Credit access and relational contracts: An experiment testing informational and contractual frictions for Pakistani farmers

Board of Governors of the Federal Reserve System International Finance Discussion Papers ISSN 1073-2500 (Print) ISSN 2767-4509 (Online) Number 1363 December 2022 Credit access and relational contracts: An experiment testing informational and contractual frictions for Pakistani farmers M. Ali Choudhary and Anil K. Jain Please cite this paper as: Choudhary, M. Ali and Anil K. Jain (2022). “Credit access and relational contracts: An experiment testing informational and contractual frictions for Pakistani farmers,” International Finance Discussion Papers 1363. Washington: Board of Governors of the Federal Reserve System, https://doi.org/10.17016/IFDP.2022.1363. NOTE: International Finance Discussion Papers (IFDPs) 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 International Finance Discussion Papers Series (other than acknowledgement) should be cleared with the author(s) to protect the tentative character of these papers. Recent IFDPs are available on the Web at www.federalreserve.gov/pubs/ifdp/. This paper can be downloaded without charge from the Social Science Research Network electronic library at www.ssrn.com.

Credit access and relational contracts: An experiment testing informational and contractual frictions for Pakistani farmers M. Ali Choudhary∗ StateBankofPakistan& CenterforEconomicPerformance,LondonSchoolofEconomics Anil K. Jain FederalReserveBoard December2022 Abstract Credit access is limited in rural areas, especially in developing economies. Using a novel two-stage experimental design in Pakistan, first, we document that bank lending only serves a small fraction of rural credit demand. Second, we test the importanceofinformationandenforcementtechnologyfrictionsforlimitingbanklendingbyrandomlyvaryingloancontractualtermsacrossfarmersandfindthatenforcementtechnologyistheprimaryfriction. Third,usinganendlinesurvey,wedocument that farmers tend to correctly identify the financial consequences of non-repayment. ∗ThisresearchwouldnothavebeenpossiblewithoutthesubstantialgroundsupportofShafqatAliShah. WewouldliketothankAtifBajwa,AbhijitBanerjee,AndrewBird,BastianvonBeschwitz,ShahidAmjad Chaudhry,SamarHasnain,NaveedHamid,MushtaqKhan,LoganLewis,AnnieMcCrone,IjazNabi,Imran Rasul, PiruzSaboury, andseminarparticipants attheAFBC meeting(virtual), AmericanEconomic Association(virtual),ColombianCentralBank,IFABSconference(Angers),MWIEDCconference(Purdue),8th RelationalContractWorkshop(virtual),IGC-SBPConference(Karachi)andADConference(LSE,Lahore). We would like to acknowledge the exceptional research assistance from Mukhtar ul Hasan and Khurram AshfaqueBaloch. AlldatacollectionwasfundedbytheStateBankofPakistan. Thefindingsandconclusionsinthispaperaresolelytheresponsibilityoftheauthorsandshouldnotbeinterpretedasreflectingthe viewsoftheBoardofGovernorsoftheFederalReserveSystem, theviewsofanyotherpersonassociated withtheFederalReserveSystem,ortheStateBankofPakistan. 1

Fourth,ourresultssuggestonepossiblesolutiontoovercomethisfinancialfriction—a motivatedandinterlinkedintermediaryandtheuseofrelationalcontracts. JELClassification: C93,G21,O16 Keywords: Creditmarkets,banks,asymmetricinformation,contracts 2

1 Introduction Credit access, and financial inclusion more broadly, can greatly improve consumer welfare,yetuniversalaccessremainsseverelylacking,especiallyinruralareasindeveloping countries. Greater access to credit can increase income (Karlan and Zinman [2009b]), reduce inequality (Solis [2017]), increase insurance (Udry [1994]), smooth consumption (Gross and Souleles [2002]), and increase entrepreneurship (Banerjee et al. [2017]). However, credit in rural areas is limited, expensive, and unreliable in many less economically developed countries (Besley [1994]). Despite having been highlighted for decades (Braverman and Guasch [1986]), the problem of having no access to financial access still affects over half of the global population (Cull et al. [2013]). Some argue that low profitability from serving low income households limits formal financial participation but there is significant empirical evidence that this does not preclude financial access (Chaia etal.[2009]). Therefore,consideringboththevirtuesoffinancialaccessagainstthedearth of financial opportunities raises an important academic and policy question: What are thekeyfinancialobstaclestogreaterfinancialaccessandarethereinnovativemethodsto overcome these obstacles? This paper focuses on information asymmetries and enforcementfrictionsthatlimitcreditaccessinruralPakistan. The existing literature outlines various theories for the cause of the lack of formal credit accessindevelopingcountriesbuttheempiricalevidenceislimited(KarlanandMorduch [2010]). Thispaperfillsthisgapintheliteraturethrougharandomizedcontrolledtrialin ruralPakistanthatfindslargecreditdemandbutlimitedbankcreditsupply. Ourpaper’s maincontributionsarethat,inoursetting,weidentifypoorbankenforcementtechnology as a key financial friction that limits greater credit access. Moreover, our experiment suggeststhatonepossiblesolutiontoovercomethisfriction—amotivatedandinterlinked intermediaryandtheuseofrelationalcontracts. To identify reasons for limited bank lending in rural areas, we use a novel two-stage experimental design with an endline survey and three key stakeholders: a set of sugarcane farmers,anon-traditionalcreditintermediary(asugarmill),andabank. In the first stage, we document the extent of credit demand that the bank and the nontraditional intermediary (the sugar mill) would be willing to service. Specifically, we ask these institutions to screen a set of sugarcane farmers that have signalled an interest in procuring a loan. For the non-traditional intermediary, we ask whether they would be willingto guarantee abankloan tothefarmer. Forthe bank,weask whethertheywould 3

be willing to offer a direct loan to a farmer (that is, with no guarantee from any other party). To ensure accurate due diligence, we inform the non-traditional intermediary that they will be the residual claimant for a strict subset of the loans that they positively screened. In the second stage, we test the relative importance of information frictions and enforcement technology frictions for limiting credit access. Due to the close geographic, financial, and economic links between the sugar mill and the sugarcane farmers, relational contracting—informal agreements sustained by the value of the future relationships (Baker et al. [2002])—and relationship banking suggests that the non-traditional intermediary (the mill) will have superior enforcement technology (Levin [2003] and Boot[2000]). Furthermore,thetheoryofinterlinkedrelationships(BravermanandStiglitz [1982]) suggests that the mill has superior information than the bank for lending to the farmers. Conversely, given the lack of credit specialization, the non-traditional intermediaryislikelytobehamperedbyhigherfundingcosts. Inourexperiment,usingtheresultsofthefirststageoftheexperiment,wedividefarmers intogroupsaccordingtowhetherthebankwaswillingtolendtothefarmerandwhether the mill was willing to guarantee a loan to the farmer. Subsequently, we randomize the loan’s contractual terms across these groups.1 Specifically, some farmers get loans direct from the bank and some bank loans are guaranteed by the mill. By comparing the loan outcomes for farmers that are selected as creditworthy by the same institution, but get different loan contracts, we aim to isolate differences in enforcement technology. By comparing the loan outcomes for farmers that are screened as creditworthy by the different institutions, but get the same loan contract, we aim to isolate differences in screening technology. Our research methodology is inspired by Karlan and Zinman [2009a] who empirically identify moral hazard and adverse selection by randomly varying loan offer contractual terms before and after the consumer has accepted the loan terms, as well as, randomizingcontractualtermsbeforetheloan’srenewal. Inthefinalstage,weimplementanendlinesurveytouncoverthereasonsforthesuperior performance of loans that were guaranteed by the mill. Specifically, we interview the farmers to understand how the institutions tried to collect repayment and to understand 1Specifically, using the results in the first stage of the experiment, we can classify farmers into four mutuallyexclusivegroups: thosefarmersthat(i)thebankwasunwillingtolendandthemillwasunwilling toguaranteealoan,(ii)thebankwasunwillingtolendandthemillwaswillingtoguaranteealoan,(iii)the bankwaswillingtolendandthemillwasunwillingtoguaranteealoan,and(iv)boththebankwaswilling tolendandthemillwaswillingtoguaranteealoan. 4

whythefarmersdid,ordidnot,repay. Turningtoourexperiment’sresults: Inthefirststage,wefindthatthebankisonlywilling to service a small portion of the demand for loans (27 percent), and though the intermediary is willing to guarantee more farmers (52 percent), there is still a sizeable majority thatisdeemednotcreditworthybyeitherinstitution(39percent).2 In the second stage, we identify that the key financial friction limiting bank lending in rural areas is poor bank enforcement—rather than screening—technology. Specifically we find that the overdue rates were nearly 60 percentage points lower for loans where the mill was the residual claimant (that is, the mill guaranteed the bank loan to the farmer) than for those loans where the bank was the residual claimant (that is, the bank made a direct loan to the farmer. Moreover, the mill’s enforcement power led to low default rates with under five percent of those loans that were guaranteed by the mill overdue more than 90 days—a default rate that is economically sustainable for all stakeholders. Somewhatsurprisingly,wedonotfindthatthemillhadsuperiorscreeningtechnologythan the bank; specifically, repayment rates were similar for those farmers that were deemed creditworthybydifferentinstitutionsbutgotthesameloancontract. Given the large statistically and economically significant effects in stage 2 of our experiment, we consequently decided to add an endline survey of farmers. From this survey we learn that farmers perceived different costs of non-repayment according to the loan contract. For farmers with loans guaranteed by the mill, consistent with the predictions of relational contract theory, the majority of farmers stated non-repayment would damagetheirfuturerelationshipwiththemill. Whereas,forfarmerswithdirectbankloans,a relativelylargerfractionbelieveditwouldcausehigherinterestratesinthefuture. Moreover, there is a large difference between the enforcement technology utilized by the mill and the bank. To ensure the mill is not responsible for losses on loans that the mill guaranteed,themillcollectstheloanpaymentatthetimeofmillingthefarmer’ssugarcane. In contrast,thebankusesamixtureofmessagestothefarmerandcostlyvisitstothefarmer attheendoftheharvestingseason. Our paper aims to understand a key policy and academic question in empirical economics: Whyisbanklendinginruralareasindevelopingcountriessolimited? Thetheoretical literature—building from the classical papers on information asymmetries, moral hazard and adverse selection (Jaffee and Russell [1976] and Stiglitz and Weiss [1981])— 2Note these probabilities do not need to sum to 100 percent since the bank and the intermediary both denotedsomefarmersascreditworthy. 5

haveraisedanumberofpossiblecausesforthemarketfailure;including,costlyorinsufficientenforcementtechnology(HoffandStiglitz[1998]),limitedcommitment(Matthews [2001]), limited liability (Innes [1993]), differentially informed lenders (Ghosh and Ray [2016]), insufficient borrower risk capacity or tolerance (Boucher and Carter [2001] and Boucheretal.[2008]),aswellas,largetransactioncosts(Gine´ [2011]). Moreover,asnoted by Banerjee and Duflo [2011], these frictions often cause a multiplier effect, whereby the higher administrative costs cause higher interest rates, which in turn, amplify the initial friction, necessitating even larger rises in the interest rates. However, these frictions are notoriously difficult to identify in practice (Karlan and Morduch [2010]). Understanding the key cause for this problem is imperative for designing the optimal policy response giventhelargepotentialbenefitsfromcreditaccess. Ourexperimentfindsthatakeyfinancialfrictionforbanklendinginruralareasisbanks’ inability to enforce the loan contract. Further one solution that reduces this friction— andimportantlytopotentiallyeconomicallysustainablelevels—isthroughleveragingan interlinked intermediary (in this case, the sugar mill) with strong pre-existing economic relationships. The success of this financial innovation is evident with the bank, following our experiment, independently expanding the trial of using mills as loan guarantors to two other mills, and earmarking an expansion of the scheme from an initial $0.7 million to $35 million in credit disbursements. At some level, our experiment is formalising an existing trilateral lending relationship as, in a survey of moneylenders in rural Pakistan, Aleem [1990] discovers that that 30 percent of informal moneylenders’ funds come directly or indirectly from institutional sources, such as banks, wholesalers, and cotton mills. Our results contribute to the growing empirical work in developing countries that find that strategic default is a key binding constraint for lending (Blouin and Macchiavello [2019]) and demonstrate that social and economic relationships can overcome this constraint. The closest paper to our work is Bryan et al. [2015] that uses an innovative randomized control trial to separate peer information from peer enforcement effects in consumerlending. Byrandomizing(exanteandexpost)referralbonusesforloanapprovals and loan repayments, similar to our paper, they find strong peer enforcement effects but little evidence for peer information. However, the magnitude of the estimated effect is substantially smaller. Specifically, they find that a small referral bonus (less than 3 percentofthetotalloansize),whichiscontingentonthereferredborrowerrepayingtheloan, increased repayment rates by only 10 percentage points. In our setting, the bank induces 6

substantially larger incentives for the mill to successfully enforce the contract (the mill is the residual claimant for any loss on the loan contract therefore the mill’s enforcement incentive is both the loan’s principal and interest, so around 113 percent of the loan size) with much larger increases in repayment (close to 60 percentage points). Moreover, the bank’scosttoinducethesestrongerincentivesandhigherrepaymentisonly2percentof theloansize(thesizeoftheloanguaranteepaidbythebanktothemill). Our paper also contributes to the growing literature on the effectiveness of relational contracts in developing countries. Developing countries have significantly weaker economic institutions, (Acemoglu et al. [2005]), for instance, weaker creditor protections cause banks to issue smaller loans, shorter maturities, and higher interest rates (Bae and Goyal [2009]). To overcome these friction there is growing attention on the role of relational contracts. Macchiavello and Morjaria [2021] studies competition and relational contracts in the Rwandan coffee industry and find that greater competition impedes the useofrelationalcontracts,andcrucially,tosuchanextentthatallpartiesaremadeworse off. Ourpaperdirectlycomparestheeffectivenessofrelationalcontractsbetweenafarmer and a mill with a formal contract between a farmer and a bank. Consistent with the results in Macchiavello and Morjaria [2021], in our setting, we find that the relational contractbetweenthemillandthefarmerisfarsuperiortothatoftheformalcontractbetween the bank and the farmer. More generally, there is a long literature describing how relationships can overcome weak contractual protections in developing countries, such as McMillan and Woodruff [1999] in the Vietnamese informal credit market and Banerjee andDuflo[2000]intheIndiansoftwaremarket. The results of our experiment highlight one possible avenue for overcoming the lack of formal financing in rural areas in developing countries—a key problem that has been highlightedintheacademicandpolicyliterature. Forinstance,inadequateformalfinancing forces individuals to rely on less formal forms of credit such as moneylenders; trade credit;rotatingcreditandsavingsassociations(ROSCAs);andinterpersonalfinancialnetworks (Collins et al. [2009], Rutherford [2000]). However, these informal institutions often suffer from high prices, limited supply, poor resilience to financial shocks, and may exacerbateeconomicinequality.3 3In a survey of six rural villages in Kerala and Tamil Nadu, Dasgupta [1989] finds that average interestchargedbyprofessionalmoneylendersisabout52percentandprovidedalmosthalfoftotalavailable credit. Moreover, Aleem [1990] documents there are significant financial frictions for moneylenders too, withruralPakistanimoneylenders’averagecostofcapitalbeingover30percent—significantlyhigherthan the prevailing 10 percent for bank deposit rates at that time. There is significant evidence that interper- 7

In recent years, given the dearth of effective lending methods, there is interest in new mechanisms of increasing credit access, often utilizing key aspects of the borrower’s socialandprofessionalnetwork. Theleadinginnovationisthemicrofinanceindustry(with an estimate of $60 to $100 billion with 120 million clients in 2015). However, there is emerging evidence for some of the drawbacks for wider use of microfinance. Some of the innovations that facilitate cheaper financing, such as regular, frequent repayments and group repayments (Feigenberg et al. [2013]) likely reduces the benefits of credit access. Microfinance is less optimal for large, seasonal borrowings—for example, farmers borrowingforsowingseason—becauserepaymentsareusuallyrequiredtostartimmediately. Moreover, some microfinance requires group lending, and consequently does not necessarily facilitate equal access and possibly rules out some of the least socially connected individuals. Other forms of formal financing are also being trialled such as the useofcommunitychosenloan-officer(Maitraetal.[2020]),innovativeformsofassetcollateralization (Jack et al. [2016]), and greater use of non-traditional data or fintech data (Bharadwaj et al. [2019]). Our paper introduces another potential solution that leverages themonopsonypowerofamill. The rest of the paper proceeds as follows: Section (2) presents the research design of our experiment, including outlining the sample frame, borrower summary statistics, timing, and empirical strategy. Section (3) presents and analyzes the results of our experiment. Section (4) presents a summary of our findings and concluding remarks for future research. 2 Research Design Thispaperreportstheresultsfromatwo-stagefieldexperimentwithanendlinesurveyin Pakistan. In the first stage, we analyze differences in screening decisions between different institutions. In the second stage, by randomly varying the farmer’s loan contractual terms, we identify the relative importance of different institution’s information and enforcement technology for determining loan repayment. Finally, we survey the farmers to understandwhytheydidordidnotrepaytheirloans. sonalcreditandrisk-sharingnetworkshavelimitedeffectiveness,especiallyduringcorrelatedshocks(for example, natural disasters) Dercon [2002]. Finally, Blumenstock et al. [2016] finds that following a large earthquake,betterconnected(largersocialnetworksandmorecentrallylocated)andwealthierindividuals werelikelytoreceivegreaterfinancialtransfers. 8

Our experiment started in June 2016 in 60 villages in the district of Matiari, Sindh, Pakistan.4 Thisruralareais22milesfromtheclosestcityofHyderabad,withthemainoccupationbeingsugarandwheatfarming. Matiariissignificantlylessdevelopedandpoorer than the rest of Pakistan, with the average monthly household income only around Pakistani Rupee 21,000 (about 209 US dollars) in 2015—30 percent lower than the average income in Pakistan—and with a literacy rate of only 45 percent, significantly below the nationalaverageof60percent(PakistanBureauofStatistics[2015]). Our experiment uses three key stakeholders: (i) the bank, (ii) the sugar mill, and (iii) the sugarcane farmers. To obtain a baseline survey for our sampling frame, we asked all sugarcanefarmersintheMatiaridistrictiftheywouldlikeabankloanforthenextsugar growing season.5 A few key characteristics stand out. Of the 528 farmers that wanted a loan,mostfarmerssellthemajorityoftheirsugarcanecroptothissugarmill(onaverage 73percentoftheirtotalcrop),andsugarisoneoftheirmaincrops(23percentoftheircrop sales) suggesting that the sugar mill plays a substantial economic role for these farmers. Table(2)providesmoredetailsaboutthefarmers. Our sampling frame is the set of sugar farmers in Matiari who expressed an interest in takingabankloan(528farmers). 2.1 Key details on sugarcane farming and relational contracts ThestudyofthePakistanisugarcaneindustryiswellsuitedtoanalyzetheeffectivenessof relationalcontractstoovercomeinformationalandenforcementfrictionsincreditmarkets foreconomicandlogisticalreasons. The key economic reason to analyze sugarcane farming is that sugarcane mills are particularly well-suited to enforcing contracts. The high fixed cost of building mills ensures there are only a small number in any particular area (in the area we conduct our experiment, Matiari district, there are only two mills). Moreover, sugarcane must be quickly processedafterharvesting(evenifrefrigerated,sugarcanecan,atmost,bestoredfortwo weeks). Therefore, the limited competition for a farmer’s sugarcane and the necessity to quicklyselltheharvestedsugarcane,ensuresthatthemillshavesignificantmarketpower 4Weusethewordvillagetodescribewhatareknownas“deh”inthelocallanguage,Sindhi. Formally, dehareareasthathaveatleastoneofthreekeycharacteristics(i)aseparaterecord-of-rights,(ii)havebeen separatelyassessedtolandrevenue,or(iii)Pakistan’sBoardofRevenuedeclaretobeaDeh. 5TheoriginallistofsugarcanefarmersintheMatiaridistrictcamefromaStateBankofPakistan2011- 2012door-to-doorsurveyoffarmersinthatarea. 9

atthetimeofpurchase. The key logistical reason to analyze sugarcane farming in Pakistan is the time from sowing to harvesting sugarcane is relatively short (only 12 to 18 months) thereby facilitating a faster experiment that is both cheaper and more likely to have lower attrition. Moreover,theauthorshaveaccesstotheofficialPakistanicreditregistryandsubsequentlycan combinedatafromtheexperimentwithofficialfarmercredithistory. Appendix (A.1) provides additional background detail on the (i) sugar industry in Pakistan; (ii) cultivation, harvesting, and processing of sugarcane; and (iii), sugar farming in Matiaridistrict. 2.2 Experiment design: Stage 1–Testing differences in designations of creditworthiness To understand how different institutions assess credit risks, we first asked the sugar mill andthebanktoindependentlyassessthecreditworthinessofallthefarmersthatexpressed aninterestingettingaloan. The bank was asked to assess whether the bank would be willing to offer a loan (with no loan guarantee from the mill) to the farmer. The mill was asked whether they would guarantee a loan to the farmer.6 Importantly, even though both institutions are asked to assessthe“creditworthiness”ofthefarmers,thisclassificationwillvaryacrossinstitution. For the bank, creditworthiness is defined as whether the farmer would repay the bank. Forthemill,creditworthinessisdefinedaswhetherthefarmerwillrepaythebankconditionalonthemillguaranteeingtheloan;thatis,wouldthefarmerbewillingtopotentially renegeontherelationalcontractbetweenthefarmerandthemill. Toensureaccurateassessments,themillwasinformedthattheywouldguaranteeastrict subset of those farmers they said they were willing to guarantee (by randomizing the set of farmers that got a loan guarantee we can identify the importance of the institution’s enforcementandinformationtechnology—thefocusofsection(2.3)). To ensure consistency across the experiment—apart from the loan size and whether the loanwasguaranteedbythemill—theloancontract’stermswerestandardacrossallfarm- 6Theresultsofthisbankandmillinvestigationwerenotsharedwiththefarmer; rather,dependingon the bank’s and the mill’s assessment and the subsequent randomization, the farmer was either offered a specificloancontractornoloan(moredetailsinthenextsection). 10

ers. The farmer was charged an interest rate of 13 percent on each loan, irrespective of whether the loan was guaranteed by the mill. If the mill guaranteed a loan, the bank would pay a 2 percent fee (of the loan amount) to the mill, and the mill would repay the bank if the loan was not repaid. Therefore, if the farmer took a $100 loan from the bank that was guaranteed by the mill, at the end of the loan term the farmer would owe $113 tothebank. Iftheloanwasfullyrepaid,themillwouldreceive$2fromthebank,andthe bank would earn a net return of $11. If the farmer did not repay the loan, the mill would paythebank$111,andthebankwouldstillearnanetreturnof$11.7 Toidentifythefarmercharacteristicsthatareimportantforeachinstitution’sclassification ofafarmer’screditworthiness,werunregression(1)separatelyforeachinstitutiontype. Creditworthy = β Farmercharacteristics +(cid:101) (1) lf l f lf Where‘Creditworthy’isabinaryvariable,definedaswhetherthebank(mill)waswilling to offer a loan (guarantee a bank loan) to farmer f. The institution (either the bank or the mill)isdenotedby l. To explore similarities and differences across the institution’s choice, we define four mutually exclusive groups: group (A) “creditworthy only mill ” (farmers that were solely def notedascreditworthybythemill),group(B)“creditworthyboth ”(farmersthatweresolely f denoted as creditworthy by the mill and the bank), group (C) “creditworthy only bank ” f (farmers that were solely denoted as creditworthy by the bank), and group (D) “creditworthy neither ” (farmers that were denoted by both the bank and the mill as not creditf worthy). Figure (1) illustrates the possible set of each borrower. Out of the 528 farmers, 185 farmers were in group (A), 88 farmers in group (B), 51 farmers in group (C) and 204 farmersingroup(D). 2.3 Experiment design: Stage 2—Testing asymmetric information and differing enforcement technology The second stage of our experiment varied the loan contractual terms across farmers to identify the relative importance of (i) different institution screening technology and (ii) 7Analternativeexplanationisthatifthemillwasguaranteeingtheloanforthefarmer,thecontractual net payments for the mill are identical to the scenario where the mill borrowed $100 from the bank at an interestrateof11%,andsubsequentlylent$100tothefarmerataninterestrateof13%. 11

Figure1: Possibleclassificationofeachfarmerbetweenthemillandthebank D Sample frame: 528 Both the bank and the mill classified the farmer as not creditworthy A B C Only the mill Both the Only the bank classified the bank and the classified the farmer as mill classified farmer as creditworthy the farmer as creditworthy creditworthy Note: The528farmersthatrequestedaloanwasscreenedbyboththemillandthebank. Farmers could be allocated to four mutually exclusive groups: (A) only the mill defined the farmer as creditworthy (185 farmers), (B) both institutions defined the farmer as creditworthy (88 farmers), (C)onlythebankdefinedthefarmerascreditworthy(51farmers),or(D)noinstitutiondefinedthe farmerascreditworthy(204farmers). institutionenforcementtechnologytodetermineloanrepaymentoutcomes. To explore the effect on repayment, using the results from the willingness to lend experiment (stage 1), we randomized loan contract offers to farmers depending on their group (stratifying on the number of acres each farmer planned to grow sugarcane). The ideal experimentwouldrandomizecontracts(directloanfromthebankorguaranteedloanby the mill) across all possible groups (A) to (D). However, in practice, we were only able to partiallyrandomizecontractsacrossthesegroups. In the interest of greater experimentation, the bank was willing to offer some loans to farmersthattheyhadpreviouslydenotedasnotcreditworthy. Specifically,inadditionto thefarmersthatthebankdeemedcreditworthy,thebankwasalsowillingtoofferadirect bank loan to farmers that the mill classified as creditworthy. However, the bank was unwillingtolendtofarmersthatneitherthemillorthebankhaddenotedascreditworthy. Unfortunately, the mill was only willing to guarantee loans that the mill had previously 12

denotedascreditworthy. Overall,farmersingroups(A)and(B)receivedbothloancontractsandfarmersingroup (C) received only direct bank loans. Since both the bank was unwilling to directly lend andthemillunwillingtoguaranteeloanstofarmersingroup(D),theywereomittedfrom the rest of the study. For reference, tables (A.11) and (A.12) in Appendix (A.2) report the definitionofeachgroup. Across the various treatment arms, the bank was willing to lend a total of approximately $700,000 in direct and guaranteed loans. Given that the initial interest in receiving loans waslargerthantheavailablefundsforlending,werandomizedonwhetherfarmerswere actually offered a loan, and if so, the loan’s contractual terms. To ensure the allocation of farmers across treatments were more likely to be balanced ex-post, we stratified farmers on the number of acres that the farmer planned to plant with sugarcane. Since we were unable to randomize contracts to farmers in group C—thereby limiting the potential inference—we selected a greater fraction of farmers in groups A and B, than group C. In total, just over 90 percent of farmers that were screened as creditworthy in groups A and B were offered a loan, and just over 50 percent of farmers that were screened in groupCwereofferedaloan. Intotal,asdescribedlaterinmoredetail,around75percent offarmersacceptedtheloanoffers. Overall, following the randomization of loan offers, there were five mutually exclusive groups that took loans. From the set of borrowers that the mill were willing to guarantee (groupA),oneset(groupA1)wereofferedbankloanswithaloanguaranteefromthemill andasecondset(groupA2)wereoffereddirectloansfromthebank. NotethatgroupA2 were not classified as creditworthy farmers by the bank—a key part of our experimental design. Fromthesetoffarmersthatboththemillandthebankclassifiedascreditworthy (group B), one set (group B1) were offered bank loans with a loan guarantee from the mill, and a second set (group B2) were offered direct loans from the bank. Finally, the set of borrowers that only the bank classified as creditworthy (group C) were offered only directbankloans. Figure(2)illustratesthepossiblegroups,andforreference,table(A.12) inAppendix(A.2)repeatsthedefinitions. Tominimizetheex-antedifferencesacrossourtreatmentgroupswestratifiedoursample on the number of acres each farmer planned to grow sugarcane. As evidence for successful randomization, tables (A.14) and (A.15) in the appendix show that there were no statistically significant differences within each group (that is, between farmers in groups 13

A1andA2,andbetweenfarmersingroupsB1andB2). An additional key statistical concern is selective attrition. As is common in randomized control trials, even though some farmers were approved to receive a bank loan, some farmers—following the randomization—decided to not exercise their bank loan option. We find no evidence for selective attrition. First, table (A.16) in the Appendix finds that farmersthattooktheloanweremostlysimilaronobservables(onmeasuressuchasassets, farm size) to the farmers that did not take the loan but were offered a loan. Second, our experiment had high take-up rates with close to 75 percent of farmers taking the offered loan. Finally, table (A.17) in the Appendix shows that the loan take-up rate across the various treatment groups is similar. Moreover, the Chi-squared test at the 10 percent significancelevelisunabletorejectthenullhypothesisthatthetake-upratesarethesame acrosstreatmentgroups. Overall, a total of 204 loans were disbursed with an average loan size of $3,400—a relatively large sum given that the average monthly household income in this area was $209 (PakistanBureauofStatistics[2015]). 2.3.1 Empiricalspecifications Enforcementtechnology To test the importance of each institution’s enforcement technology, we compare the repaymentratesforthosefarmersthatweredefinedascreditworthybythesameinstitution but got different loan contracts. Our research design is similar to Karlan and Zinman [2009a]’s seminal paper but tests a different economic question. Given our research design,wecanundertaketwosimilartestsbutondifferentsetsoffarmers. First, we analyse differences in repayment rates for the farmers defined as creditworthy by onlythe mill, therefore holding the credit screening technology constant, butreceived different contracts (groups A1 and A2). Specifically, using the group of growers which the mill exclusively selected we compare the repayment rates between (i) farmers that received a guaranteed loan by the mill (group A1) and (ii) farmers that received a direct bankloan(groupA2). Second,weanalysedifferencesinrepaymentratesforthecommon set of farmers passing the creditworthy criterion of both the mill and the bank (groups B1 andB2),specifically(i)farmersthatreceivedaguaranteedloanbythemill(groupB1)and (ii) farmers that received a direct bank loan (group B2). Specifically, in table (4), we run 14

Figure2: ExperimentalDesign NO Not part of the Would you like a loan? experiment. YES (528) Would [Bank/Mill] be willing to give a loan or offer a guarantee? BANK MILL BANK MILL BANK MILL BANK MILL Group A Group B Group C Group D Loan (185) (88) (51) (204) guaranteed (G) G NG NL G NG NL NG NL Loan not NL guaranteed (NG) A1 A2 B1 B2 C No loan (NL) (75) (54) (38) (25) (12) Note: The numbers in parenthesis are the number of farmers in each group. For example, there were528farmerswhowereinterestedinabankloan,andofthese528farmers,185farmerswere classifiedascreditworthybythemillbutnotbythebank(GroupA).88farmerswereclassifiedas creditworthy by the mill and the bank, and 51 were classified as creditworthy by only the bank. The farmers in groups A, B, and C, were randomly allocated to three possible treatments: direct bankloan(loannotguaranteed),bankloanguaranteedbythemill,ornoloan. Giventhatneither thebanknorthemillclassifiedthefarmersingroupDascreditworthy,noneofthesefarmerswere offeredloans. Somefarmerseventhoughtheyinitiallydescribedinterestintakingaloan,decided againsttakingaloan,asdescribedinsection2.3. regressionssimilartothefollowingformontheselectedsamples: Overdue = β Loanguarantee +β Farmercharacteristics +(cid:101) (2) f E f 1 f f where “Overdue ” is a binary variable equal to zero farmer f repaid the bank on time f and one otherwise. “Loan guarantee ” is a dummy variable equal to one if the farmer f receivedaloanguaranteefromthemill. The coefficient of interest, β , estimates whether those farmers that were selected to re- E ceive a loan guarantee had higher default rates than those farmers that received a direct bank loan. This regression tests the hypothesis that the mill’s loan enforcement 15

technology—viatheloanguarantee—issuperiortothebank’senforcementtechnology. Screeningtechnology To test the importance of each institution’s screening technology, we compare the repayment rates between those farmers who received the same loan contract—therefore holding constant the enforcement ability—but were classified as creditworthy by different institutions. Importantly, even though both institutions are asked to assess the creditworthiness of the farmers, this classification will vary across institution. For the bank, creditworthinesswillmeanwillthefarmerrepayadirectloan. Forthemill,creditworthiness will be whether the farmer will repay the bank conditional on the mill guaranteeing the loan; that is, would the farmer be willing to potentially renege on the relational contract between the farmer and the mill. By comparing borrowers that were selected by different institutions but were offered the same loan terms, we can identify the relative effectivenessofeachinstitution’sscreeningtechnology. First, for farmers that received a guaranteed bank loan, we analyse differences in repayment rates between (i) farmers who only the mill defined as creditworthy (group A1) and (ii) farmers who both the bank and mill defined as creditworthy (group B1). Second, we conduct a symmetric test and compare repayment rates for the farmers who got a direct bankloan(groupsA2,B2,andC)butweredefinedascreditworthybydifferentinstitutions. Specifically, in table (6), we run regressions similar to the following form on the selected samples: Overdue = β CreditworthyonlyMill +β Farmercharacteristics +(cid:101) (3) f S f 1 f f where “Overdue ” is a binary variable equal to zero if farmer f repaid the bank on time f and one otherwise. “Creditworthy only Mill ” is a dummy variable equal to one if only f themillselectedfarmer f ascreditworthy,andzerootherwise. Thecoefficientofinterest, β ,estimateswhetherthosefarmersthatwereselectedascred- S itworthy by the mill had lower default rates than those farmers that were selected by the bank. This regression tests the hypothesis that the mill’s loan screening technology is superiortothebank’stechnology. Finally,intable(7),weconductapooledregressionthatexaminestheeffectofthescreen- 16

ing technology but controls for the different loan contract. By controlling for the loan contract, we can increase our sample size and consequently increase the power of our test. 2.4 Experiment design: Stage 3–Endline survey Following the completion of our experiment, we were struck by the large effects that our experiment uncovered. To explore these results in more detail, we conducted an additional survey that asked the farmers that received loans for why they did or did not repaytheloan,andspecificallyaskedquestionsonwhattheyexpectedtheconsequences of not repaying the loan would be. In March 2020, we tried to interview all farmers that took a loan, and we successfully managed to interview (via telephone) 128 farmers, of which60farmersreceivedaguaranteedloanfromthesugarmilland68farmersreceived adirectloanfromthebank.8 2.5 Timing The experiment started in mid-2016 and ended in early-2020, and figure (3) shows the timeline of the experiment. In June 2016, we interviewed all sugar farmers that grew sugarcane in the district of Matiari to assess whether they were interested in availing a new bank loan. In total we interviewed 1,455 sugar farmers. Of these 1,455 farmers only 528 farmers were interested in taking a bank loan. Of the farmers that were interested in aloan,weconductedabaselinesurvey. In July and August 2016, the bank and the mill simultaneously (and independently) screenedthesetoffarmersthatwereinterestedintakingaloan. By August 26 2016, we had received the lending decisions of the bank and the mill, and subsequently randomized loan contracts across the appropriate farmers on the basis of the baseline surveys. By September 15 2016, farmers were notified of the decision of whether they would receive a loan, and received the details of the loan contract where applicable. Thegrowersweregiven15daystodecideonwhethertoaccepttheloanoffer. Loans were disbursed from late September 2016 through February 2017 (that is, before sugarcane is sowed) with a due date for March 31 2018 (that is, a tenor of approximately 8Asuccessfulcallbackrateofover60percent. 17

18 months and after the harvesting season for sugarcane). Finally, in March 2020, we surveyedfarmersontheirreasonsforrepayingornotrepayingtheirloans. Figure3: Timelineoftheexperiment Aug 2016: Researchers Jun 2016: Baseline randomize farmers across March 2020: survey of 1455 sugar groups, stratifying on acres Sep 2016-Feb 2017: Final survey growers of sugarcane. Loans dispersed implemented Jul-Aug 2016: Bank and the Sep 2016: Farmers Mar 2018: Loan mill simultaneously screen decide whether to take due date all 528 farmers that the loan. requested loan. 2.6 Data Our paper utilizes three key sources of data: hand-collected survey data, credit registry data from the State Bank of Pakistan’s e-CIB database (this data has also been used in Khwaja and Mian [2005] and Choudhary and Jain [2020]), and administrative data providedbythesugarmill. Toprocuremoredataonthefarmersinoursample,weconductedsurveysinthreewaves; first,todetermineinterestforcredit,weconductedatelephonicsurveyofallsugarfarmers in Matiari; second, to understand the characteristics of those farmers that wanted a loan, we solicited information on farmers’ assets, crops, and demographics. Third, we conductedanendlinesurveytolearnmoreaboutwhyfarmersdidordidnotrepay.9 To ascertain better knowledge on farmers’ access to credit, we matched each farmer to their corresponding credit registry entry in Pakistan’s national credit registry, e-CIB, run by the State Bank of Pakistan. If farmers had previously held any formal credit in the last five years, their details would be in the registry. Of the 528 farmers in our sample, only 34 percent (180 farmers) had entries in the Pakistan’s credit registry, reinforcing the 9AllcostsrelatedtodatacollectionandsurveyingwerefundedbytheStateBankofPakistan. 18

evidence for a lack of formal credit markets in rural Pakistan. If a farmer’s details are in the registry, we match details on total amount of credit the farmer previously had, and whetherthefarmerrepaidtheloanontime. Finally,mostfarmershaveapre-existingrelationshipwiththemill,sowesupplementthe farmer’s survey data on the farmer’s sales to the mill in previous years. We use information on the length of the farmer’s relationship with the mill, the fraction of produce that is sold to the mill, the value of produce sold to the mill, and the distance of the farmer to themilltocompletetheevidenceforthefarmer’srelationshipwiththemill. 3 Results In section (3.1), we start by examining the differences in the mill and the bank’s determination of a farmer’s creditworthiness. We explore what characteristics of the farmer that makethebankandthemillmorelikelytolend. Insection(3.2),weanalyzethedifference inrepaymentratesfollowingourrandomizationofcontracttypesacrossfarmers. 3.1 Stage 1: Screening technology To examine how the set of farmers who are judged to be creditworthy across different institutions with different information and enforcement technology, we asked each institution to independently assess the creditworthiness of each farmer that had expressed interest in procuring a loan. We proceed in three steps. First, we outline significant aggregate differences in the set of farmers each institution chooses. Second, we document differencesinfarmercharacteristicsacrossthedifferentselectedgroups. Finally,wemore formallytestthekeyfarmercharacteristicsthatarecorrelatedwithwhethertheinstitution definedthefarmerascreditworthy. Starting with the aggregate differences in the institutions’ choices, in table (1), we immediatelyidentifythatthesetoffarmersthatweredefinedascreditworthybetweenthetwo institutionssignificantlydifferedinsizeandcomposition. The mill was significantly more willing to be the residual claimant on loans to farmers thanthebank. Outofthesetfarmersthatexpressedaninterestinaloan,themilldefined almost double the number of farmers as creditworthy (group A and group B) than the 19

bank(groupBandgroupC).10 Additionally, the bank and the mill evaluate farmers on significantly different characteristics. Therewasalargenumberoffarmers(235farmers)thatonlyoneinstitutiondefined as creditworthy (group A and group C). Whereas, only 88 farmers were selected as creditworthybyboththebankandthemill(groupB). Table1: Sizeofeachgroup Group A B C D Definedascreditworthyby: Onlythemill Both Onlythebank Neither Numberoffarmers 185 88 51 204 Fractionoffarmers 35% 17% 10% 39% Observations 528 GroupAisthesetoffarmersthatonlythemilldefinedascreditworthy,groupBisthesetoffarmersthat boththemillandthebankdefinedascreditworthy,groupCisthesetoffarmersthatonlythebankdefined ascreditworthy,andgroupDisthesetoffarmersneitherinsitutiondefinedascreditworthy. Turning to the differences in farmer characteristics across each group, in table (2), we report summary statistics for farmers according to which institutions defined the farmer ascreditworthy. Inthefinalcolumn(“total”),wereportaggregatestatisticsforthesample offarmerswhoexpressedaninterestinprocuringaloan. The starkest differences in the characteristics of farmers that each institution defined as creditworthyareapparentwhencomparinggroupsAandC—thefarmersthatwereonly chosen by one institution. The bank was significantly more likely to lend to farmers that had larger farms and owned expensive farming equipment (so more likely to be richer) and have good credit history. Whereas, the mill was more likely to guarantee loans to farmersthatsoldrelativelyandabsolutelymoreoftheiroutputtothemill(somorelikely to have a deeper economic relationship and more likely to be able to sustain a relational contract). Surprisingly, the length of the farmer’s relationship with the mill (the number of years selling to the mill) was lower for the farmers selected by the mill than for the bank. 10Consistentwiththedefinitionsusedintheearlierpartofthepaper,groupAarefarmersthatonlythe milldefinedascreditworthy,groupBarefarmersthatboththemillandthebankdefinedascreditworthy, groupCarefarmersthatonlythebankdefinedascreditworthy,andgroupDarefarmersthatnoinstitution definedascreditworthy. 20

Table2: Baselinecharacteristicsandsummarystatistics(mean)bygroup A B C D Total Sugarcaneplanted(acres) 15.14 24.69 23.48 16.03 17.46 Incomefromagriculture(percent) 85.88 82.58 89.81 86.58 85.93 Incomefromsugar(percent) 21.97 22.53 21.15 22.83 22.47 Relativevalueofcropsales(decile) 4.381 6.580 6.333 5.718 5.456 Educ. belowhighschool(Y=1;N=0) 0.470 0.506 0.481 0.560 0.519 Farmsize(acres) 21.15 42.75 49.53 33.10 31.62 Yearssellingtothemill 11.95 22.22 21.22 11.86 13.96 Salestothemill(pctoftotalsales) 76.77 85.69 35.72 42.71 59.78 Within5kmofthemill(Y=1;N=0) 0.571 0.840 0.481 0.393 0.523 Relativesalestothemill(decile) 6.399 7.531 4.593 3.881 5.278 Formalcredithistory(Y=1;N=0) 0.298 0.432 0.444 0.345 0.348 Prev. bankloanoverdue(Y=1;N=0) 0.0357 0.0370 0.0000 0.0159 0.0246 Own-tractor(Y=1;N=0) 0.190 0.309 0.407 0.345 0.294 Own-thresher(Y=1;N=0) 0.185 0.321 0.296 0.302 0.267 Own-blade(Y=1;N=0) 0.208 0.346 0.407 0.357 0.311 Own-cultivator(Y=1;N=0) 0.214 0.321 0.407 0.357 0.309 Own-raja(Y=1;N=0) 0.214 0.346 0.444 0.353 0.313 Own-gobal(Y=1;N=0) 0.208 0.309 0.370 0.357 0.303 Own-bundmaker(Y=1;N=0) 0.143 0.247 0.259 0.246 0.214 Own-harvestingmachine(Y=1;N=0) 0.0298 0.0494 0.0741 0.0159 0.0284 Observations 528 Note: GroupAisthesetoffarmersthatonlythemilldefinedascreditworthy,groupBisthesetoffarmers that both the mill and the bank defined as creditworthy, group C is the set of farmers that only the bank defined as creditworthy, and group D is the set of farmers no one defined as creditworthy. The final column shows the aggregate value across all farmers. Table (A.13) in Appendix (A.2) describes the variable 21 definitions.

To formally test the differences in the mill and the bank’s choice of creditworthy farmers, we regress the determination of the farmer’s creditworthiness by each institution on the farmer’s observable characteristics in table (3). In column 1 (column 2) we regress whether the farmer was designated as creditworthy by the mill (bank) on a set of farmer characteristics and in the final column we compute the average difference between these twomeasures.11 The results in table (3) are broadly similar with the results in table (2). The bank was relatively more willing to lend to farmers that had observable characteristics that are highly likely to be correlated with greater farmer wealth; for instance, the bank positivelyscreenedfarmer’swithhighcropsales,largefarmsizes,andownedexpensivefarm machinery. In contrast, consistent with theory of relational contracts, relatively more important characteristics for the mill were those that suggested a deeper or more valuable partnership with the mill. Therefore, those farmers who were closer to the mill, or sold more of their crop to the mill were relatively more likely to be classified as creditworthy by the mill. Surprisingly, those farmers that had had been selling to the mill for a greater time were relatively more likely to be defined as creditworthy by the bank. One possible explanation for this counter-intuitive result is that the farmers that have been growing sugar the longest are also likely the richest farmers, who are relatively preferred by the bank. 11To be precise, we do the following regression Creditworthy = β Farmercharacteristics + lf 1 f β Farmercharacteristics ∗mill +(cid:101) , and report the coefficients β in table (3). As the model is satu- 2 f l lf 2 rated, the coefficient β is the exact difference between the regression coefficients reported in columns 1 2 and2. Tocalculatethet-statisticsincolumn3,weclusterthestandarderrorsatthefarmer-level. 22

Table 3: Analyzing the observable characteristics that determine the farmer’s creditworthiness (1) (2) (3) Mill Bank Difference Sugarcaneplanted(acres) 0.001 0.001 0.000 (0.001) (0.001) (0.001) Valueofcropsales(decile) -0.024∗∗∗ 0.006 -0.030∗∗∗ (0.008) (0.007) (0.011) Educ. belowhighschool(Y=1;N=0) -0.011 -0.018 0.007 (0.038) (0.032) (0.048) Farmsize(acres) -0.001∗∗ 0.001 -0.002∗∗ (0.000) (0.001) (0.001) Yearssellingtothemill 0.002 0.029∗∗∗ -0.028∗∗∗ (0.003) (0.002) (0.004) Salestothemill(pctoftotalsales) 0.001 0.000 0.001 (0.001) (0.001) (0.001) Within5kmofthemill(Y=1;N=0) 0.137∗∗∗ 0.052 0.085 (0.041) (0.032) (0.053) Relativesalestothemill(decile) 0.070∗∗∗ 0.003 0.067∗∗∗ (0.010) (0.009) (0.014) Formalcredithistory(Y=1;N=0) 0.020 0.028 -0.008 (0.039) (0.035) (0.051) Prev. bankloanoverdue(Y=1;N=0) 0.005 -0.201∗∗∗ 0.205∗∗ (0.081) (0.063) (0.096) Own-tractor(Y=1;N=0) -0.028 0.300∗∗∗ -0.328∗∗ (0.124) (0.094) (0.156) Own-thresher(Y=1;N=0) 0.131 -0.079 0.210∗ (0.086) (0.074) (0.119) Own-blade(Y=1;N=0) -0.096 0.049 -0.145 (0.121) (0.138) (0.184) Own-cultivator(Y=1;N=0) -0.008 -0.248 0.239 (0.158) (0.170) (0.231) Own-raja(Y=1;N=0) -0.029 0.291∗∗ -0.320 (0.154) (0.124) (0.218) Own-gobal(Y=1;N=0) -0.004 -0.316∗∗∗ 0.313∗ (0.147) (0.107) (0.178) Own-bundmaker(Y=1;N=0) -0.047 -0.080 0.033 (0.075) (0.070) (0.104) Own-harvestingmachine(Y=1;N=0) 0.062 0.043 0.019 (0.108) (0.132) (0.184) Constant 0.159∗∗∗ -0.250∗∗∗ 0.409∗∗∗ (0.056) (0.041) (0.063) Observations 528 528 1056 Standarderrorsinparentheses ∗ p <0.10,∗∗ p <0.05,∗∗∗ p <0.01 Note: Theresultincolumn1(column2)istheregressionofthemill’s(bank’s)classificationofthefarmer’s creditworthinessonfarmercharacteristics.Thestandarderrorsincolumns1are2robuststandarderrors.In column3,wereportthedifferencebetweenthesetwocoefficients. Standarderrorsarecalculatedusingthe 23 followingregressionCreditworthy = β Farmercharacteristics +β Farmercharacteristics ∗mill +(cid:101) lf 1 f 2 f l lf andareclusteredatthefarmerlevel. Table(A.13)inAppendix(A.2)describesthevariabledefinitions.

3.2 Stage 2: Contractual frictions To preview our main result, figure (4) shows overdue rates by treatment arm. Across all arms, the striking result is direct bank loans (the orange bars) had substantially greater overdue rates than the loans that were guaranteed by the mill (the blue bars). The average overdue rate for direct bank loans was over 65 percent, yet, for loans that were guaranteed by the mill, the average overdue rate was just over 4 percent. This result demonstratesthatthebankfacessevereenforcementfrictions. Therestofthissectionformallyanalyzesthisresultingreaterdetail. First,insection(3.2.1),weanalyzecontractual frictions by testing how farmers that were selected by the same institution but were randomly assigned different contracts affected farmer repayment. Second, in section (3.2.2), we analyze information frictions by testing how farmers that received the same contract butwereselectedbydifferentinstitutionsaffectedfarmerrepayment. Figure4: Overdueratesbycontractandscreenedgroup Note: This graph shows overdue rates (percent) for each treatment arm in our study. The orange bars correspondtooverdueratesforloansreceiveddirectlyfromthebank,andthebluebarsforloansthatwere guaranteedbythemill. Thekeyinferencefromthisfigureisthatdirectbankloanshadsubstantiallyhigher overdueratesthantheloansthatwereguaranteedbythemill. 24

3.2.1 Enforcementfrictions To understand whether the mill has superior enforcement technology than the bank, we test for whether the repayment rates are relatively higher for those farmers that were guaranteed by the mill. To ensure we are identifying the effect of differences in institutions’ enforcement technology, we compare repayment rates between farmers that receiveddifferentcontracts(differentresidualclaimantonloanproceeds)butwereinitially selectedascreditworthybythesameinstitution. Intable(4),wereporttheresultsfromregressingrepaymentonwhetherthefarmer’sloan wasguaranteedbythemill,andadditionalcontrols.12 Columns1and2restrictattention to those farmers that were identified as creditworthy by only the mill (that is, farmers in group A). Columns 3 and 4 restrict attention to those farmers that were identified as creditworthy by both the mill and the bank (that is, farmers in group B). Columns 2 and 4includecontrolsforfarmercharacteristics. As previewed in figure (4), there is strong evidence that the mill has superior enforcement technology than the bank. Across all four specifications, those farmers that had a loanguaranteefromthemillwerearound60percentagepointslesslikelytobeoverdueon their loan than those farmers that received a direct bank loan. The high rate of overdue loansforbanks(andsignificantlyhigherthantheinterestrateontheloans)illustratesthe severity of the banks’ problem to offer rural loans. Moreover, the results are robust to including additional controls for farmer characteristics (columns 2, 4, and 6), separately estimating the results on the various classifications of farmer creditworthiness—farmers thatwereidentifiedascreditworthyonlybythemill(columns1and2),byboththebank andthemill(columns3and4),andestimatingtheresultsonallfarmersthatwereidentifiedascreditworthybythemill(columns5and6). Interestingly,therearenosignificantdifferencesintherepaymentbehaviorbetweenthose farmers that were solely selected by the mill, and those farmers that were selected by both the mill and the bank, after controlling for contract type. This lack of difference in repayment behaviour is evident by the small coefficient estimate for “creditworthyonly mill”, an indicator variable for whether the farmers were in group A—that is, those farmers that were only selected by the mill. Therefore, this surprising result suggests that from the set of borrowers chosen by the mill, the bank does not choose borrowers that are, on average, more likely to repay the bank. Specifically, you may expect that the 12Moredetailontheempiricalstrategyisprovidedearlierinsection(2.3.1). 25

bank’sscreeningtechnologywouldfacilitatethebanktochoosehigherqualityborrowers leadingtohigherrepayments. Table4: Doesthemillhavesuperiorenforcementtechnologythanthebank? (1) (2) (3) (4) Overdue Overdue Overdue Overdue LoanGuarantee -0.59∗∗∗ -0.62∗∗∗ -0.57∗∗∗ -0.55∗∗∗ (0.071) (0.070) (0.10) (0.10) Sugarcaneplanted(acres) -0.0034 -0.00091 (0.0024) (0.0014) Relativevalueofcropsales(decile) 0.018 -0.027 (0.017) (0.021) Educ. belowhighschool(Y=1;N=0) -0.0075 -0.015 (0.071) (0.084) Farmsize(acres) 0.00071 0.0012 (0.0011) (0.0011) Yearssellingtothemill -0.010∗∗ -0.020∗ (0.0047) (0.011) Salestothemill(pctoftotalsales) 0.00032 -0.0042∗∗ (0.0011) (0.0019) Within5kmofthemill(Y=1;N=0) 0.018 0.19 (0.068) (0.12) Relativesalestothemill(decile) 0.0032 0.0041 (0.020) (0.027) Formalcredithistory(Y=1;N=0) 0.016 0.071 (0.075) (0.080) Prev. bankloanoverdue(Y=1;N=0) 0.019 -0.063 (0.058) (0.096) Observations 129 129 63 63 Groups A1&A2 A1&A2 B1&B2 B1&B2 Standarderrorsinparentheses ∗ p <0.10,∗∗ p <0.05,∗∗∗ p <0.01 Note: Table(A.13)inAppendix(A.2)describesthevariabledefinitions. Havingfoundthatthemillhassuperiorenforcementtechnology,weinvestigateforwhich farmers the mill’s enforcement technology is superior. To do so, in table (5), we supplement the regressions in table (4) with additional independent variables. Specifically, we interact the dummy variable ‘whether a farmer had a loan guarantee’, with the farmer’s characteristics. Overall, the regressions suffer from a lack of power, which is not surpris- 26

ing given the small sample size. Nonetheless, there are two notable results. The mill was relativelybetteratenforcingrepaymentbythosefarmersthatplantedlesssugarcane(coefficient on the regressor “Guarantee X Sugarcane planted” is consistently positive and statisticallysignificantinsomespecifications)andthosefarmersthatownedlargerfarms (coefficient on the regressor “Guarantee X Farm size” is consistently negative and statistically significant in some specifications). These results suggest that the mill has superior technology for those farmers with larger farms, and for whom sugar is a small part of the farmer’s crop portfolio. Somewhat surprisingly, we do not see significant coefficients on some of the variables that you may expect to be correlated with the mill’s superior technology, namely, those variables that suggest a stronger relationship between the mill and the farmer (such as the fraction of sales to the mill, short distance from the mill, and thetenureoftherelationshipwiththemill). 27

Table5: Forwhichfarmersdoesthemillhavesuperiorenforcementtechnology? (1) (2) (3) Overdue Overdue Overdue ∗∗ GuaranteexSugarcaneplanted 0.0057 0.0054 0.0087 (0.0071) (0.0057) (0.0036) ∗∗∗ GuaranteexRelativevalueofcropsales(decile) -0.023 0.16 0.013 (0.038) (0.047) (0.030) GuaranteexEduc. belowhighschool 0.075 -0.00045 -0.057 (0.16) (0.21) (0.13) ∗∗∗ ∗∗∗ GuaranteexFarmsize(acres) -0.0019 -0.0074 -0.0042 (0.0027) (0.0016) (0.0015) GuaranteexYearssellingtothemill 0.0074 0.036 0.0068 (0.011) (0.030) (0.0086) ∗∗∗ GuaranteexSalestothemill(pctoftotalsales) -0.0013 0.0095 0.0016 (0.0026) (0.0033) (0.0020) GuaranteexWithin5kmofthemill -0.13 0.060 -0.12 (0.16) (0.30) (0.13) GuaranteexRelativesalestothemill(decile) 0.014 -0.030 0.0041 (0.047) (0.052) (0.036) GuaranteexFormalcredithistory -0.048 -0.014 -0.0047 (0.17) (0.16) (0.13) CreditworthyonlyMill -0.042 (0.066) ∗∗ ∗∗∗ ∗∗∗ Constant 0.66 2.91 0.96 (0.26) (0.60) (0.25) Observations 129 63 192 Groups A B A&B AdditionalControls Yes Yes Yes Standarderrorsinparentheses ∗ p <0.10,∗∗ p <0.05,∗∗∗ p <0.01 Note: We have interacted the farmer characteristics with an indicator variable for whether the farmer receivedaloanguaranteefromthemill;thesevariablesareprefixedwiththeword’Guarantee’,forinstance, ”GuaranteexAcresSugarcane”referstothevariable”LoanGuarantee”interactedwiththevariable”Acres Sugarcane”. Inalltheregressionsweincludeadditionalcontrolvariables,specifically,weincludethenoninteractedvariables(thatis,Loanguarantee,acressugarcane,valueofcropsales,etc.) butintheinterestof spaceandsimplicitywedonotreportthecoefficientresults. Table(A.13)inAppendix(A.2)describesthe variabledefinitions. 28

3.2.2 Screeningtechnology: Informationfrictions To understand whether the mill has a superior information about farmer’s creditworthiness than the bank, we test for whether the repayment rates are relatively higher for those farmers selected by the mill than for the bank. To ensure we are identifying the effectofdifferencesininstitutions’screeningtechnology,wefirstcomparerepaymentrates between farmers that received the same contract but were selected as creditworthy by differentinstitutions. In table (6), we report the results from regressing repayment on whether the farmer was chosen by the mill, and additional controls.13 Columns 1 and 2 restrict attention to those farmers that received a guaranteed loan from the mill and were selected by the mill as creditworthy(thatis,farmersingroupsA1andB1). Columns3and4restrictattentionto thosefarmersthatreceivedadirectbankloanandwereselectedbythemillascreditworthy (that is, farmers in groups A2 and B2). Finally, columns 5 and 6 restrict attention to all the farmers that received a direct bank loan (that is, farmers in groups A2, B2, and C). Columns2,4,and6,includecontrolsforfarmercharacteristics. Overall, we find there is little economic or statistical significance for the mill possessing superior information on farmer’s creditworthiness than the bank. Across the set of regressions in table (6), the estimated casual effect from the mill’s greater information on repaymentratesvariesaroundzero,withthehighestestimatedeffectbeingincreasingrepayment by nearly eight percentage points (column 6) or reducing repayment by nearly fivepercentagepoints(column3). Analyzing the results for the control variables, it is evident that across the regressions there was little systemic differences by farmer characteristics. There are only two variables that are both statistically significant in some specifications and consistent in the coefficient’s sign across the various specifications. The farmers that plant more sugarcane (”acres sugarcane”) and have been selling for more years to the mill (”years selling to the mill”), on some specifications, had statistically significant higher repayments rate, suggestingthesefarmersare—observationally—morecreditworthy. 13Moredetailontheempiricalstrategyisprovidedearlierinsection(2.3.1). 29

Table6: Doesthemillhavesuperiorinformationtechnologythanthebank? (1) (2) (3) (4) (5) (6) Overdue Overdue Overdue Overdue Overdue Overdue CreditworthyonlyMill 0.027 -0.044 0.048 -0.038 -0.028 -0.078 (0.037) (0.030) (0.12) (0.16) (0.10) (0.15) Sugarcaneplanted(acres) -0.00030 -0.0090∗∗ -0.0042 (0.00028) (0.0037) (0.0041) Relativevalueofcropsales(decile) 0.0049 -0.0076 0.0067 (0.012) (0.028) (0.023) Educ. belowhighschool(Y=1;N=0) -0.0058 0.051 -0.012 (0.041) (0.13) (0.12) Farmsize(acres) -0.00020 0.0040∗∗ 0.0023∗∗ (0.00017) (0.0015) (0.0011) Yearssellingtothemill -0.0067∗ -0.013 -0.014 (0.0037) (0.0090) (0.0086) Salestothemill(pctoftotalsales) 0.00026 -0.0014 -0.0011 (0.00074) (0.0019) (0.0018) Within5kmofthemill(Y=1;N=0) -0.031 0.084 0.12 (0.049) (0.13) (0.11) Relativesalestothemill(decile) 0.0063 0.0021 -0.014 (0.015) (0.033) (0.026) Formalcredithistory(Y=1;N=0) 0.0065 0.012 0.036 (0.049) (0.12) (0.11) Prev. bankloanoverdue(Y=1;N=0) -0.018 (0.026) Observations 113 113 79 79 91 91 Groups A1&B1 A1&B1 A2&B2 A2&B2 A2,B2&C A2,B2&C Additionalcontrols No Yes No Yes No Yes Contract: LoanGuarantee Yes Yes No No No No Contract: Directbankloan No No Yes Yes Yes Yes Standarderrorsinparentheses ∗ p <0.10,∗∗ p <0.05,∗∗∗ p <0.01 Note: Table(A.13)inAppendix(A.2)describesthevariabledefinitions. Table (7) builds on the regressions in table (6). We increase the power of our tests by pooling the farmers across their assigned contract. Specifically, we include an indicator variable (”Loan Guarantee”), that is “one” if the loan was guaranteed by the mill. Columns 1 and 2 only include those farmers that the mill identified as creditworthy, 30

whereas, columns 3 and 4 include all farmers. The results in table (7) are consistent with the results in table (6), that is, holding the contract fixed, there is no statistical evidence that those farmers selected as creditworthy by the mill were more likely to repay their loans. In all the regressions, the coefficient for the indicator variable ”creditworthy only mill”isnotstatisticallysignificantandthecausalestimateisclosetozero. 31

Table7: Pooledregressions: Doesthemillhavesuperiorinformationtechnologythanthe bank? (1) (2) (3) (4) Overdue Overdue Overdue Overdue CreditworthyonlyMill 0.036 -0.038 0.0016 -0.058 (0.053) (0.064) (0.051) (0.063) ∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ LoanGuarantee -0.59 -0.59 -0.62 -0.61 (0.058) (0.057) (0.054) (0.054) Sugarcaneplanted(acres) -0.0023 -0.0018 (0.0016) (0.0013) Relativevalueofcropsales(decile) 0.0042 0.0089 (0.013) (0.012) Educ. belowhighschool(Y=1;N=0) 0.012 -0.0047 (0.053) (0.052) Farmsize(acres) 0.00097 0.00083 (0.00065) (0.00053) ∗∗ ∗∗ Yearssellingtothemill -0.0090 -0.0092 (0.0041) (0.0041) Salestothemill(pctoftotalsales) -0.00061 -0.00069 (0.00089) (0.00087) Within5kmofthemill(Y=1;N=0) 0.053 0.059 (0.058) (0.058) Relativesalestothemill(decile) -0.0052 -0.012 (0.016) (0.014) Formalcredithistory(Y=1;N=0) 0.026 0.027 (0.056) (0.054) Prev. bankloanoverdue(Y=1;N=0) 0.011 0.0093 (0.040) (0.039) Observations 192 192 204 204 Groups A&B A&B A,B&C A,B&C Contract Both Both Both Both Standarderrorsinparentheses ∗ p <0.10,∗∗ p <0.05,∗∗∗ p <0.01 Note: Table(A.13)inAppendix(A.2)describesthevariabledefinitions. 32

3.3 Endline survey: Survey exploring the large differences in repayment rates In this section, we present the findings from an endline survey that investigates why farmers didor did notrepay the loan. For farmers torepay their loansthey must beboth able and willing to repay the loan. If farmers’ inability to pay (for example, due to a poor sugar harvest) was the primary reason for high default rates, we would expect to see similardefaultratesacrosstherandomlyallocatedcontractualtreatmentsbecausethere’s no reason to think that the farmers that were randomly allocated to the direct bank loan would have different abilities to pay. However, in section (3.2.1), we found strikingly large differences across treatment groups, specifically farmers repaid the bank markedly less often than the mill. Taking these results together suggests that the net benefits of repaying the mill must be higher than repaying the bank. This section explores possible reasonsforthisresult. 14 3.3.1 Differencesincollectiontechnology Table(8)presentstheresultsofthequestion,”Howdidthelenderencouragerepayment?” As is quickly evident, the sugar mill and the bank used substantially different methods toencouragerepayment. Thesugarmillleverageditspositioninthefarmer’sproduction process to procure repayment at the time of milling the sugarcane. Whereas, the bank relied on predominantly calling the farmer (57 percent of all loans), and in some cases alsovisitingthefarmer(anadditional38percentofloans). Visiting individual farmers is expensive because (i) the bank branch staff is located in a nearbycityofHyderabadsome22miles(35kilometers)awayfromthesugarmilland(ii) the sugar farmers are on average disbursed over 17 miles radius (30 kilometers) and the terrain is poorly connected with narrow dirt roads. For example, the remotest farm from the sugar mill is located at 28 miles (45 kilometers). Even though the bank visited many farmers,insection(3.2.1),wefoundthatfarmerswere60percentagepointsmorelikelyto be overdue on direct bank loans than loans guaranteed by the mill. Therefore, even with the careful attention of the bank loan officer, the loan officer was unable to get high—or even close to profitable—repayment rates. In contrast, the mill’s collection process was 14As described earlier, this part of the experiment was added after the experiment after observing a sufficientperiodofnonrepayment,aswewantedtofurtherexplorethereasonsfarmersdefaulted. 33

relatively inexpensive as the mill collected repayments directly from the farmer at the timeofmillingthesugarcane. Table8: Endlinesurveyquestion: Howwaspaymentcollected? Bank(mean) Mill(mean) Diff. StandardError ∗∗∗ CollectedatMill 0.00 0.92 -0.92 0.04 ∗∗∗ In-personVisitandCall 0.38 0.05 0.33 0.08 ∗∗∗ OnlyCall 0.57 0.03 0.54 0.08 Other 0.05 0.00 0.05 0.03 Observations 100 T-testsignificanceleveldenotedby*p<0.05,**p<0.01,***p<0.001 Note: This table presents the results from an endline survey that was administered to farmers after the duedateoftheirloans. Thefirsttwocolumnspresentthefractionsoffarmersthatreportedhowthebank (column1)orthemill(column2)collectedtheloanrepayments. Thecolumnlabelled‘Bank’reportsresults fromthosefarmersthatreceivedadirectbankloan(thatis,thosefarmersingroupsA2,B2,andC)andthe columnlabelled‘Mill’reportsresultsfromthosefarmersthatreceivedabankloanthatwasguaranteedby themill(thatis,thosefarmersingroupsA1andB1). Standarderrorsforthet-testarecomputedassuming thatthevariancesofthevariableofinterestmaydifferbetweenthebankandthemill. 3.3.2 Statedreasonsforrepayingontime Table (9) presents the results of the question, ”Why did you repay on time?” Overall, farmersstatedrelativelysimilarreasonsacrossbothtreatmentswithmostfarmersstating the importance of ‘responsible behavior’, that is, the farmers’ felt that repaying a loan is an obligation because it is the ‘right’ thing to do. In terms of some of the differences, relativelymoreofthefarmerswithloansguaranteedbythemillsuggestedthattheywere financially capable of repaying. Interesting, even though the mill enforced the farmer contract by collecting the loan repayment at the time of milling the sugar (see table (8)), very few (less than ten percent of farmers that received a loan guaranteed by the mill) farmersnotedthatwasthereasontheyrepaidontime. 34

Table9: Endlinesurveyquestion: Whydidyourepayontime? Bank(mean) Mill(mean) Diff. StandardError ∗ Responsiblebehaviour 0.62 0.40 0.22 0.10 ∗ Goodcropandfinancialliquidity 0.05 0.22 -0.17 0.07 Futurerelationship 0.12 0.20 -0.08 0.07 Financialcost 0.14 0.09 0.05 0.07 ∗ Deductedatmilling 0.00 0.09 -0.09 0.04 Other 0.07 0.00 0.07 0.04 Observations 97 T-testsignificanceleveldenotedby*p<0.05,**p<0.01,***p<0.001 Note: Thistablepresentstheresultsfromanendlinesurveythatwasadministeredtofarmersafterthedue date of their loans. The column labelled ‘Bank’ reports results from those farmers that received a direct bank loan (that is, those farmers in groups A2, B2, and C) and the column labelled ‘Mill’ reports results fromthosefarmersthatreceivedabankloanthatwasguaranteedbythemill. Standarderrorsforthet-test arecomputedassumingthatthevariancesofthevariableofinterestmaydifferbetweenthebankandthe mill. 3.3.3 Theperceivedconsequencesofnotrepaying Table (10) presents the results of the question: ”What do you think the consequences of not repaying the loan would be?” Interestingly, if the farmer did not repay the loan guaranteedbythemill,theyperceivedthemaincostasjeopardizingtheirfuturerelationshipwiththemill,whereas,thosefarmerswithadirectbankweremoreconcernedabout higher financing costs in the future. Not surprisingly, given the limited effectiveness of the legal system in rural Pakistan, in both treatments, the threat of legal actions was not ratedhighly.15 The results to this question reiterate the importance of relational lending. Farmers with loans that were guaranteed by the mill were clearly cognizant of the potential hit to their relationship with the mill if they were to default on the loan. By the mill guaranteeing the loan, the mill was bundling its production process and intermediating in the credit process. The farmers’ answers to this question show that the farmers were aware of this bundlingandthepotentialadverseimpactontheirrelationshipwiththemill. Incontrast, 15Djankovetal.[2003]documentsthedifficultiesofevictingatenantinvariouslegalsystems.InPakistan, theyfoundthatittook365daystoevictatenant,versusamedianof180daysinlegalsystemswithsimilar origins,or49daysintheUnitedStates. 35

those farmers with direct bank loans were relatively more concerned with potential adverse financial impacts from not repaying the loan (higher future interest rates, and/or penalties). Table 10: Endline survey question: What do you think the consequences of not repaying theloanwouldbe? Bank(mean) Mill(mean) Diff. StandardError ∗ Jeopardizefuturerelationship 0.44 0.70 -0.25 0.10 ∗∗ Higherfutureinterestrates/penalities 0.44 0.15 0.29 0.09 Legalactions 0.12 0.15 -0.04 0.07 Observations 98 T-testsignificanceleveldenotedby*p<0.05,**p<0.01,***p<0.001 Note: Thistablepresentstheresultsfromanendlinesurveythatwasadministeredtofarmersafterthedue date of their loans. The column labelled ‘Bank’ reports results from those farmers that received a direct bank loan (that is, those farmers in groups A2, B2, and C) and the column labelled ‘Mill’ reports results fromthosefarmersthatreceivedabankloanthatwasguaranteedbythemill. Standarderrorsforthet-test arecomputedassumingthatthevariancesofthevariableofinterestmaydifferbetweenthebankandthe mill. 3.3.4 Summaryofthefinalsurvey Taking the results togetherfrom the final survey we learn that the millhas two comparative advantages in ensuring high repayments. First, the mill is able to leverage its role in milling the sugarcane to physically collect repayments in an efficient and cheap manner. Second, finance emphasises the importance of relational lending to ensure repayment; in this context, it’s clear that the farmers highly value the importance of their relationship withthemill,anddonotwanttojeopardizetheirinterlinkedrelationship. 4 Conclusion On a promising note for increasing credit provision, leveraging less traditional credit intermediaries may increase credit efficiency and improve credit access, specifically for large productive investments. The non-traditional intermediary (the mill) was willing to guaranteeloanstoanadditional184farmersthatthebankwasunwillingtolendto—more 36

than doubling the number of farmers that were eligible for loans. We find that the mill’s key advantage over traditional forms of lending for improving credit access is the mill’s superior enforcement technology. Moreover, the mill’s enforcement technology was sufficiently effective that the trilateral credit relationship between the farmer, the mill, and thebankwasbothincentive-compatibleandeconomicallysustainable. Intheexperiment, boththemillandthebankwereabletomakepositiveprofitsandthefarmerreceivedlow interest rates.16 The success of this financial innovation is evident with the bank, following our experiment, expanding the trial of using mills as loan guarantors to other two other mills, and earmarking an expansion of the scheme from the initial $0.7 million to $35millionincreditdisbursements. Intermsofscreening,somewhatsurprisingly,themillandthebankhadsimilareffectiveness at screening farmer’s creditworthiness, that is, the mill did not seem to have better screeningtechnologyeventhoughthefarmerhashadsignificantlymoreinteractionwith themillthanthebank. On a less promising note for credit provision, a substantial number of farmers were classified as not ‘creditworthy’. Out of 528 farmers that requested loans more than 200 were not selected as creditworthy by either institution. Moreover, this statistic is likely an underestimate for the number of farmers that would be excluded because the high level of loan defaults for the bank suggests that the bank’s selection criteria was insufficiently rigorousgiventhebank’senforcementcapabilities. The key additional finding is that banks struggle with high farmer defaults even from farmers with large farms or have more education. The key constraint seems the banks’ limitedcapacityorhighcosttoenforcethecreditcontract. Additionalpoliciesandmethods that might alleviate this constraint are worth exploring; prior literature have examinedwhetheradditionalpledgeablecollateral,suchasincreasingpropertyrights,orusing socialcapitalmayincreasethefarmers’incentivetorepay,butmorecreativesolutionsare worthconsidering. Oneavenuethatisgaininggreaterattentionisthegreateruseoftransactionaldata—andadigitalfootprintmorebroadly(seeBergetal.[2020],Bharadwajetal. [2019],Frostetal.[2019]). We should also touch upon some of the drawbacks of our study. Our experiment mostly 16The rate of overdue loans on the guaranteed loans was slightly above the fee earned by the mill— therefore there was no net profit for guaranteeing the loans for the mill (and likely some additional administrativecosts). However,guaranteeingtheloansindirectlybooststhemill’sprofitbecauseitlikelyled farmerstogrowmoresugarcane,inturn,leadingtoadditionalincomeforthemillfromthemill’smargin fromprocessingthefarmer’ssugarcaneintosugar 37

only examined a one-shot game, that is, a big Pakistani bank offered loans, but did not explicitlyofferfuturebankloans. Akeyfindingofthecreditliteratureisthatrelationship lending (see Boot [2000] for an overview), specifically the promise of a future loans, increasesconsumerrepayment. Therefore,ourexperimentmayhaveledtoahigherdefault rate than we may have in equilibrium because we did not explicitly link the repayment of this bank loan to future bank loans. In mitigation, the farmers that successfully repaid (either the direct bank loan or the guaranteed bank loan) would have a positive mark in their credit report that should increase their ability to procure future loans. Moreover, the results of the experiment’s stage 3 demonstrated that borrowers are cognizant of the potentialconsequencesfromnon-repayment. Our study may also over-estimate the number of borrowers the bank was willing to lend to. A key design choice of our experiment was that the bank was required to do due diligence on all the borrowers that requested a bank loan. It is highly possible that the screening cost for all farmers that request a loan is too high given the potential expected returns from those borrowers that actually procure loans, thereby causing a missing financialmarket. 38

References D. Acemoglu, S. Johnson, and J. A. Robinson. Institutions as a fundamental cause of long-rungrowth. Handbookofeconomicgrowth,1:385–472,2005. I. Aleem. Imperfect information, screening, and the costs of informal lending: a study of aruralcreditmarketinpakistan. TheWorldBankEconomicReview,4(3):329–349,1990. K.-H. Bae and V. K. Goyal. Creditor rights, enforcement, and bank loans. The Journal of Finance,64(2):823–860,2009. G. Baker, R. Gibbons, and K. J. Murphy. Relational contracts and the theory of the firm. TheQuarterlyJournalofEconomics,117(1):39–84,2002. A. V. Banerjee and E. Duflo. Reputation effects and the limits of contracting: A study of theindiansoftwareindustry. TheQuarterlyJournalofEconomics,115(3):989–1017,2000. A. V. Banerjee and E. Duflo. Poor economics: A radical rethinking of the way to fight global poverty. PublicAffairs,2011. A.V.Banerjee,E.Breza,E.Duflo,andC.Kinnan. Docreditconstraintslimitentrepreneurship? heterogeneityinthereturnstomicrofinance. HeterogeneityintheReturnstoMicrofinance(September1,2017).GlobalPovertyResearchLabWorkingPaper,(17-104),2017. T. Berg, V. Burg, A. Gombovic´, and M. Puri. On the rise of fintechs: Credit scoring using digitalfootprints. TheReviewofFinancialStudies,33(7):2845–2897,2020. T.Besley. Howdomarketfailuresjustifyinterventionsinruralcreditmarkets? TheWorld BankResearchObserver,9(1):27–47,1994. P. Bharadwaj, W. Jack, and T. Suri. Fintech and household resilience to shocks: Evidence from digital loans in kenya. Technical report, National Bureau of Economic Research, 2019. A. Blouin and R. Macchiavello. Strategic default in the international coffee market. The QuarterlyJournalofEconomics,134(2):895–951,2019. J. E. Blumenstock, N. Eagle, and M. Fafchamps. Airtime transfers and mobile communications: Evidenceintheaftermathofnaturaldisasters. JournalofDevelopmentEconomics, 120:157–181,2016. 39

A. W. Boot. Relationship banking: What do we know? Journal of Financial Intermediation, 2000. S. R. Boucher and M. R. Carter. Risk rationing and activity choice in moral hazard constrainedcreditmarkets. Technicalreport,2001. S. R.Boucher, M.R. Carter, andC. Guirkinger. Riskrationing and wealtheffects incredit markets: Theory and implications for agricultural development. American Journal of AgriculturalEconomics,90(2):409–423,2008. A.BravermanandJ.L.Guasch. Ruralcreditmarketsandinstitutionsindevelopingcountries: Lessons for policy analysis from practice and modern theory. World development, 14(10-11):1253–1267,1986. A. Braverman and J. E. Stiglitz. Sharecropping and the interlinking of agrarian markets. TheAmericanEconomicReview,72(4):695–715,1982. G. Bryan, D. Karlan, and J. Zinman. Referrals: Peer screening and enforcement in a consumercreditfieldexperiment. AmericanEconomicJournal: Microeconomics,7(3):174–204, 2015. A. Chaia, A. Dalal, T. Goland, M. J. Gonzalez, J. Morduch, and R. Schiff. Half the world isunbanked,2009. M. A. Choudhary and A. K. Jain. How public information affects asymmetrically informed lenders: Evidence from a credit registry reform. Journal of Development Economics,2020. D.Collins,J.Morduch,S.Rutherford,andO.Ruthven. Portfoliosofthepoor: howtheworld’s poorliveon$2aday. PrincetonUniversityPress,2009. R. Cull, R. J. Cull, A. Demirgu¨c¸-Kunt, and J. Morduch. Banking the World: Empirical FoundationsofFinancialInclusion. MITPress,2013. A.Dasgupta. Reportsoninformalcreditmarketsinindia: Summary. NewDelhi: National InstituteofPublicFinanceandPolicy,1989. S. Dercon. Income risk, coping strategies, and safety nets. The World Bank Research Observer,17(2):141–166,2002. 40

S.Djankov,R.LaPorta,F.Lopez-deSilanes,andA.Shleifer. Courts. TheQuarterlyJournal ofEconomics,118(2):453–517,2003. B. Feigenberg, E. Field, and R. Pande. The economic returns to social interaction: Experimental evidence from microfinance. Review of Economic Studies, 80(4):1459–1483, 2013. J. Frost, L. Gambacorta, Y. Huang, H. S. Shin, and P. Zbinden. Bigtech and the changing structureoffinancialintermediation. EconomicPolicy,2019. P.GhoshandD.Ray. Informationandenforcementininformalcreditmarkets. Economica, 83(329):59–90,2016. X. Gine´. Access to capital in rural thailand: An estimated model of formal vs. informal credit. JournalofDevelopmentEconomics,96(1):16–29,2011. D. B. Gross and N. S. Souleles. An empirical analysis of personal bankruptcy and delinquency. Reviewoffinancialstudies,15(1):319–347,2002. K. Hoff and J. E. Stiglitz. Moneylenders and bankers: price-increasing subsidies in a monopolistically competitive market. Journal of Development Economics, 55(2):485–518, 1998. R. Innes. Financial contracting under risk neutrality, limited liability and ex ante asymmetricinformation. Economica,pages27–40,1993. W. Jack, M. Kremer, J. De Laat, and T. Suri. Borrowing requirements, credit access, and adverseselection: Evidencefromkenya.Technicalreport,NationalBureauofEconomic Research,2016. D. M. Jaffee and T. Russell. Imperfect information, uncertainty, and credit rationing. The Quarterly Journal of Economics, 90(4):651–666, 1976. URL http://qje. oxfordjournals.org/content/90/4/651.short. D. Karlan and J. Morduch. Access to finance. In Handbook of development economics, volume5,pages4703–4784.Elsevier,2010. D.KarlanandJ.Zinman. Observingunobservables: Identifyinginformationasymmetries with a consumer credit field experiment. Econometrica, 77(6):1993–2008, 2009a. URL http://onlinelibrary.wiley.com/doi/10.3982/ECTA5781/abstract. 41

D. Karlan and J. Zinman. Expanding credit access: Using randomized supply decisions toestimatetheimpacts. ReviewofFinancialstudies,pagehhp092,2009b. A. I. Khwaja and A. Mian. Do lenders favor politically connected firms? rent provision in an emerging financial market. The Quarterly Journal of Economics, 120(4):1371–1411, 2005. J.Levin. Relationalincentivecontracts. AmericanEconomicReview,93(3):835–857,2003. R. Macchiavello and A. Morjaria. Competition and relational contracts in the rwanda coffeechain. TheQuarterlyJournalofEconomics,136(2):1089–1143,2021. P. Maitra, S. Mitra, D. Mookherjee, and S. Visaria. Decentralized targeting of agricultural credit programs: Private versus political intermediaries. Technical report, National BureauofEconomicResearch,2020. S. A. Matthews. Renegotiating moral hazard contracts under limited liability and monotonicity. JournalofEconomicTheory,97(1):1–29,2001. J.McMillanandC.Woodruff. Interfirmrelationshipsandinformalcreditinvietnam. The QuarterlyJournalofEconomics,114(4):1285–1320,1999. PakistanBureauofStatistics. Pakistansocialandlivingstandardsmeasurement. 2015. S.Rutherford. Thepoorandtheirmoney. OxfordUniversityPressNewDelhi,2000. A.Solis. Creditaccessandcollegeenrollment. JournalofPoliticalEconomy,125(2):000–000, 2017. J. E. Stiglitz and A. Weiss. Credit rationing in markets with imperfect information. The Americaneconomicreview,71(3):393–410,1981. C.Udry. Riskandinsuranceinaruralcreditmarket: Anempiricalinvestigationinnorthernnigeria. TheReviewofEconomicStudies,61(3):495–526,1994. 42

A Appendices A.1 Background on the sugar industry SugarinPakistan Sugar is a major global commodity that is produced in over 100 countries and is a material component of Pakistan’s economy. The sugar industry in Pakistan consists largely of growing sugarcane and the production, manufacture, and marketing of white sugar. In Pakistan, the sugar manufacturing industry has expanded substantially. At the time of independence in 1947, there were only two sugar mills in Pakistan. Following the setup of a government commission in 1957, the industry started to expand substantially, reaching around 50 mills by 1990 and then further expanding to 89 mills. All told, the daily crushing capacity has increased in 1990 from 90,000 metric tons to its current capacity of 505,000 tons. Sugarcane is the second most important cash crop in Pakistan. Its production contributes just over 3 percent in agriculture’s value addition and just under 1 percentofPakistanigrossdomesticproduct. Sugarcanefarmingisprotectedbythegovernmentwithaminimumpriceforthefarmers. This policy plays an important role in persuading growers to cultivate sugarcane over othercrops. Province level data shows that Punjab is the largest sugarcane producing province in Pakistan but production in Sindh is, on average, more efficient. Sugarcane production in Sindh province—where the experiment takes place—has a cultivation area less than half ofPunjab. However,productioninSindhisgenerallymoreefficientwithsugarcaneyield per hectare and the recovery ratio (the ratio of the amount of nutrient in the harvested croptotheamountofnutrientapplied)higherthanPunjab. Cultivation,harvestingandprocessingofsugarcane Therearefivedistinctstepsintheproductionofsugarfromsugarcane. The first stage of the cultivation is tilling or weeding the land. If sugarcane was planted in the previous year, the land solely needs to be weeded. Otherwise, the land must be tilled,whichisusuallydonewithanownedorrentedtractor. The second stage is planting and caring for the plants. During the growth process the 43

sugarcane plant receives care in the form of pesticide protection, fertilizer, and water. Sugarcanestrawscangrowuptotwometresinabouttwelvetoeighteenmonths The third stage is harvesting. The harvesting of sugarcane in Pakistan is generally done byhiringlocalmanuallabor. Sugarcaneisharvestedwhentheplantisaroundtwometers high,atwhichpointtheplantiscutneartherootallowingthestemtoregrowforafuture crop. The fourth stage is the transportation and sale of the sugarcane to the sugar mill. In Pakistan, the sugarcane is manually loaded onto trailers or trucks. Once at the mill, the sugarcaneisweighedandsold. The fifth stage is the milling of the sugarcane. To begin, the sugarcane is crushed into a liquid. Subsequently, this liquid is heated and filtered to form sugar crystals, which is then followed by centrifugation to remove excess water. Finally, the refined sugar is sorted,packaged,anddistributed. SugarfarminginMatiaridistrict,SindhProvince In the district of Matiari—the district where the experiment is conducted—sugarcane is grown under well-irrigated conditions where water is available either through surface canals or underground pumping facility using tube wells. Moreover, the district is well situatedinthefavorablesubtropicalweatherconditionsandgoodannualrainfall. Atthetimeofoursurvey,therearetwosugarmills,MatiariSugarMill—theintermediary in the experiment—and Mehran Sugar Mill, that are located close to the farmers in our experiment. The two mills were only 30 kilometers apart. In our survey area, typically, a mill would buy sugarcane from farmers in the radius of 25-30 kilometers. Similar to other areas in Pakistan, in Matiari, sugarcane farming competes with other cash crops such as, wheat, cotton, onion and chilies. In discussions with farmers, farmers report they prefer growing sugarcane for several reasons. First, the guaranteed government minimum support price. Second, easy access to multiple buyers (two sugar mills in our case. Third, lower manual effort to grow sugarcane relative to other cash crops, such as cotton. Fourth,sugarcaneplantisrelativelyresilienttoweathershocks. Finally,thereisa well-developedlocalmarketforrequiredinputs,suchasamanuallabormarket. 44

A.2 Variable and group definitions TableA.11: Groupdefinitionsforstage1oftheexperiment Group Definition A Screenedascreditworthybyonlythemill(Creditworthyonlymill) B Screenedascreditworthybyboththemillandthebank(Creditworthyboth) C Screenedascreditworthybyonlythebank(Creditworthyonlybank) D Screenedascreditworthybynoinstitution(Notcreditworthy) TableA.12: Groupallocationsforstage2oftheexperiment Loanguaranteedbymill Directbankloan (1) (2) Group Creditworthyonlymill (A) A1 A2 Creditworthyboth (B) B1 B2 Creditworthyonlybank (C) omitted C Notcreditworthy (D) omitted omitted As part of the experiment design, the mill was only willing to guarantee those farmers that the mill had screened as creditworthy (hence guaranteed bank loans were only offered to some farmers in groups A and B). Whereas, the bank was willing to offer direct bank loans to any farmer that had been denoted as creditworthybythebankorthemill(hencedirectbankloanswereofferedtosomefarmersingroupsA,B, andC). 45

TableA.13: Variabledefinitions Variable Valuetype Description Loanguarantee Binary Labels the type of contract the farmer receives: ”0”equalsdirectbankloanand”1” equalsloanguaranteedbythemill Sugarcaneplanted(acres) Continuous Totalnumberofacresofsugarcaneplanted bythefarmerin2016-17 Income from agriculture Continuous be- Farmer’sincomefromagricultureasaper- (percent) tween0and100 centoftotalincome Income from sugar (per- Continuous be- Farmer’s income from sugar as a percent cent) tween0and100 oftotalincome Relative value of crop Discrete, values A (decile) ranking of each farmers’ total sales(deciles) from1to10 cropsalesinrupees,with10beingthemost sales Educ. belowhighschool Binary Education level of the farmer: ”1” if the level of the farmers’ highest education is belowhighschool,”0”otherwise Farmsize(acres) Continuous Total number of acres owned for plantationbythefarmerin2016-17 Yearssellingtothemill Discrete Number of years the farmer has been sellingtothemill Sales to the mill (percent Continuous be- Farmer’s sales (by value) to the mill as a oftotalsales) tween0and100 percentoftotalfarmersales. Within5kmofthemill Binary Value equal to ”1” if the farmer’s home is within 5 Km miles of the mill, ”0” otherwise Relative sales to the mill Discrete, values A(decile)rankingofeachfarmers’fraction (decile) from1to10 of sales of sugarcane to the mill, with 10 beingthemostsales Own-”farmequipment” Binary Labelswhetherthehouseholdowns”farm equipment”: ”1”yes,”0”otherwise. Formalcredithistory Binary Labelswhether thefarmerappears inPakistan’s formal credit registry: ”1” equals formalcredithistoryexists,”0”otherwise. Prev. bankloanoverdue Binary Labels whether the farmer had a previous overdue loan in Pakistan’s formal credit registry: ”1”equalsyes,”0”otherwise. 46

A.3 Baseline characteristics and balance between treatment arms TableA.14: BaselinecharacteristicsandbalancebetweenfarmersingroupsA1andgroup A2 Diff. Sugarcaneplanted(acres) 0.991 (0.28) Incomefromagriculture(percent) -2.284 (-0.72) Incomefromsugar(percent) -2.542 (-1.04) Relativevalueofcropsales(decile) -0.733 (-1.60) Educ. belowhighschool(Y=1;N=0) -0.0374 (-0.47) Farmsize(acres) -9.831 (-1.87) Yearssellingtothemill 0.576 (0.46) Salestothemill(pctoftotalsales) -3.404 (-0.62) Within5kmofthemill(Y=1;N=0) -0.0567 (-0.72) Relativesalestothemill(decile) -0.465 (-1.20) Formalcredithistory(Y=1;N=0) -0.0482 (-0.67) Prev. bankloanoverdue(Y=1;N=0) 0.0151 (0.30) Own-tractor(Y=1;N=0) 0.0307 (0.49) Own-thresher(Y=1;N=0) -0.0338 (-0.56) Own-blade(Y=1;N=0) 0.00103 (0.02) Own-cultivator(Y=1;N=0) -0.00887 (-0.14) Own-raja(Y=1;N=0) -0.0337 (-0.52) Own-gobal(Y=1;N=0) 0.00103 (0.02) Own-bundmaker(Y=1;N=0) -0.0142 (-0.26) Own-harvestingmachine(Y=1;N=0) 0.0250 (0.86) Observations 168 tstatisticsinparentheses ∗ p <0.05,∗∗ p <0.01,∗∗∗ p <0.001 Thistablereportsdifferencesinsummarystatistics(mean)ofvariousfarmercharacteristicsbetweenfarmers in group A1 and group A2 (where a positive number indicates a larger number for farmers in group A1). 47

TableA.15: BaselinecharacteristicsandbalancebetweenfarmersingroupsB1andB2 Diff. Sugarcaneplanted(acres) 2.473 (0.32) Incomefromagriculture(percent) 1.807 (0.41) Incomefromsugar(percent) -4.868 (-1.68) Relativevalueofcropsales(decile) 0.694 (1.12) Educ. belowhighschool(Y=1;N=0) 0.196 (1.75) Farmsize(acres) 1.297 (0.11) Yearssellingtothemill 0.0459 (0.04) Salestothemill(pctoftotalsales) -10.85 (-1.49) Within5kmofthemill(Y=1;N=0) -0.0446 (-0.52) Relativesalestothemill(decile) 0.207 (0.40) Formalcredithistory(Y=1;N=0) -0.0944 (-0.84) Prev. bankloanoverdue(Y=1;N=0) -0.0612 (-1.77) Own-tractor(Y=1;N=0) -0.0453 (-0.43) Own-thresher(Y=1;N=0) 0.0376 (0.35) Own-blade(Y=1;N=0) -0.00319 (-0.03) Own-cultivator(Y=1;N=0) -0.0140 (-0.13) Own-raja(Y=1;N=0) 0.0485 (0.44) Own-gobal(Y=1;N=0) 0.0580 (0.54) Own-bundmaker(Y=1;N=0) 0.0568 (0.56) Own-harvestingmachine(Y=1;N=0) -0.0300 (-0.64) Observations 81 tstatisticsinparentheses ∗ p <0.05,∗∗ p <0.01,∗∗∗ p <0.001 48

Table A.16: Baseline characteristics and balance between the farmers approved for loans whotookloansandthosefarmersthatdidnottakeloans Diff. Sugarcaneplanted(acres) 6.778 (1.41) Incomefromagriculture(percent) 3.469 (1.31) Incomefromsugar(percent) -0.0952 (-0.05) Relativevalueofcropsales(decile) 0.683 (1.69) Educ. belowhighschool(Y=1;N=0) 0.00572 (0.08) Farmsize(acres) 9.170 (1.47) Yearssellingtothemill 0.472 (0.40) Salestothemill(pctoftotalsales) -3.422 (-0.66) Within5kmofthemill(Y=1;N=0) 0.0531 (0.82) Relativesalestothemill(decile) -0.0547 (-0.15) Formalcredithistory(Y=1;N=0) -0.0621 (-0.97) Prev. bankloanoverdue(Y=1;N=0) 0.0310 (0.68) Own-tractor(Y=1;N=0) 0.0613 (1.00) Own-thresher(Y=1;N=0) 0.000817 (0.01) Own-blade(Y=1;N=0) 0.0507 (0.81) Own-cultivator(Y=1;N=0) 0.0743 (1.18) Own-raja(Y=1;N=0) 0.0784 (1.23) Own-gobal(Y=1;N=0) 0.0891 (1.42) ∗ Own-bundmaker(Y=1;N=0) 0.126 (2.14) Own-harvestingmachine(Y=1;N=0) 0.0212 (0.71) Observations 276 tstatisticsinparentheses ∗ p <0.05,∗∗ p <0.01,∗∗∗ p <0.001 49

TableA.17: Requestedloans,offeredloans,andfinaltake-up Classifiedas Classifiedas Classifiedas creditworthyby creditworthybythe creditworthyby onlythemill millandthebank onlythebank (A) (B) (C) Guaranteed Direct Guaranteed Direct Direct (A1) (A2) (B1) (B2) (C) Requested 185 88 51 Offered 101 67 49 32 27 Take-up 75 54 38 25 12 Take-up(%) 74% 81% 78% 78% 44% Note: Thistableoutlinesthenumberofindividualsthatrequestedaloanandwereclassifiedascreditworthybyeitherthemill,thebank,orboth(firstrowinthetable,labelled“Requested”). Ofthesefarmersthat requestedaloan,duetoconstraintsonthetotalstudysize,onlyasubsetofthesefarmerswereofferedaloan (thesecondrowinthetable,labelled“Offered”). Therewassomeattritionwithonlysomefarmerssubsequentlytakingaloan(thethirdandfourthrowsinthetable,labelled“Take-up”and“Take-up(%)”,where “take-up(%)”isthefractionoffarmersthatwereofferedaloanthateventuallytooktheloan). Overall,we donotfindthetake-upratesarestatisticallydifferentacrosstreatments(theChi-squaredtestisunableto rejectthenullhypothesisthattake-upresultsarethesameacrosstreatmentsatthe10%significancelevel). 50