hypotheses now defined, it is the purpose of this chapter to analyze results. Firstly, I provide a brief overview of the evolution of prices over the observation period. Figure 2 provides a graph of the evolution of the minimum, average and highest price as well as the standard deviation from day to day. What is immediately evident is that there is clearly a wide distribution of posted prices in the market. Whilst the lowest offered price changes modestly if at all from day-to-day, the highest price appears to fluctuate wildly and extremely, which is somewhat perplexing. Variation in the identity of the firm offering the lowest price is less surprising given that firms which offer the lowest price will likely have the highest sales. The magnitude of the fluctuations in highest price is more surprising and it appears implausible to attribute such prices to factors such as superior service for which consumers are willing to pay such a high premium. Given that such high prices persist and that posting prices on Amazon.com comes at a cost, it is suitable to conjecture that such price offerings are serious. Firms listing artificially high prices are unlikely to generate sales from the site to justify the associated fees of listing. A discussion based on these findings will be devoted Section 6.
24
5.2 Price Stability
It is important now to check whether sellers’ prices remain stable from day-to-day or whether or not there is evidence of frequent price changes. In physical markets there are menu costs to take into account of when sellers contemplate making price changes and it often takes time to react to competitors’ pricing decisions. However, in online markets such changes can be enacted almost instantaneously and it is very easy to react quickly to price changes. Intuitively, one would thus expect to see more frequent price changes online and more variation in the identity of both the lowest and highest priced seller. I thus observed whether there is individual price variability arising from the fact that a seller might increase or decrease its price over time.
Of the 51 sellers in the sample, 19 of the sellers are observed to make at least one price modification during the observed period. This represents 39% of the sellers in the sample. Additionally, across the observation period, there were six different sellers who posted the lowest price on a given day. Although the lowest price appears to change only marginally, this suggests that firms compete regularly to become the lowest priced seller. At the other end of the distribution, a total of four different sellers offered the highest price on any given day. This suggests that there is no single firm who will consistently offer the lowest or highest price. Perhaps surprising is the frequency of price changes. Table 4 tabulates all sellers who made price changes during the observation period, documenting the frequency of price changes, magnitude and the number of times the seller was the lowest or highest priced-seller. As can be seen, for a number of firms the magnitude of price changes for certain sellers were very high indeed. However, individual price changes cannot be the primary source of price dispersion as for the remaining firms in the sample price changes are symmetric.
5.3 Preliminary Results
Preliminary analysis of the data provided some interesting insights into the relationship between reputation and price. Looking at the relationship between the log of net price and log rating it appears somewhat counter intuitively that the two are negatively correlated. Looking however at the relationship between the number of ratings the log of total price there appears
25
to be little relationship between the two. The number of ratings was expected to be an indication of how well-established a seller was in the market. However, further extensive analysis is required before any firm conclusions can be derived from results.
Figure 2 - Relationship between Experience and Net Price Figure 3- Relationship between -Reputation and Net Price
5.4 Regression Analysis
It is the purpose of this section to more formally test hypotheses set out in Chapter 3. In order to do this, more formal econometric modeling is required. Having now established that there is significant variance in prices charged by sellers, I now wish to understand what types of sellers charge low or high prices. I wish to test the link between seller price levels and a specific set of measurable characteristics. The empirical model outlined hereafter seeks to determine the relationship between a seller’s posted net price and a series of variables which represent each seller’s characteristics. I use the net price (including shipping cost) as the dependent variable of analysis as prices (inclusive of shipping) are chronological and customers are likely to consider shipping cost in their purchasing decision as part of the final total. As stated previously, I take the natural log of both the continuous dependent and independent variables in order to normalize the distribution and account for both large variation and skewness. This renders an equation which is log-linear.
Moreover, given that a seller’s rating and number of ratings differ little in the short run and that prices generally remain stable for a large number of sellers, I take the average value of total price, seller rating and number of ratings for each seller. In equation (1), this is accounted for with a bar sign above the variable. In addition, the number of products sold by
26
a seller was dropped from the estimation as this variable was found to be highly correlated with the number of ratings. Moreover, using the Breusch-Pagan test for heteroskedacity, I found that heteroskedacity was present in the estimation. Thus, in order to overcome this I estimate the model using OLS with hetresokedacity-robust standard errors. With the model now defined, it is now possible to explore whether the detected variance in prices of a specific digital camera were related to the variables in Equation (1).
(1)
The results of the regression are outlined in Table 3. In addition, a summary of the results of the price level regression with respect to the predictions of hypotheses outlined can be found in Table 4. Turning attention to the key variable of interest, reputation, I find counter intuitively that a seller’s average price and average rating and negatively correlated. What’s more, this variable is significant (p value of 0.009) at the 1% confidence interval. Thus, H1 is not supported.
Table 3 - Regression results
Variable Coefficient SE t Pr > |t|
Log Rating -1.6722*** 0.06925 -2.70 0.009
Log no. of Ratings 0.0091 0.01267 0.72 0.476
Electricity Specialist 0.7226 0.06679 1.08 0.285 Camera Specialist 0.1704 0.09676 1.11 0.273 Website 0.0076 0.06900 0.11 0.913 Storefront -0.0610 0.09891 -0.62 0.540 Contact -0.0798 0.05906 -1.23 0.226 N=51 R-squared: 0.26
Note: significance levels are defined as *=10%, **=5% and ***=1%
Looking at the results against H2a, I find that there is a slightly positive relationship between a seller’s number of ratings and net price, however the variable is not significant. This result is somewhat surprising given that there was a large variation in the number of ratings across sellers. This could indicate a mix of established and un-established sellers who operate
27
different pricing strategies, however results here suggest that is not the case or that the number of seller ratings is not an accurate measure of a seller’s establishment in the marketplace. For instance, a seller with few ratings on Amazon might only be new on the Amazon marketplace but will already be well established and known to consumers on other marketplaces. This is not picked up by the number of ratings as an approximation for seller experience.
Looking at the results of the binary variables, none of the factors were found to be significant, indicating that there are likely other more important factors not included here in predicting prices online. However, given the locality of the study a more in depth analysis would be required before the significance of these binary variables can be ruled out. Indeed, only 51 sellers were present in the marketplace and the proportion of specialized electronic sellers and sellers who operated a physical storefront was small. Overall, an R-squared of 26% shows that the model used accounts for some of the variation in prices across sellers yet there are other factors at play which likely play an important role in price dispersion. It is the purpose of the following chapter to discuss these findings in more detail.
Table 4 - Variable Outcome Summary
Dependent Variable Abbreviation
Log of Total Price LnPrice
Independent Variables Abbreviation Predicted sign Actual sign Significant? Supported?
Rating No. of Ratings Camera Specialist Electronics Specalist Storefront Website Contact LnRep LnExp Cam Elec Store Web Cont + + + + + + + - + + + - + - Yes No No No No No No No Yes Yes Yes No Yes No
28
6. Discussion
The aim of this section is to align the findings with existing theory. Results outlined in the previous section reveal some interesting peculiarities in the online market. As expected and in line with similar previous empirical studies, price dispersion is both significant and persistent. This result is entirely expected but what is unexpected and peculiar are the extremely high posted prices observed. The question remains whether there is any semblance of order to this apparent chaotic pricing? Attention will turn firstly to the strategy of high-priced sellers and to what extent we can find answers in extant literature before discussion of results found in Chapter 5.4.
6.1 Multiproduct Competition and Price ‘Signaling’
In Varian’s seminal ‘A model of Sales’, he proposes a mixed pricing strategy where firms charge high prices sometimes and low prices in others in order to extract profits from both uninformed and informed consumers. A common result of his model is that there are no firms with static prices and an important assumption is that firms are profit maximizing. In a marketplace where firms are sequentially ranked from lowest to highest price and there is little scope for price obfuscation or information asymmetry, one can conjecture that transactions will be concentrated around the lowest price. Certainly analysis of transaction prices supports this stance (see e.g. Ghose & Yao, 2010). The distribution of prices in a selected digital market shown here shows that certain firms were observed to post prices well above both the lowest and mean prices. A firm is only in a position to set a price if it is viable and a profit maximizing decision to do so and setting artificially high prices with zero probability of making a sale certainly does not appear to represent profit maximizing behavior.
It is important to emphasize however that firms compete not just in one singular product market, but across several other markets simultaneously. Therefore, there is both multiproduct competition and multi-market contact. As Chellappa et al (p88, 2011) argue ‘most firms sell a variety of products, and do not price them with a myopic view of achieving a one-time sell, or necessarily react to their competitor’s price in Bertrand-competition fashion.’ Sellers make a conscious decision in their pricing strategies is likely to impact the distribution of prices in the market over and above other relevant factors.
29
In addition, it is plausible to believe that sellers on the Amazon marketplace try to induce sales through a particular low priced product in a bid to increase sales in high margin products on Amazon, through other channels and beyond. Such practices are relatively common across other industries: for instance, restaurants can lure consumers with low food prices but high prices for drinks. Price dispersion is in the end a firm decision making. Whilst this explanation may hold for firms who offer marginally higher prices than competitors, it is unlikely to hold for firms which set ‘extreme’ high prices.
Alternatively, one potential explanation for this ‘extreme’ price setting could be related to price signaling. A seminal paper by Akerlof (1970), ‘The Market for “Lemons”’ first theoretically argued that posted prices were a signal to consumers of high quality products and perceived “lemons”, the low quality products. In markets, price ‘signaling’ thus helps mitigate the adverse selection problem. Certainly when a firm offers higher quality service at higher cost, feedback mechanisms are vital to in keeping sellers from being wiped out by price competition. Yet, in online markets feedback mechanisms are in place in order to avoid the requirement for such price setting behavior. However, although a degree of information asymmetry is mitigated through feedback mechanisms, sellers remain better informed on both product and service quality. In a structured marketplace such as Amazon in a market where products are chosen as “new”, high prices may be due to price signaling through extremes. There still remains a degree of uncertainty and thus sellers might attempt to signal quality through the choice of price (Adrianna & Deidda, 2011). It is however beyond the scope of this paper to investigate the true rationale behind such ‘extreme’ price setting behavior in depth.
6.2 Regression Analysis Results
A key finding in the previous section was that a negative linear relationship was observed between seller rating and net price which appears counterintuitive at a glance. However, it is important to bear in mind that the numerical rating of a seller is a reflection of overall satisfaction with a seller. It is possible that seller ratings reflect not only a seller’s post sale service e.g. delivery and packaging, but satisfaction with the seller’s price. Lower prices therefore lead to lower ratings and this helps lower priced sellers achieve higher ratings. In addition, it was expected that the number of ratings would have a positive correlation with price. However, a marginally positive yet insignificant relationship between the number of
30
ratings and total price was found. The number of ratings appears a good indicator of seller size as it documents the number of transactions in a seller’s history. It was used here as a factor of analysis as it is a measure which is both publicly available and fully transparent. However, it appears at least in this study that such a measurement is ineffective.
Finally, results found that amongst all binary variables used in the analysis, none were found to be statistically significant. This could reflect that there are other more important seller indicators which explain the large variance in prices. However, the sample size was small and therefore it is difficult to ascertain whether such factors do in fact play a more important role. This study provides only a snapshot of price dispersion and therefore it is difficult to infer significant conclusions.
31
7. Limitations and Directions for Future Research
The preceding chapter has attempted to align findings found in this study with existing literature. However, it is important to take into account several limitations of this study and take caution in inferring conclusions from the results. It is the purpose of this chapter to highlight drawbacks of the study as well as look to directions for future research in order to overcome such limitations.
Firstly, the results of this study are applicable to one particular product market on one specific marketplace. It is not possible to generalize these results across other product categories, nor even the digital camera market itself. It would be very useful to extend this analysis across other digital camera markets and where applicable other homogeneous goods markets. This would establish whether factor variables contribute to price dispersion on a wider scale even though they appear insignificant locally. Secondly, the observation period is short and therefore I am unable to assess how the drivers develop over time. It would be interesting for future research, particularly on reputation, to determine whether the relationship between reputation and price is stable over a long period of time.
Thirdly, there are limitations with respect to the data. Whilst the explanatory variables assessed highlight some of the ways in which sellers are able to differentiate themselves, they do not provide a full picture. One notable absence in the dataset is data on seller costs. Costs are likely to vary across sellers and by not accounting for seller costs; there is a possibility for bias in the results. Unfortunately, such data is rarely available and often theoretical models of price dispersion must use simplifying assumptions of homogeneous costs across sellers, which evidently do not hold in real markets. Such data would prove invaluable in a more in depth study of price dispersion.
Fourthly, this study has shown that many firms on Amazon.co.uk also operate an external sales website. Whilst it was proposed that operating a sales website reflected an indicator of a seller’s size, it was not verified whether a seller also offered an identical product and their posted price. As discussed in the preceding chapter, there is a degree of multi contact and multi-product competition in online marketplaces. This represents an avenue not yet fully investigated in price dispersion literature.
