1

Switching Cost and Brand Loyalty in Electronic Markets:

Evidence from On-Line Retail Brokers

Pei-Yu (Sharon) Chen University of Pennsylvania, Wharton School

1300 Steinberg Hall-Dietrich Hall Philadelphia, PA 19104

pychen@wharton.upenn.edu Phone: 215-898-8047 Fax: 215-898-3664

Lorin M. Hitt University of Pennsylvania, Wharton School

1318 Steinberg Hall-Dietrich Hall Philadelphia, PA 19104

lhitt@wharton.upenn.edu Phone: 215-898-7730 Fax: 215-898-3664

We would like to thank Robert Josefek, Eli Snir, Arun Sundararajan, and seminar participants at New York University, the University of Texas at Austin, the University of Pennsylvania, the Workshop on Information Systems and Economics, the International Conference on Information Systems (ICIS) and three anonymous reviewers from ICIS for comments on earlier drafts of this paper. We would also like to thank Media Metrix and Gomez Advisors for providing essential data. This research was funded by the Wharton eBusiness Initiative (WeBI).

Keywords: Switching cost; Brand loyalty; Electronic markets; Electronic commerce; eBusiness; Measures;

Empirical research; on-line retail broker.

2

Switching Cost and Brand Loyalty in Electronic Markets:

Evidence from On-Line Retail Brokers

Abstract

The ability to retain and lock-in customers in the face of competition is a major concern for e-commerce businesses, especially those that invest heavily in advertising and customer acquisition. In this paper, we propose a method for measuring the magnitude of switching costs and brand loyalty for on-line service providers based on the random utility modeling framework. Applying this approach to the on-line brokerage industry, we find significant variation (as much as a factor of 2) in measured switching costs. We find that customer demographic characteristics have little effect on switching, but that customer behavior (such as usage, changes in usage, and adoption of multiple brokers) are better predictors of switching behavior. We also find that firm characteristics such as product line breadth and quality reduce switching and may also reduce customer attrition. Overall, we conclude that on-line brokerage firms appear to have different abilities in retaining customers and have considerable control over their switching costs through product and service design and various unobserved retention strategies.

3

1. Introduction

Many emerging e-commerce companies, especially those focused on business to consumer

(B2C) e-commerce, are in an aggressive phase of recruiting new customers in what analysts

have called a “land grab”. These firms devote a large amount of their resources to advertising

and promotion, and increasingly to outright customer subsidies. For example, E*trade was

offering $400 in free computer merchandise for new customers who signed up between

January and March, 2000. E*trade also spent about $400 million in 1999 on selling and

marketing, representing over 60% of their non-interest expenses and over 45% of net revenue.

Customer acquisition costs, which are estimated to range from about $40 per customer for

Amazon.com to over $400 for some on-line brokers (McVey, 2000), are probably the largest

contributors of costs to new B2C startups and represent a substantial portion of the initial

financial losses these firms typically incur. Clearly, the expectation is that these early

investments in customer acquisition will result in a long-term stream of profits from loyal

customers to offset these costs.

Essential to this strategy is that customers experience some form of “lock-in” or switching

costs to prevent them from defecting to another provider. These switching costs arise from a

variety of factors, including the general nature of the product, the characteristics of customers

that firms attract, or deliberate strategies and investments by product and service providers.

By creating or exploiting switching costs, firms can soften price competition, build a “first

mover” advantage, and earn supranormal profits on advertising or other investments

(Schmalensee, 1974, 1982, 1986; Klemperer, 1987, 1989,1995; Beggs and Klemperer, 1992;

Farrell and Shapiro, 1988; Nilssen, 1992). The ability to create switching costs and build

customer loyalty has also been argued to be a major driver of success in e-commerce

businesses (Reinchheld and Schefter, 2000; Reinchheld et al. 2000). However, if switching

costs are inherently low and firms are unable to lock-in customers, long-term profitability

may be difficult to attain, especially in many B2C e-commerce environments with low entry

barriers (other than customer acquisition costs) and limited differentiation.

4

Despite the critical role of switching costs in e-commerce strategy, there is surprisingly little

empirical evidence about the presence, magnitude or impact of switching costs on customer

behavior. This appears to be true more broadly – despite a robust theory literature, there are

only a limited number of empirical analyses on the measurement of switching costs (Zephirin,

1994; Handelsman and Munson, 1989; Elzinga and Mills, 1998) and even fewer that consider

how firms might influence their customers’ switching costs. A few studies in the information

systems and e-commerce literature have looked at related questions, such as price premia for

branded retailers (Brynjolfsson and Smith, 2000), the relationship between visit frequency and

website experience (Moe and Fader, 2000), customers’ propensities to search (Johnson et al.,

2000) and the relationship between customer satisfaction and loyalty in on-line and offline

environments (Shankar et al., 2000). However, these studies only investigate some aspects of

switching or brand loyalty, typically do not provide explicit measures of switching costs or

examine the specific sources of switching cost variation.

In this paper we make three specific contributions. First, we develop a model that allows

differences in switching costs across providers to be estimated utilizing web site traffic data,

based on the well-known random utility/discrete choice modeling framework (McFadden,

1974a).1 Second, we apply this model to study the on-line brokerage industry – a large and

important on-line industry where switching cost and customer acquisition are critical parts of

firm strategy and performance. Finally, we examine customer and firm-specific

characteristics that affect switching as well as adoption behavior and attrition. This enables us

to study how firm practices affect switching as well as to understand any secondary effects

that these practices might have on attracting and retaining new customers.

Using “clickstream” data on over 2000 individuals who utilize the 11 largest on-line broker

sites provided by Media Metrix, we find that there is substantial heterogeneity in switching

costs across providers, and that this variation is robust over time and after correcting for

measurement biases and heterogeneity in customer characteristics. Moreover, we also show

that firm characteristics, including those observed by a third-party consumer rating agency

1 This model is applicable to any setting in which a customer’s relationship with multiple service providers can be precisely observed. However, these data are typically difficult to obtain in the offline world because few datasets exist that can comprehensively capture customer interactions with multiple, competing businesses.

5

(Gomez Advisors), as well as customer behavioral characteristics, affect switching, customer

acquisition and attrition.

2. Literature Review and Background

2.1 Brand Loyalty and Switching Costs

In many markets, consumers face non-negligible costs of switching among different brands of

products or services. As classified by Klemperer (1987), there are at least three types of

switching costs: transaction costs, learning costs, and artificial or contractual costs.

Transaction costs are costs that occurred to start a new relationship with a provider and

sometimes also include the costs necessary to terminate an existing relationship. Learning

costs represent the effort required by customers to reach the same level of comfort or facility

with a new product as they had for an old product. Artificial switching costs are created by

deliberate actions of firms: frequent flyer programs, repeat-purchase discounts, and

“clickthrough” rewards are all examples. Besides these explicit costs, there are also implicit

switching costs associated with decision biases (e.g., the “Status Quo Bias”) and risk aversion,

especially when customers are uncertain about the quality of other products or brands.

Switching costs have been theoretically shown to influence prices, profits, and entry decisions,

and have been linked to a variety of competitive phenomena such as price wars and subsidies

for new customers (Klemperer, 1987, 1989, 1995; Schmalensee, 1974, 1986; Beggs and

Klemperer, 1992; Farrell and Shapiro, 1988; Nilssen, 1992).

The marketing literature identifies a specific manifestation of switching cost, termed “brand

loyalty”. Brand loyalty is usually defined as the minimum price differential needed before

consumers who prefer one brand switch to some competing brand (Pessemier, 1959; Raju,

Srinivasan and Lal, 1990). Brand loyalty may be due to real switching costs, decision biases,

uncertainty about the quality of other brands, or other psychological issues that drive

customers’ seemingly “unreasonable” behaviors. Much of this extensive literature

emphasizes the identification of loyal customers (Jacoby and Chestnet, 1978) by individual

behaviors such as repeat purchases or expressed preferences in surveys or focus groups.

6

Typically, loyalty is treated as an intrinsic characteristic of consumers (Jacoby and Chestnet,

1978; Grover and Srinivasan, 1987; Colombo and Morrison, 1989; McCarthy et. al., 1992).

Less emphasis has been placed on explicit measurement of switching costs due to firm

practices.

2.2 Brand Loyalty and Switching Costs in Electronic Markets

With the advent of the Internet, it has been argued that a competing firm is “just a click away”

(Friedman, 1999), and thus consumers face minimal barriers to switching product or service

providers. Nonetheless, recent research suggests that brand loyalty still exists in on-line

channels. Brynjolfsson and Smith (2000) found, using data from a price comparison service

(the DealTime “shopbot”), that customers were willing to pay premium prices for books from

the retailers they had dealt with previously. Johnson, et. al. (2000) showed that 70% of the

CD and book shoppers were loyal to just one site, and consumers tended to search fewer sites

as they become more experienced with on-line shopping over time. One possible explanation

for these findings is that firms have found ways to retain customers in the on-line channel that

introduce new “frictions” where old ones, such as difficulty in searching and making

comparisons, have been removed. Examples include frequent-purchaser programs, use of

user profiles for personalization, “clickthrough” rewards, and affiliate programs (Varian, 1999;

Smith et al. 1999; Bakos, 2001). Alternatively, on-line firms may try to attract and retain

their customers through product/services offerings and website design, for example, by

providing a compelling on-line environment (Hoffman and Novak, 2000; Novak et al., 2000)

and building customer trust (Hoffman et al., 1999) or by simply replicating “old economy”

strategies for customer retention by investing in quality, convenience or service to enhance

customers’ satisfaction and thus overcome the ease of switching providers. Irrespective of the

source, the benefits (or absence of benefits) from these strategies are difficult to evaluate

without a consistent framework for the measurement of switching costs that is compatible

with available data.

2.3 The On-line Brokerage Industry

7

Retail brokers provide individual investors with the ability to purchase and trade stocks,

bonds and other financial instruments. On-line brokers differ from their traditional

counterparts in the discount brokerage segment by conducting the vast majority of their

transactional activity using the Internet.

This industry is interesting to study for a number of reasons. First, the market is large and

significant and is considered to be one of the “killer applications” in B2C electronic

commerce (Varian, 1998; Bakos et al., 2000). There were over 140 on-line retail brokers by

the end of 1999 and they managed just over $1 trillion in customer assets in 2000. By year-

end 1999, these accounts represented about 15% of all brokerage assets and 30% of all retail

stock trades (Saloman Smith Barney, 2000). Second, this industry is known to have very

aggressive customer acquisition tactics and requires large advertising expenditures – for

example, E*trade spent over $154 million and Schwab spent $101 million on advertising in

the first quarter of 2000 alone (McVey, 2000). Third, there is reason to believe that switching

costs have a substantial influence on customer behavior. Executing a stock trade is a

potentially significant financial transaction, and there is some complexity in learning to use

the tools and interface of a new broker. This deters customers from switching brokers. In

addition, switching brokers either requires that customers convert their entire portfolios into

cash (incurring significant trading costs and possibly tax liabilities) or transfer the assets

individually, which requires a waiting period of up to six weeks for completion. Finally, the

lifetime value of an active account is large (>$1,000) and customers of on-line brokers are

found to trade much more than customers of traditional full-service brokers (Varian, 1998),

which implies that firms may be willing to make significant investments in customer retention.

2.4 Hypotheses

For our purposes, we define switching costs as any perceived disutility a customer would

experience from switching service providers – including both explicit (e.g, fees, time and

effort) and implicit (e.g., quality uncertainty) costs. From prior research, we can view

switching costs as influenced by the nature of the product, consumer characteristics, and firm

practices. However, to the extent that the fundamental characteristics of the underlying

8

product are governed by regulation and tend to be similar across firms, there should be little

variation in switching cost created by fundamental product characteristics. That leaves

customer characteristics and firm practices as the primary drivers of switching cost variation,

if any, in this study. Our starting null hypothesis is thus to test whether there is variation in

measured switching costs across firms:

H1: There is no variation in measured switching cost across firms

To the extent that this hypothesis can be rejected and switching costs vary, our analysis will

focus on distinguishing the role of firm and customer effects because they are associated with

different observable variables and have different strategic implications. If switching behavior

is driven solely by customer characteristics, the challenge for firms is to target and prescreen

customers who are more likely to be loyal. If it is solely due to firm practices, the challenge is

to design their service offerings and products such that they either attract loyal customers or

lock-in customers once they are acquired.

We first consider customer characteristics. For discussion purposes, it is useful to divide

customer characteristics into intrinsic characteristics (which are unchanging) and observed

behaviors (which can change over time). In terms of the former, if we view switching as

resulting from “dissatisfaction,” then we would expect little effect of unchanging

demographic characteristics. If customers are well informed about the match between their

preferences and the offerings of various service providers when they make their initial

choices, they are unlikely to switch due to factors that are unchanging. However, we might

expect greater switching when usage patterns change, since the characteristics of their existing

provider may not be optimal given their revised usage pattern. If, instead, switching is a

result of an intrinsic “loyalty characteristic” this could directly affect switching. In our data

these effects can only be measured to the extent that it is correlated with other observables

(like demographics or behavior). This suggests that demographics overall may have an effect,

but does not provide any particular guidance to which characteristics will matter or even the

signs of the effects. However, we do have one behavioral proxy for intrinsic loyalty – the

number of service providers that the customer adopts. Overall, to the extent that these

9

characteristics and behaviors capture a significant share of the relevant characteristics of

customers (or are correlated with factors we do not measure directly) and that the “type” of

customer varies across brokers, we can investigate the extent to which variation in switching

cost is due to variation in customers. Thus, we generalize our earlier hypothesis to consider

customer-specific factors:

H2a: There is no variation in switching costs after controlling for customer characteristics

We can also directly examine the effects of customer characteristics on switching

(summarizing the discussion above):

H2b: Customers who use multiple brokers or change their usage patterns are more likely to switch. Unchanging demographics (age, income, gender, race) may impact switching in an unspecified way.

If we are still able to reject H2a, it suggests that switching is also driven by firm-specific

practices.2 We can also directly test the effects of broker attributes on switching through

examining the propensity to switch as a function of the broker characteristics.

Here we are forced to consider the practical issues of what firm characteristics can be

measured, particularly over time. Our starting point will be characteristics of brokers tracked

by Gomez Advisors, one of the leading on-line consumer rating services. Gomez tracks firm-

level characteristics in five dimensions: cost, consumer confidence (related to an abstract

notion of “quality”), on-line resources (breadth of offerings), relationship services (equivalent

to a degree of personalization), and ease of use. These factors are broadly representative of

the factors used by other consumer rating services, but have the advantage that they are

defined by analysts rather than consumers (thus removing possible biases of customer

heterogeneity), measured consistently over time, and are at a finer level of granularity than

other rating services that principally provide an “overall satisfaction” score. The Appendix

lists the definitions of these factors as publicly described by Gomez (no data were available 2 One could argue our rejection of H2a was due to an incomplete list of customer characteristics. However, even in this case, the test requires that these customers with different characteristics be distributed unevenly across firms. Thus, it could still be considered, in part, a firm-level effect.

10

on the subcomponents of these scores that they use internally). We also include a measure of

the required initial investment to establish an account since it may be directly relevant for

screening customers and is easily measured. We can also examine the effects of other

“unobserved” broker characteristics on switching through broker dummies.

Prior literature and reasoning provide directional predictions for many of these variables.

Firms can invest in relationship services and personalization technologies that allow users to

customize their user interfaces, potentially leading to greater lock-in (Crosby and Stephens,

1987; Pearson, 1998; Mobasher et al., 2000; Cingil et al., 2000). We would therefore expect

that “relationship services” would have positive effects on switching (that is, they reduce

switching). Service quality has been linked to higher customer retention both theoretically and

empirically (Boulding et al., 1993; Gans, 2000a,b) suggesting a positive link. In our analysis,

this is proxied by Gomez’s “consumer confidence” measure that likely captures a wide variety

of unspecified quality attributes. Breadth of offering may also be important for

accommodating fluctuations in user preferences, making it less likely that a customer will

switch to obtain access to other complementary products or services. Cost has no clear

predictions -- as long as customers are well informed about prices (which seems reasonable to

assume) they will make the “right” adoption choices on average and thus are unlikely to

switch as a result of the level of cost, especially since prices have been fairly stable over our

sample period. Similarly, minimum balance is not likely to have a direct effect since it is no

longer a constraint once the customer has adopted a particular broker, although it may be

interesting to see if there are secondary effects screening customers by minimum balance.

“Ease of use” may actually be associated with higher switching as it does not require users to

invest in learning an interface (this is the flip side of the argument on retention made by

Johnson et. al., 2000). Summarizing the discussion above, we expect:

H3: Switching is reduced by personalization, quality and breadth of offerings. Ease of use may increase switching, and cost and account minimums have no effect on switching behavior controlling for customer characteristics.

11

In addition to the focus on switching, it may also be useful to consider the closely related

issue of customer attrition (where a customer ceases to use all service providers entirely). We

will employ the same techniques used to test H3 to the examination of this factor.

The predictions on attrition largely parallel those of switching. By using the same arguments

as for switching, we would generally expect that personalization, quality, and relationship

services reduce attrition, and cost should have little effect. There are two critical differences:

ease of use may reduce attrition since customers may be more likely to depart because the

interface is too difficult to use; and minimum balances act as a screen against customers who

intend only to collect new user subsidies and, thus, should reduce attrition. As before, we

have no strong predictions for demographics, although we would expect higher volume users

and those with multiple brokers to be less likely to disappear, suggesting that some behavioral

characteristics will matter. We therefore expect:

H4: Customer attrition will be reduced by personalization, quality, relationship services and account minimums; cost should have little effect. Usage and number of brokers adopted will be associated with reduced attrition.

3. Model

Our model exploits a natural experiment that is present in most (if not all) on-line

environments. At any point in time, we observe new customers entering the market and

making service provider choices, and we also observe existing customers making choices on

whether to continue their relationship with their existing provider or adopt a different provider.

Normally it is difficult to disentangle switching costs from the differences in benefits of

different service providers, especially since benefits might vary over time. However, if we

use adoption behavior of new customers as a surrogate of overall utility of the various

providers, and are willing to assume that preferences (on average) of new customers and

existing customers are the same, we can separate out switching cost effects from “quality”

differences.

12

Under these conditions, we can estimate the extent of switching costs by comparing the rate

of adoption of different service providers for new customers to the rate of switching faced by

each provider. In the extreme case, if there were no switching costs at all, then in each period

customers could reconsider their product choices at no cost and would migrate to new

providers in the same proportion as new adopters (assuming that their preferences as a group

were equivalent). If, however, customer defections have a different distribution across

different providers than new customer adoption, this is indicative of variations in switching

costs.

The mathematical structure for the analysis is provided by the random utility modeling

framework (McFadden, 1974a), which has been extensively applied in studying consumer

choice behavior among multiple products (McFadden, 1974b; Schmidt and Strauss, 1975a,b;

Boskin, 1974; Guadagni and Little, 1983; Brynjolfsson and Smith, 2000). Consider a set of

consumers who have preferences over a set of goods that are comprised of two parts: a

systematic component related to the observable and unobservable characteristics of the good

(ν), and a random component which is idiosyncratic to an individual customer and arises due

to specific tastes or random error in selection (ε).

We further subdivide the systematic component ν into a price index (r), a vector of non-price

attributes (x), and a collection of (possibly) unobservable firm-specific effects (γ). For a set

of N consumers choosing among M firms, we write the utility of a particular consumer if she

chooses firm j ( [1,2,... ]j M∈ ) as:

j j j j j j ju v x rε γ β α ε= + = + − + (1)

If we observe a customer choosing firm j, we can infer that this choice provides the consumer

with the highest utility over the set of M firms. That is, the probability that a consumer will

choose firm j is determined by the relative utility level:

p prob u u kj j k= ≥ ∀.( , ) (2)

13

In the original theoretical work on this model (McFadden, 1974a) and a large body of

empirical work that followed, it was typical to assume that the error term is independent and

identically distributed with Weibull distribution, also known as "extreme value" distribution

(that is, prob e whereje.( ) ,ε ε ε

ε≤ = − ∞ < < ∞− −

). This type of error structure arises when

choice behavior is governed by independence of irrelevant alternatives (IIA) – that is, the

relative choice probability of any two products does not depend on the presence or absence of

an alternative choice. In markets where all alternatives are available to all customers at all

times (a reasonable approximation of the on-line brokerage market), the IIA assumption is

more likely to hold. Under IIA, we have a simple expression for the choice probabilities (or

the fraction of consumers that choose provider j):

1 1

j j j j

l l l l

v x r

j M Mv x r

l l

e epe e

γ β α

γ β α

+ −

+ −

= =

= =∑ ∑

(3)

This expression is known as the multinomial logit model (MNL). While we will use this

model in our empirical analysis, it is not necessary for our general switching cost framework.

In other work (Chen and Hitt, 2000b) we show that a similar model can be utilized based on

the same concept where consumer preferences can be expressed by the more flexible

“generalized extreme value” model (McFadden, 1978, 1980, 1981) -- however the derivations

are complex and do not add significantly to this study since we find that the required

assumptions for the multinomial logit model hold in our data.

We now introduce switching costs in this framework. Let sk be the cost (disutility) a

consumer must overcome when switching from firm k ( [1, 2,... ]k M∈ ) to another firm. Note

that we have implicitly assumed that the switching cost does not depend on the firm the

customer switches to, but only on the firm she switches from.3

Consider a two-period setting where some consumers choose a firm in period 1, while others

do not. In period 2, some of the early adopters stay, some of the early adopters switch firms,

3 Based on this assumption, we only have to estimate M parameters rather than Mx(M-1). As a result, this is a common assumption in literature. This assumption is also testable, and shown to be satisfied by our data.

14

and new customers who had not previously adopted enter the market. For a consumer who

chose firm k in period 1, her utility from choosing to stay with firm k in period 2 is:

|k k k k k k k ku v x rε γ β α ε= + = + − + (4)

where the notation ua|b denotes the utility a customer gets if she chooses firm b in period 1 and

then switches to firm a in period 2, and note that kε is different for each customer.

However, if she decides to leave firm k and switch to another (say, firm j, [1,2,... ]j M∈ ), she

incurs a switching cost ks , and as a result, the utility of choosing j is:

|j k j k j j j j k ju v s x r sε γ β α ε= − + = + − − + (5)

If the user is new in period 2, then there is no switching cost, so the utility of choosing any j is

(where we denote the subscript “n” as representing new users):

|j n j j j j j ju v x rε γ β α ε= + = + − + (6)

Again, we have implicitly assumed that the preferences for new customers over price and

other attributes is the same for new customers as for existing customers on average, except for

switching costs and the customer-specific utility (or unobserved taste distribution) ε.

However, customers who switch differ from new adopters in that they incur the disutility from

switching as represented by ks . Given the utility expressions derived in equations (4)-(6) and

the choice probability expression (3), the probability that a user from firm k will choose to

stay with firm k is:

|

k

k m k

v

k k v v sm k

ePe e −

≠

=+∑

(7)

It is useful to rewrite this in the form of an “odds ratio”:

|

|1

k k k

m k m

v s vk k

v s vk k

m k m k

p e e ep e e−

≠ ≠

= =− ∑ ∑

(8)

Similarly the equivalent odds ratio for a new adopter can be found based on equation (6) or by

setting the switching cost parameters to zero in equation (8). Thus the odds ratio of choosing

firm k as a new user is given by:

15

|

|1

k

m

vk n

vk n

m k

p ep e

≠

=− ∑

(9)

Comparing equation (8) and (9), firm k’s level of switching cost can be calculated

immediately by dividing the two odds ratios:

|

|

|

|

1

1

k

k k

k k s

k n

k n

pp

epp

−=

−

(10)

These values of ks can be directly measured using the ratios described in equation (10) using

only market share data (the share of new customers, and the fraction of adopters who remain

at the same firm). If we have disaggregated individual choice data rather than simply market

data, we can also obtain switching costs estimates and standard errors of the estimates by

estimating the choice system given by equations (4)-(6) directly.4 Our analysis will first be

conducted with the ratio approach to demonstrate its characteristics – we will then shift over

to regression analysis to facilitate hypothesis testing.

The relationship implied by equations (4)-(6) (assuming an MNL choice formulation) is

known as the “conditional logit model”. We use the standard formulation of this model that

include individual (i) characteristics (zi) as well as firm (j) attributes (xj and rj), plus an

additional set of dummy variables to capture switching costs. The underlying model we

estimate is:

[1,2.. ], [1,2.. ], {[1, 2.. ] }ij j j j j i k k iju x r z s W i N k M l M jγ β α λ ε= + − + − + ∀ ∈ ∀ ∈ ∈ ≠ (11)

Several notes about this formulation are in order. First, this model is typically estimated by

simultaneously estimating individual logistic choice equations for each firm for each

customer– this yields a total of M firms x N individuals or MxN data points. In the discussion

that follows we will sometimes refer to one of the individual firms’ equations. Second, the 4 Of course, one could also generate standard errors for the ratio analysis by bootstrapping, but this has the same data requirements as the regression analysis and is much less flexible for hypothesis testing. When individual choice data are not available, the only other approach is to utilize repeated observations over time to generate standard errors, although this risks confounding changes in switching costs with time series variation in utility of various brokers.

16

model includes a dummy variable for each firm ( jγ ), a vector of quality and price parameters

( , )β α common across firms and individuals as before, and a vector of customer characteristic

parameters for each firm (the set of vectors jλ ). This treatment of individual attributes is the

standard formulation for conditional logit models (Green, 1997, p. 914) to accommodate the

fact that because individual characteristics are constant across all alternatives they are not

otherwise identifiable unless interacted with a variable that varies by firm. Finally, our only

deviation from the standard model is the inclusion of a vector of dummy variables, one

element per firm, kW , which takes on the value of 1 if customer i is a potential switcher from

firm k and zero otherwise. In other words, the dummy variable is 1 whenever a customer

adopting a particular firm would face a switching cost if she chose to adopt another.5 The

estimated values of the parameters on this set of dummy variables ( kW ) are the switching cost

parameters ( ks ) for each firm. All our hypothesis tests will be performed using parameter

restriction tests (Wald tests using an asymptotic covariance estimate) in the conventional way.

4. Data and Analysis

4.1 Data: Site Usage

Our primary data for this study are drawn from a panel of “clickstream” data provided by

Media Metrix. Media Metrix has a panel of more than 25,000 households that have an applet

installed in their computers which tracks the user, time and URL of every page request they

make on the world-wide web. They also collect demographic information from the users

(gender, household income, age, education level, occupation, race and others). This enables

us to use these data for individual-level control variables, and also enables Media Metrix to

ensure that their panel is demographically consistent over time and representative of the US

Internet-using population. Our analysis is focused on four consecutive quarters of data from

July, 1999 to June, 2000 which we will label Q399, Q499, Q100 and Q200. We restrict our

analysis to customers who are tracked by Media Metrix in all four quarters so that we can get

5 A numerically identical approach would be to set the dummy corresponding to firm, say, k to 1 if customer i stays with firm k. This simply reverses the sign of the coefficient.

17

proper estimates of the number of first period non-adopters and track customers’ flow during

this time frame.

Using analyst reports (Salomon Smith Barney and Morgan Stanley Dean Witter) we identified

the 11 largest retail brokers,6 which account over 95% of all on-line brokerage accounts, and

extracted all page references to these sites. We use the number of days that a broker is

accessed and total time spent in a quarter as a proxy for activity at the broker. We restrict our

analysis to individuals who are registered account holders at these brokers – individuals who

browse broker sites, but do not have accounts, are excluded. To determine whether a customer

is an account-holder or not, we examine the individual URLs that each customer visited – if

she accessed any pages that were restricted to account-holders for at least 5 seconds7 during

the period, we defined the customer as an account holder. We corroborated our estimates of

overall market share with other sources (Salomon Smith Barney and Morgan Stanley Dean

Witter) and found them to be remarkably consistent with a Pearson correlation over 90% and

98% rank order correlation.

There are two key limitations of these data. First, while we can tell whether the customer is

an account holder, we cannot determine their trading volume since, in general, we cannot

identify whether a page view corresponds to a trade. However, previous studies have

suggested a positive relationship between visiting frequency and purchase propensity (Roy,

1994; Moe and Fader, 2000). This suggests that people who visit a broker more frequently

may be more likely to trade.

A second issue is that our data covers home usage but not work usage. Given that significant

trading activity in many accounts occurs during the daytime when the financial markets are

open, our visit frequencies may not be indicative of trading activity. However, as long as there

is positive connection between visiting frequency and trading propensity, which is very likely

to be true, then visiting frequency still contains valuable information. More importantly, to the 6 These brokers are Ameritrade, Datek, DLJDirect, E*Trade, Fidelity, Fleet (which owns QRonline and Suretrade), Morgan Stanley Dean Witter Online, Schwab, TDWaterhouse, Vanguard, and National Discounted Brokerage. 7 We identified over 2000 unique URLs on the sites that we classified. The 5 second limit was utilized to catch non-customers who reached a restricted page and were automatically redirected.

18

extent that most users utilize these sites for both trading (during market hours) and research

and financial management in the off-hours, we are not likely to be missing the overall

adoption decision. There are also possible errors introduced by the presence of financial

services aggregators (e.g., Yodlee.com) that enable customers to manage their accounts

without visiting the sites, but these services are used by less than 1% of the customers in our

analysis. We address the general problem of missing some customer usage of these sites by

aggregating our data to calendar quarters – this way, if a customer accesses these sites during

the quarter, we will properly capture her broker choices.

In our sample, 80% of the customers have only one broker at a point in time. For the

remaining 20%, we define a “major broker” for each customer, who is the broker for which

the customer visits the brokers account holders’ pages most often. Our switching analysis

therefore focuses on customers who change their major broker. We chose this strategy for

several reasons, the most important of which is that it allows us to accommodate users with

multiple brokers – a natural feature of this market. Second, it enables us to compare the

switching behavior of multiple broker users to other customers, since we would generally

believe that these customers face lower switching costs. Finally, our results do not appear to

be sensitive to this assumption, as similar switching cost estimates were found in earlier work

that tracks all accounts (Chen and Hitt, 2000a).

4.2 Data: Broker Characteristics

We also utilize additional data to determine the attributes of the sites we study from Gomez

Advisors, an on-line market research firm. Gomez utilizes consumer surveys and a research

staff to catalog and evaluate the quality and characteristics of various e-commerce sites,

including retail brokers. These data were described earlier in Section 2.4. We have

corresponding data from Gomez for each quarter considered in our study. In addition, we

include minimum initial balance required for each broker, which we obtained from each

broker’s registration page.

4.3 Summary Statistics and Preliminary Analysis

19

Among our four datasets from Media Metrix, we have a total of 28,807 households, of which

11,397 households are tracked throughout the period of interest (this includes both web users

who use on-line brokers and web users who do not). Restricting our sample to only

individuals who appear in all datasets, we have 1,249 broker users in Q399, 1,393 in Q499,

1,780 in Q100 and 1,586 in Q400. Overall, we have 2,902 unique broker users, including

1,653 new adopters during the year of interest. Figure 1 shows the movement of customers

among different categories between two consecutive quarters.

Tables 1a and 1b summarize market share in each period and market share for new adopters.

The Table shows that market shares are relatively stable with the exception of Ameritrade and

Datek who appear to be growing in share over time. As mentioned earlier, these market share

numbers are consistent with estimates obtained from other sources (Salomon Smith Barney,

2000; and Morgan Stanley Dean Witter, 1999).

Among the 2,902 unique users (from 2,257 households) we examined, 303 of them changed

their major broker during the year we tracked. Table 2 shows user flow by broker. For

example, among all E*Trade users, 55.7% of them remained active and stayed with E*Trade,

10.6% of them switched out while 33.7% of them became inactive. Note that we define

inactive as not returning to a broker at any time in the future through the end of our data

period (as opposed to having no access during an intermediate period and then returning in a

later period). The variation on switching and attrition rates described in the Table is easily

seen in Figure 2. Schwab and Datek have a higher retention rate than DLJDirect, E*Trade,

MSDW and Vanguard. For the top three brokers, E*Trade has the most serious problem of

customer departure. Their switching rate is more than 1.5 times that of Fidelity and Schwab,

and attrition rate is the highest across all brokers we examine. These differences in flow rates

are both economically and statistically significant ( 2χ (20)=69.32, p<0.001). Moreover

retention rates, switching rates and attrition rates are all economically and statistically

different (p<0.001).

4.4 Variation in Switching Costs

20

The switching cost framework we introduce in Section 3 provides a method of calculating

switching costs that does not rely on a specific functional form – only that the assumptions of

MNL random utility models hold and that new users and existing users as a group8 have the

same preference structure absent the influence of switching costs. Because the units of the

switching cost measure are ambiguous (partly due to the scaling of variables used in the

analysis9, and partly due to assumptions embedded in the MNL formulation – see Chen and

Hitt, 2000b for further discussion of the latter issue), we treat these estimates as relative

values. Because this is a differentiated market and all brokers we examine are available to all

users throughout the study period, we expect the independence assumption to hold making the

multinomial logit the appropriate formulation. Using the regression-based analysis (discussed

later), we test the independence (IIA) assumption and find that it is satisfied in our data10

(p>0.05).

The calculated switching cost values based on equation (10) for each pair of quarters are

presented in Table 3 (Columns 1-3). Note that the switching cost measures for each pair of

quarters appear to be very stable over time. The Pearson correlation is over 90% and rank

order correlation is over 75% for each pair. The data in the Table are also depicted

graphically in Figure 3, which highlights the similarity of the measures over time. This

suggests that we can pool our data over time, increasing statistical power. The switching cost

estimates from the pooled analysis are reported on Table 3, column 4.

This measured switching cost is the disutility perceived by a representative user of each

broker regardless of source (customer or firm characteristics). Based on these estimates, we

know that NDB and Datek have the “stickiest” sites, with switching costs over 2 times that of

E*trade. Among the three largest brokers, it is about 1.5 times more difficult to induce a

representative customer of Schwab to switch than to induce a representative customer of

8 Or one can say that the representative new user and representative existing user have the same preferences. 9 The value we use for firm attributes are relative scores as recorded by Gomez Advisors, which are 0-10 scales. 10 We test for satisfaction of IIA by the so-called universal logit model (McFadden, 1974) or mother logit model (Kuhfeld, 1996) including all cross effects. A cross effect represents the effect of one alternative on the utility of another alternative. If the joint hypothesis test that all cross effects are zero cannot be rejected, that provides evidence that the independence assumption is not violated.

21

E*Trade to switch, ceteris paribus. This simple analysis cannot distinguish switching due to

customer characteristics and switching due to firm characteristics. Nonetheless, this does

provide an estimate of overall switching costs and it is consistent with earlier work that used a

different subset of the data and slightly different methods (Chen and Hitt, 2000a).

The ratio analysis has the advantage that it has minimal data requirements – one need only

calculate the fraction of customers who stay at each broker, and the market share for new

adopters. However, more complicated tests (such as validating the IIA assumption,

controlling for customer heterogeneity and many others) are much more easily conducted in a

regression framework using disaggregated individual choice data. We begin by estimating a

simple conditional logit model that computes switching cost using the model described in

equation (11), including controls for firm attributes, firm-specific dummy variables and time

dummy variables. This analysis yields switching cost estimates (Table 3 column 5) that are

very close to our ratio analysis shown in columns 1-4; this is depicted graphically in Figure

3b. This is not surprising, as they utilize the same underlying model. To examine H1

(equivalence of switching cost), we test whether that all firms have the same switching cost –

this is clearly rejected ( 2 (10) 189, 0.0001Pχ = < ). The estimated switching costs, with 95%

confidence intervals, are shown in Figure 4a.

More interesting, the results (Table 3, Column 6) are not substantially changed if we include a

full set of demographic controls in the analysis (age, gender, income, education, market size,

race, household size, marital status, occupation, number of brokers and frequency of visit), as

revealed in Figure 3b, regression 1 versus regression 2. Based on the regression results, we

are again easily able to reject the hypothesis that switching costs are identical across brokers

even controlling for demographics and individual customer characteristics

( 2 (10) 162, 0.0001Pχ = < ). For example, we find that E*Trade has significant lower

switching cost than all the other brokers we track (Figure 4b). It is also notable that Figure 4a

(no customer characteristics) and Figure 4b are quite similar, suggesting that the overall effect

of customer characteristics on switching is small. We therefore can also reject H2a

(equivalence of switching cost, controlling for customer characteristics).

22

Overall, this suggests that there is a significant firm-specific component of switching cost.

We will explore this variation in the latter sections of the paper, once we examine the

robustness of our estimates in the next section.

4.5 Robustness of the Switching Cost Estimates

In formulating the model, we have made several key assumptions, most of which are testable

in our regression framework. As stated earlier, we can test whether the independence

assumption (IIA) holds, and find evidence that IIA is supported for our data (see footnote 10

for the method). Second, we can examine the assumption that switching costs depend only on

the firm a customer switches from, and not the destination firm by estimating a more general

switching cost model that includes all possible combinations of previous and new brokers.

Our analysis suggests that the hypothesis that the adoption distribution for switching

customers is the same for all brokers cannot be rejected ( 2 (110) 59.24, 1.00Pχ = = ),

validating this assumption. These two tests are actually closely linked since both are implied

by IIA, but differ in implementation.

A potentially problematic assumption is that choice behavior for new customers and existing

customers is similar. One specific way in which this can be violated is if brokers have “new

user” subsidies or other benefits that can only be realized by new adopters. Since these

subsidies may entice more customers to adopt but do not increase the incentive to stay, it

could skew adoption rates relative to retention rates and lead to an underestimate of switching

costs for those brokers that have the greatest subsides.11 Because we do not have customer-

specific data or even complete broker-specific data on subsidies, and most brokers have a

wide variety of subsidy programs for new users, we cannot measure subsides directly.

However, analogous to switching costs (a cost to leave), we can define and measure an

“adoption cost” for each broker (which is directly related to the effectiveness of firms’

subsidy program), and examine whether firms have different adoption costs after controlling

for customer and broker characteristics. When adoption is subsidized, we generally expect

this to be a negative “cost” for customers that may or may not vary by brokers. If this cost 11 We thank Arun Sundararajan for noting this issue.

23

varies by broker or appears to be substantial, we will have to elaborate our switching cost

framework to adjust for these differences.

Identification of adoption cost proceeds in the same way as our switching cost measure. We

assume that a user does not face any adoption cost (ac) if he does not switch ( |k k k ku v ε= + ),

but adoption cost influences a new user’s adoption decision ( |k n k k ku v ac ε= − + ) and a

switcher’s “where to switch to” decision ( |j k j j k ju v ac s ε= − − + ). Because we have three

types of customers and three unknown parameters (v, ac, s), adoption cost can be identified.

This can be captured using a second dummy variable ( kV ) defined in a way that parallels the

switching cost dummy ( kW ) from equation (11). Let 1kV = if the customer is either a new

adopter in the adoption equation for broker k, or a customer who is an existing customer at a

different broker. It is zero otherwise. Regression estimates of adoption costs are presented in

Table 4. The test result shows that while Ameritrade and Datek seem to have lower adoption

costs these differences are not significant, and the magnitude of adoption costs is minimal

compared to the switching cost level (which are measured in the same units). This suggests

that our earlier switching results are robust to the assumption of common adoption behavior

between new and existing customers.

Taken collectively, our results suggest that we have a robust finding that switching costs vary

across brokers and that most observable customer characteristics (e.g. demographics and

usage) have only a limited influence on switching costs. This implies the presence of a strong

firm effect. In the following sections we explore the drivers of switching as well as customer

acquisition to understand how brokers influence customers’ switching behavior (and other

aspects of building a customer base).

4.5 Predictors of Switching Behavior

To predict direct effects of firm attributes on switching cost, we can proceed in two ways.

First, we can compute switching cost estimates for each firm and regress these on firm

characteristics. However, this strategy is limited in this context by the small number of firms

24

and time periods (a total of 33 estimates across 3 quarters) and thus may have low statistical

power. It also does not enable direct comparisons with adoption or attrition predictors nor can

it easily examine customer-specific effects. Alternatively, rather than test the direct effect of

firm attributes on switching costs, we can test how firm attributes influence customers’

switching behaviors. That is, we predict switching as a function of customer characteristics

and firm attributes using logistic regression. Formally, we estimate the model:

Pr( ) s s s sk k k i iSwitch x r zγ β α λ ε= + − + + (12)

Switch is a variable that is 1 if the customer switches, and zero otherwise. The parameters

( , , ,s s s sγ β α λ ) are analogous (but not the same as) to the parameters included in the

switching cost estimation model (11) – we use the superscript s to distinguish these

coefficients from those in the earlier analysis. These parameters represent the influence of

time-invariant firm specific switching effects, the effects of firm practices and the effect of

customer characteristics on switching rates, respectively.

In Table 5 we estimate three variants of this model. First, we examine a regression with only

customer characteristics (column 1). We find that demographics have little effect on overall

switching behavior as expected. However, more specific indicators of individual differences

are much better predictors of switching. Customers who have adopted fewer brokers are less

likely to switch – this is consistent if we interpret this measure as capturing unobserved

propensity to be loyal. We also find that changes in usage affect switching. Interestingly, we

find that level of activity is associated with reduced switching, which is consistent with a

story that greater experience with a service provider creates implicit lock-in through learning

(as suggested by Johnson et al., 2000). These results lend support to our arguments

summarized in H2b.

In column 2 of Table 5, we also add characteristics of the brokers to the analysis (the

measures are the characteristics of the broker a customer used in that period). Overall, we find

that higher customer confidence reduces switching while ease of use has a negative effect on

customer retention (albeit barely significant). Surprisingly, relationship services do not reduce

25

switching which is inconsistent with the idea that personalization leads to greater customer

lock-in.

In column 3, Table 5, broker dummies are added to the regression to capture any firm level

effects on switching. This eliminates the effects of levels of the broker attributes and thus

changes the coefficient interpretation of the broker characteristics variables to the effects of

changes in these factors. We find that in this analysis both the confidence and ease of use

factors disappear, suggesting that these do not vary highly over time. We do find a strong

beneficial effect of increasing resources (breadth of product line). The relationship services

coefficient is now marginally significant, but has the “wrong” sign. Thus, we find some

support for the assertions in H3 (firm-level determinants of switching), except for the result

on personalization, in both levels and fixed effects regressions.

This analysis indicates that high quality and increasing product line breadth are helpful in

reducing switching, but that most other factors have little (positive) influence on switching.

However, we still find large firm-level variation in switching as evidenced by our earlier

results and the strong significance levels of the broker dummies (Table 5 column 3) even

when control variables for specific practices are included. This suggests that while we have

identified some of the mechanisms by which firms influence (or might be able to influence)

their switching costs, firms still have significant control over their switching costs through

various “unobserved”12 strategies.

4.6 Drivers of Customer Attrition

In our earlier analysis, we found that attrition (customers who open a brokerage account and

do not return to any broker in the future) is a significant problem. Figure 2 shows that

attrition rates range from 33.7% for E*trade to 25% for Schwab. There are a variety of

reasons for attrition, most notably customer experimentation, especially experimentation

encouraged by subsidies. Using an identical model to equation (12) in the previous section

we can conduct a parallel analysis for attrition that we performed for switching.

12 In the sense that we don’t have them in our data.

26

In Table 6 we present 4 variations of the base model (the three considered previously for

switching, but also an additional model that includes only firm-specific dummy variables

which parallels that baseline switching cost estimates using ration analysis). Across all

specifications we find that younger people and women are more likely to become inactive.

The gender effect appears consistent with a recent study by Barber and Odeon (1999) that

found that men trade on-line significantly more frequently than women, so it is not surprising

that women are more likely to become inactive. Interestingly, our visit frequency data are

fairly consistent with Barber and Odeon’s study – our data show that single men visit their

brokers 58% more than single women, while their corresponding number using trading

volume is 67%. This suggests that our visit frequency information may not be a bad proxy for

trading behavior, at least when examining aggregate behavior.

Behavioral variables tend to be much better predictors of attrition: frequent visitors and

people with more accounts are less likely to become inactive. Unlike switching, there are

strong seasonal effects in attrition – attrition rates rose dramatically in Q2 2000. In addition,

the average visit frequency in Q200 was only 85% of that in Q100. This finding is consistent

with a recent report cited in Business Week (Gogoi, 2000): “overall, on-line trading volume

fell more than 20% in the second quarter (2000)….”. This seasonal effect is likely driven by

market conditions, since the Nasdaq and Internet stocks in particular had significant declines

over this period. Besides these observed variables, we find that E*Trade and Ameritrade have

significantly greater attrition rates than Schwab (Table 6 column 2 and column 4). Greater

minimum deposits are effective in reducing attrition rate (column 3) and, as before, ease of

use has a negative effect on attrition. These results are, again, largely consistent with our

prior arguments (summarized in H4), although there appear to be some clear demographic

effects which were not present in the previous analyses.

5. Discussion

Overall our analysis suggests that, using a variety of techniques, there are substantial

differences in switching cost across brokers. Our analysis is robust to our basic assumptions,

27

such as independence of irrelevant alternatives (IIA) and the switching cost depending only

on the existing broker. In addition, we find that heterogeneity in demographics of adopters

and unobserved features offered to new adopters do not significantly affect the results. We do

find that usage and changes in usage are good predictors of switching and attrition. We also

find that firm characteristics such as minimum balances or those tracked by Gomez that

measure overall “quality” and cost, also influence customers’ behaviors.

Ideally, a firm would like the highest acquisition rate and lowest switching rate and attrition

rates. Our analysis enables us to make comparisons on the types of factors that might be

generally more desirable in building a large and loyal customer base, while identifying others

that involve tradeoffs among practices or may be unexpectedly undesirable. For example,

high levels of customer service may increase acquisition and reduce switching and attrition,

while low minimum account requirements may improve acquisition at the expense of

increasing switching or attrition. Using analyses we discussed earlier, we can summarize

these effects in a single table (Table 7) using a consistent set of control variables. Note that

we have altered the signs of the coefficients such that “+” is good, and “-“ is bad; a factor is

coded as “0” if it is not statistically significant (p>0.05).

The results in Table 7 suggest that breadth of product offering appears to be universally

beneficial (if it can be provided at reasonable costs). Others have tradeoffs – low minimum

balances increase acquisition at the expense of attrition. Interestingly, technology strategies

that should be beneficial do not appear to be helpful: ease of use appears either ineffective or

negative, and investments in personalization (“relationship services”) appears to have no

effect at best. For ease of use, it may suggest that improvements in ease of use reduce

functionality, or it could be possible that sophisticated interface design creates lock-in due to

the time (cost) of learning a complex interface. For relationship services, it may simply

reflect that firms do not invest enough in these services to be effective – a situation that may

change over time as personalization technology improves. It is also important to note that

demographics typically are not good predictors of behavior except for a few isolated results

on attrition. Thus, for targeting consumers it is much more important to focus on behavioral

variables (particularly volume of usage and change of behaviors) to identify good customers.

28

Since the price of trading services substantially exceeds marginal cost and there is very little

unpriced customer service activity, higher volume customers are typically more profitable.13

Therefore, for example, it may be worthwhile to subsidize customers who show a high level

of use at a competitor (since they face higher switching costs) rather than new adopters.

Our findings are also consistent with prior work on on-line shopping behavior. Lohse et at.

(1999) found that past behavior tends to be a better predictors of future on-line behavior than

demographics. Moe and Fader (2000) show that changes in behavior form a potentially useful

basis for profitable customer segmentation.

5. Conclusion

Previous theoretical work has shown that the presence of switching costs, either generally or

in specific firms, can have a substantial effect on profitability. However, the creation of

switching costs requires substantial and deliberate investments by the firm in customer

retention. By understanding the magnitude of these switching costs, it is then possible to

understand tradeoffs between investments in loyalty and retention programs and other types

of investments such as advertising (for building new customer acquisition rates) and service

level improvements or price reductions which raise both the acquisition and retention rates

simultaneously. Our analysis reveals that implied switching costs vary substantially across

brokers, but this does not appear to be explained by differences in customer characteristics.

This suggests that factors under the firm’s control may influence these switching costs. Our

initial analyses using firm attributes identifies some of these factors, but there is still

substantial heterogeneity, suggesting that firms have significant control over their switching

costs through various kinds of retention strategies. By linking the switching costs due to firm

specific retention strategies to the implementation costs, we can check the effectiveness of

different retention strategies, and see how well each firm is in converting the investment into

effects. It is a promising avenue for future research in this and other industries where

13 This stands in contrast to other financial industries, such as banking, where transaction volume is typically associated with lower customer profitability.

29

customer retention is critical or where firm growth is such that all new customer adoption

necessarily results from switching rather than growth in the industry as a whole.

The method and approach used by this paper are readily applicable to analyze other on-line

markets or industries. The method proposed here is especially suitable for on-line businesses

because we are able to observe all the products a customer considered and know the options

that were available at the time the customer made an adoption choice. Our goal is to develop

a platform that enables the components of switching cost to be separated, measured, and

correlated with firms’ attempts to retain customers. Our initial results for the on-line

brokerage industry suggest that this is a promising approach.

30

Figures: Figure 1: Customer flow diagram (sample period: Q3-99 to Q4-99) Figure 2: Customer flow rates

Customer flow chart

0%

10%

20%

30%

40%

50%

60%

70%

80%

90%

100%

AMERITRADE

DATEK

DLJDIR

ECT

ETRADE

FIDELIT

Y

FLEET

MSDWNDB

SCHWAB

TDWATERHOUSE

VANGUARD

attritionswitchstay

Q399-Q200 Media Metrix

datasets: 28,807

households

Householdstracked

overtime: 11,397

households

Broker Users: 1,249

Non- brokerusers

Q3-99 Q4-99

Stayers: 834

Switchers:146

Non-broker

users: 269

New adopters:

605

31

Figure 3a: Stability of switching cost measures over time

Note: we do not include Q499 NDB switching cost measure due to insufficient data. Figure3b: Switching cost estimates

Regression 1: switching cost measures without control for customer heterogeneity. Regression 2: switching cost measures after control for customer heterogeneity.

Switching cost by broker

0123456789

ETRADE

FIDELIT

Y

SCHWAB

AMERITRADE

VANGUARD

DLJDIR

ECT

TDWATERHOUSE

MSDWFLEET

DATEKNDB

Q499Q100Q200

Switching cost by broker

0

1

2

3

4

5

6

ETRADE

FIDELIT

Y

SCHWAB

AMERITRADE

VANGUARD

DLJDIR

ECT

TDWATERHOUSE

MSDWFLEET

DATEK

PooledRegression 1Regression 2

32

Figure 4a: Switching cost measure with 95% confidence interval without controlling for demographics

switching cost measure without controlling for demographics

0123456789

10

ETRADE

FIDELIT

Y

AMERITRADESCH

VANGUARD

DLJDIR

ECT

TDWATERHOUSE

MSDWFLE

ET

DATEKNDB

Figure 4b: Switching cost measure with 95% confidence interval after control for customer heterogeneity

Switching cost after control for customerheterogeneity

02

46

810

12

ETRAD

EFI

DELIT

YAM

ERIT

RADE

SCH

VANGUAR

DDLJ

DIREC

TMSD

W

DATEK

TDW

ATER

HOUSE

FLEE

T

NDB

33

Tables Table 1a: Market Share by Broker

BROKER Q399

USERS SHARE Q499

USERS SHAREQ100

USERS SHARE Q200

USERS SHARE AMERITRADE 88 7.05% 101 7.25% 151 8.48% 147 9.27% DATEK 40 3.20% 45 3.23% 63 3.54% 78 4.92% DLJDIRECT 53 4.24% 54 3.88% 74 4.16% 73 4.60% ETRADE 278 22.26% 300 21.54% 375 21.07% 341 21.50% FIDELITY 328 26.26% 384 27.57% 449 25.22% 392 24.72% FLEET 37 2.96% 40 2.87% 50 2.81% 48 3.03% MSDW 17 1.36% 19 1.36% 22 1.24% 13 0.82% NDB 14 1.12% 27 1.94% 26 1.46% 20 1.26% SCHWAB 216 17.29% 222 15.94% 278 15.62% 251 15.83% TDWATERHOUSE 87 6.97% 103 7.39% 130 7.30% 98 6.18% VANGUARD 91 7.29% 98 7.04% 162 9.10% 125 7.88% Total 1249 100.00% 1393 100.00% 1780 100.00% 1586 100.00% Table 1b: Market share of new adopters

BROKER Q499 NEW ADOPTERS SHARE

Q100 NEW ADOPTERS SHARE

Q200 NEW ADOPTERS SHARE

AMERITRADE 41 6.78% 64 9.68% 44 11.37%DATEK 20 3.31% 24 3.63% 24 6.20% DLJDIRECT 28 4.63% 29 4.39% 14 3.62% ETRADE 161 26.61% 149 22.54% 98 25.32%FIDELITY 156 25.79% 157 23.75% 85 21.96%FLEET 18 2.98% 23 3.48% 12 3.10% MSDW 10 1.65% 8 1.21% 4 1.03% NDB 14 2.31% 8 1.21% 3 0.78% SCHWAB 73 12.07% 86 13.01% 48 12.40%TDWATERHOUSE 41 6.78% 37 5.60% 19 4.91% VANGUARD 43 7.11% 76 11.50% 36 9.30% TOTAL 605 100.00% 661 100.00% 387 100.00%

34

Table 2: User flow by broker (all time periods pooled)

BROKER

PERCENTAGE OF

CUSTOMERS THAT STAY

PERCENTAGE OF

CUSTOMERS THAT SWITCH

PERCENTAGE OF

ATTRITION AMERITRADE 59.12% 9.41% 31.47% DATEK 66.22% 7.43% 26.35% DLJDIRECT 58.56% 11.05% 30.39% ETRADE 55.72% 10.60% 33.68% FIDELITY 65.81% 6.55% 27.65% FLEET 60.63% 10.24% 29.13% MSDW 50.00% 22.41% 27.59% NDB 65.67% 2.99% 31.34% SCHWAB 67.04% 7.54% 25.42% TDWATERHOUSE 65.63% 8.13% 26.25% VANGUARD 54.99% 7.12% 37.89%

Table 3: Estimated Switching Cost (various methods)

1 2 3 4 5 6 Q499 Q100 Q200 Pooled Regression 1 Regression 2AMERITRADE 4.64 3.89 3.93 4.15 4.07 (0.20) 3.97(0.22) DATEK 5.10 5.29 5.71 5.34 5.30 (0.34) 5.18 (0.37) DLJDIRECT 4.06 4.85 5.66 4.77 4.70 (0.27) 4.79(0.29) ETRADE 2.20 3.03 3.14 2.78 2.77 (0.12) 2.72 (0.13) FIDELITY 3.25 3.50 3.66 3.46 3.53 (0.13) 3.60(0.15) FLEET 5.01 5.16 5.42 5.19 5.29 (0.33) 5.49 (0.41) MSDW 4.56 5.87 5.03 5.11 5.22 (0.39) 5.26(0.47) NDB na 7.11 7.62 7.27 7.36 (0.75) 8.80 (1.23) SCHWAB 3.97 4.25 4.21 4.13 4.05 (0.16) 4.02(0.18) TDWATERHOUSE 4.57 4.63 5.78 4.86 4.87 (0.23) 5.09 (0.27) VANGUARD 4.38 4.15 4.52 4.31 4.34 (0.23) 4.38(0.25)

Columns 1-4 are ratio analysis-based measures Regression 1: switching cost measures without control for customer heterogeneity. Standard error in parenthesis. Regression 2: switching cost measures after control for customer heterogeneity. Standard error in parenthesis. Table 4: adoption cost measures (baseline: Schwab)

BROKER ESTIMATE STD ERROR AMERITRADE -0.576* 0.274 DATEK -0.777* 0.389 DLJDIRECT -0.042 0.320 ETRADE 0.140 0.206 FIDELITY -0.102 0.216 FLEET -0.275 0.392 MSDW 0.575 0.437 NDB -1.638 1.082 SCHWAB 0.025 0.306 TDWATERHOUSE -0.414 0.287

* - Significantly different from zero at P<0.05

35

Table 5: Predictors of Switching Behavior (negative is less switching) REGRESSION 1 REGRESSION 2 REGRESSION 3 Intercept -2.2375***

(0.550) -0.6148 (1.180)

-0.7803 (1.572)

Age -0.0091 (0.005)

-0.0080 (0.005)

-0.0070 (0.005)

Female 0.0859 (0.140)

0.1142 (0.142)

0.1009 (0.144)

hhsize (Household size)

0.0282 (0.056)

0.0351 (0.056)

0.0282 (0.057)

race1 (white) -0.1965 (0.251)

-0.2095 (0.254)

-0.1990 (0.257)

race3 (oriental) 0.3913 (0.325)

0.3159 (0.329)

0.4517 (0.336)

hhinc (Household income)

-0.0015 (0.002)

0.0015 (0.002)

0.0017 (0.002)

Education -0.0813 (0.065)

-0.0823 (0.065)

-0.0859 (0.066)

mktsize 2χ (4)=4.86 P=0.30

2χ (4)=4.58 P=0.33

2χ (4)=5.03 P=0.28

Marrital status 2χ (2)=0.02 P=0.99

2χ (2)=0.17 P=0.92

2χ (2)=0.17 P=0.92

Occupation 2χ (5)=3.88 P=0.57

2χ (5)=3.80 P=0.58

2χ (5)=3.66 P=0.60

Previous frequency -0.0783*** (0.010)

-0.0779*** (0.011)

-0.0748*** (0.010)

Change ratio (change in usage)

0.0562*** (0.015)

0.0563*** (0.016)

0.0579*** (0.016)

Previous number of brokers

1.0766*** (0.096)

1.0742*** (0.097)

1.0473*** (0.098)

Q100 -0.3845** (0.147)

-0.407** (0.150)

-0.3563* (0.155)

Q200 -0.4777** (0.151)

-0.2611 (0.191)

-0.3066 (0.194)

Easeofuse 0.0989* (0.050)

0.0231 (0.072)

Confidence -0.1959* (0.096)

-0.1658 (0.100)

Resources -0.0495 (0.116)

-0.4879*** (0.137)

Services -0.1096 (0.146)

0.3227* (0.160)

Cost 0.00737 (0.064)

0.1559 (0.095)

Minimum deposit 0.0568 (0.071)

Broker dummies 2χ (10)=35.8 P<0.0001***

N 2824 2824 2824 2χ 264.06*** 278.09*** 316.73***

Standard errors in parenthesis; *- P<0.05 **- P<0.01 ***- P<0.001

36

Table 6: Attrition analysis (negative is less attrition) Regression 1 Regression 2 Regression 3 Regression 4 Intercept 1.4291***

(0.372) 1.0916** (0.401)

1.2518 (0.939)

1.1759 (2.940)

Age -0.0106*** (0.003)

-0.0093** (0.003)

-0.0091** (0.003)

-0.00947** (0.003)

female 0.2415** (0.089)

0.2615** (0.090)

0.2698** (0.090)

0.2506** (0.090)

Education -0.0901* (0.046)

-0.0935* (0.046)

-0.0910* (0.0.046)

-0.0952* (0.046)

Occupation 2χ (5)=10.75 P=0.06

2χ (5)=10.45 P=0.06

2χ (5)=11.55 P=0.04*

2χ (2)=11.14 P=0.05*

Previous frequency (usage)

-0.2549*** (0.016)

-0.2536*** (0.016)

-0.2543*** (0.016)

-0.2548*** (0.016)

Previous number of brokers

-0.2161 (0.118)

-0.2512* (0.121)

-0.2361* (0.120)

-0.2402* (0.121)

Q100 0.2110 (0.110)

0.2089 (0.110)

0.1451 (0.112)

0.1017 (0.120)

Q200 1.1391*** (0.103)

1.1355*** (0.103)

1.0697*** (0.134)

1.187*** (0.164)

easeofuse

0.1359*** (0.040)

0.3462*** (0.080)

confidence

0.0525 (0.066)

-0.0396 (0.104)

resources

0.1126 (0.083)

-0.1391 (0.186)

services

0.0073 (0.101)

-0.00711 (0.165)

Cost

-0.004 (0.045)

-0.2397 (0.167)

Minimum deposit

-0.101* (0.047)

Ameritrade 0.4934** (0.188)

2.3448* (0.938)

Datek 0.0824 (0.260)

1.4091 (1.041)

DLJDirect 0.2554 (0.231)

0.877 (0.535)

E*Trade 0.5008*** (0.147)

1.0587* (0.461)

Fidelity 0.1303 (0.140)

0.6848** (0.217)

Fleet 0.3517 (0.268)

1.8843* (0.860)

MSDW 0.2669 (0.381)

0.1476 (0.803)

NDB 0.1965 (0.351)

1.0515 (0.633)

TDWaterhouse 0.099 (0.204)

2.1112** (0.810)

Vanguard 0.3981* (0.18)

1.675 (0.922)

N 3634 3634 3634 3634 2χ 1009.24*** 1029.83*** 1037.28*** 1052.14***

Standard errors in parenthesis; *- P<0.05 **- P<0.01 ***- P<0.001; Some insignificant demographic variables omitted from table due to space considerations but included in the analysis (hhsize, race, hhinc, mktsize and marital status)

37

Table 7: Summary of factors that affect acquisition, switching and attrition Model includes broker characteristics, customer characteristics and time controls

*- p<0.05 ** - p<0.01 *** - p<0.001 Note that many factors in the acquisition model are insignificant due to the large number of demographic control variables interacted with firm dummy variables (see Section 3 for a discussion). When individual effects are not included, the remaining Gomez factors (cost, ease of use, confidence, relationship management) are all positive and significant, as would be expected.

Acquisition Switching AttritionCost 0 0 0Ease of Use 0 -* -***Confidence 0 +* 0Resources +*** +*** 0Relationship Mgmt. 0 0 0Minimum Deposit -** 0 +*Usage varied +*** +***Change in Usage na -*** n/amultiple brokers varied -*** +*Demographics varied minimal Age (+**)

Women (-**)Education (+*)

Overall Fit χ2(186)=1430 χ2(29)=278 χ2(28)=1031Model Conditional logit Logisitic Logisitic

38

Appendix: Variable Definitions Variable description and code: VARIABLE DESCRIPTION CODE sex Gender F: female; M:male mktsize Market size- MSA 3=50,000--499,999; 4=500,000 999,999; 5=1,000,000-

2,499,999; 6=2,500,000 and over; 9=Non-MSA. race 1=White, 3=Oriental, 4=Black and other. married Marital Status 1=married, 3=widowed or divorced or separate,

4=single. occ Occupation 1=Professional; 2=Proprietors, Managers, Officials;

4=Sales; 5=Craftsmen, Foremen or operative; &=Retired, Unemployed; o=others.

frequency Visiting frequency (number of days visit)

Change ratio 2 . 1 .

1period freq period freq

period freq−

Easeofuse A composite measure of firms’ ease of use, including useful account demos and tutorials, easily accessible new account forms, well-integrated features, and some basic customization capability, etc.

Confidence A measure of customers confidence for the site: do they have a large capital base, track their Web site availability and phone response time, have a large well trained customer service staff, and disclose key information about trading rules and executions to their customers, etc.

Resources A measure of on-line resources available besides quotes-charts-news, including useful screening tools for stocks and mutual funds, portfolio analytics and a broad product line.

Services A measure of firms’ relationship services with their customers, including electronic relationships with real-time account information, sophisticated alerts, advice oriented functionality, and educational content.

Cost A measure of whether firms offers low-cost equity trading across market and limit orders and fund trading, and competitive margin and interest rates.

Deposit Minimum deposit required to open an account.

39

References: 1. Anderson, S.P., A. DePalma and J.-F. Thisse. Discrete choice theory of product differentiation, 1992,

MIT Press, Cambridge, MA. 2. Bakos, Y., H. C. Lucas, Jr., W. Oh, G. Simon, S. Viswanathan and B. Weber. "The Impact of Electronic

Commerce on the Retail Brokerage Industry." Working Paper, Stern School of Business, New York University. 2000.

3. Bakos, Y. “The Emerging Landscape for Retail E-Commerce,” Journal of Economic Perspectives, Jan. 2001.

4. Barber, B and T. Odean. “Boys will be Boys: Gender, Overconfidence, and Common Stock Investment, “ Working Paper, UC-Davis, 1999.

5. Beggs, A and P. Klemperer. "Multi-Period Competition with Switching Costs," Econometrica (60:3), 1992, pp. 651-666.

6. Ben-Akiva, M and S. R. Lerman. Discrete Choice Analysis: Theory and Application to Travel Demand, 1985, MIT Press, Cambridge, MA.

7. Berry, S. “Estimating discrete-choice models of product differentiation,” The Rand Journal of Economics (25), 1994, pp.242-262.

8. Besanko, D, S. Gupta and D. Jain. “Logit Demand Estimation Under Competitive Pricing Behavior: An Equilibrium Framework,” Management Science (44), 1998, pp.1533-1547.

9. Boskin, M. J. “A Conditional Logit Model of Occupational Choice,” Journal of Political Economy (82), 1974, pp.211-264.

10. Boulding, W., A. Kalra, R. Staelin, and V. A. Zeithaml. "A Dynamic Process Model of Service Quality; From Expectations to Behavioral Intentions," Journal of Marketing Research, 30, 1993, pp.7-27.

11. Bresnahan, T., S. Stern and M. Trajtenberg. “Market Segmentation and the Sources of Rents From Innovation: Personal Computers in the Late 1980’s,” Rand Journal of Economics, 1996.

12. Brynjolsson, E. and M. Smith. “The Great Equalizer? Consumer Choice Behavior at Internet Shopbots,” Working paper, MIT-Sloan, 2000.

13. Chen, P-Y. and L. M. Hitt. “Switching Cost and Brand Loyalty in Electronic Markets: Evidence from On-Line Retail Brokers,” in Proceedings of International Conference on Information Systems, 2000a.

14. Chen, P-Y. and L. M. Hitt. “Measurement of Switching Costs and Brand Loyalty in Electronic Markets,” Working paper, The Wharton School, 2000b.

15. Cingil, I., A. Dogac and A. Azgin. “A broader approach to personalization,” Communications of the ACM (43:8), 2000 Aug, pp.136-141.

16. Colombo, R. A. and D. G. Morrison. “A Brand Switching Model with Implications for Marketing Strategies,” Marketing Science (8:1), 1989, pp. 89-99.

17. Crosby, L. A. and N. Stephens. "Effects of Relationship Marketing on Satisfaction, Retention, and Prices in the Life Insurance Industry," Journal of Marketing Research, 24 , 1987, pp.404-11.

18. Elzinga, K. G. and D. E. Mills. “Switching costs in the wholesale distribution of cigarettes,” Southern Economic Journal (65:2), 1998 Oct., pp. 282-293.

19. Farrell, J. and C. Shapiro. "Dynamic Competition with Switching Costs," Rand Journal of Economics (19:1), 1988, pp.123-137.

20. Friedman, T. L. “Amazon.you,” New York Times, Feb. 26, 1999, p. A21. 21. Gans, N. “Customer Loyalty and Supplier Strategies for Quality Competition: II,” Working Paper, The

Wharton School, 2000b. 22. Gans, N. and Y. Zhou. “Customer Loyalty and Supplier Strategies for Quality Competition: I ,” Working

Paper, The Wharton School, 2000a. 23. Gogoi, Pallavi. “Rage Against On-line Brokers,” Business Week, Nov. 20 2000, pp. 98-102. 24. Green, W. H. Econometrics Analysis, Prentice-Hall, NJ, 1997. 25. Grover, Rajiv and V. Srinivasan, “A Simultaneous Approach to Market Segmentation and Market

Structuring,” Journal of Marketing Research (24), 1987, pp. 139-153. 26. Guadagni, P. M. and J. Little. “A Logit Model of Brand Choice Calibrated on Scanner Data,” Marketing

Science (2), 1983, pp. 203-238. 27. Handelsman, M. and J. M. Munson. "Switching Behaviours from Credit Card to Cash Payment Among

Ethnically Diverse Retail Customers," International Journal of Retailing (4:1), 1989, pp.31-44. 28. Hitt, L. M. and F. Frei. “Customer Heterogeneity in Electronic Markets: The Case of PC Banking,”

Working paper, Harvard Business School, 1999.

40

29. Hoffman, D. L. and T. P. Novak. “Marketing in Hypermedia Computer-Mediated Environments: Conceptual Foundations,” Journal of Marketing (60:3), 1996, pp. 50-68.

30. Hoffman, D.L., T.P. Novak, and M.A. Peralta. "Building Consumer Trust On-line," April, Communications of the ACM (42:4), 1999, pp.80-85.

31. Jacoby, J and R. W. Chestnut. Brand Loyalty Measurement and Management, John Wiley and Sons, NY, 1978.

32. Johnson, E. J., Wendy Moe, Peter S. Fader, Steven Bellman, and Jerry Lohse. "On the Depth and Dynamics of On-line Search Behavior," Working Paper, The Wharton School, 2000.

33. Klemperer, P. “Markets with Consumer Switching Costs,” The Quarterly Journal of Economics (102:2), 1987, pp.375-394.

34. Klemperer, P. “Price Wars Caused by Switching Costs,” The Review of Economic Studies (56), 1989, pp. 405-420.

35. Klemperer, P. “Competition when Consumers have Switching Costs: An Overview with Applications to Industrial Organization, Macroeconomics, and International Trade,” The Review of Economic Studies (62), 1995, pp. 515-539.

36. Kuhfeld, W. F. “Multinomial Logit, Discrete Choice Modeling,” Technical report, 1996, SAS Institute Inc. 37. Lohse, J., S. Bellman and E. Johnson. “Predictors of On-line buying: Findings from the Wharton Virtual

Test Market”. Communications of the ACM, Dec. 1999. 38. McCarthy, P. S., P. K. Kannan, R. Chandrasekharan and G. P. Wright. “Estimating Loyalty and Switching

with an Application to the Automobile Market,” Management Science (38:10), 1992, pp. 1371-1393. 39. McFadden, D. “Conditional Logit Analysis of Qualitative Choice Behavior,” in Frontiers in Econometrics,

P. Zarembka, eds. Academic Press, NY, 1974a. 40. McFadden, D. “The Measurement of Urban Travel Demand,” Journal of Public Economics (3), 1974b, pp.

303-328. 41. McFadden, D. “Modelling the Choice of Residential Location.” In Spatial Interaction Theory and

Residential Location. A. Karlquist et al., eds. North Holland, Amsterdam, 1978, pp. 75-96. 42. McFadden, D. “Econometric Models for Probabilistic Choice among Products,” Journal of Business

(53:3), 1980, pp. 13-29. 43. McFadden, D. “Economic Models of Probabilistic Choice,” Structural Analysis of Discrete Data with

Econometric Applications, ed. C. Manski and D. McFadden, 1981, MIT Press, Cambridge MA. 44. McVey, Henry. On-line Financial Services, presentation by Morgan Stanley Dean Witter, August 2000.

(available at: http://www.msdw.com/) 45. Mobasher, B., R. Cooley and J. Srivastava. “Automatic personalization based on Web usage mining.”

Communications of the ACM (43:8), 2000 Aug., pp.142-151. 46. Moe, W and P. S. Fader, "Capturing Evolving Visit Behavior in Clickstream Data," Working Paper, The

Wharton School, 2000. 47. Morgan Stanley Dean Witter. The Internet and Financial Service Report. 1999. 48. Nilssen, T. “Two Kinds of Consumer Switching Costs,” Rand Journal of Economics (23:4), 1992, pp.579-

589. 49. Novak, T. P, D. L. Hoffman and Y-F. Yung. “Measuring the Customer Experience in On-line

Environments: A structural Modeling Approach.” Marketing Science, special issue on “Marketing Science and the Internet”, 2000.

50. Pearson, M. “Attractors: Building Mountains in the Flat Landscape of the World Wide Web,” Director (51:12), 1998 Jul., pp. 81.

51. Pessemier, E. A. “A New Way to Determine Buying Decisions,” J. Marketing (24), 1959, pp.41-46. 52. Raju, J., V. Srinivasan and R. Lal. “The Effects of Brand Loyalty on Competitive Price Promotional

Strategies,” Management Science (36), 1990, pp.276-304. 53. Reichheld, F. F. and P. Schefter. “E-loyalty: Your secret weapon on the Web,” Harvard Business Review

(78:4), 2000 Jul/Aug, pp. 105-113. 54. Reichheld, F. F., R. Markey and C. Hopton. “The loyalty effect - the relationship between loyalty and

profits,” European Business Journal (12:3), 2000, pp. 134-139. 55. Robinson, W. T. “Sources of Market Pioneer Advantages: The Case of Industrial Goods Industries,”

Journal of Marketing Research (25:1), February 1988, pp.87-94. 56. Roy, A.. “Correlates of Mall Visit Frequency,” Journal of Retailing, 70 (2), 1994, pp.139-161. 57. SalomanSmith Barney research reports on on-line brokers, February 2000.

41

58. Schmalensee, R. “Brand Loyalty and Barriers to Entry,” Southern Economic Journal (40), 1974, pp.579-588.

59. Schmalensee, R. “Product Differentiation Advantages of Pioneering Brands,” American Economic Review (72), 1982, pp.349-365.

60. Schmalensee, R and R. Willig. Handbook of Industrial Organization, North-Holland, Amsterdam, 1986. 61. Schmidt, P., and R. Strauss. “The Prediction of Occupation Using Multiple Logit Models,” International

Economic Review (16), 1975a, pp. 471-486. 62. Schmidt, P., and R. Strauss. “Estimation of Models with Jointly Dependent Qualitative Variables: A

Simultaneous Logit Approach,” Econometrica (43), 1975b, pp. 745-755. 63. Shankar, V., A. K. Smith and A. Rangaswamy. “Customer Satisfaction and Loyalty in On-line and Offline

Environments,” Working paper, Penn State University, 2000. 64. Smith, M. D., J. Bailey, and E. Brynjolfsson. “Understanding Digital Markets: Review and Assessment,”

in Understanding the Digital Economy, E. Brynjolfsson and B. Kahin, eds. 1999, MIT Press. Cambridge MA.

65. Varian, H. “Effect of the Internet on Financial Markets,” Working paper, UC-Berkeley, 1998. 66. Varian, H. “Market Structure in the Network Age,” Working paper, UC-Berkeley, 1999. 67. Zephirin, M. G. “Switching Costs in the Deposit Market,” The Economic Journal (104), 1994, pp.455-461.