Risk, Convenience, Cost and Online Payment Choice: A Study of eBay Transactions

Risk, Convenience, Cost and Online Payment Choice:
                         A Study of eBay Transactions

                                                Haizheng Li*
                                            School of Economics
                                       Georgia Institute of Technology
                                          Atlanta, GA 30332-0615

                                               Richard Ward
                                       DuPree College of Management
                                       Georgia Institute of Technology
                                            Atlanta, GA 30332

                                                 Han Zhang
                                       DuPree College of Management
                                       Georgia Institute of Technology
                                            Atlanta, GA 30332


* The corresponding author. The research was supported in part by the iXL Electronic Commerce Center of DuPree
College of Management, Georgia Institute of Technology. We are grateful to eBay traders who participated in our
survey, and to Isaac Springfield, Haibo Huang and Lynn Yang for their helpful comments.

       Online payments are a critical issue in electronic markets. This research investigates

online payment choices using probit and nested logit model based on the survey data we

collected from eBay users. We develop a theoretical framework to model payment choices

between the trader partners based on risk, convenience and cost dimensions. Then, we analyze

how product attributes, traders’ characteristics and payment attributes affect the payment choice.

Our findings suggest that the cost and inconvenience associated with a payment method

discourages its use in online transactions. Product attributes, especially uncertainties associated

with the product quality, appear to have stronger effect in affecting payment choices than traders'

characteristics. We also find that a seller’s reputation rating does not have a significant effect on

actual payment choices, but affects the payment options offered by a seller.

J.E.L Code:     L14, L86

Key Words:     Payment Method, Online Auction, eBay, Nested Logit Model
1. Introduction

        As an efficient and flexible sales channel, online auction businesses are becoming an

internationally successful phenomenon. Individuals use auction sites as a market to conduct

online “garage sales”; companies use auction sites to liquidate unwanted inventory, as well as to

assist in pricing new products, acquiring new markets for low-margin items, and reaching

markets that would otherwise be too expensive to reach.

        Presently, eBay, Yahoo!, and Amazon.com are the major players in online auction

markets. According to eBay, its site has over 49 million registered users. Forrester Research

projects that online consumer marketplaces will conduct over 25% of all online sales by 2006.1

Recently, Forrester Research changed its definition of online retail to include auctions, because

survey respondents make 10 percent of their purchases on eBay.2

        As a radically different channel for consumer purchases, online auction markets have

drawn significant attention from the research community. Concentrating on the asymmetric

information problem (Akerloff 1970) in electronic markets, much literature on online auctions

addresses trust issues (e.g., Kollock 1999, Brynjolfsson and Smith 2000, Ba et al. 2003).

Particularly, considerable attention has been devoted to the effects of reputation systems on

online auctions. For example, Resnick and Zeckhauser (2001), using a large data set from eBay,

have found that sellers with better reputations were more likely to sell their items, but they

enjoyed no boost in price. Melnik and Alm (2002) studied the effects of seller’s reputations on

the willingness of buyers to bid on their items, and find that reputations have a small, but

positive and statistically significant impact on price. Ba and Pavlou (2002) studied the extent to

  Forrester Research 2001, “SMEs Present Latent Opportunity For Europe's Online Auctions, Says
Forrester”, http://www.forrester.com/ER/Press/Release/0,1769,498,00.html ,February 13
  CNET 2002, Sept. 4, “Channel Advisor Announces First Annual Channel Advisor Strategy Summit”,
which trust can be induced by proper feedback mechanisms in electronic markets and find that

feedback systems can generate price premiums for reputable sellers.3

        Yet, one critical area in online auction markets has remained unexplored: how do trading

partners select a payment method for settling their transactions? Payment systems are an

essential part of electronic commerce. Although common media of exchange are used in

electronic markets, the spatial distances between buyers and sellers in electronic markets

increase their anxieties about the successful completion of their transactions. According to the

Internet Fraud Watch, operated by the National Consumers League, online auction sales have

remained the number one source for Internet fraud in the past years. In 2001, 70% of the fraud

cases reported to the Internet Fraud Watch were online auction related. The average loss per

person in online auction fraud rose from $326 in 2000 to $411 in 2001.4

        Currently, there are different payment methods available in online auction markets, such

as personal check, money order, cashier’s check, credit card and so on. These payment methods

differ in risk, convenience and cost. But what factors cause online traders to choose a particular

payment method? Does the seller’s reputation affect payment preference? This study attempts

to fill the research gap in the literature and to address these questions.

        The determinants of payment choices will have important implications for e-commerce.

First of all, to facilitate the development of electronic markets, it is vital to have an efficient

online payment system. Alan Greenspan, Chairman of the Federal Reserve, commented that

“payment systems are critical to the functioning of a modern monetary economy” (Greenspan

1996, p. 695). Understanding the choice of payment transactions between online traders will

  Similar results can be found from Lucking-Reiley et al. (1999), Brynjolfsson and Smith (2000), and Houser and
Wooders (2001).
  Internet Fraud Watch. 2002. Available: http://www.fraud.org/internet/2001stats10mnt.htm.

contribute to designing such an efficient system. Secondly, for online payment processing

agencies, knowing the factors that affect payment choices will help them to meet the needs of

online traders and to form strategies that provide them greater shares of their competitive market.

       Using the survey data we collected from eBay users, we study the choice of payment

methods in online transactions. We first develop a simple framework to model seller and buyer’s

behavior in terms of balancing risk, inconvenience and costs associated with each payment

method. Then we apply probit and nested logit model to analyze the factors that affect the

probability of choosing a particular payment method.

       The paper is organized as follows. Section 2 discusses risk, convenience and cost related

to payment methods. Section 3 presents a theoretical analysis of payment method selection in

online auction markets. Section 4 discusses the survey. Section 5 and 6 discuss the results from

probit and nested logit models. Section 7 concludes the paper.

2. Risk, Convenience, and Cost of Online Payment

       It appears that risk, convenience, and cost are among the most important issues in online

payment transactions. Abrazhevich (2002) conducted a survey of user attitudes towards

electronic payment systems, and found that ease of use, convertibility of funds, security and trust

are among the most important features. Carow and Staten (1999) investigated consumer

preferences among debit cards, credit cards, and cash for gasoline purchases and found that

convenience was the greatest determinant for using credit cards. Mantel (2000) also found that

low cost, convenience, privacy and security are among the key factors affecting consumers’

payment decisions. Notably, in the G10 working report on electronic money, developing low-

risk, low-cost, and convenient payment and financial services for consumers and businesses is

listed as a major objective for these countries’ banking and financial authorities.5

        In general, online payment methods differ in terms of risk, convenience, and cost. Most

of the risk associated with online transactions arises from the physical separation between buyers

and sellers. Therefore, asymmetric information problems exist between the trading partners.

First, sellers and buyers cannot see each other, and they know each other not by real

identification, but by virtual ID. Furthermore, online auction sites usually excuse themselves

from the responsibility of user authentication. For example, eBay’s User Agreement states:

“Because user authentication on the Internet is difficult, eBay cannot and does not confirm each

user's purported identity.”

        Second, in electronic markets, buyers cannot closely examine the product before

purchase. When bidders view a product listed at an online auction site, their experience of the

product’s quality is constrained by the limitations of the electronic medium. Online auction sites

often claim that they “have no control over the quality, safety or legality of the items advertised,

the truth or accuracy of the listings,” thus exposing bidders to potential risks and fraudulent


        The risk in online transaction is also affected by the payment method itself. As

summarized in Table 1, for buyers, credit cards generally provide more protection than cash-

equivalents (e.g., cash, money order, cashier’s check, etc.), because under the Fair Credit Billing

Act, buyers have the right to withhold payment on poor-quality or damaged merchandise

  “Report of the Working Party on Electronic Money”, Source: Paper version from U.S. Department of the Treasury,
Public Affairs. Electronic version: http://www.bis.org/publ/gten01.pdf. The G10 countries are: France, Germany,
Italy, Switzerland, the Netherlands, Belgium, Sweden, the UK, the US, Canada and Japan. The G10 grouping of
countries have set up various committees, including the Committee on Payment and Settlement Systems, which
report to, the G10 central bank Governors. The G10 established the famed Basel Capital Accord, which sets
standardized capital requirements for G10 banks, often regarded as the cornerstone of the global financial system.
  Directly quoted from eBay’s User Agreement: http://pages.ebay.com/help/basics/f-agreement.html.

purchased with a credit card.7 However, not all consumers like to use credit cards online due to

various security concerns (Ba et al. 2000). For instance, buyers are afraid that their credit card

information may be intercepted by a malicious party during transmission over the Internet, or

that the sellers may misuse their credit card information.

         From the seller’s point of view, however, cash-equivalents are more secure for on-line

transactions. If a buyer uses a stolen credit card number, the seller will suffer a loss, because the

seller will have to refund the sales price, in addition to losing the product.8

         Besides the transaction security, trading partners are also concerned about transaction

convenience. Convenience is becoming an increasingly important issue in modern society. In

fact, “Every day, people are adopting new technologies that trade a small amount of their privacy

for greater convenience”.9 A major function of Internet business is to provide convenience in

information gathering and order processing (Kim et al. 2002). And shopping convenience plays

an important role in the success of Internet commerce (Torkzadeh and Dhillon 2002).

         Payment methods have various levels of convenience. For example, a credit card is more

convenient for the buyer than a money order or a cashier’s check. A credit card can be

electronically processed, but money orders and cashier’s checks first must be obtained from an

issuer and then mailed.10 Moreover, the buyer has to wait longer to receive the product. For the

seller, credit cards are also more convenient, since the buyer’s payments are automatically

collected and forwarded to the seller’s bank quickly via electronic transmission. In addition, the

  People usually think that payment method’s float is an important feature to protect buyers. Float is defined as the
point from reception of the payment by the seller until the point, where funds are withdrawn from the buyer's bank
account. In this research, we do not treat it as “protection” if a buyer just changes his/her mind during the float
duration before receiving the product.
  In general, in online transactions, sellers can ship the product after receiving the fund. Thus, sellers are exposed to
less risk with non-credit card payments.
  The Economist (2002), "The telephone is the tool – Digital cash schemes favor the mobile phone,” London, April.
   Western Union offers online money order service through www.Bidpay.com. In this case, the buyer does not
need to take a trip to buy a money order. However, the money order still needs to be mailed physically to the seller.

seller receives some money management services, such as account detail and summary

information. For sellers with high selling volume, such features are particularly helpful.

           In addition to risk and convenience, payment methods are also associated with costs.

Such costs are referred to as transaction fees. As listed in Table 1, either the seller or the buyer

pays for the transaction fees depending upon the payment method. For example, a money order

fee is charged to a buyer by an issuer; but a credit card transaction fee is charged to a seller by a

credit card merchant services company like PayPal.

           In the event of an unsatisfactory product, the buyer requests a refund from the seller and

ships the product back to the seller. The original transaction fees may not be recovered and new

transaction fees for the reverse payment may be expended. For example, for credit cards, these

reverse transaction fees, called chargeback fees, are from $10 to $20. Additionally, if a seller’s

cumulative chargeback fees pass a certain threshold, the seller’s merchant services provider will

also increase the seller’s future transaction fees. As discussed in some literature, fraud costs to

online merchants include: chargeback fees, higher transaction fees, and fines imposed by credit

card associations.11

           Therefore, payment methods for online transactions differ in risk, convenience, and cost.

As transactions vary, different payment methods become more or less attractive to buyers or

sellers. In the next section, we will develop a simple model for payment choices.

3. A Simple Model

           Since online payment methods differ from each other, buyers and sellers may each have

different preferences. Therefore, the seller and buyer jointly decide on a payment choice.

     Lee, Mie-Yun, “ Low-cost measures can prevent high-cost online fraud,” Bizjournals, November 20, 2000.

Presumably, if they find a mutually acceptable payment method, the transaction will proceed, or

the online sale will fail. We define a complete transaction as one where the buyer receives and

does not return the product and the seller receives payment in full amount. If a transaction is

completed, the seller gains profits from the sale and the buyer gains utility from the product.12

        For the seller, assume that the potential gross profit for the sale is a function of the selling

price p and is defined as f(p). Besides price, the gross profit also depends on the cost the seller

paid of acquiring the product.13 During the transaction, the seller incurs certain sunk cost Co

associated with selling the product, such as the listing fee to an online auction agency and

advertisement cost (like photos). In addition, the seller incurs payment transaction costs. If

payment method j is used for the money transaction, the transaction costs for the seller include

direct cost Cdjs and the implicit cost Cej. The direct cost is the fee paid for using a particular

payment service, such as the fee charged by a credit card company. Implicit cost is the indirect

cost resulting from the inconvenience associated with using the payment method j. For example,

if paid by personal check, the seller will experience a delay before receiving the payment and

will have to physically deposit the check. Therefore, the sellers' potential net profit after the

transaction becomes:

                 Yj =f(p) - Co - Cdjs - Cej.

        However, it is possible that the product may be returned by the buyer or the buyer may

use a fraudulent form of payment. If the probability that the product will be returned by the

buyer is Prj, and the seller’s direct cost for refunding is Crj (for example, the charge back fee for

a credit card), the seller’s total loss when the product is returned is (Co + Cdjs + Cej + Crj).14

   To focus on payment method, we ignore the shipment issue.
   For example, on eBay, the three basic listing fees are insertion fees, final value fees and listing upgrade fees.
   Some fee components may be waived, for example, with PayPal, if the refund is through PayPal system, the seller
can re-capture Cdj and escape Crj.

Finally, assume the probability of buyer fraud is Pfjb (for example, when the buyer pays

for the product using a stolen credit card number). In this case, the total loss for the seller is (p +

Co + Cdjs + Cej).15

         Therefore, for the seller, the expected net profit associated with the transaction for using

payment j is (the probability of return and fraud may depend on the payment method adopted.):

         πj= [f(p) - Co - Cdjs - Cej]⋅(1-Prj-Pfjb) - (Co +Cdjs + Cej + Crj)⋅Prj - (p + Co + Cdjs + Cej)⋅Pfjb.

         The seller will choose payment option j for the transaction if

                  πj= max{πs1 , πs2 …πsk}.

         On the other hand, if the seller and buyer cannot agree on payment method, the

transaction will fail, and the loss for the seller will be the sunk cost Co. Thus, the seller has an

incentive to make the transaction if the expected profit is higher than the seller’s sunk cost, i.e.,

πj > - Co.

         From the buyer’s perspective, he or she will gain a certain level of utility from consuming

the product. The utility depends on the price p and ∂u⁄∂p
can be viewed as the respective inconvenience and risk concerns experienced by the buyer,

thereby lowering the total utility.

           If the product is returned to the seller, the buyer does not consume the product and the

utility becomes urj = -u(Cdjb, Vtj+Vrj, Rtj), where Vrj is the additional buyer inconvenience

associated with waiting for and cashing a refund.

           In the case of seller fraud, the seller does not ship the product after receiving the money.

The buyer’s utility becomes ufj=-u(p+Cdjb, Vtj, Rtj), where the money loss is specified as

equivalent to an increase in direct payment cost,

           Assume that the probability of returning the product is Prj, and the probability of seller

fraud is Pfjs. Then, the expected utility for the buyer is

                     Uj= utj⋅(1-Prj -Pfjs)+urj⋅Prj+ufj⋅Pfjs.

           Thus, among the k different payment options, the buyer will choose j if

                              Uj= max{U1 , U2 …Uk}.

           We assume that the buyer’s utility is zero if the transaction stops at this stage.16 Thus the

buyer would accept a payment option j, if Uj > 0.

           Although the buyer and the seller can accept any payment option j as long as πj> - Co and

Uj>0, the actual payment method chosen depends on the bargaining between the seller and the

buyer. Among k different payment options, define Minπ as the minimum expected profit for the

seller and MinU as the minimum expected utility for the buyer, thus the seller’s surplus of

choosing payment j can be defined as ∆ξj=(πj - Minπ), and the buyer’s surplus of choosing

payment j can be defined as ∆φj=g(Uj - MinU), where g is a function that converts utility

measure into money measure.

     This assumption implies that there is no disutility associated with the searching process for the buyer.

Assume that the relative bargaining power for the seller is represented by λ and λЄ[0,1],

and the bargaining power for the buyer is (1-λ). The payment choice should be determined by

the weighted total surplus between the seller and the buyer. Thus, assuming a unique maximum

exists, the payment option l will be adopted if the weighted surplus

                       Tl=Max{(1-λ)·∆φj+λ·∆ξj, j=1, 2, …k}.

       If the seller completely dominates the bargaining, λ=1, the payment method used will be

the best for the seller. If the seller has no bargaining power, λ=0, the buyer will decide the

payment method. In general case, both the buyer and the seller will influence the payment

choice. If the best option for the buyer and the seller happens to be the same, it will certainly be

adopted for the transaction.

       Therefore, based on the simple model, the seller’s expected profit and the buyer’s

expected utility is affected by the risk, inconvenience and cost associated with each payment

method. Yet, the perceived risk, inconvenience and actual cost also depend on the attributes of

the product in the transaction and the characteristics of the traders. Therefore, the weighted

surplus for the trader partners Tj will be a function of product attributes Xp, the characteristics of

the trader partners Xt, and the features of a particular payment method Xm. For transaction i via

payment method j, the weighted surplus can be represented by:

               Tij=h(Xpij, Xtij, Xmij) + εij, j=1,2,…,k,

where εij is the random error. Thus, Tij is the latent variable that determines the observed choice

of the payment method in a transaction. This equation forms a basis for our empirical analysis.

       In general, seller’s and buyer’s characteristics include experience in online transactions,

reputation ratings, and even age and education. Product attributes include product price and

uncertainties associated with the product quality. Specifically, product price may affect the

perceived risk and the actual costs for a payment transaction. Moreover, the higher the

uncertainties associated with the product quality, the higher the risk involved. On the Internet, it

is easy to judge the quality of commodity products (e.g., oil, paper clips), but difficult to judge

“look and feel” products (e.g., suits, art) (De Figueiredo 2000), because they differ in non-digital

attributes, which can only be evaluated by perceiving them in person (Lal and Sarvary 1999).

When purchasing online, buyers generally have more confidence in the quality of commodity

products, but relatively less confidence in the quality of “look and feel” products due to

information asymmetry. In addition, product uncertainties will also affect the perceived

probability of returning the product. Therefore, product attributes, such as whether under

warranty, whether used or not, generally provide additional information to reduce the


4. Survey and Data

        In order to investigate the payment choice empirically, we conducted a survey of eBay

traders. eBay is the major player in the online auction market, with 85 percent market share.17

eBay makes its money by charging sellers fees for listing their items, plus commissions based on

a percentage of an item’s selling price. After an auction closes, the winning bidder and the seller

are responsible for completing the sale.

        With the rapid development of online auctions, many online payment processing systems

have emerged. For example, PayPal (www.paypal.com) is a trusted third party enabling

consumers to send and receive payment online. Escrow (www.escrow.com) is emerging as

another type of trusted third party, which provides online escrow service for transferring

  Cisneros, O.S., 2000, “eBay Accused of Monopolization”, WiredNews. Available:
http://www.wired.com/news/business/0,1367,37871,00.html, July 31.

products and payments between online sellers and buyers.18 eBay also launched its own online

payment system, called Billpoint.19 In addition, some sellers have their own credit card

processing capability to take payments directly. Thus, on eBay, bidders and sellers may choose

the following different payment methods: cash, personal check, money order, cashier’s check,

credit card (via PayPal, eBay, or seller’s own facility), debit card, electronic check and escrow


        Ideally, to study the factors that affect payment method selection, the questionnaire

should be designed for a particular buyer and seller pair involved in an online transaction.

However, it is practically impossible to contact pairs of buyers and sellers. Therefore, the

questionnaire was designed to solely target the eBay sellers.

        Based on the simple model, we focused on the effect of payment features, product

attributes, and traders’ characteristics. The questionnaire was composed of the following

sections: 1) eBay experience, 2) Last eBay sales transaction, including various product attributes,

and 3) personal demographic information. The questionnaire was sent to five eBay “Assistance

Traders” and five student eBay sellers for pretest.20 Based on the feedback received, revisions

were made to the questionnaire. The questionnaire was then encoded into a web page using

Microsoft’s Active Server Pages web interface technology, and a Microsoft Access database was

connected to the web page.

   Escrow service works as follows: First, the escrow company collects payment for the merchandise from the buyer.
When the payment clears, the seller is notified to ship the item. Second, the buyer notifies the escrow company
when the merchandise is received and is satisfactory. Finally, the escrow company will then release the payment to
the seller.
   eBay announced the completion of its acquisition of PayPal in October 2002. Billpoint remained a payment
option on eBay through the end of 2002 and will begin to be phased out in early 2003. When we did the survey,
Paypal was an independent company.
   On eBay, “Assistance Traders” are experienced sellers and volunteers to help other people in online transactions.
Their phone numbers, email addresses and eBay usernames are available online.

The original research plan attempted to follow Ba and Pavlou’s (2002) survey

methodology. First, we randomly selected sellers on eBay based on different types of products;

second, we sent a message, which contained the questionnaire’s URL, to the selected sellers

through eBay’s email system, and invited them to participate in our survey. Unfortunately, after

we sent out about 20 emails to the selected sellers in March 2002, we received a warning

message from eBay, indicating that we were violating an eBay rule by sending a URL through

eBay’s email system.

        Therefore, we turned to post our questionnaire on discussion forums on eBay and related

auction sites, such as Auctionwatch (www.auctionwatch.com) and Auction Guild

(www.auctionguild.com). We did not offer a monetary reward for participation in the survey,

but we promised to make our research outcome available online. Fortunately, this method of

data collection was very satisfactory and many respondents were quite enthusiastic and even

offered many good comments, suggestions and questions about this research.

        When we first posted the survey online, we asked the respondents to provide their eBay

ID, which would help us obtain the sellers’ reputation ratings on the eBay website (eBay posts its

members’ reputation ratings). Unlike the pretest, many respondents did not want to reveal their

eBay ID, even though we promised that their answers were completely confidential. Therefore,

we had to add one question about the seller’s eBay reputation rating (total number or ratings and

total number of positive ratings). Still, some respondents did not answer this question.

        A total of 260 unique eBay Sellers completed the survey online.21 The descriptive

statistics are summarized in Table 2. Among sellers, the average age is 41 years old, average

education is 15 years, and 52% sellers are female. The average eBay selling experience is 3.8

  Regarding the randomness of the sample here, we do not believe that those who choose to respond are correlated
with their preferences on payment methods. Thus, we do not see a sample selectivity problem here.

years. Only 2% sellers do not have any buying experience on eBay. Their average number of

monthly transactions for the past six months is 124, and the average monthly sale is $2,193.

Among the 260 respondents, 110 respondents reported their eBay reputation ratings.22 The

average total reputation rating is 447, and the average positive reputation rating is 439.

         For the last product sold by the seller (the most recent transaction) on eBay, 33% are

under warranty, 57% are used products, and the average price for the products was $130 with a

standard deviation of $420. In addition, in order to measure the “look and feel” nature of the

product in the transaction, we include a question asking the seller to estimate the buyer's desire to

physically examine the product before purchasing on a 5-to-1 scale (high-to-low desires). The

average scale is 1.8.

         It appears that 72% of payments are rendered using credit cards, the majority of which

being rendered via PayPal (61.5%).23 Money orders account for 12.7% of transactions, while

personal checks account for 10.8%. Three questions are asked on the payment method(s)

initially proposed by the seller, the buyer, and the actual method adopted in the transaction.

Detailed information on payment choice is summarized in Table 3.

5. Probit Analysis

         In our sample, since escrow service is not used, all payment choices can be naturally

grouped into two categories: cash equivalents (cash, money order, cashier’s check, personal

check) or credit cards (via eBay or via Paypal or via seller processing). The common

   It is possible that those with higher reputation ratings are more likely to report them. Yet the smaller sample that
contains information on reputation does not pose the sample selection problem in the regression analyses in the
following sections, given that reputation rating is exogenous to payment choices and that there is always some
chance of not observing any reputation ratings, see Wooldridge (2002, Chapter 17) for the discussion.
   For comparison, a Gartner report states that credit cards account for 93% of online purchase payments
(CommerceNet Newsletter, “The Public Policy Report,” Vol. 3, No.5 May 2001). Note that Gartner report is not
limited to online auction.

characteristic is that cash equivalents must be physically mailed from the buyer to the seller,

while comparable credit/debit card transfers are electronic. Credit cards are generally more

convenient for both buyers and sellers.

         We apply the bivariate probit model to investigate the choice between these two broad

categories.24 The advantage of the probit analysis is that the effects of product attributes and

traders’ characteristics on payment choices can be easily investigated. In a more complicated

model such as nested logit, discussed in the next section, it becomes necessary to interact these

variables with choice dummies in order to identify their effects. Thus, it greatly increases the

number of parameters to be estimated, and in addition, only relative effects can be identified (i.e.,

by normalizing one choice specific parameter to zero.)

         According to the simple model in Section 3, the weighted surplus between the seller and

the buyer determines the actual payment choice, and thus can be viewed as the underlying latent

variable in the probit model.25 In the model, sellers’ characteristics include average number of

transactions per month, years of eBay experience, reputation rating, gender, age, and education;

and product attributes include price, whether the product is used or new, whether still under

warranty, and its degree of being a look and feel product.

         Table 4 reports estimated coefficient and marginal effects for different models: Model 1

(seller characteristics only), Model 2 (product attributes only), Model 3 (both), and Model 4

(with seller’s reputation rating). Marginal effect is evaluated at the sample average of the

   Because sellers usually cannot tell if buyers use credit card or debit card if the sellers get the payment through
PayPal or eBay payment system, we cannot separate credit card from debit card. However, they are equivalent in
terms of convenience and cost. As for protection, based on Federal Reserve Bank of Philadelphia, using credit card
is protected up to a point by provisions of Regulation Z, which incorporates the provisions of the Fair Credit Billing
Act. For debit card, certain provisions of Regulation E provide some protection. This protection is not as extensive
as that provided by Regulation Z, but many debit card issuers voluntarily provide protection similar to that provided
by Regulation Z for credit card purchases. Thus, the difference between the two is small.
   Based on the payment selection model, the seller and the buyer can choose not to transact if they cannot agree on
payment methods. Since such data are not available, our model here excludes this possibility.

independent variables. Log-likelihood value, likelihood ratio test, and Pseudo R-squared are

reported with the results. The Likelihood Ratio tests show that each model is highly significant,

with P-values of Chi-square statistics less than 1%. Based on the log-likelihood values, Model 3

fits the best because it controls for both seller’s characteristics and product attributes. Since

these two groups of attributes are not correlated with each other, as expected, the results from

Model 3 are almost identical to Model 1 and 2 (minor changes in significance are caused by

multicollinearity and reduced degree of freedom).

       Among seller characteristics, the monthly volume of transactions is significant at the 5%

level (Model 1). If the number of transactions increases by 100, the chance of selecting a credit

card increases 4.6 percentage points. In general, a seller should prefer cash equivalents given

their low costs and high safety nature. However, in terms of convenience, credit cards certainly

have advantages over payment collecting and record-keeping. Holding all else constant, if sales

volumes increase, the costs of inconvenience associated with payments increase, and expected

profits are lowered. Therefore, sellers with larger sale volumes prefer a credit card.

       The effect of seller's type is almost significant at the 10% level. If a seller has never

bought anything through eBay, he or she is more likely to push for a cash-equivalent. For a

buyer, credit cards are generally preferable because of their protections, convenience, and low

cost relative to cash-equivalents. Thus, it is possible that the buying experience of a seller helps

him or her to understand the buyer’s preference for credit cards, and thus increases the seller’s

willingness to accept credit cards. On the other hand, years of selling experience do not appear

to affect payment choices.

       Age and education are not significant at the 10% level. Their signs indicate a negative

effect on the probability of choosing a credit card. The same effect is found for gender. Based

on the sign, perhaps, older sellers and female sellers are relatively more conservative and risk

averse; educated sellers are generally aware of credit card risk better. Thus, their expected profit

of using credit card is lower, and they prefer cash-equivalents.

        Among product attributes, product warranty is highly significant in affecting payment

(Model 2). If a product is still under warranty, the probability of choosing a credit card increases

17.8 percentage points. On the other hand, the higher the degree of the look and feel nature of

the product being sold, the more likely that a cash-equivalent method of payment will be chosen

(almost significant at 10% level). Used products, compared to new products, also increase the

probability of choosing a cash-equivalent payment (although statistically insignificant).

        These results are consistent with the model in Section 3. Unwarranted products, used

products, and products possessing a high look and feel nature all increase product uncertainty.

Thus, the buyer’s probability of returning the product will be higher. In this case, a credit card

leads to additional chargeback costs for the seller and also makes the return of the product easier

for the buyer, thereby reducing the seller’s expected profit. A strategic seller thus will push for a

cash-equivalent. When product uncertainty is reduced because of a warranty, a new product or

low-degree look and feel nature, the seller becomes more willing to use credit cards, and thus

increase credit card use in actual transactions. Based on this result, it appears that the seller has

more bargaining power in deciding the payment method. Interestingly, the price does not have a

significant effect.26 This is probably due to a conflict of interest between buyer and seller.

Specifically, from the seller’s perspective, the higher the price, the higher the credit card

transaction fee and the higher the risk associated with credit card fraud. Thus, cash-equivalents

  Since the buyer first bids on price and then negotiates payment method with the seller, price should not be
endogenous. To check this possibility, we estimate model 2 and 3 without price. The results are almost the same
for variables included.

are preferable. From the buyer perspective, the higher the price, the higher the risk associated

with the seller fraud, such as non-delivery.27 Therefore, credit cards are preferable.28

         Since several studies find that reputation ratings have important implications for online

trading, it is interesting to investigate whether reputation ratings have any effects on payment

choices. Model 4 (Table 4) includes reputation rating as an additional seller’s characteristic.

Because the average percentage of positive ratings is very high (above 95%), we use a dummy

variable indicating whether a seller has at lease one negative reputation rating.29 However, the

effect of reputation turns out to be not statistically significant. Most likely, a buyer is concerned

most about a seller’s negative rating, and thus pushes for using a credit card, in order to have

more protection and easier product return capability; but the seller with a negative rating may

push for a cash-equivalent for self-protection. Thus, the net effect becomes ambiguous.

         On the other hand, since the actual payment method used is determined jointly by the

seller and the buyer, it is desirable to investigate whether a seller would have behaved differently

if he or she can dictate the payment choice. This can also provide some insight on the payment

negotiation procedure. Therefore, we use probit model to study whether the seller offered a

credit card option. In this model, the underlying latent variable for the probit model is just the

seller’s surplus. The results are reported in Table 5.

         Similarly, based on Model 1 in Table 5, the seller transaction volume is significant at the

10% level in increasing the chance of the seller’s extending a credit card option. Gender and

education variables change signs, but they are not significant. Statistically, the seller’s type (i.e.,

   It is possible that price is correlated with some product attributes such as warranty. Yet, this is merely a
multicollinearity problem.
   Credit card processors use cumulative transaction limits or single transaction limits to minimize risk exposure.
For example, PayPal requires information on a buyer’s bank account, if the buyer’s cumulative purchases reach a
$2000 limit. Since the average price is much lower than the limits, the effect, if any, should be negligible.
   In alternative specifications, we used percentage of positive rating, percentage of negative rating, and number of
negative rating. The results do not vary much.

whether the seller has any buying experience on eBay) has a stronger effect on actual credit card

usage than in offering to accept credit cards, possibly because of the negotiation process.

        The effect of product attributes on a seller’s offering a credit card option is presented in

Model 2. Again, used products have a highly significant negative effect upon offering credit

cards, while a product warranty has a highly significant positive effect. For example, a used

product (relative to a new one) will reduce the probability of offering credit cards by 9.7

percentage points. Consistent with the discussion for Table 4, with higher uncertainty associated

with the product, the chance of the return is higher and the expected profit for the seller from

using credit cards are lower; therefore, the seller is more reluctant to offer credit card options.

        It appears that the degree of a product’s look and feel nature becomes statistically less

important in choosing to offer credit cards, compared with the actual use of credit cards. In the

actual transaction, the effect on credit card usage is almost significant at 10% level. Perhaps,

during the interaction with the buyer in the transaction, the seller senses the buyer’s concern

about product uncertainty and the high potential for dissatisfaction, which increases the chance

of a returned product and credit card chargeback. In such a transaction, the seller has a stronger

preference for a cash-equivalent.

        When combining both seller characteristics and product attributes, the results are not very

different, but the model fits better (the results are not reported in the table). In addition, for

Model 1, the likelihood-ratio test cannot reject the null hypothesis that all slope parameters are

zero, but strongly rejects the null hypothesis for Model 2. Thus, it indicates that product

attributes are more important than seller’s characteristics for credit card offerings.

        The effect of reputation ratings on credit card offerings is reported in Model 3 (Table 5).

It appears that a negative rating reduces the chance of a seller’s offering to accept credit cards,

and is significant at almost the 15% level. Statistically, this effect is much stronger than that

affecting actual credit card use. Perhaps, a seller with negative ratings is more self-protective

and thus prefers cash-equivalents, and the effect shows up statistically when there is no influence

from the buyer. A seller’s sale volume and years of selling experience both show a positive

effect on credit card offering, and is significant (or almost significant) at the 10% level. In

addition, seller education has a negative effect, significant at the 5% level.

6. Payment Choice Analysis with Nested Logit Models

       The probit model provides a convenient tool to investigate the effects of product

attributes and seller’s characteristics on the choice of payment method. Moreover, it can also

investigate their effects on a seller’s offerings of payment methods. Thus, we can gain some

insight on the negotiation between the seller and the buyer regarding the payment transaction.

Yet, there are some limitations for the binary probit model. In particular, within one category,

one payment method is chosen more often than others, and the probit model cannot explain the

difference. For example, in the sample, traders use the PayPal credit card service more

frequently than using eBay credit card service; and among cash equivalents, money orders are

used most of the time. In addition, in the probit framework, we cannot estimate the effects of

payment-specific attributes on payment choice, because these variables perfectly predict the

choice; for example, if the seller’s fee is zero, the choice must be cash equivalents.

       Therefore, a natural extension is to apply discrete choice analysis beyond the binary

choice model. Such a framework will allow us to investigate the choice probabilities for each

payment option, and moreover, to evaluate how the cost and inconvenience affects such choices.

One option would be to use multinomial logit model (or conditional logit model). However, this

model requires that the relative probabilities between choices must be independent of attributes

or existences of other alternatives (the IIA property).30 The similarity among different types of

credit card services and among cash equivalents makes the IIA assumption unlikely. Therefore,

we adopt a more general approach by using nested logit model (McFadden 1984).31

        In a nested Logit model, similar choices are put into different groups (nests), and

correlated errors among choices within a group is allowed. The model process is to choose a

group (nest) first, then choose a particular alternative within the group. For alternatives in the

same nest, the IIA property is still required; while for alternatives in different nests, IIA property

in general is not required (Train, 2000). The resulting probability of choosing an option j is the

product of the marginal probability of choosing the group that contains the choice j and the

probability of choosing j conditional on choosing that group. In particular,

                  p( y = j ) = p( y = j | y ∈ B( j )) ⋅ p( y ∈ B( j ))

where B(j) is the nest to which alternative j belongs. The corresponding choice probability for

alternative j can be derived by assuming generalized extreme value (GEV) distribution. See,

among others, Train (2000) for the choice probability function. The nested logit model here

cannot be viewed as a random utility model because the payment choice is a joint decision

between the seller and the buyer.32 As discussed in the simple model (section 3), the actual

choice is determined by the latent variable—the weighted surplus among alternatives.

   The IIA assumption follows from the initial assumption that the disturbances are independent and homoskedastic
for the random utility model.
   Another alternative is multivariate Probit model. However, it has practical difficulty of computing the
multinormal integral and estimating an unrestricted correlation matrix.
   There are two different specifications for the choice probability for nested logit model in the literature due to
different normalizations. In some cases, one of these specifications is not consistent with the random utility model
(see Koppelman and Wen 1998, Hensher and Greene 2000, and Heiss 2002 for discussions). This particular
specification is adopted in software package Stata, which will be used for our estimation. Since our model is not
based on random utility model, such a specification does not cause a concern here. Note that both specifications are
equivalent in terms of the implied marginal effects and elasticities.

In our sample, to apply a nested logit model, we exclude the choice of seller processed

credit card (4 observations) because it depends on whether the seller has the facility to process

credit card directly, and it is not a choice available to every transaction.33 We also exclude those

who send cash directly (3 observations). It is very rare to send cash directly in online

transactions due to its irretrievable nature.34 The use of cash should be under very special

circumstances and the determinants are very different from those of choosing other payment

methods.35 After excluding those observations, the sample contains 253 observations.

         The corresponding tree structure for the nested logit model is specified in Figure 1. In

this structure, personal check enters as a degenerated nest, i.e., it has only one alternative. An

alternative structure would be to combine personal check with money order and cashier’s check.

However, it seems unlikely that the ratio of probability of choosing personal check to that of

choosing cashier’s check is independent of the existence of money order, so the IIA assumption

is violated in such a nest. Thus a separate nest for personal check is more appropriate.

         In the nested logit model, three new alternative-specific variables are added. They are:

buyer’s fee, seller’s fee and payment time. These variables can directly evaluate the effect of

cost and convenience on payment choice. Buyer’s fee is specified as 0 for credit cards and

personal check, and $6 for cashier’s check, and fees for money order are listed in Table 1.36

Seller’s fee is 0 for money order, cashier’s check and personal check, and credit card transaction

fees are listed in Table 1.37 Finally, payment time is defined as the time between that the buyer

sends the payment and that the seller receives the money and ships the product. This variable

   We have no information on whether every buyer has a credit card either, and thus assume this is the case.
   The 3 observations using cash have a price of $4, $6, and $9. This is at the lower end of the prices.
   One seller indicated to us that he used cash because the buyer was in the same town and he just dropped off the
   We do not have information on whether the seller passes the transaction fees on to the buyer.
   The credit card transaction fees are based on their standard rate. Both eBay and PayPal have complicated
incentive discount schedule for sellers based on their volumes.

partially measures the convenience of a payment method: the shorter the pay time, the quicker

the seller will get the payment and the buyer will get the product. Pay time is specified as 0 for

credit cards, 3 days for money order and cashier’s check (average shipping time), and 6 days for

personal check (average shipping time plus 3 days average clearing time).

         In a probit model, the alternative-specific variables perfectly predict the choice and thus

their effects cannot be identified. In our nested logit models, these variables are treated as

generic variables, i.e., each coefficient is the same for all alternatives. Such a structure helps to

save degrees of freedom. Because seller’s fee and buyer’s fee vary across choices and

observations, these two variables are used in the bottom level of the tree structure to explain the

choices of alternative payment methods. The pay time variable, however, does not vary within a

nest, thus is used at the first level to explain the choice of nest.

         Individual-specific variables include product attributes and seller’s characteristics. Since

these variables do not vary across choices, it is necessary for them to interact with choice dummy

or nest dummy. Since the number of parameters increases rapidly with such variables, only

product warranty and seller’s monthly volume are included in the nested logit model. Based on

the probit model, they appear to be significant in most cases. These two individual-specific

variables interact with nest dummies in the model (the second branch--cash branch is left out for

normalization), because they appear to be more relevant to the choice of nests and such a

specification saves degrees of freedom.

         The results of the nested logit model are reported in Table 6.38 Model 1 is a standard

specification for a nested logit model with four choice dummy variables. Because the variable

pay time is perfectly correlated with choice dummies, it is interacted with product price. The

 The nested logit model is estimated using Stata, which does not calculate the marginal effect for nested logit

model is highly significant based on the likelihood ratio test. Yet, most estimated parameters are

insignificant, indicating a high degree of multicollinearity. In fact, buyer’s fee and seller’s fee

are highly correlated with payment dummies.39 Based on the signs and magnitudes for the

payment dummy variables, it appears that PayPal is more attractive than eBay, and money order

and cashier’s check are almost indifferent. Buyer’s fee, seller’s fee, and pay time reduces the

probability of choosing the payment method.

           To reduce the degree of multicollinearity, Model 2 drops dummy variables for alternative

choices while adding a dummy for each nest. As expected, the value of the log likelihood

function decreases substantially, but the model is still highly significant. Buyer’s fee and seller’s

fee is very significant. The higher the transaction fee is for the seller or the buyer, the lower the

probability choosing a payment method. Therefore, the payment cost does have a negative effect

on payment choices. Similar to Model 1, seller’s fee shows a stronger effect than buyer’s fee in

terms of magnitude. This result indicates that seller more strongly resists a higher fee, and thus it

is likely that the seller has more power in deciding the payment method.

           On the other hand, pay time appears to have a negative effect on a payment method too.

As the pay time for a particular payment option grows longer, the probability of choosing it is

lowered. Clearly, traders prefer a more convenient way (less waiting time) to transfer money.

But it is only significant at the 25% level.

           Based on the coefficients for nest dummies, credit cards seem to be more attractive than

money orders and cashier’s checks. Since costs and convenience are partly controlled in the

model, the attractiveness of credit cards can be attributed to the protections they provide.

     Since the fees for money order changes with price, buyer’s fee is not perfect correlated with payment dummies.

Personal checks do not seem to be significantly different from money orders and cashier’s

checks. This is expected after controlling for the difference in cost and convenience.

         The coefficients for individual-specific variables are normalized as a difference from the

coefficient for the nest of money order and cashier’s check (cash nest). The sign shows that a

warranty increases the probability of choosing credit cards relative to cash group, and it is

significant at the 10% level. As discussed in the probit section, a warranty reduces product

uncertainties, and thus reduces the chance of returning the product by the buyer. Therefore, it

increases the expected profit for the seller for using credit cards. On the other hand, product

warranty does not show a statistically significant difference between choosing check and the

cash nest. Such a result is expected. The similar results can be found for seller’s sales volume,

but the magnitude is very small, indicating little difference among cash, credit cards, and

personal check in terms of the effect of sales volume.40

         Since the specification of interacting pay time and product price implicitly assumes that

the effect of pay time changes with product price, such a restriction may not be desirable.

Therefore, in Model 3, we use pay time directly in the model. To avoid perfect multicollinearity,

we also drop two nest dummies. The changes in other variables are small, but for pay time it is

substantial. The effect of pay time becomes highly significant in reducing the probability of

choosing a payment method. Yet, in this case, it is possible that the effect of pay time also

includes some other unobserved fixed effects of a payment method.

         The parameter of the inclusive value (IV) for each nest is also reported in the table.

Based on Heiss (2002), in a nested logit specification that is consistent with random utility

  When the measure “look and feel” nature of the product is added in the model, the results for other variables do
not change much. The effect of the “look and feel” nature is not statistically significant. The signs show that the
higher degree of uncertainty, the lower probability of choosing either credit cards. This is consistent with the probit
result that “look and feel” nature raises the chance of using cash equivalents.

model, the IV parameter measures the dissimilarity in each nest, and thus should be within a

[0,1] interval. In our results (based on Stata command), the IV parameters measure both

dissimilarity and the relative importance of the generic variables in the respective nest, thus the

interpretation is more difficult.

        Finally, using the sub-sample, we include seller’s reputation measure in nested logit

models. As in the probit model, reputation does not seem to have any statistically significant

effect on the choice of payment method. The magnitude is also much smaller than that for

product warranty.

7. Conclusions

        Due to physical separation between buyers and sellers, payment method is especially

crucial in online transactions for electronic commerce in general and for online auctions in

particular. Using the survey data collected from eBay users, we apply probit and nested logit

model to investigate the choice of payment methods in online auction markets.

        The main findings of this study can be summarized as follows. First, the cost of a

payment method discourages its use in online transactions. The transaction fees for either buyers

or sellers reduce the probability of choosing the payment method. It appears that sellers have

stronger resistance to transaction fees. Second, the inconvenience associated with a payment

method, as measured by pay time, also reduces the chance of using this payment method.

        Moreover, product attributes appear to have stronger effects on payment choices than

traders' characteristics. In general, if the uncertainties surrounding a product can be reduced, the

probability of using credit cards will increase, otherwise cash-equivalents will be more likely to

You can also read
NEXT SLIDES ... Cancel