*We thank Jeff Campbell, Ed Green, Tim Hannan, Beth Kiser, John Krainer and seminar participants at the 2004 NBER Summer Institute and the Federal Reserve Banks of Minneapolis and Chicago for helpful comments. Two anonymous referees and the Editor provided excellent constructive suggestions. Carrie Jankowski, Kaushik Murali and Anson Soderbery provided excellent research assistance. The authors gratefully acknowledge the financial support of the NET Institute.
If consumers value ‘mix and match’ combinations of network complements, incompatibility between different sellers' components should affect prices. In ATM markets, a 1996 governance change exogenously generated such incompatibility, by allowing banks to impose surcharges when other banks' deposit customers use their ATM's. In our data, incompatibility makes the relationship between deposit account pricing and own ATM's more positive, and makes the relationship between deposit account pricing and competitors ATM's more negative. The level effect on prices is positive. The pattern of results is more pronounced in high population density markets, where customers may care more about ATM's.
In order to perform an automated teller machine (atm) transaction, consumers must use both an ATM card and an ATM. Because banks share access to their ATM's, consumers can use ATM's owned by competing banks as well as ATM's owned by their home bank. In the parlance of the literature on network economics, these are ‘mix and match’ transactions in which consumers assemble a system good from component products (cards and ATM's) sold by different firms.1 Many other products are mix and match: computer hardware and software, operating systems and spreadsheets, and different audio/visual systems are a few examples.
In mix and match markets, compatibility between components sold by different firms is valuable to consumers because it increases the scope for mix and match transactions. In its brief history, the ATM market has exhibited varying degrees of compatibility. Before 1996, the largest shared networks barred banks from imposing surcharges on non-customers using their ATM's, mandating compatibility; the networks rescinded the ban in 1996, after which surcharges became widespread.2 That significantly reduced the compatibility of banks' ATM's with their competitors' cards, leaving the compatibility of own ATM's and cards unchanged. This represents a quasi-natural experiment that motivates our empirical work: a study of how incompatibility affects pricing.
We estimate the effect of incompatibility via a set of regressions linking deposit account (ATM card) pricing to the availability of ATM's. Our data consist of bank/year observations for a panel of banks competing in local markets across the United States from 1994–1999. For each bank, we observe its average deposit account fees, ATM density in its markets, and ATM fees. We can also measure competitors' ATM's available to the same bank's customers. Because consumers pay surcharges only when using a competitor's ATM, incompatibility for a given bank depends on the surcharges imposed by its competitors. It is this latter variable that changes most dramatically after 1996 and provides identification.
We find that incompatibility increases the strength of the link between deposit account prices and own ATM availability, and reduces the strength of the link between deposit prices and other banks' ATM availability. There is also a level effect of incompatibility on prices, which is positive. The net effect is that deposit account prices are positively correlated with incompatibility, and are roughly 20% higher overall after the advent of surcharging. The effects are stronger for large banks (who increase their ATM deployment faster after 1996) and in markets with high population density (where consumers appear to value ATM's more highly after 1996).
The results shed light on a question that previous work has found difficult to address empirically: how does an exogenous change in compatibility affect pricing for system goods? There is a small literature taking incompatibility as given, generally seeking to establish the existence of network effects.3 Existing work on incompatibility is essentially limited to the study of competition between incompatible networks and has employed fairly limited data on incompatibility.4 Our work benefits from the ability to observe within-market changes in incompatibility and a measure of incompatibility that is continuous rather than discrete, although in practice our identification strategy relies on a fairly discrete shift toward incompatibility.
The paper also adds to the growing literature on ATM markets, particularly the papers by Ishii , Knittel and Stango [2006, 2008], Massoud, Saunders and Scholnick  and Hannan . Ishii's work is most closely related, as she estimates the deposit market effects of incompatibility taking it largely as exogenous; her work focuses on deposit account interest rates rather than fees, and more fully explores the equilibrium effects that occur in the deposit account market (endogenizing ATM fleet size, for example). Knittel and Stango  estimate the value that consumers attach to compatibility, and also establish that consumers value both own and competitors' ATM's, but do not directly estimate the equilibrium effects on prices. Knittel and Stango  examine banks' decisions to be incompatible via high surcharges, and estimate whether such decisions are motivated by strategic concerns common in theoretical models of incompatibility.5Massoud, Saunders and Scholnick  and Hannan  both estimate links between surcharging and deposit account market share/structure. Other work on ATM's generally does not focus on ATM fees as incompatibility, although it does in some cases test hypotheses that relate to network effects.6
II(i). Deposit Accounts, ATM Services and Pricing
While banks' strategic behavior is not the focus here, it is worth highlighting the most important features of competition between banks. In addition to deposit account services, banks offer savings account services and a wide variety of other financial services such as loans, brokerage and investment services and insurance. In principle, consumers can (and often do) purchase these services from separate banks. In the United States, approximately 10,000 commercial banks compete for deposit account customers in their local markets.7 Smaller banks often operate only within a small geographic area such as a county, in many cases using a single branch. The largest banks conduct operations in many states or even nationally, and can have thousands of branches and ATM's. Markets are typically assumed to exist at the county level, a convention that we adopt in our analysis in identifying banks' competitors.8 There is considerable heterogeneity in market structure across regions, with rural markets typically being more concentrated than urban markets.
ATM cards are generally sold as part of the service bundle attached to a consumer's deposit account. The standard deposit account agreement also offers customers unlimited access to the bank's own ATM's. Banks locate their ATM's ‘on-premise’ at bank branches, and also ‘off-premise’ at locations such as convenience stores, movie theaters, bars, and other locations where consumers typically need cash. Even within markets, there is considerable variation in banks' ATM strategies—some banks blanket their markets with ATM's, while others deploy them sparingly. The most relevant systematic difference is that large banks deploy ATM's more aggressively than small banks (relative to maintaining branches, for example). Another is that ATM deployment is largely concentrated in areas of high population density.
While at its inception, the market exhibited complete incompatibility because ATM's accepted only ATM cards issued by their home bank, over the 1980's banks formed shared networks that allowed customers to use their cards at other banks' ‘foreign’ ATM's.9 By the early 1990's, all banks essentially subscribed to shared networks, creating almost full compatibility between cards and competitors' ATM's.10 From that time until 1996, the only cost for customers to use a foreign ATM was a foreign fee levied by the customer's home bank. While foreign fees reduce compatibility, they change little over our sample period. The primary source of identification in the data is the post-1996 advent of surcharging. Customers making a foreign transaction pay the surcharge to the bank owning the foreign ATM.
Bank customers therefore purchase from their home bank a bundle of services associated with the deposit account. The bundle includes an ATM card, unlimited access to that bank's ATM's and (possibly) access to other ATM's in the local market. These bundles are differentiated both horizontally and vertically. Horizontal differentiation primarily stems from geography; consumers strongly prefer banks with branches and ATM's that are conveniently located.11 Services other than deposits provided by banks can confer both horizontal and vertical differentiation. These complementary services include savings and money market accounts, loans ranging from mortgages to credit cards, and brokerage services. Large banks are more likely to offer these services, although they become more widely available at banks of any size over our sample period. Vertical differentiation also exists across features of the deposit account; banks vary in quality of customer service, for example. A good deal of vertical differentiation stems from ATM availability; banks often use the size of their ATM fleet as a component of their marketing strategy.
The institutional details of the market motivate our empirical strategy. For any given deposit account bundle, customers will value both the own ATM's associated with their account, and the degree to which they can use other firms' ATM's. This depends on both the availability of those ATM's in the local market, and on the compatibility between cards and other banks' machines. Incompatibility in turn is a function of the fees imposed by other banks for such use. The importance of incompatibility may vary cross-sectionally, based on geography or local market characteristics. We now discuss the ways in which compatibility and prices might be linked.
II(ii). Mix and Match Markets, Incompatibility and Prices
A substantial theoretical literature examines markets where consumers mix and match component products offered by different sellers.12 Considering the institutional detail of the ATM market, the most applicable theories are those in which integrated firms sell both components of the system.13 The compatibility issue then becomes whether Firm A's components will function with Firm B's complementary components, and vice versa. In such settings a general result is that ceteris paribus, incompatibility limits consumer choice and therefore reduces willingness to pay. This does not per se imply lower equilibrium prices, however, because incompatibility can change the locus of competition. With compatibility, competition occurs among components, while incompatibility shifts competition to the system level. Overall, a change in compatibility may therefore either intensify or weaken price competition, which means that equilibrium prices may rise or fall.14
A second implication of the literature is that incompatibility shifts the relative importance of components. In our setting, we can see this intuition by considering an environment with no ATM fees. In that case customers would value the only the total density of ATM's, without regard to their ownership. With incompatibility (high surcharges), own ATM's become relatively more valuable than competitors' ATM's. Thus, we would expect that the advent of surcharging would change the relative importance of ATM's in deposit account pricing.
III. PRICES, INCOMPATIBILITY AND ATM'S
Early research on the relationship between compatibility and prices views compatibility as a product characteristic affecting willingness to pay, and hence prices.15 This is almost certainly a first-order effect, and motivates the use of hedonic regressions as a framework for asking how characteristics are correlated with prices.16 One could take a similar view of our research question, which not only asks how compatibility is linked to prices, but also how other product characteristics—ATM's and competitors' ATM's—are linked to prices, and how these factors interact. However, it is well-accepted that only under very strict assumptions will a simply specified hedonic regression estimate underlying utility parameters.17Pakes  notes that generally we should view a hedonic relationship between prices and product characteristics as a reduced form specification of a richer model in which ‘the hedonic function is the expectation of marginal costs plus that of the markup conditional on ‘own-product’ characteristics.’18 With this in mind, we estimate models of the form:
These specifications model bank i's deposit account price in year t as a function of account characteristics Xit, a set of fixed effects (αi, αt), and a vector Zit containing variables that may be correlated with costs, demand, and markups. Excluding the Zit vector reduces the model to a hedonic regression, while including the Zit allows other covariates to affect prices.
III(i). Prices, ATM's and Other Characteristics
The dependent variable in our empirical model is measured by dividing annual revenue associated with deposit accounts by year-end balances in these accounts:
This measure estimates the annual price per dollar of deposit account balances, across all of a bank's local markets.19 The numerator includes revenue from monthly account fees, fees on bounced checks, per-check transaction charges, extra fees for returned checks, and in rare cases fees for the use of teller services.20 It also includes ‘foreign fee’ revenue; we discuss the implications of this below. It does not include income from surcharges, as surcharge revenue is collected from non-customers and therefore falls into a separate revenue category. While this measure of prices is commonly used in the banking literature, it is imperfect because it may not fully reflect the marginal prices that consumers pay for deposit accounts, particularly when each bank may offer a menu of account options with different implicit and explicit prices for deposit account services. It is also likely measured with error. In our case, this is less of a concern because we use price as the dependent variable; measurement error will therefore only bias our coefficients of interest to the degree that it is correlated with our right-hand side variables. We also include bank fixed effects, meaning that we require only that within-bank variation in this measure is correlated with within-bank variation in prices.21 We have estimated our model using alternative price measures; our results are robust to the use of these alternative measures (see Table V).
Table V. ALTERNATIVE PRICE MEASURE
Notes: Baseline uses specification from column 1, Table II. Model 2 uses level of fee income per transaction account as price.
ln(own ATM density)
ln(own ATM density) × Incompatibility
ln(competitors' ATM density)
ln(competitors' ATM density) × Incompatibility
Bank characteristics include those associated with ATM's and those associated more generally with deposit accounts. Account characteristics include the number of branches per square mile owned by the bank in all of its local markets (counties), the number of employees per branch, and the average salary per employee. The last variable proxies for service quality, although it is also clearly correlated with average costs. To capture the possibility that consumers value obtaining banking services outside their home county, we also include the number of counties in which the bank operates. We also observe a number of bank-level characteristics that do not vary over time (or vary only slightly). These characteristics include whether the bank is a subsidiary of a large bank holding company, whether the bank offers credit card, money market or brokerage accounts, and a dummy variable indicating whether the bank operates in multiple counties. While we can not include these characteristics in our hedonic regression because we also include bank fixed effects, we use a two-stage procedure to estimate the relationship between these characteristics and prices.22Appendix B outlines this procedure and presents results, which we summarize below.
ATM characteristics include both the ATM's owned by bank i, and the ATM's owned by its competitors in the local markets in which it competes. In the absence of incompatibility, we specify the relationship between prices and own/competitors' ATM density as:
We measure density as ATM's per square mile over all counties in which the bank operates. If the bank is part of a holding company and we only observe ATM deployment for the holding company, we disaggregate the data by allocating ATM's to each bank in the holding company proportionally to the number of branches in each bank. We use logs to reflect the fact that each additional machine reduces the expected travel distance to use an ATM by a successively smaller amount. One issue with this specification is that we have only partial data on competitors' ATM's. Our data source provides information regarding the ATM deployment of the largest 300 issuers in the United States; these issuers collectively hold roughly sixty percent of all ATM's during our sample period, but not all. In order to deal with this, we estimate the number of ATM's deployed by each bank's competitors for which we do not observe actual deployment.23 We have used a variety of techniques for this estimation, an issue discussed in Knittel and Stango .24 We outline the estimation procedure in Appendix A, and discuss its implications below.
III(ii). Specifying Incompatibility
Incompatibility should change the relative importance of own and competitors' ATM's. It makes competitors' ATM's relatively less attractive by increasing the explicit costs associated with using a competitors' ATM. It also makes own ATM's more relevant because on the margin, consumers will likely make more transactions at their bank's ATM's. We model these effects by interacting each ATM term with incompatibility:
The model also includes the level of incompatibility. This may capture changes in competition correlated with the shift toward surcharging. It may also capture other changes in either consumer behavior or bank characteristics correlated with changes in compatibility.
Our primary measure of incompatibility estimates the average surcharge a consumer would pay for using another bank's ATM. This measure is:
The average surcharge is weighted, where the weights are the shares of total ATM's held by other banks in bank i's local market(s). The motivation for this specification is an assumption that consumers know something about the distribution of ATM's and ATM fees in their local market, but do not have perfect knowledge regarding either specific fees at each ATM or the locations in which they will experience an unanticipated need for cash.25 Because we possess surcharge data for only the largest ATM issuers in each market, constructing this average requires making an assumption about the surcharging behavior of smaller issuers. We outline these assumptions and discuss robustness in Appendix A; our results are quite robust to different assumptions about the behavior of smaller issuers.26
We also measure incompatibility using a dummy variable equal to one after the lifting of the surcharge ban. This variable is equal to one for all observations after 1996, and zero prior to 1997. While this measure loses information on cross-sectional variation in the level of surcharges, it is less likely to be endogenous. We have also use an incompatibility measure that includes the bank's own foreign fees; while we do not present those results here, they are in an earlier version of the paper (Knittel and Stango ) and are not different from those we present here.
Finally, while the largest networks barred surcharging prior to 1997, a number of states overrode the bans prior to 1997. While we do not observe the extent of surcharging in these states, we can identify their effects through a dummy variable AllowSurchit. In some specifications below we also include this other measure. This is particularly useful given that most variation in incompatibility occurs discretely in 1997; the AllowSurchit variable provides a test against the hypothesis that the value of ATM's changed after 1996 due to some unobserved factor.
III(iii). Other Covariates
The preceding discussion defines those variables that enter our baseline (hedonic) regressions. In our full model, we also include other covariates that may affect markups via their relationship to costs and/or demand.27 We include three such variables: the ratio of noninterest expenses to assets for the bank, its net interest margin on all of its loans, and the average savings rate on its deposits.28 Each of these variables is measured in percentage points.29 While the noninterest expense ratio may include both fixed and variable costs, it may be correlated with marginal cost. The interpretations of the net interest margin and savings rate are less clear, as they likely measure components of both costs and willingness to pay. The savings rate, for example, represents both an opportunity cost of funds for the bank (affecting its costs) and for consumers (affecting their substitution between checking and savings accounts). The net interest margin operates similarly. Thus, we remain agnostic about the expected signs of these coefficients. To explore the possibility that incompatibility affected the influence of these variables on prices, we also estimate a specification that interacts them with a dummy variable equal to one after 1996:
III(iv). Econometric Issues
One econometric concern is that while these data are in principle very rich, we face measurement issues in both our competitors' fees and ATM density variables. The presence of measurement error generally will bias the estimated coefficients toward zero.30 However, our empirical focus is more on testing whether there was a shift in the relationship between prices and ATM density following surcharging than on obtaining accurate point estimates of our coefficients. In other work where we focus more on the latter concern, we implement a statistical correction for measurement error and undertake a variety of robustness checks of the technique.31
A second econometric concern is endogeneity, particularly of competitors' surcharges. As we noted above, the advent of surcharging is plausibly exogenous; we therefore estimate the model defining incompatibility using a post-1996 dummy variable. That model using is identified solely from the exogenous removal of the surcharge ban. We also estimate the model using only post-1996 data; that model is identified solely from (likely endogenous) cross-sectional variation in firms' equilibrium surcharges. Comparing the models is instructive regarding how much of the identification in our full model is coming from the exogenous pre and post-1996 transition, and how much is coming from endogenous post-1996 variation in the cross section. This approach is only informative regarding endogenous surcharges, of course; ATM density, branch density and our other bank-level characteristics might also be endogenous; the work by Ishii  treats ATM density as endogenous, for example. In our case, it is difficult to think of an appropriate (and large enough) set of instruments, although in other work we use higher moments of the observable variables as instruments with some success.32 We therefore acknowledge that endogeneity is a potential concern for these other variables, and interpret the coefficients that may reflect endogeneity bias cautiously.
Omitted variable bias may also be a concern; some factor other than ATM's or incompatibility may be driving price changes in our sample. While this certainly might be true, any common factor across banks leading to changes in prices would be controlled for by our fixed year effects.33 It would therefore have to be the case that the factor would be correlated both with within-bank changes in ATM density and within-bank price changes. Considering this, the most likely source of omitted variable bias is that a bank may simultaneously expand its ATM fleets and broaden its service offerings. If service offerings are an omitted variable, this would lead to spurious correlation between own ATM density and deposit account prices. Given our data, we can not rule this explanation out. We note, however, that such a correlation would not explain a positive relationship between competitors' ATM density and deposit prices, or a change in the strength of the own ATM/deposit price relationship after the advent of surcharges. Nor would it predict a difference across areas of high/low population density (something we find and document below).
Another omitted variable concern is that during our sample most ATM cards serve as debit cards; the value to consumers of an ATM card therefore depends on the availability of point-of-sale terminals as well as ATM's. The availability of point-of-sale terminals is generally viewed as having accelerated after 1996 as consumers increasingly used debit to avoid surcharges. In the specification we use here, changes in deposit account prices stemming from debit/POS markets may affect our coefficients—both those on ATM's (negatively), and on the level of incompatibility (positively). We expect similar effects from the fact that we do not measure ISO ATM deployment, which rose sharply after 1996.
III(v). Data and Descriptive Statistics
Table I presents descriptive statistics for our sample. Appendix A outlines definition and measurement issues for these variables. We take our data from a variety of sources. The ATM-related characteristics come from the Card Industry Directory, an annual trade publication listing data on ATM fleets and fees for the largest three hundred ATM issuers. Many of those issuers are holding companies consisting of multiple banks; this gives us data for roughly 3,700 bank/years over the sample period 1994–1999.
Table I. DESCRIPTIVE STATISTICS
Sources: Federal Reserve Reports of Condition and Income (Call Reports), various years; FDIC Summary of Deposits, various years; Card Industry Directory, various years. Values are medians for ATM- and branch-related variables, means for all others.
Competitors' ATMs/square mile
Deposit fees ($ per $1000 of deposits)
Foreign fee ($)
Expected competitors' surcharge ($)
Salary per employee ($1000)
Employees per branch
Number of counties
In most cases we report median values for our data, because the data are highly skewed. One source of skewness is bank size; for example, while the median bank size (in deposits) is $326 million, the mean is $2.3 billion. The tenth and ninetieth percentiles are $58 million and $5.8 billion. Another source of skewness is geographic diversity, realized largely through differences in branches and ATM's per square mile. The only variables for which we report means are those that are not skewed: deposit fees, ATM fees and our analogous measures for competitors, salary per employee and employees per branch.
The top rows show data by year regarding branches, ATM's, fees and the other variables included in the hedonic specification. Branches remain roughly constant, but ATM's (and competitors' ATM's) grow significantly over the sample period. The average level of deposit fees remains constant, although this is a bit misleading; we show below that among banks whose ATM fleets grew rapidly, prices rose as well. Foreign fees remain roughly constant, while surcharges become quite prevalent between 1997 and 1999, which nearly doubles a customer's expected costs for using a foreign ATM. Salary per employee and employees per branch remain essentially constant. The number of counties the typical bank operates in grows over time, reflecting the cross-market consolidation that occurred in banking markets after their deregulation in the 1980's and 1990's.
Table II presents the results of our regressions examining the relationship between incompatibility, bank/ATM characteristics and pricing. We show results from six specifications. The first is our baseline model including ATM's, incompatibility and other deposit account characteristics. The second also includes the AllowSurch variable equal to one in states permitting surcharges before 1997. The third model includes only data from after 1996, while the fourth measures incompatibility using a dummy variable equal to one after 1996. The final two columns include our Z vector of other covariates, and interactions of those with a post-1996 dummy variable.
Table II. PRICE AND INCOMPATIBILITY REGRESSIONS
Notes: Models 1–3, 5 and 6 use competitors' surcharges to measure incompatibility. Model 4 uses a post-1996 dummy variable. Model 3 uses only post-1996 observations.
The results suggest that prior to 1996, the link between deposit account prices and ATM's was fairly weak; the coefficients on the levels of own ATM's are positive, but significant in only two of the five relevant specifications (the third does not use pre-1996 data and therefore does not estimate the relationship). Competitors' ATM density is not significant in any of the specifications that use pre-1996 data. After 1996, those links become stronger. In the specifications that do not include the Z vector, the results suggest that incompatibility makes the correlation between own ATM's and deposit account prices more positive, and makes the relationship between competitors' ATM's and account prices more negative. The results are also robust to different price measures (See Table V). The own ATM density results are not significant in the last two columns, which include the Z vector; while this suggests that the results are fairly weak, we shed more light on this issue in the next table, which splits the sample by high/low population density. The pre-1996 incompatibility coefficients have similar signs, and although they are not individually significant, they are jointly statistically significant at the 0.10 level.
While we do not discuss them in detail, the other coefficients show an intuitive relationship between bank characteristics and prices.34 Salary per employee and employees per branch are positively related to prices, although the coefficients are not significant in every specification and small in economic terms. Branch density is not significant in any specification. Nor does there appear to be any systematic relationship between geographic breadth (as measured by number of counties) and prices. Referring to the results in Appendix B, we also find a positive relationship between price and whether the bank is a member of a bank holding company, operates in multiple counties, and offers brokerage services.
IV(i). Subsamples by Population Density
The specifications above pool the data, effectively assuming that the relationships in the data are identical across the entire cross-sectional range of markets and firms. However, most models of ATM usage view travel costs as an important determinant of the relationship between incompatibility (fees) and deposit account pricing.35 While we can not measure travel costs directly, anecdotally such costs are higher in areas with high population density. In order to explore this possibility, we stratify our sample by the population density of banks' local markets and explore the differences between high and low density markets.36 We categorize as ‘high density’ any bank operating in areas with an average population density above the sample median, and the remainder as operating in ‘low density’ areas.37
Table III presents results from our regression models using the density subsamples. The difference is striking. In the high-density subsamples, the relationship between ATM's and deposit prices is extremely strong, while in the low-density subsamples the relationships are essentially nonexistent. This is consistent with a view that travel costs increase the importance of ATM access, and also increase the importance of compatibility between cards and own ATM's. The coefficient on the level of incompatibility is also much larger in high-density markets, and close to zero in low density markets.
Table III. REGRESSIONS FOR BANKS IN HIGH AND LOW POPULATION DENSITY MARKETS
Notes: Dependent variable is ln(Deposit Fees). All specifications include fixed year and bank effects. Model uses competitors' surcharges to measure incompatibility. Models (1) and (3) use observations from high-density markets. Models (2) and (4) use observations from low-density markets.
ln(own ATM density)
ln(own ATM density) × Incompatibility
ln(competitors' ATM density)
ln(competitors' ATM density) × Incompatibility
ln(own branch density)
ln(salary per employee)
ln(employees per branch)
ln(number of counties)
Non-interest expense ratio
Net interest margin
Non-interest expenses × post-1996
Net interest margin × post-1996
Savings rate × post-1996
Number of Observations
While it seems sensible that the results should be stronger in high-density markets, it does seem a bit surprising that the results are so weak in low-density markets. While one possibility is simply that travel costs are low enough to render ATM's (and by extension incompatibility) irrelevant, another possibility is simply that our model is well-specified for high-density markets and poorly specified for low-density markets. Evidence in favor of this comes from the other coefficients. Branch density, for example, is positively and significantly related to prices in high-density markets but not in low-density markets; our priors tell us that branches would be relatively more important in markets where consumers do not value ATM's (though this is only a conjecture).38 Similarly, the number of counties is positively related to prices in high-density markets—a result we find intuitive—but negatively related to prices in low-density markets. This pattern seems to suggest that specification error may be a problem in low-density markets. Given this inconclusive evidence, we interpret our results as finding strong relationships between incompatibility and pricing in high-density markets, while in low-density markets we are unsure whether our (non-)results reflect lower travel costs or specification error for these markets.
While do not report the results here, we have also estimated models splitting the sample by bank size. Generally, doing so yields results similar to those in Table III because large banks tend to be located in high-density areas. However, further splits by size and density suggest that the results are strongest for large banks in high-density areas. The coefficients on ATM's and incompatibility are not significant either for large banks in low-density markets, or for small banks in high-density markets (though the sizes of these subsamples are small, reducing the power of the test).
IV(ii). Evidence on Magnitudes
Given the results above, we now present some supplementary information on the magnitudes of the results. Table IV shows summary data for our sample stratified in two ways. First, we separate banks by population density. We also separate large and small banks, treating as ‘large’ any bank in the subsample with a share of the local ATM market larger than the median (for that subsample). These data show a clear pattern that is not only informative regarding the travel cost story but also sheds light on variation in the data that identifies our earlier results. The greatest changes following the advent of surcharging are by large banks in dense areas. The most dramatic changes are in ATM density, which doubles for large high-density banks but is unchanged for smaller high-density banks. This is associated with equally dramatic changes in prices. Large banks charge higher deposit fees. More importantly, this deposit fee gap grows significantly after the advent of surcharging, from $0.60 in 1995 to $1.66 in 1999. There is little evidence of such change in low density areas. While there are differences between large and small banks, they are not nearly so dramatic. Nor do they change very much after the advent of surcharging.
Table IV. SUMMARY STATISTICS BY ATM SHARE AND POPULATION DENSITY
Notes: Incompatibility is expected competitors' surcharge. High and low density are above and below sample median. Large/small banks are those above/below median deposit market share for density subsample.
large bank, high density
small bank, high density
large bank, low density
small bank, low density
large bank, high density
small bank, high density
large bank, low density
small bank, low density
large bank, high density
small bank, high density
large bank, low density
small bank, low density
To learn more about the equilibrium effects of incompatibility on prices, we calculate changes in fitted prices (in natural log) implied by the coefficients in Tables II and III, and the changes in both ATM density and surcharges between 1996 and 1999. These fitted values are calculated using variants of the following equation:
Figure 1 shows two kernel densities of these equilibrium changes, using the coefficients from Model 1 of Table II.39 The left-most (dotted line) distribution computes the change in ln p̂ considering only the change in surcharges (keeping ATM densities at their 1996 level), while the right-most (solid) distribution incorporates the changes in ATM density over time. The shift in the distribution therefore reflects the fact that ATM density increases in a way that predicts higher prices. The mean of the right-most distribution suggests that, on average, prices increase 14 per cent due to changes in incompatibility and ATM density; much of this net effect comes from the large level shift implied by the incompatibility variable. There is significant variation across banks; the 25th percentile is 9 per cent, while the 75th percentile is 19 per cent. Figure 2 presents the same data as in the right-most density of Figure 1, but separates small and large banks; on average, the predicted price increases are much more positive for larger banks. Because both densities use the same coefficients, the difference stems from differences in the change in incompatibility (greater for large banks) and ATM density (also greater for large banks). Finally, Figure 3 compares predicted price changes for banks in markets with high population density using both the full sample coefficients (Column 1 of Table II) and coefficients from the subsample of banks from high density markets (Column 1 of Table III). The difference in the sets of coefficients suggests economically significant differences in the estimated price changes.
V. DISCUSSION AND CONCLUSION
Overall, our results suggest that the shift to incompatibility in ATM markets—as measured by the prevalence of surcharges on foreign transactions—had an economically significant effect on deposit account prices. These changes include both a level effect, a change in the relative importance of own and competitors' ATM's, and changes in the levels of ATM's available to deposit account customers. They are economically large, and more so in areas with high population density. These results generally square with other work examining ATM fees and links to deposit account pricing. Knittel and Stango , for example, find that one motive for high ATM fees is to intentionally create incompatibility and raise deposit account prices—but that this motive is only operative for large banks.
Overall, the results suggest that taking a simple view—that compatibility affects prices only through its static effect on consumer choice sets—is probably incorrect. Our results are more consistent with a model in which willingness to pay, the intensity of competition, and both own and competitors' product characteristics are affected by changes in compatibility. Drawing firm conclusions about the specific policy question in ATM markets—whether the advent of fees increased or reduced consumer welfare—still remains difficult, but one thing that clearly emerges from this and other studies is that considering the deposit account effects of ATM fees is necessary before reaching any conclusions about welfare.
APPENDIX SOURCES AND VARIABLE CONSTRUCTION
A(a) Primary Data Sources
We take our data from four principal sources. The first is the Card Industry Directory, an annual trade publication listing detailed data on ATM and debit card issuers. The Card Industry Directory contains data for the largest 300 ATM card issuers, who collectively owned roughly 80 per cent of the nation's ATM fleet during our sample period. These issuers are most often commercial banks, although some are bank holding companies, credit unions or thrifts. The sample period covered in our data set runs from 1994 to 2002. Data are measured on January 1 of each year.
We also take data from the FDIC Reports of Condition and Income, or Call Reports. The Call Report data are collected quarterly by the FDIC for every commercial bank in the country. The Call Reports contain detailed balance sheet and income data for each bank. They also indicate which bank holding company owns the bank. Thus, if the Card Industry Directory contains a listing regarding ATM issuance for a bank holding company, we can match that data with the corresponding data for each bank owned by the holding company. The Call Reports do not contain data for credit unions or thrifts; we drop them from the sample.
We supplement the above with data from the FDIC Summary of Deposits Database (SOD). The SOD lists the location of branches for every bank and thrift in the country. It also lists the deposits held at each branch.
A(b) An Observation in Our Data
By cross-referencing the data sets above, we obtain observations at the issuer level describing each issuer's balance sheet activity and ATM activity. We also use the geographic data from the SOD to derive information about the market(s) in which the issuer competes. Because the data are measured at different times, we must establish a concordance between the dates in the different data sets. We establish the concordance based on the fact that our analysis includes deposit prices as LHS variables, and ATM-related variables as RHS variables. While these may be jointly determined, to mitigate the endogeneity problem, we match ATM-related data for each January with six-month ahead data from the other data sets. Thus an observation from 1994 contains ATM-related data from January, 1994, while all other data are from June, 1994. We describe these data below.
A(c) Pricing for Accounts and ATM's
For each issuer, we observe its income associated with deposit accounts over the year preceding the observation date. The primary component of such income is income from monthly service charges on deposit accounts. It also includes foreign fee income paid by its customers stemming from the use of other issuers' ATM's. It also includes a variety of other fees such as NSF fees for bounced checks and other penalty fees on deposit accounts. If the issuer is a bank holding company, we sum its deposit fee income for all banks in the holding company.
To develop our measure of prices, we divide income on deposit accounts by the end-of-year dollar value of deposits (in thousands). This price measure therefore represents the average fees paid per dollar of deposits. This measure omits the additional opportunity cost of holding deposits in checking, which is the forgone savings interest income. However, it is likely that the measurement error associated with omitting this component of ‘prices’ is similar across banks, and within banks over time.40
Another issue associated with using this price measure is that banks typically offer consumers account options with lower explicit fees in exchange for maintaining higher minimum balances. If banks differ systematically in the composition of their customer bases, we will understate fees at banks with high deposits per customer (assuming those customers sort into accounts designed for them).
A practical difficulty with using this measure of fees is that the numerator is a flow measure over the previous year, while the denominator is a stock measure at end-of-year. This creates measurement error for banks with large deposit acquisitions or divestitures during the year. Indeed, there are a significant number of observations with implausibly small or large fee measures. To check that these were outliers stemming from measurement error, we measured the year-to-year percentage change in deposits for observations with exceedingly small or high fee measures; we found that in most cases such observations were for banks that experienced extremely large changes in deposits (more than fifty per cent in absolute value).41 We drop these observations.
For each issuer in the Card Industry Directory, we also observe its foreign ATM fee and surcharge at the beginning of the year for the observation. In some cases, the bank lists a range for these fees. In that case, we use the highest fee reported. In the empirical work, this tends to understate the true relationship between fees and our other variables of interest.
Because our measure of prices is somewhat aggregate and rough, we explore the robustness of our results to alternative measures of price. Table 5 shows results using one alternative, measuring prices as the level of fee income per transaction. Knittel & Stango  shows results using other alternatives.
A(d) Competitors' ATM's and Surcharges
For each issuer, we observe its total deposits, ATM's and branches. We also observe the distribution of its deposits and branches across individual counties. Obtaining data on county size in square miles allows us to calculate the density of branches/ATM's per square mile within each county. For banks operating in multiple counties, we calculate the average number of branches and ATM's across all counties in which the bank operates. For issuers who own multiple banks, we allocate ATM's to each bank in the holding company based on branches.
In order to construct competitors' ATM's, we estimate the total number of competitors' ATM's in each county. We do this by estimating a within-sample regression of ATM's on branches, year dummies and year/branch interaction terms. To control for the fact that larger FIs have a greater ratio of ATM's to branches, we also interact the branch variables with the log of issuer size (in deposits). We then construct fitted values of ATM's for each FI for which we do not have ATM data. In order to check the sensibility of this procedure, we compared the fitted total number of ATM's from this procedure to aggregate data on ATM deployment. The figures match fairly closely. We also conduct in Knittel and Stango  a wide variety of robustness checks, involving different methods of estimating competitors' ATM's.
As we discuss in the text, our measure of competitors' ATM's omits ATM's deployed by Independent Service Operators (ISO's). This introduces measurement error, and may bias our measures of competitors' ATM density. This would be particularly important in urban markets where ISO ATM deployment was quite rapid after 1996.
For competitors' surcharges we undertake a similar procedure.
A number of bank-level characteristics are fixed at the bank level over time. This precludes their inclusion in the hedonic regressions, which also include bank fixed effects. However, we can learn something about the value of these other characteristics by examining their relationship to the fixed effects.
Starting with our estimates of the bank fixed effects, we construct the vector Πi of time-invariant bank characteristics. The first set of such characteristics describes the product offerings of each bank; there are dummies equal to one if the bank offers a credit card, money market accounts or brokerage services. We also include a dummy if the bank has branches in multiple counties, and a dummy equal to one if the bank is part of a larger bank holding company. Some of these variables (particularly the product offering variables) vary over time for a small subset of banks. For these banks we use the average value of the dummy variable over the sample period as the independent variable. Table VI shows results of these models.
Table VI. SECOND STAGE HEDONICS WITH TIME-INVARIANT CHARACTERISTICS
Notes: Specifications use first-stage results from model (3) of Table II. Model 1 uses entire sample. Model 2 uses observations from high-density markets. Model 3 uses observations from low-density markets.
2 Nine states overrode the ban prior to 1996; we account for this in the empirical work below. See Prager  for an examination of this episode. One state (Iowa) maintained its ban after 1996, but our sample contains no data from banks in Iowa.
3 Examples of work taking incompatibility as given include Gandal, Greenstein and Salant , who study the link between operating system values and software availability in the early days of the microcomputer market. They find evidence supporting the existence of indirect network effects. More recent work by Gandal, Kende and Rob  tests for indirect network effects in the adoption of Compact Disks (CD's) and CD players. Rysman  provides evidence supporting the existence of complementary demand relationships in a two-sided platform market (Yellow pages). More recent work by Shankar and Bayus , Nair, Chintagunta and Dube  and Karaca-Mandic  applies structural econometric techniques to test for the existence of network effects in markets where compatibility is fixed.
4Gandal [1994, 1995] and Brynjolfsson and Kemerer , find that computer spreadsheets compatible with the Lotus system commanded higher prices during the early 1990's. Our work differs from this early work, in that it estimates the effects of compatibility across different components of the network. It also differs in that it primarily relies on within-firm and within-market rather than cross-sectional variation in compatibility for identification. More precisely, the analyses in Gandal [1994, 1995] and Brynjolfsson and Kemerer  do not separate within-firm from cross-sectional effects of compatibility. The datasets are panels, but too small to allow the examination of within-firm variation. In one other piece of work examining a different market, Greenstein  finds that mainframe buyers prefer to upgrade to compatible systems, a result suggesting that compatibility between past and future hardware is important.
6Hannan  uses the fact that Iowa maintained its ban after 1996 to test whether equilibrium market structure is related to surcharging. Hannan, et al.  examine banks' propensity to impose surcharges as a function of a variety of characteristics, although they do not explicitly link their analysis to deposit account pricing. Prager  tests whether small banks lost market share in states that allowed surcharges prior to 1996; this is implicitly a test of whether incompatibility favored banks with high-quality ATM fleets, although she does not pose the question in those terms. Hannan and McDowell [1984a, 1984b, 1990] explore the relationship between market concentration and ATM adoption. They find that markets in which large banks adopted ATM's became more concentrated during the 1980's, although they do not discuss their finding in terms of network economics. Saloner and Shepard  examine the diffusion of ATM's from 1972–1979 and find that adoption occurred earliest for firms with many branches and deposits, a result they interpret as consistent with the existence of indirect network effects in demand. Gowrisankaran and Krainer  estimate the welfare effects of the increase in ATM deployment stemming from the surcharge ban, although their model does not incorporate network effects.
7 Our data omit observations for credit unions and thrifts. However, these institutions collectively hold only a small share of the deposit market.
8 Some work treats multi-county MSA's rather than individual counties as markets in urban areas—in our case, doing so makes no difference empirically. This is due to our focus on within-market changes rather than cross-sectional differences. Any differences between the levels of variables measured at the county rather than MSA are likely to be constant for a given bank over time, and are swept out by bank fixed effects. Recently, the question of whether banking markets have become less local has come to light (see Radecki , Biehl  and Heitfeld and Prager  for discussions). While this may be true for products such as mortgages, it is unlikely to be true for consumers' ATM usage, which is necessarily local.
9 The networks themselves are typically joint ventures formed by banks in order to share the fixed costs of interconnection infrastructure. Banks usually pay a fixed monthly or annual membership fee to the network. They also pay a ‘switch fee’ for each transaction made by one of their customers on another bank's ATM's; the switch fee is roughly $0.40 on average during our sample, and does not vary significantly across networks or regions. Part (on the order of $0.10) of the switch fee is paid to the network, and the remainder flows to the ATM's owner in order to compensate it for providing services to a non-customer.
10 In most cases access to these ‘foreign’ ATM's is incomplete because it only allows consumers to withdraw cash; more complex transactions such as making deposits are not permitted through the shared network.
11 See Stavins  for a discussion of the characteristics that consumers favor when making their deposit account choices.
18Feenstra  goes beyond this general point to explicitly model the exact relationship between prices, costs, markups and characteristics for particular functional forms of costs and utility.
19 Here we assume that banks operating in multiple markets charge the same prices (or offer the same menu of account options) across all of their markets. See Radecki  and Heitfeld  for evidence on this point as it relates to interest rates. In some cases an ‘issuer’ in our data is a holding company owning multiple banks. In those instances we assign the issuer the average price (weighted by deposits) across the banks in the holding company.
20 It does not include interest paid because most checking accounts do not pay interest. However, in the empirical work we control for savings rates and the bank's net interest margin.
21 Because we measure prices at the bank level, this variation exists only over time, not across markets.
22 As suggested by Chamberlain , we regress the estimated fixed effects from the first stage on the fixed bank characteristics.
23 We do not have data on ATM deployment by Independent Service Operators (ISO's), who began deploying machines after the advent of surcharging. Aggregate data indicate that by 1999, ISO-deployed ATM's comprised ten percent of ATM's nationwide. The effect of these ATM's on prices is an omitted variable in our specifications.
24 Almost all smaller issuers deploy roughly one ATM per branch, with deployment growing slightly over time. Aggregate data from the Card Industry Directory confirm this; in every year between 1994 and 1999, the roughly 10,000 issuers outside the top 300 deploy a total of 10,000 ATM's.
25 This is analogous to the modeling assumptions in Massoud and Bernhardt .
26 The behavior of the issuers we need to forecast (smaller banks) is fairly easy to explain. We explain 60 per cent of the variation in surcharges for the sample and 40 per cent for smaller deployers. Furthermore, because the measure is weighted by the issuers ATM share, mismeasurement of surcharge behavior for small banks will have a small effect on the weighted average.
27Pakes  also notes that markups may depend on competitors' product characteristics. In unreported specifications, we include a variety of such characteristics (such as the fraction of competitors offering credit cards, money market funds and brokerage services). None are statistically significant.
28 We obtain these variables from the Call Reports. All variables are annualized. Noninterest expenses are yearly expenses divided by total assets. The net interest margin is aggregate loan income minus aggregate loan losses divided by total loan balances, all measured in dollars. The savings rate is total interest expense on savings accounts divided by total savings balances, all measured in dollars.
29 The sample means for the noninterest expense ratio, the net interest margin and the savings rate are 1.63%, 1.21% and 2.58% respectively.
30 The direction of bias may be away from zero if measurement error is correlated with other right-hand side variables. See Fuller  for a discussion of the problem and some solutions.
33 The year effects would also control for the increasing popularity of credit unions during the sample period, and for any general changes in markups due to banking mergers (which were significant across markets but did not substantively change local market structure).
34 Note that the bank-level fixed effects capture bank characteristics that do not vary over time.
36 This sample split will also measure any systematic difference in the value of using ATM's or consumers' use of cash across the two types of area.
37 The sample median population density is 201 per square mile (measured at the county level). This is a density typical of a small urban area on the East Coast such as Merrimack County, New Hampshire (which contains Manchester and Nashua). Because our data cover only the largest three hundred ATM issuers, and these issuers operate primarily in metropolitan areas, the sample of markets is disproportionately high-density; fewer than 500 of the more than 3,000 counties nationwide have a density above 200.
38 We have also estimated specifications that interact own branches with incompatibility. The results suggest that own branches became relatively more important after 1996. While there is little reason to believe that branches per se changed, branches are the locations of ‘on-premise’ ATM's. If these ATM's are more valuable to consumers, the increased value of ATM's may reflect the relative increase in importance attached to them.
39 The densities are not weighted by size. Doing so would shift the distribution to the right—see the small/large split in Figure 2.
40 In Knittel and Stango , we include a measure of the opportunity cost of funds in our price measure. Using the broader measure has little effect on the results in that paper.
41 We conjecture that many of these reflect merger-related changes, but can not identify mergers in the data.