Available online at www.sciencedirect.com
Journal of Banking & Finance 32 (2008) 1661–1676
www.elsevier.com/locate/jbf
Demand estimation and consumer welfare in the banking industry q
Astrid A. Dick
INSEAD Business School, Economics and Political Science Department, Boulevard de Constance, 77305 Fontainebleau, France
Received 8 June 2007; accepted 3 December 2007
Available online 15 December 2007
Abstract
This paper estimates a structural demand model for commercial bank deposit services in order to measure the effects on consumers
given dramatic changes in bank services throughout US branching deregulation in the 1990s. Following the discrete choice literature,
consumer decisions are based on prices and bank characteristics. Consumers are found to respond to deposit rates, and to a lesser extent,
to account fees, in choosing a depository institution. Moreover, consumers respond favorably to the branch staffing and geographic density, as well as to the bank’s age, size, and geographic diversification. Consumers in most markets experience a slight increase in welfare
throughout the period.
Ó 2007 Elsevier B.V. All rights reserved.
JEL classification: G21; L11; L89; C25
Keywords: Demand; Discrete choice; Consumer welfare; Product differentiation; Banking; Deregulation
1. Introduction
Following the removal of regulatory barriers to the geographic expansion of the banking firm, the US banking
industry experienced considerable growth and consolidation in the 1990s, with significant entry and exit. In particular, the 1994 Riegle-Neal Interstate Banking and
Branching Efficiency Act allowed for nationwide branching
by letting banks open branches in almost any US state, and
as such dramatically changed the strategic possibilities of
the firms in the industry.
The purpose of this paper is to measure the impact on
consumer welfare following significant changes in banking
services in the period. In order to measure consumer welfare, I develop a structural model of demand for commercial bank deposit services that allows not only for the
changes observed in prices, but also those in service charac-
q
This paper was reviewed and accepted while Prof. Giorgio Szego was
the Managing Editor of The Journal of Banking and Finance and by the
past Editorial Board.
E-mail address: [email protected]
0378-4266/$ - see front matter Ó 2007 Elsevier B.V. All rights reserved.
doi:10.1016/j.jbankfin.2007.12.005
teristics, such as the size of the branch network and the
geographic diversification.
While what interests us here is the effect of these changes
on consumers, regardless of their cause, it is nevertheless
interesting to review the background related to the removal
of geographic restrictions in banking. While no causal relationship can be established, the Riegle-Neal Act of 1994 is
likely to have played an important role in the expansion of
branch networks and other changes in bank prices and services throughout the 1990s. For many years, firms and government agents debated about the best regulatory
framework regarding the geographic expanse of a bank’s
activities. Those in favor of deregulation usually argued
that it would bring greater efficiency and competition
among banks, with resulting benefits to consumers. Those
against deregulation commonly alleged that the removal
of geographic restrictions would lead to highly concentrated banking markets and high profits in detriment of
consumer welfare. In terms of the theory, support can be
found for both views based on the different assumptions
one is willing to make about bank competition, such as
the degree of product differentiation and the nature of
the production technology. In previous empirical research,
1662
A.A. Dick / Journal of Banking & Finance 32 (2008) 1661–1676
the lifting of geographic restrictions in banking has been
linked to an improvement of economic conditions (Jayaratne and Strahan, 1996); bank performance and efficiency (Jayaratne and Strahan, 1998; Stiroh and Strahan,
2002); increase in service quality, costs and fees accompanied with no effect on market structure (Dick, 2006); significant bank entry (Amel and Liang, 1992); and an increase
in bank stability (Calomiris, 2000). In terms of the political
process of the phasing out of the heavy geographic regulation on banking activities, Kroszner and Strahan (1999)
find that small banks were the most resistant to branching
deregulation and therefore the most likely to suffer from it.
The industrial organization literature has gone a long
way in recent times in the estimation of structural models
of demand that take into account product differentiation,
and, given their microfoundations, are particularly useful
to address the effects from changes in policy or the market
environment. This paper estimates a discrete choice model
of demand for banking services by making use of some of
these techniques. While this paper was the first to implement this machinery to banking, much work has been
reported recently applying it to answer other important
policy questions in the industry. Adams et al. (2007) estimate deposit demand for banks as well as thrifts in order
to determine whether they are close substitutes, an important question for antitrust regulation given its implications
for the definition of the relevant geographic market. In her
exploration of ATM networks, Ishii (2005) estimates a
structural model of deposit demand and bank behavior in
order to determine the effects of surcharges – fees charged
to unaffiliated customers – on demand, ATM investment
and competition. Along a similar vein, Knittel and Stango
(2004) estimate a deposit demand to determine the effects
of ATM-fee induced incompatibility on ATM deployment.
Given a variety of banks in a market – defined as a
Metropolitan Statistical Area – a consumer is assumed to
choose one bank for deposit services. This decision depends
on the prices offered by the bank, checking account fees
and deposit interest rates paid, and non-price characteristics such as the size of the branch network, branch personnel, and geographic diversification. As a result, the model
can capture the net effect on consumers from the changes
in all of these features throughout the period.
Following the discrete choice literature, consumer preferences for bank services are identified from aggregate market shares across markets in the US by assuming a
distribution for the unobserved consumer taste. The discrete choice approach, by defining consumer preferences
over characteristics as opposed to actual products or firms,
incorporates product differentiation explicitly while avoiding the estimation of a large number of substitution parameters across firms. The model is estimated for the US
commercial banking sector over 1993–1999, using a data
set that combines information from several industry
sources. The Riegle-Neal branching deregulation occurred
between 1994 and 1997. This sample is chosen as 1993 predates the deregulation and 1999 follows it, thereby allowing
for changes in banking services to take place, while keeping
the link with deregulation strong.
Based on the estimation of logit-based models, the
results indicate that consumers respond to deposit rates,
and to a lesser extent, to account fees, in choosing a depository institution. Moreover, consumer demand responds
favorably to the staffing and geographic density of local
branches, as well as to the age, size, and geographic diversification of banks. The paper also finds important differences across markets in the demand for banking services,
with higher income areas being more responsive to prices
and bank size, and less to location characteristics, relative
to lower income areas. This could be related to a number
of factors, such as competition being less fierce and branch
networks smaller in lower income areas.
In light of the changes in bank services throughout the
period, I find that the net effect on consumer welfare is
positive in most markets. The consumer in the median market experienced a gain in welfare of $0.005–0.01 per dollar
(depending on the model), representing an annual gain of
$8–18 for a consumer with an average deposit balance.
Even in markets where prices increased, the improvement
in service characteristics usually made up for the detrimental effect of the price increase. As consumers are found to
value several bank attributes other than price, this exercise
is at least suggestive of the bias that might arise in welfare
inferences based solely on prices and concentration measures. In particular, the usual policy approach of focusing
on the price effects in the case of mergers might need to
acquire a broader perspective.
The paper is organized as follows. Section 2 provides an
overview of the banking industry and the deregulation. In
Section 3, the empirical framework is outlined. In Section
4, I describe the data and estimation. Results are presented
in Section 5, while Section 6 concludes.
2. The US banking industry: An overview
Throughout the last three decades, and particularly in the
1990s, the US banking industry underwent several changes
in both its structure and regulation. Regulatory restrictions
affecting the ability of banks to diversify geographically
and the range of products offered decreased dramatically.
Deregulation of unit banking and limited branch banking
occurred gradually throughout 1970–1994.1 In 1994, the
Riegle-Neal Interstate Banking and Branching Efficiency
Act was passed, permitting nationwide branching as of June
1997. As states had the option to ‘‘opt in” earlier than the
June 1997 federal deadline, the Act became effective gradually among the US states between 1994 and 1997.2 Banks
1
Intrastate branching deregulation began in some states even before the
1970s, while interstate banking (through Banking holding companies)
started as early as 1978. See Berger et al. (1995) for the evolution of the
industry throughout 1979–1994.
2
Only Texas and Montana opted out of the federal regulation, though
allowing for interstate branching with neighboring states.
A.A. Dick / Journal of Banking & Finance 32 (2008) 1661–1676
appear to have responded to these regulatory changes in a
significant way.3 The number of commercial banks
decreased by the thousands, reaching around 8000 in 1999
from over 11,000 a decade earlier, usually as the result of
merger and acquisition activity. The number of mergers
has averaged around 400 per year throughout the 1990s.
Throughout the same period, the industry increased to over
65,000 branches and 190,000 automated teller machines.
Moreover, the distribution of bank size changed. While the
share of assets of large banking organizations increased to
almost 30%, that of small banks decreased to less than
5%.4 Based on accounting rates of return, profit margins
were also high throughout the 1990s. As the sector increased
its concentration, return on equity remained above 15%.5
The sample period employed here covers the years
1993–1999. This sample is chosen as 1993 predates the deregulation and 1999 follows it, thereby allowing for a pre-deregulation and post-deregulation measure of welfare. This
allows for changes in banking services related to deregulation to take place, while keeping the link with deregulation
strong.6 Banking markets are defined as geographically
local, in particular, as Metropolitan Statistical Areas
(MSA). In the US there are more than 300 urban markets,
representing over 80% of US dollar deposits. While there
was significant increase in concentration at the regional level
and a sharp reduction in the number of banking institutions,
the structure of local banking markets remained virtually the
same throughout the period, as shown in Dick (2006). The
number of banks in most markets stayed the same, with
the average number in a market decreasing from 21 to 20
banks, and no change in the median – though given the
increase in regional concentration, there was more overlap
in the corporate identities of banks across markets in a given
region in 1999 relative to 1993. The average Herfindahl–Hirschman index (HHI) in an urban market was around 1900 in
1999, a slight increase since 1993.7 In terms of the entire distribution, concentration levels increased, usually slightly, in
only 47% of the markets, thereby decreasing slightly or stay3
The information described in this section has been constructed using
data on bank accounting and mergers and acquisitions from the Federal
Reserve Board, as well as branch deposit data from the Federal Deposit
Insurance Corporation.
4
Large banking organizations are defined as those with assets over $100
billion. Small banks are defined as those with assets below $100 million.
Despite the large number of banks in the US, the 10 largest banks hold
almost a third of national deposits.
5
The profit margins are used only in a descriptive manner here. See
Fisher and McGowen (1983) on the misuse of accounting rates of return.
6
Note that adding more years of data might be useful to estimate the
direct and long run effects of deregulation. However, here I am estimating
a demand model, and only indirectly drawing the link with deregulation.
Also, the demand model, estimated over the period, implicitly assumes
that consumer preferences do not change over the period, which is a
reasonable assumption as long as the sample period is small enough.
7
The HHI is a concentration measure constructed as the sum of the
squares of the market share of deposits at the local market level. Here,
following the practice of the Antitrust Division, I multiply it by a factor of
10,000, which is the index of a monopolist in a market. The Antitrust
Division defines the threshold of a highly concentrated market at 1800.
1663
ing the same in most markets. The average number of
branches, however, increased significantly, going from 131
to 140 on average in a local market.
Furthermore, bank characteristics also changed. Table 1
shows market averages for some bank attributes, for 1993
and 1999. Service fees increased throughout the period, but
deposit rates also increased (though not as much as shown
in the table if adjusting for the risk-free rate). Banks
increased their branches in local markets, as well as their
geographic expanse as they began operations in more
states. The number of employees per branch decreased,
probably as a result of the greater branch density per bank.
Not surprisingly, the average bank age rose and the distribution of bank size shifted to the right, usually as the result
of merger and acquisition activity.
3. Empirical framework
3.1. Definitions for a model of deposit services demand
3.1.1. Consumer decision
The demand model focuses on deposit services, which
include checking, savings, and time deposit accounts.
While one might model demand for these products separately, deposit services cannot be disaggregated at the
bank–market level. Detailed product data exist at the bank
level, but the only information collected at the branch level
is for total deposits.
While this is certainly a constraint, the evidence on consumer behavior suggests that it is reasonable to assume that
consumers cluster their purchases for deposits services
within one depository institution. Based on the Survey
of Consumer Finances, consumers show preference for
Table 1
Sample statistics: 1993 versus 1999 (before and after Riegle-Neal Act)
Variable
Prices
Service fees
Deposit interest rate
Service characteristics
Number of employees per branch
Branch density
Age of bank
Number of states of bank’s operations
Big (1 = yes)
Medium (1 = yes)
1993
Pre Act
1999
Post Act
0.63%
(0.18)
2.75%
(0.36)
0.67%
(0.16)
3.06%
(0.33)
32
(39)
0.011
(0.0196)
79
(24)
1.0
(0.1)
72.3%
(23.1)
16.6%
(16.9)
26
(20)
0.012
(0.0166)
87
(21)
4.0
(2.2)
82.0%
(14.9)
11.6%
(11.2)
Based on the deposit market share weighted averages. Standard deviations
in parenthesis. See Appendix for a description of the variables.
1664
A.A. Dick / Journal of Banking & Finance 32 (2008) 1661–1676
acquiring banking services together, in particular deposit
services, and appear to cluster their purchases with their primary financial institution.8 Evidence on small- and mediumsized business behavior provides similar insight.9 Therefore,
having to lump together these products should not be very
restrictive since the products are similar and consumers are
likely to perceive them as a bundle.
Alternatively, one could justify this approach by assuming that consumers have demands for multiple banking services, and incur a fixed cost for each new firm they have to
deal with. For sufficiently high fixed costs, consumers consolidate deposit services with a single bank. There is a
growing literature in banking that documents that consumers switching costs are significant, using consumer surveys
as well as firm and demographic data.10
A limitation of the data is that it does not allow us to
discriminate among the two main consumer groups that
make up the deposit data: households and nonfinancial
businesses.11 However, consumer and business surveys
indicate that the behavior of these two groups is similar.
Both appear to cluster their purchases, especially of deposit
services, with their primary financial institution, which is
typically situated close to their address.12
ken (1990) find that 93% of both small and medium-sized
businesses use a local commercial bank.
3.2. Relevant geographic market
3.4. Demand model
My approach is to define the relevant banking market as
geographically local, at the level of the Metropolitan Statistical Area (MSA), based on the available evidence on consumer purchases.13 Antitrust analysis has relied on the
definition of a banking market at the MSA level. Using
data from the Survey of Consumer Finances, Amel and
Starr-McCluer (2001) find that households obtain most services, especially checking accounts, at local depository
institutions. In particular, they find that 90% of checking
accounts, savings accounts and certificates of deposits are
acquired within the local market. Kwast et al. (1997) find
that over 94% of small businesses use a local depository
institution. Though a bit outdated, Elliehausen and Wol-
My model of deposit services demand is designed to
reflect as closely as possible the nature of consumer decision making in choosing a depository institution, given
the constraints in the available data. As mentioned earlier,
the format of the data drives a number of modelling
choices, but seems unlikely to significantly affect the interpretation of the results. Demand is derived following a discrete choice approach. By defining consumer preferences
over product characteristics, as opposed to specific products or firms, the approach avoids the estimation of a large
number of substitution parameters across firms.
Consumers are interested in purchasing deposit services
from a bank.17 Assume that t ¼ 1; . . . ; T markets are
observed, each with i ¼ 1; . . . ; I t consumers and
3.3. Output quantity and commercial bank competitors
I define market share on the basis of dollar deposit data
collected at each bank branch in the US.14 In modeling the
deposit demand for commercial banks, it is important to
consider purchases of deposit services from outside the
set of commercial banks, which can be achieved by allocating some consumers to an outside good. I use thrifts and
credit unions to build the market share of the outside good,
since, as depository institutions, are likely bank competitors.15 Note that this definition has its limitations, since
some people choose not to have a deposit account at all,
or other deposit alternatives than those considered here.
Another possibility is to define the potential size of the market, to capture the true outside good. There is a trade-off
between the measures: the current definition does not fully
account for the outside good, while the alternative relies on
the usually ad hoc estimation of the potential market size.
The results are actually robust to using the alternative definition of market share.16
8
See Amel and Starr-McCluer (2001). In fact, the share of households
having more than one service at their primary institution rose from 57% in
1989 to 64% in 1998. See also Elliehausen and Wolken (1992).
9
See Elliehausen and Wolken (1990) and Kwast et al. (1997).
10
Kiser (2002a) finds that the average household stays with the same
bank for ten years, and that the most frequently cited motivation for
changing banks is a household relocation. See also Kiser (2002b), Sharpe
(1997), Calem and Mester (1995) and Stango (2002).
11
At the aggregate level, more than two thirds of checkable deposits are
owned by businesses while 95% of time and savings deposits belong to
households. Within business customers, 75% of time and savings deposits
and 35% of checkable deposits are held by the nonfarm noncorporate
sector, while the rest is held by the corporate sector (percentages based on
the Federal Reserve Flow of Funds Accounts data for the year 1999).
12
Based on data from the Federal Reserve Board’s Survey of Consumer
Finances and Survey of Small Business Finances.
13
The format of the available data on deposits actually allows for any
definition of market, given that the dollar amount of deposits is available
for each bank branch in the US.
14
For a given bank, there is unfortunately no data on the number of
bank accounts per market. Data on bank accounts is reported only for the
bank as a whole. Defining market share by number of accounts would thus
require an allocation of accounts by market. In earlier work (Dick, 2002),
I use market shares defined in terms of accounts. The results are robust to
this alternative definition. Note that the current definition, based in
dollars, should capture the average of annual flows, including accounts
that open and close throughout the year.
15
According to Amel and Starr-McCluer (2001), 98% of households uses
a depository institution as of 1998. Thrifts and credit unions comprise less
than 6% in terms of US deposits as of 1999. Amel and Hannan (1999)
provides evidence for leaving these institutions outside the deposit product
market.
16
See Dick (2002).
17
I assume that consumers choose the proportion of assets they allocate
to deposit services prior to choosing a bank. This is a reasonable
assumption in light of the fixed costs a consumer incurs in dealing with a
given bank.
A.A. Dick / Journal of Banking & Finance 32 (2008) 1661–1676
j ¼ 0; . . . ; J t firms (where j ¼ 0 is the outside good). Let
the utility function take on a linear form such that the conditional indirect utility of consumer i from choosing bank
j’s services in market t is
uijt djt þ ijt pdjt ad psjt as þ xjt b þ nj þ ijt ;
ð1Þ
where pdjt and psjt represent interest rate paid by banks on
deposits and service charges on checking accounts,
respectively, xjt is a K-dimensional row vector of observed characteristics, nj represents unobserved bank
characteristics (depicted as a mean across consumers),
and ijt is a mean zero random disturbance. The K + 2dimensional vector hD ¼ ðad ; as ; bÞ represents the taste
parameters.
If we assume that the consumer heterogeneity term,
ijt , follows an extreme value distribution of the form
expð expðÞÞ, we can derive the market share for bank
j based on the probability that consumer i will choose j
conditional on bank characteristics (we drop market subscript for convenience). The predicted market share for
expðd Þ
bank j is then given by ~sj ðdÞ ¼ PJ j . Thus, the
k¼0
expðdk Þ
derived market shares depend only on mean utility levels
d, such that a simple structural relationship between the
marginal utilities and the observed market shares is
obtained. Following Berry (1994), by setting the predicted market shares equal to the observed market shares
and normalizing the mean utility of the outside good to
zero one obtains
lnðsj Þ lnðs0 Þ ¼ pdj ad psj as þ xj b þ nj :
ð2Þ
Given the simple linear model derived above, one can estimate the parameters in (2) with ordinary least squares, by
regressing lnðsj Þ lnðs0 Þ on ðxj ; pdj ; psj Þ, as well as deal with
the potential endogeneity of prices using standard linear
instrumental variables techniques.
While the above is certainly convenient, it imposes restrictive substitution patterns, as own- and cross-price elasticities
depend only on market shares. The nested logit model
reduces this problem by allowing consumer preferences to
be correlated within product categories. In particular, it
allows the distribution of consumer characteristics to
depend on the unknown parameter r, such that market
shares and the implied mean utilities vary with r as well. This
requires an a priori grouping of products into G + 1 exhaustive and mutually exclusive sets (including the outside good).
For product j 2 Gg , consumer i’s utility is given by
uij dj þ 1ig þ ð1 rÞij ;
ð3Þ
where 1ig is shared among products in the group and has a
distribution that depends on r 2 ½0; 1Þ. As r approaches
one, the correlation of utility across products in group g
approaches one as well. While this is certainly a more flexible model than the basic logit, it comes at the cost of an
increase in the parameters to be estimated and so the number of required instruments. See Berry (1994) for a detailed
outline of this model.
1665
To carry this estimation, I divide banks into two groups
based on their geographic diversification: multi-states
banks, which operate in more than one state, and banks
that have presence in a single state. This grouping should
capture important differences in geographic presence.
Banks that operate in no more than one state tend to be
in only a single local market within the state as well,
thereby being mostly ‘‘local” banks. Those that have presence in more than one state tend to operate in many local
markets within each state.18 Given such grouping, under
the nested logit model expression (2) becomes
lnðsj Þ lnðs0 Þ ¼ xj b þ pdj ad psj as þ r lnðsj=g Þ þ nj ;
ð4Þ
where lnðsj=g Þ represents the market share of bank j, which
belongs to group g, as a fraction of the total group share.
This term is clearly endogenous and as a result, instruments
are necessary to obtain a consistent estimate of r. As indicated in Berry (1994), the characteristics of other firms in
the group can be used for such purpose.
4. Data and estimation
4.1. Data
The data come from several sources. The data on bank
characteristics derived from balance sheet and income
statement are taken from the Report on Condition and
Income (‘‘Call Reports”) from the Federal Reserve Board.
The data on branch deposits used in the construction of
local market shares, as well as the number of branches,
are obtained from the Federal Deposit Insurance Corporation (FDIC). Demographic data at the MSA level are taken
from both the US Census and the Bureau of Economic
Analysis. The sample covers the period 1993–1999.19 All
urban markets in the territorial US are included in the sample, with 318 MSA markets per year.20
An observation is defined as a bank–market–year combination in the estimation exercises. The bank attributes
are chosen based on the available data and on the belief
that they are important and recognizable to the consumer.
Summary statistics are provided in the Appendix, which
also contains a description of the variables. Because most
of the available data measure observed bank characteristics
at the bank rather than bank–market level, most attributes
for banks that operate in more than one market offer no
market variation within a given year. For instance, price
18
Note that while nationwide branching deregulation began in 1994,
even as early as 1993 (the beginning of the sample), most states had
agreements with neighboring states that allowed for banks to cross state
borders.
19
The data are taken from the second quarter reports of each year. I
choose this quarter because one of the variables of interest is only reported
then.
20
For 1999, for instance, the largest MSA in the sample is Los AngelesLong Beach, CA, with more than 9 million people, and the smallest is
Enid, OK with approximately 57,000. The average MSA population is
around 660,000.
1666
A.A. Dick / Journal of Banking & Finance 32 (2008) 1661–1676
for bank j is the same across all markets within a year. This
may be less restrictive than it first appears. Radecki (1998),
for example, finds that the current practice among banks in
New York and other large states is to set uniform retail
deposit and consumer loan rates across an entire state or
large regions of a state. The general conclusion has been
that multi-market banks set uniform rates over large geographic areas, allowing relatively little autonomy to their
branches (Park and Pennacchi, forthcoming).21
Two prices are observed: service charges on checking
accounts and the interest rate paid on deposits. Following
common practice, these are imputed from deposit revenues,
in the case of service charges, and from deposit expenses, in
the case of the rate paid on deposits, adjusted by the stock
of deposits.22
A central part of this work is to illustrate the extent to
which consumers view banks as heterogeneous. Several
observed bank characteristics are included, starting with
the number of local branches per square mile.23 I also
include the number of employees per branch, which might
be important to consumers through its correlation with
waiting time. The variable might also capture the value
of human interaction to consumers who find technological
access to their bank more intimidating, and/or the types of
services specific to bank branches.
I control for bank size through a set of dichotomous
variables: large, medium, and small (the latter omitted in
the regressions).24 The reason for including these is not
to capture size itself, but rather those features associated
with larger banks, such as larger infrastructure, diversity
of products, and know-how. I also introduce the number
of states in which the bank has presence, which should
measure the value attached to network size and geographic
diversification. The age of the bank, measured by the number of years since the bank first began operations, is
intended to capture the importance of branding and the
perceived degree of experience and expertise of a bank.
4.2. Instruments
Prices are likely to be correlated to the unobserved bank
characteristics n, since while the researcher does not
observe the values of n, banks and consumers do.25 Unobserved bank characteristics are variables such as the bank’s
service quality, reputation for customer service and financial soundness and prestige.
Identification of the demand parameters is obtained
from the variation in bank–market shares corresponding
to the variation in bank characteristics. Assume that the
demand unobservable is mean independent of both
observed bank characteristics and cost shifters. Letting
both observed bank characteristics x and cost shifters w
enter the matrix of instruments z, one has
E½nj jz ¼ 0:
Bank characteristics are taken as given and therefore provide instruments for themselves. Supply side variables that
shift a bank’s marginal costs as well as variables that shift
markups are used as instruments for prices.
In the set of costs shifters I include variables related to
four main components of marginal costs: labor, rental
and other operating costs, funding costs, and several environmental variables to capture differences in marginal costs
from different institutional characteristics.26 Local labor
costs come from wage data from the Bureau of Economic
Analysis.27 Rental prices are proxied by the housing
price index from the Office of Federal Housing Enterprise
25
21
Seventy percent of banks actually operate in a single local market, and
as a result their headquarter’s data exactly fit local market data.
22
Data on actual interest rates is sometimes collected by surveys, but
based on a small sample of around 300 banks.
23
Given lack of data at the bank level, I do not include the number of
ATMs, even though it is likely to be relevant. As consolation, we might
consider the fact that the correlation between the number of ATMs and
branches of a given bank is around 80%, based on a sample of almost 1500
bank–market observations in 1998 obtained from CIRRUS (a large ATM
network).
24
The size categories are defined as follows: banks with assets between
100 and 300 million are medium-sized, while those with assets above 300
million are large. The definition is based on the FFIEC form that banks
report to the regulatory authority. The market share a bank has in a given
market should have no feedback effect on its size category. There is also
almost no variation in terms of which size category a bank belongs to
across the years in the sample, as only 17% of banks ever change category
in the sample. Moreover, while there is variation in terms of the market
shares that a given bank has across the local markets it serves, size is
defined at the bank level (and in terms of assets as opposed to deposits).
ð5Þ
We assume that the rest of the observed bank characteristics, unlike
prices, are not correlated to unobserved demand shocks. On the one hand,
prices can be changed rather rapidly, and when banks decide prices, they
are likely to take the quality level as given. Thus, any unobserved demand
shock (say, as a result of the bank’s advertising campaign) is likely to be
correlated with price, and that is why instrumental variable estimation is
required. On the other hand, observed bank attributes such as the branch
network and geographic coverage, take longer to be modified. Indeed, it is
understood that in the long run, all variables are endogenous, but it is this
difference in the dynamics of price and, say, the branch network across
states and in a local area, that makes the assumption of exogeneity on the
latter a reasonable one.
26
Whenever these instruments are based on local market variables, they
are constructed as the weighted average over the markets in which the
bank operates (where the weight is the bank’s deposit share in each market
out of its total deposits). Prices are available at the bank level and are
likely to respond proportionally to the local conditions in those markets.
27
While the Call Report also contains salary data, these are likely to
include a hidden quality component, since more skilled employees will
tend to be more expensive, thus violating the independence assumption.
Also, note that the correlation between the mean wage and the bank teller
wage (a better measure of bank labor costs) in a market is around 70%,
based on data from the Occupational Employment Statistics of the Bureau
of Labor Statistics. I use the mean wage instead because the bank teller
wage is available for only the last three years of my sample.
A.A. Dick / Journal of Banking & Finance 32 (2008) 1661–1676
Oversight, as well as city density, which should be also correlated with rental costs.28 To control for other operating
expenses, I use the expenses on premises and fixed assets,
usually referred to as the occupancy rate, from the Call
Report, which include expenses on lease payments, depreciation, utilities, building maintenance, legal fees, insurance, amortization of assets, and ordinary repairs.
Funding costs from borrowed money other than deposits
is measured by the market-average price of purchased
funds from the Call Report, as well as credit risk, measured
as the average of non-performing loans.29 As environmental variables to control for differences in marginal costs due
to diverse production characteristics, I use the following:
an indicator variable for whether the bank belongs to a
banking holding company (banking holding companies
should provide easier access to funds), the proportion of
commitment loans, which are unused credit lines, and affect
the production technology of the firm in planning its
resource allocation to manage loan demand on call, the
degree of bank capitalization, as measured by the ratio of
equity over assets, and an indicator variable for whether
the bank operates in at least one rural area.30
In the set of markup shifters, I use the standard variables
in the discrete choice literature, which are the characteristics
of other products in the market as instruments for price (see
Berry, Levinsohn and Pakes, 1995). I refer to these as BLP
instruments. This relies on the assumption that product
characteristics other than price are exogenous, and therefore orthogonal to unobserved demand. Given the location
of products on the characteristics space, price will be correlated with the characteristics of other products. The argument is that products that have close substitutes will have
lower markups while other products located further away
from rival ones will have higher prices relative to cost.
As mentioned in Section 3.4, the nested logit model,
while adding flexibility to the demand model, requires addi-
28
The correlation between rental rates per square foot and the housing
price index is close to 50% based on a sample of 62 cities over 20 years (I
thank Thomas Davidoff and Ashok Bardhan for providing me with NREI
data). Nevo (2001) uses city density similarly in his study of demand for
breakfast cereal.
29
Purchased funds include federal funds, subordinated notes, demand
notes issued to the US Treasury, trading liabilities and other borrowed
money. As far as credit risk, one situation in which such would not be an
appropriate instrument is if it stands for a particular type of specialization
that consumers find valuable. It is unlikely that credit risk is of significance
to potential depositors, given that deposits are insured in the US up to
$100,000 by the FDIC for member banks (the sample contains only FDIC
insured banks). A more subtle story where the exogeneity assumption
would be violated is for the case of banks engaging in mass marketing of
specific consumer risk types. Even though consumers might not be aware
of the bank’s risk portfolio, prices could still be correlated with the
unobserved demand component. This occurrence, however, is not
expected to be prevalent.
30
All of these variables should have little meaning from the perspective
of a potential depositor, who is unlikely to be aware of them. As a result,
they should not be correlated with unobserved demand shocks.
1667
tional instruments for the identification of the additional
parameter r. Following the literature, I use the characteristics of other banks in the nest. In particular, I use the
branch density, branch personnel, bank age, and size of
other banks in the group as instruments for the withingroup share of a given bank.
5. Results
Table 2 presents the estimation results. All columns correspond to the logit model, except for the last two, based
on the nested logit model. The standard errors are adjusted
for heteroskedasticity and for the fact that there may be
correlation between errors of the same bank. All the regressions include year and regional effects (state or market
dummy variables), with some specifications introducing
bank fixed effects as well. The dependent variable in all
of the specifications is a function of the bank’s market
share and the share of the outside good, where the latter
is defined as credit unions and thrifts. In particular, the
logit specification reduces the model to the estimation of
lnðsjt Þ lnðs0t Þ on prices and bank characteristics, where
sjt is bank j’s market share in market t and s0t is the market
share jointly held by credit unions and thrifts in the market.
Columns (i), (ii) and (iii) show results for the OLS estimation, while the rest (columns iv–viii), present those for the
IV model, where the price variables, service fees and the
deposit rate, are instrumented for. Throughout the columns, different fixed effects are included, as indicated at
the bottom of each column. The various columns are examined to illustrate the importance of instrumenting for prices
and controlling for unobserved bank characteristics.31
Both the OLS and the IV model show that consumers
respond to prices, as their coefficients are highly significant
and of the expected sign. Service fees enter utility negatively while the interest rate received by consumers on their
deposits enters positively.
5.1. OLS regressions
The regression in column (i) includes observed bank
characteristics but not bank fixed effects, so that the error
term includes the unobserved bank characteristics earlier
referred to as nj . Given the logit demand structure, which
allows each bank to have a different price elasticity, the
mean of the distribution of own-price elasticities across
the sample observations is 0.15 for service fees and 0.30
for the deposit rate. By introducing bank dummy variables,
columns (ii) and (iii) control for this unobserved firm
31
Note that all the results are robust to various exercises, such as the
removal of certain bank characteristics and instruments. They are also
robust to the use of an alternative measure of deposit market share based
on the estimation of the potential market size and average deposit balance,
as previously discussed.
1668
A.A. Dick / Journal of Banking & Finance 32 (2008) 1661–1676
Table 2
Estimation results
Explanatory variable
Service fees
Deposit rate
Employees per branch
Branch density
Bank age
Number of states
Big
Medium
OLS
IV
(i)
(ii)
(iii)
(iv)
(v)
(vi)
(vii)
(viii)
26.557
(6.877)**
10.276
(2.535)**
0.001
(0.000)**
32.606
(5.854)**
0.006
(0.001)**
0.068
(0.019)**
1.957
(0.070)**
0.927
(0.050)**
22.456
(5.635)**
7.624
(0.930)**
0.001
(0.000)*
19.762
(4.235)**
0.000
(0.002)
0.046
(0.018)**
0.530
(0.064)**
0.297
(0.034)**
18.308
(5.186)**
6.903
(0.878)**
0.001
(0.000)*
51.797
(7.222)**
0.003
(0.003)
0.046
(0.017)**
0.439
(0.056)**
0.270
(0.032)**
89.033
(21.826)**
100.227
(10.472)**
0.001
(0.000) 34.966
(6.399)**
0.007
(0.001)**
0.115
(0.024)**
2.053
(0.077)**
1.190
(0.063)**
88.134
(18.579)**
77.255
(6.458)**
0.001
(0.000)*
19.831
(4.239)**
0.003
(0.002)
0.009
(0.018)
0.419
(0.064)**
0.404
(0.037)**
66.816
(15.916)**
78.062
(5.744)**
0.001
(0.001)*
51.703
(7.225)**
0.001
(0.003)
0.012
(0.017)
0.329
(0.056)**
0.383
(0.033)**
74.259
(18.747)**
54.194
(6.781)**
0.001
(0.000)
14.283
(4.300)**
0.001
(0.002)
61.862
(16.544)**
59.178
(5.897)**
0.001
(0.001)**
41.045
(7.499)**
0.001
(0.002)
0.296
(0.067)**
0.276
(0.041)**
0.291
(0.042)**
0.245
(0.058)**
0.284
(0.036)**
0.223
(0.036)**
44,948
0.36
State
44,948
0.77
State
Bank
44,948
0.88
Market
Bank
44,948
44,948
44,948
44,948
44,948
State
State
Bank
Market
Bank
State
Bank
Market
Bank
lnðsj=g Þ
Observations
R-squared
Fixed effects
Time effects are included in all specifications. Estimated standard errors, robust and corrected for within bank dependence, are in parentheses. The logit
model consists of the estimation of the log difference of bank j’s market share sj and the outside good’s market share s0 ðlnðsj Þ lnðs0 ÞÞ on prices and bank
characteristics; the nested logit model is similar but also includes, as an explanatory variable, the market share of bank j, which belongs to group g, as a
fraction of the total group share. See text and Appendix for description of variables and instrument sets.
Significant at 10%.
*
Significant at 5%.
**
Significant at 1%.
component, and as a result the R-squared more than doubles in both cases. Column (iii) includes market instead of
state fixed effects, to control for time-invariant conditions
at the market level, such as demographic characteristics.
Throughout all of these OLS regressions, the price coefficients are similar and rather low in terms of their implied
elasticities (even lower elasticities for (ii) and (iii)).
5.2. IV regressions
The rest of the columns present the two-stage estimation
results, which are of most interest to us here. If prices are
higher when unobserved quality is higher, for instance, one
might not observe market share respond to higher prices,
and therefore instrumenting for prices is crucial if the
demand parameters are to be consistently estimated. Column (iv) is similar to the specification in column (i), but service fees and the deposit rate are now instrumented for. Not
surprisingly, the coefficients on price increase substantially,
in particular the deposit interest coefficient which increases
almost tenfold, with the mean of the distribution of ownprice elasticities increasing to 0.50 in the case of service fees
and 2.96 in the case of the deposit rate. The latter is particularly remarkable, since now more than 95% of the deposit
rate elasticities are elastic, compared to none being elastic
under OLS estimation. This is quite reassuring. While there
are several potential reasons why it might be reasonable for
the service fee elasticity to be below unity, a deposit rate elasticity below unity would be more worrisome, as will be later
discussed. The introduction of instrumental variables generates considerable increase in the absolute magnitude of the
price coefficients as would be expected if what is in the error
term is something that increases consumer valuation of the
bank, such as quality. Also, note that the introduction of
instrumental variables leaves the rest of the coefficients more
or less unchanged.
Column (v) introduces bank dummy variables (over 6000)
to fully control for the unobserved bank characteristics.
Specifying a bank fixed effect removes bank characteristics
from the error term and is particularly important because,
given data constraints, one might expect that many characteristics relevant to the consumer’s decision will not be
accounted for explicitly in the empirical model, such as
advertising and product variety. Following Nevo (2001),
suppose one lets the error term be instead nj þ Dnjts þ ijts ,
where s is now a time subscript. A bank dummy variable
would capture the unobserved bank characteristics that do
not vary across market–year combinations within a given
bank, nj . Note, however, that even when bank fixed effects
are introduced, the time deviation from mean demand char-
A.A. Dick / Journal of Banking & Finance 32 (2008) 1661–1676
acteristics, Dnjts , is left in the error term. Since the latter
should be correlated with price, one still needs to instrument
for this leftover component. Whenever bank dummies are
introduced in this setup, the fit of the model more than doubles, with the model explaining around 80% or more of the
variation in market share.
Column (vi) is similar but replaces state fixed effects with
over 300 market dummies. Controlling for local characteristics is important if consumer valuations are correlated with
demographic characteristics at the local market level. Thus,
introducing a dummy variable for the market could potentially get rid of omitted variable bias. When market instead
of state dummies are specified, the R-squared increases from
0.77 to 0.88 (similar to OLS case). Note, however, that the
price coefficients are similar (not statistically different from
each other), regardless of whether we introduce state or market fixed effects here. Moreover, they all increase considerably relative to their OLS counterparts.
1669
to 0.1–0.2% increase in market share.32 Bank size, which
might capture product diversity and other service differences
associated with large banks (not controlled elsewhere), is
always highly significant and enters utility positively, indicating that the value of a bank relative to the outside option
increases with larger bank size. The number of employees at
the branch is always positive and usually significant at the 5%
level. Bank age, which might capture bank experience and
branding, as well as geographic diversification (as measured
by the number of states of operation), both enter utility positively and significantly in the baseline specifications, but
usually lose significance when bank fixed effects are introduced. This is not surprising, given that there is little variation left in the variables to identify the coefficients once
bank fixed effects are included. Indeed, given their nature,
one should interpret these coefficients from the regressions
without bank fixed effects.
5.5. First-stage regressions
5.3. Nested logit model
The nested logit model is more flexible than the logit, as it
allows for interactions between product and consumer characteristics while remaining fairly simple (though a nesting
structure must be chosen a priori). Consistent estimation
of the r parameter does require additional instruments, as
discussed earlier. In particular, I use the BLP-type instruments, that is, the characteristics of other banks in the
group, as an instrument for the within-market share of a
given bank. Columns (vii) and (viii) of Table 2 report results
from estimating the IV nested logit model. Both include
bank fixed effects and regional effects (state dummies in
(vii) and market dummies in (viii)). The results in terms of
price sensitivity are similar to the logit results. Note that
the variable on geographic diversification has now been
dropped as an explanatory variable, since this is the criterion
used in the nesting strategy. The correlation parameter r is
significant and precisely estimated, with a value between
0.2 and 0.3, indicating that the nesting strategy is applicable.
In particular, consumer preferences appear to be correlated
across the set of multi-state banks, on the one hand, and
across the set of ‘‘local” banks, on the other. Importantly,
the finding suggests there is some market segmentation
between multi-state banks and those that operate within a
single state, as consumers appear to group them separately.
Adams et al. (2007) also find limited substitutability between
geographically diversified banks (defined as multi-market,
as opposed to multi-state) and local banks (defined as single-market).
5.4. Bank characteristics
In terms of the observed bank characteristics, branch density is always highly significant and enters utility positively,
as expected. For instance, the average elasticity suggests that
a 1% increase in branch density in an urban area would lead
The first-stage regressions show a reasonable fit, with the
exogenous variables explaining at least 30% of the variation
in prices (results shown in the Appendix), with the fit increasing to up to 90% when bank fixed effects are introduced. Price
instruments are usually significant and of the expected sign.
Fees increase with wage increases, as well as with operating
expenses, and fall with an increase in the cost of funding.
Deposit rates paid to consumers are negatively associated
with wages and expenses, and positively correlated with the
costs of funds. Among other relationships, both fees and rates
appear to be negatively related to the level of bank capitalization, and positively associated with the level of credit risk.
5.6. Price elasticities
To facilitate the interpretation of the coefficient magnitudes, Table 3 presents the distribution of price elasticities
based upon the estimates obtained earlier for the logit and
nested logit model.33
The median elasticity of service fees is negative and
between 0.3 and 0.4, while the median deposit rate elasticity is positive and between 1.8 and 3.0. In other words, a
1% increase in service fees should lead to a 0.3–0.4%
32
As a robustness exercise, I use Cirrus data to estimate the model
including the number of ATMs of the bank as an explanatory variable
(though the sample is significantly reduced). It is comforting to find that
the results are robust to the inclusion of the number of ATMs of the bank
(and in particular in terms of the magnitude of the price coefficients), and
that the coefficient is positive and significant at the 1% level. I thank Beth
Kiser, Timothy Hannan and Robin Prager for providing the raw Cirrus
data.
33
Under the logit model, the price elasticities are
apj ð1 sj Þ ifj ¼ k;
osj pk
ð6Þ
gjk ¼
¼
opk sj
otherwise:
apk sk
For the derivation of elasticities under the nested logit model, see Berry
(1994).
1670
A.A. Dick / Journal of Banking & Finance 32 (2008) 1661–1676
Table 3
Price elasticity percentiles
Table 4
Estimation results
Price
10%
25%
Median
75%
90%
Service fees
Logit (state FE)
Logit (mkt & bk FE)
Nested (mkt & bk FE)
0.87
0.65
0.60
0.64
0.48
0.44
0.44
0.33
0.31
0.29
0.22
0.20
0.17
0.13
0.12
Deposit rate
Logit (state FE)
Logit (mkt & bk FE)
Nested (mkt & bk FE)
1.94
1.51
1.14
2.44
1.90
1.44
2.99
2.33
1.77
3.52
2.74
2.08
3.98
3.10
2.35
The entries correspond to the indicated percentiles of the distribution of
own-price elasticities across banks in the sample, based on the estimation
results shown in Table 2, columns (iv), (vi) and (viii).
decrease in the bank’s market share in the local market,
while a 1% increase in the deposit rate paid by the bank
should lead to around a 2% to 3% increase in its market
share. Elasticities are usually lower in the nested logit
model, though, reassuringly, they are very similar across
all IV models.34 Adams et al. (2007) find higher deposit rate
elasticities across MSA markets, with a median of 3.47.
Note, however, that they do not include service fees in their
estimation, and this could potentially account for the difference if consumers take both prices into account when
making a purchase decision.
The elasticity of service fees, which is below unity, is
clearly not profit maximizing: given this elasticity value,
banks should respond by raising fees until they reach an
elastic portion of the demand curve. However, there are
many plausible explanations for why banks might choose
not to raise service fees. First, there is abundant anecdotal
evidence about banks using low or zero-fee checking
accounts as ‘‘loss leaders”, that is, as a way to attract
and lock in consumers that will later on proceed to purchase other services offered by the bank. Second, banks
may consider raising service fees as too conspicuous to
both consumers as well as regulators. Yet another possibility is that the relevant price elasticity is simply that of the
bundle of all deposit products.
5.7. Market differences
Table 4 shows estimation results based on splitting the
sample by income per capita at the market level.35 The
results are shown for the logit and the nested logit specifi34
This might suggest several things, such that the nests are not very good
(the correlation parameter, while estimated precisely, is rather low), that
the additional instruments used to identify the parameter are not good
enough, or simply that elasticities are lower (if the model is better than the
basic logit in explaining reality).
35
Note that the number of observations between the two subsamples
differs greatly since on average there are, not surprisingly, many more
banks in high income markets, as compared to low income markets. Note
that similar results are obtained if the sample is divided by population
density, with little difference between high (low) population density areas
and high (low) income areas.
Explanatory variable
Service fees
Deposit rate
Employees per branch
Branch density
Bank age
Number of states
Big
Medium
High income
(i)
(ii)
(iii)
(iv)
41.446
(14.782)**
79.058
(4.990)**
0.001
(0.000)*
41.464
(5.694)**
0.002
(0.002)
0.030
(0.013)*
0.381
(0.054)**
0.380
(0.029)**
53.492
(16.532)**
63.757
(5.607)**
0.001
(0.000)*
35.108
(5.837)**
0.001
(0.002)
25.101
(37.897)
67.262
(15.178)**
0.013
(0.003)**
194.770
(11.659)**
0.038
(0.023) 0.022
(0.017)
0.198
(0.129)
0.345
(0.086)**
15.861
(36.971)
49.792
(14.637)**
0.011
(0.003)**
149.637
(16.415)**
0.023
(0.024)
lnðsj=g Þ
Observations
Fixed effects
Low income
30,817
Market
Bank
0.313
(0.056)**
0.302
(0.033)**
0.165
(0.032)**
30,817
Market
Bank
14,131
Market
Bank
0.128
(0.131)
0.242
(0.088)**
0.247
(0.063)**
14,131
Market
Bank
Time effects are included in all specifications. Estimated standard errors,
robust and corrected for within bank dependence, are in parentheses. The
results are based on two subsamples defined as high and low income areas.
All regressions are estimated by two-stage least squares. The logit model
consists of the estimation of the log difference of bank j’s market share sj
and the outside good’s market share s0 ðlnðsj Þ lnðs0 ÞÞ on prices and bank
characteristics; the nested logit model is similar but also includes, as an
explanatory variable, the market share of bank j, which belongs to group
g, as a fraction of the total group share. See text and Appendix for
description of variables and instrument sets.
Significant at 10%.
*
Significant at 5%.
**
Significant at 1%.
cations with both market and bank fixed effects. This exercise suggests that consumers in high income markets are
more responsive to prices than consumers in low income
areas.36 Another stark difference between the groups is in
terms of the marginal utilities for branch density and number of employees at the branch, with low income areas
responding much more strongly to them. For instance,
the mean elasticity for branch density is 0.3% for low
income areas versus 0.03% in high income areas. These
results might be related to the fact that low income areas
tend to have smaller branch networks and branches, so that
an additional branch/bank employee has a greater impact
on consumer utility there.37 It could also be that wealthier
consumers rely less on branches to carry out financial
transactions with their bank. High income areas might also
36
Note that within a market there is likely a lot of variation in income.
Thus, the regressions that separate markets into low and high income are
rather a crude way to obtain how income affects choices. Instead, one
might view this exercise as simply cutting the data by a given market
characteristic, which in the least, helps explore the stability of coefficients.
37
Dick (2007) finds that larger and richer markets offer denser branch
networks.
A.A. Dick / Journal of Banking & Finance 32 (2008) 1661–1676
1671
Table 5
Local market consumer welfare change percentiles (1993–1999)
Logit (state FE)
Logit (mkt & bk FE)
Nested (mkt & bk FE)
10%
25%
Median
75%
90%
$0.0030
(0.0009)**
$0.0010
(0.0006)
$0.0013
(0.0004)**
$0.0070
(0.0004)**
$0.0024
(0.0003)**
$0.0038
(0.0002)**
$0.0106
(0.0003)**
$0.0056
(0.0004)**
$0.0064
(0.0003)**
$0.0143
(0.0003)**
$0.0091
(0.0003)**
$0.0092
(0.0004)**
$0.0185
(0.0007)**
$0.0127
(0.0008)**
$0.0126
(0.0004)**
The entries correspond to the indicated percentiles of the distribution of consumer welfare change from 1993 to 1999 across banking markets, based on the
P
equivalent variation calculation: EV ¼ S s ðp0 ; x0 ; hD Þ S s1 ðp; x; hD Þ where Sðp; x; hD Þ ¼ ln½ Jj expðdj ðpj ; xj ; hD ÞÞ=a. Estimates are based on columns (iv),
(vi) and (viii) of Table 2. Bootstrap standard errors are in parenthesis (based on 100 repetitions).
*
Significant at 5%.
**
Significant at 1%.
experience more fierce competition among banks.38 In
addition, wealthier consumers appear to value more heavily larger-sized banks, also a reasonable result as one might
expect bigger banks to be able to offer features that higher
income clients may find particularly useful, such as wider
product diversification and service expertise, in which larger banks might have a competitive advantage.
change in the sample period. Following Small and Rosen
(1981), in the context of the discrete choice model, welfare
effects of changes in the choice set between periods s and
s 1 in a given market are measured as the expected equivalent variation ðEV Þ of the changes. The latter is defined as
the amount of money that would make consumers indifferent, in expectation, between facing the two choice sets.
Then, one has that
5.8. Consumer welfare effects
EV ¼ S s ðp0 ; x0 ; hD Þ S s1 ðp; x; hD Þ;
The passage of the Riegle-Neal Act in 1994 allowed
banks to open branches in almost the entire US territory.39
This opened up the opportunity for banks to redefine their
position on the differentiation space, as a bank could now
choose to remain local or provide a larger, even national
network to its clients. One would expect banks to respond
to the deregulation, and indeed, the entry and exit that
ensued in banking markets is evidence of that. With such
dramatic reorganization of the firms in the market, one
natural concern is the effect of these changes on consumers.
Our results suggest that product differentiation matters in
banking, as consumers value several bank attributes other
than price. Following the shake-out in the industry, we
cannot hope to provide any reasonable answer on consumer effects by focusing solely on price changes. As seen
earlier in Table 1, while service fees have increased slightly
since 1993, other bank attributes such as the number of
branches (even controlling for population growth) and
the geographic expansion have also increased. While a
price increase affects consumers adversely, the expansion
in other service dimensions enhances consumers’ utility,
such that the net effect on consumer welfare is ambiguous.
Thus, to explore the effect on consumers of the changes
in banking services, I carry out a calculation of welfare
38
Consumers in richer markets also appear to value relatively more the
bank’s geographic diversification and bank age, based on the (unreported)
specifications without bank fixed effects, where it is possible to correctly
interpret the coefficients on these bank characteristics, as mentioned
earlier. This might suggest that the relevant bank market is larger for
higher income clients.
39
Montana and Texas opted out of the federal regulation, but allowed
interstate branching with some neighboring states.
ð7Þ
P
where Sðp; x; hD Þ ¼ ln½ Jj expðdj ðpj ; xj ; hD ÞÞ=a.40
Table 5 shows percentiles for the distribution of welfare
change in local markets between 1993 and 1999, that is,
pre- and post-Riegle-Neal Act, based on the expected
equivalent variations using the choice sets facing consumers before and after deregulation. It is worth noting that
while many changes in bank product attributes, such as
the expansion of the branch network, could only have happened with deregulation, they could also have responded to
other factors changing throughout the period as well.
Clearly, the exercise does not intend to establish a definite
causal relationship going from deregulation to consumer
welfare, but only recognizes the likelihood that deregulation played a central role in the changing attributes and
choices available to consumers throughout the period
(what interests us here is the effect of changes in prices
and characteristics, regardless of why they changed, on
consumers). The percentiles in the table are shown for three
different models, and the results are very similar among
them. Based on the distribution of welfare change, I find
that consumers in most markets (over three quarters of
the distribution) experience a gain in welfare, with a mean
of $0.005–0.01 per consumer per year, and the changes are
statistically significant. The annual deposit account balance
for the average consumer can be estimated to be around
$1700 (based on an estimated and statistically significant
coefficient of 0.065 between personal income and deposit
balances from the Survey of Consumer Finances carried
by the Federal Reserve Board, and the average income
40
The marginal utility of income, represented by a, will be the coefficient
on service fees in our calculations of welfare.
1672
A.A. Dick / Journal of Banking & Finance 32 (2008) 1661–1676
Table 6
Local market consumer welfare change percentiles (1993–1999) – if only prices allowed to change
Logit (state FE)
Logit (mkt & bk FE)
Nested (mkt & bk FE)
10%
25%
Median
75%
90%
$0.0156
(0.0017)**
$0.0153
(0.0016)**
$0.0138
(0.0010)**
$0.0087
(0.0004)**
$0.0095
(0.0006)**
$0.0176
(0.0007)**
$0.0046
(0.0004)**
$0.0046
(0.0004)**
$0.0035
(0.0003)**
$0.0010
(0.0004)**
$0.0013
(0.0003)**
$0.0003
(0.0003)
$0.0019
(0.0003)**
$0.0010
(0.0002)**
$0.0021
(0.0004)**
The entries correspond to the indicated percentiles of the distribution of consumer welfare change from 1993 to 1999 across banking markets, based on the
P
equivalent variation calculation: EV ¼ S s ðp0 ; x0 ; hD Þ S s1 ðp; x; hD Þ where Sðp; x; hD Þ ¼ ln½ Jj expðdj ðpj ; xj ; hD ÞÞ=a. Estimates are based on columns (iv),
(vi) and (viii) of Table 2. Bootstrap standard errors are in parenthesis (based on 100 repetitions).
*
Significant at 5%.
**
Significant at 1%.
per capita in a market of approximately $26,000). With a
welfare gain of $0.005, for instance, a consumer carrying
an average balance experiences an annual benefit of
$8.50. Given these results, one can at least conclude that
the observed changes in bank services, which altered dramatically the corporate identities of the firms in each market as well as the spectrum and quality of services, did not
have any significant adverse effect on the average consumer
that uses banking services.
Among the markets with the largest positive change in
welfare there is Jersey City, NJ and New Haven, CT. The
former experienced a 30% decrease in service fees and the
number of firms in the market went from 9 to 11 throughout the period.41 New Haven, CT is an interesting case,
because there the number of banks decreased from 25 to
18 and fees actually increased slightly. However, consumer
welfare went up due to an improvement in service features,
with a 30% increase in branch density, 35% increase in
branch personnel, a larger proportion of large banks and
a fourfold increase in the geographic diversification of the
average bank. Among the markets with the lowest welfare
change (a loss, in fact), there is Portland, ME and New
York, NY. Service fees increased and deposit rates fell in
Portland, accompanied by a deterioration in service
through a decrease in branch density and branch personnel. The number of firms decreased as well, which was also
the case in New York, where deposit rates and branch density decreased.42
To better understand what is the source of such welfare
effects across all markets, I carry out a simulation where
41
Change is service fees is calculated as a deposit market share weighted
average per market.
42
It is also of interest to point out that two of the biggest mergers in the
period affected two markets that are on opposite sides of the distribution
of welfare change. On the one hand, New York City, which saw mergers
such as that of Chemical-Chase, is among the markets with the lowest
change in welfare. On the other hand, Boston, which was affected by the
Fleet-BankBoston merger, also large and that attracted a lot of attention
in the press, is among the markets that experienced the largest increase in
welfare. We cannot infer anything about the effect of mergers from our
analysis, but it is interesting to see how two large banking markets that
experienced large mergers do have different results based on our
estimation.
bank characteristics are fixed at their 1993 levels, while
prices are allowed to change to their levels in 1999. Thus,
the welfare change calculation provides in this case a measure of how consumers would have been affected if only
prices had changed throughout the period whereas the
expansion in services – through larger branch networks,
for instance – had not taken place.43 Table 6 shows the distribution of consumer welfare change from this exercise. If
banks had not been able to expand the way they did following deregulation, but prices had changed they way they
have, consumers would have then suffered in most markets
(from over 75% to 90% of them depending on the model),
with the average consumer suffering a loss of around $8–10
(based on the median of the distributions for the various
models). This is interesting because it indicates that the
realized welfare effects are not just driven by price effects,
but rather the vertical and horizontal components of banking services play a role, sometimes even offsetting the negative effect of price increases on utility. Thus, this type of
exercise is at least suggestive of the kind of bias that might
arise in welfare analysis based solely on prices and concentration measures.
In terms of whether there is a pattern relating a particular market structure with the realized consumer welfare
change, there does not appear to be one, at least in terms
of market concentration and size. Fig. 1 shows a scatter
plot of market concentration, as measured by the Herfindahl index, and welfare change over the period for each
market. The relationship between the variables appears to
be slightly downward sloping with a low correlation of
0.11 (or lower, depending on the demand model), thereby
requiring huge increases in concentration for any meaningful effect on welfare. Fig. 2 shows a scatter plot of market
size, as measured by population, and the welfare change.
Similarly, the relationship is also negative and small, with
a correlation of 0.13 (or lower, depending on the demand
model). Thus, there is no evidence that concentrated or
43
This is for heuristics purposes only, as this is certainly not derived as an
equilibrium.
Welfare Change 1993-99 (annual $)
A.A. Dick / Journal of Banking & Finance 32 (2008) 1661–1676
1673
60
50
40
30
20
10
0
-10
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
-20
-30
Herfindahl Index
Welfare Change 1993-99 (annual $)
Fig. 1. Market concentration and welfare change.
60
50
40
30
20
10
0
0
10
20
30
40
50
60
70
80
90
100
-10
-20
-30
Population (hundred thousands)
Fig. 2. Market size and welfare change.
large markets were associated with low or negative welfare
change.44
These results suggest that consumers on average gained
from the branching expansion observed after deregulation.
The large number of mergers (an average of 360 per year
throughout the period) tended to reduce the number of
smaller banks in the market, with the size of the average
bank in a market increasing. In some markets, this led to
a reduction in the number of local banks–usually referred
to as community banks. This could raise concern if these
44
The states of Montana and Texas were the only two to opt out of the
federal regulation for nationwide branching, though they both allowed
interstate branching with neighboring states, which, in the early stages of
deregulation, might not have been much of a constraint for the geographic
expansion of local banks. Nevertheless, it is interesting to explore what the
experience for these states has been in terms of consumer welfare. Based
on our results, the markets in both Montana and Texas experienced
similar changes in consumer welfare throughout the period, with an
average of $0.01 or $18 annually per consumer, which is also the average
for the other markets in our sample (based on the logit, for example). This
is not surprising since it is unlikely that many banks found the regulatory
constraint binding in the first few years of deregulation. It would be
interesting, however, to examine this difference for a later period, as banks
have continued to expand throughout the US territory.
local banks were able to provide a service no longer provided by the larger banks. It is possible that consumers
value certain aspects of small community banks not captured in the set of observed characteristics included in the
demand estimation. For example, they might value a local
bank’s specialized knowledge of the local area, its clients
and the specific products that might be tailored to their
needs. In such situation, it would be of particular importance to include bank fixed effects in the regressions, as
was done earlier, to control for these unobserved characteristics. Nevertheless, some consumers could have lost from
not having any longer the personal contact with their bank
that they enjoyed before a large bank purchased it. All we
can say is that, on average, consumers were not worse off.
While markets are assumed here to be geographically
local, the results might still provide insight into the relevant
market definition, which is a central issue to regulatory
agencies in terms of designing regulation, investigating
potential antitrust cases and carrying out merger analysis.
In particular, local market bank variables, such as branch
density, appear to be highly significant for the consumer
decision. Presumably, this is a sign that banking markets
continue to be geographically local, at least for the bulk
of consumers. In fact, the generally good fit of the model,
1674
A.A. Dick / Journal of Banking & Finance 32 (2008) 1661–1676
which is defined at the level of the MSA, might be a fair
indication that the relevant geographic market is demarcated appropriately.
The demand estimates are also useful in analyzing the
supply side, in terms of the implied bank incentives for
horizontal differentiation.45 For example, given consumer
tastes for branches, over time we might expect banks to
continue to increase their branch networks, either
through direct investment or through mergers that would
enhance this attribute. Also, they might want to increase
the staffing of their branches. Furthermore, banks may
strategically decide to merge with neighboring state
banks, as consumers value geographic diversification.
This is an interesting implication, since the antitrust
authorities have tended to be more permissive when it
comes to allowing mergers between banks whose markets
do not overlap. In fact, once geographic restrictions were
lifted in the 1990s, banks appear to have responded to
these incentives. Not only there was a dramatic merger
wave in the industry following deregulation, but also
banks ended with more branches in each local market,
covering more local markets, and with presence in a larger number of states.
analysis of policy. The estimates of consumer preferences
across bank characteristics can be used to analyze the
effects of potential mergers or various other changes in
the environment on consumer welfare. This kind of
counterfactual exercises can be complemented with the
modeling of the supply side. In the banking industry, this
would be particularly interesting in light of the extensive
applied research that has been carried out in banking
cost function estimation. While there is usually very little
prior knowledge on the technical characteristics of the
production side, the banking industry enjoys a wealth
of results and methods from the empirical literature on
cost functions.46
6. Concluding remarks
Appendix
The purpose of this paper has been to estimate the
demand for deposit services in the US commercial banking industry in order to asses the effect on consumers
from the significant changes in banking services throughout the 1990s. The model is able to accommodate the
various changes that have taken place in banking markets, both in terms of service prices and characteristics,
and in particular those that occurred following deregulation, such as the increase in the geographic expanse of a
bank’s service.
The results provide insight on consumer behavior in
choosing a deposit institution, as consumers are found to
respond not only to prices but also to several bank attributes as well. Despite all the changes in the industry and
the fears that some had about the potential harmful effects
on competition, the results suggest that consumers, if anything, benefited from nationwide branching. Clearly, the
paper is not able to establish a definite causal relationship
from deregulation to consumer welfare. While the increase
in the branch network was only possible under deregulation, the actual form this expansion took could have
responded to several other factors changing throughout
the period as well.
Understanding the form of demand and consumer
behavior in banking has several immediate uses. The
use of a structural model of demand, which incorporates
product differentiation, provides a framework for the
Acknowledgements
The author is grateful to Susan Athey, Nancy Rose and
anonymous referees for their insightful comments. She
thanks Peter Davis, Glenn Ellison, Sara Fisher Ellison,
Timothy Hannan, Paul Joskow, Whitney Newey, Scott
Stern, Philip Strahan, Andrew Sweeting, and Jeff Wilder
as well as seminar participants at various universities and
government agencies. Any errors are the author’s.
Summary statistics
Variable
Mean
St.
dev.
Min
Max
Market share
Outside good share
Service fees
Deposit interest rate
Number of employees
per branch
Branch density
Bank age
Number of states
Big
Medium
Mean market wage
(000s)
Housing price index
Expenses of premises
and fixed assets
Funding costs
Banking holding
company indicator
City density (00s per
sq mile)
Total commitments/
total loans
0.0406
0.1600
0.0059
0.0308
23
0.0690
0.1094
0.0038
0.0080
66
0.0000
0.0000
0.0001
0.0050
2
0.8836
0.8635
0.0495
0.0970
4187
0.0040
60
1
0.37
0.23
28.684
0.0121
44
2
0.48
0.42
5.176
0.0000
0
1
0
0
17.044
0.4069
215
16
1
1
60.090
10.641 0.895 7.843 14.436
0.0053 0.0027 0.0000 0.0913
0.3234 0.3188 0.0017 2.9234
0.80
0.40
0
1
6.87
9.93
0.05
0.2844 0.7974 0.0000 120.3061
45
While observed bank attributes are assumed exogenous in the
econometric specification, over time banks are expected to change the
characteristics of their services.
46
118.38
See Berger and Mester (1997) and the references therein.
A.A. Dick / Journal of Banking & Finance 32 (2008) 1661–1676
Appendix (continued)
Appendix (continued)
Variable
Mean
Non-performing loans/
total loans
Equity/assets
Bank operates in at
least one rural area
BLP number of
employees per branch
BLP branch density
BLP bank age
BLP Number of states
BLP big
BLP medium
0.0255 0.0246
0.0000 0.6040
0.0926 0.0363
0.36
0.48
0.0010 0.8306
0
1
1886
3693
17
23,557
0.1183
2729
63
14
15
0.1370
3203
66
14
22
0.0008
95
2
0
0
2.0985
16,303
343
71
124
Observations
44,948
St. dev. Min
Max
Source: Federal Reserve Report on Condition and Income; US census;
Bureau of Economic Analysis.
Description of variables
Variable
Description
Market share
Bank’s market dollar deposits/total
market deposits
Outside good
Credit union + thrifts deposits/total
share
market deposits
Service fees
Service charges on deposit accounts/
deposits
Deposit interest Interest expense on deposits (includes
rate
interest on time, savings and NOW
accounts)/deposits
Employees per Number of bank employees/number of
branch
branches
Branch density Number of branches in local market/
square miles of local market
Bank age
Years since beginning of bank’s
operations
Big (1 = yes)
Bank with assets over US$300M
Medium
Bank with assets of US$100M–300M
(1 = yes)
First-stage results
Variable
Number of
employees per
branch
Branch density
Bank age
Number of states
Big
Medium
Mean market
wage
1675
Service fees
Deposit rate
Coef.
Std.
Err.
Coef.
0.0000
0.0000** 0.0000
0.0000**
0.0149
0.0000
0.0002
0.0005
0.0006
0.0000
0.0014**
0.0000**
0.0000**
0.0000**
0.0000**
0.0000
0.0028**
0.0000**
0.0000**
0.0001*
0.0001**
0.0000**
0.0159
0.0000
0.0002
0.0002
0.0027
0.0001
Std.
Err.
Variable
Housing price
index
Expenses on
premises and
fixed assets
Funding costs
Banking holding
company
indicator
City density
Total
commitments/
total loans
Non-performing
loans/total
loans
Equity/assets
Bank operates in
at least one
rural area
BLP Number of
employees per
branch
BLP branch
density
BLP bank age
BLP number of
states
BLP big
BLP medium
Observations
R-squared
Fixed effects
Service fees
Deposit rate
Coef.
Coef.
Std.
Err.
0.0002 0.0000** 0.0001
0.3918
Std.
Err.
0.0001
0.0060** 0.3804 0.0123**
0.0013 0.0001** 0.0053
0.0001**
**
0.0001 0.0000
0.0004 0.0001**
0.0000
0.0003
0.0000** 0.0000
0.0000
0.0000** 0.0003 0.0000**
0.0052
0.0006** 0.0259
0.0013**
0.0107 0.0004** 0.0234 0.0009**
0.0006
0.0000** 0.0007
0.0001**
0.0000
0.0000** 0.0000
0.0000**
0.0027 0.0006** 0.0005 0.0011
0.0000**
0.0000**
0.0000
0.0000
0.0000
0.0000
0.0000** 0.0000
0.0000
0.0000
44,948
0.34
State
0.0000** 0.0001
0.0000**
0.0000
0.0001 0.0000**
44,948
0.39
State
Note. Year effects included. Estimated standard errors, robust and corrected
for within bank dependence, are in parentheses. BLP refers to the markup
shifters built from an oligopoly model (for a given bank characteristic, the
instrument is the sum of the characteristics of other banks in the market).
*
Significant at 5%.
**
Significant at 1%.
References
Adams, R., Brevoort, K., Kiser, E., 2007. Who competes with whom? The
case of depository institutions. Journal of Industrial Economics 55,
141–167.
Amel, D.F., Hannan, T.H., 1999. Establishing banking market definitions
through estimations of residual deposit supply equations. Journal of
Banking and Finance 23, 1667–1690.
Amel, D.F., Liang, N.J., 1992. The relationship between entry into
banking markets and changes in legal restrictions on entry. Antitrust
Bulletin 37, 631–649.
Amel, D.F., Starr-McCluer, M., 2001. Market definition in banking:
Recent evidence. Federal Reserve Board Finance and Economics
Discussion Series, 2001–2016.
1676
A.A. Dick / Journal of Banking & Finance 32 (2008) 1661–1676
Berger, A.N., Mester, L.J., 1997. Inside the black box: What explains
differences in the efficiencies of financial institutions? Journal of
Banking and Finance 21, 895–947.
Berger, A.N., Kashyap, A.K., Scalise, J.M., 1995. The transformation of
the US banking industry: What a long, strange trip it’s been. Brooking
Papers on Economic Activity 2, 55–218.
Berry, S.T., 1994. Estimating discrete-choice models of product differentiation. RAND Journal of Economics 25, 242–262.
Berry, S., Levinsohn, J., Pakes, A., 1995. Automobile prices in market
equilibrium. Econometrica 63, 841–890.
Calem, P.S., Mester, L.J., 1995. Consumer behavior and the stickiness
of credit-card interest rates. American Economic Review 85, 1327–
1336.
Calomiris, C.W., 2000. US Bank Deregulation in Historical Perspective.
Cambridge University Press, New York.
Dick, A.A., 2002. Demand estimation and consumer welfare in the
banking industry. Federal Reserve Board Finance and Economics
Discussion Series, 2002–2058.
Dick, A.A., 2006. Nationwide branching and its impact on market
structure, quality and bank performance. Journal of Business 79, 567–
592.
Dick, A.A., 2007. Market size, service quality and competition in banking.
Journal of Money, Credit, and Banking 39, 49–81.
Elliehausen, G.E., Wolken, J.D., 1990. Banking markets and the use of
financial services by small and medium-sized businesses. Federal
Reserve Bulletin, 801–817.
Elliehausen, G.E., Wolken, J.D., 1992. Banking markets and the use of
financial services by households. Federal Reserve Bulletin, 169–181.
Fisher, F.M., McGowen, J.J., 1983. On the misuse of accounting rates of
return to infer monopoly profits. American Economic Review 73, 82–
97.
Ishii, J., 2005. Compatibility, competition, and investment in network
industries: ATM networks in the banking industry, Stanford GSB
Working Paper.
Jayaratne, J., Strahan, P.E., 1996. The finance-growth nexus: Evidence
from bank branch deregulation. Quarterly Journal of Economics 111,
639–670.
Jayaratne, J., Strahan, P.E., 1998. Entry restrictions, industry evolution,
and dynamic efficiency: Evidence from commercial banking. The
Journal of Law and Economics 41, 239–273.
Kiser, E., 2002a. Household switching behavior at depository institutions:
Evidence from survey data. The Antitrust Bulletin 47, 619–640.
Kiser, E., 2002b. Predicting household switching behavior and switching
costs at depository institutions. Review of Industrial Organization 20,
349–365.
Knittel, C.R., Stango, V.R., 2004. Incompatibility, product attributes and
consumer welfare: Evidence from ATMs. NBER Working Paper No.
10962, December.
Kroszner, R.S., Strahan, P.E., 1999. What drives deregulation? Economics
and politics of the relaxation of bank branching restrictions. Quarterly
Journal of Economics 114, 1437–1467.
Kwast, M., Starr-McCluer, M., Wolken, J., 1997. Market definition and
the analysis of antitrust in banking. The Antitrust Bulletin 42, 973–
995.
Nevo, A., 2001. Measuring market power in the ready-to-eat cereal
industry. Econometrica 69, 307–342.
Park, K., Pennacchi, G., forthcoming. Harming depositors and helping
borrowers: The disparate impact of bank consolidation. Review of
Financial Studies.
Radecki, L.J., 1998. The expanding geographic reach of retail banking
markets. Economic Policy Review 4, 15–34.
Sharpe, S.A., 1997. The effect of consumer switching costs on prices: A
theory and its application to the bank deposit market. Review of
Industrial Organization 12, 79–94.
Small, K.A., Rosen, H.S., 1981. Applied welfare economics with discrete
choice models. Econometrica 49, 105–130.
Stango, V., 2002. Pricing with consumer switching costs: Evidence from
the credit card market. Journal of Industrial Economics 50, 475–492.
Stiroh, K., Strahan, P.E., 2002. Competitive dynamics of deregulation:
Evidence from U.S. banking. Journal of Money, Credit and Banking
35, 801–827.