The Importance of Technology in Banking during a Crisis

Finance and Economics Discussion Series Federal Reserve Board, Washington, D.C. ISSN 1936-2854 (Print) ISSN 2767-3898 (Online) The Importance of Technology in Banking during a Crisis Nicola Pierri, Yannick Timmer 2022-020 Please cite this paper as: Pierri, Nicola, and Yannick Timmer (2022). “The Importance of Technology in Banking duringaCrisis,”FinanceandEconomicsDiscussionSeries2022-020. Washington: Boardof Governors of the Federal Reserve System, https://doi.org/10.17016/FEDS.2022. 2022020. NOTE: Staff working papers in the Finance and Economics Discussion Series (FEDS) are preliminary materials circulated to stimulate discussion and critical comment. The analysis and conclusions set forth are those of the authors and do not indicate concurrence by other members of the research staff or the Board of Governors. References in publications to the Finance and Economics Discussion Series (other than acknowledgement) should be cleared with the author(s) to protect the tentative character of these papers.

The Importance of Technology in Banking during a Crisis * NicolaPierri† YannickTimmer‡ ForthcomingintheJournalofMonetaryEconomics Abstract Whataretheimplicationsofinformationtechnology(IT)inbankingforfinancialstability? Data onUSbanks’ITequipmentandthebackgroundoftheirexecutivesrevealsthathigherpre-crisisIT adoptionledtofewernon-performingloansandmorelendingduringtheglobalfinancialcrisis.EmpiricalevidenceindicatesadirectroleofITadoptioninstrengtheningbankresilience;thisincludes instrumental variable estimates exploiting the historical location of technical schools. Loan-level analysisshowsthathigh-ITbanksoriginatedmortgageswithbetterperformance,indicatingbetter borrowerscreening. Noevidencepointstooffloadingoflow-qualityloans, differencesinbusiness models,orenhancedmonitoring. JELCodes:O3,G21,G14,E44,D82,D83 Keywords:Technology,FinancialStability,ITAdoption,Non-PerformingLoans,Screening *ThispaperisarevisedversionoftheIMFWP20/14withtitle"TechinFinbeforeFinTech:BlessingorCurseforFinancialStability?".Weare gratefultoUrbanJermann(editor),JosephVavra(theassociatededitor),andananynomousrefereeforinvaluablefeedbackthatsignificantly improvedthepaperduringtherevisionprocess. WethankTobiasAdrian,AndreaAjello,DavidAikman(discussant),UfukAkcigit,Tobias Berg,BarbaraBiasi,MarkusBrunnermeier,HansDegryse(discussant),GiovanniDell’Ariccia,EnricaDetragiache,DouglasDiamond,Ehsan Ebrahimy,AndreasFuster(discussant),AnnHarrison,PeterHoffmann,DenizIgan,DivyaKirti,SimoneLenzu,DavideMalacrino,AtifMian, SoleMartinezPeria,CyrilPouvelle(discussant),AndreaPresbitero,DamienPuy,LevRatnovski,GustavoSuarez,SuchananTambunlertchai, JamesVickery(discussant),GregorWeiss,TorstenWezel,BrianWolfe(discussant),andseminarandconferenceparticipantsatThirdBergen FinTechconference2020,EBAPolicyResearchWorkshop,BIS/BoE/CEPRConferenceonFinancialInnovation:ImplicationsforCompetition, RegulationandMonetaryPolicy,EFA,IWH-FIN-FIREWorkshoponChallengestoFinancialStability,CAFRWorkshoponFintech,CassBusiness School,ECB,Fed,IMF,VillanovaUniversity,andRESMFannualconferencefortheirinsightfulcomments.NicolaPierriisgratefultoNickBloom forguidanceandsupportduringtheearlystagesoftheproject.WethankChenxuFu,HalaMoussawi,andHuyNguyenforexcellentresearch assistanceandGladysChengandSylviePoirotforhelpwithdatasources.Allremainingerrorsareoursoleresponsibility.Theviewsexpressed inthepaperarethoseoftheauthorsanddonotnecessarilyrepresenttheviewsoftheIMF,itsExecutiveBoard,oritsManagementnorofthe FederalReserveBoardortheFederalReserveSystem. †InternationalMonetaryFund.1900PennsylvaniaAvenueNW,WashingtonDC.Email:npierri@imf.org ‡FederalReserveBoard.Email:yannick.timmer@frb.gov

1 Introduction Banks’massiveinvestmentsinInformationTechnology(IT),andthemorerecentworldwideemergence ofFinTech, havegeneratedadebateonthefinancialstabilityimpactofIT(Claessensetal.,2018;FSB, 2019; Carletti et al., 2020; Boot et al., 2021). FinTech and the latest technological developments have beenchangingthewayinformationisprocessedandtherelativeconsequencesforcreditallocationand performance(Bergetal.,2019;DiMaggioandYao,2021;Fusteretal.,2019,e.g). However,aparamount questionthatremainslargelyunaddressediswhetherITadoptionaffectsbanks’resilienceduringperiodsoffinancialstress. Tounderstandthepotentialimpactofhighertechnologyintensityinlendingonfinancialstability, we study the non-performing loans on the balance sheet of US banks with a heterogeneous degree of ITadoptionduringtheGlobalFinancialCrisis(GFC).ThesignoftherelationshipbetweenITadoption andnon-performingloansisa-prioriambiguous. Advancesintechnologycanimprovemonitoringand screeningthankstotheenhancedabilitytocollect,store,communicate,andprocessinformation(Liberti andPetersen,2018).However,bankswithmoreITadoptionmightrelytoomuchon“hard”information, whichareeasiertoreportandcommunicate,inducingthemtoneglect“soft”information(Rajan,2006; Rajanetal.,2015). WefindthatUScommercialbankswhichwereleadersinITadoptionbeforetheGFCweresignificantlymoreresilientduringthecrisis.Figure1illustratesthisstrikingpattern:high-andlow-ITadoption bankshadthesamelevelofNPLoverassetsbeforethecrisis,butassoonasthecrisishit,high-ITadoption banks experienced a significantly smaller increase of NPLs compared to their peers. Regression analysis reveals that a one standard deviation higher pre-GFC IT adoption is associated with 16 basis pointslowerNPLtoassetsratiointheyearsbetween2007and2010. Thisrepresentsa10%reduction with respect to the cross-sectional average and 14% of the cross-sectional standard deviation. In the panel dimension, there is no significant correlation between pre-crisis IT adoption of banks and their non-performingloansoutsidethecrisis. However,oncethecrisishit,higherITadoptionpredictsfewer NPLs: ourestimatesimplythatifbankshadaonestandarddeviationhigherITinvestmentbeforethe crisis,thesurgeinNPLscouldhavebeenlowerby15%withrespecttopre-crisislevels. The level of NPLs has widely been considered an important indicator for banking sector distress (Demirgüç-KuntandDetragiache,2002)andastrongincreaseisassociatedwithsevereadversemacroe- 2

conomicconsequences(PeekandRosengren,2000;Caballeroetal.,2008). ConsistentwithITadoption partially shielding banks’ ability to support the real economy, we find low IT banks tightened lending significantlymorethanhighITbanksduringandseveralyearsaftertheGFC. We then explore the channels through which IT can enhance banks’ resilience to the crisis. Using loan-leveldataonsecuritizedmortgages,wefindevidenceinfavorofimprovedborrowerscreeningby highITbanks.Instead,wefindnoevidenceforotherpotentialexplanationssuchasdifferencesinbusinessmodelorspecialization,offloadingoflower-qualityloans,orenhancedmonitoring. OurmainmeasureofITadoptioninbankingiscloselyrelatedtoseveralseminalpapersonITadoptionfornon-financialfirms, suchasBloometal.(2012), Beaudryetal.(2010), Bresnahanetal.(2002), andBrynjolfssonandHitt(2003). Weaccessdataonthenumberofpersonalcomputers(PCs)andthe numberofemployeesinabankbranch.Followingthispreviousliterature,weusetheratioofPCsperemployeewithinabranchastherelevantmeasureofbranch-levelITadoption. Astherevolutionarypower ofITstemsfromitsbeingamulti-purposetechnology,wefollowthispreviousliteratureandstudygeneral adoption of information technology rather than specific technologies (e.g. ATMs, online banking, orMortgageElectronicRegistrationSystems,asinHannanandMcDowell(1987),Hernández-Murilloet al.(2010),orLewellenandWilliams(2021)). Therefore,ouranalysisshedslightontheoveralleconomic impactofITadoption,ratherthanonspecificITapplications.Nonetheless,wetestthereliabilityofPCs asapredictoroftheuseofotherITtechnology.Weconfirmthatthereisastrongcorrelationbetweenthe shareofPCsperemployeeandothermeasuresofITadoption,suchasITbudgetoradoptionoffrontier technologiesin2016;thesealternativemeasuresareunfortunatelyunavailableinourdatabefore2007.1 Tothebestofourknowledge,thisisthefirstpapertousethistypeofdatatostudyfinancialfirms. Wedocumentthatthemostpowerfulpredictorofabranch’sITadoptionisthebankgroupitbelongs to.WemapthebankbranchestoBankHoldingCompanies(BHCs)andestimateaBHCfixedeffectafter controllingforthegeographyofthebranch(throughcountyfixedeffects)andothercharacteristics,such asthesizeofthebranch.ThesefixedeffectsserveasourmeasureofITadoptionontheleveloftheBHC ,whichwealsorefertoasbankinterchangeablyinthetext. Pre-crisistechnologyadoptionmaybecorrelatedwithotherbankcharacteristicswhichimpactnonperformingloansduringthecrisis.Mostimportantly,demandforITequipmentanditsproductivityhas 1Infact,laterwavesofthesamedatasetprovideadditionalinformationonIT-budgetandadoptionofCloudComputingat theestablishmentlevel: thenumberofPCsperemployeeisastrongpredictoroftheseothermeasuresofITadoption. For example,thecorrelationbetweenthepercapitashareofPCsandtheITbudgetis65%onthebranchleveldata. 3

beenassociatedwithfirms’organizationalformsandmanagerialquality(Bresnahanetal.,2002;Bloom etal.,2012).Thus,themainconcernforacausalinterpretationofourresultsisthatITintensebankswere simply“betterrun”andsuperiormanagementpracticesshieldedthemfromtheimpactofthecrisis. OuranalysisuncoversseveralpiecesofevidenceinfavorofadirectroleofITadoption,thusmitigatingtheconcernthatthecorrelationbetweenITandbankresilienceisdrivenbyunobservablefactors, such as management quality. First, we find that our measure of IT adoption–which is purged of local variationandbranchcharacteristics–isnotsignificantlycorrelatedwithbanks’ex-anteexposuretothe GFCintermsoftheirgeographicalfootprintorbusinessmodelasmeasuredbyfundingsources,assets composition,andotherbalancesheetcharacteristics. OurmeasureofITisalsouncorrelatedwithemployees’averagewagesorexecutives’compensation,whichcanbethoughtofasmeasuresofworkforce humancapital(Becker,2009).TheabsenceofacorrelationbetweenITadoptionandalltheseobservable characteristicsisausefulfalsificationtest: itsuggeststhatourmeasureisunlikelytobecorrelatedwith otherunobservablepredictorsofexposuretothecrisis. Second,wefindthattheestimatedimpactofITonNPLsisunaffectedbytheinclusionofarichsetof variablesascontrols.Exploitingthiscoefficient’sstability,wefollowAltonjietal.(2005)andOster(2019) toprovideformaltestingforthepresenceofbiasfromunobservablebank-levelcharacteristics,finding noevidenceofasizeablebias. Third,wecomplementouranalysiswithasetofinstrumentalvariable(IV)specificationsbasedon the distance between a bank headquarter and land-grant colleges and universities. These institutions wereestablishedattheendofthenineteenthcenturyinallUSstatestoprovidetechnicaleducation.We showthattheirstudentsaresignificantlymorelikelytomajorintechnicalsubjectsandlesslikelytomajorinbusinessandmanagementsciences,suggestingthatthesecollegesareashifteroftheavailability oftechnicalknowledgeratherthanmanagerialcapabilities. Inaddition,thelocationofland-grantcollegesdoesnotpredictthepresenceofBHCheadquartersinacounty, indicatingthedistancebetween locationsisindependentwithrespecttothemostrelevantfactorsimpactingthebankingbusiness. We thenshowthatbankswhoseheadquartersareclosertothesecollegeshavegenerallyahigherlevelofIT adoption, supportingtheideathattechnicalknowledgeisanimportantfactorinfosteringtechnology adoption. However,theexplanatorypowerofthedistancebetweenabank’sheadquarterandtheclosest land-grant colleges is low. We thus rely on weak instrument techniques which provide confidence intervalsforcausaleffectsinabsenceofastronginstrument(AndrewsandStock,2018). Theestimated 4

intervalsdonotcontainzero,rejectingthenullofnocausalimpactofITonNPLs,andconfirmingthe mainfindingofthepaper. Fourth,toshedfurtherlightontheroleofITversusmanagerialquality,westudythebiographiesof thebanks’topmanagement. Infact, thepersonalcharacteristicsandexperienceofleadersmatterfor theoutcomesoftheirorganizations(BenmelechandFrydman,2015). Weapplyasimpletext-analysis algorithm to the biographies of top executives hired before 2007. We search for specific tech-related keywordsandusethemtomeasurethemanagers’predispositiontowardIT.Wefindthatbanksledby more“tech-oriented”executivesadoptedITmoreintensivelyandwerealsomoreresilientinthecrisis. Interestingly,whenweestimatetheimpactoftech-savvinessofexecutivesovertime,wefindastrikingly similarpatterncomparedtotheoneestimatedwiththebaselinemeasureofITadoption: banksrunby tech-orientedexecutiveshadstatisticallyindistinguishablelevelsofNPLscomparedtotheirpeersinany yearbeforethecrisis;however,oncethecrisishit,theirNPLsincreasedsignificantlylessthanbanksled byexecutiveswithnotech-background. Theresultsholdaftercontrollingforboththeircompensation andothercharacteristics,suchastenureorpost-tertiaryeducation.Aslongasexecutive“overall”quality is,atleastpartially,pricedintheircompensation,theseresultssuggestthatitis“techorientation”that mattersandnotmanagerialtalent. ThesefindingssupportthehypothesisthatITadoptioninbanking, whichcanbepartlycausedbyexecutives’personalexperienceandinclinations,ledtomoreresilience duringthecrisis. Adopting a “weight of evidence” approach, this collection of results points toward IT itself as the causeoflowerNPLsratherthanaspuriouscorrelationbetweenthetwovariablescreatedbyunobserved bankcharacteristics,suchasmanagerialquality. TurningtothemechanismthroughwhichhigherITadoptionincreasesbanks’resiliencewetestfor fourpossiblechannels:(i)adaptationofthebusinessmodel,(ii)offloading,(iii)monitoring,(iv)screening. ITadoptionmayimpactNPLsbychangingbanks’businessmodel,forinstance,ifitincreasesbanks’ focusonmarketsegmentsthatarelessaffectedbyfinancialcrises.Wefindnoevidenceforsuchamechanism. We find instead that high IT adopters experienced lower NPLs during the GFC for any of the majorloancategories(C&I,CommercialRealEstate,ResidentialRealEstate).Consistent,wefindnocorrelationbetweenoriginators’ITadoptionandborrowers’creditscoreintheGSEdata: highITadoption banksdonotsimplyfocusonlessriskymarketsegments. 5

To test for the other potential channels, we analyze the performance of mortgages originated before 2007 and sold to Freddie Mac and Fannie Mae, the two large government-sponsored enterprises (GSEs). Wefindthatmortgagessoldbyhigh-ITadoptionbanksweresignificantlylesslikelytobecome delinquentduringtheGFCthantheonessoldbyotherbanks,rulingoutthathigh-ITbankssimplyoffloaded worse loans and therefore had fewer NPLs on their balance sheet. This also implies that the betterperformanceofhigh-ITadoptersduringthecrisisisdriven–atleastinpart–bytheoriginationof betterloans,leavingscreeningormonitoringaspotentialexplanations. However,wefindnoevidence thattheITadoptionofthemortgageservicerisassociatedwithlowerdelinquencyrateorhighermodificationprobabilityofdelinquentmortgages,pointingtowardbetterscreeningratherthanmonitoringas thechannelthroughwhichITadoptionincreasesbankresilience. Thisresulthasimportantimplicationsforfinancialstability. Ifhigh-ITadopterswereonlybetterat offloading their bad loans to GSEs, then IT intensity would lead to risk-shifting and exacerbate moral hazard,ratherthanenhancefinancialstability.Thisisanimportantconcernbecausesecuritizationmay reduce the incentives of banks to screen and monitor borrowers (Keys et al., 2009, 2010, 2012) and IT adoptionmayfacilitatesecuritization. Arelatedconcernisthosehigh-riskindividuals, whichwererejectedbytechnologyadopters,borrowedfrombankswithlessIToperatinginthesamearea.Wetestfor these spillover effects and find no evidence, either. Both of these results suggest that IT adoption had positiveaggregateeffectsonthestabilityofthefinancialsystemandwasnotassociatedwithatransfer ofriskacrossparties. WefurtherhighlightthefinancialstabilityimplicationsofITinbankingbyprovidingevidencethat technologyadoptioninbankingdoesnotonlylowertheNPLsonthebalancesheetofthebankbutalso increasestheirlendingvolumesduringthecrisis. Thesepositiverealeffectssuggestthattherecentrise inFinTechcouldbebeneficialforfinancialstabilitythroughbetterscreeningabilities. RelatedLiterature This paper is related to the finance literature on technology adoption, which has beenthrivinginrecentyearsthankstothesurgeofFinTech(Fusteretal.,2018,2019;Bergetal.,2019; DiMaggioandYao,2021;Buchaketal.,2018,e.g.).WecontributebyevaluatingtheimpactofITadoption inlendingonfinancialstabilityandbystudyingtheimpactoftechnologyacrossasamplethatcoversthe majorityofUSbanklending. ClosetousareafewpapersthatanalyzecertainfeaturesofITadoptioninbankingduringnormal 6

times (Hannan and McDowell, 1987; Berger, 2003; Bofondi and Lotti, 2006; Hernández-Murillo et al., 2010; Bostandzic and Weiss, 2019; D’Andrea and Limodio, 2019; He et al., 2021). We contribute by focusingontheeffectofoverallITadoptionacrossbanksontheirperformancewhenasystem-wideshock hits.Morerecently,afewstudieshavefollowedourapproachtostudytheroleofITduringtheCOVID-19 pandemic and found that banks adopting IT more intensely performed better even in a crisis of nonfinancialnature(Branzolietal.,2021;Kwanetal.,2021;Dadoukisetal.,2021)andthatfirms’ITfosters theresilienceofthelocallabormarkets(PierriandTimmer,2020). WealsoprovidedirectevidencethatITadoptionamelioratesthescreeningofborrowers,consistent withtheliteratureoninformationinlendingwhicharguesthatadvancesinITimprovetheprocessingof informationbyhelpingfirmstogather,store,distribute,andanalyzeinformation(LibertiandPetersen, 2018;PetersenandRajan,2002;DegryseandOngena,2005;PetersenandRajan,1994). Therestofthepaperisstructuredasfollows. Insection2wedescribetheseveraldatabasesused;in section3wepresentthemainresultsonITadoptionandNPLs;insection4weprovideevidenceonthe rootsofITadoptionandproposeaninstrumentalvariablestrategy; insection5wepresentadditional resultsonmortgagesperformancetoshedlightonpotentialmechanisms;insection6weconclude. 2 DataandMeasurement ITAdoption The IT data comes from an establishment survey on personal computers per employee by CiTBDs Aberdeen (previously known as “Harte Hanks”) for years 1999, 2003, 2004, 2006, and 2016. Fortheyear2016,wealsohaveinformationontheITbudgetandtheusageofcloudcomputingofthe establishment.Thedataalsocontainsinformationaboutthetypeofestablishment–i.e.whetheritisthe headquarter(HQ),abranch,oraback-endoffice–thenumberofemployeesintheestablishmentaswell as the location. The correlation between the IT budget of the establishment and the number of computers as a share of employees is very strong for later years, e.g. 65% in 2016. There is also a positive correlationbetweenPCsperEmployeeandtheadoptionofcloudcomputing. ThesecorrelationsprovideassurancethatthenumberofpersonalcomputersperemployeeisagoodmeasureofITadoption, evenmorerecently, butlikelyevenmoresoinearlieryearswhenotherformsofITadoptionwereless common. (Asafurthercheck,weuseOECDdataonbusinesses’useoftechnologiesindifferentcountries,sizecategories,andindustries,andseeastrongcorrelationbetweentheuseofPCsbyemployees 7

andvariousotherITmeasures,suchastheuseofbroadbandconnections,BigDataanalytic,orCloud Computing,asgraphicallyillustratedbyFigureA1.) Wefocusonlyonestablishmentsinthebankingsector(basedonSIC2classification)anddropsavingsinstitutionsandcreditunions(basedonSIC3). Afterthesecleaningsteps,weendupwith143,607 establishment-yearobservations.WemapbankbranchesfromtheAberdeendatasettotheBHCdataby usingbanks’namesandtheBHCstructure. OurmeasureofITadoptionisbasedonaregressionoftheshareofpersonalcomputersonabank fixedeffectcontrollingforthegeographyoftheestablishmentandothercharacteristics. Bydoingsowe can control for several characteristics that may be correlated with the number of personal computers peremployeeofthebankbutarenotinformativeaboutwhetherthebankhasbeenatthetechnological frontier. ThisapproachfollowsBeaudryetal.(2010),whomeasureITadoptionontheregion-levelcontrollingforestablishments’industryandsize. Weestimatethefollowingregressionfortheyears1999, 2003,2004,and2006: PCs/Emp i,t =I(cid:103)T b +θ type +θ c +θ t +γ·Emp+(cid:178) i,t (1) wherePCs/Emp istheratioofcomputersperemployeeinbranchi surveywavet (cappedattop i,t 1%), I(cid:103)T b isabank(i.e. BHC)fixedeffect,θ type isaestablishment-type(HQ,standalone,branch)fixed effects,θ isacountyfixedeffect,θ isayearfixedeffectandEmpisthelognumberofemployeesinthe c t establishment. TheR-squaredoftheregressionis42%. Themainpartoftheexplainedvariationiscapturedbythe bankfixedeffect(60%).Thelocationoftheestablishmentexplains27%ofthevariation,yearfixedeffect explains11%,whilenumberofemployees’andestablishmenttypeshavesmallerexplanatorypower. Our measure of IT adoption, IT hereafter, is a standardized version of the bank fixed effect. It is b obtainedbydividing I(cid:103)T b byitsstandarddeviationaftersubtractingitsmean. Thisadjustmentisdone considering the summary statistics for the sample of banks that we are able to match with BHC data only.ThebottompanelofFigureA2plotsthecross-sectionaldistributionofIT .ForeachBHC,wealso b constructstheITadoptionofotherbanksoperatinginthesamelocation. 8

RegulatoryDataonBHCs Weusebankbalancesheetinformationfrombankholdingcompanies(BHCs) toassesstheresilienceofbankstotheGFC.ThedataiscollectedbytheFederalReserveBankofChicago. We use the Financial Institution Reports which provides consolidated balance sheet information and incomestatementsfordomesticBHCs. OurmaindependentvariableistheamountofNPLsscaledbytotalassets.2 Wechecktherobustness of the main results of the paper to other definitions of NPLs (e.g. including loans with shorter delinquencyperiods)andalternativescalingchoices(e.g. theuseofloansasdenominator), seesection3.3 FigureA2showsthedistributionoftheaverageNPLsratiobetween2007and2010acrossbanks. Most bankshaveanNPLratioofaround1%inthecrisisperiod,butthereisalongrighttailinthedistribution. Forsomebanksalmost5%oftheirbalancesheetconsistsofNPLs. InadditiontoNPLsweconstructthefollowingvariablesasbank-levelcontrols. Theshareofloans overtotalassets(Loans),thelogofassets(inthousandsofUSDollars)(Size),equityoverassets(Capital), wholesalefundingoverassets(Wholesale),thereturnonassets(ROA),andtheaveragelogwagepaid to employees (in thousands of US Dollars) (LogWage). All variables are averaged between the years 2001and2006.Wewinsorizeallbank-levelratioattop2.5percentbeforetakingaverages,butresultsare robusttodifferenttreatmentofoutliers. OursampleiscomposedofBHCwhichwecanmergebynamewiththeITAberdeendataandsuch thatallthecontrolsvariabledescribedabovearenon-missing. Thesamplecoversabout80%ofallpre- GCF bank lending, andevenmoreforeC&Iand Residential Real Estate loans (see Figure A3). We also match77outofthe100largestBHCintermsof2006loans. BiographiesofExecutives WeobtaindataonthebiographyofexecutivesfromS&PGlobalMarketIntelligence. We have information on the Chief Executive Officer, the Chief Financial Officer, the Chief OperatingOfficer, andthePresidentofthebank. WefocusontheexecutivesthathavebeenhiredbeforetheGFC.Wesearchthebiographyforthefollowingwordstocharacterizewhetheranexecutiveis tech-prone: technology, engineering, math, computer, machine, system, analytic, technique, method, 2OurbaselineNPLsaredefinedfollowingHirtleetal.(2018): Totalloans,leasingfinancingreceivablesanddebtsecurities andotherassets-pastdue90daysormoreandstillaccruing(bhck5525)+Totalloans,leasingfinancingreceivablesanddebt securitiesandotherassets-nonaccrual(bhck5526)-Debtsecuritiesandotherassets-pastdue90daysormoreandstill accruing(bhck3506)-Debtsecuritiesandotherassets-nonaccrual(bhck3507). 3WerelyonassetsasascalingvariableforNPLs(ratherthanloans)sincelendingisendogenoustoITadoptionandNPLs duringthecrisis,aswedocumentinsubsection3.4.Moreover,assetsarecommonlyusedtonormalizethebank-levelvariables. ThemainrawpatternsandresultsarerobusttousingloansinsteadofassetstonormalizeNPLs. 9

process,stem,efficiency,efficient,software,hardware,data,informatic.Wecountthenumberofoccurrencesofthesewordsforeachexecutiveinthebiographyandscalethenumberbythetotalnumberof wordsinthebiography. Foreachbank,wetaketheaverageacrossexecutivestoconstructabank-level measureoftheITintensityoftheirexecutives.Inadditiontothebiography,wealsousedataonthetotal compensationoftheexecutivesfromtheStandard&Poor’sExecutiveCompensationdatabase. House Price, Land-Grant, and County-level Data We compute BHCs’ exposure to the downturn in houseprices,HP Exposure.Weobtaincounty-levelhomevalueindexfromZillow.Foreachcountywe constructthe(oppositeof)percentagechangeinhousepricesfrom2007Q4to2012Q3. ForeachBHC wecomputetheweightedaveragedecreaseinhousepricewheretheweightsarethelocalfootprintof theBHCacrosscounties,asmeasuredbydepositsintheFDICdataset.4 Weobtainthelistofallthe70land-grantcolleges(anduniversities)establishedintheUSstatesduringthenineteenthcentury(1862and1890)fromthewebsiteoftheUSDepartmentofAgriculture. We obtain data on enrolment by major and test scores from IPEDS (Integrated Postsecondary Education DataSystem)surveyfor1996and2018forseveralhighereducationinstitutions. Weobtaincounty-level demographiccharacteristicsfromthe2000USCensusandfromtheAmericanCommunitySurveyfrom 2001 to 2006. We obtain information on employment by industry in each county from the Quarterly WorkforceIndicators(wesetthesevariablesequaltonationalaverageforcountiesoutsidethedataset, e.g.inPuertoRico).County-levelvariablesareaveragedbetween2000and2006. GSEData AsHurstetal.(2016)wepooldatafromFreddieMac’sSingleFamilyLoan-LevelDatasetand FannieMae’sSingleFamilyLoanPerformanceDataset.Theseloan-leveldatasetscoverstheperformance onmortgagesthatthetwoGSEsboughtstartingin1999. Thedataincludeshigher-qualityloanswhich hadtoconformtoagencyguidelines(Adelinoetal.,2016). 3 ITadoption,NPLs,andLending Inthissection,weinvestigatetherelationshipbetweenbanks’ITadoptionbeforetheGFCandtheirNPLs duringandoutsidethecrisis. Asapreview, Figure1showstheevolutionoftheratioofNPLstoassets 4Wetaketheaveragedepositsacrossthepre-crisisyears.ForthreeBHC,whichwedonotmergetotheFDICdataset,weuse insteadtheITestablishmentdatatocomputethelocalfootprint. 10

from1996to2014forbanksinthebottomandtopquartileoftheITadoptiondistribution.Thisrawdata showsthatthetwoseriesarevirtuallyindistinguishableuntil2007. However,in2008–asNPLsstartto surge–thetwolinesdiverge. ThegrowthinNPLsisconsiderablymorepronouncedforbankswithlow ITadoption.TheNPLspeakin2010andthetwoseriesstartconvergingagainfrom2011.5 3.1 Panel Inoursample,thesharpriseofNPLsoverassetsoccurredintheyearsfrom2007to2010. Therefore,we definetheseyearsasthe“crisis”period.ToinvestigatewhetherbankswithdifferentlevelsofITadoption experienceddifferentlevelsofNPLsduringthisperiod,werelyonthefollowingpanelequation: NPL =α +δ +βIT ·crisis +(X ·crisis ) (cid:48)γ+(cid:178) (2) b,t b t b t b t b,t where NPL is the share of non-performing loans relative to assets for bank (BHC) b in year t. b,t IT is our bank-level measure of IT adoption before the crisis as defined in section 2, α and δ are b b t bankandyearfixedeffects,respectively.Theformercapturebanktime-invariantheterogeneitywhilethe lattercapturetime-varyingaggregateshocks,suchasbusinesscyclefluctuations. X isavectorofbankb levelvariables(pre-crisisaverages)thatmaybeassociatedwithNPLsduringthecrisis,asdescribedin section2.Weincludeobservationsforyearsbetween2001and2014andkeeponlyobservationforwhich wehaveallthevariablesinX .Weareleftwith4,608observationson337banks. b Table1presentstheresultsfromestimatingdifferentversionsofEquation2viaOLS,togetherwith standarderrordouble-clusteredatthebankandyearlevel. Wefirstpresentalesssaturatedversionof the above equation. Column (1) shows the result of Table 1 without the inclusion of bank fixed and yearfixedeffectsaswellaswithoutcontrols.Thebaseeffectoftechnologyadoptiononnon-performing loansisnegativebutsmallandnotstatisticallysignificant. WedonotfindthatITadoptionsignificantly affectsnon-performingloansduringnormaltimes. However,theinteractionbetweenthecrisisdummy and IT adoption is negative and statistically significant. In the time of the crisis banks that adopted moreITbeforethecrisishadasignificantlylowershareofnon-performingloansthanbankswithless ITadoption. Thisresultisrobusttotheinclusionofbankandyearfixedeffectsandvariouscontrols. In 5ThedynamicsofNPLsforabankwithintermediateadoptionliesbetweenthelow-andhigh-adoptersinmostyears,see FigureA4. Tocheckthatthispatternismainlydrivenbythenumeratoroftheseries(NPLs)ratherthanthedenominator (Assets),wefixthevalueofassetstothebank-specificpre-crisisaverageandplottheadjustedratioinFigureA5,findingavery similarpattern. 11

addition,thecoefficientisstableacrossspecifications,suggestingalowcorrelationbetweenthecontrols included and the measure of IT adoption. A one standard deviation higher IT adoption is associated with a between 13 and 17 basis points lower NPL share increase during the crisis. The average share of NPLs was 1.5 percent in the crisis period, while its standard deviation was 1.13. Therefore, a one standarddeviationhigherITadoptionledtoareductioninNPLsbetween9and11%withrespecttothe mean and between 12 and 15% with respect to the cross-sectional standard deviation. Moreover, the increasebetweenthepre-crisisaverageandthecrisisNPLshareis1.05percentagepoints. Therefore,if we ignore potential heterogeneity in the effect of IT adoption, spillover between banks (which we test for,seebelow),andgeneralequilibriumeffects,wefindthataonestandarddeviationuniformincrease inITadoptionacrossallbankswouldhavediminishedthesurgeinNPLsbetween12and16%. Columns (5) introduces additional controls to the baseline specification with only bank and year fixedeffects, leavingourresultsunchanged. Forbrevity, weonlydisplaytheIToflocalcompetitorsas a control variable to shed light on whether there are negative spillover effects of IT adoption on other banks.Individualswhowanttoborrowbutarerejectedbyahigh-ITbankcouldapplyforaloanatalow- ITbankinthesamearea. Ifthelow-ITbankdoesnotidentifytheborrowerasrisky,thebankmaygrant aloan,whichdefaultsduringthecrisisandleadstoanincreaseinNPLsforthisbank.Ifthismechanism isatwork,wewouldstillfindasignificantdifferencebetweenhigh-andlow-ITbanksintermsoftheir NPLsduringthecrisis,buttheaggregateincreaseinNPLswouldbethesameifallbanksadoptedmore ITwithambiguousimplicationsforfinancialstability.However,theinteractionshowsthatbanks,which arebasedinareaswheretheircompetitorsadoptedmoreIT,didnotsufferastrongerincreaseintheir NPLsrelativetootherbanks. ThisevidencesuggeststhatITadoptiondoesnothavenegativespillover effectsonlocalcompetitors. Finally, abatteryofrobustnesstestsarereportedinTableA1, confirming ourresults.6 Next,weallowtheimpactofpre-crisisITadoptiononNPLstovaryeachyearbetween1996and2014 byyearbyestimatingthefollowingequation: (cid:88) NPL b,t =α b +δ t + β τ ·IT b ·1[t=τ]+(cid:178) b,t (3) τ(cid:54)=2006 6Forinstance,inarobustnessexerciseweuseaBHC-levelbootstrapproceduretore-estimatethestandarderrors,finding verysimilart-statisticsasthoseinTable1(seecolumn10ofTableA1).Becausethisprocedurere-estimatestheITmeasurefor eachofthebootstrapsamples,itaddressestheconcernthatourstandarderrorssufferfroma“generatedregressor”problem. 12

Thecoefficientof2006isnormalizedtozero.ResultsareillustratedinFigure2.TheeffectofITadoptiononNPLsisinsignificantinthepre-crisisperiodbetween1996and2007,exceptforasmallnegative effectwhichisstatisticallysignificantatthe5%levelin2002, likelyduetotheearly2000recession. As showninTable1bankswhichadoptedmoreITbeforethecrisishadsignificantlylowerNPLsthantheir counterpartsinthecrisis. Inparticular, between2007until2010theeffectisnegativeandstatistically significantatthe5%level,andin2009and2010theeffectisevenstatisticallysignificantatthe1%level. Thecoefficientreachesitsmaximumin2010with-0.3. Inotherwords,aonestandarddeviationhigher ITadoptionwasassociatedwith30basispointslowerNPLsin2010. Theimpactisstillnegativein2011 and2012althoughnotstatisticallysignificantanymore. Wedetectnoimpactinthetwolatestyearsof thesample,2013and2014. 3.2 Cross-sectionalanalysis Inthissection,weanalyzetherelationshipbetweenbank-levelITadoptionandvariousotherbankcharacteristicsinthecross-section. WeapplyOLStothefollowingequation: Y =α+β·IT +(cid:178) (4) b b b where Y is either the share of NPLs over assets in the crisis period or one of the control variables b inthesetX describedaboveandtheindependentvariableisthepre-crisisITadoption. Collapsingthe b datainthisavoidstheunderestimationofstandarderrorsthatcanarisewhenestimatingadiff-in-diff specificationusingpaneldata(Bertrandetal.,2004). Table2presentstheresults. Consistentlywiththepanel(Table1andFigure2)technologyadoption is strongly negatively correlated with NPLs in the crisis period (column 1) with an R-squared of 2.6%. Themagnitudeofthecoefficient(18basispoints)isslightlyhigherbutsimilartotheoneestimatedin thepanelregressions. Columns(2)-(8)testwhetherITadoptiononthebankleveliscorrelatedwithotherbank-levelvariables– describedinsection2–thatcouldbeimportantindrivingNPLsinthepost-crisisperiod. WefindthatIT adoptionisnotsignificantlyrelatedtoanyofthesecharacteristics.7 Moreover,theR-squaredofcolumn 7Wecomputeadditionalvariables,suchastheshareorresidentialorpersonalloansoverthetotalamountofloans,andfind 13

(1) is much larger (at least 4 times) than the ones of columns (2)-(8). We, therefore, conclude that IT adoptionisnotcorrelatedwithanyimportantbank-levelcharacteristicsthatcouldpredicttheirexposuretotheGFC.Thisresultisacomforting“balancing”testsinceitsuggestsITisunlikelytobecorrelated tootherunobservablecharacteristicsthatwouldalsomakethemmoreexposedtothefinancialshocks andrelatedrecession. The(counterintuitive)lackofcorrelationbetweenITadoptionandpre-crisisROAisinlinewithpreviousliteraturewhichdocumentsweakproductivitygainsfromITinbankinginnormaltimes(Beccalli, 2007), andanegativeornullcorrelationbetweenprofitabilityandadoptionofATMoronlinebanking (HannanandMcDowell,1984;Hernández-Murilloetal.,2010). Wedonotfindevidencethatbanksize isadeterminantofbranch-levelITadoption.Kovneretal.(2014)documentsizeableeconomiesofscale (also)forITexpenditure.Thismaybeduetothefactthatlargerbuyerscanbargainforlowerprices.Our measure,however,capturesITequipment,notITexpenditure. Whileexpenditureshavetheadvantage toconvey,throughprices,someinformationonthequality(andnovelty)ofITpurchases,lookingonlyat thequantityofPCsavoidsthebiasthatmaybecausedbytheheterogeneouspurchasingpowerrelated tobanksize.Otherstudiesfoundthatthetimingoftheadoptionofaspecifictechnologyisusuallyfaster forlargerbanks(HannanandMcDowell,1984;Hernández-Murilloetal.,2010). Thosestudies,however, focusonverycoarseadoptionvariables(haveanyATMornot,haveawebsiteornot)andaretherefore pronetorewardlargerbanks;ourITvariable,instead,aimstocapturehowwidespreadisthegeneraluse ofcomputationalsystemswithinabankbranchesnetworkbymeasuringtheavailabilityofequipment peremployee. It must also be noticed that our measure of IT adoption is purged by local variation (and branch characteristics)toeliminatepotentialconfoundingfactorsclusteredatthelocallevelanddonotcomparebanksoperatinginverydifferentpartsofthecountry(ourmainresultsarerobustnottoperform suchadjustments,seeTableA1). Thus,thelackofcorrelationwithothercharacteristicsdoesnotnecessarilyimplyalackofcorrelationwiththerowproxiesofITadoption. Incolumn(10)weincludeallthebankcharacteristicsascontrolsinaregressionofpost-crisisNPLs on IT adoption. We find a very small change in the coefficient of IT adoption, despite the seven-fold increase in R-square. This suggests that the equation of column (1) does not suffer from an omitted nocorrelationofthesevariableswithITadoptioneither. Additionally,insection5wepresentdirectevidencethattheimpact ofITonNPLsisnotdrivenbythelocationoflendingactivities. 14

variablebiasandpointstowardsacausalrelationshipbetweenITadoptionandNPLs(Altonjietal.,2005; Oster,2019),whichwealsotestformallyfollowingOster(2019)andfindthatourresultsarerobusttothe presenceofunobservablevariables.8 column(9),instead,showsapositivecorrelationbetweenaBHC’s ITadoptionandIToflocalcompetitors. 3.3 LoanCategories Which loan categories are responsible for the lower NPLs during the crisis? For most of BHC in our sample,wecancomputeNPLsacrossthreedifferentloancategories:residentialrealestate,commercial realestate,andC&Ilending.AsreportedbyTable3,highITadoptionbanksreportedlowerNPLsontheir balancesheetinallthreecategories.TheresultsholdwhetherwenormalizetheNPLsbytheloansinthe samecategoryorbythetotalassets. ThesefindingssuggestthatthelowerNPLsforhighITadoptionbanks(Table2)arenotdrivenbythe fact that IT changed banks business model and induces them to focus less on lending categories that weremoreimpactedbythecrisis,suchascommercialrealestate.TheseresultsalsohighlightthatIThas playedanimportantrolenotonlyforbankslendingtohouseholdsbutalsoforcorporatecredit(Petersen andRajan,2002;Ahnertetal.,2021;Heetal.,2021). 3.4 BankLending HighlevelsofNPLsweighonbanks’profitabilityandcanconstraintheirlending, depressingrealeconomicactivity.AsITadoptionimprovesbanks’resilience,itmayalsoshieldtheirabilitytoprovidecredit tocustomersduring(andrightafter)financialturmoil. Figure4reportstheshareoftotalloans(normalizedbypre-GFCassets)forbanksinthehigh-and low-ITadoptiongroupsfrom2001to2014. Thetwoseriesareindistinguishableupto2006,consistent withITnotbeingaveryimportantfactorinthepre-crisisperiod. From2007 on, theamountofloans providedbylow-ITadoptersisremarkablylowerthantheoneprovidedbythemoreIT-intensecounterparts. Thetwoseriesstartconvergingin2012butthedifferenceisstillpresentin2014. Thesepatterns, 8 Infact,Oster(2019)providesformalstatisticalprocedurestoassessthestabilityofOLScoefficientstotheinclusionof relevantcontrolandtestforthepotentialbiasarisingfromthepresenceofotherunobservablevariables. FollowingOster (2019)jargon, wesetthe“hypotheticalR-square”to1, whichisthemostconservativechoice. Wefindarelativedegreeof selectionabove1,indicatingthat,undertheassumptionof“proportionalselectionofobservablesandunobservables”(Altonji etal.,2005;Oster,2019),theimpactofITadoptiononNPLscannotbenon-negative.Thatis,thevaluesoftheOLScoefficient compatiblewithadditionalunobservablecovariatesareallbelowzero. 15

which are confirmed by regression analysis (see Table A2), indicate that heterogeneity in IT adoption, arguably through the impact on NPLs documented in section 3, helped banks providing credit during (andafter)theGFC.ThissuggestsITinbankingcanhaveanimpactontherealeconomy. 4 TheRootsofITAdoption WhydosomebanksadoptlessITthanothersdespiteitsbeneficialeffects? Economistshavedocumentedthatfrictionsofdifferentnature–suchaslackofinformationoragency frictions(Bloometal.,2013;Atkinetal.,2017)–canpreventorslowdownfirms’adoptionofbeneficial practices and technologies. In this section, we study two factors that could help overcome these frictionsandfosterbanks’ITadoption: thebackgroundandpersonalinclinationofaBHCtopexecutives and the closeness of its headquarters to the land-grant colleges established in the nineteenth century acrosstheUnitedStates. Thefocusonthebank’stopexecutivesandheadquarterlocationisbasedon thedescriptivepatternsdocumentedinsection2:theexplainedvariationintechnologyadoptionatthe branch-levelisdrivenbyBHCcharacteristics(60%)relativetogeographiccharacteristics(27%). 4.1 Executives’Background The characteristics and background of top executives impact firms’ outcomes (Benmelech and Frydman, 2015; Bertrand and Schoar, 2003). We consequently conjecture that top executives with a more tech-pronebackgroundandorientationmaypromoteahigherdegreeofITadoptioninthebanksthey lead. To capture the tech orientation of the top executives (CEOs, CFOs, COOs, and Presidents) hired before 2007, search for tech-related keywords in their biographies (see section 2). We extract several othercharacteristicsfromthebiographies,suchaswhethertheyobtainedapost-graduatedegree(Ph.D. or Master’s), how long they have been in their current position (tenure), their age, their gender, and whether they have an educational background in management or business administration (e.g. completedanMBA).Lastly,weobtaindataontheirtotalcompensationfromtheStandard&Poor’sExecutive Compensation database. We then average these variables for each bank group and match regulatory data,executivebiographies,andsalariesfor156BHC. Totestourconjecture,wethenestimatethefollowingcross-sectionalregressionmodel: 16

Y =α+βExecIT +X (cid:48)γ+(cid:178) (5) b b b b where b is a bank in our sample, ExecIT is the “tech-orientation” of b’s executives, and the deb pendent variable Y is either the level of NPLs over assets during the crisis period or the pre-crisis IT b adoption. ResultsarepresentedinTable4. Column(1)showsapositiveassociationbetweenthetech orientationoftheexecutivesandourbaselineITmeasures. Column(2)showsthatbanksledbymore tech-orientedexecutivesexperiencedlowerNPLsduringthecrisis. Column(3)showsthatbankswith higher-paidmanagershavealsolowerNPLsduringthecrisis,indicatingthatexecutiveswithmorehumancapitalperformedbetterduringthecrisis.Theimpactof“tech-orientation”maythusbedueto–for instance–bettermanagementpracticesunrelatedtoITitself. Incolumn(4)weaddbothvariablestothe regressionandfindthat–oncewecontrolfortech-orientation–theimpactofcompensationisnotstatisticallysignificantanymoreandinsteaddominatedbythetech-savvinessoftheexecutives:itisspecifically thetech-orientation,andnotgeneralqualityorskills,thatmattersforbankperformanceduringthecrisis.Incolumns(5)and(6)weadditionallycontrolforothercharacteristicsoftheexecutives,leavingour coefficient of interest on the tech-orientation statistically significant (the bank controls of Table 2 are also added, but coefficients are left unreported). Interestingly, while most of the characteristics of the executivesdonotseemtoplayaroleforthebankperformanceduringthecrisis,wefindevidencethat banks that had more women and younger leaders among the top executives experienced fewer NPLs. Thesefindingsareconsistentwithtech-savvyexecutivesboostingtheadoptionofITwhich,inturn,improvesbanks’performanceduringthecrisis.Theinterpretationoftech-savvinessasarootofITadoption comeswiththecaveatthatanalternativeexplanationisthatbanksmorepronetoITarealsomorelikely tohire/promotemoretech-proneexecutives. WealsoreestimateEquation3,whichshedslightonthetime-varyingimpactonITadoptiononNPLs, replacingourbaselineITmeasurewiththetechorientationoftheexecutives. Figure3showsstrikingly similarresults. Banksledbymoretech-orientedexecutivesexperiencedasignificantlymorelimitedincreaseinNPLsduring–andrightafter–theGFC.Theresultsarealsosimilarintermsofeconomicmagnitudes. WhileaonestandarddeviationhigherITadoptionusingourbaselinemeasureisassociatedwith 30basispointslowerNPLsduringthepeakofthecrisis,aonestandarddeviationhigherITadoptionin termsofthetech-savvinessofthemanagersisassociatedwith21basispointslowerNPLratio. Asimilar 17

patternemergesfromtherawdata(FigureA6). 4.2 TheLand-GrantColleges Inthissection,westudyapotentiallyexogenousshifteroftechnicalknowledgeandinclinationofbanks’ headquarter decision-makers and other employees. The Morrill Act of 1862 endowed federal land to states to found universities. The focus was to teach science, agriculture, and other technical subjects, duetoanationwidedemandformoretechnicalskills. Whilesomeland-grantuniversitiesoffernowadays degrees in both arts and science, their focus remains on technical subjects. Indeed, in appendix A1weshowthatstudentsatland-grantcollegesanduniversitiesarestillmuchmorelikelytomajorin engineeringandlesslikelytomajorinnon-technicalfields,suchaseducationorbusiness. Wealsofind thatSATscoresinmatharehigherforstudentsofland-grantuniversitystudents,buttheirwritingscores arenot. Thepresenceofland-grantuniversitieshasbeenusedasaninstrumentforthesupplyofskilledlabor inametropolitanarea(Moretti,2004)astheirexactlocationislargelyduetohistoricalaccidents.9 The locationofmanybanks’headquartersisalsorelatedtotheirhistoricalheritageandusuallypredatesthe ITrevolution.10 Forinstance,BankofAmerica’sheadquarterlocationinCharlotte(NorthCarolina)was establishedin1874bythefoundationofthe“CommercialNationalBank”(BlytheandBrockmann,1961). Moregenerally,thepresenceofaBHCheadquarterinaUScountyisuncorrelatedwiththepresenceofa land-grantcollege(seeappendixA1),indicatingland-grantlocationisplausiblyexogenouswithrespect tothemostimportantfactorsaffectingthebankingindustryandheadquarterchoice. Land-grantcollegesanduniversitiescanimpactthetechnicalknowledgeandinclinationofdecisionmakers–and thus their attitude towards more aggressive adoption of IT–in several ways. Directly, by increasingthelikelihoodofhiringworkersthataremoretech-inclinedasgraduatesoftheseinstitutions aremorelikelytodirectlybepartoftheheadquarterpersonnel. Butalsoindirectly,throughspilloverof knowledgeandideasfromthecampuses. As many socio-economic phenomena, including internal migration or knowledge and technology 9Asanexample,thechoiceofIthacaoverSyracusefortheestablishmentofCornellUniversitywasduetothefactthatone ofthetwofounderswasrobbedwhilevisitingSyracuse(Andrews,2019). Land-grantcollegesaredistributedevenlywithina stateandindependentofCensusregions,andnotestablishedinareasthatwerericherduetonaturalresourcesorotherfactors; workersinareasclosetoaland-grantcollegeareshowntobesimilarintermsofracialanddemographiccharacteristicsand haveverycloseArmedForcesQualificationTestscoresforagivenlevelofeducation(Moretti,2004;Shapiro,2006) 10ForeachBHC,wetaketheheadquartercountyfromregulatoryfilingsin1995orearliestavailableyear. 18

diffusion, tend to follow gravity-like patterns (Keller, 2002; Santacreu, 2019), our specification adopts a gravity approach to incorporate both channels: we compute the set of variables D , which are b(c),j equaltothedistanceinlogmiles(plusone)betweenthecountyofeachland-grantcollege j andBHC’s headquartercounty,weightedbythelogsizeofthecollege(STEMenrollment). Under the assumption that land-grant colleges impact banks’ performance during crises only becausetheyfostertheirtechnologyadoption,D canbeusedasasetofinstrumentsforEquation4.As b(c),j closenesstothesecollegesincreasestheeducationofthelocallaborforce,anobviousthreattothisexclusionrestrictionisthatbankslendinareasclosetotheirheadquarterandtheseareasaremoreresilientto shocksthroughtheeffectonoveralleducationandeconomicprosperity.SincewefocusonBHC,whose lendingportfolioisusuallygeographicallydiversified, thisislessofaconcernthanifweweretofocus onsmallerbanks.Wefurthermitigatethisconcernbycontrollingdirectlyforlocaleducation,household income,countysize,andpopulationdensityofthecountywheretheBHCisheadquartered.Wealsoadd statefixedeffects,sowecompareBHCsheadquarteredintheareasthataresimilarforthelevelofeducation,income,anddensity,areinthesamestate,butthathavedifferentexposuretotechnicalknowledge becauseoftheland-grantcolleges. Wealsocontrolfortheshareofemploymentinmanufacturingand construction,becausetheformerisasectorthatsuffersmoreduringrecessions,whilethelatterplayed aparticularlyimportantroleduringtheGFC.Finally,thefindingthatthesecollegeshavefewerstudents majoringinbusinessandmanagementsciencemitigatestheconcernthattheyimpactbanksthrough bettermanagementpracticesratherthanITadoption. WeimplementtheIVempiricalstrategybyestimatingthefollowing2SLSmodel: IT =δ+ (cid:88) ρ ·D +X (cid:48)γ +η (6) b j b(c),j b 1s b j NPL =α+β·IT +X (cid:48)γ +(cid:178) (7) b b b 2s b whereX includestheusualsetofbank-levelpre-crisiscontrols,plusthesetofcounty-levelcontrols. b Includingallcolleges j intothefirststagecouldbeproblematicbothstatisticallyandeconomically,we thuspickonlythethreeclosesttoeachBHC’sheadquarter: intheappendix(FigureA9)weillustratethe negative correlation between the average D for the three closest colleges to b(c) and IT adoption, b(c),j andalsoapositivecorrelationwithNPLsduringthecrisis.11 2SLSarereportedbyTable5:weestimatea 11Asanalternativewaytotesttheplausibilityoftheinclusionrestrictionweestimatethesetofequations: IT b =δ+ρ j · 19

negativeandlargecoefficientofITadoptiononBHCNPLsduringthecrisis(column1).Theinclusionof thecontrols(columns2to4)decreasesthesizeoftheestimatedcoefficientbyabout30%.Thecoefficient ofcolumn(2),whilelargerthanOLS(-0.168),isnotstatisticallydifferentfromit(p-value12%). Similar results are found by using only the closest two colleges or excluding BHC headquartered on the West Coast,Alaska,Hawaii,orPuertoRico.Therefore,thefarthestawaypartofthecountryorthelessdensely populated West Coast does not drive the results. While these results are in line with the OLS analysis, theeffectiveF-statbyOleaandPflueger(2013),whichisrecommendedinmultipleinstrumentssetting (AndrewsandStock,2018),iswellbelowtherule-of-thumbthresholdof10,revealingaweakinstruments problem. Because our instruments are weak, we apply weak instrument techniques following Andrews and Stock (2018). Rather than producing point estimates, these techniques aim to provide confidence intervalsfortheparameterofinterestwhiletakingintoaccounttheextravariabilityintroducedbythelow powerofthefirststage.Eachpointisincludedinthesetifacertainstatisticaltestcannotrejectthatvalue fortheparameter(i.e. constructedthrough“test-inversion”). Asthereisnoconsensusonthebesttest toapplyinoursetting(multipleinstrumentsandlackofhomoskedasticity),weusefourtestsproposed bytheliterature(Moreira,2003;Andrewsetal.,2006;AndrewsandStock,2018):theWaldtest,theconditionallikelihoodratiotest,theK,andtheK-Jtests.12 Table5reportstheseintervals,using90%and95% confidencelevels.Suchintervalsarelarge,asintuitivelyexpectedgiventhelowF-stat.However,theyall rejectthenullofazero(orpositive)impactofITonNPLsduringthecrisis,withoneexception(outof 16)ofthe95%confidenceintervals,providingnoevidencefortheconcernthatOLSestimatesaresolely drivenbysomespuriouscorrelation. Inconclusion,theland-grantIVanalysispresentssomelimitations. Inparticular,theweakIVtechniquesonlyprovideconfidenceintervalsanddonotofferpointestimatesforthe2SLSregressions.Also, stronger technical knowledge of the workforce could improve the organizational quality of the firm in general, improving their ability to screen, and its resilience during a crisis. This analysis nonetheless offersadditionalempiricalsupportfortheclaimthattechnologyadoptioncanhaveacausalimpacton NPLsduringacrisis. D b(c),j +(cid:178) b,j wherej=1,..70isoneoftheland-grantcolleges.Wefindthatρ j isstatisticallydifferentthan0in32casesoutof 70anditisnegative90%ofthesecases. 12TheWaldisnotsuitedforcaseofweakinstrumentssoitisreportedonlyforillustrativepurposes. 20

5 Loan-LevelAnalysis Toshedfurtherlightonthechannelsthroughwhichhigh-ITadoptionbankswereabletolimitthesurge inNPLs,westudytheperformanceandcharacteristicsofmortgagesoriginatedbybankswithheterogeneousdegreesofITadoptionandsoldtoGSEs. WepoolloandatafromFreddieMacandFannieMae Single-FamilyLoansDatasetsandestimatethefollowingloan-levellinearprobabilitymodel: Delinquent =α +β·IT +X (cid:48)γ+η (8) l z(l),o(l) b(l) l l wherel isamortgageheldbyaGSEsandoriginatedbetween2000and2006byacommercialbankin ourITsample;Delinquent isadummyvariableindicatingwhethertheloanhaseverbeendelinquent l (pastdue90daysormore)upto2010. IT isbank-leveltechnologyadoptionofthesellerbanks. X isa b l vectorofmortgagecharacteristicsatorigination.Itincludesborrower’sFICOscore,Loan-to-Value(LTV) ratio, andDebtserving-to-Income(DTI)ratio, adummyvariableformultipleborrowers, asetoffixed effectsfortheoriginationchannelsandthemortgage purpose. Wealso includeα , asetoffixed z(l),o(l) effectsforproperty’szipcodeinteractedwiththeoriginationyeartocontrolforlocalheterogeneitythat canarise,forinstance,fromtheseverityoftheGFCandorfromdifferenthousemarketdynamics. Columns(1)-(4)ofTable6reportsOLSestimatesofEquation8,togetherwithstandarderrorclustered atthesellerlevelwiththedependentvariablemultipliedby100.13 Column(1)reportsthatmortgages originated by banks with a one-standard deviation higher IT adoption are 0.12 percentage points less likelytobedelinquent,or1.6%oftheaverage(whichis7.4percentagepoints). Incolumn(2)weallow thecoefficientofITadoptiontobedifferentforborrowersaboveandbelowthemediancreditscore(735), findingthatonlymortgagesgiventorelativelyriskierborrowersareimpactedbylenders’ITadoption.In column(3)weexpandthesampletoincludeallmortgagesoriginatedsince2000, findingqualitatively similarresults. Thenumberofbanksinthesampleincreasesto27from18incolumns(1)and(2),expandingthevariationusedtoestimatetherelationshipbetweenITandmortgageperformance. These results highlight that at least part of the effect we document in subsection 3.1 is due to the originationofmoreresilientloansbeforethecrisis. Importantly, itshowsthathigh-ITadoptionbanks were not offloading low-quality loans to GSEs. If technology-prone banks were simply better able to securitize and offload their bad loans, IT adoption would lead to lower on-balance sheet NPLs during 13Alternativewaysofclustering,e.g.onthestateorpostalcodelevel,leadtosmallerstandarderrors. 21

thecrisis,withoutreducingtheamountofNPLsinaggregate(Acharyaetal.,2013). Ifthiswasthecase, technology adoption would only lead to risk shifting and increase moral hazard issues and would not enhancefinancialstability. However, the impact of IT adoption on the delinquency of mortgages offloaded to GSEs is much smaller than the impact on NPLs kept on the balance sheet (9.5% of the mean, see Table 3). This is consistentwithhighITbanksoffloadingmortgageswhichareworsethanwhatisheldontheirbalance sheet. As the original pool of originated mortgages is better to start with, arguably because of better screening,eventhese“negatively”selectedmortgagesarebetterthantheonesoffloadedbylowITbanks. Wethenaskwhetherthebetterloanperformancederivesfrombetterborrowerscreeningorex-post monitoring. Agarwaletal.(2017)documentalargedispersionacrossmortgageservicersinthetakeup ofHAMP,averylargepublicly-fundedhomemodificationprogram, potentiallyimpactingloanperformance. The authors also conjecture that part of this heterogeneity in loan modifications could stem fromdifferencesinITinvestments. Incolumn(4)wethusincludetheITadoptionofthefirmthatservices the loan, which we can match to the IT data for most of the sample. We find no evidence that mortgagesservicedbyhighITfirmsarelesslikelytobecomedelinquent.Suchtesthoweverhastwolimitations: about90%ofloansareservicedandsoldbythesamebank,soestimatingvariationislimited, andloanmodification–whichcanhelplimitinglender’slosses–oftenhappenaftertheloanhasbecome delinquent.Incolumn(5)werestrictoursampletoonlyloansthathavebeendelinquentandshowthat mortgageshandledbyhighITadoptionservicersarenotmorelikelytobemodified(anytimeupto2010). Therefore,theresultsincolumn(4)and(5)pointtowardsscreeningratherthanmonitoringasbeingan importantfactorindetermininghighITbanks’performanceduringthecrisis.14 Inthelastcolumn,wetestwhetherbankswithdifferentITadoptionfocusedondifferentsegments ofthemortgagemarkets.Wefindnoevidencethatborrowerswithdifferentcreditscoreswereservedby bankswithdifferentIT.Thisisadditionalevidenceagainstthepotentialexplanationofourfindingsthat ITchangedbanks’businessmodelandmakethemfocusonlessriskyborrowers. Wefind,instead,they werebetteratscreeningborrowers,especiallyamongthehigherrisksegment. AsthecorrelationbetweenITandNPLs(andothermeasuresofbankperformance)outsidethecrisis issmallandinsignificant(seeTable1andTable2),improvementsinscreeningappeartobeparticularly 14Forexample,moreITmighthaveallowedthesebankstosustainmorereliableinternalratingsystems. WerefertoBerg (2015)andBergetal.(2020)foradescriptionofinternalratingsystems. 22

valuablewhenoveralldelinquenciesrise,boostingNPLsandhamperingbanks’performance.Thebanks thathavebetterscreeningability–suchasthebanksthatinvestedmoreonIT–areabletolimittherisein delinquencyandkeeplendingtotherealeconomy. The mortgage data allows us to control for additional characteristics of the loan, which also sheds more light on the channel through which IT adoption can affect NPLs, such as the postal code of the underlying property and the year of origination. The results confirm that the impact of IT adoption onNPLsisnotfully drivenbyhigh-ITadopterslendingtoareasthatwerehitlessby delinquencyand foreclosuresororiginatingalargeramountofloansinaparticularyear. 6 Conclusion As the financial industry becomes more and more reliant on Information Technology, it is extremely policy-relevanttounderstandtheconsequencesforfinancialstabilityofamoreintenseuseoftechnologyinlendingdecisions. Inthispaper,wemeasuretheheterogeneousdegreeofITadoptionofUScommercialbanksbefore the GFC using a novel dataset. We show that high-IT-adopters experienced a significantly smaller increaseinNPLsontheirbalancesheetsandprovidedmorecredittotheeconomyduringthecrisis. We present evidence pointing towards a causal impact of IT on banks’ performance during the crisis. We evaluate different potential mechanisms for the impact of IT, including better screening of borrowers, better monitoring, differences in business models, and offload of risks to GSEs. A loan-level analysis points towards IT improving borrower screening, while we do not find any evidence in favor of other potentialexplanations. References Acharya,ViralV,PhilippSchnabl,andGustavoSuarez,“Securitizationwithoutrisktransfer,”JournalofFinancialEconomics, 2013,107(3),515–536. Adelino,Manuel,AntoinetteSchoar,andFelipeSeverino,“Loanoriginationsanddefaultsinthemortgagecrisis:Theroleofthe middleclass,”TheReviewofFinancialStudies,2016,29(7),1635–1670. Agarwal,Sumit,GeneAmromin,ItzhakBen-David,SouphalaChomsisengphet,TomaszPiskorski,andAmitSeru,“Policyinterventionindebtrenegotiation:Evidencefromthehomeaffordablemodificationprogram,”JournalofPoliticalEconomy, 2017,125(3),654–712. Ahnert, Toni, SebastianDoerr, NicolaPierri, andYannickTimmer, “DoesIThelp? Informationtechnologyinbankingand entrepreneurship,”IMFWorkingPapers,2021,2021(214). 23

Altonji, JosephG,ToddEElder, andChristopherRTaber, “Selectiononobservedandunobservedvariables: Assessingthe effectivenessofCatholicschools,”JournalofPoliticalEconomy,2005,113(1),151–184. Andrews,DonaldWK,MarceloJMoreira,andJamesHStock,“Optimaltwo-sidedinvariantsimilartestsforinstrumentalvariablesregression,”Econometrica,2006,74(3),715–752. Andrews,IsaiahandJamesHStock,“Weakinstrumentsandwhattodoaboutthem,”(NBERSummerInstitute2018-Methods Lectures),2018. Andrews,MichaelJ,“LocalEffectsofLandGrantCollegesonAgriculturalInnovationandOutput,”NBERworkinpaper,2019. Atkin,David,AzamChaudhry,ShamylaChaudry,AmitKKhandelwal,andEricVerhoogen,“Organizationalbarrierstotechnologyadoption: Evidencefromsoccer-ballproducersinPakistan,”TheQuarterlyJournalofEconomics,2017,132(3),1101– 1164. Beaudry,Paul,MarkDoms,andEthanLewis,“Shouldthepersonalcomputerbeconsideredatechnologicalrevolution? EvidencefromUSmetropolitanareas,”JournalofPoliticalEconomy,2010,118(5),988–1036. Beccalli,Elena,“DoesITinvestmentimprovebankperformance?EvidencefromEurope,”JournalofBanking&Finance,2007, 31(7),2205–2230. Becker,GaryS,Humancapital:Atheoreticalandempiricalanalysis,withspecialreferencetoeducation,UniversityofChicago press,2009. Benmelech,EfraimandCarolaFrydman,“Militaryceos,”JournalofFinancialEconomics,2015,117(1),43–59. Berg, Tobias, “Playingthedevil’sadvocate: Thecausaleffectofriskmanagementonloanquality,” TheReviewofFinancial Studies,2015,28(12),3367–3406. ,ManjuPuri,andJorgRocholl,“LoanOfficerIncentives,InternalRatingModelsandDefaultRates,”ReviewofFinance,2020. ,ValentinBurg,AnaGombovic´,andManjuPuri,“OntheriseofFinTechs–Creditscoringusingdigitalfootprints,”TheReview ofFinancialStudies,2019. Berger,AllenN,“Theeconomiceffectsoftechnologicalprogress:Evidencefromthebankingindustry,”JournalofMoney,Credit andBanking,2003,pp.141–176. Bertrand,MandASchoar,“TheManagingwithStyle: TheEffectofManagersonCorporatePolicy,”QuarterlyJournalofEconomics,2003,118(4),1169. Bertrand, Marianne, Esther Duflo, and Sendhil Mullainathan, “How much should we trust differences-in-differences estimates?,”TheQuarterlyjournalofeconomics,2004,119(1),249–275. Bloom,Nicholas,BennEifert,AprajitMahajan,DavidMcKenzie,andJohnRoberts,“Doesmanagementmatter?Evidencefrom India,”TheQuarterlyJournalofEconomics,2013,128(1),1–51. ,RaffaellaSadun,andJohnVanReenen,“AmericansdoITbetter:USmultinationalsandtheproductivitymiracle,”American EconomicReview,2012,102(1),167–201. Blythe,LeGetteandCharlesRavenBrockmann,Hornets’Nest:TheStoryofCharlotteandMecklenburgCounty,PublicLibraryof CharlotteandMecklenburgCounty,1961. Bofondi,MarcelloandFrancescaLotti,“Innovationintheretailbankingindustry: thediffusionofcreditscoring,”Reviewof IndustrialOrganization,2006,28(4),343–358. Boot,Arnoud,PeterHoffmann,LucLaeven,andLevRatnovski,“Fintech:what’sold,what’snew?,”JournalofFinancialStability, 2021,53,100836. Bostandzic,DenefaandGregorNFWeiss,“InnovatingBanksandLocalLending,”AvailableatSSRN3421123,2019. Branzoli,Nicola,EdoardoRainone,andIlariaSupino,“TheRoleofBanks’TechnologyAdoptioninCreditMarketsduringthe Pandemic,”AvailableatSSRN3878254,2021. 24

Bresnahan,TimothyF,ErikBrynjolfsson,andLorinMHitt,“Informationtechnology,workplaceorganization,andthedemand forskilledlabor:Firm-levelevidence,”TheQuarterlyJournalofEconomics,2002,117(1),339–376. Brynjolfsson,ErikandLorinMHitt,“Computingproductivity:Firm-levelevidence,”ReviewofEconomicsandStatistics,2003, 85(4),793–808. Buchak,Greg,GregorMatvos,TomaszPiskorski,andAmitSeru,“Fintech,regulatoryarbitrage,andtheriseofshadowbanks,” JournalofFinancialEconomics,2018,130(3),453–483. Caballero,RicardoJ,TakeoHoshi,andAnilKKashyap,“ZombielendinganddepressedrestructuringinJapan,”AmericanEconomicReview,2008,98(5),1943–77. Carletti,Elena,StijnClaessens,AntonioFatás,andXavierVives,BarcelonaReport2-TheBankBusinessModelinthePost-Covid- 19World,CentreforEconomicPolicyResearch,2020. Claessens, Stijn, JonFrost, GrantTurner, andFengZhu, “Fintechcreditmarketsaroundtheworld: size, driversandpolicy issues,”BISQuarterlyReviewSeptember,2018. Dadoukis,Aristeidis,MaurizioFiaschetti,andGiuliaFusi,“ITadoptionandbankperformanceduringtheCovid-19pandemic,” EconomicsLetters,2021,204,109904. D’Andrea, AngeloandNicolaLimodio, “High-speedinternet, financialtechnologyandbankinginafrica,” BAFFICAREFIN CentreResearchPaper,2019,(2019-124). Degryse,HansandStevenOngena,“Distance,lendingrelationships,andcompetition,”TheJournalofFinance,2005,60(1), 231–266. Demirgüç-Kunt,AsliandEnricaDetragiache,“Doesdepositinsuranceincreasebankingsystemstability?Anempiricalinvestigation,”JournalofMonetaryEconomics,2002,49(7),1373–1406. FSB,“FinTechandmarketstructureinfinancialservices:Marketdevelopmentsandpotentialfinancialstabilityimplications,” Basel:FinancialStabilityBoard,2019. Fuster,Andreas,MatthewPlosser,PhilippSchnabl,andJamesVickery,“Theroleoftechnologyinmortgagelending,”TheReviewofFinancialStudies,2019,32(5),1854–1899. ,PaulGoldsmith-Pinkham,TarunRamadorai,andAnsgarWalther,“Predictablyunequal?theeffectsofmachinelearningon creditmarkets,”JournalofFinance,2018. Hannan,TimothyHandJohnMMcDowell,“Thedeterminantsoftechnologyadoption: Thecaseofthebankingfirm,”The RANDJournalofEconomics,1984,pp.328–335. and ,“Rivalprecedenceandthedynamicsoftechnologyadoption:anempiricalanalysis,”Economica,1987,pp.155–171. He,Zhiguo,SheilaJiang,DouglasXu,andXiaoYin,“InvestinginLendingTechnology: ITSpendinginBanking,”Universityof Chicago,BeckerFriedmanInstituteforEconomicsWorkingPaper,2021,(2021-116). Hernández-Murillo,Rubén,GerardLlobet,andRobertoFuentes,“Strategiconlinebankingadoption,”JournalofBanking& Finance,2010,34(7),1650–1663. Hirtle,Beverly,AnnaKovner,andMatthewCPlosser,“Theimpactofsupervisiononbankperformance,”FRBofNYStaffReport, 2018. Hurst,Erik,BenjaminJKeys,AmitSeru,andJosephVavra,“RegionalredistributionthroughtheUSmortgagemarket,”American EconomicReview,2016,106(10),2982–3028. Keller,Wolfgang,“Geographiclocalizationofinternationaltechnologydiffusion,”AmericanEconomicReview,2002,92(1),120– 142. Keys,BenjaminJ,AmitSeru,andVikrantVig,“Lenderscreeningandtheroleofsecuritization:evidencefromprimeandsubprimemortgagemarkets,”TheReviewofFinancialStudies,2012,25(7),2071–2108. 25

,TanmoyMukherjee,AmitSeru,andVikrantVig,“Financialregulationandsecuritization:Evidencefromsubprimeloans,” JournalofMonetaryEconomics,2009,56(5),700–720. , , ,and ,“Didsecuritizationleadtolaxscreening? Evidencefromsubprimeloans,”TheQuarterlyJournalofEconomics,2010,125(1),307–362. Kovner,Anna,JamesIVickery,andLilyZhou,“Dobigbankshaveloweroperatingcosts?,”EconomicPolicyReview,2014,20(2), 1–27. Kwan,Alan,ChenLin,VesaPursiainen,andMingzhuTai,“StressTestingBanks’DigitalCapabilities:EvidenceFromtheCOVID- 19Pandemic,”mimeo,2021. Lewellen, StefanandEmilyWilliams, “Didtechnologycontributetothehousingboom? EvidencefromMERS,”Journalof FinancialEconomics,2021. Liberti,JoséMaríaandMitchellAPetersen,“Information:Hardandsoft,”ReviewofCorporateFinanceStudies,2018,8(1),1–41. Maggio,MarcoDiandVincentYao,“FintechBorrowers:LaxScreeningorCream-Skimming?,”TheReviewofFinancialStudies, 2021,34(10),4565–4618. Moreira,MarceloJ,“Aconditionallikelihoodratiotestforstructuralmodels,”Econometrica,2003,71(4),1027–1048. Moretti,Enrico,“Estimatingthesocialreturntohighereducation: evidencefromlongitudinalandrepeatedcross-sectional data,”Journalofeconometrics,2004,121(1-2),175–212. Olea,JoséLuisMontielandCarolinPflueger,“Arobusttestforweakinstruments,”JournalofBusiness&EconomicStatistics, 2013,31(3),358–369. Oster,Emily,“Unobservableselectionandcoefficientstability:Theoryandevidence,”JournalofBusiness&EconomicStatistics, 2019,37(2),187–204. Peek,JoeandEricSRosengren,“Collateraldamage: EffectsoftheJapanesebankcrisisonrealactivityintheUnitedStates,” AmericanEconomicReview,2000,90(1),30–45. Petersen,MitchellAandRaghuramGRajan,“Thebenefitsoflendingrelationships: Evidencefromsmallbusinessdata,”The JournalofFinance,1994,49(1),3–37. and ,“Doesdistancestillmatter?Theinformationrevolutioninsmallbusinesslending,”TheJournalofFinance,2002,57 (6),2533–2570. Pierri, NicolaandYannickTimmer, “ITShields: TechnologyAdoptionandEconomicResilienceduringtheCOVID-19Pandemic,”IMFWorkingPapers,2020,2020(208). Rajan,RaghuramG,“Hasfinancemadetheworldriskier?,”EuropeanFinancialManagement,2006,12(4),499–533. Rajan,Uday,AmitSeru,andVikrantVig,“Thefailureofmodelsthatpredictfailure:Distance,incentives,anddefaults,”Journal ofFinancialEconomics,2015,115(2),237–260. Santacreu,AnaMaria,“WhatDrivesInternationalTechnologyDiffusion? AGravityApproach,”FRBSt.LouisWorkingPaper, 2019,(2019-31). Shapiro,JesseM,“Smartcities:qualityoflife,productivity,andthegrowtheffectsofhumancapital,”Thereviewofeconomics andstatistics,2006,88(2),324–335. 26

Figure1:NPLsoverAssetsbypre-GFCITadoption ThisFigureplotsthemedianshareofNPLsoverassetsforhigh-andlow-ITadopters. “HighITadoption"isthemedianshare ofNPLsoverassetsforbankswithIT abovethe75thpercentile. “Low-ITadoption"isthemedianshareofNPLsoverassets b forbankswithIT belowthe25thpercentile.Weincludeonlybanksforwhichwehaveregulatorydataforatleast14years.See b subsection3.1andsection2formoredetails. 27

Figure2:Time-varyingEffectofITadoptiononNPLs ThisFigureplotsthecoefficientandthe95%and99%confidenceintervalsofβτfromthefollowingestimatedequation: NPL b,t =α b +δ t + (cid:88) βτIT b ·1[t=τ]+(cid:178) b,t τ(cid:54)=2006 whereb isabank(BHC), t oneyearbetween1996and2014, α b arebankfixedeffects, andδ t areyearfixedeffects. The dependentvariableNPL istheshareofNPLsoverassetsinb’sregulatoryfilingforyeart. IT isthepre-crisisITadoption b,t b ofbankbestimatedasdescribedinsection2.Thecoefficientof2006isnormalizedtozero.Confidenceintervalsarebasedon double-clusteredstandarderrorsatthebankandyearlevel.Seesubsection3.1andsection2formoredetails. 28

Figure3:Time-varyingEffectoftech-backgroundofexecutivesonNPLs 2. 1. 0 1.- 2.- 3.- 4.- 1996 1998 2000 2002 2004 2006 2008 2010 2012 2014 Year 99% Confidence Interval 95% Confidence Interval Point Estimates ThisFigureplotsthecoefficientandthe95%and99%confidenceintervalsofβτfromthefollowingestimatedequation: NPL b,t =α b +δ t + (cid:88) βτExecIT b ·1[t=τ]+(cid:178) b,t τ(cid:54)=2006 whereb isabank(BHC), t oneyearbetween1996and2014, α b arebankfixedeffects, andδ t areyearfixedeffects. The dependentvariableNPL istheshareofNPLsoverassetsinb’sregulatoryfilingforyeart. ExecIT istheaverage“techb,t b orientation”ofbank’sbtopexecutives(CEOs,CFOs,andPresidents).The“tech-orientation”ofabanks’executivesiscomputed bydividingthetotalamountof“tech-related”keywordsoverthetotalamountofwordsintheirbiographies,seesubsection4.1 andsection2formoredetails.Thecoefficientof2006isnormalizedtozero.Confidenceintervalsarebasedondouble-clustered standarderrorsatthebankandyearlevel.Seesubsection3.1andsection2formoredetails. 29

Figure4:Loansoverpre-crisisAssetsbypre-GFCITadoption ThisFigureplotsthemedianshareoftotalloansscaledbyaveragepre-crisis(2001-2006)assetsforhigh-andlow-ITadopters. “HighITadoption"isthemedianshareofLoanoverpre-crisisassetsforbankswithIT abovethe75thpercentile. “LowIT b adoption"isthemedianshareofLoanoverpre-crisisassetsforbankswithIT belowthe25thpercentile. Weincludeonly b banksforwhichwehaveregulatorydataforatleast14years.Seesubsection3.4andsection2formoredetails. 30

Table1:PanelRegressions (1) (2) (3) (4) (5) NPL NPL NPL NPL NPL IT-adoption -0.0239 -0.0283 (0.017) (0.018) ∗∗ ∗∗ crisis 0.811 0.793 (0.349) (0.346) IT-adoption×crisis -0.160 ∗∗ -0.168 ∗∗ -0.157 ∗∗ -0.170 ∗∗ -0.151 ∗∗ (0.063) (0.065) (0.066) (0.068) (0.067) IToflocalcompetitors×crisis 0.0309 (0.044) (Within)R-squared 0.112 0.140 0.0111 0.00997 0.0482 N 4608 4608 4608 4608 4608 BankFE No Yes No Yes Yes YearFE No No Yes Yes Yes Bankcontrols×crisis No No No No Yes Resultsofestimatingthefollowingequation: NPL b,t =α b +δ t +βIT b ·crisis+(X b ·crisist) (cid:48)γ+(cid:178) b,t whereb isabank(BHC), t oneyearbetween2001and2014, crisist adummyvariableindicatingyears2007to2010, α b arebankfixedeffects, andδ t areyearfixedeffects. ThedependentvariableNPL b,t istheshareofNPLsoverassetsinb’s regulatoryfilingforyeart. IT isthepre-crisisITadoptionofbankbestimatedasdescribedinsection2. Thebank-levelset b ofcontrolsX includesthepre-crisis(2001-2006)averageof:theloanstoassetsratio,thecapitaltoassetsratio,thewholesale b fundingratio,ROA,the(logof)averagewagesinthousandsofUSD,andthe(logof)assetssizeinthousandsofUSD.X also b includestheaverageITadoptionoflocalcompetitorsandameasureofexposuretothehousepriceshocks(HPExposure) basedonthecombinationobservedpercentagechangeinprices(2006Q4-2010Q4)ineachcountyandthelocationofbanks’ branches. Columns(1)and(3)excludebankfixedeffect, whilecolumn(1)and(2)excludeyearfixedeffects. Column(5) includesinteractedcontrolsbutonlydisplaystheonebetweenIToflocalcompetitors. Seesubsection3.1andsection2for moredetails. Samplesizeiskeptconstantbydroppingobservationswithmissingvaluesforanyvariable. Standarderrors(in parentheses)aredouble-clusteredonbankandyearlevel.*p<0.1,**p<0.05,***p<0.01 31

Table2:Cross-SectionalRegressions (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) NPL Loans Size HPExposure Capital Wholesale ROA LogWage IToflocalcomp. NPL IT-adoption -0.183∗∗∗ -0.648 -0.0931 0.131 -0.195 -0.0459 -0.0282 -0.0227 0.275∗∗∗ -0.168∗∗∗ (0.061) (0.700) (0.057) (0.108) (0.420) (0.372) (0.049) (0.018) (0.083) (0.061) R-squared 0.0262 0.00220 0.00712 0.000875 0.000427 0.0000383 0.00107 0.00414 0.0750 0.186 N 337 337 337 337 337 337 337 337 337 337 Mean 1.54 62.69 13.9 .64 13.02 15.92 2.55 4.84 .01 1.54 Std.Dev. 1.13 13.8 1.1 4.41 9.43 7.41 .86 .35 1 1.13 Resultsofestimatingthefollowingequation: Y =α+βIT +(cid:178) b b b wherebisabank(BHC)andIT isthepre-crisisITadoptionofb,estimatedasdescribedinsection2.Thedependentvariable b Y iseithertheshareofNPLsoverassetsinbankbregulatoryfiling(averagedover2007to2010)oroneofthevariablesofthe b setX ,definedasfollow. X includesthepre-crisis(2001-2006)averageof:theloanstoassetsratio,thecapitaltoassetsratio, b b thewholesalefundingratio,ROA,the(logof)averagewagesinthousandsofUSD,andthe(logof)assetssizeinthousandsof USD.X alsoincludestheaverageITadoptionoflocalcompetitorsandameasureofexposuretothehousepriceshocks(HP b Exposure)basedonthecombinationoftheobservedpercentagechangeinprices(2006Q4-2010Q4)ineachcountyandthe locationofbanks’branches. Incolumn(10)thedependentvariableistheshareofNPLsoverassetsandthesetofcovariates X areincludedascontrols. Seesubsection3.2andsection2formoredetails. Samplesizeiskeptconstantbydropping b observationswithmissingvaluesforanyvariable.Robuststandarderrorsarereportedinparentheses.*p<0.1,**p<0.05,*** p<0.01 Table3:NPLsbyLoanCategory (1) (2) (3) (4) (5) (6) NPLsduringcrisis: ResidentialRE CommercialRE C&I ResidentialRE CommercialRE C&I Normalizedby Loansinthecategory TotalAssets IT-adoption -0.196∗∗ -0.396∗∗∗ -0.256∗∗∗ -0.0311∗∗ -0.129∗∗ -0.0253∗∗∗ (0.0927) (0.138) (0.0878) (0.0139) (0.0642) (0.00946) R-squared 0.0108 0.0195 0.0279 0.00995 0.0154 0.0243 N 284 285 285 285 285 285 AverageNPLsduringcrisis 2.062 3.469 1.662 0.332 1.062 0.165 NormalizedbyaverageNPLs -0.0949 -0.114 -0.154 -0.0936 -0.121 -0.153 Resultsofestimatingthefollowingequation: Yk=α+βIT +(cid:178) b b b wherebisabank(BHC)andIT isthepre-crisisITadoptionofb,estimatedasdescribedinsection2.Thedependentvariable b Y istheshareofNPLsinloancategorykinbankbregulatoryfiling(averagedover2007to2010). Column(1)and(4)report b theresultsforresidentialrealestateNPLs,columns(2)and(5)forcommercialrealestateand(3)and(6)forcommercialand industrialloannpls.Incolumns(1)-(3)wedividebytheamountofloansbythebankintherespectivecategoryandincolumns (4)-(6)wedividebytotalassetsofthebank. b/meandisplaystherespectivecoefficientrelativetothemeanoftheyvariable. b/sddisplaystherespectivecoefficientrelativetothestandarddeviationoftheyvariable.Seesubsection3.2andsection2for moredetails.Robuststandarderrorsarereportedinparentheses.*p<0.1,**p<0.05,***p<0.01 32

Table4:Executives’“tech-orientation” (1) (2) (3) (4) (5) (6) IT NPL NPL NPL NPL NPL Tech-orientation 0.104∗ -0.255∗∗∗ -0.237∗∗∗ -0.131∗∗ -0.142∗ (0.057) (0.069) (0.068) (0.065) (0.073) LogCompensation -0.131∗ -0.108 -0.0611 -0.00878 (0.069) (0.068) (0.074) (0.076) Management -0.0518 -0.0958 (0.115) (0.133) Post-Grad 0.344 (0.332) Longtenure -0.232 (0.208) Female -0.761∗ (0.440) Age 0.0239∗ (0.013) R-squared 0.0136 0.0455 0.0205 0.0592 0.276 0.306 N 149 156 156 156 156 156 Resultsofestimatingthefollowingequation: Y =α+βExecIT +(cid:178) b b b whereb isabank(BHC).ThedependentvariableY iseithertheratioofNPLstoassetsaveragedbetween2007and2010 b (columns(2)-(6)), orthepre-crisisITadoption(column1), estimatedasdescribedinsection2. Theindependentvariable ExecIT isthe“tech-orientation”ofbank’sbtopexecutives(CEOs,CFOs,andPresidents). The“tech-orientation”ofabanks’ b executivesiscomputedbydividingthetotalamountof“tech-related”keywordsoverthetotalamountofwordsintheirbiographies,seesubsection4.1andsection2formoredetails. LogCompensationisthelogaveragecompensationoftheexecutives. Managementiscomputedbydividingthetotalamountof“management”keywordsoverthetotalamountofwordsin theirbiographies. Post-gradistheshareofexecutiveswithapost-graddegree. Long-tenureistheshareofexecutiveswithan abovemediantenureatitscurrentposition.Femaleistheshareoffemaleexecutives.Ageistheaverageageoftheexecutives. Columns(5)and(6)addthesetofpre-GFCbanklevelcontrolsdescribedinsection2. Robuststandarderrorsarereportedin parentheses.*p<0.1,**p<0.05,***p<0.01 33

Table5:Land-GrantCollegesInstruments (1) (2) (3) (4) NPL NPL NPL NPL IT-adoption -1.343** -0.859** -0.877** -0.850* (0.560) (0.430) (0.409) (0.434) EffectiveF-stat 3.14 2.42 2.7 3.84 Instruments 3closest 3closest 3closest 2closest Controls No Yes Yes Yes N 337 337 306 337 Sample All All ExclWest All 90%ConfidenceIntervals(RobusttoWeakInstruments) WaldCI: [-2.26,-.42] [-1.57,-.15] [-1.55,-.20] [-1.56,-.14] CLRCI: [...,-.50] [-2.40,-.21] [-2.21,-.29] [-2.20,-.21] KCI: [...,-.50] [-2.07,-.28] [-2.02,-.34] [-2.05,-.24] K-JCI: [...,-.45] [-2.21,-.24] [-2.15,-.30] [-2.20,-.21] 95%ConfidenceIntervals(RobusttoWeakInstruments) WaldCI: [-2.44,-.25] [-1.70,-.02] [-1.68,-.08] [-1.70,.00] CLRCI: [...,-.34] [-3.51,-.03] [-2.94,-.15] [-2.96,-.06] KCI: [...,-.36] [-2.63,-.16] [-2.50,-.21] [-2.62,-.11] K-JCI: [...,-.31] [-2.85,-.11] [-2.69,-.18] [-2.85,-.08] Resultsofestimatingthefollowing2SLSequation: IT b =δ+(cid:88)ρ j ·D b(c),j +η b j NPL =α+β·IT +(cid:178) b b b wherebisabank(BHC).NPL istheratioofNPLstoassetsaveragedbetween2007and2010. IT ispre-crisisITadoption, b b estimatedasdescribedinsection2. D isthedistanceinlogmiles(plusone)betweenthecountyofland-grantcollege b(c),j j andBHCb’sheadquartercounty,weightedbythelogsizeofthecollege. Wealsoinclude–inbothfirstandsecondstage–a setofpre-crisisBHC-levelcontrolsplusasetofcounty-levelcontrolsforheadquartercounty.EffectivefirststageF-statsfrom OleaandPflueger(2013)aredisplayed. Thebottompanelsreportthe90%and95%weak-instrumentconfidenceintervals constructedwithdifferentstatistics(AndrewsandStock,2018).The...incolumn(1)indicatethereisnoestimatedlowerbound tothoseconfidenceintervals.Tosearchforalowerbound,weexpandthegriduptoan“unreasonablylow”valuefortheimpact ofITonNPLs:thevaluesuchthataonestandarddeviationhigherITadoptionwouldbepredictedtooffsetalltheNPLsfor99% ofBHC(i.e.,-4.71);suchvaluecannotberejectedforthespecificationwithnocontrols(column1). leavingthelowerbounds undefined. Column(3)excludeBHCheadquarteredontheWestCoast(Alaska,California,Oregon,Washington),Hawaii,or PuertoRico.Seesubsection4.2formoredetails.Robuststandarderrorsarereportedinparentheses.*p<0.1,**p<0.05,*** p<0.01 34

Table6:Loan-LevelRegressions EverDeliquent90Days+ Modified CreditScore (1) (2) (3) (4) (5) (6) ∗ ∗∗∗ ∗∗∗ ITadoption-originator -0.118 -0.302 -0.325 -0.601 (0.059) (0.080) (0.069) (0.587) ITadoption×HighCreditScore 0.0327 (0.109) ITadoption×LowCreditScore -0.271 ∗∗ (0.113) ITadoption-servicer 0.337 -1.502 (0.509) (0.932) Origination 2004/2006 2004/2006 2000/2006 2004/2006 2004/2006 2004/2006 R2 0.0889 0.0890 0.0765 0.0871 0.103 0.115 Observations 5,063,032 5,063,032 16,406,595 4,387,965 303,809 5,063,032 Resultsofestimatingthefollowingequation: Delinquent =α +βIT +X (cid:48)γ+η l z(l),o(l) b(l) l l wherel isamortgageheldbyFreddieMacorFannieMaeandoriginatedbefore2007,α arepostal-code∗originationz(l),o(l) yearfixedeffectsoftheunderlyingloan. IT istheisthepre-crisisITadoptionofthebankwhichsoldthemortgageto b(l) FreddieMacorFannieMae,estimatedasdescribedinsection2. ThedependentvariableDelinquent isadummyvariable l indicatingwhetheraloanwaseverdelinquentfor90+days(multipliedby100)anytimeupto2010. α areorigination z(l),o(l) zipcode×yearfixedeffects,whilethevectorofcontrolsX includesthecreditscorescore,thedebtservicingtoIncome(DTI), l theLoan-to-Value(LTV)ratiosatorigination,theoccupancystatus,theloanpurpose,adummyforaloanwithmultipleborrowers,thelogloanamount,andtheoriginationchannel.Allindependentvariablesarestandardizedtohaveameanof0and astandarddeviationof1.Column(2)interactsbanks’ITadoptionwithadummyvariableindicatingwhethertheborrowerhas aCreditScoreaboveorbelowthemedian. Column(3)includesmortgagesoriginatedsince2000,sotoexpandthesampleof originationbanks:28versus18forcolumns(1)-(2).Column(4)addstheITadoptionoftheservicingfirm.Column(5)replaces thedependentvariablewithadummyiftheloanhaseverbeenmodified. Itincludesonlyloansthathavebeendelinquentat leastonce. Column(6)replacesthedependentvariablewiththecreditscoreoftheborrower. Seesection5andsection2for moredetails.Standarderrors(inparentheses)areclusterattheoriginationbank-level.*p<0.1,**p<0.05,***p<0.01 35

Supplemental Materials A1 Land-GrantCollegesCharacteristicsandHeadquarterSelection Inthissectionwestudythecharacteristicsofland-grantcolleges’studentsandhowthelocationoflandgrantcollegesimpactbanks’headquarterslocation. Theseresultsarecomplementarytotheanalysisin subsection 4.2. The IPEDS survey provides data on the major and SAT scores for students enrolled in morethan1,400highereducationinstitutionsintheUSduringfall2018.Wethenestimatethefollowing regressions: Share =α +β Landgrant +(cid:178) u,M M M u u,M and SAT_Score =α+βLandgrant +(cid:178) u u u whereShare istheshareofstudentsineachof6fieldsofstudyM wehaveinformationon(we u,M dropdentistryschool)ininstitutionu,SAT_Score isthe25thor75thpercentileofSATscoreinreading u ormathoftheenrolledstudents,andLandgrant isadummyflaggingwhetheruisaland-grantcollege u oruniversity. ResultsarereportedinTableA3andTableA4. Land-grantinstitutionshaveamuchlarger shareofstudentsenrolledinengineeringandslightlymorestudentsinotherscientificdisciplines,such asbiologyandphysics.Conversely,theyhavemuchfewerstudentsinbusinessandmanagementscience andalsolessstudentsineducation. Moreover,studentsatland-grantcollegeshavesignificantlyhigher math scores (whether we look at the 25th or 75th percentile) but similar reading scores. These results indicatethatland-grantcollegesaremainlytechnicalschools. Werepeattheanalysisusingdatafrom fall1996(sincewetakebankheadquarterlocationin1995whenpossible)andfindverysimilarresults. Wethenmovetoanalyzewhetherthedistancefromland-grantcollegespredictsheadquarters’location.Weestimatethefollowinglinearprobabilitymodel: BankHQ =α+βDistance_Landgrant +γX +(cid:178) c c c c whereBankHQ isadummyvariableindicatingwhetherthecountyc hosttheheadquarterofoneof c the337BHCofourmainsample,and X issetofcontrolsincludingpre-GFCeducation,income,and c 1

statefixedeffects.Distance_Landgrant isoneoffourmeasureofdistance(inlogofmilesplusone)of c countyc toland-grantcolleges. Thefirstthreemeasuresareclosestcollege,medianacrossallcolleges, mean across all colleges. The fourth measure is, instead, a linear combination of the distance of the countyfromallland-grantcolleges.Theparametersofsuchlinearcombinationarechoseninaprevious estimationstagewherewerelyonLASSOtopredicttheITadoptionofaBHCwiththedistanceofthe BHC’sheadquarterfromallland-grantcolleges.ThisfourthmeasureissalientasLASSOextractthevariationinthecounty-collegesdistancesvectorthatismoreimportanttoexplainourvariableofinterest, thatisBHC’sITadoption.ResultsarepresentedinTableA5.Nomeasurehasstatisticallysignificantpredictivepower.Theresultsarerobusttoestimatingaprobitmodelratherthanalinearprobabilitymodel (unreported). 2

FigureA1:PCusageandotherITMeasures TheFigureplotsbinscatterplotsbetweenthevariousITmeasuresandtheshareofemployeesusingaPCusingcountry-classdatafromtheOECDICTAccessandUsage Businessesdatasetfor2018,whereaclassisdefinedbyindustryorfirmsize. Pairwiselinearcorrelationsrangebetween31%and64%. Samepatternsemergebyselectinga differentyearorpoolingdataacrossallyears.WethankFrancescoManaresiforsuggestingthisexercise. 3

FigureA2:Cross-sectionaldistributionofNPLsoverAssets(crisis)andITadoption(pre-crisis) ThisFigureplotsthecross-sectionaldistributionoftheratioofNPLstoassetsaveragedbetween2007and2010(toppanel)and ofthepre-crisisITadoptionIT .Seesection2formoredetails. b FigureA3:Coveragebyloantype RatioofloansheldbyBHCinthematchedanalysissampleversustotalloansintheregulatorydataset,bytypeofloans(in 2006). 4

FigureA4:NPLsoverAssetsbypre-GFCITadoption ThisFigureplotsthemedianshareofNPLsoverassetsforhigh,medium,andlow-ITadopters.“HighITadoption"isthemedian shareofNPLsoverassetsforbankswithIT abovethe75thpercentile. “LowITadoption"isthemedianshareofNPLsover b assetsforbankswithIT belowthe25thpercentile. “MedianITadoption"isthemedianshareofNPLsoverassetsforbanks b withIT betweenthe25thpercentileandthe75thpercentile.Weincludeonlybanksforwhichwehaveregulatorydataforat b least14years.Seesubsection3.1andsection2formoredetails. FigureA5:NPLsoverpre-GFCAssetsbypre-GFCITadoption ThisFigureplotsthemedianshareofNPLsscaledbyaveragepre-crisis(2001-2006)assetsforhigh-andlow-ITadopters.“High ITadoption"isthemedianshareofNPLsoverpre-crisisassetsforbankswithIT abovethe75thpercentile.“Low-ITadoption" b isthemedianshareofNPLsoverpre-crisisassetsforbankswithIT belowthe25thpercentile.Weincludeonlybanksforwhich b wehaveregulatorydataforatleast14years.Seesubsection3.1andsection2formoredetails. 5

FigureA6:NPLsoverpre-GFCAssetsbybanktopexecutives’technologyorientation ThisFigureplotsthemedianshareofNPLsscaledbyaveragepre-crisis(2001-2006)assetsforbankswithhighandlowexecutives’“tech-orientation”. “High-ITexecutive"isthemedianshareofNPLsoverpre-crisisassetsforbankswithexecutives “tech-orientation”abovethe75thpercentile. “Low-ITexecutive"isthemedianshareofNPLsoverpre-crisisassetsforbanks withexecutives“tech-orientation”atorbelowthe25thpercentile. Weincludeonlybanksforwhichwehaveregulatorydata foratleast14years. The“tech-orientation”ofbanks’executivesiscomputedbydividingthetotalamountof“tech-related” keywordsoverthetotalamountofwordsintheirbiographies. Wethencomputeabank-levelmeasurebyaveragingoverthe topexecutives(CEOs,CFOs,COOs,andPresidents)hiredbefore2007.Seesubsection4.1andsection2formoredetails. FigureA7:ProvisionforcreditlossesoverAssets % 4 3 2 1 0 1995 2000 2005 2010 2015 Year High IT adoption Low IT adoption ThisFigureplotsthemedianshareofprovisionforcreditlossesscaledbyassetshigh-andlow-ITadopters.“HighITadoption" isthemedianshareofNPLsoverpre-crisisassetsforbankswith IT abovethe75thpercentile. “Low-ITadoption"isthe b medianshareofNPLsoverpre-crisisassetsforbankswithIT belowthe25thpercentile.Weincludeonlybanksforwhichwe b haveregulatorydataforatleast14years.Seesubsection3.1andsection2formoredetails. 6

FigureA8:RobustnessoftheExecutives’resultstochangesinthekeywordslist ThisFigureplotsthecoefficientofcolumns(2)-(3)ofTable4fordifferentmeasuresofbanktopexecutives’technologyorientation.Foreachwordusedindefiningthetechnologyorientationofexecutives,wecreateanewmeasureinwhichweleaveout thisparticularwordandbuildthemeasurebasedonallremainingwords.Thedashedlinereflecttheestimatesofcolumns(2) and(3)ofTable4.Seesubsection4.1andsection2formoredetails. FigureA9:Averagedistancetonearbyland-grantcolleges,ITadoption,andNPLsduringGFC ThefigureplotstheaverageITadoptionandNPLsduringGFCforeachof20binsdefinedaccordingtotheaveragelogdistance fromthethreeclosestland-grantcolleges(weightedbysize). AllvariablesareresidualizedaftercontrollingforBHCpre-crisis characteristics,BHC’sheadquartercountycharacteristics,andstatefixedeffects. Thedistanceisnormalizedtotohavemean zeroandunitstandarddeviation. 7

TableA1:RobustnessofMainPanelRegression (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) NPL NPL NPL NPL NPL NPL NPL NPL NPL NPL IT-adoption×crisis -0.165 ∗∗ -0.243 ∗ -0.158 ∗∗ -0.161 ∗∗ -0.242 ∗∗ -0.214 ∗∗ -0.380 ∗ -0.882 ∗∗ -0.165 ∗∗∗ -0.165 ∗∗∗ (0.068) (0.120) (0.069) (0.063) (0.095) (0.080) (0.183) (0.404) (0.051) (0.054) Exercise Baseline PCsperEmp HWIT HWNPLs Loans Broaddef. Asof2006 Provisions BankClustering Bootstrappeds.e. R-squared 0.00944 0.00376 0.00794 0.0108 0.00867 0.00993 0.00530 0.00158 0.00944 N 4692 5035 4692 4692 4692 4692 4655 4548 4692 4692 Resultsofestimatingthefollowingequation: NPL b,t =α b +δ t +βIT b ·crisis+(cid:178) b,t wherebisabank(BHC),t oneyearbetween2001and2014,crisist adummyvariableindicatingyears2007to2010,α b arebankfixedeffects,andδ t areyearfixedeffects. ThedependentvariableNPL istheshareofNPLsoverassetsinb’sregulatoryfilingforyeart.IT isthepre-crisisITadoptionofbankbestimatedasdescribedinsection2. b,t b Incolumn(2)theITadoptionismeasuredbytheaveragePCsperemployeeinbankb’sbranches. Incolumn(3)theITadoptionmeasureiswinsorizedafterestimationat5 percentoneachside.Incolumn(4)theNPLsarewinsorizedat5percentoneachside.Incolumn(5)NPLsarenormalizedbytheamountofloansratherthanassets.Incolumn (6)NPLsaredefinedaccordingtoabroaderdefinition,whichincludesloanswithshorterdelinquencyperiod. Incolumn(7)wenormalizedNPLsbytheaverageamountof assetsthateachbankhadinthepre-crisisperiod(2001to2006)ratherthancontemporaneousassets. Incolumn(8)thelefthandsidearetheprovisionforcreditlossesover totalassets.Incolumn(9)weclusterstandarderrorsonlyonthebank-level.Incolumn(10)wereportbootstrapstandarderrorsbasedon500simulations.(Witheachrandom sample,wefirstre-estimatethefirststage–Equation1–toobtainanewestimateofbank-levelITadoption,andthenweestimatetheequationofinterest. Standarderrorsare thencalculatedasthestandarddeviationofthebootstrapcoefficientsofinterests.) Standarderrors(inparentheses)aredouble-clusteredonbankandyearlevelforcolumns (1)-(7).Seesubsection3.1andsection2formoredetails.*p<0.1,**p<0.05,***p<0.01 8

TableA2:LendingRegressions (1) (2) ∆Lending ∆Lending ∗∗∗ NPLpost-crisis -0.951 (0.161) ∗ IT-adoption 0.364 (0.187) R-squared 0.100 0.0115 N 336 336 Resultsofestimatingthefollowingequation: ∆Loans GFC=α+βX +(cid:178) b b b wherebisabank(BHC).Thedependentvariable∆LoansGFCistheloangrowthoverassetsinbankbregulatoryfiling(averaged b over2007to2010). X iseitherIT isthepre-crisisITadoptionofb,estimatedasdescribedinsection2ortheshareofNPLs b b overassetsinbankbregulatoryfiling(averagedover2007to2010). Seesubsection4.1andsection2formoredetails. Robust standarderrorsarereportedinparentheses.*p<0.1,**p<0.05,***p<0.01 TableA3:EnrollmentbyMajor ShareofEnrollmentbyMajor Biology Business Education Engineering Medicine Physics (1) (2) (3) (4) (5) (6) Landgrant 0.0167* -0.124*** -0.0892*** 0.191*** -0.00226 0.00730** (0.009) (0.012) (0.013) (0.017) (0.002) (0.003) R-squared 0.000726 0.0154 0.0100 0.0660 0.000103 0.000720 N 1,468 1,468 1,468 1,468 1,468 1,468 Resultsofestimatingthefollowingequation: Shareu,M =α M +β MLandgrantu +(cid:178) u,M whereuisanhigher-educationinstitution,andM isamajorofstudy. ThedependentvariableShareu,M istheratioofenrollmentinmajorM toenrollmentinalldegreesinFall2018. TheindependentvariableLandgrantu isadummyvariable thattakesthevalueoneiftheinstitutionisaland-grantcollegeandzerootherwise. Robuststandarderrorsarereportedin parentheses.*p<0.1,**p<0.05,***p<0.01 9

TableA4:SATScore SATScoreReading SATScoreMath 25thpercentile 75thpercentilescore 25thpercentile 75thpercentile (1) (2) (3) (4) Landgrant 10.19 10.31 21.61** 25.35** (8.652) (9.922) (10.922) (11.650) R-squared 0.00116 0.00112 0.00461 0.00644 N 1,144 1,144 1,144 1,144 Resultsofestimatingthefollowingequation: SAT_Scoreu =α+βLandgrantu +(cid:178) u whereuisahigher-educationinstitution. ThedependentvariableSAT_Scoreu isentrySATscoreforeithermathorreading forthe75thor25thpercentile. TheindependentvariableLandgrantu isadummyvariablethattakesthevalueoneifthe universityisaland-grantcollegeandzerootherwise.Robuststandarderrorsarereportedinparentheses.*p<0.1,**p<0.05, ***p<0.01 10

TableA5:BankHeadquarterLocation (1) (2) (3) (4) DependentVariable:HQofBank Averagedistancefromland-grantcolleges -0.0291 (0.066) Mediandistancefromland-grantcolleges 0.0178 (0.063) Distancefromclosestland-grantcollege -0.00691 (0.006) LASSOland-grantcollegeIV 0.159 (0.115) Education 0.326*** 0.323*** 0.306*** 0.328*** (0.075) (0.075) (0.077) (0.075) Income 0.0627*** 0.0632*** 0.0625*** 0.0623*** (0.005) (0.005) (0.005) (0.005) R-squared 0.146 0.146 0.147 0.147 N 3144 3144 3144 3144 Resultsofestimatingthefollowingequation: BankHQc =α+βDistance_Landgrantc +(cid:178) c wherecisacounty.ThedependentvariableBankHQcisadummythatequalsoneifoneoftheBHCofourmainsamplehasits headquarterinthecountyandzerootherwise.Distance_Landgrantciseithertheaveragedistancetoallland-grantcolleges (column1),themediandistancetoallland-grantcolleges(column(2),thedistancetotheclosestland-grantcolleges(column 3),ortheLASSOland-grantcollegeIVasdescribedinsubsection4.2andTable5in(column4).Weincludeasetofcountylevel controls"theshareofpeoplewithbachelordegrees,thelogaveragehouseholdincome,andstatefixedeffects.Robuststandard errorsarereportedinparentheses.*p<0.1,**p<0.05,***p<0.01 11