Commercial Lending and Distance: Evidence from Community Reinvestment Act Data

Commercial Lending and Distance: Evidence from Community Reinvestment Act Data Kenneth P. Brevoort(cid:3) Timothy H. Hannan Federal Reserve Board Mail Stop 149 Washington, DC 20551 Kenneth.P.Brevoort@frb.gov Timothy.H.Hannan@frb.gov February, 2004 Abstract Innovationssuchascreditscoringhaveincreasedthe abilityofbankstolendtodistantbusiness borrowers, which could expand the geographic market for small business loans. However, if this e(cid:11)ect is limited to a few large banks, the market may become segmented and lending distance at local banks actually decrease. This paper, using a new data source and a spatial econometric model, empirically estimates the relationship between distance and commercial lending and how this relationship is evolving over time. We (cid:12)nd distance is negatively associated with the likelihood of a local commercial loan being made and that the deterrent e(cid:11)ect of distance is consistently more important, the smaller the size of the bank. We (cid:12)nd no evidence that distance is becoming less important in the United States in recent years. In fact, the bulk of the evidence suggeststhat distance may be ofincreasing importance in localmarketlending. (cid:3)The opinions expressed in this paper are those of the authors and do not necessarily reflect the views of the Board of Governors of the Federal Reserve System or its sta(cid:11). The authors would like to thank Bob Adams, Ron Borzekowski,HendrikHakenes,BethKiser,RobertMarquez,MackOtt,RobinPrager,JohnWolken,andparticipants ofthe2003CompetitioninBankingConferenceattheKatholiekeUniversiteitLeuveninBelgiumandthe2003ASSA meetings in Washington, DC. Onka Tenkean provided outstanding research assistance and James Tedrick provided programming support. Anyerrors are theresponsibility of the authors. 1

1 Introduction The geographic area over which banks are willing to extend credit has important implications for competition in bank lending and the application of antitrust policy. There is some evidence that improvements in information technology, most notably credit scoring, may have increased the ability of banks to lend to distant business borrowers. Such a shift could expand the size of a geographic market, but if this e(cid:11)ect is limited to larger lenders, a possible consequence is that the market becomes segmented and that lending distance at more local banks actually decreases. This paper, using a new data source and a spatial econometric model, empirically estimates this relationship between distance and commercial lending and how this relationship is evolving over time. A recentstudy by Petersen& Rajan(2000)uses the 1993Survey ofSmall Business Finance (SSBF) to examine how the distance between borrowers and bank lenders have been changing overtimeandreportsthataveragedistancesareincreasing. Thereisreasontobelieve,however, that the phenomenon of increasing distance has not been uniform. In comparing data from the 1993and1998SSBFs,Wolken&Rohde(2002)(cid:12)ndthat,whileaveragedistancebetweenlenders and small business borrowers increased, median distances changed only minimally, suggesting thatthephenomenonofincreasingdistancebetweenlenderandborrowermaybeoccurringonly at the extreme tail of the distribution. Neither of these studies addresses the possibility that the technological advancements that led to these changes, most notably credit scoring, may have been adopted more fully by larger banks. Thustheincreaseintheaveragedistancebetweenborrowerandlendermaybetheresult ofasmallnumberofbankswho,becauseofthenewtechnology,havee(cid:11)ectivelybecomenational or regional lenders. Theories presented by Dell’Ariccia & Marquez (forthcoming) and Hauswald & Marquez (2002)suggestthatgreatercompetitionbythesedistantlendersmaycauselocallenderstofocus ontheloansforwhichtheyhaveaninformationaladvantage,andthismayinvolveshorter,rather 2

than longer, distances between these lenders and their borrowers. While a relatively uniform loweringofthecostsoflendingatgreaterdistancesforallbankswouldclearlyimplyabroadening of the geographic markets relevant to small business lending, a change in lending costs that applies to only a few large banking organizations could, because of these considerations, imply something quite di(cid:11)erent for antitrust market de(cid:12)nitions and antitrust policy. To examine whether the relationship between distance and small business lending has been increasing across the board, or whether it has evolved asymmetrically across lenders, we exclusively examine within-market lending, where we consider only loans from local banks to local businesses. This approach diminishes the e(cid:11)ect of large banks that make loans nationwide, which may have driven the changes in average distance reported by Petersen & Rajan (2000) and Wolken & Rohde (2002). Instead, our paper examines whether the changes in technology have been adopted by a wide enough spectrum of banks to influence not only national lending, but local lending as well. Thispaperanalyzesanewsourceofdatacollectedannuallyasaresultofchangesintheregulations implementing the Community Reinvestment Act (CRA), that provide detailed spatial information of the type needed to investigate within-market spatial characteristics of commercial lending. As a result of these regulatory changes, lending institutions above a modest size reported annually on the small business loans that they originate. Most importantly, the loan data is reported by geographiclocationof the borrower. This, combinedwith detailed informationonthe locationsofthe o(cid:14)cesoflending institutions, makesitpossible tocalculate relevant distances between the lender and the borrower. Speci(cid:12)cally, in this paper we employ annual CRA data over the period 1997 - 2001 for nine di(cid:11)erent metropolitan areas in the United States to investigate the role of distance in withinmarket lending and how it has changed over time. One metropolitan statistical area (MSA) or consolidated metropolitan statistical area (CMSA) was randomly selected from the population of MSAs and CMSAs1 with populations in excess of one million for each of the nine census regions in the U.S. For each city, data were compiled for all CRA-reporting banks that had 3

establishedbranches within the boundariesof the MSAs. The data indicate the number of loan originations made by each bank to each of the census tracts within the MSA for each of the years 1997-2001. Aprobitmodelisemployedtoestimatethelikelihoodofabankmakingaloantoatleastone (cid:12)rm in each census tract in the metropolitan area, given the identify of the bank, the number of small businesses in the census tract, and the distance between the center of the census tract and the location of the nearest branch of that bank. Because of the paucity of data available at the census tract level, we employ spatial econometric methods. Speci(cid:12)cally, to account for unobservedvariablesthatmaybespatiallycorrelatedacrosscensustracts,theprobitisspeci(cid:12)ed, as in McMillen (1992), to allow for spatially dependent errors and to correct for the resulting heteroscedasticity. The results of this estimation procedureareused to test hypotheses relating to the role of distance in commercial lending and also to examine how that role has changed over time. Using these data, we (cid:12)nd that, even within areas generally considered to represent a local market, distance is negatively associated with the likelihood of a commercial loan being made. Further, this deterrent e(cid:11)ect of distance is consistently more important the smaller the size category of the bank. In other words, distance matters in explaining where a bank chooses to lend in a metropolitan area, and it matters more for smaller banks than for larger banks. With respect to the changing importance of distance over time, the results of this study suggest that there has been no discernable increase in the distance between lenders and their local borrowers in the United States in recent years. In fact, the bulk of the evidence suggests thatdistancemaybeofincreasingimportanceinlocalmarketlending. This(cid:12)ndingisconsistent with the theoreticalpredictionsofbothDell’Ariccia &Marquez(forthcoming)andHauswald& Marquez (2002). The plan of the paper is as follows. Section 2 summarizes the theoretical reasons why distance may be important in lending and reviews the empirical evidence, focusing on those 1Hereafter, theterm \MSA" will be used to refer to both MSAs and CMSAs. 4

studies that havespeci(cid:12)cally addressedhow the relationshipbetween distance and lending may be changing over time. Section 3 describes the estimation procedure employed to account for spatially correlatederrors. Section 4 discussesthe data usedand Section5 describes the results obtained. A (cid:12)nal section summarizes the analysis and presents conclusions. 2 Literature Studies have consistently demonstrated the importance of distance in the provision of banking services. Using data from the 1993 Survey of Small Business Finance (SSBF), Kwast, Starr-McCluer & Wolken (1997) report that 92.4 percent of small businesses use a depository institutionthatiswithinadistanceof30miles. Furthermore,theseauthors(cid:12)ndthatthemedian distance between a small business and its lender is six miles or less for lines of credit, mortgage loans, equipment loans, motor vehicle loans, and other loans. The only traditional credit product that had a greater median distance was capital leases at 39 miles. Using more recent data from the Credit, Banks and Small Business Survey, conducted by the National Federation of Independent Business, Scott(2003)(cid:12)nds that in2001,the averagedistance (measuredintravel time) between a small business and its primary (cid:12)nancial institution was 9.5 minutes, with a median of 5 minutes. Thetheoreticalliteraturehasputforthtwodi(cid:11)erentrationalesforwhydistanceshouldserve as a deterrent to lending. The (cid:12)rst, drawn from traditional models of spatial competition, is borrowertravelcost. Prospectiveborrowersmustincurtravelcoststodobusinesswithalender, inmuchthesamewaythatismorecommonlyassertedforadepositorthatchoosestodobusiness with one depository institution rather than another. Papers by Chiappori, Perez-Castrillo & Verdier (1995) and Park & Pennacchi (2003) are examples of this type of rationale. The second rationale, more speci(cid:12)c to the case of commercial credit, concerns the advantage that proximity may give lenders in screening perspective borrowers and monitoring loans, particularly in the case of loans to small businesses. Lenders, lacking the \hard" information 5

provided by detailed public (cid:12)nancial statements typically available for large (cid:12)rms, have to rely on \soft" information informally collected through relationships between the lender and the borrower.2 The collection of this soft information is costly to the lender, as it may require multiple site visits by a loan o(cid:14)cer to the small business or specialized knowledge of the local marketin which the (cid:12)rm operates. Additionally, banks may acquire informationon small (cid:12)rms through the provision of non-loan-related banking services, such as checking accounts that are mostoftenprovidedbylocalsuppliers. Intheseinstances,a(cid:12)rmwouldbemorelikelytoreceive favorable loan terms from lenders in closer proximity to the (cid:12)rm, as close lenders would incur lower costs to gather soft information.3 Almazan (2002) provides an example of this type of model. Most directly applicable to the issues relating to how changes in the competitive environment might alter the relationship between distance and lending are the theoretical models of Dell’Ariccia & Marquez (forthcoming) and Hauswald & Marquez (2002). If large banks have access to a cheaper source of funds (Kiser forthcoming), this may allow these banks to extend loans to more distant markets, even though the large banks might be at an informational disadvantage relative to closer, local (cid:12)nancial institutions. Dell’Ariccia & Marquez (forthcoming) provideamodelofthistypeofsituationandexaminehowtheextensionofcreditinlocalmarkets might be a(cid:11)ected by changes in either the cost advantages of the less-informed banks or in the degreeofinformationalasymmetryamong(cid:12)nancialinstitutions. Theauthorsshowthatgreater competition from outside lenders will cause local banks to reallocate credit towards borrowers for whom the local lenders possess an informational advantage (an e(cid:11)ect that Dell’Ariccia and Marquez term a \flight to captivity"). If proximity confers an informational advantage, this suggeststhatgreatercompetitionfromoutsidelendersmightresultinlocallendersreducingthe distance over which they extend credit to businesses. In a related paper, Hauswald & Marquez (2002) present a model that focuses on \informa- 2For a review of issues involved in small business lending, see Berger & Udell (1998). 3The importance of relationships in small business lending has received increasing attention in the academic literature in recent years. For a review of relationship banking see Boot (2000). 6

tional distance" and its relationship to investments in information acquisition technology by lenders. An implication of their model is that, as competition increases, banks may respond by shifting their resources to loans involving greater informational proximity. If informational proximity translates to physical proximity, then a reduction in the distance between borrower and lender as competition increases is a possible outcome. Together, the papers by Dell’Ariccia & Marquez (forthcoming) and Hauswald & Marquez (2002) suggest that the competitive changes brought about by technological changes in bank lending may have asymmetric e(cid:11)ects on the relationship between distance and lending. In particular,whiletechnologicalchangesmayleadsome(cid:12)nancialinstitutionstolendbeyondtheir local markets, the resulting changes in the competitive environment might lead local lenders to restrict their lending activities to a smaller geographic area. While these theoretical studies have established why distance would be important and how changesinthecompetitiveenvironmentmightaltertheimportanceofdistance,veryfewstudies have empirically examined the evolving relationship between distance and bank lending. One study by Petersen& Rajan(2002)examines how the relationshipbetweendistance andlending is changing over time. Using the 1993 SSBF, the authors conclude that the distance between small (cid:12)rms and their lenders is increasing over time and that this phenomenon is correlated with improvements in bank productivity. These (cid:12)ndings were based on the \synthetic panel" that the authors constructed from the 1993 SSBF cross section, wherein the distance between borrower and lender is compared with the time at which respondents in the 1993 SSBF report the lending relationship began. A more direct comparison over time that examines more recent years was conducted by Wolken & Rohde (2002), who compare results of the 1993 and 1998 SSBF surveys. They (cid:12)nd thatforsmallbusinessloansingeneral,theaveragedistancebetweenthebusiness’sheadquarters and the (cid:12)nancial institution making the loan increased from 115 miles in 1993 to 244 miles in 1998, while the median distance increased from only 9 to 10 miles during the same period. It is clear that this sharp distinction between the mean and median changes is driven by a sharp 7

increase in distance exhibited at the upper tail of the distribution. Decomposition according to type of loan reveals that this phenomenon has been particularly pronounced for capital leases and motor vehicle loans.4 A third study that addressed the relationship between distance and lending over time was conducted by Degryse & Ongena (2002). In a paper that examines how the distance between a borrower and a lender a(cid:11)ects the interest rate paid on loans from a large Belgian bank, the authors address the issue of whether the distances between borrowers and the bank have been changing over time. They conclude that the distance between the bank in their study and the European (cid:12)rms it served did not increase substantially between 1975 and 1997. Taken together, the results of the (cid:12)rst two empirical studies are consistent with the notion that there has been an increase over time in the distance between lender and borrower for \intermarket" loans, or loans made over longer distances. While this increase in distance may be the result of innovations in broad-basedscreening techniques, they provide little evidence of atrendindistancesformorelocal,\relationship-driven"loans. Similarly,theresultsofDegryse and Ongena (2002) suggest that the trend in the importance of distance may not be occurring for all banks. While it is possible that the innovations that have resulted in the rise of longer-distance \intermarket" loans may be at work to increase distances associated with shorter-distance, \intramarket" loans, this need not be the case. Indeed, the results of the existing empirical literature are not inconsistent with the predictions of Dell’Ariccia & Marquez (forthcoming) andHauswald&Marquez(2002)relatingtothe asymmetrice(cid:11)ectsonthe relationshipbetween distance and lending. 4Using the distinction made by Berger & Udell (1995), capital leases and motor vehicle loans are \transactiondriven" as opposed to \relationship-driven," which may partially explain why changes in the markets for those loan productsdi(cid:11)ered from other typesof small business lending. 8

3 The Empirical Model This paper employs a probit model that allows for spatially-dependent errors and corrects for heteroscedasticity. The probability that a bank extends credit to a census tract is a function of theattractivenessofthecommercialcreditmarketconditionsinthatcensustract. Attractiveness cana(cid:11)ectedbythedemandcharacteristicsofpotentialborrowersinthatcensustract,theability of banks to reach potential borrowers in that census tract, the degree of adverse selection the bank faces, or other considerations. A bank, b, makes a loan to census tract c in year t if the latentvariableL bct,whichrepresentstheattractivenessofthelendingenvironmentofthecensus tracttobankb,isgreaterthanzero. Thislatentvariableisassumedtofollowalinearfunctional form given by L bct =X bct (cid:12)+u bct (1) where u bct is a random normaldisturbance term, (cid:12) is a kx1 vector of coe(cid:14)cients, and X bct is a 1xk vector of explanatory variables reflecting the characteristics of the bank making the loan, thecensustractintowhichtheloansaremade,andtheinteractionbetweenthebankandcensus tract (such as the distance between the bank and the census tract). The random disturbance terms are assumed to have a spatial autoregressive structure, so P that u bct = (cid:26) j w cj u bjt+(cid:15) bct, where (cid:26) is a parameter indicating the strength of the spatially autoregressive process, w cj is a measure of spatial contiguity, and (cid:15) bct is an i.i.d. draw from a standard normal distribution. The parameter (cid:26) therefore reflects the influence of unobserved characteristics of neighboring census tracts that are correlated across space, rather than the impact of observations across time. Written in matrix notation, the model speci(cid:12)cation can be expressed as 9

L=X(cid:12)+(I−(cid:26)W) −1(cid:15) (2) and Euu0 =[(I−(cid:26)W) 0 (I−(cid:26)W)] −1: (3) W is the spatial weight matrix that describes the spatial pattern of the autoregressive process. In this application w cj is equal to the proportion of census tract c0s border that is shared with census tract j. If census tracts c and j are nonadjacent, then w cj is equal to zero. Following convention in the spatial econometrics literature, w cc is equal to zero.5 The complex error structure in this model makes the use of direct maximum likelihood methods infeasible, so we employ the expectation-maximization (EM) algorithm, an iterative method of estimating (cid:12) and (cid:26).6 First, we begin with starting values of (cid:12)^ and (cid:26)^. Using these values, the expected value ofthe latent variableL^ bct is calculated,conditionalon whether bank b actually was observed making a loan to census tract c in year t. New values of (cid:12)^ and (cid:26)^ are then selected to maximize the likelihood function derived by McMillen (1992), X max=−0:5(L bct −X(cid:12)) 0 (I −(cid:26)W) 0 (I −(cid:26)W)(L bct −X(cid:12))+ ln(1−(cid:26)! i); (4) (cid:12);(cid:26) i where ! i is the i-th eigenvalue of the spatial weight matrix. These values of (cid:12)^ and (cid:26)^are then used to update the expected value of the latentvariable L^ bct. The EMalgorithmcontinues this process until the change in the values of (cid:12)^ and (cid:26)^fall below a speci(cid:12)ed tolerance.7 5For a review of theuse of spatial weight matrices in spatial econometrics, see Anselin (1998) or LeSage (1998). 6Amemiya (1985) providesan overview of theapplication of theEM algorithm to discrete choice models. 7The tolerance used in theestimations reported in thispaper was 10−8. 10

4 Data ThedataonsmallbusinesslendingpatternsarecollectedbytheFederalReserveBoardtosatisfy theCommunityReinvestmentAct(CRA).Underthisact,independentlendinginstitutionswith total assets greater than $250million, or institutions of any size if they are owned by a holding company with more than $1 billion in total assets, must report annually on the small business loans that they originate or hold. Both the number and volume of loans are recorded for each reporting bank for the census tract to which the loans were made. Due to the large number of census tracts in the United States and the computational complexityinvolvedinestimatingaspatialprobitmodel,astudythatencompassestheentireUnited States is not feasible. As a result, this study focuses on lending patterns in nine cities located throughoutthe U.S.8 These cities wereselectedatrandomfromthe universeofMSAs andCM- SAs withpopulations over1 million. One city wasselectedfromeachofthe nine censusregions to make the sample geographically diverse. The MSAs selected include the cities of Atlanta, GA;Denver,CO;Indianapolis,IN;KansasCity,MO;Nashville,TN;Providence,RI;Rochester, NY;SanAntonio,TX;andSeattle,WA.Thecompletenamesandcodenumbersofeachofthese MSAs are provided in Table 1. CRA data were compiled for each of these MSAs. Each of the CRA-reporting banks in each of the nine cities was checked against the FDIC’s Institution Directory to determine if that bank had established at least one o(cid:14)ce within the MSA. For those banks with o(cid:14)ces in the MSA, a list of addresses of each bank’s branches in that MSA was compiled for each year. These addresses were then geocoded to the longitude andlatitude coordinatesofthe branchaddress. Banksthatweremakingloanswithinthe MSA, but that did not have a local branch presence (\out of market lenders"), were excluded from the sample to focus exclusively on local market lending. 8Over the years, the academic literature has provided substantial support for the existence of geographically limited markets for both retail deposits and small business loans. The bulk of the evidence has come in the form of price-concentration studies, wherein bankdeposit ratesor bankloan rates areregressed onmeasures of local market concentration,calculatedundertheassumptionthatMSAsandnon-MSAcountiesapproximatelocalbankingmarkets. See, for example, Berger & Hannan (1989), Calem & Carlino (1991), and Hannan (1991). 11

FIPS Code Name 0520 Atlanta, GA MSA 2082 Denver-Boulder-Greeley, CO CMSA 3480 Indianapolis, IN MSA 3760 Kansas City, MO-KS MSA 5360 Nashville, TN MSA 6480 Providence-Fall River-Warwick, RI-MA MSA 6840 Rochester, NY MSA 7240 San Antonio, TX MSA 7602 Seattle-Tacoma-Bremerton, WA CMSA Table 1: Metropolitan Statistical Areas Used in this Study AmongloansmadebyCRA-reportinginstitutions,theshareofloansmadebywithinmarket lenders has fallen substantially overthe time frame ofthis study.9 This decline, which is shown in Table 2, is demonstrative of the growing importance of out-of-market lenders. While the market share of the within market lenders has fallen in terms of number of loans made each year, market share in terms of loan volume (shown in Table 3) has been much more stable. In 2001,the mostrecentyearofdata,the within-marketlendersexaminedinthis study accounted for at least 80 percent of dollar volume of loans reported by CRA-reporting banks in each of the nine MSAs examined. For each MSA, the model was estimated using a panel data set consisting of census tract observations over the years 1997 to 2001. For each of the within market lenders, the log of the great circle distance (or \distance as the crow flies") between the centroid of each census tract and the closest branch of the bank in that MSA was calculated. Since this study is primarilyinterestedinhowthe importanceofdistanceischangingovertime, thelogofdistance is also interacted with a linear time trend. To allow for the possibility that distance may di(cid:11)er in importance to banks of di(cid:11)erent sizes, both of these variables are interacted with dummy variables for medium (assets between $500 million and $5 billion) and large size banks (assets 9Thesmall businessloansreportedby(cid:12)nancialinstitutionsunderCRArequirementsincludeloansmadethrough credit cards, and such loans tend to involve substantial distances between lender and borrower. Also, the loans of institutions not large enough to come under CRA reporting requirements tend to be more local. For both of these reasons, the share of CRA-reported loans made by within-market institutions will underestimate the true share of within-market non-credit card loans. Nonetheless, it is instructive to examinethese observed shares. 12

Variable 1997 1998 1999 2000 2001 Atlanta 47.1 53.6 46.9 25.0 25.1 Denver 50.7 56.4 54.9 30.2 29.0 Indianapolis 57.4 52.9 53.4 32.1 28.9 Kansas City 53.1 51.4 44.8 29.3 29.7 Nashville 67.0 61.3 61.8 36.4 40.0 Providence 43.1 46.8 36.8 18.6 24.7 Rochester 70.0 66.0 61.5 42.8 43.4 San Antonio 62.6 55.0 40.4 23.1 22.7 Seattle 71.0 65.7 56.6 33.7 25.9 Table 2: Market Share of Loans Made by CRA-Reporting Institutions Operating within the MSA Variable 1997 1998 1999 2000 2001 Atlanta 83.5 88.8 86.2 80.7 83.9 Denver 82.7 85.3 85.8 79.4 81.4 Indianapolis 89.4 88.7 88.6 84.6 80.2 Kansas City 88.2 90.5 86.1 82.7 85.5 Nashville 90.6 91.9 90.0 86.0 87.6 Providence 84.0 85.3 81.6 70.2 80.8 Rochester 94.5 92.2 92.1 89.0 90.5 San Antonio 88.0 85.9 84.8 76.6 82.6 Seattle 95.1 93.3 91.0 84.1 85.5 Table 3: Market Share of Loan Volume by CRA-Reporting Institutions Operating within the MSA over $5 billion). Also, to account for bank-speci(cid:12)c heterogeneity, bank-speci(cid:12)c (cid:12)xed e(cid:11)ects are employed in each of the estimations performed. Factors other than distance and bank characteristics { for example, the characteristics of potentialborrowersinthecensustract{areimportantinbanklendingdecisions. Aconsiderable problem encountered in conducting analyses of bank lending at the a census-tract level is the lack of available data at that level. One source is Dunn & Bradstreet, which compiles data on the number of businesses located in each census tract in the United States over the time period of this study. The log of the number of (cid:12)rms is included as anindication of the strength of demand for loans in each census tract. In addition, to account for additional unobserved heterogeneityacrosscensustracts,census-tract-speci(cid:12)c(cid:12)xede(cid:11)ectsareemployedinsomeofthe model speci(cid:12)cations. 13

Variable Atlanta Denver Indianapolis KansasCity Nashville Providence Rochester SanAntonio Seattle Tracts 495 507 331 443 204 335 260 255 613 Observations 73,260 107,991 35,086 62,020 14,892 29,815 18,980 24,480 77,238 Banks 45 58 32 41 24 22 20 34 38 Bank-years 148 213 106 140 73 89 73 96 126 Small 68 140 28 66 12 21 21 14 45 Medium 42 44 45 50 15 48 17 37 48 Large 38 29 33 25 46 20 35 45 33 MeanBankAssets 26.680 4.4541 9.1747 3.5048 52.249 9.6832 39.1395 52.9236 22.7584 Std. Dev. BankAssets 66.425 14.065 15.844 11.750 110.11 28.227 60.3034 109.121 79.5468 MeanDistance 20.764 11.805 10.846 13.096 9.6631 10.624 14.1808 9.6763 17.8570 Std. Dev. Distance 16.874 11.869 9.7368 11.781 9.6803 9.0375 14.0914 9.9317 17.2731 MeanFirmCount 351.0 235.9 175.24 158.99 263.126 180.38 179.064 215.863 296.0536 Std. Dev. Firms 322.4 236.5 176.89 144.81 234.742 135.19 143.148 209.019 266.1723 Notes: Meansandstandarddeviationsforassets,distance,and(cid:12)rmcountsarefor2001. Assetsaremeasuredinbillionsofdollars. Table 4: Summary Statistics 14

Summary statistics for the data set employed in this paper are provided in Table 4. The summary statistics show a substantial amount of heterogeneity across MSAs. In terms of the number of census tracts, the size of the MSAs range from Nashville with 204 tracts to Denver which has almost twice as many. Even greater disparities are seen in cross-MSA comparisons of the remaining variables. The distribution of observations across bank size categories also is substantially di(cid:11)erent across the MSAs examined. These statistics suggest that the MSAs selected for use in this study represent a broad cross section of large MSAs. 5 Results Table5listsandde(cid:12)nesallvariablesusedintheestimationsreportedinthesucceedingtables.10 Complete estimation results are presentedin Tables 12 through 20 in the appendix. Eachtable presented in the text provides a summary of the results across di(cid:11)erent speci(cid:12)cations for those coe(cid:14)cients of particular interest. Variable Description D1998-D2001 Year dummy variables S Small (cid:12)rm dummy (assets < $500 million) M Medium (cid:12)rm dummy ($500 million < assets < $5 billion) L Large (cid:12)rm dummy (assets > $5 billion) LnD Log of distance between census tract and nearest bank branch (in miles) T Linear time trend (1997 = 0) LnFirms Log of (cid:12)rm count in census tract Table 5: Variable List Multiple model speci(cid:12)cations were estimated for each of the nine MSAs. Each estimation includes bank-speci(cid:12)c (cid:12)xed e(cid:11)ects, year dummies (D1998−D2001), the log of one plus the number of small (cid:12)rms in each census tract as reported by Dunn & Bradstreet (LnFirms)11, and the log of the distance from the centroid of each census tract to the nearest branch of the 10The dummy variables used for bank-level and census-tract-level (cid:12)xed e(cid:11)ects are not listed in the tables nor reported in with the results to conservespace.’ 15

bank (LnD). Twoalternativespeci(cid:12)cationswereestimatedtoallowforthepossibilitythattheimportance of distance may be changing over time. In speci(cid:12)cation I, LnD is interacted with a linear time trend that is constant across banks. The second speci(cid:12)cation allows for the possibility that the changing importance of distance might di(cid:11)er across size category of (cid:12)nancial institution by interacting LnD with separate linear time trends for small, medium, and large (cid:12)nancial institutions. To accountfor the possibility of unobservedheterogeneity acrosscensus tracts eachof these twospeci(cid:12)cations wasalsoestimatedwith andwithout time-invariant,census-tract(cid:12)xede(cid:11)ects (TFE).Inaddition,totestthesensitivityofthemodeltotheassumptionofspatially-correlated errors, we also estimate probit models where the residuals are assumed to be independently distributed. The resultsofboththe spatialandaspatialprobitsarequalitativelysimilar. While thediscussioninthefollowingsectionsfocusesexclusivelyontheresultsfromthespatialprobit, the coe(cid:14)cients from both models are presented in the Appendix. 5.1 Speci(cid:12)cation I: Single Linear Time-Distance Trend Theestimationsthatincludedasinglelineartimetrendontheimpactofdistanceacrosssizeclass of (cid:12)nancial institution provide consistent results regarding the determinants of bank lending to census tracts. Table 6 provides a summary of the coe(cid:14)cients on LnFirms for the speci(cid:12)cation that includes a single linear time-distance trend for the cases in which tract (cid:12)xed e(cid:11)ects are included (\TFE") and not included (\No TFE"). For the estimations that do not include tract (cid:12)xed e(cid:11)ects, the coe(cid:14)cient on LnFirms is uniformly positive and highly signi(cid:12)cant in each of the nine MSAs. When census tract (cid:12)xed e(cid:11)ects areincluded, inwhichcaseidenti(cid:12)cationofthe e(cid:11)ect of LnFirmscomes fromvariations 11Additionalestimationsthatdecomposethenumberofbusinessesintotheirone-digitSICclassi(cid:12)cationswerealso performed. Theresultsfromtheseestimationsaresimilartotheresultsreportedhere,thoughthecollinearityamong the SIC variables was such that many of the SIC coe(cid:14)cients had negative or inconsistent signs. The coe(cid:14)cients on theothervariables werelargely una(cid:11)ected. Tosavespace,thesespeci(cid:12)cations are notreported in thepaper,though they are available from the authorsupon request. 16

MSA No TFE TFE Atlanta 0.6527 0.1020 (74.9016) (2.4671) Denver 0.7922 0.1925 (88.2126) (5.1061) Indianapolis 0.5567 0.1288 (53.7225) (3.1433) Kansas City 0.6056 0.0275 (69.7249) (0.7799) Nashville 0.4805 0.0252 (24.9078) (0.3286) Providence 0.4906 0.0767 (42.5173) (1.3793) Rochester 0.7668 -0.0722 (46.6951) (-0.8234) San Antonio 0.6093 -0.0063 (46.4820) (-0.1258) Seattle 0.5279 0.0490 (65.7587) (1.4696) Notes: TFE denotes the coe(cid:14)cients are from estimations that employed census tract (cid:12)xed e(cid:11)ects. T-statistics are in parentheses. As with McMillen (1992), the standard errors used to calculate the t-statistics in the estimations with spatially-dependent errors are conditional on (cid:26) and so that the t-statistics overestimate the unconditional t-statistics. Table 6: Coe(cid:14)cients on Log Number of Firms in Estimations with Spatially-Correlated Errors and a Single Time-Distance Trend in the number of (cid:12)rms overtime, the e(cid:11)ect is not as strong and statistically signi(cid:12)cantat the 1 percent level for only Atlanta, Denver, and Indianapolis. The results suggest, not surprisingly, that as the number of small businesses in a census tract increases,the probability of each bank extending credit to (cid:12)rms in that tract increases. The coe(cid:14)cients on the log of distance are presented in Table 7. Like the coe(cid:14)cients on the number of (cid:12)rms, these coe(cid:14)cients also tell a consistent story across the di(cid:11)erent speci(cid:12)cations and MSAs. The coe(cid:14)cient on the baseline distance measure, LnD, is negative and highly signi(cid:12)cant in every case, implying that the greater the distance between a census tract and a bank’snearestbranch,thelesslikelythebankwillhaveextendedaloantoa(cid:12)rminthatcensus tract. 17

No TFE TFE LnD LnD*M LnD*L LnD LnD*M LnD*L Atlanta -1.6009 0.9095 1.4368 -1.5666 0.6691 0.9575 (-61.4929) (27.6890) (47.7137) (-61.2260) (20.6880) (31.7892) Denver -0.7638 0.2824 0.5852 -1.0249 0.2149 0.4500 (-52.0910) (13.5252) (27.5067) (-63.6699) (10.5362) (21.4020) Indianapolis -1.2034 0.5052 0.7793 -1.2990 0.3845 0.5992 (-30.1241) (12.3742) (18.5308) (-32.1569) (9.4401) (14.3407) Kansas City -0.7877 0.2917 0.5248 -1.0202 0.2256 0.4375 (-36.7291) (11.4324) (19.4873) (-44.8768) (8.9259) (16.2805) Nashville -0.8682 0.4503 0.7821 -1.1283 0.2801 0.5870 (-12.3822) (5.6968) (11.9143) (-16.5094) (3.6743) (9.0109) Providence -0.8200 -0.1394 0.3521 -1.0893 -0.0607 0.2295 (-22.0269) (-3.6569) (8.1159) (-29.1095) (-1.6419) (5.5327) Rochester -1.5347 0.5214 1.0324 -1.6730 0.4499 0.8154 (-24.8780) (6.9621) (16.0307) (-26.0984) (5.7258) (11.7985) San Antonio -0.5982 0.2501 0.4183 -0.8360 0.1814 0.2725 (-10.8635) (4.3144) (7.5817) (-14.8758) (3.2030) (4.9717) Seattle -1.0104 0.1971 0.8053 -1.1277 0.1268 0.6162 (-43.3017) (7.7701) (30.9358) (-46.8186) (4.9405) (22.8121) Notes: TFE denotes the coe(cid:14)cients are from estimations that employed census tract (cid:12)xed e(cid:11)ects. T-statistics are in parentheses. As with McMillen (1992), the standard errors used to calculate the t-statistics in the estimations with spatially-dependent errors are conditional on (cid:26) and so that the t-statistics overestimate the unconditional t-statistics. Table 7: Coe(cid:14)cients on Log Distance in Estimations with Spatially-Correlated Errors and a Single Time-Distance Trend For eight of the nine MSAs, distance was less of a deterrent at medium-sized (cid:12)nancial institutions than at small institutions, a di(cid:11)erence that was statistically signi(cid:12)cant at the 99 percentlevelineachcase. IntheremainingMSA,Providence,distancewasmoreofadeterrentat medium-sized institutions than at small institutions, though this di(cid:11)erence was only signi(cid:12)cant in the estimation without tract (cid:12)xed e(cid:11)ects. At larger institutions, the deterrent e(cid:11)ect of distance was lower than at either small or medium institutions across all nine MSAs for both the estimations with and without tract (cid:12)xed e(cid:11)ects. With a single linear time-distance trend across bank size categories, the deterrent e(cid:11)ect of distance in bank lending does not appear to be diminishing over time. The coe(cid:14)cients for the estimations involving spatially-correlated errors are summarized in Table 8. The results presentedinthattableindicatethat,ifanything,theimportanceofproximityhasbeengrowing 18

over time, though these results are not as clear as those for the number of (cid:12)rms or distance. The time-distance coe(cid:14)cient (the coe(cid:14)cient on the interaction of T and LnD) is negative and signi(cid:12)cant at the 1 percent level for three cities (Atlanta, Nashville, and Providence), positive and signi(cid:12)cant for Indianapolis, and otherwise either insigni(cid:12)cant or negative and signi(cid:12)cant only in the estimations without tract (cid:12)xed e(cid:11)ects. MSA No TFE TFE Atlanta -0.0477 -0.0551 (-8.5391) (-10.1045) Denver 0.0006 0.0026 (0.1353) (0.5764) Indianapolis 0.0436 0.0550 (5.5568) (7.0598) Kansas City -0.0174 0.0022 (-2.8543) (0.3645) Nashville -0.0628 -0.0418 (-4.3098) (-2.9410) Providence -0.0359 -0.0388 (-4.3087) (-4.7124) Rochester -0.0129 -0.0008 (-1.2393) (-0.0778) San Antonio -0.0260 0.0156 (-2.5378) (1.5164) Seattle -0.0086 0.0012 (-1.5666) (0.2162) Notes: TFE denotes the coe(cid:14)cients are from estimations that employed census tract (cid:12)xed e(cid:11)ects. T-statistics are in parentheses. As with McMillen (1992), the standard errors used to calculate the t-statistics in the estimations with spatially-dependent errors are conditional on (cid:26) and so that the t-statistics overestimate the unconditional t-statistics. Table 8: Coe(cid:14)cients on Time*Distance in Estimations with Spatially-Correlated Errors and a Single Time-Distance Trend While the results discussed above indicate the direction of the e(cid:11)ects and trends examined, theyprovidelittleinformationaboutthedegreetowhichthesetrendsande(cid:11)ectsareofeconomic importance. Todetermine the economicimportanceofdistance andhowrapidlyits importance is changing, if at all, over time, marginal e(cid:11)ects were calculated for distance and time. These marginale(cid:11)ects indicate the changein the probability ofone ormore loansbeing extended to a 19

census tract that would result from a one mile increase in distance or from a one year increase in time. These marginal e(cid:11)ects are calculated for 2001 using the average values presented in the summarydata(Table4)asvaluesfortheexplanatoryvariables. Furthermore,whencalculating the probabilities used in calculating the marginal e(cid:11)ects, we assume that the spillover e(cid:11)ects P from neighboring census tracts balance out, so that (cid:26) j w cj u bjt = 0. All (cid:12)xed e(cid:11)ects are assumed to take on their averagevalues across (cid:12)nancial institutions and census tracts. For estimations that include a single linear time-distance term, distance has a substantial impact on the probability that a (cid:12)nancial institution extends credit to a census tract. The medianmarginale(cid:11)ects(acrossmodelspeci(cid:12)cationsandMSAs)ofdistancearelargeratsmaller institutions thanatlargerinstitutions. Speci(cid:12)cally,a one-mileincreaseindistance isassociated with a 1.75 percentage point decrease in the probability of credit being extended to a census tract at small institutions, compared to decreases of 1.4 and 0.8 percentage points at medium and large institutions, respectively.12 While the coe(cid:14)cients on the time-distance trend variable provided a somewhatinconsistent indication of how distance is changing over time, the marginal e(cid:11)ects provide a clearer view. While individual marginal e(cid:11)ects move in either direction, the median marginal e(cid:11)ect of time is a decrease across each of the three institution size classes. A one-year increase corresponds to a decrease in the probability of a (cid:12)nancial institution extending credit to a census tract of 0.12 percentage points at small institutions, 0.14 at medium institutions, and 0.07 at large institutions. While these results suggest that distance is becoming more of a deterrent to local bank lending, the magnitude of the increase in the deterrent is very small. 12To reduce the number of tables included in the paper, the complete set of marginal e(cid:11)ects are not reported here. A full set of marginal e(cid:11)ects for (cid:12)rms, distance, and the distance-time interaction can be found at http://ken.brevoort.com or are available from theauthors upon request. 20

5.2 Speci(cid:12)cation II: Multiple Linear Time-Distance Trends The results from the model speci(cid:12)cations that allow for separate linear time-distance trends for each of the three institution size classes are presented in Tables 9 through 11. Table 9 presents the coe(cid:14)cients on LnFirms for both the estimations with and without census tract (cid:12)xede(cid:11)ects. As wasthe casefor the estimations witha singlelinear time trend, the coe(cid:14)cients on LnFirms from the estimations that omitted census tract (cid:12)xed e(cid:11)ects are all positive and signi(cid:12)cantatlevels less than 1 percent. For the estimations that included tract(cid:12)xed e(cid:11)ects the resultswerepositiveandsigni(cid:12)cantatthe1percentlevelforthreeMSAs (Atlanta,Denver,and Indianapolis) and otherwise insigni(cid:12)cant. MSA No TFE TFE Atlanta 0.6537 0.1108 (74.9123) (2.6743) Denver 0.7947 0.1825 (88.3409) (4.8298) Indianapolis 0.5572 0.1173 (53.7251) (2.8582) Kansas City 0.6058 0.0286 (69.7280) (0.8097) Nashville 0.4800 -0.0000 (24.8984) (-0.0006) Providence 0.4907 0.0810 (42.5195) (1.4543) Rochester 0.7664 -0.0813 (46.6711) (-0.9246) San Antonio 0.6103 -0.0013 (46.5209) (-0.0251) Seattle 0.5299 0.0642 (65.8874) (1.9223) Notes: TFE denotes the coe(cid:14)cients are from estimations that employed census tract (cid:12)xed e(cid:11)ects. T-statistics are in parentheses. As with McMillen (1992), the standard errors used to calculate the t-statistics in the estimations with spatially-dependent errors are conditional on (cid:26) and so that the t-statistics overestimate the unconditional t-statistics. Table 9: Coe(cid:14)cients on Log Number of Firms in Estimations with Spatially-Correlated Errors and Multiple Time-Distance Trends As with the coe(cid:14)cients on LnFirms, the coe(cid:14)cients on LnD in estimations with multiple 21

linear time-distance trends tell a very consistent story (Table 10). Indeed, the coe(cid:14)cients from the multipletime-distancetrendestimationsexhibitthesamegeneralpatternasthecoe(cid:14)cients from the estimations where a single time-distance trendwas estimated. Speci(cid:12)cally, in all cases distancehasastatisticallysigni(cid:12)cantnegativeimpactontheprobabilitythatabankwillextend credit to a census tract. Furthermore,once againdistance is less of a deterrent to bank lending atmedium-sizedinstitutionsthanatsmallerinstitutionsforeveryMSAbutProvidence. Andin all cases distance has a greater deterrent e(cid:11)ect at large banks than at small or medium banks. No TFE TFE LnD LnD*M LnD*L LnD LnD*M LnD*L Atlanta -1.7618 1.1447 1.6101 -1.6532 0.8029 1.0525 (-46.9569) (21.4704) (33.0718) (-45.8293) (15.5226) (22.1860) Denver -0.7452 0.0902 0.7203 -0.9946 0.0001 0.5771 (-40.1410) (2.8614) (22.0695) (-51.6661) (0.0039) (18.1084) Indianapolis -1.3793 0.7811 0.8895 -1.4207 0.6266 0.6297 (-19.8564) (10.4609) (10.9960) (-20.1897) (8.3009) (7.7856) Kansas City -0.7637 0.2481 0.4965 -0.9806 0.1587 0.3876 (-26.9537) (6.0592) (11.7880) (-33.5189) (3.9059) (9.2515) Nashville -0.5903 0.3547 0.3841 -0.8478 0.1583 0.1883 (-5.4015) (2.7338) (3.1889) (-8.0416) (1.2661) (1.6022) Providence -0.7539 -0.2442 0.3041 -1.0437 -0.1352 0.2010 (-13.0527) (-3.7464) (4.2719) (-18.3909) (-2.1198) (2.9247) Rochester -1.6262 0.6435 1.1273 -1.7515 0.5861 0.8798 (-19.9808) (6.8103) (11.1606) (-20.9051) (5.9812) (8.4752) San Antonio -0.6775 0.2749 0.5474 -0.8748 0.1784 0.3468 (-8.5623) (3.0136) (6.1856) (-10.8117) (1.9613) (3.8945) Seattle -1.0782 0.4812 0.7931 -1.1833 0.4121 0.5772 (-32.4111) (10.6834) (18.4302) (-34.1836) (8.9494) (12.9643) Notes: TFE denotes the coe(cid:14)cients are from estimations that employed census tract (cid:12)xed e(cid:11)ects. T-statistics are in parentheses. As with McMillen (1992), the standard errors used to calculate the t-statistics in the estimations with spatially-dependent errors are conditional on (cid:26) and so that the t-statistics overestimate the unconditional t-statistics. Table10: Coe(cid:14)cientsonLogDistanceinEstimationswithSpatially-CorrelatedErrorsandMultiple Time-Distance Trends While the coe(cid:14)cients on the time-distance trend variables exhibit a pattern that is similar to the results from the estimations that employed a single time-distance trend (Table 11), the resultsvaryacrossbanksizecategoriesandMSAs. Forsmallbanks,thetime-distancecoe(cid:14)cient was positive and signi(cid:12)cant at the one percent level for the Indianapolis estimations, negative 22

and signi(cid:12)cant for Nashville, Providence, and the Kansas City estimation without tract (cid:12)xed e(cid:11)ects, and otherwise statistically insigni(cid:12)cant and of mixed sign. No TFE TFE LnD*T*S LnD*T*M LnD*T*L LnD*T*S LnD*T*M LnD*T*L Atlanta 0.0217 -0.0817 -0.0542 -0.0181 -0.0760 -0.0590 (1.7094) (-5.8088) (-6.3744) (-1.4826) (-5.7407) (-7.0179) Denver -0.0075 0.0798 -0.0717 -0.0106 0.0861 -0.0701 (-1.0739) (8.6400) (-8.0947) (-1.6017) (9.6349) (-8.0239) Indianapolis 0.1066 -0.0092 0.0738 0.0988 -0.0096 -0.0963 (4.8676) (-0.7228) (5.7571) (4.4911) (-0.7814) (7.6390) Kansas City -0.0285 -0.0082 -0.0153 -0.0161 0.0149 0.0071 (-2.7012) (-0.7690) (-1.3789) (-1.5155) (1.3984) (0.6592) Nashville -0.1884 -0.1873 -0.0070 -0.1676 -0.1496 0.0143 (-5.2946) (-5.7347) (-0.3835) (-4.9368) (-4.6606) (0.7834) Providence -0.0650 -0.0177 -0.0462 -0.0583 -0.0251 -0.0484 (-3.0494) (-1.3658) (-3.3113) (-2.8395) (-1.9865) (-3.4950) Rochester 0.0300 -0.0331 -0.0146 0.0365 -0.0339 -0.0162 (1.1377) (-1.3143) (-0.8578) (1.3651) (-1.7010) (-1.1728) San Antonio 0.0109 0.0001 -0.0515 0.0332 0.0351 -0.0024 (0.3910) (0.0073) (-3.4789) (1.1959) (2.1267) (-0.1625) Seattle 0.0179 -0.1018 0.0292 0.0227 -0.0987 0.0452 (1.6463) (-9.3174) (3.3914) (2.0108) (-9.0180) (5.1936) Notes: TFE denotes the coe(cid:14)cients are from estimations that employed census tract (cid:12)xed e(cid:11)ects. T-statistics are in parentheses. As with McMillen (1992), the standard errors used to calculate the t-statistics in the estimations with spatially-dependent errors are conditional on (cid:26) and so that the t-statistics overestimate the unconditional t-statistics. Table 11: Coe(cid:14)cients on Time*Distance in Estimations with Spatially-Correlated Errors and Multiple Time-Distance Trends The results for medium and large banks were also more often negative when statistically signi(cid:12)cant, but again not uniformly. When signi(cid:12)cant, the time-distance e(cid:11)ect among medium banks was negative for three cities (Atlanta, Nashville, and Seattle) and positive for one (Denver). Likewise, the time-distance e(cid:11)ect among large banks was negative and signi(cid:12)cant across both speci(cid:12)cations three times (Atlanta, Denver, and Providence) and positive once (Seattle). Additionally, the time-distance e(cid:11)ect for Indianapolis was positive and signi(cid:12)cant in the estimation without tract (cid:12)xed e(cid:11)ects and negative and signi(cid:12)cant in the estimation with (cid:12)xed e(cid:11)ects. Marginal e(cid:11)ects were also calculated for the estimations including multiple time-distance 23

e(cid:11)ects. These marginale(cid:11)ects exhibita trendthat is once againverysimilarto the results that were estimated for the case of the single time-distance trends. Speci(cid:12)cally, the marginal e(cid:11)ect of a one-mile increase in the distance between a bank and a census tract with characteristics equal to the sample means (presented in Table 4) is a decrease of 1.7 percentage points in the probability of extending a loan at small banks, 1.4 at medium banks, and 0.7 at large banks. The marginal e(cid:11)ect of a one-year increase in time are of similar magnitude to the results from the single time-distance trend results. The median e(cid:11)ect of a one-year increase in time is an increase in the probability of a bank making a loan to a census tract of 0.06 percentage points for small banks, and decreases of 0.19 and 0.15 percentage points for medium and large banksrespectively. Theseresultssuggestthatanychangethatmightbeoccurringinlocalbank lending patterns over time is relatively minor. Neither the results for the single or multiple time-distance trend estimations support the (cid:12)ndingthatthedeterrente(cid:11)ectofdistancehasbeenincreasingovertimeforwithin-marketloans. In fact, to the extent thatthe results ofthe estimations performedas partofthis study provide evidence ofatrendacrosscities,it ismoreoftenthe casethat distanceis becomingincreasingly importantinlocallending,althoughthemarginale(cid:11)ectsforthetime-distancevariable,forboth the estimationswithsingle andmultiple trends, suggestthat anysuchchangesarelikelyminor. 24

6 Conclusions We employ annual CRA data over the period 1997 - 2001 for nine di(cid:11)erent metropolitan areas in the United States to investigate the role of distance in within-market lending and how it has changed over time. A probit model that accounts for possible spatial correlation is used to estimate the likelihood of a bank making a loan to at least one (cid:12)rm in each census tract in the metropolitan area,given the identity of the bank, the number of small businesses in the census tract, and the distance between the center of the census tract and the location of the nearest branch of that bank. We (cid:12)nd that, even within areas generally considered to represent a local market, distance betweenbankandcensustractisnegativelyassociatedwiththe likelihoodofacommercialloan being made. Further, this deterrent e(cid:11)ect of distance is more important, the smaller is the size category of the bank. In other words, distance matters in explaining where a bank chooses to lend in a metropolitan area, and it matters more for smaller banks than for larger banks. Theobservedchangesintheroleofdistanceoverthetimeperiodexaminedarenotasrobust across metropolitan areas or size categories of the banks. However, we do (cid:12)nd that within a majority of the metropolitan areas examined, distances between borrower and lender became even more negatively associated with the likelihood of an observed commercial loan over time, consistentwiththenotionthat,ascompetitionincreasesfromlenderslocatedoutsidethemarket, local lenders reallocate resources toward loans in which they enjoy a locational advantage. While this study doesnotconstitute anexplicit structuraltestofthe theoreticalpredictions of either Dell’Ariccia & Marquez (forthcoming) or Hauswald & Marquez (2002) the results presented here are, in general, consistent with their predictions. Our (cid:12)ndings, when combined withtheexistingempiricalliteraturesuggestthatthetechnologicalchangesincreditmarketsare havingasymmetrice(cid:11)ects acrossinstitutionsizecategoriesontherelationshipbetweendistance and lending at (cid:12)nancial institutions. 25

References Almazan, A. (2002). Amodel ofcompetition in banking: Bank capitalvs.expertise,Journal of Financial Intermediation 11: 87{121. Amemiya, T. (1985). Advanced Econometrics, Harvard University Press. Anselin,L.(1998). Gisresearchinfrastructureforspatialanalysisofrealestatemarkets,Journal of Housing Research 9(1): 113{133. Berger, A. N. & Hannan, T. H. (1989). The price-concentration relationship in banking, The Review of Economics and Statistics 71(2): 291{299. Berger, A. N. & Udell, G. F. (1995). Relationship lending and lines of credit in small (cid:12)rm (cid:12)nance, The Journal of Business 68(3): 351{381. Berger, A. N. & Udell, G. F. (1998). The economics of small business (cid:12)nance: The roles of private equity and debt markets in the (cid:12)nancial growth cycle, Journal of Banking and Finance 22(6-8): 613{673. Boot, A. W. A. (2000). Relationship banking: What do we know?, Journal of Financial Intermediation 9: 7{25. Calem, P. S. & Carlino, G. A. (1991). The concentration/conduct relationship in bank deposit markets, The Review of Economics and Statistics 73(2): 268{276. Chiappori, P.-A., Perez-Castrillo, D. & Verdier, T. (1995). Spatial competition in the banking system: Localization, cross subsidies and the regulation of deposit rates, European Economic Review 39(5): 889{918. Degryse, H. & Ongena, S. (2002). Distance, lending relationships, and competition. Mimeo. Dell’Ariccia, G. & Marquez, R. (forthcoming). Informationand bank credit allocation,Journal of Financial Economics . Hannan,T.H.(1991).Bankcommercialloanmarketsandtheroleofmarketstructure: Evidence from surveys of commercial lending, Journal of Banking and Finance 15: 133{149. 26

Hauswald,R.&Marquez,R.(2002).Competitionandstrategicinformationacquisitionincredit markets. Mimeo. Kiser,E.K.(forthcoming). Modelingthewhole(cid:12)rm: Thee(cid:11)ectofmultipleinputsand(cid:12)nancial intermediation on bank deposit rates. FEDS Working Paper. Kwast, M. L., Starr-McCluer, M. & Wolken, J. D. (1997). Market de(cid:12)nition and the analysis of antitrust in banking, The Antitrust Bulletin . LeSage, J. P. (1998). Spatial econometrics. Mimeo available at http://www.spatialeconometrics.com. McMillen, D. P. (1992). Probit with spatial autocorrelation, Journal of Regional Science 32(3): 335{348. Park, K. & Pennacchi, G. (2003). Harming depositors and helping borrowers: The disparate impact of bank consolidation. Mimeo. Petersen, M. A. & Rajan, R. G. (2000). Does distance still matter? the information revolution in small business lending. NBER Working Paper No. 7685. Petersen, M. A. & Rajan, R. G. (2002). Does distance still matter? the information revolution in small business lending, Journal of Finance 57: 2533{2570. Scott, J. A. (2003). Credit, banks and small business { the new century. Mimeo. Wolken, J. & Rohde, D. (2002). Changes in the location of small businesses’ (cid:12)nancial services suppliers between 1993 and 1998. Federal Reserve Board Memo. 27

7 Appendix I II III IV I II III IV Spatial Er- Y N Y N Y N Y N rors Tract FE N N Y Y N N Y Y Bank FE Y Y Y Y Y Y Y Y Constant -0.169 -0.451 0.262 -0.050 -0.168 -0.440 0.268 -0.038 (-2.975) (-6.240) (1.768) (-0.309) (-2.891) (-6.118) (1.805) (-0.232) D1998 -0.192 -0.161 0.024 0.040 -0.202 -0.177 0.015 0.028 (-6.794) (-6.372) (0.749) (1.221) (-6.953) (-6.898) (0.452) (0.837) D1999 -0.217 -0.182 0.096 0.105 -0.237 -0.212 0.080 0.084 (-7.731) (-7.002) (2.740) (2.723) (-7.786) (-7.844) (2.182) (2.123) D2000 -0.307 -0.270 0.122 0.124 -0.341 -0.318 0.096 0.091 (-10.495) (-9.727) (2.992) (2.659) (-9.940) (-10.739) (2.186) (1.873) D2001 -0.290 -0.250 0.220 0.219 -0.332 -0.311 0.189 0.177 (-9.898) (-8.770) (4.951) (4.203) (-9.007) (-9.972) (3.845) (3.263) LnD -1.601 -1.457 -1.567 -1.425 -1.762 -1.631 -1.653 -1.516 (-61.493) (-53.763) (-61.226) (-51.659) (-46.957) (-37.773) (-45.829) (-34.994) LnD*M 0.910 0.879 0.669 0.636 1.145 1.114 0.803 0.759 (27.689) (28.104) (20.688) (19.689) (21.470) (20.417) (15.523) (13.511) LnD*L 1.437 1.204 0.958 0.742 1.610 1.404 1.053 0.851 (47.714) (43.127) (31.789) (24.722) (33.072) (28.838) (22.186) (16.795) LnD*T -0.048 -0.057 -0.055 -0.064 (-8.539) (-10.357) (-10.105) (-10.933) LnD*T*S 0.022 0.018 -0.018 -0.025 (1.709) (1.181) (-1.483) (-1.649) LnD*T*M -0.082 -0.085 -0.076 -0.078 (-5.809) (-7.404) (-5.741) (-6.520) LnD*T*L -0.054 -0.070 -0.059 -0.072 (-6.374) (-9.972) (-7.018) (-9.513) LnFirms 0.653 0.608 0.102 0.105 0.654 0.609 0.111 0.114 (74.902) (64.791) (2.467) (2.218) (74.912) (64.842) (2.674) (2.388) rho 0.500 0.439 0.499 0.439 Notes: T-statistics are in parentheses. As with McMillen (1992), the standard errors used to calculate the t-statisticsintheestimationswithspatially-dependenterrorsareconditionalon(cid:26)andsothatthet-statistics overestimate the unconditional t-statistics. Table 12: SDE Probit Estimates - Atlanta 28

I II III IV I II III IV Spatial Er- Y N Y N Y N Y N rors Tract FE N N Y Y N N Y Y Bank FE Y Y Y Y Y Y Y Y Constant -0.204 -0.691 -1.048 -1.436 -0.214 -0.686 -1.057 -1.427 (-3.563) (-20.151) (-7.976) (-8.341) (-3.744) (-19.837) (-8.022) (-8.215) D1998 0.047 0.037 0.046 0.038 0.033 0.025 0.033 0.027 (2.289) (1.657) (2.239) (1.591) (1.599) (1.112) (1.619) (1.128) D1999 0.019 0.013 0.074 0.063 -0.005 -0.009 0.052 0.043 (0.916) (0.571) (3.545) (2.534) (-0.257) (-0.413) (2.497) (1.729) D2000 -0.155 -0.145 -0.038 -0.041 -0.180 -0.169 -0.059 -0.061 (-7.584) (-6.124) (-1.705) (-1.436) (-8.527) (-7.120) (-2.593) (-2.154) D2001 -0.114 -0.116 0.077 0.058 -0.141 -0.142 0.056 0.037 (-5.475) (-4.766) (3.047) (1.771) (-6.407) (-5.834) (2.164) (1.128) LnD -0.764 -0.653 -1.025 -0.914 -0.745 -0.632 -0.995 -0.882 (-52.091) (-43.481) (-63.670) (-54.673) (-40.141) (-33.132) (-51.666) (-42.896) LnD*M 0.282 0.191 0.215 0.148 0.090 -0.004 0.0001 -0.067 (13.525) (10.994) (10.536) (8.081) (2.861) (-0.150) (0.004) (-2.241) LnD*L 0.585 0.446 0.450 0.341 0.720 0.590 0.577 0.475 (27.507) (23.137) (21.402) (16.293) (22.070) (19.768) (18.108) (14.904) LnD*T 0.001 -0.001 0.003 0.001 (0.135) (-0.260) (0.576) (0.158) LnD*T*S -0.008 -0.011 -0.011 -0.013 (-1.074) (-1.526) (-1.602) (-1.809) LnD*T*M 0.080 0.078 0.086 0.084 (8.640) (9.901) (9.635) (10.181) LnD*T*L -0.072 -0.078 -0.070 -0.075 (-8.095) (-9.480) (-8.024) (-8.646) LnFirms 0.792 0.727 0.193 0.183 0.795 0.729 0.183 0.174 (88.213) (65.750) (5.106) (3.739) (88.341) (65.895) (4.830) (3.571) rho 0.411 0.348 0.410 0.347 Notes: T-statistics are in parentheses. As with McMillen (1992), the standard errors used to calculate the t-statisticsintheestimationswithspatially-dependenterrorsareconditionalon(cid:26)andsothatthet-statistics overestimate the unconditional t-statistics. Table 13: SDE Probit Estimates - Denver 29

I II III IV I II III IV Spatial Er- Y N Y N Y N Y N rors Tract FE N N Y Y N N Y Y Bank FE Y Y Y Y Y Y Y Y Constant -0.019 -0.514 0.494 -0.040 -0.032 -0.533 0.475 -0.067 (-0.258) (-12.678) (2.548) (-0.248) (-0.432) (-13.074) (2.439) (-0.416) D1998 0.147 0.130 0.169 0.156 0.151 0.139 0.176 0.167 (3.796) (4.173) (4.542) (4.803) (3.829) (4.432) (4.680) (5.125) D1999 0.164 0.136 0.238 0.209 0.172 0.151 0.252 0.230 (4.220) (4.204) (6.253) (6.062) (4.259) (4.649) (6.468) (6.619) D2000 0.127 0.086 0.283 0.240 0.141 0.111 0.311 0.278 (3.216) (2.603) (6.993) (6.318) (3.287) (3.325) (7.218) (7.201) D2001 0.038 -0.005 0.220 0.174 0.055 0.026 0.255 0.222 (0.967) (-0.142) (5.275) (4.338) (1.203) (0.760) (5.471) (5.380) LnD -1.203 -1.071 -1.299 -1.165 -1.379 -1.219 -1.421 -1.256 (-30.124) (-27.468) (-32.157) (-28.587) (-19.856) (-13.957) (-20.190) (-14.138) LnD*M 0.505 0.427 0.385 0.321 0.781 0.680 0.627 0.535 (12.374) (11.062) (9.440) (7.956) (10.461) (7.457) (8.301) (5.747) LnD*L 0.779 0.582 0.599 0.429 0.890 0.643 0.630 0.408 (18.531) (15.493) (14.341) (10.716) (10.996) (7.004) (7.786) (4.335) LnD*T 0.044 0.042 0.055 0.053 (5.557) (6.028) (7.060) (7.204) LnD*T*S 0.107 0.095 0.099 0.086 (4.868) (3.316) (4.491) (2.954) LnD*T*M -0.009 -0.014 -0.010 -0.013 (-0.723) (-1.320) (-0.781) (-1.167) LnD*T*L 0.074 0.083 0.096 0.105 (5.757) (8.411) (7.639) (10.149) LnFirms 0.557 0.533 0.129 0.124 0.557 0.533 0.117 0.112 (53.723) (53.436) (3.143) (3.207) (53.725) (53.440) (2.858) (2.883) rho 0.357 0.317 0.356 0.313 Notes: T-statistics are in parentheses. As with McMillen (1992), the standard errors used to calculate the t-statisticsintheestimationswithspatially-dependenterrorsareconditionalon(cid:26)andsothatthet-statistics overestimate the unconditional t-statistics. Table 14: SDE Probit Estimates - Indianapolis 30

I II III IV I II III IV Spatial Er- Y N Y N Y N Y N rors Tract FE N N Y Y N N Y Y Bank FE Y Y Y Y Y Y Y Y Constant -1.587 -1.7655 -1.362 -1.562 -1.589 -1.766 -1.366 -1.566 (-32.799) (-27.200) (-8.338) (-10.055) (-32.823) (-27.206) (-8.360) (-10.075) D1998 0.076 0.061 0.084 0.070 0.079 0.062 0.088 0.073 (2.859) (2.361) (3.161) (2.574) (2.939) (2.417) (3.313) (2.690) D1999 0.028 0.022 0.126 0.110 0.032 0.025 0.133 0.115 (1.030) (0.833) (4.527) (3.781) (1.160) (0.922) (4.730) (3.934) D2000 -0.218 -0.198 0.030 0.026 -0.212 -0.194 0.040 0.032 (-7.748) (-7.012) (0.989) (0.742) (-7.351) (-6.819) (1.265) (0.934) D2001 -0.075 -0.077 0.210 0.182 -0.070 -0.074 0.220 0.188 (-2.592) (-2.584) (6.388) (4.780) (-2.278) (-2.451) (6.467) (4.908) LnD -0.788 -0.675 -1.020 -0.906 -0.764 -0.654 -0.981 -0.872 (-36.729) (-34.025) (-44.877) (-41.946) (-26.954) (-24.246) (-33.519) (-30.686) LnD*M 0.292 0.226 0.226 0.172 0.248 0.184 0.159 0.110 (11.432) (9.924) (8.926) (7.351) (6.059) (4.857) (3.906) (2.848) LnD*L 0.525 0.373 0.438 0.307 0.497 0.354 0.388 0.268 (19.487) (17.001) (16.281) (13.274) (11.788) (9.896) (9.252) (7.235) LnD*T -0.017 -0.018 0.002 0.001 (-2.854) (-3.002) (0.365) (0.174) LnD*T*S -0.029 -0.027 -0.016 -0.015 (-2.701) (-2.624) (-1.516) (-1.391) LnD*T*M -0.008 -0.007 0.015 0.014 (-0.769) (-0.734) (1.398) (1.365) LnD*T*L -0.015 -0.018 0.007 0.003 (-1.379) (-2.096) (0.659) (0.350) LnFirms 0.606 0.546 0.028 0.021 0.606 0.546 0.029 0.022 (69.725) (59.288) (0.780) (0.511) (69.728) (59.298) (0.810) (0.537) rho 0.372 0.344 0.372 0.344 Notes: T-statistics are in parentheses. As with McMillen (1992), the standard errors used to calculate the t-statisticsintheestimationswithspatially-dependenterrorsareconditionalon(cid:26)andsothatthet-statistics overestimate the unconditional t-statistics. Table 15: SDE Probit Estimates - Kansas City 31

I II III IV I II III IV Spatial Errors Y N Y N Y N Y N Tract FE N N Y Y N N Y Y Bank FE Y Y Y Y Y Y Y Y Constant 0.459 -0.026 0.871 0.421 0.557 0.053 0.935 0.468 (2.605) (-0.268) (3.319) (2.134) (2.962) (0.546) (3.519) (2.371) D1998 -0.112 -0.093 -0.025 0.001 -0.127 -0.110 -0.037 -0.013 (-1.500) (-2.007) (-0.356) (0.014) (-1.716) (-2.362) (-0.532) (-0.256) D1999 -0.092 -0.074 0.045 0.056 -0.090 -0.076 0.054 0.063 (-1.220) (-1.545) (0.606) (1.014) (-1.205) (-1.586) (0.727) (1.123) D2000 -0.075 -0.043 0.156 0.169 -0.037 -0.010 0.203 0.213 (-0.900) (-0.872) (1.840) (2.759) (-0.445) (-0.203) (2.355) (3.448) D2001 0.086 0.133 0.387 0.405 0.148 0.193 0.462 0.480 (1.011) (2.629) (4.300) (6.158) (1.714) (3.738) (4.953) (7.149) LnD -0.868 -0.852 -1.128 -1.132 -0.590 -0.605 -0.848 -0.882 (-12.382) (-14.625) (-16.509) (-17.838) (-5.402) (-7.725) (-8.042) (-10.643) LnD*M 0.450 0.404 0.280 0.240 0.355 0.321 0.158 0.137 (5.697) (6.161) (3.674) (3.438) (2.734) (3.440) (1.266) (1.401) LnD*L 0.782 0.724 0.587 0.535 0.384 0.360 0.188 0.168 (11.914) (12.889) (9.011) (8.850) (3.189) (4.219) (1.602) (1.875) LnD*T -0.063 -0.075 -0.042 -0.052 (-4.310) (-7.182) (-2.941) (-4.501) LnD*T*S -0.188 -0.186 -0.168 -0.165 (-5.295) (-7.042) (-4.937) (-6.010) LnD*T*M -0.187 -0.184 -0.150 -0.150 (-5.735) (-7.282) (-4.661) (-5.697) LnD*T*L -0.007 -0.020 0.014 0.004 (-0.384) (-1.513) (0.783) (0.311) LnFirms 0.481 0.491 0.025 0.049 0.480 0.491 -0.000 0.022 (24.908) (30.814) (0.329) (0.811) (24.898) (30.774) (-0.001) (0.352) rho 0.422 0.338 0.408 0.328 Notes: T-statistics are in parentheses. As with McMillen (1992), the standard errors used to calculate the t-statisticsintheestimationswithspatially-dependenterrorsareconditionalon(cid:26)andsothatthet-statistics overestimate the unconditional t-statistics. Table 16: SDE Probit Estimates - Nashville 32

I II III IV I II III IV Spatial Er- Y N Y N Y N Y N rors Tract FE N N Y Y N N Y Y Bank FE Y Y Y Y Y Y Y Y Constant -0.703 -1.015 0.283 -0.122 -0.698 -1.005 0.281 -0.117 (-13.187) (-19.236) (1.444) (-0.626) (-12.904) (-18.957) (1.430) (-0.598) D1998 -0.069 -0.050 -0.103 -0.077 -0.072 -0.054 -0.107 -0.083 (-1.883) (-1.311) (-2.877) (-1.902) (-1.961) (-1.431) (-2.980) (-2.046) D1999 -0.038 -0.025 -0.015 0.001 -0.044 -0.035 -0.022 -0.011 (-1.037) (-0.651) (-0.411) (0.018) (-1.156) (-0.902) (-0.592) (-0.262) D2000 -0.164 -0.148 -0.078 -0.063 -0.168 -0.158 -0.084 -0.076 (-4.309) (-3.766) (-1.904) (-1.312) (-4.189) (-3.989) (-1.985) (-1.558) D2001 0.081 0.081 0.251 0.247 0.078 0.069 0.246 0.232 (2.097) (2.086) (5.671) (4.721) (1.831) (1.751) (5.273) (4.386) LnD -0.820 -0.718 -1.089 -0.965 -0.754 -0.651 -1.044 -0.920 (-22.027) (-20.845) (-29.110) (-25.301) (-13.053) (-12.087) (-18.391) (-15.981) LnD*M -0.139 -0.122 -0.061 -0.053 -0.244 -0.240 -0.135 -0.142 (-3.657) (-3.552) (-1.642) (-1.439) (-3.746) (-3.874) (-2.120) (-2.180) LnD*L 0.352 0.180 0.230 0.097 0.304 0.152 0.201 0.090 (8.116) (4.895) (5.533) (2.402) (4.272) (2.393) (2.925) (1.338) LnD*T -0.036 -0.043 -0.039 -0.046 (-4.309) (-5.108) (-4.712) (-5.235) LnD*T*S -0.065 -0.072 -0.058 -0.065 (-3.049) (-3.546) (-2.840) (-3.124) LnD*T*M -0.018 -0.018 -0.025 -0.025 (-1.366) (-1.509) (-1.987) (-2.046) LnD*T*L -0.046 -0.062 -0.048 -0.065 (-3.311) (-4.630) (-3.495) (-4.610) LnFirms 0.491 0.449 0.077 0.075 0.491 0.450 0.081 0.080 (42.517) (34.726) (1.379) (1.167) (42.520) (34.728) (1.454) (1.231) rho 0.415 0.302 0.413 0.300 Notes: T-statistics are in parentheses. As with McMillen (1992), the standard errors used to calculate the t-statisticsintheestimationswithspatially-dependenterrorsareconditionalon(cid:26)andsothatthet-statistics overestimate the unconditional t-statistics. Table 17: SDE Probit Estimates - Providence 33

I II III IV I II III IV Spatial Er- Y N Y N Y N Y N rors Tract FE N N Y Y N N Y Y Bank FE Y Y Y Y Y Y Y Y Constant 0.093 -0.029 0.559 0.425 0.080 -0.043 0.532 0.399 (0.935) (-0.275) (2.552) (2.002) (0.779) (-0.401) (2.412) (1.876) D1998 -0.156 -0.152 0.110 0.100 -0.156 -0.151 0.118 0.108 (-3.104) (-3.950) (1.975) (2.115) (-3.030) (-3.867) (2.077) (2.238) D1999 -0.192 -0.197 0.199 0.176 -0.194 -0.195 0.214 0.192 (-3.657) (-4.956) (3.175) (3.131) (-3.337) (-4.591) (3.149) (3.220) D2000 -0.312 -0.310 0.227 0.205 -0.318 -0.309 0.247 0.227 (-5.873) (-7.343) (3.062) (2.976) (-4.840) (-6.358) (2.927) (3.011) D2001 -0.259 -0.266 0.316 0.284 -0.271 -0.268 0.337 0.309 (-4.879) (-6.307) (4.142) (3.969) (-3.709) (-5.061) (3.660) (3.787) LnD -1.535 -1.551 -1.673 -1.685 -1.626 -1.607 -1.752 -1.733 (-24.878) (-23.403) (-26.098) (-24.390) (-19.981) (-18.223) (-20.905) (-18.983) LnD*M 0.521 0.633 0.450 0.563 0.644 0.714 0.586 0.657 (6.962) (8.030) (5.726) (6.898) (6.810) (7.070) (5.981) (6.299) LnD*L 1.032 0.979 0.815 0.771 1.127 1.032 0.880 0.802 (16.031) (14.619) (11.799) (10.835) (11.161) (10.421) (8.475) (7.768) LnD*T -0.013 -0.017 -0.001 -0.004 (-1.239) (-1.861) (-0.078) (-0.367) LnD*T*S 0.030 0.010 0.037 0.020 (1.138) (0.336) (1.365) (0.671) LnD*T*M -0.033 -0.034 -0.038 -0.034 (-1.314) (-1.701) (-1.528) (-1.634) LnD*T*L -0.015 -0.016 0.007 0.005 (-0.858) (-1.173) (0.385) (0.341) LnFirms 0.767 0.747 -0.072 -0.057 0.766 0.746 -0.081 -0.066 (46.695) (43.959) (-0.823) (-0.655) (46.671) (43.925) (-0.925) (-0.759) rho 0.434 0.430 0.433 0.428 Notes: T-statistics are in parentheses. As with McMillen (1992), the standard errors used to calculate the t-statisticsintheestimationswithspatially-dependenterrorsareconditionalon(cid:26)andsothatthet-statistics overestimate the unconditional t-statistics. Table 18: SDE Probit Estimates - Rochester 34

I II III IV I II III IV Spatial Er- Y N Y N Y N Y N rors Tract FE N N Y Y N N Y Y Bank FE Y Y Y Y Y Y Y Y Constant 0.346 -0.158 1.147 0.814 0.357 -0.146 1.140 0.808 (3.213) (-2.909) (4.524) (3.375) (3.295) (-2.700) (4.498) (3.349) D1998 -0.123 -0.114 -0.026 -0.023 -0.124 -0.115 -0.028 -0.026 (-2.429) (-3.337) (-0.533) (-0.646) (-2.451) (-3.377) (-0.573) (-0.720) D1999 -0.442 -0.410 -0.236 -0.214 -0.449 -0.417 -0.243 -0.221 (-8.486) (-11.605) (-4.487) (-5.224) (-8.619) (-11.787) (-4.603) (-5.395) D2000 -0.413 -0.388 -0.051 -0.041 -0.422 -0.397 -0.060 -0.050 (-7.568) (-9.915) (-0.846) (-0.819) (-7.664) (-10.109) (-0.983) (-1.000) D2001 -0.398 -0.370 0.060 0.068 -0.408 -0.382 0.050 0.057 (-6.957) (-8.933) (0.917) (1.214) (-6.981) (-9.151) (0.744) (1.008) LnD -0.598 -0.574 -0.836 -0.811 -0.678 -0.637 -0.875 -0.835 (-10.864) (-12.169) (-14.876) (-15.812) (-8.562) (-8.632) (-10.812) (-10.676) LnD*M 0.250 0.231 0.181 0.166 0.275 0.232 0.178 0.138 (4.314) (4.744) (3.203) (3.313) (3.014) (2.817) (1.961) (1.612) LnD*L 0.418 0.341 0.273 0.208 0.547 0.459 0.347 0.274 (7.582) (7.260) (4.972) (4.279) (6.186) (5.771) (3.895) (3.298) LnD*T -0.026 -0.019 0.016 0.022 (-2.538) (-2.133) (1.516) (2.325) LnD*T*S 0.011 0.011 0.033 0.033 (0.391) (0.398) (1.196) (1.156) LnD*T*M 0.0001 0.011 0.035 0.046 (0.007) (0.788) (2.127) (3.191) LnD*T*L -0.052 -0.047 -0.002 0.001 (-3.479) (-3.913) (-0.163) (0.074) LnFirms 0.609 0.584 -0.006 -0.018 0.610 0.585 -0.001 -0.013 (46.482) (44.833) (-0.126) (-0.363) (46.521) (44.878) (-0.025) (-0.265) rho 0.311 0.313 0.310 0.312 Notes: T-statistics are in parentheses. As with McMillen (1992), the standard errors used to calculate the t-statisticsintheestimationswithspatially-dependenterrorsareconditionalon(cid:26)andsothatthet-statistics overestimate the unconditional t-statistics. Table 19: SDE Probit Estimates - San Antonio 35

I II III IV I II III IV Spatial Er- Y N Y N Y N Y N rors Tract FE N N Y Y N N Y Y Bank FE Y Y Y Y Y Y Y Y Constant -3.006 -2.996 -3.037 -3.042 -3.027 -3.015 -3.072 -3.075 (-24.125) (-7.617) (-14.028) (-6.049) (-23.605) (-7.454) (-13.815) (-5.919) D1998 -0.123 -0.115 0.034 0.030 -0.116 -0.109 0.039 0.032 (-4.730) (-4.464) (1.224) (1.016) (-4.415) (-4.226) (1.381) (1.090) D1999 -0.071 -0.072 0.230 0.205 -0.070 -0.073 0.228 0.199 (-2.722) (-2.767) (7.107) (5.860) (-2.603) (-2.779) (6.838) (5.584) D2000 -0.174 -0.168 0.164 0.144 -0.170 -0.167 0.165 0.140 (-6.692) (-6.197) (4.746) (3.701) (-6.044) (-5.992) (4.516) (3.508) D2001 -0.218 -0.216 0.177 0.149 -0.231 -0.229 0.159 0.128 (-8.141) (-7.727) (4.657) (3.443) (-7.763) (-7.896) (3.932) (2.883) LnD -1.010 -0.907 -1.128 -1.020 -1.078 -0.976 -1.183 -1.080 (-43.30) (-37.671) (-46.819) (-40.459) (-32.411) (-25.388) (-34.184) (-27.495) LnD*M 0.197 0.168 0.127 0.098 0.481 0.443 0.412 0.375 (7.770) (6.827) (4.941) (3.783) (10.683) (9.127) (8.949) (7.532) LnD*L 0.805 0.643 0.616 0.467 0.793 0.629 0.577 0.432 (30.936) (24.768) (22.812) (16.372) (18.430) (13.608) (12.964) (8.932) LnD*T -0.009 -0.008 0.001 0.001 (-1.567) (-1.290) (0.216) (0.111) LnD*T*S 0.018 0.019 0.023 0.024 (1.646) (1.448) (2.011) (1.773) LnD*T*M -0.102 -0.097 -0.099 -0.095 (-9.317) (-9.109) (-9.018) (-8.550) LnD*T*L 0.029 0.031 0.045 0.045 (3.391) (3.716) (5.194) (5.102) LnFirms 0.528 0.494 0.049 0.053 0.530 0.496 0.064 0.068 (65.759) (54.843) (1.470) (1.372) (65.887) (54.958) (1.922) (1.775) rho 0.345 0.320 0.343 0.318 Notes: T-statistics are in parentheses. As with McMillen (1992), the standard errors used to calculate the t-statisticsintheestimationswithspatially-dependenterrorsareconditionalon(cid:26)andsothatthet-statistics overestimate the unconditional t-statistics. Table 20: SDE Probit Estimates - Seattle 36