Annotated Bibliography

ShippingFeesorShippingFree-ATaleofTwoPrice.pdf

Home >Information Systems homework help >Annotated Bibliography

Shipping Fees or Shipping Free? A Tale of Two Price Partitioning Strategies in Online Retailing

Mehmet Gümüş, Shanling Li Desautels Faculty of Management, McGill University, Montreal, Quebec H3A 1G5, Canada,

Wonseok Oh College of Business, Korea Advanced Institute of Science and Technology, 85 Hoegiro Dongdaemoon-Gu, Seoul, 130-722, Korea,

[email protected]

Saibal Ray Desautels Faculty of Management, McGill University, Montreal Quebec H3A 1G5, Canada, [email protected]

I n this article, we study the price partitioning decisions of online retailers regarding shipping and handling (S&H) fees. Specifically, we analyze two partitioning formats used by retailers in this context. In the first scenario, retailers present

customers with a price that is partitioned into a product price and a separate S&H surcharge (the PS strategy); in the second, customers are offered free shipping through a non-partitioned format where the product price already includes the shipping cost (the ZS strategy). We first develop a stylized game-theoretic model that captures the competitive dynamics between (and within) these two formats. Analysis of the model provides insights into how both firm and product level characteristics drive a retailer’s strategic choice regarding which partitioning format to adopt and, hence, determines the equilibrium mar- ket structure in terms of proportion of ZS and PS retailers. Subsequently, we conduct empirical analyses, based on product and S&H prices data for two different product categories (digital cameras and printers) collected from online retailers, to validate all the results of our theoretical model. We establish that PS retailers charge lower product prices than ZS ones, but the total price (product + S&H) charged is higher for the first group. The S&H charge for PS retailers can be significant—it is, on average, 5.4% (printers) and 3.0% (digital cameras) for our two product categories. Furthermore, retailers which are popular and/or face risky cost environment are more likely to opt for the ZS strategy, while retailers whose portfolio mostly includes large or heavy products with high cost (S&H)-to-price ratios usually choose the PS strategy. Lastly, our empirical study also illustrates that the price adjustment behavior of retailers is affected by their shipping-fee policies—for example, ZS retailers change their product prices almost 1.5 times more frequently than PS ones.

Key words: retail; E-commerce pricing strategy; price partitioning; free shipping; shipping and handling costs History: Received: May 2009; Accepted: December 2010, after 2 revisions.

1. Motivation

Nowadays, electronic transactions constitute a sub- stantial portion of retail activities. Online retail sales in the United States reached about $35 billion during the second quarter of 2008. This is an increase of almost 9.5% compared to the second quarter of 2007; during the same period, total US retail sales only increased 2.5% (US Census Bureau News 2008). One of the salient characteristics that differentiates online retailers from their offline counterparts (for physical products) is that, with the former, products are deliv- ered to customers by the retailer, while, in the latter, customers usually visit the retailer to buy the product. This means that the shipping and handling (S&H) cost1

is an important consideration for online retailers. This additional responsibility compels online retailers to tackle a new set of decisions. They need to choose

whether to partition the total price charged to custom- ers into a product price and a S&H surcharge (the PS strategy) or to offer them free shipping by charging only the product price and include delivery-related costs—either partly or wholly—in that price (the non- partitioned ZS strategy).2 Obviously, retailers need to decide how much to charge for the product and, if applicable, for S&H. The primary goal of this paper is to study how the firm and the product characteristics drive the strategic price partitioning decisions of online retailers. We also investigate how the price adjustment behaviors of such retailers differ based on their strategic choices. In order to address the above issues, it is important

to understand that E-commerce related S&H costs can indeed be substantial. Such expenses can account for more than 30% of the total cost in a number of sectors such as groceries and toys (Barsh et al. 2000). While it

758

is true that efficient order fulfillment through better supply chain network design and inventory manage- ment has reduced S&H costs for online retailers (refer to Johnson and Whang 2002, Netessine and Rudi 2006, and references therein), such costs can still be quite high. In fact, Amazon.com’s delivery-related operations incurred losses of $630M in 2008 and $849M in 2009 because of the costs associated with free shipping offers (Amazon.com Annual Report 2009, page 27).3 In terms of absolute value, per order costs for S&H are about $15 for prescription drugs and $26 for toys (Lewis et al. 2006). There are practi- tioner experts from the online retail industry who suggest that the lost income due to not charging for S&H costs might be as large as 35%–40% of the oper- ating contribution per order for some retailers (Bolot- sky 2008, Brown 2008, and references therein). So, in order to recoup such costs, retailers might need to add a surcharge for S&H on top of the product price. However, any S&H surcharge could have a signifi-

cant negative effect on the purchase decisions of con- sumers. Studies have repeatedly shown S&H fees to be one of the consumers’ main complaints about online retailing. In annual surveys by the e-tailing group, consumers consistently respond that the number one reason for them not buying more prod- ucts online is the high cost of S&H, and this trend is increasing over the years (Freedman 2008). In a com- score survey from 2008, 72% of those surveyed responded that if an e-commerce site starts charging S&H fees, they would use another site that offered free shipping (Kawamoto 2008). Similar sentiments have been expressed in other surveys by Paypal and Forrester.4 There is a segment of customers who view S&H surcharges as sneaky and unfair strategy used by a retailer to make additional profits. This group is referred to as shipping-charge skeptics (Schindler et al. 2005). Consequently, these consum- ers are very reticent when it comes to paying any such charges. Indeed, research suggests that online consumers are more sensitive to S&H surcharges than other additional charges such as taxes (Smith and Brynjolfsson 2001). The choice between PS and ZS strategies is especially relevant in this context. The literature has established that an end consumer’s mental cost of moving from zero to a positive price for any service/product is more than an “equal” change in the positive price range. For example, a decrease in price from a non-zero value to zero, say from $1 to 0, increases demand more than the same decrease in the positive price range, say from $2 to $1. This is termed the zero price effect (Shampanier et al. 2007).5 We would also like to refer the readers to Anderson (2009) for real-life examples which show that free offers generate much more demand than even a very low price.

When deciding on their optimal pricing strategies, retailers, then, must carefully trade off the negative impact of delivery-related expenses on their total costs against the positive effect of not charging for S&H on demand (and hence, revenue). This results in online retailers resorting to the two distinct pricing formats explained above—the “partitioned pricing” format or the PS strategy and the “non-partitioned pricing” format or the ZS strategy.

1.1. Research Questions and Main Results Motivated by the issues discussed above, this study addresses the following research questions:

1. How do online retailers decide whether to add an extra surcharge for S&H services or offer free shipping (i.e., adopt PS or ZS strategy, respectively) in a competitive environment? What equilibrium market structure, in terms of proportion of PS and ZS retailers, will such strategic interactions result in for a particular setting?

2. How do the particular characteristics of a re- tailer, or the characteristics of the products being sold by the retailer, affect the retailer’s strategic decision on price partitioning?

3. How do the average retail price and the price adjustment behavior (over time) for PS firms compare to those of ZS firms?

In spite of the significant importance of S&H in online retailing, the extant literature is silent about the above issues. The primary motivation of this research is to address this gap. We use theoretical and empirical approaches to answer the research questions. We first construct a demand model for an oligopoly frame- work that incorporates how consumer purchases are affected by partitioning decisions as well as actual prices. The resulting profit functions for ZS and PS retailers capture the basic trade-off that retailers need to consider when deciding on optimal price-partition- ing strategies. Subsequently, we utilize a two-stage game-theoretic approach, where retailers first decide whether or not to offer free shipping, and then decide what prices to charge. This helps us to determine the competitive equilibrium partitioning strategy that will be adopted by each retailer and, hence, the equilib- rium proportion of ZS and PS retailers in the online market. We then theoretically analyze how both product

and retailer characteristics shape the optimal parti- tioning strategy. We first assume all retailers to be symmetric to delineate the role of product characteris- tics and then study a duopoly setting with asymmet- ric retailers to examine the role of the firms’ characteristics. We show that retailers adopting the ZS strategy charge more in terms of unit price, but

less in total, than those adopting the PS strategy. Moreover, as the number of shipping-charge skeptics in the market decreases, the product and/or its S&H gets more costly, or when the future cost environment is expected to be stable, it is advantageous for most retailers in the market to implement the PS strategy. On the other hand, as an online retailer becomes more popular or if it is mostly selling relatively light/small products, adopting the ZS strategy is its optimal choice. Finally, we conduct an empirical study to validate

our theoretical findings. The data for the empirical analysis consists of actual product and S&H prices collected from a large number of online retailers for two product categories—digital cameras and printers. The results suggest that a unit price (i.e., price without S&H fees) charged by a ZS retailer is indeed higher than a unit price charged by a PS retailer. The reverse is true when total prices (i.e., prices with S&H fees) are considered. Furthermore, our results indicate that the volume of a product as well as the total cost of a product is negatively associated with the proportion of ZS retailers in the market. On the other hand, we note that the variance in cost at the retailer-product level and the extent of a retailer’s popularity both have a significant positive association with a retailer’s propensity to offer free shipping. Our empirical study also demonstrates that the ZS retailers change their prices more frequently than PS ones, albeit the magni- tude of the changes is not significantly different between the two groups of retailers. In summary, our analysis sheds light on the status of different S&H fee-related partitioning strategies adopted by online retailers as well as the underlying reasons behind the differences in adoption level of such strategies.

2. Review of Related Literature

Although the specific research questions of this study have not been addressed before, there are certain streams of existing literature that are relevant to our research. The first one deals with the effects of parti- tioning strategies on purchasing decisions of individual customers. To a rational customer, partitioning should not matter, because their purchase decisions ought to be based on total price, irrespective of how it is apportioned. However, behaviorial marketing liter- ature shows that customers do not weigh product price and surcharges equally. Several studies (e.g., Cheema 2008, Morwitz et al. 1998, and references therein) suggest that with partitioned pricing custom- ers tend to more heavily anchor on the product price and overlook the small surcharge. Consequently, cus- tomers are likely to underestimate the total price, resulting in a relatively more positive effect on demand. However, there are also studies that contrast

this viewpoint (e.g., Schindler et al. 2005; Thaler 1985). They argue that partitioning sometimes induces customers to pay attention, in addition to the product price, to the surcharged secondary attributes which they would have otherwise ignored. Conse- quently, partitioned pricing can result in reduced demand compared to a non-partitioned format. We use the insights from this literature stream in develop- ing our demand framework. However, while prior research focuses on the effects of partitioning on indi- vidual customer behavior, we investigate the equilib- rium partitioning strategy adopted at the firm level in a competitive market setting. The second stream of related literature consists of

studies dealing with supply chain issues in the E-commerce world. Its focus has been on using the differing level of price and time sensitivities of online customers to optimally differentiate the available price and delivery choice options (e.g., Zhao et al. 2008), or using customized delivery strategies for online customers to improve inventory management (e.g., Cattani and Souza 2002), or better order fulfill- ment through effective inventory management and network design (e.g., Johnson and Meller 2002, Netes- sine and Rudi 2006, Pyke et al. 2001). Johnson and Whang (2002) provide a detailed review of this stream. While S&H fees are discussed in some of the above studies, none of them deals with the issue of price-partitioning. The study from this stream most relevant to us is Leng and Parlar (2005), which ana- lyzes the issue of free shipping—specifically, what should be the optimal threshold for (total) purchase order value above which free shipping should be offered by an online retailer. Subsequently, Becerril- Arreola et al. (2009) extends Leng and Parlar (2005) to also include retail prices and inventories as decision variables. Clearly, our objective, to determine the equilibrium proportion of free-shipping (ZS) and non-free-shipping (PS) retailers in an oligopolistic framework, is quite distinct. Moreover, most of the above studies are analytical ones and do not empiri- cally validate their results (only Becerril-Arreola et al. 2009 validate some of their results using data from Lewis et al. 2006). There are also a few empirical studies concerning

S&H charges that are related to our research. Lewis et al. (2006) use data from an online retailer specializ- ing in grocery and drugstore items to show that S&H charges, and especially free shipping offers, signifi- cantly affect frequency and sizes of customer orders. On the other hand, Lewis (2004) compares the effects of loyalty programs with other marketing instru- ments, including S&H fees, on customer retention. This study suggests that loyalty programs might be more effective than other marketing schemes. Clay et al. (2002) studied pricing strategies of online book

retailers and showed that some retailers offer low product prices with high S&H charges while others offer high prices and low shipping fees. Note that the above studies focus on retail level strategy (and how it impacts customer behavior) and not on the equilib- rium market structure, as we do in this study. More- over, in contrast to us, they do not use analytical modeling to develop insights which are then vali- dated by an empirical study. Lastly, there is a mature information system literature

stream pertaining to the temporal pricing behavior of online retailers. The existing literature along this line has empirically examined a variety of issues, includ- ing the degree of price dispersion between online and offline retailers (Bakos et al. 2005) as well as among online retailers (Clemons et al. 2002) and the fre- quency and magnitude of price adjustments (Oh and Lucas 2006). Despite the divergence in terms of their scope and context, the consensus among these studies is that electronic markets allow sellers to promptly adjust their pricing tactics according to fluctuating market conditions. Although the present study also explores pricing behavior of online retailers, our focus is quite unique. We study how such behavior of online retailers is driven by their price-partitioning policies, based on the differences in price change pat- terns between PS and ZS retailers. The rest of the article is organized as follows. In sec-

tion 3, we develop our game-theoretic model frame- work and present both product- and retailer-based analyses. In section 4, we empirically validate our the- oretical results and also discuss the temporal pricing behavior of ZS and PS retailers. Section 5 focuses on the managerial implications of our results. In section 6, we present our concluding remarks.

3. Theoretical Model Formulation and Analysis

In this section we develop an analytical model to understand how online retailers decide whether to offer free shipping or not (i.e., opt for the ZS or the PS strategy) in a competitive environment and what factors drive their choices. The basic model framework involves N( � 2)

retailers selling a particular product online and engaging in a (Nash) price competition among themselves for the potential consumer base. In our model, each retailer makes two decisions. In the first stage, each of them simultaneously decides on whether it should adopt the PS strategy or the ZS strategy. Subsequently, in the second stage, each retailer, given its partitioning decision, simulta- neously determines the exact values of the product price and S&H fee (if applicable) that consumers

need to pay. We assume that ZS retailers offer free shipping regardless of the end consumer’s purchase value. In reality, such offers sometimes require that the value exceed a threshold (e.g., for Amazon.com it is $25). Our framework is an approximation of reality—the PS strategy is equivalent to the thresh- old value being very high, and the ZS strategy is equivalent to the threshold being zero. Moreover, for analytical purposes, we assume that the model horizon is relatively short, so that any transient effects (e.g., customer loyalty dynamics, customer learning) do not significantly impact the retail deci- sions of our interest. Before analyzing the above game, we need to

develop the demand and profit functions for the retailers. Let Nzs and Nps be the number of retailers that adopt ZS and PS partitioning strategies, respec- tively, in the first stage ðNzs þ Nps ¼ NÞ. We denote this decision for retailer i by fi 2 fZS; PSg. Moreover, suppose pi and si are the product and S&H prices charged by retailer i in the second stage. If retailer i opts for ZS, that is, fi ¼ ZS, then it does not charge anything for delivery, that is, si ¼ 0. On the other hand, if retailer i opts for PS, that is, fi ¼ PS, then we assume that si ¼ s�pi, where s (s < 1) is the (exoge- nous) percentage of the unit product price levied as S&H surcharge. Demand and Profit Functions: Let 0 \ ci \ 1 be

the relative popularity of retailer i. A higher value of ci compared to cj signifies that, ceteris paribus, more customers will prefer retailer i over retailer j. One of the primary reasons behind this asymmetry in attractiveness is brand equity.6 For example, Ama- zon.com has a higher brand equity, and hence a higher ci, compared to most other online retailers. The retailers procure the product by paying a price cpðciÞ per unit, and incur a cost csðciÞ per unit for S&H (total cost cðciÞ ¼ cpðciÞ þ csðciÞ).7 We assume that a retailer with higher popularity pays less for buying the product and/or as S&H cost (due to economies of scale and/or scope). Specifically, sup- pose cðciÞ decreases in ci. In order to capture the effects of both partitioning

and pricing decisions in the demand model, we use results from the behaviorial marketing literature. As far as the effect of partitioning decision is concerned, Schindler et al. (2005) and Lewis et al. (2006) have established that that there is a group of consumers who are very sensitive to any S&H charges (shipping- charge skeptics), while others are willing to pay a “reasonable” amount for delivery services (non-ship- ping-charge skeptics).8 Moreover, they suggest that the size of this skeptic group depends on the physical characteristics of the product. Specifically, ceteris pari- bus, customers would be more willing to pay S&H charges for larger and/or heavier products, implying

that for such products the skeptic group would be relatively small. Suppose w represents the physical characteristics of a product where higher w implies heavier and/or larger products. We assume that online customers consist of two groups: 0 < a(w) < 1 proportion of them are shipping-charge skeptics and the remaining (i.e., 1�a(w)) portion is non-shipping- charge skeptics. Obviously, based on our above discussion, a(w) is decreasing in w.

As far as the aggregate demand functions from the above two groups are concerned,9 they depend on the total prices (i.e., product price + S&H charge, if any) of all retailers in the market. We model the price effect via an exponential functional form (e.g., like in Jeuland and Shugan 1988), where the demand for a ZS retailer (resp., PS retailer) is decreasing in its own total price, and increasing in the total prices of its competitors, both ZS and PS. Because of the heteroge- neous characteristics inherent in the two groups, we assume that their demand sensitivities towards any change in S&H charges are different—specifically, skeptics are more sensitive than the other group. We represent this in our model by parameters snsc (0 \ snsc � 1 Þ and ssc ( � 1), which represent the rela- tive weight that non-skeptics and skeptics place on any change in S&H charges, respectively. We also assume that customers in the skeptics group are not willing to buy the product from any PS retailer; rather, they only check free-shipping ZS retailers and decide whether or not to buy from a retailer in that set. The remaining group, who is willing to pay S&H surcharges, considers all retailers as potential pur- chase sources. Moreover, the customer segment who considers only ZS retailers (resp., both ZS and PS retailers) is divided among Nzs retailers (resp., Nzs þ Nps ¼ N retailers) based on the relative popu- larity levels of the retailers, that is, ci. Note that our model captures the fact that if the total price charged

by a ZS retailer (resp., PS retailer) is too high, then not only do some of the customers move to competing ZS retailers (resp., PS retailers), but some others might indeed defect to PS retailers (resp., ZS retailers) also. Obviously, some others might decide not to buy the product at all.10 Therefore, in our framework, the overall demand function for an arbitrary retailer i is as shown below (recall that fi is the strategic choice for retailer i):11

where u and v represent the direct and cross effects of prices on demands for each type of retailer (u > v). We can then develop the profit functions for the two types of retailers. The profit for retailer i in the ZS set is given by (recall that ZS ones offer free shipping)

piNzs;Npsðpi; p�ijfi ¼ ZSÞ ¼ diNzs;Npsðpi; p�ijfi ¼ ZSÞðpi � cðciÞÞ;

ð3Þ

while that for each PS retailer can be represented by

piNzs;Npsðpi; p�ijfi ¼ PSÞ ¼ diNzs;Npsðpi; p�ijfi ¼ PSÞðð1 þ sÞpi � cðciÞÞ:

ð4Þ

The above profit functions capture the basic trade- off of our interest indicated in the previous section. Free shipping by ZS retailers results in their loss of S&H surcharge revenues, but gain in demand since they are able to attract both shipping-charge skeptics and non-skeptics. In contrast, PS retailers receive extra S&H revenues, but lose out on demand from shipping-charge skeptics.

3.1. Competitive (Market) Equilibrium Partitioning Strategy The analysis in this section will provide insights as to how product characteristics affect the equilibrium proportion of PS and ZS retailers in the market. With-

diNzs;Npsðpi; p�ijfi ¼ ZSÞ ¼ ciP

i2Nzs ci aðwÞ exp �upi þ

X j2Nzsnfig

vpj þ X j2Nps

vpjð1 þ sscsÞ 0 @

1 A

0 @

1 A

zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl}|fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{Demand from shipping-charge skeptics

þ ciP i2N ci

ð1 � aðwÞÞ exp �upi þ X

j2Nzsnfig vpj þ

X j2Nps

vpjð1 þ snscsÞ 0 @

1 A

0 @

1 A

zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl}|fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{Demand from non-shipping-charge skeptics ð1Þ

diNzs;Npsðpi; p�ijfi ¼ PSÞ ¼ ciP i2N ci

ð1 � aðwÞÞ exp �upið1 þ snscsÞ þ X

j2Npsnfig vpjð1 þ snscsÞ þ

X j2Nzs

vpj

0 @

1 A

zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl}|fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{Demand from non-shipping-charge skeptics ; ð2Þ

out loss of generality, for the rest of the article we assume

P i2N ci ¼ 1. In order to focus on the product

level analysis, for the time being, suppose that the retailers are symmetric in terms of their popularity, that is, ci ¼ 1N 8i (refer to section 3.2 for asymmetric analysis). This means that their (unit) product, S&H, and total costs are also equal (that is, cpðciÞ ¼ cp, csðciÞ ¼ cs, and cðciÞ ¼ c, ∀i). Recall that the retailers compete by first choosing their partitioning strategies, that is, fi, and then by setting their prices, that is, pi. We solve the game by backward induction starting from the last stage. The following proposition charac- terizes the equilibrium strategy for each retailer. Proofs for all propositions and corollaries are pro- vided in section A.1 of the (online) Appendix.

PROPOSITION 1. There is a unique pure strategy Nash equilibrium to the game. Specifically:

� Suppose that the equilibrium proportion of retailers who offer free shipping to customers, that is, adopt the ZS strategy, is given by b� (so 1 � b� propor- tion adopts the PS strategy). Then, b� ¼ N�zs=N ¼ minðb̂; 1Þ, where b̂ is the unique solution to the following equation:

b̂ exp b̂vNs c

1 þ s þ u

1 þ snsc

� � ðssc � snscÞ

� �

¼ aðwÞ 1 � aðwÞ

1 þ snscs 1 þ s

expðvNsð c 1þs þ u1þsnscÞðssc � snscÞÞ

exp ðu þ vÞ csð1�snscÞ 1þs

h i� � � 1þsnscs

1þs :

ð5Þ

� Let p�zs be the equilibrium price charged by a retailer of the set who opts for the ZS strategy in the first stage, that is, f�i ¼ ZS. Then, p�zs ¼ c þ 1u per unit (s�zs ¼ 0).

� Let p�ps and s�ps be the equilibrium product and S&H prices, respectively, charged by a retailer of the set who opts for the PS strategy in the first stage, that is, f�i ¼ PS. Then, p�ps ¼ c=ð1 þ sÞ þ 1= ðuð1 þ snscsÞÞ per unit and s�ps ¼ s�p�ps ¼ cs= ð1 þ sÞ þ s=ðuð1 þ snscsÞÞ per unit.

The above proposition shows the equilibrium proportion of PS and ZS retailers, and the correspond- ing equilibrium prices, for online markets. Moreover, we can show that b� [ 0, that is, there will always be some ZS retailers in the market. Note that Harrington and Leahey (2007) analyze an oligopoly setting simi- lar to ours in the presence of search cost for S&H charges and show that in equilibrium all retailers will offer free shipping (b� ¼ 1). However, this is not the case in reality, where PS and ZS strategies co-exist. In their model, effectively, all customers are assumed to be shipping-charge skeptics (that is, a = 1). Evidently,

as a approaches 1, the above proposition suggests that b� ¼ 1, i.e., all retailers would opt for the ZS strategy. Our framework is able to reconcile the disparity between theory and practice—when a is not very high, indeed both ZS and PS retailers would exist in a competitive online market. Before proceeding further, it is worthwhile to com-

pare the equilibrium prices that customers will end up paying under the two strategies.

COROLLARY 1. The following are true:

(i) The unit product price is higher for ZS retailers than PS ones, that is, p�ps \ p

� zs.

(ii) The unit total price is lower for ZS retailers than PS ones, that is, p�ps þ s�ps � p�zs.

The above corollary demonstrates that, while cus- tomers generally assume that they are saving on delivery charges due to free shipping by ZS retailers, they are indeed paying higher product prices when purchasing from such retailers (although total price is still lower). Consequently, ZS retailers are “padding” their product prices to counterbalance, at least in part, the loss of revenues from free shipping offers. Their primary motivation for doing so is to attract the cus- tomer segment who is averse to paying any S&H surcharge. As is evident from Proposition 1, the equilibrium

proportion of PS and ZS retailers depends on factors such as the proportion of shipping-charge skeptics in the market (a), the physical characteristics of the par- ticular product (w), and total cost c. We next study how changes in these parameters affect the equilib- rium proportion.

PROPOSITION 2. The following are true:

1. [(i)] b� increases in a and decreases in w, that is, the equilibrium proportion of ZS retailers increases (resp., decreases) in the proportion of shipping- charge skeptics in the market (resp., in the weight and/or volume of the products).

2. [(ii)] Suppose that N is sufficiently large.12 Then, the equilibrium proportion of ZS retailers, that is, b�, decreases in the total cost c.

The first part of the above proposition implies that if the product being sold by the retailer is small and/ or light (i.e., low w), then a sufficiently large number of customers in the market would be skeptical of S&H surcharges (i.e., high a(w)). So, it is optimal for most retailers to offer free shipping to customers (charging for S&H will lead to a substantial demand loss in this case) and vice versa if the product being sold is large and/or heavy. The second part suggests that the higher the product and/or the S&H cost, the more likely it is that most of the retailers would add a S&H

surcharge to recoup some of those costs. We can then expect more premium product categories (higher product costs) to have a larger proportion of PS retailers.

What if the Product Cost Is Random? While the above proposition demonstrated the effects of prod- uct cost c on the optimal partitioning strategy, the assumption there is that online retailers have precise knowledge about the value of c. However, in reality, this is rarely the case. So, one issue of managerial interest is to understand the effects of cost uncertainty facing retailers on their optimal partitioning strate- gies. For example, uncertainty in S&H costs may arise due to fluctuation in prices charged by delivery com- panies, while uncertainty of product costs might be due to fluctuating raw material costs. In order to address the above issue we analyze a

framework similar to the one before (so N symmetric retailers). The only difference is that now we assume the total selling horizon is divided into T periods, and the total cost faced by retailer i at the beginning of each period is cit; t ¼ 1; 2; . . .; T per unit, where cit are i.i.d. random variables (represented by c with mean lc and standard deviation rc). At the beginning of the selling horizon (i.e., t = 0), all retailers simultaneously decide whether each of them would adopt the PS or the ZS strategy. Once they decide on a strategy they cannot change it during the selling horizon. Sub- sequently, the retailers observe the realized total cost per unit at the beginning of each period and set their unit product and S&H prices accordingly. All other assumptions remain the same as before. Analysis of this game leads to the following conclusion:13

PROPOSITION 3. The equilibrium proportion of retailers who decide to adopt the ZS strategy is given by b�ran ¼ minðb̂; 1Þ, where b̂ is defined as follows:

b̂¼ aðwÞ 1�aðwÞ

Ec expðð1�b̂ÞvNð c1þsþ 1uð1þsnscÞÞðssc �snscÞsÞ h i

1þs 1þsnscs

Ec expð�ucð1þsnscsÞ1þs Þ½ � Ec expð�ucÞ½ �

Ec expð�vcð1þsnscsÞ1þs Þ½ � Ec expð�vcÞ½ � �1

� � : ð6Þ

The equilibrium product price to be charged by each re- tailer of ZS and PS sets in period t, t = 1,2,…,T, is the same as in Proposition 1 with c replaced by cit, the actual realization of total cost for retailer i in period t.

In what follows we provide a more detailed under- standing of the effects of cost variability on equilib- rium proportion by assuming that the total product

cost is uniformly distributed between lc � ffiffiffi 3

p rc and

lc þ ffiffiffi 3

p rc, that is, c � Uðlc �

ffiffiffi 3

p rc; lc þ

ffiffiffi 3

p rcÞ,

implying that the mean product cost is lc and the

standard deviation is rc. This yields the following result:

PROPOSITION 4. If c � Uðlc � ffiffiffi 3

p rc; lc þ

ffiffiffi 3

p rcÞ, then

the following are true:

(i) The equilibrium proportion of retailers who opt for the ZS strategy is given by b�ran ¼ minðb̂; 1Þ, where b̂ is defined as follows:

b̂ ¼ aðwÞ 1 � aðwÞ exp

ð1 � b̂ÞvNðssc � snscÞs uð1 þ snscÞ

! f1

1þs 1þsnscs f2 � 1

where f1 ¼ Ec exp ð1�b̂ÞvNðssc�snscÞs1þs c � �h i

and

f2 ¼ Ec expð� ucð1þsnscsÞ1þs Þ h i Ec expð�ucÞ � Ec expð�vcð1þsnscsÞ1þs Þ

h i Ec expð�vcÞ � .

(ii) Moreover, for a constant lc, b � ran is increasing in

rc. So, the proportion of ZS retailers increases as the coefficient of variation (covc ¼ rc=lc) of the total cost becomes larger.

This proposition proves that the degree of cost uncertainty, as represented by coefficient of variation, affects partitioning strategies of online retailers. The intuition behind the above result is as follows. In our model, firms first decide on whether to adopt the ZS or PS strategy and then adjust their prices (unit and S&H, if applicable) by competitively responding to random changes in (unit) product and S&H costs. Note that the uncertainty in cost implies that the equi- librium ZS and PS prices are also uncertain, which, in turn, implies uncertain demand streams for ZS and PS firms. In general, the higher is the degree of cost uncertainty, the higher is the demand risk for online retailers. In that scenario, the extra demand lift pro- vided by a free shipping offer is very desirable since this enables a ZS firm to clear out its stock faster com- pared to a PS firm (thus reducing the risk), even though the ZS firm loses out on S&H revenue. Obvi- ously, if the uncertainty is relatively low, then the benefit of demand lift in terms of reduction of risk is not significant. In that case, PS strategy becomes more desirable because the surcharge revenue can help firms recoup their S&H costs. Therefore, expectation of a relatively stable cost environment results in higher proportion of PS retailers, and as the possibility of cost uncertainty increases, more and more firms start adopting the ZS strategy to attract more demand.

3.2. Retailer-Based Analysis of Partitioning Strategy In order to focus on the effects of product characteris- tics on the equilibrium segmentation in the market among PS and ZS firms, the last section assumed that all retailers are symmetric. In this sub-section our goal

is to study how asymmetry among the retailers, in terms of their relative popularity, that is, ci, affect their partitioning strategies. For expository reasons, we assume in this sub-section that the market consists of two retailers (duopoly), who are different only in terms of their popularity levels (i.e., they are selling the same product). Specifically, the popularity of retailer i = 1,2 is denoted by ci 2 ½0; 1�, and c1þ c2 ¼ 1. This also implies that the product and S&H costs for the two retailers are now different. The unit total cost for retailer i is cðciÞ, and cðciÞ decreases in ci (the decrease in cost might be due to product and/or shipping costs).14 All other model details remain the same as before. That is, both retailers first decide on whether to adopt the ZS or the PS strategy and sub- sequently compete on prices. We can then develop the demand and profit functions for the two retailers as shown in the (online) Appendix. Based on the profit functions, we then analyze the two-stage game, again using backward induction. Keeping in mind the space constraints, we focus only on the more interest- ing case where at least one retailer’s popularity is large enough to exhibit significant cost benefits. The equilibrium strategy for the two retailers is given in the following proposition:

PROPOSITION 5. If at least one retailer’s popularity is above a threshold popularity level ��c, that is, when maxðci; cjÞ � ��c, where ��c is the unique solution to the following equation:

exp usð1 � snscÞ

1 þ s cð ��cÞ

� � ¼ 1 þ snscsð1 � aðsÞÞð1 þ sÞ

then the equilibrium partitioning strategies for the two retailers are as follows:15

1. If ci � ��c � cj or cj � ��c � ci, then the retailer with lower popularity should adopt the PS strategy, and the one with higher popularity should opt for the ZS strategy.

2. If both firms’ popularities are greater than ��c, that is, minðci; cjÞ � ��c, then both retailers should adopt the ZS strategy.

The optimal prices under each strategy are the same as in Proposition 1 (with c replaced by cðciÞ).

The above proposition suggests that popular retail- ers should take advantage of their low S&H costs (due to economies of scale/scope) and offer free ship- ping, while retailers who are less popular should add S&H surcharges to recoup their (relatively high) delivery costs. Offering free shipping will allow pop- ular retailers to get a further demand lift by attracting shipping-charge skeptics. This, in turn, will allow them to reduce their costs even further, making free

shipping offers even more attractive for them. The above proposition also helps us to understand how the physical characteristics of the product, i.e., w, affect the partitioning strategies of the two asymmet- ric retailers (the result below is similar in spirit to Proposition 2(a)).

COROLLARY 2. For given ci and cj, as w decreases, it becomes more likely that both retailer i and retailer j would adopt the ZS strategy.

In summary, our theoretical analysis in this section identifies the strategic rationale as to why some retail- ers opt for partitioned pricing with S&H surcharges (the PS strategy) while others prefer non-partitioned strategy with free shipping offers (the ZS strategy). Clearly, differences in retailer as well as product char- acteristics shape the optimal partitioning strategy that results in the varying level of PS-ZS segmentation observed in online retail markets.

4. Empirical Validation

In this section, we empirically validate the insights obtained from game-theoretic perspective in section 3. Our empirical analysis also seeks to study whether an online retailer’s strategic partitioning decision related to S&H surcharge is associated with its tempo- ral pricing behavior.

4.1. List of Hypotheses Before proceeding further, in Table 1, we summarize our analytical results of section 3 in the form of hypotheses, which will subsequently be validated empirically.

4.2. Description of Empirical Methods Data Following previous studies (e.g., Baye et al. 2004, Smith 2002), we chose a price comparison site Biz- Rate.com (http://www.bizrate.com/) to gather price and non-price data over a 4-month period, from December 1, 2006, to March 31, 2007. BizRate is one of the most popular online central exchanges, on which many sellers post their up-to-date prices for numerous product categories. To facilitate data collection, we developed a proprietary software program that automatically downloads the target HTML pages and parses their content down to extract prices and other relevant information on a daily basis. Automated proprietary software agents such as ours, often called “intelligent bots” or “spiders,” are widely used as the main data collection mechanisms in the literature (Baye et al. 2004, Oh and Lucas 2006, Pan et al. 2002). To ensure the accuracy of this program, we tested it for 2 weeks prior to the actual

data collection and manually inspected its reliability. This validation process affirmed that the software is 100% accurate. Because S&H costs are an important aspect of

our study, we chose two product categories (digital cameras and printers) that differ substantially in terms of weights and sizes and, therefore, S&H costs. For each product category, we selected 12 different products randomly from CNET’s list (http://www.cnet.com), which contains comprehen- sive categories of computer-related products. A detailed list of all the 24 products, along with their shipping weights and volumes, is provided later on in Tables 12 and 13. As one can imagine, the printers in the sample are, on average, substantially heavier and larger than the digital cameras. The number of retailers posting their prices varied from product to product; some specialist vendors do not offer an entire range of products, but limit their sales to particular products (e.g., digital cameras). Consistent with Oh and Lucas (2006), New York City (zip code 10012) was used as the default desti- nation city and regular delivery as the default delivery option, to compute the S&H charges. During the 2-week pre-sampling period, we used another zip code (78712; Austin, Texas) to deter- mine whether the choice of zip code significantly affects S&H fees. The results show that the calcu- lated S&H charges were almost the same between the two cities and, therefore, only New York City was employed as the target destination during the data-collection period. Regarding the characteristics of the retailers in our

sample, 71% are “pure” online retailers who operate only in online markets and 29% are hybrid retailers who offer products in both online and offline chan- nels. As for the types of products sold, 60% of the retailers in the sample sold a wide range of products, including digital cameras and printers, whereas the remaining 40% are specialized in only one specific product category. Finally, many retailers in our sample are popular retailers with established name brands (e.g., Amazon.com, Dell.com, Buy.com), but many others can be considered small or medium-

sized retailers who have operated for a relatively short period of time.

4.2.1. Variables and Measurements. Tables 2–4 describe the definition and operationalization of the variables used in the analysis. First of all, to vali- date Hypotheses 1a and 1b, UNIT_PRICE and TOTAL_PRICE were used as the dependent variables, respectively, while SHIP_POLICY represented the independent variable (retailer-product-time level analysis; refer to Table 2). We used linear mixed effect models to control for the heterogeneities that could arise from both product and time characteristics. A

Table 1 Summary of Analytical Results and Hypotheses

Hypothesis Description

Hypothesis 1a The average product price per unit charged by a PS retailer is less than the average product price per unit charged by a ZS retailer (Corollary 1a)

Hypothesis 1b The average total price per unit of a PS retailer is higher than the average total price per unit of a ZS retailer (Corollary 1b) Hypothesis 2a The proportion of S&H charge skeptics is positively associated with the proportion of ZS retailers in the market (Proposition 2) Hypothesis 2b The cost of products is negatively associated with the proportion of ZS retailers in the market (Proposition 2) Hypothesis 3a The degree of uncertainty in unit costs is positively associated with the propensity of a retailer to adopt the ZS strategy

(Proposition 4b) Hypothesis 3b The propensity of a retailer to adopt the ZS strategy increases as the popularity of the retailer grows (Proposition 5)

Table 2 Variables, Descriptions, and Measurements (Retailer-Product- Time Level)

Variable name Description / measurement

UNIT_PRICE (i,j,t) The price excluding S&H fees of a product j offered by retailer i in time t

TOTAL_PRICE (i,j,t) The total price (i.e., price with S&H fees) of a product j offered by retailer i in time t

SHIP_POLICY (i,j,t) A binary variable indicating whether retailer i offers free shipping (i.e., zero S&H fee) for a product j in time t. Zero S&H charge is coded as 1 and positive S&H charge is coded as 0

Table 3 Variables, Descriptions, and Measurements (Retailer-Time Level)

Variable name Description / measurement

ZS_PROP (j, t) The proportion of retailers who offer free shipping for product j in time t. For each product and each day we calculated the proportion of retailers offering free S&H to the total number of retailers offering the product. A total of 2,711 out of 2,904 samples (24 products 9 121 days) were collected.

PROD_COST (j, t) Product j’s average total price across all retailers selling the product in time t. As explained later on, we use price as a proxy for cost

PROD_VOL (j) Product j’s actual volume in cubic inches used as a proxy to operationalize consumers’ skepticism to pay S&H fees

DEC_EFCT (t) A measurement of seasonal effect. December is coded as 1 and the other months are coded as 0

detailed description of model specifications for all of the empirical analyses will be provided later. Note from Table 2 that, for a given product, TOTAL_PRICE can be computed as the sum of UNIT_PRICE and S&H charges. SHIP_POLICY is a binary variable indi- cating whether the product is offered by ZS retailers (coded as 1) or PS retailers (coded as 0). To test Hypotheses 2a and 2b, ZS_PROP was

employed as the dependent variable, which repre- sents the proportion of retailers in the market who offer free shipping for product j in time t, i.e., b�

(retailer-time level analysis; refer to Table 3). The reason we used a product’s average total price as a proxy for cost (PROD_COST) in the analysis is that, at the aggregate level, cost is one of the primary determinants of the price levels set by retailers, espe- cially in a competitive setting like ours, and therefore they are highly correlated (Blinder et al. 1998). More- over, we used products’ volume in cubic inches (PROD_VOL) as proxies to operationalize the pro- portion of shipping-charge skeptics in the market. Although further validation is necessary, it is reason- able to assume that there will be more consumers willing to pay S&H fees when they purchase larger products; consequently, products with larger vol- umes would have less proportion of shipping-charge skeptics.16 That is, PROD_VOL represents w, and like in section 3, the proportion of skeptics a(w) decreases in w. Note that PROD_COST and PROD_ VOL represent the main variables to be tested in the analysis. In addition, we also included a dummy variable (DEC_EFCT) that captures the possible sea- sonality effect. This variable was included to reflect the fact that many online retailers rely heavily on December sales, and subsequently use different S&H policies during that period. Hypotheses 3a and 3b were tested at the retailer-

product level (refer to Table 4). The dependent variable, FREE_PROPENSITY, measures retailer i’s propensity to offer free shipping for a product j. Because some retailers differentiate their S&H strate-

gies over time, we used a continuous scale, with 100% representing free shipping for product j during the entire four-month period. An index of 75% (resp., 0%) indicates that a product was available with free ship- ping for 3 (resp., 4) months, but for 1 month the retail- er charged S&H fees for that product. Among the independent variables, COST_VAR, indicating the co-efficient of variation of price for a given product sold by a particular retailer, was calculated as indi- cated in Table 4. The degree of a retailer’s popularity, POPULA, was measured as follows. On the first day of each month during the data collection period, we assessed the popularity of each retailer in our sample based on the mechanism used by Linkpopulari- ty.com. This website measures a website’s “popular- ity” by counting the total number of websites that link to it. More specifically, assuming that more popular retailers or websites appear more frequently in search results, Linkpopularity.com queries all three major search engines (Google, Yahoo, and MSN) and returns each URL’s total link counts. We computed the monthly average index based on the average counts reported by the three major search engines. We then log-transformed the average index and used it to gauge the popularity of each retailer in our sam- ple. Lastly, we added two control variables (PROD_ VOL and CHANNEL) to the analysis as controls. PROD_VOL is likely to affect a retailer’s actual costs associated with S&H, while a retailer’s shipping- related decisions are likely to be influenced by its channel strategy (CHANNEL).

4.3. Results Because our data are based on cross-sectional and longitudinal observations, we employed panel data analytics suited to exploring empirical regularities associated with shipping-charge strategies.

4.3.1. Retailer-Product-Time Level Analysis (Hypo- theses 1a and 1b). We used multivariate analyses to validate Hypotheses 1a and 1b (refer to Table 1)

Table 4 Variables, Descriptions, and Measurements (Retailer-Product Level)

Variable name Description / measurement

FREE_PROPENSITY (i,j) Retailer i’s propensity to offer free shipping for a product j across time. It is a continuous variable with the range between 0 and 1; 0 and 1 indicate that retailer i offers free shipping for product j none and all of the time, respectively. A value of 0.5 means that the retailer offers free shipping for half of the time of the product being sold

POPULA (i) Retailer i’s popularity. Linkpopularity.com (http://www.linkpopularity.com) rates websites’ popularity based on search results. We calculated the average value of the search results from three major search engines (Google, Yahoo, and MSN) to reflect each retailer’s popularity and then performed log-transformation

COST_VAR (i,j) Co-efficient of variation of price (across time) for each retailer i-product j combination. It was computed for each such combination by dividing the standard deviation of total price across time by the average price. For example, for retailer 1 - product 1 combination, we computed the standard deviation and average of total prices charged by retailer 1 for product 1 over the data collection period and divided the standard deviation by the average

PROD_VOL (j) As defined in Table 1. The retailers’ actual costs for shipping would be related to this product characteristics CHANNEL (i) Whether or not retailer i offers only online sales: pure online retailer and hybrid retailer are coded as 0 and 1, respectively

regarding the impact of shipping policies on product price. We employed a mixed-effects linear regression to test these hypotheses because both fixed (e.g., product) and random (e.g., time) effects are present in our model. A Durbin–Watson test suggests that an ordinary least squares (OLS) approach is inappropri- ate in our case since using this approach will violate the assumption of independence (Greene 2008). Since product characteristics (e.g., volume or weight) are likely to impact a retailer’s shipping strategy, we used fixed effects to control for the product effect. Moreover, the variation across time is assumed to be random and uncorrelated with the independent variable. In addition, prices of digital cameras and printers tend to be time-sensitive, as they vary over time. Therefore, we used random effects to control for the time factor. By controlling for these fixed and random effects, we can assess the predictor’s (ship- ping strategy) net effect on prices. For model estima- tions, we employed restricted maximum likelihood (REML), which generates unbiased estimates of variance and covariance parameters (Cressie and Lahiri 1993). These statistical procedures address the potential problems of heteroskedasticity and autocor- relation that might arise when analyzing cross- sectional and longitudinal data. We chose linear mixed models instead of general linear models because the former have superior capabilities than

the latter for handling correlated data and unequal variances (McCulloch and Searle 2001). STATA/SE 9.2 was used to perform the analyses. We present the linear effect models based on the form recommended by Laird and Ware (1982):

UNIT PRICEi;j;t ¼ b0 þ b1SHIP POLICYi;j;t þ Vt þ ei;j;t þ Uj ð7Þ

TOTAL PRICEi;j;t ¼ b0 þb1SHIP POLICYi;j;t þVt þei;j;t þUj

ð8Þ

where i, j, and t are retailer, product, and time indices, respectively, b0 is the constant, b1 is the fixed-effect co-efficient, UNIT PRICEi;j;t and TOTAL PRICEi;j;t are the dependent variables, Uj is the product-specific error (unobserved heterogeneity due to products), SHIP POLICYi;j;t is the regressor, Vt � Nð0; w2kÞ is the random-effect variable for time t assumed to be multi- variately normally distributed, w2k are the variances among the random effects assumed to be constant across times, and ei;j;t � Nð0; r2ki;j;tÞ and r2ki;j;t are the covariances between errors in time t. Tables 5 and 6 present the results of Equations (7)

and (8), respectively. The Wald’s chi-squared statistic indicates that the model fit is strong ((Prob > v2) < 0.01) and the coefficients in the model are

Table 5 Mixed-Effects REML Regression (Retailer-Product-Time Level)

Number of obs = 84,091, Number of groups = 24, Group variable: product Wald v2ð1Þ ¼ 78:02, Log restricted-likelihood = �496,975.99, Prob [ v2 ¼ 0:0000 UNIT_PRICE Coef. SE z P > z 95% Conf. Interval SHIP_POLICY 5.508448 .6236456 8.83 0.000 4.286125 6.730771 CONSTANT 603.8902 83.09398 7.27 0.000 441.029 766.7514 Random-effects parameters Estimate Std. Err. 95% CI product:independent sd(time) 0.2470781 0.0367285 0.1846299 0.3306485

sd(_cons) 407.0624 60.01901 304.8991 543.4577 sd(Residual) 89.0359 .2171707 88.61127 89.46256

LR test vs. linear regression: v2(2) = 2.6e+05 , Prob > v2 = 0.0000

Dependent variable: UNIT_PRICE. Significance levels: �p \ 0:05; ��p \ 0:01.

Table 6 Mixed-effects REML regression (Retailer-Product-Time Level)

Number of obs = 84,091, Number of groups = 24, Group variable: product Wald v2(1) = 617.02, Log restricted-likelihood = �497,064.4, Prob > v2 = 0.0000 TOTAL_PRICE Coef. SE z P > z 95% CI

SHIP_POLICY �15.50747 .6242954 �24.84 0.000 �16.73107 �14.28387 CONSTANT 625.3421 84.27606 7.42 0.000 460.1641 790.5202 Random-effects parameters Estimate Std. Err. 95% CI Product:independent sd(time) 0.2542482 .0377416 0.190065 0.3401053

sd(_cons) 412.8535 60.87275 309.237 551.1889 sd(Residual) 89.12857 .2173967 88.7035 89.55568

LR test vs. linear regression: v2(2) = 2.6e+05, Prob > v2 = 0.0000

Dependent variable: TOTAL_PRICE. Significance levels: �p \ 0:05; ��p \ 0:01.

significantly different from zero. The results show that free shipping is positively associated with UNIT_ PRICE (i.e., price without S&H fees) at the 99% con- fidence level (z = 8.8, p < 0.01), indicating that for a given product a unit price charged by ZS retailers is higher than a unit price charged by PS retailers (i.e., p�ps \ p

� zs). Consequently, Hypothesis 1a is supported

by our data. We also found strong support for Hypothesis 1b. The analysis reveals that free shipping is negatively related to TOTAL_PRICE (i.e., price with S&H fees) (z = �24.84, p < 0.01). This result, there- fore, validates Hypothesis 1b that for a given product the total price charged by a PS retailer is higher than the total price of a ZS retailer (i.e., p�ps þ s�ps � p�zs).

4.3.2. Retailer-Time Level Analysis (Hypotheses 2a and 2b). As shown in Table 1, Hypotheses 2a and 2b state that the proportion of ZS retailers in the market is associated with several factors, including proportion of shipping-charge skeptics (Hypothesis 2a), and the total cost of a product (Hypothesis 2b). Regarding our data structure, we included 24 prod- ucts (12 digital cameras and 12 printers), each of which resides within a time series (121 days). There- fore, the maximum sample size is 2904 ( = 249121), 193 of which were omitted (6.6% of the total) due to product discontinuation. Consequently, 2711 sam- ples remained in the analysis. The dependent vari- able for the analysis indicates the proportion of ZS retailers in the market, for a given product and time. Table 7 shows the descriptive statistics and the corre- lation coefficients of the variables included in this analysis. The mean of ZS PROP was 0.48—that is, for a given product and in a given day, approximately 48% of products were offered with free shipping. The average price of the products was $615.67, while the average of PROD_VOL and DEC_EFCT was 5.50 and 0.25, respectively. As for the bi-variate correla- tion coefficients, ZS_PROP has a significant negative association with PROD_COST and PROD_VOL, while PROD_COST was positively correlated to PROD_VOL. Similar to the procedures used to test Hypotheses

1a and 1b, we developed a mixed effect model in which the proportion of ZS (ZS_PROP), the depen- dent variable, is thought to be affected by both fixed and random effects. More specifically, we assumed

that PROD_COST, PROD_VOL, and DEC_EFCT all affect the proportion of ZS; therefore, these regres- sors’ effects are thought to be fixed. However, the regression results might also vary depending on how we sampled the products and time periods, and there- fore we subsequently assumed the presence of ran- dom effects on products and time periods. Our regression model is as follows:

ZS PROPj;t ¼ b0 þb1PROD COSTj;t þb2PROD VOLj;t þb3DEC EFCTj;t þUi þVt þej;t

ð9Þ

where j and t are product and time indices, respec- tively, b0 is the constant, bk, k = 1,…,3 are the fixed effect coefficients, Ui is the product-specific error (unobserved heterogeneity due to products), Vt � Nð0; w2kÞ is the time-specific error component, w2k are the variances among the random effects assumed to be constant across times, and finally, ej;t � Nð0; r2kj;tÞ affects only the particular observa- tion where r2kj;t are the covariances between errors in time t. Note that Ui is unobserved heterogeneity specific to products, while Vt is time specific hetero- geneity, which is peculiar to all observations for that time period t. The error term, ej;t, is uncorrelated with the time specific component and the product specific error. We included these terms to control for possible biases arising from unobserved hetero- geneities inherent to sample procedures. Table 8 presents the results of the linear mixed

effect model (9) developed at the market level. The model as a whole has statistically significant predic- tive capability, as reflected in the strong F-value (Prob > v2 ¼ 0:0000). This Wald’s chi-squared sta- tistic reveals that all the coefficients in the model are different from zero. The two-tail p-values tests indi- cate that all of the regressors have a significant influ- ence on the dependent variable, ZS_PROP. Consistent with our analytical results, the proportion of ship- ping-charge skeptics (recall that higher PROD_VOL implies lower proportion of skeptics) is negatively related to the proportion of ZS retailers in the market, that is, b� decreases in w (z = �5.31, p < 0.01). Conse- quently, Hypothesis 2a was strongly supported. Our data also strongly support Hypothesis 2b, showing

Table 7 Summary Statistics and Correlation Coefficients (Retailer-Time Level) (N = 2711)

Mean SD 1 2 3 4

1. ZS_PROP 0.4781003 0.1051404 1 2. PROD_COST 615.6742 411.1025 �0.3060�� 1 3. PROD_VOL 5.496703 2.764333 �0.6851�� 0.4024�� 1 4. DEC_EFCT 0.2559941 0.436499 �0.0470 0.0016 �0.0004 1 ��Correlation is significant at the 0.01 level (two-tailed). �Correlation is significant at the 0.05 level (two-tailed).

that products’ total costs are negatively related to the proportion of ZS firms in the market, that is, b�

decreases in c (z = �2.48, p < 0.05). This finding sug- gests that, ceteris paribus, low-cost products are less likely to be subject to S&H charges than their high- cost counterparts.

4.3.3. Retailer-Product Level Analysis (Hypo- theses 3a and 3b). In Table 9, we present the descrip- tive statistics and the correlation coefficients of the variables included in this analysis. The mean of FREE_PROPENSITY, which indicates the average propensity that a retailer offers free S&H, was 0.49. The average values of COST_VAR, POPULA, and PROD_VOL were 0.035, 8.388, and 5.047, respectively. Based on the bivariate correlation analysis, we found that FREE_PROPENSITY is significantly associated with POPULA, PROD_VOL, and CHANNEL at the 95% confidence level. Interestingly, POPULA is posi- tively correlated to COST_VAR. These results suggest that the more popular a retailer is, the higher the degree of price variability within a retailer over time will be. To test Hypotheses 3a and 3b (refer to Table 1), we used a fixed effect model (Equation (10)) since the data are collected across retailers and prod- ucts. Given that time is not of our interest in this test, a mixed effect model with random effects was not necessary. The fixed effect eliminates the heteroske-

dasticity that is present due to the unobserved retailer characteristics.

FREE PROPENSITYi;j ¼ b0 þ b1COST VARi;j þ b2POPULAi þ b3PROD VOLj þ b4CHANNELi þ

X Retaileri þ ei;j

ð10Þ

where i and j are retailer and product indices, respectively, b0 is the constant, bk, k = 1,…,4 are the fixed effect coefficients, FREE PROPENSITYi;j is retailer i’s propensity to offer free shipping for a product j across time,

P Retaileri is for fixed effects

that control for unobserved heterogeneity due to retailer characteristics, and ei;j � Nð0; r2Þ. Table 10 shows the regression results at the retailer-

product level given by Equation (10). We present the results of two models (model 1 and model 2), which differ only in terms of the presence of product volume as a control variable. Hypothesis 3a posits that the degree of uncertainty in costs (COST_VAR) is posi- tively associated with the propensity of the retailer to offer free S&H (Table 1). The results of both models in Table 10 statistically support this hypothesis, albeit at different significance levels. Finally, the results show that the popularity of a retailer (POPULA) is positively associated with its propensity to offer free S&H at the 99% confidence level (t = 4.20, p < 0.01).

Table 8 Mixed-Effects REML Regression (Retailer-Time Level)

Number of obs = 2711, Number of groups = 24, Group variable: product Wald v2(5) = 80.64, Log restricted-likelihood = 3998.254, Prob > v2

= 0.0000

ZS_PROP Coef. SE z P>z 95% CI PROD_COST �0.0000648 0.0000262 �2.48 0.013 �.0001161 �.0000135 PROD_VOL �0.021033 0.0039608 �5.31 0.000 �.028796 �.01327 DEC_EFCT �0.0189404 0.0034452 �5.50 0.000 �.0256928 �.012188 CONSTANT 0.6461126 0.0241895 26.71 0.000 0.598702 0.6935232 Random-effects parameters Estimate SE 95% CI product:Independent sd(time) 0.0006454 0.0000979 0.0004795 0.0008687

SD(_cons) 0.0487997 0.0080176 0.0353646 0.0673388 SD(Residual) 0.0531127 0.0007282 0.0517045 0.0545592

LR test vs. linear regression: v2ð2Þ ¼ 1798:14, Prob [ v2 ¼ 0:0000

Dependent variable: Proportion of ZS (ZS_PROP).

Table 9 Summary Statistics and Correlation Coefficients, n = 978 (Retailer-Product Level)

Mean SD 1 2 3 4 5

1. FREE_PROPENSITY 0.4887423 .4815146 1 2. COST_VAR 0.0346662 0.0882884 �0.0129 1 3. POPULA 8.388421 3.244646 0.1614�� 0.1810�� 1 4. PROD_VOL 1871.508 (5.047) 2734.832 (2.758) �0.1709�� 0.1911�� 0.034 1 5. CHANNEL 0.3137652 0.4642572 0.0959�� 0.0185 0.2113�� -0.172�� 1

��Correlation is significant at the 0.01 level (two-tailed). The values in parentheses indicate log transform of actual values.

That is, the more popular a retailer is, the greater its likelihood to offer free S&H for a given product. Regarding the two control variables, in model 2 both PROD_VOL and CHANNEL are negatively associ- ated with the dependent variable. Before proceeding further, we will summarize our

analysis to this point. To this effect, Table 11 provides a snapshot of our theoretical results and their empiri- cal support. In addition to the above statistical analysis, our

study also offers some economic implications related to retailers’ S&H charge strategies (refer to Tables 12 and 13). First, based on over 84,000 samples, we found that the total price of a digital camera, charged by PS firms, is (on average) 3.4% higher than the total price

charged by their ZS counterparts; for printers, the total price of PS firms is (on average) 4.5% higher than the one charged by ZS firms. Therefore, ZS retailers with free S&H indeed offer consumers something “free” in the form of lower prices, compared to PS retailers who charge such fees. For example, consum- ers who buy a $1000 printer can save $45, on average, when they use a ZS retailer. If the empirical regulari- ties found in this study prove to be a general phenom- enon across all product categories, PS retailers tend to take advantage of these “opaque” charges to improve their revenues substantially. To gain additional insights into PS retailers’ S&H

related tactics, we analyzed the extent to which PS retailers charge S&H for the two product types.

Table 10 Regression Results (Retailer-Product Level)

Retail-product level variables Model 1 (n = 978) Model 2 (n = 978)

Cons 0.2107363� (0.1220657) 0.2102114� (0.1210264) COST_VAR 0.4530033�� (0.1772018) 0.2998289� (0.1797218 ) POPULA 0.0617392�� (0.014693) 0.0647958�� (0.0145875) PROD_VOL �0.0000114�� (2.82e-06) CHANNEL �0.6624503�� (0.0878865) �.6718424�� (.0871691) F-value 59.67�� 60.19��

Adjusted R2 0.8362 0.8390

Dependent Variable: FREE_PROPENSITY. Numbers in parentheses indicate standard errors. Significance Levels: �p \ 0:1;��p \ 0:05; ��p \ 0:01.

Table 11 Summary of Theoretical and Empirical Results

Hypothesis Unit of analysis Theoretical result Empirical support

H1a Retailer-Product-Time Corollary 1: p�ps \ p � zs Supported

H1b Retailer-Product-Time Corollary 1: p�ps þ s�ps � p�zs Supported H2a Retailer-Time Proposition 2: Equilibrium proportion of ZS retailers (i.e., b�) decreases in w Supported H2b Retailer-Time Proposition 2: b� decreases in c Supported H3a Retailer-Product Proposition 4b: b� increases in covc Supported H3b Retailer-Product Proposition 5: Equilibrium partitioning strategy (i.e., f �i ) switches from PS to ZS as ci increases Supported

Table 12 Product Characteristics (Printer)

Printer Volume (cu-in) Shipping weight (lbs) ZS total ($) PS total ($) (PS-ZS)/ZS Shipping %

HP LaserJet 4250N Laser Printer 4590 59 1176.82 1203.98 0.023 0.035 HP LaserJet 4250DTN Laser Printer 7140 84 1714.5 1748.38 0.020 0.037 HP LaserJet 1320 Laser Printer 6101 30 381.49 404.01 0.059 0.057 HP LaserJet 4250TN Laser Printer 6120 78 1430.48 1453.07 0.016 0.036 HP LaserJet 4250 Laser Printer 9903 58 876.96 898.73 0.025 0.044 HP DeskJet 460c Inkjet Printer 274 7 225.6 230.44 0.021 0.053 HP LaserJet 1160 Laser Printer 1918 28 320.5 322.98 0.008 0.077 Epson Stylus Photo R1800 Photo Inkjet 2774 36 536.26 557.8 0.040 0.047 Epson Stylus Photo R2400 Photo Inkjet 2592 37 811.63 853.08 0.051 0.035 Epson Stylus Photo R800 Photo Inkjet 3863 23 381.85 401.79 0.052 0.054 Brother DCP-7020 Laser Printer 3076 30 200.88 225.79 0.124 0.101 Brother MFC-8220 Laser Printer 4284 38 285.68 314.31 0.100 0.073 Median 4074 37 459.055 480.905 0.033 0.050 Average 4386 42 695.220 717.863 0.045 0.054

The volume and weight information for Tables 12 and 13 is collected from Amazon.com.

Interestingly, the results indicate that PS retailers charge a higher percentage of S&H fees in relation to the total price for large/heavy products, such as printers, than they do for small/light products, such as digital cameras. For example, for a digital camera, S&H fees constitute (on average) 3% of the total price of a given product. However, PS retailers charge almost twice as much (on average, 5.4% of the total price) in S&H fees for printers. Note that, in this com- parison, we calculate S&H fees as a percentage of the total price of a given product. Furthermore, we found that, compared to digital cameras, a larger variation exists among PS retailers in S&H charge percentages. This percentage ranges from 1.6% to 4.7% for the 12 digital cameras in our sample. However, the variation increases substantially for printers, ranging from 3.5% to 10.1%. What this might imply is that for larger/ heavier products (e.g., printers) depending on where a consumer purchases such products, the S&H fees vary significantly, even within PS retailers. In con- trast, a relatively moderate “risk” exists when con- sumers purchase smaller products (e.g., digital cameras) from PS retailers. All these findings indicate that S&H fees have an important economic implica- tion for consumer welfare and price competition.

4.3.4. Price Adjustment Behavior of PS and ZS Retailers. Although studying the price adjustment behaviors of retailers in detail is beyond the scope of this study, we provide initial insights into such phe- nomena. ZS retailers tend to change prices more frequently than PS retailers. More specifically, for a given product, the retailers in our sample changed the price, on average, 2.2 times in a month, that is, one price change every 13.6 days. When comparing the two types of retailers in question, ZS retailers changed prices for a particular product, on average, 2.6 times in a month (or every 11.5 days), whereas PS retailers adjusted them, on average, only 1.8 times during the

same period (or every 16.7 days). In other words, the prices posted by PS retailers “live” approximately 30% longer than the prices offered by ZS retailers. Interestingly, no significant difference was observed between ZS and PS retailers with respect to the direc- tion (i.e., positive or negative) and magnitude of the price changes. For example, 64% and 62% of price changes by PS and ZS retailers, respectively, were associated with price decreases, while the (absolute) magnitude of price changes was 6.2% for ZS retailers and 7.3% for PS retailers. Why do ZS retailers change prices more frequently

than PS retailers? We offer two possible explanations for this phenomenon. The first is the negative senti- ment of consumers in response to price changes. The kinked demand curve theory (Okun 1981) posits that a decrease in price, unless by a substantial amount (e.g., over 25% of the original price), does not help a firm attract many consumers. In contrast, an increase in price even by small amount is likely to cause it to lose many consumers. We found that a high correla- tion exists between product prices and S&H fees (r = 0.85) for PS retailers, which suggests that for such retailers S&H fees are in proportion to product prices. From a consumer’s point of view, price changes occur on both components of the price when dealing with PS retailers. This “double whammy” is likely to have an additional negative impact on consumers’ pur- chasing decisions. Moreover, a small increase in S&H fees might discourage shipping-charge skeptics from buying the product, but a small decrease in S&H fees is unlikely to have a major impact on their purchase decisions. For these reasons, PS retailers are less active in changing prices relative to ZS retailers. The second explanation is related to the difference in firm size. Based on a survey, Buckle and Carlson (2000) found that large firms are more likely to raise their prices in response to cost increases than small firms, although no significant difference was detected in the

Table 13 Product Characteristics (Digital Camera)

Digital camera Volume (cu-in) Shipping weight (lbs) ZS total ($) PS

total ($) (PS − ZS)/ZS Shipping %

Canon PowerShot SD600 6.43 2 231.05 240.24 0.033 0.047 Olympus Stylus 720 SW 7.1 2 327.37 341.62 0.053 0.034 Canon PowerShot SD700 IS 7.92 3 331.16 352.23 0.054 0.032 Canon PowerShot u710 IS 15.81 2 321.59 340.40 0.067 0.034 Canon PowerShot SD630 6.34 3 269.33 282.59 0.046 0.041 Canon PowerShot u640 21.24 3 347.73 366.18 0.055 0.033 Canon EOS Digital Rebel XT 46.25 4 617.27 604.33 �0.019 0.023 Canon PowerShot SD900 9.11 2 418.85 422.80 0.017 0.030 Canon Digital Rebel XTi 48.1 4 757.87 783.69 0.028 0.021 Canon PowerShot G7 19.99 3 529.23 552.62 0.045 0.024 Canon EOS 30D 69.43 5 1251.61 1248.08 �0.003 0.016 Olympus EVOLT E-330 46.2 6 839.41 858.99 0.029 0.020 Median 17.90 3.00 383.29 394.49 0.039 0.031 Average 25.33 3.25 520.21 532.81 0.034 0.030

case of cost decreases. According to these authors, the shorter price duration in large firms occurs because they, compared to small firms, are quicker to respond to demand and cost changes and gain greater benefits from making price changes. Our data from the popu- larity analysis show that PS retailers are generally less popular than ZS retailers (implying that PS retailers are also possibly smaller than ZS retailers). This sug- gests that PS retailers are slower than ZS retailers in responding to market changes and receive relatively little gain from price adjustments. Consequently, PS retailers are less “motivated” than their ZS counter- parts in making price adjustments.

5. Managerial Implications

In the last two sections we theoretically and empiri- cally analyzed the price partitioning strategies of online retailers. In this section, we discuss key mana- gerial implications of our results and also show how certain real world phenomena are consistent with our results. Our analysis suggests that if a retailer wants to offer

free shipping, it ought to be aware that such a tactic has a cost associated with it. The retailer in that case is subsidizing, at least partly, the S&H costs, which can erode or even possibly destroy its margins. Moreover, if the retailer adds some of the S&H costs to product price, such an action might adversely affect demand, especially for price-sensitive customers. This is indeed consistent with the opinion of some of the practitioners actually making decisions about whether to offer free shipping or not (Bolotsky 2008, Brown 2008). We suggest that such an offer makes sense if: (i) the retailer is operationally efficient so that its S&H costs are low and the “cost of free shipping offers” is not significant; (ii) if the retailer is large (i.e., popular) enough so that it can negotiate a low prod- uct and/or S&H costs with its suppliers; (iii) the demand loss from not providing free shipping is sub- stantial; and (iv) there is a high possibility of a risky cost environment in the future. This means that online retailers need to think about

customizing their price partitioning strategies based on the attributes of the products that they are selling as well as retailers’ own characteristics. For example, cost (S&H)-to-price ratios for larger/heavier/fragile prod- ucts are relatively high. For such products, there are less shipping-charge skeptics since customers will incur high transportation cost if they want to buy the products offline; so, demand loss from charging S&H fees for these products would not be high, and retailers can opt for PS strategy. Free shipping offers perhaps ought to be targeted toward light/small, premium- price products for which customers expect the S&H costs to be relatively low and retailers to bear such

costs. At the very least, retailers need to develop an aggregate measure of cost (S&H)-to-price ratio (i.e., if most of the products in their portfolios have high or low cost (S&H)-to-price ratios) in order to decide on the partitioning strategy. Similarly, retailers also need to think about their operational efficiency and/or bar- gaining power while deciding on shipping policies. When retailers are large or efficient enough such that their S&H costs are low, offering free shipping makes sense. But for a relatively less popular retailer or one whose delivery operation is not properly optimized, such offers might be “too costly.” The retailers not only need to worry about average product and/or S&H costs, but also about the potential variability in those costs. For example, if retailers expect that the volatility in fuel costs would make S&H costs charged by carri- ers like UPS, FedEx, and USPS more variable or that the fluctuation in prices of commodity metals would make product costs less stable, our model suggests that we should see more retailers opting for ZS strat- egy. On the other hand, expectation of a stable cost structure going forward (at least in the short term) would induce most retailers to opt for PS strategy. Indeed, our data suggests that about 40% of the retail- ers in our sample differentiated their shipping policies across products and across time. Note that the shipping charge strategy adopted by

Amazon.com seems to be consistent with our analy- sis. Amazon.com now uses a threshold S&H-charge policy; that is, if the total purchase value is above that threshold, then customers get free shipping, other- wise not. However, if we look at the history of Amazon.com, it did not offer any free shipping from its inception in July 1995 until November 1999. Even when it started offering free shipping, the threshold was $100, and only later on (August 2002) did Amazon.com reduce the threshold to $25 (Lewis et al. 2006). As indicated earlier, in our setting, a reduction in the threshold value is (approximately) equivalent to changing from the PS to the ZS strategy. So, as our model suggests, Amazon.com indeed started with the PS strategy, but as it established its brand equity and improved its S&H operations, it took advantage of economies of scale/scope and opted for the ZS strategy.17

Lastly, recall that, in the last section, we demon- strated that ZS retailers change prices more frequently than PS ones. However, managerial decisions about price adjustments ought to be determined based on the optimal equilibrium between the economic gains from “flexible pricing” and the “consumer costs” (Goodfriend 1997) that arise from the consumer’s negative reaction in response to price changes. ZS retailers should be aware that the benefits accrued from dynamic pricing tactics might be short-lived, as consumers discover the “randomness” in ZS retailers’

prices. One potential negative consequence is that this might encourage strategic behavior on the part of con- sumers. For example, ZS customers might delay their purchases in the anticipation of price cuts. These delayed purchases might eventually result in demand– supply imbalances.

6. Concluding Remarks

Pricing is an important element in any firm’s strategy, whether it operates in online or offline markets. For online firms, inherent S&H costs pose an additional strategic challenge. This study sought to provide an analytical and empirical vantage point from which to observe and interpret competitive actions related to S&H charges. More specifically, we developed an oli- gopoly model of (Nash) price competition among online retailers, which captures the basic trade-off between PS and ZS strategies—the ZS strategy attracts more customers but loses out on S&H charge revenues, while the PS strategy earns more revenue per unit at the cost of lower demand. Analysis of our theoretical model illustrated the effects of both prod- uct and firm level characteristics on the relative domi- nance of these two partitioning strategies among the retailers. Our results first showed that PS retailers should

charge lower product prices than ZS retailers, but their total prices (including S&H) should be higher than ZS ones. As far as the effects of product characteristics are concerned, this study showed that the optimal strategy for retailers with a large segment of ship- ping-charge skeptics is to offer free shipping to their customers. The same ZS strategy should be ideal for retailers selling low-priced products and products associated with high levels of cost uncertainty. A retailer’s popularity and its product portfolio are also important factors in choosing between PS and ZS strategies. When the retailer is relatively less popular or its product portfolio consists of a large number of products with a high cost-to-price ratio, then the opti- mal strategy should be to add a S&H surcharge. One distinguishing feature of our study is that we vali- dated our theoretical results using an empirical study. Data were collected for 4 months about actual product and S&H prices charged by a large number of online retailers for two product categories—one of which is relatively small and light but expensive (digital cam- eras), whereas the other one is relatively larger, hea- vier, and cheaper (printers). All of our theoretical results were supported by the empirical data. Our empirical results also showed that the price adjust- ment behavior of ZS firms differs from that of PS firms. ZS firms tend to have frequent price changes, while the pricing policy of PS retailers is relatively more stable. But there was no significant difference

between PS and ZS firms in terms of magnitude and direction of price changes. There are three worthwhile avenues for pursuing

future extensions of our research. First, we assume the model horizon to be short and so do not include transient behavior of customers and retailers in our framework. A more comprehensive model should capture how a retailer’s pricing and partitioning strat- egy over time affects its demand and brand equity. It would require a dynamic modeling perspective, and in that case the data collection period for the empirical study should also be sufficiently long. Second, it would be useful to empirically study the performance implications of the two shipping modes in terms of financial measure like profitability. Third, a demand model that captures more complex consumer behav- ior would perhaps be helpful for developing real-life pricing decision support systems, although analyzing it would be a quite challenging endeavor. Note that one of the assumptions in our theoretical

modeling framework is that shipping-charge skeptics only consider ZS firms while making purchase deci- sions. An alternative formation might be to assume that all customers consider all firms (i.e., both ZS and PS) while deciding to buy. The only difference between the two customer segments is that skeptics weigh any S&H charge more than non-skeptics (e.g., they weigh this charge by ssc and snsc, respectively, where ssc � snsc). We discuss and analyze such a framework in details in (online) Appendix A2, which turns out to be quite complex. Some of the results (e.g., parts of Proposition 1 and Corollary 1) can still be analytically established. However, because of the involved expressions for certain equilibrium deci- sions, analytically proving some other characteriza- tions is quite difficult. Our extensive numerical study clearly demonstrates that almost all the qualitative the- oretical insights of the paper hold true even for that model, although the equilibrium values are different. As far as our empirical study is concerned, most such stud- ies have several common inherent limitations, such as endogeneity and reverse causality. Regarding endo- geneity, our independent variables do not by any means cover all known factors influencing a retailer’s free shipping offer. Future research should identify other key factors influencing such a decision. Regard- ing reverse causality, our empirical models define the causal relationships between the variables in such ways that they provide empirical insights into the findings we obtained from analytical assessments. However, reverse causality could occur. For example, price-partitioning strategies might drive a retailer’s popularity rather than the other way around. Although this reverse relationship was not the objec- tive of our study, future research can expand along this line of inquiry.

In closing, we would like to point out that while the reasons for an online retailer’s choice between offer- ing free shipping and charging shipping fees can be complex, consumers’ attitudes toward shipping fees, product features, and the retailer’s market status might be key determinants of a firm’s price partition- ing strategy and are therefore critical in sustaining a competitive advantage.

Acknowledgments

The authors thank EICs for the special issue, an anony- mous senior editor, and three anonymous referees for constructive comments that led to significant improve- ments in this paper. The authors’ works were supported in part by research grants from the Natural Sciences and Engineering Research Council of Canada (NSERC RGPIN 355570-09 for Mehmet Gumus, NSERC 222773 for Shan- ling Li, NSERC 202889 for Saibal Ray); the Social Sciences and Humanities Research Council of Canada (SSHRC 214832 for Mehmet Gumus and Saibal Ray, SSHRC 410- 2010-1703 for Wonseok Oh); and Fonds de recherche sur la Société et la culture of Québec (FQRSC NP-132114 for Mehmet Gumus).

Notes

1Handling includes activities such as picking, packing, and customer service. Also, note that we use the terms S&H and delivery interchangeably throughout the article. 2Note that any reference to the product price in this article excludes applicable taxes. 3Amazon.com’s gross profits for its overall operations were $5531M and $4270M in 2009 and 2008, respectively (Amazon.com Annual Report 2009, p. 26). 4Paypal’s July 2009 survey showed that 46% of online con- sumers abandoned their purchases due to high S&H costs (http://blog.searchenginewatch.com/090706–160859). According to Forrester Research, 75% of the consumers prefer free shipping retailers and 58% claim that S&H prices deter them from shopping online (Mulpuru 2008). 5According to Shampanier et al. (2007), when Amazon introduced free shipping in some European countries, the price in France mistakenly was reduced not to zero but to one French franc (about 10 cents). Whereas the number of orders increased dramatically in the countries with free shipping, not much change occurred in France. When Amazon.fr rectified the mistake and offered free shipping, orders in France also jumped significantly. 6Ideally, attractiveness factor ci should be a function of product and shipping prices (as well as other tangibles and intangibles) in the long run. Since we focus on the short term, we assume ci to be exogenous. 7Suppose that there is only one option available for deliv- ering to end consumers (e.g., regular delivery). 8Indeed, as we indicated earlier, Kawamoto (2008) refers to a comscore survey where a large segment of customers indi- cated that if an online retailer starts charging an S&H fee, then they would opt for another “free shipping” retailer.

9Our demand functions are aggregate ones. Although they capture consumer preferences, they are not based directly on individual consumer choice models. 10An alternative formulation of the model might be to assume that both groups of customers consider all firms, ZS and PS, and the two groups are differentiated only by their sensitivities to S&H charge, that is, snsc and ssc. We discuss such a model in our concluding remarks (section 6) and analyze it in detail in (online) Appendix A2. It turns out that almost all our qualitative theoretical insights hold true even for that model, although not all them can then be proved analytically. 11In order to see how a price change affects demands in our framework, note the following example. If PS retailer i increases its product price (and hence its S&H charge), then such an action decreases i’s demand from the non- skeptic group. Some of these non-skeptics move to other PS retailers and also to ZS retailers. The price increase also results in more shipping-charge-skeptic customers who consider only ZS retailers (diNzs;Nps ðpi; p�ijfiÞ is affected by a price increase in a PS retailer through both skeptics and non-skeptics). Evidently, the effects of the price increase on the skeptic and non-skeptic groups are different because of how they weigh any S&H charge. 12Indeed, large N is a sufficient condition. The result might be true even for small N as shown in the Appendix. 13ExðAÞ represents expected value of expression A, where the expectation operator is over the random variable x. 14As regards product cost, many retailers get a price dis- count for high volumes from suppliers. Since S&H for many retailers is handled by third parties, higher volumes also provide cost benefits in that case (Bolotsky 2008). 15As we show in the Appendix, the condition maxðci; cjÞ � ��c is not very restrictive. It will be satisfied as long as w or u or s is relatively low and/or snsc is rela- tively high. 16Note that our results remain the same even if we use products’ weight, rather than volume, as the proxy. 17Refer to Schepp and Schepp (2009) about how Amazon improved the efficiency of its S&H operations over time and Brown (2008) about Amazon’s bargaining power with USPS. Amazon’s strategy contrasts with e-tailers like Web- van who were too aggressive about free shipping offers too early in their lifecycles, and paid the price by actually losing money on most of the items they sold.

References Amazon.com Annual Report. 2009. Amazon.com Inc., Seattle.

Available at http://phx.corporate-ir.net/phoenix.zhtml? c = 97664 & p = irol- reportsannual (accessed August 9, 2012).

Anderson, C. 2009. Free: The Future of a Radical Price. Hyperion, New York.

Bakos, Y., H. Lucas, W. Oh, G. Simon, S. Viswanathan, B. Weber. 2005. The impact of e-commerce on competition in the retail brokerage industry. Inf. Syst. Res. 16(4): 352–371.

Barsh, J., B. Crawford, C. Grosso. 2000. How e-tailing can rise from the ashes. McKinsey Quart. 3: 98–109.

Baye, M., J. Morgan, P. Scholten. 2004. Price dispersion in the small and in the large: Evidence from an internet price com- parison site. J. Ind. Econ. 52: 463–496.

Becerril-Arreola, R., M. Leng, M. Parlar. 2009. An online retailer’s optimal decisions on profit margin, free-shipping threshold and inventory for a holiday season: An empirical study with simulation. Working paper. Anderson School of Management, UCLA.

Blinder, A., E. Canetti, D. Lebow, J. Rudd. 1998. Asking About Prices: A New Approach to Understanding Price Stickiness. Rus- sell Sage Foundation, New York.

Bolotsky, D. 2008. One retailer’s view on why not to offer free shipping. Commentary by Founder and CEO of Uncommon- goods.com in Shop.org, 22 August 2008. Available at: http:// blog.shop.org/2008/08/22/one-retailers-view-on-why-not-to- offer- free-shipping/ (accessed August 9, 2012).

Brown, T. 2008. Free shipping: Customer centric or lost profit. Com- mentary by Sr. Director & GM of Timberland.com in Sho- p.org, 3 September 2008. Available at: http://blog.shop. org/2008/09/03/free-shipping-customer-centric-or-lost- profit/ (accessed August 9, 2012).

Buckle, R. A., J. A. Carlson. 2000. Menu costs, firm size and price rigidity. Econ. Lett. 66: 59–63.

Cattani, K. D., G. C. Souza. 2002. Inventory rationing and ship- ment flexibility alternatives for direct market firms. Prod. Oper. Manag. 11(4): 441–457.

Cheema, A. 2008. Surcharges and seller reputation. J. Consumer Res. 35: 167–177.

Clay, K., R. Krishnan, E. Wolff, D. Fernandes. 2002. Retail strate- gies on the web: Price and non-price competition in the online book industry. J. Ind. Econ. L 3: 351–367.

Clemons, E., I. Hann, L. Hitt. 2002. Price dispersion and differen- tiation in online travel: An empirical investigation. Manage. Sci. 48(4): 534–549.

Cressie, N., S. N. Lahiri. 1993. The asymptotic distribution of REML estimators. J. Multivar. Anal. 45: 217–233.

Freedman, L. 2008. 3rd Annual Mindset of the Multi-Channel Shopper Holiday Study. Research conducted by the e-tailing group, November 2008. Available at http://www.e-tailing.- com/content/wp-content/uploads/2008/11/atg_111108_white- paper.pdf (accessed August 9, 2012).

Greene, W. 2008. Econometric Analysis, 6th edn, Pearson Prentice Hall, Upper Saddle River, NJ.

Goodfriend, M. 1997. A Framework for the analysis of moderate inflations, J. Mon. Econs. 39: 45–65.

Harrington J. E. Jr., M. F. Leahey. 2007. Equilibrium pricing in a (partial) search market: The shopbot paradox. Econ. Lett. 94: 111–117.

Jeuland, A. P., S. M. Shugan. 1988. Channel of distribution profits when channel members form conjectures. Market. Sci. 7(2): 202–210.

Johnson, M. E., R. D. Meller. 2002. Performance analysis of split- case sorting systems. Manuf. Serv. Oper. Manag. 4(4): 258–274.

Johnson, M. E., S. Whang. 2002. E-business and supply chain management: An overview and framework. Prod. Oper. Manag. 11(4): 413–423.

Kawamoto, D. 2008. comScore offers e-commerce retailers holiday advice. C-Net news, November 6, 2008. Available at http:// news.cnet.com/8301-1023_3-10084394-93.html (accessed August 9, 2012).

Laird, N.M., J. H. Ware. 1982. Random-effects models for longitu- dinal data. Biometrics. 38: 963–974.

Leng, M., M. Parlar. 2005. Free shipping and purchasing decisions in B2B transactions: A game-theoretic analysis. IIE Trans. 37: 1119–1128.

Lewis, M. 2004. The influence of loyalty programs and short-term promotions on customer retention. J. Market. Res. XLI: 281–292.

Lewis, M., V. Singh, S. Fay. 2006. An empirical study of the impact of nonlinear shipping and handling fees on purchase incidence and expenditure decisions. Market. Sci. 25(1): 51–64.

McCulloch, C. E., S. R. Searle. 2001. Generalized, Linear and Mixed Models. Wiley, New York.

Morwitz, V. G., E. A. Greenleaf, E. J. Johnson. 1998. Divide and prosper: Consumers’ reactions to partitioned prices. J. Market. Res. XXXV: 453–463.

Mulpuru, S. 2008. Outlook For US Online Holiday Sales, 2008. 21 October 2008 Available at. http://www.forrester.com/rb/ Research/outlookforusonlineholidaysales%2C2008/q/id/47380/ t/2. (accessed August 9, 2012).

Netessine, S., N. Rudi. 2006. Supply chain choice on the internet. Manage. Sci. 52(6): 844–864.

Oh, W., H. Lucas. 2006. Information technology and pricing deci- sions: Price adjustments in online computer markets. MIS Quart 30(3): 755–775.

Okun, A. 1981. Prices and Quantities: A Macroeconomic Analysis. The Brookings Institution, Washington.

Pan, X., B. Ratchford, V. Shankar. 2002. Can price dispersion in online markets be explained by differences in e-tailer service quality? J. Acad. Market. Sci. 30(4): 433–445.

Pyke, D., M. E. Johnson, P. Desmond. 2001. E-fulfillment: Its harder than it looks. Supply Chain Manag. Rev. 5(1): 26–32.

Schepp, B., D. Schepp. 2009. Amazon Top Seller Secrets: Insider Tips from Amazon’s Most Successful Sellers, Amacom, New York, NY.

Schindler, R. M., M. Morrin, N. N. Bechwati. 2005. Shipping charges and shipping-charge skepticism: Implications for direct mar- keters’ pricing formats. J. Interact. Market. 19(1): 41–53.

Shampanier, K., N. Mazar, D. Ariely. 2007. Zero as a special price: The true value of free products. Market. Sci. 26(6): 742–757.

Smith, M. 2002. The impact of shopbots on electronic markets. J. Acad. Market. Sci. 30(4): 442–450.

Smith, M., E. Brynjolfsson 2001. Consumer decision-making at an internet shopbot: Brand still matters. J. Ind. Econ. 49: 541–558.

Thaler, R. H. 1985. Mental accounting and consumer choice. Mar- ket. Sci. 4: 199–214.

US Census Bureau News. 2008. Quarterly Retail E-Commerce Sales: 2nd Quarter 2008. August 15, 2008. Available at ftp://ftp. census.gov/retail/releases/historical/ecomm/08q2.pdf (accessed August 9, 2012).

Zhao, X., K. Stecke, A. Prasad. (2008). Lead time and price quota- tion mode selection: Uniform or differentiated? Working paper, The School of Managemeet, The University of Texas at Dallas, TX.

Supporting Information Additional Supporting Information may be found in the online version of this article:

Appendix S1. Proof for propositions and Corollaries.

Appendix S2. Extensim: An Alternative Formation of the Demand Model and its Analysis.

Please note: Wiley Blackwell is not responsible for the con- tent or functionality of any supporting materials supplied by the authors. Any queries (other than missing material) should be directed to the corresponding author for the arti- cle.

Copyright of Production & Operations Management is the property of Wiley-Blackwell and its content may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's express written permission. However, users may print, download, or email articles for individual use.