Compatibility, Competition, and Investment in Network
Industries: ATM Networks in the Banking Industry
Joy Ishii∗
Stanford University
November 17, 2005
Abstract
Pricing in network industries that discriminates between aﬃliated and unaﬃliated consumers
is a form of partial incompatibility. This paper measures the magnitude of the eﬀects this incompatibility
has on competition, welfare, and investment in the retail banking industry, where
banks’ ATM networks are connected to one another and unaﬃliated consumers face fees called
surcharges. I present a structural model of consumer and bank behavior that allows banks to
choose ATM networks and set deposit rates and allows consumers to choose banks. The estimates
conﬁrm that network eﬀects are very important in this industry and show that surchargeinduced
incompatibility has a signiﬁcant impact on demand for deposit services. The results
also suggest that there is overinvestment in ATMs relative to the social optimum. Counterfactual
experiments conditional on network size predict that a move to compatibility through
the elimination of surcharges would substantially decrease market concentration, raise average
deposit interest rates, raise consumer surplus, lower industry proﬁts, and reallocate proﬁts from
large-network banks to small-network banks. Banks are also predicted to respond by reducing
the number of their ATMs.
∗
Stanford University Graduate School of Business, Stanford, CA 94305. Email: ishii joy@gsb.stanford.edu. I
am grateful to Ariel Pakes for his advice. I would also like to thank David Blackburn, Eleanor Ericson, Kate
Ho, Catherine Jones, Julie Mortimer, Markus M¨obius, Marc Rysman, Minjae Song, and seminar participants at the
University of California-Berkeley, Brown, Chicago GSB, Harvard University, MIT, NYU Stern, Northwestern Kellogg,
the University of North Carolina Kenan-Flagler, the University of Pennsylvania, Stanford GSB, the University of
Wisconsin-Madison, and Wharton.
1 Introduction
This paper examines the eﬀects of incompatibility in network industries. In a network industry such
as telecommunications, the internet, or automatic teller machines (ATMs) in the banking industry,
ﬁrms are technologically interconnected. This interconnection can lead to more complicated pricing
structures than those observed in traditional industries, since a consumer may receive direct or
indirect services both from his chosen ﬁrm and its rivals. While interconnection increases the size
of the network available to consumers, in industries such as the banking industry, the introduction
of price discrimination between aﬃliated and unaﬃliated consumers reintroduces ﬁrm-level network
economies by reducing compatibility within the shared network. This paper measures the impact
of this incompatibility and ﬁnds signiﬁcant eﬀects on competition in the deposit market, welfare,
and investment. It also brieﬂy considers an alternative institutional structure in which provision of
ATM and deposit services is separated.
In the banking industry, the customers of one bank can use their ATM cards at ATMs owned by
other banks, but the ATM owner may charge a fee called a surcharge.1 This can be interpreted as
partial incompatibility between components of a system comprised of ATM cards (bank aﬃliation)
and ATMs. Analogous to the strong complementary relationships between CPUs and peripherals
or VCRs and video tapes, ATM cards and ATMs form complementary components of a system that
allows consumers to perform transactions on their bank accounts. Consumers can choose various
combinations of these complementary goods, but the compatibility is only partial since there is a
cost associated with use of a foreign ATM, that is, an ATM not owned by the consumer’s bank.
There is a sizeable theoretical literature on compatibility in industries with network externalities
or complementary components.2 This literature predicts that incentives for compatibility diﬀer
across ﬁrms and will be smaller for ﬁrms with larger networks, since these ﬁrms lose the competitive
advantage their network size confers under incompatibility.3 The eﬀects on consumer surplus should
diﬀer depending on the distribution of consumer characteristics and the new price equilibrium that
is reached. In turn, the eﬀects of compatibility on price competition depend on a number of
factors. In the banking industry, while partial incompatibility achieved through surcharging should
theoretically soften price competition in the deposit market by making an increase in deposit
1
This paper will focus on the types of ATM transactions that can be performed on any ATM within the shared
network such as inquiries and cash withdrawals. According to a study by Dove Consulting (2002), these make up 88
percent of all ATM transactions.
2
See, for example, Katz and Shapiro (1985) and Farrell and Saloner (1985) for the network externality approach.
Matutes and Regibeau (1988) and Economides (1989) introduce models in which there is no consumption externality
but consumers assemble systems from complementary components.
3
For example, Katz and Shapiro (1985) show this eﬀect in their classic model of network externalities, Cremer,
Rey, and Tirole (2000) show a similar eﬀect in their model of interconnection quality between internet backbones, and
Chou and Shy (1993) show that a computer hardware producer’s proﬁt decreases when it increases its compatibility
with software designed for a rival’s hardware.
1
market share come at the expense of surcharge revenue (see Matutes and Padilla 1994), one might
expect this eﬀect to be swamped by the impact of diﬀering ATM network sizes on banks’ pricing
incentives. For example, banks with small networks may be forced to raise their deposit rates
to remain competitive. The impact of incompatibility on prices, welfare, and the distribution of
market shares is therefore an empirical question. In addition, little is known about how banks’
investment incentives aﬀect the optimality of the network equilibrium.
There has been considerable debate regarding the direction and magnitude of these eﬀects since
the widespread introduction of surcharging in the U.S. in 1996. While various banks and businesses
have emphasized the positive eﬀect of surcharges on ATM deployment, some consumer advocacy
groups, politicians, and small ﬁnancial institutions have argued that surcharging is an anticompetitive
practice that threatens the survival of small banks. Calls for government intervention have
ranged from requests for regulation of fee disclosure to demands for a total ban on surcharging.
However, there has been only limited empirical evidence about the eﬀects on the retail banking
industry of the partial incompatibility induced by surcharging.
This paper presents a model of demand, price-setting, and ATM network investment in the
banking deposit market that allows me to analyze the eﬀects of surcharging as well as the implications
of the current equilibrium and alternative institutional structures. I estimate the demand
and cost-side parameters from the model of consumer and ﬁrm behavior using standard method
of moments techniques in addition to a new methodology presented in Pakes, Porter, Ho, and
Ishii (2005), the method of moments with inequality constraints. To do this, I use a new data
set that includes ownership and surcharge information for ATMs in the state of Massachusetts. I
combine this with data on consumer characteristics, bank deposits, and bank characteristics. After
estimating the model’s parameters, I ﬁnd that surcharging has a large impact on bank demand.
A one percent increase in a bank’s surcharge (and hence an increase in the surcharge costs that
a consumer avoids by choosing that bank) results in an average increase of 0.12 percent in that
bank’s demand.
Using the estimates from the banking demand and cost equations, I ﬁnd that in the current
equilibrium, banks’ direct proﬁts from ATM operations generally do not cover their estimated
ATM costs. This highlights the importance of the eﬀect of ATMs on the deposit market. Indeed,
the demand-stealing eﬀects in the deposit market associated with ATM networks and surcharging
seem to be large enough to cause overinvestment in ATMs. In the observed equilibrium, the average
addition to consumer surplus from a bank’s ﬁnal ATM is estimated to be smaller than the reduction
in total proﬁts.
Next, I consider counterfactual experiments to measure the impact of surcharge-induced incompatibility
on market shares, equilibrium prices, and consumer and ﬁrm welfare. Holding constant
2
network size, I ﬁnd that an elimination of surcharges causes average deposit interest rates to rise.
Banks with large ATM networks raise their interest rates and lose market share, and banks with
small networks lower their interest rates slightly and gain market share. The magnitude of these
eﬀects is large. In the Boston banking market, for example, the Herﬁndahl-Hirschman index is
estimated to fall by about 45 percent upon the elimination of surcharges. These interest rate and
market share eﬀects translate into a sizeable reallocation of proﬁts from large-network banks to
small-network banks with an overall loss of industry proﬁts and an increase in consumer surplus.
However, as one would expect, all banks have an incentive to reduce their network sizes.
This motivates consideration of a market design question. I brieﬂy consider a scenario in which
the provision of ATM and deposit services are separated through the use of a network operator
which charges consumers the cost of operating the network. I ﬁnd that while this is likely to reduce
consumer surplus, total industry proﬁts are estimated to rise, though by a statistically insigniﬁcant
amount, and proﬁts are redistributed away from large-network banks toward small-network banks.
In addition to this paper’s contributions to the literature on network industries, it contributes to
the empirical banking literature by accounting for ATM pricing and network size in estimates of the
demand for deposit services and by developing a structural model that accounts for heterogeneous
ﬁrm and consumer characteristics in a ﬂexible way. In particular, the model accounts for the spatial
heterogeneity generated by diﬀerent consumer and bank branch locations and the heterogeneity in
income levels across diﬀerent geographic locations. This is important in a retail industry such as
banking. It is also able to accommodate a more natural deﬁnition of deposit market shares than
those used in previous research. It allows consumers to have heterogeneous levels of income and
deposits and yields predictions about each bank’s dollar volume of deposits as well as its share of
consumers.
This paper proceeds as follows: the remainder of this section discusses the relationship of this
research to the existing literature. Section 2 discusses some of the important institutional details
and history of the ATM industry. In section 3, I describe and summarize the data that will be
used. Section 4 presents the model of consumer and ﬁrm behavior, and sections 5 and 6 describe my
implementation and estimation of it. Section 7 provides estimation results, and section 8 discusses
some implications of the estimates and the results of counterfactual experiments. Finally, section
9 concludes. Additional results are presented in the appendices.
1.1 Literature
This paper relates to the small but growing literature on ATM network competition in the retail
banking industry that has emerged to capture the distinctive features of the industry. Massoud
and Bernhardt (2002) investigate the determinants of ATM pricing in a spatial model. Massoud
3
and Bernhardt (2000) extend this model to allow banks to choose ATM concentration. They ﬁnd
that banks choose ATM concentrations that are socially excessive. McAndrews (2001) is a related
model that solves for the banks’ ATM price-setting equations, allowing for both surcharges and
foreign fees, fees charged by the consumer’s bank for use of a non-bank ATM. Other theoretical
work has examined pricing incentives for banks versus non-banks where all ﬁrms have identical
ATM networks (Croft and Spencer 2003).
There has also been some empirical research on ATM competition and the banking industry.
Most closely related to this paper is the empirical work on surcharging. Hannan, Kiser, Prager, and
McAndrews (2003) examine the determinants of surcharging using a probit regression. They ﬁnd
that banks with many ATMs are more likely to surcharge, banks in markets with many ATMs are
less likely to surcharge, and banks are more likely to surcharge in states that allowed the practice
early. Prager (2001) investigates the eﬀect of surcharging on small banks by taking advantage of the
fact that prior to 1996, some states allowed ATM networks to maintain prohibitions on surcharging
while others did not. Examining the change in market share of small banks between 1987 and 1995,
she ﬁnds no signiﬁcant diﬀerence between surcharging states and non-surcharging states. However,
the study relies on a deﬁnition of “small” banks that is based on assets rather than ATM network
size. The author is also unable to observe ATM surcharge levels, so actual fee-setting behavior in
states that allowed surcharging is unknown.
Knittel and Stango (2003) examine the eﬀects of surcharge-induced incompatibility between
parts of the shared network. They use hedonic price regressions to conﬁrm their predictions that
incompatibility between a bank’s ATM cards and foreign ATMs lowers willingness to pay for deposit
accounts and strengthens the link between a bank’s deposit account prices and its network quality
through customer willingness to pay. A structural model is presented in Knittel and Stango (2004).
The authors estimate the parameters of a nested logit demand system for deposit account and ATM
services. They conclude that although incompatibility reduces consumer welfare, the increase in
ATM deployment that followed the widespread introduction of surcharging in 1996 sometimes
completely oﬀsets this welfare reduction. However, the authors do not endogenize the pricing of
deposit account services or the choice of ATM network size. Also, while their paper uses a broad
data set with many observations, it must accommodate missing data in several places. In contrast,
this paper uses a small data set with complete market data for each observation. It also endogenizes
deposit interest rates and considers the implications of banks’ ATM costs on investment.
Finally, Gowrisankaran and Krainer (2003) examine the eﬀects of surcharging using a structural
entry model. However, they do not consider the eﬀect of the bundling of ATM services with banking
deposit services, and each entrant is limited to owning a single ATM.
4
2 Industry Background
When banks ﬁrst began to introduce ATMs in 1969, the machines were accessible only to customers
of the bank. However, in order to increase convenience and lower costs, ﬁnancial institutions began
to share ATMs with one another and develop shared ATM networks. This process of sharing and
later, the consolidation of many shared networks and the introduction of the national networks,
eventually led to total technological compatibility. Each depositor’s ATM card could be used at
virtually every ATM. In the early days of ATMs, the machines were installed only in bank branches,
but installation of ATMs on city streets or in other locations such as grocery stores or shopping
malls soon appeared. The former type of ATM is now called an “on-premise” or “in-branch” ATM,
and the latter is called an “oﬀ-premise” or “remote” ATM.4
Many of the ATM networks initially banned surcharges by preventing ATM-owning member
institutions from directly charging consumers for the use of their equipment, but these networks
began to come under legal attack in the late 1980s by some network members who claimed that the
surcharge ban constituted price-ﬁxing. By April 1996, 15 states had adopted legislation overriding
network surcharge prohibitions. The national networks, Cirrus and Plus, lifted their bans on April
1, 1996, leading to the widespread introduction of surcharges throughout the U.S.
Since then, debate on the issue has continued. Some have argued that surcharges improve welfare
because they encourage investment in ATM networks. Indeed, ATM deployment has increased since
the introduction of surcharges. According to data from the 2003 EFT Databook, while the period
from 1992 to 1996 saw a 59 percent increase in national ATM deployment, there was a 96 percent
increase between 1996 and 2000. However, others have claimed that the incompatibility induced
by surcharges distorts the deposit market and is welfare-reducing. The U.S. Congress has held
hearings, and there have been attempts to re-ban or otherwise regulate surcharges at various levels
of government. Nevertheless, although the most successful attempts have been at the state and
municipal levels, federal courts have ruled that states and localities do not have the jurisdiction to
regulate the ATM fees of national banks. Two separate federal court rulings in 2002 voided bans
on surcharging in Iowa and in the cities of San Francisco and Santa Monica on these grounds.
Surcharges are the ATM fee on which this paper focuses, but there are other fees that pass between
the consumer’s bank and the ATM owner (the interchange fee) and between the consumer’s
bank and the network (the switching fee). Along with surcharges, the interchange fee provides
compensation to ATM owners for non-customer transactions. The switching fee provides compensation
to the shared network for routing non-customer transactions to the customer’s bank. These
fees are not faced by the consumer. In addition, although ﬁnancial institutions generally do not
charge customers for using their own ATMs, a consumer sometimes faces a foreign fee, a fee set by
4
See McAndrews (1991) for a review of the history of the creation of shared ATM networks.
5
his own bank for use of ATMs not owned by his bank. I do not directly include foreign fees in the
demand system in this paper, because surcharges are the fees on which the public debate centers,
and I lack data on foreign fees. They tend to be complicated, often being non-linear, and many
banks oﬀer a variety of schedules of foreign fees.
3 Data
3.1 Data Sources
The data set used in this paper is a cross-section from the state of Massachusetts in 2002. It
has been compiled from a number of diﬀerent sources. The 2002 ATM data was provided by the
Massachusetts Division of Banks. I calculate the total number of ATMs owned by each bank in
each market and classify each one as in-branch or remote. In the analysis below, I use only unique
ATM locations when calculating the size of a bank’s ATM network in order to focus on spatial
factors. I also assign each bank in each market a surcharge level.
Deposit data for each of the inside goods comes from the Summary of Deposits for 2002. This
provides deposits by branch and the addresses of the branches for each commercial bank and thrift.
By geocoding the branches, that is, assigning each a latitude and longitude, I am able to calculate
a bank’s total deposits in a local market. As discussed below, I use credit unions as the outside
good in the demand system, and the Financial Performance Reports of the National Credit Union
Association provide me with credit union deposit data.
I derive additional bank-level data from the Call Reports and Thrift Financial Reports. These
are statements of condition that ﬁnancial institutions ﬁle with the Federal Deposit Insurance Corporation
(FDIC) or the Oﬃce of Thrift Supervision (OTS), respectively. I estimate the bank-level
deposit interest rate as the ratio of interest expense on deposits to deposits. Similarly, I calculate
the bank-level loan rate as the ratio of revenue on loans and leases to loans and leases. The interest
and loan rates are both calculated as six-month rates. I also use the bank’s total number
of branches and employees to calculate employees per branch. Finally, I use several variables to
calculate instruments that will be described in section 6.5
Data on shared network interchange and switching fees are derived from the 2003 EFT Databook.
I use the average fees in 2002 for NYCE, the main regional shared network for Massachusetts.
These were 46.5c for interchange and 7.75c for switching. Finally, I use information from the 2000
Census for consumer data. This provides me with block group-level information on median income
levels, total population, and the latitudes and longitudes of the area centroids.6 Combining this
5
The interest and loan rates use expenses or revenues from January through June 2002. The other variables,
including total deposits, loans, and leases, are all calculated as of the end of June 2002.
6
Block groups are subdivisions of census tracts. They consist of sets of blocks within a census tract having the
6
information with the geocoded locations of the bank branches, I am able to calculate distances from
each block group to each bank branch. Appendix A provides additional details about the data.
3.2 Market Deﬁnition and Data Summary
I deﬁne a market’s geographic size as a metropolitan statistical area (MSA). The product market is
deﬁned as deposits at depository institutions. Deposits include checking, savings, and time deposits,
and depository institutions include commercial banks, thrifts, and credit unions while excluding
other ﬁnancial institutions such as mortgage or ﬁnance companies. Although credit unions are
a type of depository institution, it would be diﬃcult to model them as one of the inside goods.
Eligibility for participation in a credit union is often idiosyncratic, and for the purposes of my paper,
credit unions are modelled as the outside alternative. Appendix A contains more motivation and
details about the market deﬁnition.
These geographic and product-market deﬁnitions leave me with a total of 10 markets and
291 bank observations.7 Table 1 provides a brief summary of the markets. They vary in the
number of competitors and market shares. With 148 banks, the Boston market has the largest
number of competitors and the smallest average market share. Pittsﬁeld has the smallest number
of competitors, 8 banks. The market share of the outside good has a substantial range.
Table 2 displays some summary statistics on bank characteristics. The average number of bank
branches is 6.2, and the average number of ATMs is 10.1. These distributions are somewhat rightskewed,
with medians of 3 and 4, respectively. The standard deviation of the number of ATMs is
large, because most banks have moderate numbers of ATMs but a few, such as Fleet or Citizens,
have very large networks. Banks have, on average, 5.7 in-branch ATMs and 4.4 remote ATMs.
However, many banks have no remote ATMs at all: the median is zero. The average number of
employees per branch is 19.2. The interest rate has a mean of 1.2 percent, with a standard deviation
of 0.3 percent.
The distribution of surcharge levels across all banks that own ATMs is shown in Table 3. The
surcharges range from $0 to $2 and move in discrete increments of $0.25 or $0.50. The average
surcharge is $0.78. As was also true in the data used by Hannan, et al. (2003), most of the mass
of the distribution is at $0 and $1. For those banks setting a non-zero surcharge, there is limited
variation, but the overall industry standard is $1. The table also shows that those banks setting
higher surcharges (or, essentially, those banks who surcharge) tend to own larger percentages of the
shared ATM network. This is consistent with the theoretical predictions of Massoud and Bernhardt
same initial digit.
7
I drop from the analysis two very small banks that oﬀer no ATM services. Also, a few ﬁnancial institutions listed
in the Summary of Deposits for which depository services are clearly oﬀered only as part of the company’s wealth
management services are dropped.
7
(2002) and the empirical results of Hannan et al. (2003).
Consumer characteristics are measured at the block group level. Table 4 summarizes the size
and population of the block groups in each of the markets. As might be expected, the average
area is smaller in the more densely populated markets. With an overall average area of 1.14 square
miles and an average population of 1275, the level of measurement of consumer characteristics is
fairly ﬁne. Averaging over all bank observations and block groups, the mean distance to a bank is
12.56 miles, and this ﬁgure is as high as 17.41 miles in the Boston market. In contrast, the median
distance between households and their depository institutions has recently been measured at only
3 miles (Amel and Starr-McCluer 2002). Therefore, distance to bank branches is clearly something
that consumers consider when choosing a bank. In addition, Table 4 shows that the variation across
block groups of average bank distance is substantial: in the various markets, the standard deviation
is usually between one-third and one-half of the mean. Because of the importance of distance to
consumers and the extent of the variation in distances caused by geographic diﬀerentiation, it will
be important to develop a model that incorporates this block group-level spatial diﬀerentiation.
Averaging over all block groups, median income varies across markets from about $39,000 to
$60,000. The variation across block groups within markets is also quite large. This implies that
it will be useful for the model of consumer behavior to account for the impact of heterogeneous
income levels on both consumer preferences and deposit market shares.
4 Empirical Model
In order to investigate the eﬀects of partial incompatibility in the banking industry through ATM
retail interconnection pricing, I set out a model of consumer and ﬁrm behavior. This model seeks
to incorporate the most important aspects of the industry in a way that keeps the burden of
estimation reasonable. It assumes that consumers choose banks to maximize their utility given
their own characteristics and the characteristics of each alternative. Banks are assumed to maximize
proﬁts in a two-stage process with simultaneous moves in each stage. First, they choose the size
of their ATM networks given their expectations about their rivals. Second, they select interest
rates to maximize proﬁts conditional on ATM networks in a Nash equilibrium. The use of an
estimated utility function and an interest rate-setting condition will allow for counterfactual policy
experiments in which consumers reallocate themselves across banks and new pricing equilibria are
computed. The use of a network selection stage also allows investment incentives and the optimality
of the current network equilibrium to be considered.
8
4.1 Consumer Behavior
The model assumes that consumers choose a single bank for depository services. The demand side
of the model therefore follows the discrete-choice literature and adapts the methods proposed by
Berry, Levinsohn, and Pakes (1995) (hereafter referred to as BLP). Some modiﬁcation of the usual
discrete-choice models is required for an application to retail banking, because the fundamental
quantity variable is a bank’s amount of total deposits rather than its number of customers. The
model will therefore assume that individual consumers are endowed with a level of deposits, and
that each consumer places her deposits in her bank of choice.8 Since individual deposits are not
observed, a simple assumption will be used to relate deposits to consumer observables. Deposit
market shares will be determined by aggregating over the expected deposits of individual consumers.
Although other researchers have modelled demand for depository services as a discrete choice,
this paper is the ﬁrst to use consumer-level location and demographic data. It is also the ﬁrst to
use a model that makes direct predictions about a bank’s share of deposits as distinct from its
share of consumers. These two market shares clearly diﬀer when consumers have heterogeneous
deposits, and this distinction will aﬀect the model’s implications. For example, consider two banks
in the same market, where one bank has branches located in a high-income area and the other has
branches in a low-income area. If income is related to deposits but consumers are instead assumed
to have homogeneous deposits, the large increase in deposits induced by an increase in interest
rates by the bank in the high-income area will be rationalized by a high consumer sensitivity to
interest rates, whereas the small increase in deposits induced by the same change in interest rates
by the bank in the low-income area will be rationalized by a low sensitivity. Accurate measurement
of these eﬀects are important when policy experiments are to be considered.
My model speciﬁes a consumer utility function that allows for product diﬀerentiation and heterogeneous
preferences driven by consumer characteristics. I assume that the utility of individual
i for bank j in market m is
Uijm = U(xjm, rjm, ξjm, him, ijm, θ),
where xjm is a vector of observable bank characteristics including ATM network characteristics
determined by each bank’s number of ATMs (n), rjm is bank j’s deposit interest rate, ξjm is a
bank unobservable, him is a vector of observable consumer characteristics, ijm is a consumerand
bank-speciﬁc unobservable, and θ is a vector of parameters to be estimated. Note that him
will account for consumer locations. This allows the model to include geographic information
speciﬁc to a consumer-bank combination, similarly to Davis (2001). The inclusion in the model
8
The assumption that deposits are not a choice variable is clearly a simpliﬁcation, but it is made in order to keep
the problem tractable. It is fairly reasonable, since one would expect individual deposits to be driven by transaction
needs and fundamentally by income. It is also a standard assumption in the banking literature.
9
of the structural disturbance, ξ, introduces a traditional simultaneity problem which should lead
to correlation between interest rates and this disturbance term. This endogeneity problem will be
handled below through the use of instruments.
Each consumer chooses the bank that maximizes her utility. Conditional on hi, the set of
consumer unobservables, , that leads individual i to choose bank j can be implicitly deﬁned as
Aijm = { : U(xjm, rjm, ξjm, him, , θ) ≥ maxk U(xkm, rkm, ξkm, him, , θ)}, where k indexes banks
in market m. Then if has distribution f( ), the probability that consumer i chooses bank j is
Pjm(him) =
∈Aijm
f( )d . (1)
Therefore, if h has distribution g(h), bank j’s share of consumers is
qjm = qjm(x, r, ξ, θ) =
h
Pjm(h) g(h)dh. (2)
However, in the banking industry, the primary quantity variable of interest is not the market
share of consumers, but rather the share of deposits. Therefore, the deposit market share will be
the integral over the distribution of consumer characteristics of expected deposits divided by total
market deposits. Denote individual i’s deposits by depi and the joint distribution of h and dep
by g(h, dep). Then if the inside and outside banks in market m are indexed by k = 0, ..., Jm, the
market share of bank j in market m is sjm, where
sjm = sjm(x, r, ξ, θ) =
h,dep Pjm(h) dep g(h, dep)dh ddep
Jm
k=0 h,dep Pkm(h) dep g(h, dep)dh ddep
. (3)
4.2 Firm Behavior
4.2.1 Interest Rate Choice
Before discussing the ﬁrst stage in which banks choose network size, I discuss the second stage in
which interest rates are set. Each commercial bank or thrift in my sample is assumed to set its
deposit interest rate to maximize proﬁts conditional on its ATM network and the networks of its
rivals. These banks take in deposits, which are primarily used to generate revenue through the
funding of various credit instruments. They are therefore multiproduct ﬁrms, however, this paper
focuses on the deposit side of the banking industry and does not explicitly model any decisions on
the lending side. Dropping the subscript for market m for ease of notation, bank j’s proﬁt is the
following:
Rj =(lj − rj − mcj) × Dsj(x, r, ξ, θ) + surcharge revenuej+
net interchange revenuej − switching costsj + foreign fee revenuej − ηj,
10
where the rate bank j receives on loans is lj, its deposit interest rate is rj, its marginal cost is mcj,
its ﬁxed cost is ηj, and D is total market deposits. This formulation assumes constant marginal
costs of deposit-taking, but the ﬁxed cost term allows for economies of scale. The terms involving
ATM fees are deﬁned below, and it is assumed that ATM transactions have no marginal cost.9
I assume the existence of an interior Nash equilibrium in deposit interest rates. When each
bank maximizes its proﬁts given the interest rates of its competitors, its interest rate satisﬁes the
following ﬁrst order condition:
−Dsj+D(lj − rj − mcj) ×
∂sj(x, r, ξ, θ)
∂rj
+
∂
∂rj
surcharge revenuej+
net interchange revenuej − switching costsj + foreign fee revenuej = 0.
(4)
The ﬁrst two terms of this equation have the usual intuition: by increasing the deposit interest
rate slightly, a bank loses the amount of the increase on all of the deposits it currently has, but it
gains an amount equal to the loan rate net of its interest expense and other costs on its increased
demand. As will be clear when surcharge revenue, net interchange revenue, switching costs, and
foreign fee revenue are deﬁned in equations (12) and (13), the eﬀect embodied in the third term is
that an increase in the deposit interest rate raises a bank’s share of consumers, qj, which decreases
revenue from the ﬁrst three types of fees and increases revenue from foreign fees.
4.2.2 ATM Network Choice
Moving back to the ﬁrst stage, each bank simultaneously chooses its ATM network size given its
expectations about its rivals. Bank j’s total proﬁt is
πj = π(yj, nj, n−j, bj, δ) = R(yj, nj, n−j) − (bjδ + νj)nj,
where R(yj, nj, n−j) denotes the second-stage returns to ﬁrm j when its network size is nj, its
rivals’ is n−j, and the vector of all other relevant variables is yj; bj is a vector of cost shifters; δ are
parameters to be estimated; and νj captures cost diﬀerences across banks that are unobserved by
the econometrician but known to the bank. For example, a bank may be able to negotiate favorable
terms with a contractor that maintains the machines’ cash levels.10
Bank j’s expected total proﬁt is
E[π(yj, nj, n−j, bj, δ)|Ij] = E[R(yj, nj, n−j)|Ij] − (bjδ + νj)nj, (5)
9
According to industry sources, ATM costs are almost entirely ﬁxed with respect to the number of transactions.
See, e.g., Dove (2002).
10
It is also possible to include econometrician measurement error. For example, if πo
j (·) = πj(·) + uj,nj , where
πo
j (·) is observed proﬁts and uj,nj is measurement error, then all the analysis goes through as long as uj,nj is mean
zero conditional on the instruments introduced in section 6.2.
11
where Ij is the agent’s information set at the time of the decision. Note that there can be prediction
error due to randomness in observed proﬁts that is not known at the time decisions are made. For
example, a bank’s expectation of n−j may diﬀer from the outcome. This expectational error,
denoted ej,nj , is mean zero conditional on the information set by construction, i.e., E(ej,nj |Ij) = 0,
where ej,nj = R(yj, nj, n−j) − E[R(yj, nj, n−j)|Ij].
I assume that each bank maximizes its expected total proﬁt, but rather than approximating
ATM network size as a continuous variable and taking ﬁrst order conditions, I use a method
that preserves the discrete, ordered nature of the variable.11 A necessary condition for proﬁt
maximization is that for all banks j, bank j’s expected proﬁt from choosing nj is at least as great
as its expected proﬁt from choosing nj − 1 or nj + 1. This necessary condition is also suﬃcient
when proﬁts are concave in nj. This turns out to be nearly always true at the estimated parameter
values.12
Therefore, an optimal choice of nj satisﬁes
E[π(yj, nj, n−j, bj, δ)|Ij] ≥ E[π(yj, nj − 1, n−j, bj, δ)|Ij]
E[π(yj, nj, n−j, bj, δ)|Ij] ≥ E[π(yj, nj + 1, n−j, bj, δ)|Ij].
From equation (5), this implies the following condition:
E[R(yj, nj, n−j) − R(yj, nj − 1, n−j)|Ij] ≥ bjδ + νj ≥ E[R(yj, nj + 1, n−j) − R(yj, nj, n−j)|Ij]. (6)
Note that for each bank, calculation of R(yj, nj − 1, n−j) or R(yj, nj + 1, n−j) involves use of the
structural model to compute a new vector of equilibrium interest rates.
5 Empirical Implementation
Given the full behavioral model above, the empirical speciﬁcation can now be formulated. The
empirical implementation of consumer preferences and demand is discussed ﬁrst, followed by the
supply side. To brieﬂy summarize, the demand model will assume that a consumer’s utility from
choosing a bank depends on the bank’s characteristics, the consumer’s characteristics, and characteristics
of the bank-consumer combination. On the supply side, the primary speciﬁcations for
both the marginal cost of deposit-taking and the cost of an ATM network will be simple constant
functions.
11
A continuous ﬁrst order condition method is considered in section 7 as a robustness check, and it yields similar
results. Pakes, Porter, Ho, and Ishii (2005) discusses alternative estimation strategies including the discrete choice
logit method. It also discusses why traditional ordered choice models such as ordered logit or probit are unfeasible.
12
The results are quite similar when the choice set is expanded to nj +/- 1 and nj +/- 2.
12
5.1 Demand
Preferences
The demand side is estimated using two diﬀerent functional forms for utility. The primary
speciﬁcation allows for heterogeneity in both consumer and ﬁrm observable characteristics. The
conditional indirect utility function of individual i for banks j = 1, ..., Jm in market m takes the
following form:
Uijm = λdepimrjm + dijmβ + xjmγ + amρ + ξjm + ijm. (7)
depim is individual i’s deposits, dijm is a vector of individual-speciﬁc branch distance variables for
bank j, xjm is a vector of bank-speciﬁc observables including ATM network characteristics, am is a
set of market dummy variables for market m, ξjm is a bank unobservable, and ijm is an individualand
bank-speciﬁc unobservable. The utility of the outside alternative, credit unions, is given the
form
Ui0m = ξ0m + i0m, (8)
where we can normalize ξ0m to zero.
Included in the ﬁrst term of the utility function is depimrjm, the interest revenue that individual
i receives at bank j. Since individual deposits are not observed in the data, I relate deposits to
an observable variable, income, with the following simple assumption: depim = αyim. I use the
empirical joint distribution derived from the Census data for geographic location and income.
Since there is no error included in this relationship, it is implicitly assumed that any individuallevel
error is averaged out in the block group-level data on income. As noted below, it is helpful
to keep the relationship simple since the deposits variable enters both the utility function and the
share equation.
I include in xjm bank j’s employees per branch, its number of ATMs, and the surcharge cost
variable that will be deﬁned below. Each consumer has a location, which determines his distance
to each bank branch. I assume that the distance a consumer has to travel to reach a bank branch
enters his utility quadratically. Since spatial diﬀerentiation is a very important aspect of banking
decisions, the utility function accounts for the consumer’s local geography. Included in dijm is
d1ijm (the distance for consumer i to the nearest branch of bank j), d12
ijm, d2ijm (the distance for
consumer i to the second nearest branch of bank j multiplied by a multi-branch-bank indicator),
d22
ijm, and onebranch (an indicator for single-branch banks).13 In Appendix C, I allow consumers
to have varying valuations for presence in multiple states, and I ﬁnd no signiﬁcant eﬀect.
13
Further branch distance variables do not enter signiﬁcantly. I have also tried including the size of the total
branch network to check that these branch distance variables adequately control for the convenience of the branch
network. The coeﬃcient on the added branch network variable is insigniﬁcant, and the coeﬃcient on the key ATM
network variable, the surcharge variable deﬁned below, is insigniﬁcantly diﬀerent.
13
The bank-speciﬁc unobservable, ξjm, accounts for various unobserved aspects of bank quality.
This may include things such as the variety of account oﬀerings, the quality of customer service
training, or bank advertising. It also contains account fees which I do not include in the utility
function since this data is not available for the thrifts in my sample.14 Finally, the consumer-speciﬁc
unobservable, ijm, accounts for idiosyncratic preferences such as preferences for community banks.
One would also expect the branch geography near a consumer’s workplace to ﬁgure in his bank
choice, but since I do not observe workplace locations, this factor will go into ijm.
The only variable in the utility function that remains to be deﬁned is the surcharge variable in
the xjm vector. This should represent an individual’s utility cost from paying ATM surcharges. I
use a reduced-form representation of the expected cost to a customer of bank j due to surcharges.15
We can denote the percentage of the ATM network owned by bank j as fj and bank j’s surcharge
to non-j customers as cj, and suppose that consumers use a foreign bank’s ATMs in proportion to
that bank’s share of the total network. Then if the number of transactions performed by a customer
of bank j in market m at foreign bank k equals Tfkm, where T is some constant, the surcharge costs
faced by that customer are: T Jm
k=1,k=j fkmckm = T Jm
k=1 fkmckm − Tfjmcjm = tm − Tfjmcjm,
where tm = T Jm
k=1 fkmckm. The ﬁrst summation does not include bank j’s surcharge, because
bank-j customers are exempt from them. tm is a market-speciﬁc constant that applies to all
choices, and since it enters utility linearly, it cancels out.16 Taking the negative of the remaining
term and dropping T, we can deﬁne as an element of xjm surchjm = fjmcjm, which we expect
to have a positive coeﬃcient. The surchjm variable is a measure of the surcharges a consumer
does not have to pay when she chooses bank j. Although this variable does not capture the price
versus convenience trade-oﬀ consumers will face when making decisions about ATM use, it has
the advantage of accounting in a simple way for the consumer’s ex ante knowledge at the time
of choosing a bank about all competitors’ ATM network sizes and surcharges. Since, in reality,
consumers will sometimes walk to a less convenient ATM to avoid a high surcharge or forgo use of
an ATM altogether, this expected surcharge cost variable can be interpreted as a combination of
literal surcharge costs and other losses of utility due to inconvenience.
The ATM network size and surcharge variables deﬁned above are calculated based on the entire
market, with the implicit assumption that consumers travel and experience ATM needs throughout
14
One would expect individuals with high deposit levels to have a low sensitivity to these fees, and therefore that
market demand would also be fairly insensitive. Dick (2002) indeed ﬁnds that demand elasticities with respect to
service fees are quite low in MSA markets, the type of market I examine.
15
Alternatively, one could explicitly model an additional decision-making stage in which consumers choose ATMs
based on their bank aﬃliation, ATM prices, and the distribution of all ATMs. Unfortunately, bank- or ATM-level
data on the demand for ATM transactions is unavailable, so this exercise would be extremely demanding of the
existing data.
16
I assume that tm enters linearly in the utility function for all j = 0, 1, ..., Jm. Otherwise, it would simply go into
the market dummies.
14
the market. This seems reasonable, because the markets are deﬁned based on social and economic
integration, however Appendix C brieﬂy explores the robustness of the results to the use
of ATM variables that are deﬁned more locally. I also consider the eﬀect of no-surcharging alliances.
Market Shares
In order to be able to analytically integrate over the consumer unobservables, , I make the
traditional assumption that the ij’s are distributed i.i.d. type-I extreme value over individuals
and products. From equations (1), (7), and (8), this means that the probability that individual i
chooses bank j is the following:
Pjm(him) = Pjm(depim, dijm) =
exp(λdepimrjm + dijmβ + xjmγ + amρ + ξjm)
1 + Jm
k=1 exp(λdepimrkm + dikmβ + xkmγ + amρ + ξkm)
. (9)
Bank j’s share of consumers, qjm, follows directly from these probabilities, equation (2), and
the distribution of individual characteristics (deposits and branch distances). In order to calculate
deposit market shares, we need these probabilities as well as the distributions of individual deposits
and characteristics. The assumption above that depim = αyim is particularly convenient here since
the relationship has only one parameter which cancels out of the market share equation. The
market share equation can be calculated from equation (3) as the following:
sjm = sjm(x, r, ξ, θ) =
y,d Pjm(y, d) · y · g(y, d) dy dd
Jm
k=0 y,d Pkm(y, d) · y · g(y, d) dy dd
(10)
where
Pjm(yim, dijm) =
exp(φyimrjm + dijmβ + xjmγ + amρ + ξjm)
1 + Jm
k=1 exp(φyimrkm + dikmβ + xkmγ + amρ + ξkm)
and φ = λα. It is clear that only φ will be identiﬁable and not λ or α separately. It should also be
noted that identiﬁcation does not depend on the use of this simple relationship between deposits
and income. For example, random distributions with multiple parameters that need to be estimated
can be used, and one such speciﬁcation is explored in Appendix C. However, this increases the
computational burden of estimation substantially.
A Simpliﬁcation: Logit Demand
It is useful to consider a situation in which consumers have homogenous deposit levels, and
consumer heterogeneity in preferences is restricted to enter the utility function only through the ij
term. These assumptions lead to the logit demand model of McFadden (1974). Such an alternative
utility function can be speciﬁed as:
Uijm = ˜λrjm + ˜djm
˜β + xjm˜γ + am ˜ρ + ξjm + ijm,
15
where ˜d is a weighted average of the vector of variables in d over all block groups and the weights
are block group population. For example, d1ijm, the distance for a person in block group i in
market m to the nearest branch of bank j, is averaged over all block groups in market m, weighting
by block group population. This utility function, along with homogenous deposits, implies that
deposit shares (equation 10) are equal to shares of consumers and simplify to the following:
sjm = Pjm =
exp(˜λrjm + ˜djm
˜β + xjm˜γ + am ˜ρ + ξjm)
1 + Jm
k=1 exp(˜λrkm + ˜dkm
˜β + xkm˜γ + am ˜ρ + ξkm)
. (11)
Whereas the full model described above calculates individual choice probabilities (probabilities
of choosing each bank) and then integrates over the distribution of consumers to arrive at deposit
market shares, this speciﬁcation essentially integrates over the individual-speciﬁc variables ﬁrst and
then uses these averaged variables to calculate the choice probabilities of a market-level representative
agent. This is conceptually incorrect, and the logit model also has the well-known shortcoming
of generating potentially unrealistic substitution patterns. Nevertheless, it is useful as a a familiar
basis for comparison to the full model that correctly aggregates demand over heterogenous consumers.
The logit model also provides a robustness check of the full model’s assumption about the
distribution of individual deposits.
5.2 Supply
5.2.1 Interest Rate Choice
In the interest rate choice stage of the supply side, the marginal cost function for deposit-taking
remains to be speciﬁed. Researchers have traditionally included physical product characteristics
as determinants of marginal costs, however, there are few such characteristics in a service industry
such as retail banking. Therefore, the following simple marginal cost function is used:
mcjm = w + ωjm,
where the constant, w, is a parameter to be estimated, and ωjm is an unobservable source of cost
diﬀerences. As discussed in Appendix C, more complex cost functions were considered, including
ones that allowed for per-customer ﬁxed costs. However, the additional terms were not signiﬁcant.
In order to specify the revenue from ATM transactions, I rely on the simpliﬁed model of consumer
ATM behavior presented above in which consumers use a foreign bank’s ATMs in proportion
to that bank’s share of the total network. Therefore, since a bank receives its surcharge and the
interchange fee each time a non-customer uses one of its ATMs,
surcharge revenuej + interchange revenuej = TLm(1 − qjm)fjm(cjm + int), (12)
16
where T is a scaling factor, L is the number of consumers in the market, qj is bank j’s share of
consumers, fj is its share of ATMs, cj is its surcharge, and int is the network-wide interchange fee.
Also, since a bank pays out the interchange fee and the switching fee and receives the foreign fee
each time a customer uses a foreign ATM,
interchange costsj + switching costsj − foreign fee revenuej =
TLmqjm(1 − fjm)(switch + int − forjm),
(13)
where switch is the network-wide switching fee, and forj is bank j’s foreign fee.
This implies the following ﬁrst-order condition from equation (4) for the Nash equilibrium in
interest rates:
−Dmsjm + Dm(ljm − rjm−w − ωjm)
∂sjm(x, r, ξ, θ)
∂rjm
− TLmfjm(int + cjm)
∂qjm(x, r, ξ, θ)
∂rjm
− TLm(1 − fjm)(switch + int − forjm)
∂qjm(x, r, ξ, θ)
∂rjm
= 0.
(14)
Unfortunately, I don’t observe foreign fees. In order to account for them in the proﬁt function, I
assume a level of $1 for each bank.17 Also, T can be estimated from this equation, but in the actual
estimation, the parameter is not estimated with enough precision to be useful. Therefore, I apply
an estimate of the average number of ATM transactions per period and the percentage of foreign
transactions to calibrate T.18
These assumptions about the value of T and the level of foreign fees are made purely to ensure
that the level of calculated proﬁts due to ATM fees is roughly correct. They can only aﬀect the
demand system through their impact on optimal interest rates, and there will likely be very little
impact through this channel. Since revenues from deposit-taking are metered by dollars of deposits
whereas revenues from ATM transactions depend on numbers of customers, the sheer diﬀerence in
scale between these ﬁgures limits the inﬂuence of direct ATM revenue on interest-rate behavior.19
17
The average foreign fee at banks and savings institutions in 2001 was $1.17 according to the Federal Reserve
Survey of retail fees (Hannan 2002).
18
According to the 2003 EFT Databook, there were 3308 monthly transactions per ATM in the U.S. in 2002.
Applying this ﬁgure to the total number of ATMs owned by banks in the sample and dividing by total population
yields 12.83 transactions per person in a six-month period. In contrast to the count of ATMs used in all other cases,
the total number of ATMs used here includes multiple ATMs owned by a bank in a single location. In addition, a
2002 survey of consumers conducted by Pulse found that the total percentage of foreign transactions is approximately
25 percent. Thus, since I predict that a bank-j consumer makes T(1 − fj) foreign transactions, I ﬁnd the adjustment
factor, b, such that T = 12.83b and the total predicted percentage of foreign transactions is 25 percent. This implies
b = 0.28 and T = 3.6. Clearly, this adjustment weights total ATM transactions away from foreign transactions and
towards own-bank transactions.
19
In untabulated results, it can also be seen that the parameters are quite robust to the value of T. E.g., using a
value of T twice as large leaves the cost parameter unchanged to the third digit.
17
5.2.2 ATM Network Choice
In the initial ATM network choice stage of the bank’s problem, the function R(yj, nj, n−j), the
returns to bank j when its network size is nj and its rivals’ is n−j, has now been fully speciﬁed,
but the cost of this network remains to be speciﬁed. The small size of the data set requires a focus
on parsimonious speciﬁcations, so the main speciﬁcation assumes a constant per-ATM cost of the
network, (δ + νj)nj, where νj is the cost unobservable introduced in section 4.2.2. Alternative
speciﬁcations for ATM costs are also considered in Appendix C. Equation (6) then implies that
E[R(yj, nj, n−j) − R(yj, nj − 1, n−j)|Ij] ≥ δ + νj ≥ E[R(yj, nj + 1, n−j) − R(yj, nj, n−j)|Ij]. (15)
6 Estimation
Estimation of the behavioral model will implement two major methodologies. First, the demand
and deposit-market cost parameters will be estimated using generalized method of moments (GMM)
with a non-linear optimization routine and simulation methods. Second, the ATM network cost
parameters will be estimated using a new procedure, the method of moments with inequality
constraints. Readers not interested in the econometric details can skip to the results in section 7.
6.1 Demand and Interest Rate Choice
In order to estimate the parameters from the demand model and the cost of deposit-taking, I
assume that the data-generating process satisﬁes the following covariance restrictions:
E(ξj(θ0)zD
j ) = E(ωj(θ0)zS
j ) = 0, 20 (16)
where zD
j is a vector of demand instruments, zS
j is a vector of supply instruments, and θ0 is the true
value of the parameter vector. These moments are stacked, and estimation proceeds with GMM.
The parameters are identiﬁed using exogenous variation in the distribution of consumer income
and consumer locations relative to bank locations as well as variation across banks and markets in
deposit market shares and bank characteristics.
Because the methods used here are well-understood, I will describe them brieﬂy in this section
and relegate the details to Appendix B. The estimation encounters two diﬃculties. First, the
market shares are nonlinear functions of ξ. However, they can be inverted to solve for ξ using
the contraction mapping established in BLP, since this contraction mapping is also valid given
the market share in equation (10). Second, there exists no analytic expression for the integrals in
equation (10), so market shares are simulated using s random draws of individual consumers for each
20
This notation includes in θ all parameters to be estimated from the demand and interest rate choice model,
including the marginal cost parameters, w.
18
market.21 Then the simulated share of consumers of bank j is qsim
j (x, r, ξ, θ) = (1/s) s
i=1 Pj(yi, di),
and the simulated deposit share of bank j is
ssim
j (x, r, ξ, θ) =
s
i=1 Pj(yi, di) depi
J
k=0
s
i=1 Pk(yi, di) depi
=
s
i=1 Pj(yi, di) yi
J
k=0
s
i=1 Pk(yi, di) yi
,
where Pj is deﬁned in equation (10). This implies a simulated version of the moment condition.
Estimating the Logit Demand Model
Estimation of the logit speciﬁcation of demand is substantially less complicated because the form
taken by the market share allows it to be linearized through a simple transformation. Equation
(11) implies the following expression:
ln(sj) − ln(s0) = ˜λrjm + ˜djm
˜β + xjm˜γ + am ˜ρ + ξjm (17)
The parameters, ˜λ, ˜β, ˜γ, and ˜ρ are estimated with standard linear instrumental variables techniques.
Instruments
These estimation procedures require suitable demand and supply instruments, zD and zS, respectively.
We require that elements of zD are uncorrelated with ξ at θ = θ0. Variables in zS
must similarly be uncorrelated with ω. As is traditional in demand analyses, the vector of ﬁrm
characteristics, x, is included in the demand instruments. r cannot be included because we expect
interest rates to be endogenous. In order to include instruments for the individual branch distance
variables, I also include ˜d, the vector of variables that averages over consumer branch distances,
weighting by block group population.
Cost-shifters provide another source of demand-side instruments. Firms should account for
costs when setting the interest rate, leading to correlation between these variables and the interest
rate, but many cost-shifters should not be correlated with demand. I include four such variables in
zD: expenses on premises and equipment, other expenses, a credit risk cost variable, and the wage.
The ﬁrst three of these variables are normalized by assets. Expenses on premises and equipment
and other expenses are both operating cost variables. Premises and equipment expenses include
expenses on utilities, janitorial services, repairs, furniture, etc. Other expenses include legal fees,
postage, deposit insurance assessments, directors’ fees, etc. The credit risk cost variable employed
is provisions for loan and lease losses. Finally, wages are calculated based on the bank’s labor
expenses and the number of employees.22
21
I use 200 random draws of block groups for each market, weighting the probability of drawing a block group by
its population.
22
I consider the possibility that expenses on premises and equipment, other expenses, or wage is correlated with
the demand-side unobservable, ξ, by estimating the model without these instruments and examining the correlations
between the predicted ˆξ and each instrument. In each case, the correlations are small and statistically insigniﬁcant.
19
The inclusion of a bank’s surcharge cost variable in the demand instruments as part of x
warrants further discussion. Price variables are traditionally thought to be endogenous, and indeed,
I instrument for interest rates throughout the empirical analysis. However, as is evident in Table
3, there is very little variation in surcharge levels. As was found in Hannan et al. (2003), the
surcharging decision is primarily one of whether to surcharge or not, rather than what level to
choose. In addition, I include in the demand equation what I expect to be the primary determinants
of a bank’s surcharge level: market-level dummies and a measure of the bank’s ATM network size.
ATM network size is also included in x, and one might similarly worry about the endogeneity of
this variable. A bank’s ATM investment equation should have predicted market share in it, which
could lead to correlation between ATM network size and the demand unobservable. However, in
many ways this variable is just another product characteristic, and like most demand studies, I
take product characteristics as predetermined. In Appendix C, I provide a robustness check of the
assumption of the exogeneity of ATM network size and the surcharge price variable.
6.2 ATM Network Choice
To estimate the parameters and their conﬁdence intervals from the ATM network choice equation,
I use the methods described in Pakes, Porter, Ho, and Ishii (2005) (hereafter referred to as PPHI).
This paper provides simple techniques for estimation from a set of inequality restrictions given
appropriate instruments. These techniques allow for multiple potential equilibria and endogenous
regressors. Identiﬁcation of the parameters is based on the condition that a bank’s expected proﬁts
from its observed choice are greater than its expected proﬁts from alternative choices.
We require instruments, zj, such that zj ∈ Ij, the agent’s information set, and the cost unobservable
is mean zero conditional on the instruments, that is, E(νj|zj) = 0. This condition, with
equation (15), implies that
E[R(yj, nj, n−j) − R(yj, nj − 1, n−j)|zj] ≥ δ ≥ E[R(yj, nj + 1, n−j) − R(yj, nj, n−j)|zj].
Then, if g(z) is a positive-valued function of z,
E R(yj, nj, n−j) − R(yj, nj − 1,n−j) g(zj) ≥ δ g(zj) ≥
E R(yj, nj + 1, n−j) − R(yj, nj, n−j) g(zj) .
(18)
Each element of g(z) contributes two moments that deﬁne an upper and lower bound for δ.
Since banks are interacting agents within a market, to use this restriction in estimation, I create
market-level observations by averaging over banks within markets. Before taking the average over
markets, each market average is weighted by the square root of the number of banks in the market.
This varies signiﬁcantly across markets, and we expect less noise in the market average for markets
20
having many banks. Let Jm denote the number of banks in markets indexed by m = 1, ..., M. The
sample analogue of the conditions in equation (18) is therefore
1
M
M
m=1
√
Jm
Jm
Jm
j=1
R(yj, nj, n−j) − R(yj, nj − 1, n−j) g(zj) ≥
1
M
M
m=1
√
Jm
Jm
Jm
j=1
δ g(zj) ≥
1
M
M
m=1
√
Jm
Jm
Jm
j=1
R(yj, nj + 1, n−j) − R(yj, nj, n−j) g(zj) .
All δ that satisfy this set of inequalities are included in the feasible set of parameters. If there
exists no such δ, we ﬁnd the point that minimizes the sum of the absolute value of the amount by
which each inequality is violated.
Conﬁdence intervals are constructed in two diﬀerent ways as in PPHI. First, I calculate a “conservative”
(1 − α) level conﬁdence interval. This interval is the set of parameters such that, given
the distribution of the sample moments, all moments are satisﬁed with probability (1 − α). I also
construct an alternative, simulated conﬁdence interval for the parameters. I take simulation draws
from an approximation of the limit distribution of the data and estimate the parameters at each
draw. This provides a simulated distribution of parameters from which conﬁdence intervals can
be directly calculated. PPHI contains additional details on the properties and construction of the
estimator.
Instruments
I use a constant term, market population, the number of banks in a market, and the number of
branches of a bank in a market as z variables to construct instruments. These variables are all in
the information sets of the agents at the time they make their ATM network choices. Furthermore,
they can reasonably be assumed to be independent of the νj cost disturbances in the proﬁt function.
I use simple transformations of these variables as the actual instruments, g(zj). For each element of
zj, I use one indicator function equal to one if zj exceeds its mean, and another indicator function
equal to one if zj is less than or equal to its mean.
7 Results
I ﬁrst estimate the logit demand model, because this speciﬁcation is simple to compute and will
provide a basis for comparison to the full model. Then I estimate the full model which integrates
individual demand over the distribution of consumers. Table 5 shows the results of estimating the
logit demand model, where the dependent variable is ln(sj) − ln(s0). The coeﬃcients from OLS
estimation are in the left column, and the IV results are in the right column. The eﬀect of the
endogeneity of the interest rate is immediately evident. The interest rate coeﬃcient becomes larger
21
by a factor of 3 when instruments are used. In the OLS version, we expect there to be a negative
correlation between the demand unobservable and the interest rate that biases the coeﬃcient
downwards. The mean interest rate elasticity implied by this coeﬃcient in the IV speciﬁcation is
2.49.23 This lies between the mean elasticities estimated by Dick (2002) using a logit model in U.S.
MSA markets (1.49) and all U.S. markets (5.86). In addition to diﬀerences in speciﬁcation and
instrumentation, the diﬀerences could result from higher elasticities in Massachusetts MSAs than
in MSAs in the nation as a whole or from the addition of thrifts in my data.
Recalling that surchj represents surcharge costs that a consumer avoids by choosing bank j, the
coeﬃcient on surch is positive in the IV speciﬁcation, as expected, and is statistically signiﬁcant.
It implies an average elasticity of 0.123. The number of ATMs is also estimated to enter utility
positively. The quadratic speciﬁcation for branch distance ﬁts fairly well, but three out of four
coeﬃcients are insigniﬁcant at the ﬁve-percent level. The signs of the coeﬃcients imply that a
consumer’s utility is decreasing at a decreasing rate in the distance to the nearest and secondnearest
bank branch. The dummy for single-branch banks picks up a large negative eﬀect, and the
coeﬃcient on employees per branch is positive as expected. The ﬁt of these regressions is quite
good. The OLS regression has an R2 of 0.85.
Turning now to the full demand and interest-rate-setting model, the ﬁrst column of Table 6
displays the results from estimating the demand side alone, and the second column shows the
results of estimating the demand and supply sides jointly. Almost no coeﬃcients achieve statistical
signiﬁcance when demand is estimated alone. However, as expected, joint estimation of demand
and supply increases the precision of the estimates, making each coeﬃcient signiﬁcant at the ﬁvepercent
level or lower. Each standard error is reduced by joint estimation, and some are reduceed
by a substantial amount. The addition of the supply side does not have a large impact on the
demand coeﬃcients. Indeed, there are no signiﬁcant diﬀerences, which provides support for the
supply-side speciﬁcation.
The coeﬃcient on the interaction between the interest rate and deposits implies an average
elasticity across banks of 4.20 with a median of 3.62. This is somewhat higher than the average
interest rate elasticity from the logit speciﬁcation of 2.49. The higher elasticities may result from
the fact that the shares in the full model assign a larger weight to consumers with higher income,
who are predicted to be more sensitive to interest rates because of their larger deposit balances.
Comparing the other coeﬃcients to the logit estimates in Table 5, the coeﬃcients on surch, the
dummy for single-branch banks, and employees per branch are very similar. The biggest diﬀerence
is in the branch distance coeﬃcients, where the magnitudes in the full model are much larger than
in the logit model. To investigate this further, the third column of Table 6 presents estimates from
23
See Appendix D for details on the calculation of elasticities for both the logit and the full demand model.
22
the full model where the block group-speciﬁc distance variables have been replaced by the averaged
versions ( ˜d) used previously in the logit speciﬁcation. It is evident that the model picks up a much
stronger eﬀect from branch distance when distance is measured accurately at the block group level.
The distance coeﬃcients from this alternative speciﬁcation closely resemble the coeﬃcients from
the logit model. The results using the block group-speciﬁc variables also have the sensible property
that the coeﬃcients on distance to the nearest branch have a larger magnitude than the coeﬃcients
on distance to the second nearest branch. It also seems that the coeﬃcient on the number of ATMs
decreases when the model accounts for branch distances accurately at the block group level, but
it remains signiﬁcant at the ﬁve percent level. Overall, the fact that the demand coeﬃcients move
little or in sensible ways from the logit to the full model implies that the full model’s distributional
assumption on deposits is not driving the results. Finally, the marginal cost of deposit-taking for a
six-month interval is estimated to be 1.97 cents per dollar. This coeﬃcient is estimated extremely
precisely.
To illustrate the relative magnitudes of the coeﬃcients, the eﬀect of a one-standard deviation
increase in the interest rate evaluated at the average income in the sample is equivalent to the eﬀect
of an increase in the surcharge variable of 2.8 standard deviations or an increase in the number of
ATMs of 8.5 standard deviations. This same increase in the interest rate is also roughly equivalent
to a decrease in the distance to the nearest bank branch of 1.1 miles or a decrease in the distance
to the second nearest bank branch of 3.4 miles.24
A number of alternative speciﬁcations that explore the robustness of these demand- and supplyside
results are considered in Appendix C. I allow consumers’ valuations to depend on bank presence
in multiple states; I use an alternative assumption on the deposit distribution; I generalize the cost
function speciﬁcation; I examine the potential endogeneity of the surcharge cost variable and ATM
network size; I consider more local deﬁnitions of the ATM variables in the demand system; and I
examine the eﬀect of no-surcharge alliances. The primary speciﬁcation is found to be reasonably
robust.
Finally, Table 7 presents the estimated parameters from the ATM network choice model. The
ﬁrst row contains the results from the primary cost speciﬁcation in which there is just a single
constant multiplying the number of ATMs. The estimated parameter is a singleton, i.e., there
exists no set of parameters that satisfy every inequality restriction.25 The estimated constant is
$32,492 for a six-month period, implying an ATM cost per location of about $5415 per month
with a simulated 95 percent conﬁdence interval from $4905 to $6407 per month. As expected, the
conservative conﬁdence interval is wider, ranging from $15,501 to $45,386 for a six-month period,
24
These eﬀects are evaluated at the mean branch distance, where the mean is taken over consumers and banks.
25
This is not surprising, because the inequalities are random variables. The probability that all elements are
satisﬁed can be made arbitrarily small by increasing the number of inequality restrictions.
23
which is about $2584 to $7564 per month. Recalling that these ﬁgures are based on ATM locations
rather than actual ATM numbers, and using the fact that there are an average of 1.3 ATMs per
location in the data, the average cost per ATM would be 77 percent of the ﬁgure above, or about
$4165 per month.26
As an illustration of how the estimator works in a simple case, the graph on the left in Figure
1 is a histogram of the market average across banks of the change in returns R(·) (gross of ATM
costs) when a bank adds an ATM location. The graph on the right shows the analogous change in
returns when banks subtract an ATM location. If the instrument set included just a constant, the
estimated set of cost parameters would lie between a weighted average of these diﬀerences. One can
visually see that this would be somewhere around 30,000, which is consistent with the estimates
above.
I also compare these results to those obtained by approximating the number of ATMs as a
continuous variable. The cost parameter can then be estimated using standard GMM procedures.27
The results are shown in the second row of Table 7. The estimated coeﬃcient is $38,491 per ATM
location for a six-month period with a standard error of $7,754. The point estimate is slightly larger
than the point estimate using the inequality estimator, but by less than one standard deviation.
Alternative speciﬁcations of ATM costs that depend on a number of diﬀerent bank- or marketspeciﬁc
variables are considered in Appendix C. However, these more complex speciﬁcations are
considerably less signiﬁcant. This is likely due to the small data set being used. With just ten
markets, it is only possible to identify the parameters of quite parsimonious speciﬁcations. Thus,
I use the main speciﬁcation in the ﬁrst row of Table 7 in the results that follow.
8 Implications and Counterfactual Experiments
With the estimates of the structural model above, it is possible to examine the current equilibrium
and consider a variety of policy experiments. I focus on counterfactuals relating to the institutions
governing ATM ownership and fees.
26
According to the American Bankers Association (2003), the cost of buying an ATM machine is as high as $50,000
per machine, and the annual maintenance costs range between $12,000 and $15,000. With a ﬁve-year depreciation
period, this leads to an estimated cost per period that is lower than my point estimate but within a 95 percent
conﬁdence interval. In addition, banks face installation costs, and they may occasionally face costly upgrades to
remain compliant with encryption standards or improve their software or operating systems. Costs in Massachusetts
may also exceed national averages.
27
Note that calculation of dRj/dnj must allow for changes in the interest rate with respect to increments in the
number of ATMs. I use two-sided numerical derivatives using the ﬁrst order condition for a Nash equilibrium in
interest rates. Banks with zero ATMs are dropped from this procedure. This creates only a small selection problem
since only about ﬁve percent of banks in my sample own no ATMs.
24
8.1 Current Equilibrium
Before considering any counterfactuals, I ﬁrst examine the implications of the estimated model in
the current equilibrium. Using observed values of all variables and estimated demand and cost
parameters, I evaluate each bank’s total proﬁts. These include proﬁts from the deposit market as
well as from ATM operations.
Many large banks have defended their use of surcharges by noting that ATMs are expensive
to provide, and indeed, not a proﬁt-generating enterprise. For example, Dove (2002) reports that
“aggressive deployment and intense competition have competed away surcharges’ spoils” and resulted
in negative ATM proﬁtability for many banks. The president of a small ﬁnancial institution
explains that “the machines aren’t intended to be a cash cow” but rather a service (Petersen 2002).
This claim is supported in my results. The model predicts that at the observed levels of interest
rates, almost no bank that owns ATMs makes enough revenue directly from its ATM operations to
oﬀset its costs.28 This implies that an important part of a bank’s incentive to own ATMs derives
from their eﬀect on the deposit market. The elimination of surcharges, which provide incentives for
banks to invest in their ATM networks through direct fees as well as indirectly through the deposit
market, may thus have a sizeable impact on the investment equilibrium.
However, this industry has signiﬁcant potential for an overinvestment problem due to the large
demand-stealing eﬀects in the deposit market associated with ATM networks and surcharging.
When a bank chooses its ATM network size, it does not account for the eﬀect of its choice on its
rivals’ proﬁts. To investigate this, for each bank with positive ATMs, I use the estimated demand
and cost parameters to calculate the change in total proﬁts and consumer surplus associated with
the addition of the bank’s ﬁnal ATM. I hold the bank’s rivals’ networks ﬁxed and allow for changes
in the interest rate equilibrium. I use the lower bound calculation of consumer surplus described
below in section 8.2 which holds consumers’ choices ﬁxed. Averaging across banks, a lower bound
on the average increase in consumer surplus is $6,886, while total proﬁts decrease by an average
of $60,381. This provides some evidence of overinvestment in this industry relative to the social
optimum.29 This overinvestment result will be mitigated to the extent that the true change in
28
Only one bank is predicted to cover its ATM costs through direct ATM revenue.
29
One might worry that the change in proﬁts associated with an increase in ATMs is understated here, because it
does not account for a decrease in labor costs due to substitution away from bank tellers. However, according to the
Occupational Employment Statistics (OES) from the Bureau of Labor Statistics, the average annual teller wage in my
markets in Massachusetts in 2002 was $22,517, implying a six-month wage of $11,259. If an additional ATM allowed
a bank to hire one less teller, this would therefore raise proﬁts by about $11,259, or less than a ﬁfth of the estimated
change in proﬁts. However, it is unlikely that even this amount of teller cost savings would be realized, since banks
are unlikely to substitute between ATMs and tellers one-for-one. Based on national data from the OES and the 2003
EFT Databook, for every additional ATM deployed between 1998 and 2002, banks employed only about 0.2 fewer
tellers. It is also not the case that banks were shutting down large numbers of branches, since according to the FDIC,
the number of banking oﬃces increased by approximately 2 percent during the same period. It is therefore unlikely
that the overinvestment estimate is entirely due to such costs. Another potential source of cost savings due to ATMs
might arise if the convenience of ATMs leads consumers to hold smaller stocks of cash. However, conversations with
25
consumer welfare exceeds the lower bound, but it seems unlikely that the sign of the eﬀect could
be reversed by accounting for switching between banks.30
8.2 Eliminating Surcharges
The possibility of banning surcharges has been at the heart of the public debate about ATMs since
the mid-1990s, and bans were even implemented in certain states and municipalities before being
vacated in court. However, there has been little evidence available regarding the eﬀects of such
bans. I ﬁrst use the estimated model to predict the eﬀects on market shares, equilibrium interest
rates, and welfare conditional on banks’ ATM networks. Since surcharging is expected to be an
important determinant of a bank’s optimal level of investment in its ATM network, it will not be
possible to draw conclusions from these results regarding the overall welfare eﬀects of surcharging.
Below, I also consider the eﬀect of surcharging on ATM investment incentives.31
After exogenously setting the surcharge cost variables to zero for all banks, consumers are
allowed to reoptimize in their bank choice. The new pricing equilibrium is computed numerically.
Starting at the observed interest rates, I search for the new vector of interest rates at which no
ﬁrm can gain proﬁts by deviating. Standard errors are computed using a Monte Carlo approach.
They are based on 200 draws from the estimated asymptotic distribution of the parameters and
their associated policy implications.
Using the estimated parameters from column 2 of Table 6, the results of this experiment on
market shares and prices are shown in Table 8. Allowing banks to adjust their interest rates
and weighting banks’ interest rate changes by their initial deposits, equilibrium deposit interest
rates are predicted to rise by 0.025 percentage points (2.87 percent) on average. These estimates
are signiﬁcant at the one-percent level. It thus appears that eliminating surcharges intensiﬁes
competition on interest rates. The magnitude of the price eﬀect is fairly small, indicating that
surcharging does not have a huge impact on the deposit rate paid on the average dollar of deposits.
However, this average across banks masks the fact that the size and direction of the eﬀect
depends on the relative size of the bank’s ATM network. Banks with between zero and 10 percent
of the network (“small-network banks”) constitute most of the banks. Interest rates on deposits at
these banks, on average, decrease a small amount, and these banks gain market share. The increase
industry representatives indicate that banks do not consider this to be a signiﬁcant factor. One bank executive noted
that while ATM networks have expanded considerably over the past ﬁve years, the size of the average cash withdrawal
has remained stable.
30
Using the equivalent variation of the change in consumer surplus described in section 8.2, the average estimated
increase in consumer surplus is smaller than $6,886.
31
This counterfactual experiment assumes that the interchange and switching fees set by the shared network
do not change. Although surcharges were widely introduced in 1996, network interchange fees generally remained
ﬂat or increased between 1990 and 2002 (see Debit Card Directory, 1995 and 1998, and the EFT Databook, 2003).
Therefore, it does not appear that interchange fees have been very sensitive to surcharge levels.
26
in market share is small in absolute value, but is about 25 percent of initial market share. The
largest ATM owners, those with 20 to 30 percent of the shared network (“large-network banks”,
e.g. Fleet, Citizens, Sovereign, or BankNorth), respond to the surcharge ban by raising interest
rates: the interest rate on the average dollar of deposits rises by 7.2 percent. Nevertheless, these
banks’ market share is predicted to fall by an average of 5.6 percentage points, which is 35 percent
of their original market share. The eﬀects at banks with 10 to 20 percent of the ATM network
(“medium-network banks”) lie between the eﬀects at the smallest and largest ATM owners. All
the estimated changes in market shares and interest rates are strongly statistically signiﬁcant with
the exception of the interest rate changes of small-network banks.
It is interesting to see how much the results are aﬀected by the new pricing equilibrium. In
the last two columns of Table 8 where the experiment is repeated holding interest rates ﬁxed, the
eﬀects are similar. Since the large-network banks are constrained not to raise their interest rates
in response to the change, they are predicted to lose even more market share than in the case in
which interest rates are adjustable.
The eﬀect of the shifts in shares on measures of market concentration is displayed in Table 9.
The ﬁrst two sets of columns show the one-ﬁrm and four-ﬁrm concentration ratios [C(1) and C(4)].
They are ﬁrst calculated using observed market shares, then using the predicted market shares after
surcharges are eliminated and banks adjust their interest rates. On average across markets, C(1) is
21.1 percent, and C(4) is 52.7 percent. Elimination of surcharges has the eﬀect of lowering average
concentration. C(1) falls in all markets but two, and by an average of 3.7 percentage points while
C(4) decreases in all markets, with an average decrease of 7.7 percentage points. The last column
in Table 9 shows analogous statistics using the Herﬁndahl-Hirschman index (HHI).32 Elimination
of surcharges decreases the HHI in all markets, and the average decrease is 169 points.
Since Table 8 demonstrates that banks with large percentages of the ATMs have, on average,
the greatest decreases in market share, one would expect markets in which banks with large ATM
networks also have large market shares to have the greatest decreases in concentration. Boston,
the market in which the changes in concentration are largest, is in fact the market with the highest
correlation between market share and percentage ownership of the ATM network. Across markets,
this correlation has a large negative correlation that is statistically signiﬁcant with change in C(1)
(-0.69), change in C(4) (-0.70), and change in HHI (-0.76).
Finally, Table 10 examines the eﬀect of these shifts in market shares, interest rates, and sur-
32
The HHI is deﬁned as the sum of the squared market shares multiplied by 10,000, where market shares are
calculated here ignoring credit union deposits. The Justice Department and banking authorities do not use credit
union deposits in their estimates of concentration. In addition, the elimination of surcharges causes a decrease in
aggregate demand for the inside goods, increasing estimated market shares for credit unions. This can have a large
impact on HHI if the market share of the outside good is included. In this case, the behavior of HHI is less consistent
across markets, and the average decrease is 15 points.
27
charges on welfare. Elimination of surcharges leads to a loss of industry proﬁts and a reallocation of
proﬁts from the banks with large ATM networks to those with relatively few ATMs. For example,
small-network banks are predicted to gain a total of $44 million per six months while mediumnetwork
banks lose a total of $11 million and large-network banks lose a total of $93 million. This
implies that the industry overall is predicted to lose about $59 million of proﬁts, which is 12.0
percent of proﬁts. When proﬁts are decomposed into proﬁts from the deposit market and proﬁts
from ATM fees, it can be seen that all types of banks lose proﬁts on ATM fees after the elimination
of surcharges. However, the eﬀect of proﬁts being reallocated from banks with larger ATM
networks to those with smaller networks is entirely due to changes in proﬁts from the deposit market.
The changes in interest rates and market shares lower proﬁts from deposits by $85 million
for large-network banks and $8 million for medium-network banks, while raising these proﬁts for
small-network banks by $52 million. These estimates are signiﬁcant at the ﬁve- or the one-percent
level with the exception of the total proﬁt changes at small-network banks, since the changes in
proﬁts from the deposit market and ATM fees oﬀset each other.
When interest rates are held ﬁxed, the small-network banks beneﬁt more from the elimination
of surcharges, because the large-network banks are constrained not to raise their interest rates.
This leads to an overall loss of industry proﬁts from deposits of about $31 million per six months.
The total loss of industry proﬁts is estimated at about $49 million per six months, or 9.9 percent.
It is diﬃcult to use this model to estimate consumer surplus, because there is no marginal
utility of income parameter directly estimated as part of the demand system. Therefore, I calculate
a lower bound on the change in consumer surplus deﬁned as the change in transfers from banks
to consumers holding market shares ﬁxed. This is equal to the expected change in deposit interest
received by consumers less the change in surcharges paid when consumers do not reevaluate their
bank choices. This forms a lower bound since consumers may increase their utility by switching
banks.
As displayed in Table 10, this lower bound on the change in average individual consumer surplus
is estimated to be $8.40 per six months when surcharges are eliminated and banks reoptimize their
interest rates. This estimate is statistically signiﬁcant. Total consumer surplus in the sample of
Massachusetts MSAs rises by an estimated $49.32 million per six months, or $98.64 million per
year. The magnitudes of the consumer surplus eﬀects are considerably smaller when interest rates
are held ﬁxed. Conditional on ATM networks, the estimated change in total welfare is therefore
a statistically insigniﬁcant decrease of $10.06 million per six months. The total welfare eﬀect will
be greater to the extent that consumer switching across banks leads to increases in the change in
consumer surplus beyond the lower bound.
A traditional method for estimating the total eﬀect on consumer welfare uses the equiva-
28
lent variation (EV), which is the amount of consumer income that is equivalent to the change
in welfare. For an individual consumer, the EV associated with the elimination of surcharges is
EVi = (1/µ)[maxj Uno surcharges
ij −maxj Usurcharges
ij ], where µ is the marginal utility of income. This
can be aggregated across consumers to produce the total change in consumer surplus.33 To use
this expression, we need an estimate of µ. The coeﬃcient on the expected surcharge costs can be
interpreted as the marginal utility of income multiplied by the constant T, originally deﬁned in
section 5.1. Therefore, I use the estimate of T discussed in section 5.2 to obtain a rough estimate of
marginal utility. An alternative approach would be to extract an estimate of µ from the coeﬃcient
on the product of the interest rate and income, but this would require a great deal of reliance on
the assumed relationship between income and deposits. The estimated change in consumer surplus
is smaller than the lower bound calculated above, so that lower bound is taken to be more reliable
than this approximation of the equivalent variation.34
It is clear, therefore, that surcharge-induced incompatibility has a substantial impact on market
outcomes and the distribution of welfare in the deposit market. As predicted, when compatibility
causes banks with relatively large networks to lose the advantage of their network size, they are
estimated to lose proﬁts while small-network banks gain proﬁts. The industry overall is predicted
to lose proﬁts. Equilibrium prices are aﬀected as well, with small-network banks lowering their
deposit rates slightly and large-network banks raising them. An elimination of surcharges is also
estimated to reduce market concentration and raise consumer surplus.
The analysis has thus far held ATM networks ﬁxed, but we expect banks’ investment incentives
to depend on the ATM fee structure. It would be ideal to calculate the new network equilibrium
after making a market structure change, however we might expect there to be multiple equilibria
problems in a network industry, and ﬁnding the correct new equilibrium is a complex task that is
beyond the scope of this paper. Therefore, I simply look at the direction of each bank’s investment
incentives. If the initial equilibrium were still a Nash equilibrium, then for each bank, holding its
rivals’ network sizes ﬁxed, the bank would ﬁnd it optimal to leave its network size unchanged. Using
the estimated ATM costs, I calculate each bank’s change in proﬁts from dropping one ATM from
its network, accounting for the implied new interest rate equilibrium. Perhaps not surprisingly,
I ﬁnd that every bank that owns ATMs ﬁnds it optimal to reduce its network size. The average
33
It is useful to write the conditional utility function as Uij = ˜Uij + ij, where is the extreme value unobservable.
Then the total change in consumer surplus is
∆CS = L
Z
EVi g(h)dh g( )d = L
Z
(1/µ){ln[
JX
j=0
exp( ˜Uno surcharges
ij )] − ln[
JX
j=0
exp( ˜Usurcharges
ij )]}g(h)dh,
where L is the total mass of consumers. The second equality is derived in McFadden (1981).
34
Average individual consumer surplus is estimated to rise by $3.22 per six months, and total in-sample consumer
surplus is estimated to rise by $18.90 million per six months.
29
change in proﬁts from a reduction of one ATM is about $29,000 per six months.35
Given this investment response, it is likely that the overall welfare eﬀect from eliminating surcharges
will change. A reduction in network size is expected to reduce consumer surplus, potentially
by a large amount. However, a reduction in ATM expenditures may lead to an increase in total
industry proﬁts.
8.3 Introducing a Network Operator
An alternative industry structure of interest would place maintenance and operation of the ATM
network in the hands of a third-party operator. Such a solution has been used, for example, in the
electricity industry where a centralized system operator manages the transmission network. Here,
holding the total network size ﬁxed, I consider whether ﬁrm proﬁts and consumer welfare would be
improved or worsened if a third-party operator assumed responsibility for the ATM network and
charged consumers exactly the cost of operating the network.
Based on the size of the total network, the total population, and estimated ATM costs, the cost
per consumer to the network operator would be $16.29. I assume each consumer is charged this ﬂat
fee. The change in consumer surplus without this fee should be very similar to the eﬀect described
above when surcharges are eliminated, but it will diﬀer slightly due to changes in the banks’ proﬁt
function, and therefore their interest rate-setting behavior. In the proﬁt function, I set surcharges,
interchange fees, and foreign fees to zero, since there is no longer a need for payments for foreign
transactions, but I assume that banks continue to pay switching fees.
Table 11 shows the estimated changes in interest rates, market shares, and proﬁts. The eﬀects
on interest rates and market shares are similar to the eﬀects when surcharges are eliminated but
banks continue to own the network. Looking at proﬁts, we can see that, due to the high cost
of owning and operating ATMs, total industry proﬁts rise upon the introduction of a network
operator, but the eﬀect is statistically insigniﬁcant. However, there are signiﬁcant distributional
eﬀects. Small-network banks are estimated to gain from this change, but large-network banks lose
proﬁts. For large-network banks, the gain in proﬁts from forgone ATM costs is outweighed by the
substantial loss in the deposit market.
The estimated lower bound change in consumer surplus is a total of $49.02 million, or an average
of $8.35 per consumer. Since this is much less than $16.29, it seems unlikely that consumers would
be better oﬀ under this proposed third party network operator.36 This is perhaps not surprising
given the previous result that banks’ direct revenue from ATM operations does not cover their
35
This simple analysis does not distinguish between sunk and non-sunk costs. Banks’ incentives to drop ATMs
may be lessened to the extent that many of the costs of ATM investment are sunk.
36
The change in consumer surplus is understated because it does not account for switching between banks or the
elimination of foreign fees. It is impossible to extract this foreign fee eﬀect from the utility function. These two
eﬀects would have to contribute about $8 per consumer to reverse the sign of the total eﬀect.
30
ATM costs. Banks can cover some of these costs with proﬁts from the deposit market due to the
bundling of deposit services and ATM services in the banking industry.
9 Conclusion
This paper has estimated a structural model of demand and supply for banking deposit services
that incorporates ATM networks and retail ATM interconnection pricing. In this industry, each
bank’s ATM cards are technologically compatible with all ATMs in the shared network, but the
use of surcharges leads to partial incompatibility. This implies that demand for a bank’s deposit
services should depend on its ATM network size and its surcharge, since consumers are able to
avoid a bank’s surcharge by choosing that bank. The demand parameters reﬂect this, showing that
consumers value bank ATM network size and dislike surcharges. Indeed, the results of a model of
ATM network investment imply that banks generally do not cover their ATM costs through direct
ATM fee revenue, which highlights the importance of the deposit market on banks’ ATM network
decisions. Strong demand-stealing eﬀects also appear to lead to overinvestment in ATMs in the
observed equilibrium.
The estimated demand and cost parameters imply that network eﬀects and partial incompatibility
can have large impacts on equilibrium prices, market share allocations, and welfare. A
counterfactual policy experiment conditional on network size predicts that surcharging leads to
signiﬁcantly higher concentration and lower deposit interest rates, with the interest rate eﬀects
being the greatest for large-network banks. It also reallocates market share and large amounts of
proﬁts from deposits from small-network banks to large-network banks and lowers consumer surplus.
However, surcharging has a substantial positive eﬀect on banks’ incentives to invest in their
ATM networks. A further counterfactual experiment illustrates that an attempt to move provision
of ATM services into the hands of an independent third-party operator would have distributional
eﬀects on ﬁrm proﬁts and would be unlikely to beneﬁt consumers.
This paper leaves open questions regarding the ultimate equilibrium response of ATM network
size to policy stimuli. This is of particular interest in a network industry, making this a useful topic
of future research.
31
Appendix A: Data Details
The Massachusetts Division of Banks requires ATM owners to submit information about the
locations and surcharges of each of their ATMs annually. I calculate a bank’s number of ATMs
in a local market by geocoding the ATMs.37 I also classify ATMs as in-branch or remote based
on whether the addresses match any branch addresses of the ATM owner. I assign each bank in
each market a surcharge level. Surcharges are generally a bank-level decision, and other theoretical
and empirical work has treated them this way.38 In the ATM data, for almost 90 percent of bank
observations, only one surcharge level is used across all ATMs owned in a market by that bank.
For those banks that use more than one surcharge, I use the modal surcharge.
To compute the distances between block groups and bank branches, I use the Haversine formula
from Sinott (1984) for calculating the distance between two points on a sphere.
The credit union deposit data from the Financial Performance Reports of the National Credit
Union Association is available by credit union, rather than by branch location. Therefore, I estimate
a credit union’s local market deposits as the fraction of its total deposits that is equal to the
fraction of its branches located in the market.39
Market Deﬁnition
I assume that the relevant geographic size of a market is a metropolitan statistical area (MSA).40
Antitrust analysis in banking has traditionally assumed that markets are essentially local, and
there has been substantial evidence in favor of geographically local retail deposit markets. Surveys
consistently show that consumers choose ﬁnancial institutions very close to their homes and workplaces.41
The median distance of consumers from their depository institutions was 3 miles in the
1998 Survey of Consumer Finances, and 75 percent of consumers were within 8 miles (Amel and
Starr-McCluer 2002). Studies have also documented negative relationships between deposit rates
and concentration across local areas using both recent and older data. See, for example, Heitﬁeld
and Prager (2002) and Berger and Hannan (1989). MSAs provide a convenient and commonly used
37
I use the Maptitude software package for all geocoding of street addresses.
38
E.g., Hannan et al. (2003), Massoud and Bernhardt (2000,2002), and McAndrews (2001).
39
This should be a reasonable approximation, and furthermore, in the logit demand speciﬁcation, any errors in
measurement of market-level demand for credit unions aﬀect only the market dummies. This can be seen in equation
(17), in which the share of the outside good aﬀects only the constant subtracted from the left-hand side.
40
An MSA is essentially an area with a large population center combined with adjacent communities that are highly
socially or economically integrated. In large markets, I use primary MSAs rather than the larger consolidated MSAs.
41
Although both households and businesses consume banking services, this paper uses a behavioral model of
individual consumers, since it uses data on household location and demographics. Separation of business from
household demand is often impossible in any demand model. In this application, a model of household behavior is
a reasonable simpliﬁcation, because households hold the vast majority of deposits. According to the 2000 Federal
Reserve Flow of Funds, households hold about 80 percent of total U.S. household/nonﬁnancial business deposits. In
addition, consumer and small business behavior has been found to be similar in the use of depository institutions
and the clustering of demand for ﬁnancial services at local institutions (Kwast, Starr-McCluer, and Wolken 1997).
32
local geographic delineation.
The product market is deﬁned as deposits at depository institutions, where deposits include
checking, savings, and time deposits. Geographically disaggregated data on branch deposits is
available only for total deposits. Fortunately, there is evidence that consumers cluster their demand
for depository services at their primary institution. According to the 1998 Survey of Consumer
Finances, most households obtain multiple services at their primary ﬁnancial institution (Amel and
Starr-McCluer 2002). Account balances also tend to be fungible across diﬀerent types of accounts.
For example, many banks oﬀer “relationship” accounts that link balances across deposit accounts.
Depository institutions include commercial banks, thrifts, and credit unions. Many empirical
banking papers have included only commercial banks in their market deﬁnitions, but thrifts and
credit unions have a signiﬁcant presence in Massachusetts with 16.6 percent of the deposits in
all markets. It is also useful to use this fairly broad deﬁnition because, in the behavioral model,
I assume that all consumers use a commercial bank, thrift, or credit union. The 1998 Survey
of Consumer Finance showed that over 98 percent of households with any bank-type accounts
or loans used a depository institution, so this is a reasonable approximation (Amel and StarrMcCluer
2002). While antitrust analysis has traditionally included only commercial banks in
market concentration calculations, it has accounted for competition from thrifts, credit unions,
and other institutions through the use of merger guidelines that are less restrictive than those
used in non-banking industries.42 Since legislation in 1980 has allowed for increased competition
between commercial banks and thrifts, the Federal Reserve Board now includes some or all thrifts
in their market concentration calculations, while maintaining the relaxed concentration standards
to account for competition from credit unions and other institutions (Amel 1997).
Appendix B: Estimation Details for Demand and Interest Rate Choice
Stacking the covariance restrictions in equation (16), the following moment condition can be
formed:
G(θ) = E[mj(θ)] = E


ξj(θ)zD
j
ωj(θ)zS
j

 .
The vector of demand parameters can be estimated using the demand moments alone. However,
joint estimation using the demand and supply moments is expected to increase the eﬃciency of the
demand estimates, since the interest-rate-setting equation implied by the Nash equilibrium depends
on the demand parameters. I will estimate θ0 both ways. Equation (16) implies that G(θ0) = 0. If
42
The U.S. Department of Justice Merger Guidelines state that a proposed banking merger would not be expected
to raise antitrust concerns unless it increased the Herﬁndahl-Hirschman index (HHI) by 200 or more points and led
to a level greater than 1800. In non-banking industries, mergers that raise the HHI by more than 50 points are
considered potentially problematic.
33
there are a total of N bank observations in the sample, a sample moment condition can be formed
as follows:
GN (θ) =
1
N
N
j=1
mj(θ) =
1
N
N
j=1


ξj(θ)zD
j
ωj(θ)zS
j

 .
Under some technical conditions, the estimator deﬁned by minimizing GN (θ) A, where A
is a weight matrix, is
√
N-consistent and asymptotically normal with variance-covariance matrix
(Γ AΓ)−1Γ AV AΓ(Γ AΓ)−1, where Γ = ∂G(θ0)/∂θ and V = E[m(θ0)m(θ0) ]. These standard errors
allow for arbitrary heteroskedasticity.
The two diﬃculties that will be discussed in turn are that the market shares are nonlinear
functions of ξ, and the expression for market share is non-analytic. First, individual consumer
demand, and therefore aggregated market demand, is a nonlinear function of the unobservable, ξ.
In order to use the orthogonality restrictions above, it is necessary to transform the data into a
linear function of the ξ. Berry (1994) shows the conditions under which a unique mapping from
shares to ξ exists, given the observable data and the parameter vector, and BLP demonstrates how
this mapping can be computed. In BLP, a contraction mapping is established that allows ξ to be
computed iteratively for each set of parameters, observable data, and distributions of consumer
unobservables. It can be shown that the contraction mapping used in BLP is also a contraction
mapping given the market share in equation (10).43 This should not be surprising given that the
market share used here diﬀers from the standard market share only in that it is a weighted average
of individual consumer choice probabilities, whereas the standard market share is a simple average.
Therefore, the market shares here can be inverted to solve for ξ using the BLP contraction mapping.
Second, it is not possible to explicitly compute GN (θ), because there exists no analytic expression
for the integrals in equation (10), the market share. However, simulation methods as in Pakes
and Pollard (1989) and BLP allow θ to be estimated consistently and with only minor changes in
its limiting properties. The shares are simulated as described above in section 6.1. The asymptotic
distribution of the estimator is aﬀected only through the estimated variance-covariance matrix of
the moment conditions, ˆV , which uses the simulated moments.
In the actual estimation, γ and w are “concentrated out” using the fact that the ﬁrst-order
conditions associated with the minimization problem are linear in these parameters. Then a nonlinear
search is performed over the remaining parameters using the Nelder-Mead non-derivative
simplex method.
Appendix C: Robustness
This appendix considers a number of alternative speciﬁcations to explore the robustness of both
43
Each of the conditions of the proof in the appendix of BLP can be veriﬁed.
34
the demand- and the supply-side results of section 7. First, I examine the demand and interestrate-setting
model. I allow consumers’ valuations to depend on whether banks have branches in
multiple states; I use an alternative assumption on the deposit distribution; I use a diﬀerent cost
function speciﬁcation; I consider the potential endogeneity of the surcharge cost variable and ATM
network size; I consider the possibility that ATM demand is more properly measured at a local
level; and I examine the eﬀect of no-surcharge alliances. I also consider alternative speciﬁcations
for the ATM network investment model. I allow the costs of in-branch and remote ATMs to diﬀer;
I consider the possibility that ATM costs depend on the cost of leasing space; and I allow ATM
costs to depend on the size of the network.
First, consumers could have heterogeneous preferences for “large” banks, with some consumers
preferring them and others preferring community-style banks. In column 1 of Table 12, I reestimate
the full model with the addition of a random coeﬃcient on an indicator variable for banks that
have branches in multiple states. I assume that the marginal utility of this variable is distributed
normally across consumers, and I estimate the mean and standard deviation of this distribution.
Neither parameter is estimated to be signiﬁcantly diﬀerent from zero.
Next, it was noted previously that it is possible to use more complex assumptions on the
distribution of deposits than the assumption used above. In order to allow randomness in the
distribution, I reestimate the full model assuming that, conditional on yi, deposits are distributed
exponential with mean αyi.44 Column 2 of Table 12 shows the estimated coeﬃcients. They are
qualitatively similar to the main results of Table 6’s second column, though the addition of noise
at the block group level seems to make it harder for the eﬀect of the branch distance variables to
be picked up. Their coeﬃcients are much smaller and less signiﬁcant. Nevertheless, these results,
combined with the logit results where it is assumed that consumers are homogeneous, provide
evidence that assumptions about the distribution of deposits are not driving the results.
I next examine the cost function. The speciﬁcation in section 7 included only a constant,
but one might expect there to be not only a marginal cost for deposits but also a ﬁxed cost per
customer due to costs of printing bank statements, providing customer service, etc. The model
can accommodate such costs since it yields predictions for both share of deposits (sj) and share of
consumers (qj). Column 3 of Table 12 shows the coeﬃcient estimates from the full model when the
cost function includes both a constant on deposits and a per-customer ﬁxed cost. The ﬁxed cost
estimate is insigniﬁcantly diﬀerent from zero.
I now return to the issue of the potential endogeneity of the surcharge cost variable or of ATM
network size. It can be argued that there is only minimal variation in surcharges, mitigating the
44
Note that this speciﬁcation once again identiﬁes φ = λα, where λ is the parameter in the utility function on the
product of deposits and the interest rate. One way to see this is by inverting the cdf of the exponential distribution
so that if wi = −αyiln(1 − ui) where ui ∼ U[0, 1], then wi has the same distribution as deposits, exponential(αyi).
35
scope of the surcharge endogeneity problem. In addition, the observables in the utility function
include what are expected to be the most important determinants of surcharge levels. However,
to address any remaining concerns with the surcharge variable or the number of ATMs, I estimate
the full model again, this time instrumenting for surcharge costs and the number of ATMs. With
the exception of these two variables, I use the same instruments as before, and I add two new
instruments: the number of in-branch ATMs and the number of in-branch ATMs as a percentage
of total market ATMs. In-branch ATMs are highly correlated with total bank ATMs, but they
should be closer to exogenous themselves. Many banks automatically invest in ATM machines
for their bank branches, so if the branch network can be taken as exogenous for the purposes of
this cross-sectional study, then these in-branch ATMs are close to exogenous. Also, they make an
attractive instrument, because a bank’s number of in-branch ATMs may contain information about
the bank’s ATM costs.45
Column 4 of Table 12 presents the results using the full model with demand estimated jointly
with supply. The standard errors on the surcharge cost variable and the number of ATMs increase
when these variables are not included in the instrument set, but the coeﬃcient on surcharge costs
remains signiﬁcant at the 5 percent level and the coeﬃcient on the number of ATMs is signiﬁcant at
the 10 percent level. The coeﬃcients are very similar to the results in Table 6, where the surcharge
and ATM network size variables are included in the instrument set. Indeed, none of the coeﬃcients
are signiﬁcantly diﬀerent. This provides further support for not instrumenting for the surcharge or
ATM network size variables.
Next, although I expect market-level ATM networks to be a good proxy for what consumers care
about, I consider a local version of this variable. I deﬁne an alternative ATM network size variable
as the number of ATMs within a ﬁve-mile radius of a consumer’s block group (called #ATMs5).
Similarly, I deﬁne a local version of the surcharge cost variable analogous to surch for costs within
a ﬁve-mile radius (called surch5). In order to ease the computational burden, I return to the logit
demand model. Column 1 of Table 13 shows the estimated demand coeﬃcients when these ATM
network size and surcharge cost variables are averaged over block groups, weighting by block group
population. Most of the coeﬃcients change very little, but the coeﬃcient on surch5 is smaller
(though not signiﬁcantly so) and less signiﬁcant than the coeﬃcient when market-level variables
are used. The change in the ATM network size coeﬃcient reﬂects the scaling of the variable. Since,
if anything, the coeﬃcient on the surcharge cost variable picks up less of an eﬀect when measured
locally, I maintain the assumption that ATM networks are appropriately measured at the level of
the market. The market-level variables should account for work, shopping, and other important
locations away from home.
45
In-branch ATMs are thought to be less expensive to install and operate than remote ATMs. See, for example,
Balto (1996) and Dove Consulting’s 2002 ATM Deployer Study.
36
A ﬁnal potential issue with the estimated demand system is that many banks in Massachusetts
are members of a no-surcharging alliance called SUM. Banks that choose to join SUM agree not
to surcharge each other’s customers, though they still independently set surcharges for customers
of banks that are not members. I therefore consider an alternative surcharge variable called sur
which accounts for the SUM alliance.46 In each market m, the banks can be ordered so that banks
1, ..., ˆJm are members of SUM and banks ˆJm +1, ..., Jm are not. Customers of non-SUM banks face
surcharge costs of tm −Tfjmcjm, as above, whereas customers of SUM banks should face surcharge
costs of T Jm
k= ˆJm+1
fkmckm = T Jm
k=1 fkmckm − T
ˆJm
k=1 fkmckm = tm − T
ˆJm
k=1 fkmckm. As with
surch, the constant, tm, cancels. Therefore, we can deﬁne
surjm = I{non-SUM}jmfjmcjm + I{SUM}jm(
ˆJm
k=1 fkmckm).
It will also be useful to deﬁne the following variables:
surnon−SUM
jm = I{non-SUM}jmfjmcjm
surSUM
jm = I{SUM}jm
ˆJm
k=1 fkmckm
surnojSUM
jm = I{SUM}jm
ˆJm
k=1,k=j fkmckm.
Using a logit regression and the alternative surcharge variable, sur, the second column of Table
13 shows that the coeﬃcients on all variables except for sur appear quite similar to the instrumental
variables results in Table 5. The coeﬃcient on sur is very small and completely insigniﬁcant. The
speciﬁcation in the third column of Table 13 allows for diﬀerent coeﬃcients on sur for SUM and
non-SUM banks. The coeﬃcient is again small and insigniﬁcant for the SUM banks, but is larger
and marginally signiﬁcant (with a p-value of 0.06) for the non-SUM banks. This diﬀerence in
coeﬃcients may arise if consumers do not fully internalize the eﬀect of SUM membership when
choosing banks, so that avoiding surcharges at an ATM in the SUM alliance not owned by the
consumer’s bank is not considered to be equivalent to avoiding surcharges at ATMs owned by
the consumer’s bank. The surcharge variable sur (or surSUM ) would then be mismeasured for
SUM banks. Such an eﬀect could arise if consumers are ill-informed about the SUM alliance when
choosing banks. For example, in 2002, consumers may not have understood how it works, or they
may not have known how large it was or which banks were members.47
To examine this issue further, I consider one additional speciﬁcation in the fourth column.
This speciﬁcation splits the terms in surj in a diﬀerent way. As deﬁned above, surch accounts
for the surcharges of the consumer’s own bank, and surnojSUM handles the remaining terms for
the SUM banks. There is little change in the other coeﬃcients, and the coeﬃcient on surnojSUM
46
Banks are assumed to be members of SUM if they were listed as participating institutions in May 2003 on the
SUM website, http://www.sum-atm.com.
47
Indeed, individual banks may not wish to spend signiﬁcant resources advertising the SUM network, since advertising
for the network is a form of advertising for their competitors.
37
is insigniﬁcantly diﬀerent from zero. These results imply that the expected surcharge costs at a
consumer’s own bank have a signiﬁcantly positive eﬀect on bank demand, but the surcharge costs
of SUM banks other than the consumer’s own bank have an insigniﬁcant eﬀect. In the analysis,
I therefore use surch in the vector of bank characteristics rather than the alternative surcharge
variable, sur.
I now consider some alternative speciﬁcations for the ATM network choice model. I ﬁrst consider
allowing a diﬀerent cost coeﬃcient for in-branch and remote ATMs by allowing banks to make
discrete choices over two variables: the number of in-branch ATMs and the number of remote
ATMs. One might expect the cost of installing and maintaining ATMs to be higher for remote
than in-branch ATMs, since rental costs may be higher at remote locations and oﬀ-site ATMs may
also be more diﬃcult to service. However, banks may tend to install more ATMs or more expensive
ATMs in their branch locations, so the direction of the cost diﬀerence is not clear. Rows 1 and 2
of Table 14 show that the estimated parameters are quite similar, with an in-branch coeﬃcient of
$36,649 and a remote coeﬃcient of $38,348. The remote coeﬃcient is somewhat larger than the
in-branch coeﬃcient, but the remote parameter’s 95 percent conﬁdence interval is also quite a bit
larger. It covers the entire conﬁdence interval for the in-branch parameter in both the conservative
and simulated conﬁdence interval cases.
Next, one might expect ATM costs to vary based on the cost of leasing space. I assume that
the cost of an ATM network of size nj is (δ +bjδb +νj)nj, where bj is the rental cost variable. Since
commercial real estate price data is unavailable, I use business density as a rough proxy for these
rental costs. The number of ﬁrms ﬁgure that is used to form this variable is obtained from the 1997
Economic Census, and I include business density in the instrument set for this speciﬁcation. Rows
3 and 4 show that the coeﬃcient on the business density term is not of the expected sign, and it
is insigniﬁcantly diﬀerent from zero based on the conservative conﬁdence interval. The coeﬃcient
on the constant term is larger than in the primary speciﬁcation, but its conﬁdence interval is also
much larger. The conservative conﬁdence interval extends past zero.
Finally, I consider a speciﬁcation that allows the cost per ATM to depend on the size of the ATM
network. I estimate separate coeﬃcients on the number of ATMs and the square of the number of
ATMs. One might expect there to be decreasing costs if owners of larger networks purchase their
machines at lower prices, but there may be oﬀsetting quality diﬀerences. In addition, it is possible
that banks face increasing costs as low-cost locations are ﬁlled up. Rows 5 and 6 show positive
coeﬃcients for both the number and the square of the number of ATMs. However, the coeﬃcient
on the number of ATMs is no longer signiﬁcant. This is likely due to the very high correlation
between the number of ATMs and its square (0.96).
38
Appendix D: Elasticity Calculations
In the logit model, in which market shares are as deﬁned in equation (11), the elasticity of bank
j’s share with respect to the interest rate of bank k in market m can be computed as follows:
ηjkm =
∂sjm
∂rkm
rkm
sjm
=



˜λrjm(1 − sjm) if k = j,
−˜λrkmskm otherwise.
In the full model, in which market shares are as deﬁned in equation (10), the elasticities take the
following form:
ηjkm =
∂sjm
∂rkm
rkm
sjm
=



φrjm ×
y,d Pjm(y, d)[1 − Pjm(y, d)] y2 g(y, d) dy dd
y,d Pjm(y, d) y g(y, d) dy dd
if k = j,
−φrkm ×
y,d Pjm(y, d)Pkm(y, d) y2 g(y, d) dy dd
y,d Pjm(y, d) y g(y, d) dy dd
otherwise.
References
[1] Amel, Dean F., 1997, Antitrust Policy in Banking: Current Status and Future Prospects,
Conference on Bank Structure and Competition, 33, 166-179.
[2] Amel, Dean F., and Martha Starr-McCluer, 2002, Market Deﬁnition in Banking: Recent
Evidence, The Antitrust Bulletin, 47(1), 63-89.
[3] American Bankers Association, 2003, ATM Fact Sheet.
[4] ATM and Debit News EFT Databook, 2003 edition, Thomson.
[5] Balto, David A., 1996, ATM Surcharges: Panacea or Pandora’s Box?, Review of Banking and
Financial Services, 12(17), 169-177.
[6] Berger, Allen N., and Timothy H. Hannan, 1989, The Price-Concentration Relationship in
Banking, Review of Economics and Statistics, 71(2), 291-299.
[7] Berry, Steve T., 1994, Estimating Discrete-Choice Models of Product Diﬀerentiation, RAND
Journal of Economics, 25(2), 242-262.
[8] Berry, Steve, James Levinsohn, and Ariel Pakes, 1995, Automobile Prices in Equilibrium,
Econometrica, 63(4), 841-890.
[9] Chou, Chien-fu, and Oz Shy, 1993, Partial Compatibility and Supporting Services, Economic
Letters, 41, 193-197.
[10] Cremer, Jacques, Patrick Rey, and Jean Tirole, 2000, Connectivity in the Commercial Internet,
Journal of Industrial Economics, 48(4), 433-472.
39
[11] Croft, Elizabeth W., and Barbara J. Spencer, 2003, Fees and Surcharging in Automatic Teller
Machine Networks: Non-Bank ATM Providers versus Large Banks, N.B.E.R. Working Paper
9883.
[12] Davis, Peter, 2001, Spatial Competition in Retail Markets: Movie Theaters, Working Paper,
MIT Sloan.
[13] Debit Card Directory, 1995 edition, 1998 edition, Faulkner and Gray.
[14] Dick, Astrid A., 2002, Demand Estimation and Consumer Welfare in the Banking Industry,
Working Paper, Federal Reserve Board, Washington, D.C.
[15] Dove Consulting, 2002 ATM Deployer Study, Executive Summary, February 2002.
[16] Economides, Nicholas, 1989, Desirability of Compatibility in the Absence of Network Externalities,
American Economic Review, 1165-1181.
[17] Farrell, Joseph, and Garth Saloner, 1985, Standardization, Compatibility, and Innovation,
RAND Journal of Economics, 16(1), 70-83.
[18] Gowrisankaran, Gautam, and John Krainer, 2003, The Welfare Consequences of ATM Surcharges:
Evidence from a Structural Model, Working Paper, Washington University.
[19] Hannan, Timothy H., Elizabeth K. Kiser, Robin A. Prager, and James J. McAndrews, 2003,
To Surcharge or Not to Surcharge: An Empirical Investigation of ATM Pricing, Review of
Economics and Statistics, 85(4), 990-1002.
[20] Hannan, Timothy, 2002, Retail Fees of Depository Institutions: 1997-2001, Federal Reserve
Bulletin, 405-413.
[21] Heitﬁeld, Erik, and Robin A. Prager, 2002, The Geographic Scope of Retail Deposit Markets,
Working Paper, Federal Reserve Board, Washington, D.C.
[22] Katz, Michael L., and Carl Shapiro, 1985, Network Externalities, Competition, and Compatibility,
American Economic Review, 75(3), 424-440.
[23] Knittel, Christopher, and Victor Stango, 2003, Compatibility and Pricing with Indirect Network
Eﬀects: Evidence from ATMs, Working Paper, Federal Reserve Bank of Chicago.
[24] Knittel, Christopher, and Victor Stango, 2004, Incompatibility and Consumer Demand: Evidence
from ATMs, Working Paper, UC Davis.
[25] Kwast, Myron L., Martha Starr-McCluer, and John D. Wolken, 1997, Market Deﬁnition and
the Analysis of Antitrust in Banking, The Antitrust Bulletin, 42(4), 973-995.
[26] Massoud, Nadia, and Dan Bernhardt, 2000, Endogenous ATM Concentration, Working Paper,
University of Alberta.
[27] Massoud, Nadia, and Dan Bernhardt, 2002, ATM Competition: Understanding “Rip-Oﬀ”
ATM Surcharges, RAND Journal of Economics, 33(1), 96-115.
[28] Matutes, Carmen, and Jorge A. Padilla, 1994, Shared ATM Networks and Banking Competition,
European Economic Review, 38(5), 1113-1138.
40
[29] Matutes, Carmen, and Pierre Regibeau, 1988, “Mix and Match”: Product Compatibility
Without Network Externalities, RAND Journal of Economics, 19(2), 221-234.
[30] McAndrews, James J., 1991, The Evolution of Shared ATM Networks, Business Review, Federal
Reserve Bank of Philadelphia, 3-16.
[31] McAndrews, James J., 2001, A Model of ATM Pricing: Foreign Fees and Surcharges, Working
Paper, Federal Reserve Bank of New York.
[32] McFadden, Daniel, 1974, Conditional Logit Analysis of Qualitative Choice Behavior, in P.
Zarembka, ed., Frontiers in Econometrics, 105-142, New York: Academic Press.
[33] McFadden, D., 1981, Econometric Models of Probabilistic Choice, in C. Manski and D. McFadden,
eds., Structural Analysis of Discrete Data, 198-272, Cambridge: MIT Press.
[34] Pakes, Ariel, and David Pollard, 1989, Simulation and the Asymptotics of Optimization Estimators,
Econometrica, 57(5), 1027-1057.
[35] Pakes, Ariel, Jack Porter, Katherine Ho, and Joy Ishii, 2005, Moment Inequalities and Their
Application, Working Paper, Harvard University.
[36] Petersen, Rebecca A., Automatic Fees, Ames Tribune, May 28, 2002.
[37] Prager, Robin, 2001, The Eﬀects of ATM Surcharges on Small Banking Organizations, Review
of Industrial Organization, 18(2), 161-173.
[38] Sinnott, Roger W., 1984, Virtues of the Haversine, Sky and Telescope, 68(2), 159.
41
Table 1: Markets Description
Market Banks Mean sj C(1) Biggest Bank s0
Barnstable-Yarmouth 9 11.1 25.4 Cape Cod Five 0.04
Boston 148 0.6 30.1 Fleet 4.83
Brockton 22 3.7 11.2 Rockland Trust 19.64
Fitchburg-Leominster 12 4.6 11.6 BankNorth 44.59
Lawrence 17 5.2 26.2 BankNorth 11.51
Lowell 19 4.4 12.6 Enterprise B&T 15.81
New Bedford 10 7.7 32.1 Compass Bank 14.94
Pittsﬁeld 8 9.2 25.6 Berkshire Bank 17.52
Springﬁeld 21 4.3 16.1 Fleet 9.39
Worcester 25 3.5 18.3 Fleet 12.32
Notes: Mean sj is the average market share of the inside goods, C(1) is the market share
of the largest inside good, and s0 is the market share of the outside good, credit unions.
Table 2: Bank Characteristics Description
Mean Median Std.Dev.
Number Branches 6.2 3.0 15.1
Number ATMs 10.1 4.0 40.1
In-Branch ATMs 5.7 3.0 14.8
Remote ATMs 4.4 0 25.9
Interest Rate (%) 1.2 1.2 0.3
Employees per Branch 19.2 15.0 31.6
Notes: The number of observations is 291.
Table 3: ATM Surcharges
Number Mean %
Surcharge Banks Network Owned
0.00 78 1.1
0.50 5 0.2
0.75 6 1.7
1.00 137 4.5
1.25 23 8.2
1.50 22 3.7
2.00 4 3.0
Notes: Mean % Network Owned uses the number
of ATMs as a percentage of the total ATMs
owned by banks in the market.
42
Table 4: Block Groups Description
Number Area Pop Bank Distance Income
Market Block Groups Mean Mean Mean Std.Dev. Mean Std.Dev.
Barnstable-Yarmouth 141 1.73 1153.0 5.16 2.05 46,342 11,996
Boston 2653 0.75 1280.8 17.41 6.17 60,035 27,099
Brockton 209 1.42 1222.3 5.41 1.79 52,728 18,372
Fitchburg-Leominster 102 2.73 1395.0 5.07 1.81 44,250 13,928
Lawrence 184 1.02 1439.5 4.78 1.02 53,051 27,820
Lowell 206 1.08 1411.5 5.31 1.68 58,383 22,576
New Bedford 160 1.33 1095.0 3.58 1.81 39,084 17,082
Pittsﬁeld 96 2.62 882.4 5.17 2.23 40,952 13,447
Springﬁeld 434 1.68 1364.0 6.87 2.02 41,043 16,876
Worcester 418 1.94 1202.2 7.72 3.05 48,595 18,689
All 4603 1.14 1274.9 12.56 7.52 54,625 25,061
Notes: Area is in square miles, and Pop is total population. Bank distance is the average over banks of the shortest
distance in miles from each block group to any branch of each bank. Income is median income in dollars.
Table 5: Logit Demand Model
OLS IV
Coeﬀ. Std. Err. Coeﬀ. Std. Err.
Interest Rate 71.6019** 21.1358 217.3221** 51.9314
Surch 2.5521 1.3988 3.9715** 1.4080
#ATMs 0.0065** 0.0013 0.0066** 0.0016
AvgDist1 -0.0687 0.0420 -0.0868 0.0470
AvgDist12 0.0012 0.0008 0.0018* 0.0009
AvgDist2 -0.0869* 0.0432 -0.0907 0.0470
AvgDist22 0.0013 0.0009 0.0013 0.0010
OneBranchDummy -2.5316** 0.4594 -2.6622** 0.4936
Employees/Branch 0.0114** 0.0013 0.0140** 0.0011
Market Dummies Included Included
R2 0.8504 0.8165
N 291 291
Markets 10 10
Notes: The dependent variable is ln(sj) − ln(s0). Surch is as deﬁned in section
5.1, and the average distance variables average the block group-speciﬁc branch
distance variables deﬁned in section 5.1 over block groups, weighting by block
group population. White heteroskedasticity-consistent standard errors are used.
Signiﬁcance at the 5 and 1 percent levels is indicated by * and **, respectively.
The R2
for the IV speciﬁcation is intended only to give an indication of goodness-
of-ﬁt.
43
Table 6: Full Model
(1) (2) (3)
Demand Joint demand Joint demand
& supply & supply
Coeﬀ. Std. Err. Coeﬀ. Std. Err. Coeﬀ. Std. Err.
Demand:
Interest Rate×Income 3.4379** 1.1503 5.7155** 0.9390 2.9836** 0.6479
Surch 2.0619 1.8788 4.3123** 1.3972 3.5577** 1.3807
#ATMs 0.0014 0.0014 0.0024* 0.0011 0.0074** 0.0016
Dist1 -0.9561 0.9761 -1.0476** 0.2159
Dist12 -0.0063 0.0445 0.0137** 0.0028
Dist2 -0.6850 0.4423 -0.3777* 0.1704
Dist22 0.0223 0.0257 0.0059** 0.0019
AvgDist1 -0.0824 0.0455
AvgDist12 0.0018* 0.0009
AvgDist2 -0.0849* 0.0453
AvgDist22 0.0012 0.0010
One Branch Dummy -2.9618** 1.6368 -2.3064** 0.7386 -2.6508** 0.4740
Employees/Branch 0.0178** 0.0035 0.0194** 0.0019 0.0140** 0.0011
Market Dummies Included Included Included
Marginal Cost:
Constant 0.0197** 0.0007 0.0722** 0.0017
N 291 291 291
Markets 10 10 10
Notes: Surch is deﬁned in section 5.1. The distance variables are the block group-speciﬁc branch distances,
which are also deﬁned in section 5.1. Signiﬁcance at the 5 and 1 percent levels is indicated by * and **,
respectively.
Table 7: ATM Choice Model
Cost Simulated 95% Conservative 95%
Coeﬀ. Conﬁdence Interval Conﬁdence Interval Std. Err.
1. #ATMs 32,492 29,431 38,444 15,501 45,386
FOC Method
2. #ATMs 38,491 7,754
Notes: There are 291 banks in 10 markets in every speciﬁcation.
44
Table 8: Eliminating Surcharges
Shares and Prices
Allowing r to adjust Holding r ﬁxed
Average over ﬁrms Average over ﬁrms
% Network N ∆r %∆r ∆share %∆share ∆share %∆share
(%) (%) (%)
Small: [0,.1] 258 -0.0056 -0.18 0.23** 24.72** 0.32** 31.57**
(0.0032) (0.38) (0.07) (7.53) (0.09) (8.90)
Medium: (.1,.2] 21 0.0279** 2.98** -1.74** -18.46** -2.22** -22.13**
(0.0107) (1.13) (0.55) (5.80) (0.69) (6.74)
Large: (.2,.3] 12 0.0674* 7.22* -5.62** -35.01** -7.14** -42.26**
(0.0274) (2.93) (1.64) (9.96) (1.92) (11.20)
All 291 0.0246** 2.87** -0.16** 19.14** -0.17** 24.65**
(0.0095) (1.03) (0.05) (5.85) (0.05) (6.95)
Notes: Standard errors are presented in parentheses below each statistic. They are calculated using 200 Monte
Carlo simulations. Average interest rate changes are weighted by observed deposits. Signiﬁcance at the 5 and
1 percent levels is indicated by ** and *, respectively.
Table 9: Eliminating Surcharges
Concentration Measures
C(1) C(4) HHI
Before After ∆ Before After ∆ Before After ∆
Barnstable-Yarmouth 25.4 28.0 2.6 69.0 65.8 -3.3 1567 1561 -6
Boston 30.1 20.4 -9.8 54.0 41.0 -13.0 1293 703 -590
Brockton 11.2 8.9 -2.3 32.3 29.9 -2.3 704 664 -40
Fitchburg-Leominster 11.6 9.6 -2.0 37.2 30.8 -6.4 1404 1316 -88
Lawrence 26.2 23.0 -3.2 63.2 54.2 -9.0 1629 1415 -214
Lowell 12.6 13.2 0.6 45.3 43.7 -1.6 1045 1006 -39
New Bedford 32.9 27.6 -5.3 72.3 58.9 -13.4 2177 2023 -154
Pittsﬁeld 26.2 18.4 -7.9 64.9 54.1 -10.7 1972 1795 -177
Springﬁeld 16.1 10.6 -5.5 42.6 33.7 -8.9 882 709 -173
Worcester 18.3 14.2 -4.2 46.3 37.6 -8.6 1000 786 -214
Average 21.1 17.4 -3.7 52.7 45.0 -7.7 1367 1198 -169
Notes: Figures in the “before” columns use observed market shares. Figures in the “after” columns are calculated
from the scenario that eliminates surcharges and allows interest rates to adjust. The concentration ratios and
HHI exclude the outside good.
45
Table10:EliminatingSurcharges
Welfare
AllowingrtoadjustHoldingrﬁxed
Total(M$)Total(M$)
%NetworkN∆profit%∆profit∆profit∆profit∆profit%∆profit∆profit∆profit
totaldepositsATMfeestotaldepositsATMfees
Small:[0,.1]25844.1817.55**51.52*-7.34**55.13*21.90**62.20*-7.07**
(22.70)(6.65)(22.40)(0.31)(25.00)(7.14)(24.66)(0.37)
Medium:(.1,.2]21-11.03**-25.65**-8.20*-2.83**-11.07**-25.74**-8.21*-2.86**
(3.48)(5.21)(3.45)(0.03)(3.72)(5.70)(3.69)(0.04)
Large:(.2,.3]12-92.53**-46.49**-84.74**-7.79**-93.07**-46.76**-85.16*-7.91**
(32.20)(10.96)(32.07)(0.15)(33.35)(11.49)(33.20)(0.18)
All291-59.38**-12.03**-41.42**-17.96**-49.02**-9.93**-31.17*-17.84**
(13.11)(1.63)(13.24)(0.13)(12.11)(1.62)(12.26)(0.16)
AllowingrtoadjustHoldingrﬁxed
Mean($)Total(M$)Mean($)Total(M$)
∆CS8.404**49.318**3.114**18.275**
(2.025)(11.881)
Notes:Standarderrors(presentedinparenthesesbeloweachstatistic)arecalculatedusing200MonteCarlosimulations.
Signiﬁcanceatthe5and1percentlevelsisindicatedby*and**,respectively.Theunitoftimeissixmonths.
46
Table11:IntroducingaNetworkOperator
AverageoverﬁrmsTotal(M$)
%NetworkN∆r%∆r∆share%∆share∆profit%∆profit∆profit∆profit
(%)(%)totaldepositsATMs
Small:[0,.1]258-0.0095-0.590.21**23.01**77.70**37.78**52.65**25.06**
(0.0051)(0.67)(0.07)(7.78)(24.85)(8.01)(24.11)(3.97)
Medium:(.1,.2]210.0229*2.43*-1.70**-17.95**-2.46-7.56-7.87*5.42**
(0.0097)(1.02)(0.56)(5.88)(3.39)(9.58)(3.43)(0.92)
Large:(.2,.3]120.0696**7.46**-5.56**-34.49**-58.38-37.11*-84.70*25.33**
(0.0247)(2.64)(1.68)(10.21)(31.21)(14.65)(31.70)(3.41)
All2910.0230**2.70**-0.16*17.69**15.873.99-39.92**55.80**
(0.0079)(0.85)(0.05)(6.06)(13.24)(4.48)(11.48)(8.28)
Notes:Standarderrors(presentedinparenthesesbeloweachstatistic)arecalculatedusing200MonteCarlosimulations.
Signiﬁcanceatthe5and1percentlevelsisindicatedby*and**,respectively.Theunitoftimeissixmonths.
47
Table12:RobustnessChecks,FullModel
(1)(2)(3)(4)
Includingmulti-stateChangingdepositAddingcustomerInstrumentingfor
bankeﬀectsdistributionﬁxedcostsurch,#ATMs
Coeﬀ.Std.Err.Coeﬀ.Std.Err.Coeﬀ.Std.Err.Coeﬀ.Std.Err.
Demand
InterestRate×2.5627**0.3717
Depositterm
InterestRate×Income5.6494**1.06285.7029**0.97535.5770**0.9714
Surch4.3864**1.52555.8107**1.41984.3910**1.40674.4289*1.9055
#ATMs0.0025*0.00110.0075**0.00200.0022*0.00120.00290.0017
Dist1-0.9949**0.2337-0.01470.0716-1.1748**0.2817-0.7652**0.5623
Dist120.0131**0.00280.00060.00110.0153**0.0036-0.00490.0288
Dist2-0.3797*0.1797-0.1676**0.0605-0.3719*0.1738-0.59100.3465
Dist220.0059**0.00210.0022*0.00100.0061**0.00200.01820.0178
OneBranchDummy-2.3652**0.8135-3.5518**0.7175-2.1858**0.7419-2.9733**1.3624
Employees/Branch0.0192**0.00190.0154**0.00120.0198**0.00200.0190**0.0021
Multi-StateDummy
Mean-0.07620.2868
Std.Dev.0.00013.4381
MarketDummiesIncludedIncludedIncludedIncluded
MarginalCost:
Constant0.0197**0.00080.0199**0.00050.0208**0.00080.0197**0.0007
Customerﬁxedcost-25.195716.6663
N291291291291
Markets10101010
Notes:Thespeciﬁcationincolumn1allowsarandomcoeﬃcientonadummyformulti-statebanksbyincludingthedummyanditsintereraction
witharandomvariabledistributedstandardnormalacrossconsumers.Incolumn2,thedepositsareassumedtobedistributedexponential(αyi).
Incolumn3,thecostspeciﬁcationincludesaﬁxedcostpercustomer.Theresultsincolumn4instrumentforsurchargesandATMnetworksize.
Signiﬁcanceatthe5and1percentlevelsisindicatedby*and**,respectively.
48
Table13:RobustnessChecks,LogitDemandModel
(1)(2)(3)(4)
Coeﬀ.Std.Err.Coeﬀ.Std.Err.Coeﬀ.Std.Err.Coeﬀ.Std.Err.
InterestRate217.377**52.003205.077**49.349214.655**50.557214.248**51.815
Surch52.9331.544
#ATMs50.053**0.011
Sur0.3090.325
SurSUM0.4860.368
Surnon−SUM2.2571.177
Surch4.371**1.478
SurnojSUM0.3580.321
#ATMs0.008**0.0020.008**0.0020.007**0.002
AvgDist1-0.094*0.047-0.0870.048-0.0900.049-0.0930.048
AvgDist120.002*0.0010.002*0.0010.002*0.0010.002*0.001
AvgDist2-0.0840.047-0.139**0.045-0.131**0.045-0.091*0.045
AvgDist220.0010.0010.002*0.0010.002*0.0010.0010.001
OneBranchDummy-2.590**0.489-3.191**0.466-3.102**0.481-2.654**0.486
Employees/Branch0.014**0.0010.014**0.0010.014**0.0010.014**0.001
MarketDummiesIncludedIncludedIncludedIncluded
R20.8130.8170.8130.819
N291291291291
Markets10101010
Notes:Whiteheteroskedasticity-consistentstandarderrorsareused.Signiﬁcanceatthe5and1percentlevelsisindicatedby*and**,
respectively.Surch5issurchargecostsavoidedwithinaﬁve-mileradiusand#ATMs5isnumberofATMswithinaﬁve-mileradius.
49
Table 14: Robustness Checks, ATM Choice Model
Cost Simulated 95% Conservative 95%
Coeﬀ. Conﬁdence Interval Conﬁdence Interval
1. #In-branch ATMs 36,649 32,283 38,871 10,317 58,244
2. #Remote ATMs 38,348 26,179 47,292 5,677 71,387
3. #ATMs 53,032 42,359 57,110 -1,283 131,063
4. Business density -254 -410 -119 -1,189 273
5. #ATMs 15,272 -8,901 26,215 -22,850 44,867
6. #ATMs2 950 375 2,342 80 3,042
Notes: There are 291 banks in 10 markets in every speciﬁcation.
Figure 1: Proﬁt Diﬀerences
10000 20000 30000 40000 50000 60000 70000
0
1
2
Avg[R(n+1)−R(n)]
10000 20000 30000 40000 50000 60000 70000
0
1
2
Avg[R(n)−R(n−1)]
Notes: The graph on the left is a histogram of the market average of the change in a bank’s returns from n to n + 1
ATMs. The graph on the right is the analogous histogram for the change in returns from n − 1 to n ATMs.
50