Credit Reporting, Relationship Banking, and Loan Repayment

Credit Reporting, Relationship Banking,and Loan Repayment∗

Martin Brown† and Christian Zehnder‡

February 2007

Abstract

How does information sharing between lenders affect borrower’s repayment behavior?

We show - in a laboratory credit market - that information sharing increases repayment

rates, as borrowers anticipate that a good credit record improves their access to credit.

This incentive effect of information sharing is substantial when repayment is not third-

party enforceable, and lending is dominated by one-shot transactions. If, however,

repeat interaction between borrowers and lenders is feasible, the incentive effect of

credit reporting is negligible, as bilateral banking relationships discipline borrowers.

Information sharing nevertheless affects market outcome by weakening lenders’ ability

to extract rents from relationships.

Keywords: Credit Market, Information Sharing, Relationship BankingJEL: G21, G28, D82

∗We thank Paul Calem, Hans Degryse, Richard Disney, Tullio Jappelli, Marco Pagano, Michael Kosfeld,Ernst Fehr and three anonymous referees, as well as participants of seminars at the Fed Philadelphia,CSEF Salerno, the Banca D’Italia, the University of Copenhagen, the University of Frankfurt and TilburgUniversity for helpful comments. Franziska Heusi provided excellent research assistance. Zehnder gratefullyacknowledges financial support by the national center of competence in research on “Financial Valuation andRisk Management”. The national centers in research are managed by the Swiss National Science Foundationon behalf of the federal authorities. The views expressed in this paper do not necessarily reflect those of theSwiss National Bank.

†Swiss National Bank, Borsenstrasse 15, CH-8022 Zurich, [email protected].‡Institute for Empirical Research in Economics, University of Zurich, Blumlisalpstrasse 10, CH-8006

Zurich, [email protected].

1 Introduction

In credit markets, borrowers typically have more information about their investment oppor-

tunities, their own character and their prior indebtedness than lenders. This asymmetry of

information gives rise to selection problems for lenders and potential moral hazard of bor-

rowers, which may lead to a rationing of credit (Stiglitz and Weiss, 1981). In many countries

problems of asymmetric information are aggravated by the fact that loan contracts are costly

to enforce (Levine, 1998; Jappelli et al., 2005).

One response to asymmetric information and costly enforcement in the credit market is

information sharing between lenders about characteristics and behavior of their borrowers.

Theoretical models suggest that information sharing can reduce adverse selection in markets

where borrowers approach different lenders sequentially (Pagano and Jappelli, 1993). More-

over, information sharing can also have a strong disciplining effect on borrowers. The model

of Diamond (1989) suggests that a public credit registry can motivate borrowers to choose

agreed projects. Further models show that information sharing can discipline borrowers into

exerting high effort in projects (Vercammen, 1995; Padilla and Pagano, 2000) and repaying

loans (Klein, 1992).

A recent survey by the World Bank shows that institutionalized information sharing, i.e.,

credit reporting through private credit bureaus or public credit registries, now exist in over

100 countries worldwide (World Bank, 2006).1 In the USA, where credit reporting is most

prevalent, over 3 million credit reports are issued every day (Hunt, 2005). In recent years,

many developing and transition economies have also introduced credit registries or fostered

credit bureaus in the hope of boosting credit growth (Miller, 2003). Giving the strong

growth of credit reporting worldwide and the high hopes which policy makers place in such

institutions, there is a need for empirical evidence which examines how credit reporting

affects the performance of the financial sector.

1Public credit registries are created by public authorities and typically run by central banks. It is usuallymandatory for all supervised financial institutions to submit information to a public credit registry. Inreturn, the same institutions are entitled to receive credit reports based on information available in theregistry. Private credit bureaus are typically set up by banking associations or private entrepreneurs tofacilitate voluntary information sharing between lenders.

1

In this paper we use experimental methods to examine how credit reporting affects loan

repayment and credit market performance. We examine an experimental credit market in

which loan repayment is not third-party enforceable. We first implement a market in which

there is no opportunity for information sharing between lenders. We then implement an

identical market, but with a stylized public credit registry which collects and disburses credit

information to lenders. By comparing repayment behavior and credit volumes between the

two markets we can identify the impact of a credit registry on credit market performance.

We contribute to the empirical literature on information sharing in two ways. First, this

is the only study we know of which examines the impact of information sharing on borrower

behavior. Several authors have shown that credit reports do reduce the selection costs of

lenders by allowing them to more accurately predict loan defaults (Kallberg and Udell, 2003;

Barron and Staten, 2003; Powell et al., 2004; Luoto et. al., 2004). The disciplining effect of

information sharing on borrower behavior has, however, not yet been studied.2 This is by no

means surprising, giving that with field data it is difficult to identify whether an individual

borrower has behaved differently than he would have done without the presence of a credit

registry.

The second contribution of our paper is that, in contrast to existing studies, we can

directly identify how credit reporting affects credit market performance. Current evidence

on the relation between information sharing and credit market performance relies on cross-

country comparisons using aggregate or firm-level data. Jappelli and Pagano (2002) and

Djankov et al. (forthcoming) show that aggregate bank credit to the private sector is higher

in countries where information sharing is more developed. Analyses of firm-level survey data

(Galindo and Miller, 2001; Love and Mylenko, 2003; Brown et al., 2007) further show that

access to bank credit is easier in countries where credit bureaus or registries exist. These

studies cannot clearly identify the direction of causality between information sharing and

credit volume. After all, theory suggests that information sharing will emerge where lenders

benefit more from them (Pagano and Jappelli, 1993) and this is certainly the case where

2Jappelli and Pagano (2002) show that loan defaults, measured by country risk indicators, are lower incountries where credit registries and bureaus are more developed. However, this result can obviously resultfrom better selection of borrowers rather than from actually disciplining them to repay.

2

the credit volume is higher. Thus a positive correlation between credit reporting and credit

market performance may arise simply because credit bureaus are more likely to emerge in

countries where current lending is vibrant or an expansion of credit activity is expected.

By applying experimental methods, our study allows us to circumvent this endogeneity

issue and identify how the exogenous introduction of a credit registry affects credit market

performance.

The impact of credit reporting on repayment behavior should depend on the presence of

alternative disciplining mechanisms. One alternative disciplining mechanism is relationship

banking. Theoretical models suggest that implicit contracts between lenders and borrowers,

i.e., banking relationships, can motivate high effort and timely repayments (Boot and Thakor,

1994). Empirical studies confirm that some credit market segments (in particular small

business lending) are pervaded by relationship-banking and that these relationships improve

the access of potential borrowers to credit (Petersen and Rajan, 1994, Elsas and Krahnen,

1998). Experimental studies (Brown et al., 2004; Fehr and Zehnder, 2005) also confirm

that long-term relationships are a powerful disciplinary device. In credit markets dominated

by repeated interactions (e.g. working capital loans), information sharing may therefore

not be required to discipline borrowers. In contrast, in credit markets dominated by short

term interactions (e.g. consumer credit markets when borrower mobility is high), borrowers

may only be motivated to repay if they know that their current behavior is observable by

other lenders, due to credit reporting. In this paper we examine how the impact of credit

reporting on repayment is related to the presence of relationship banking. We conduct our

experiment for two credit market environments. In one environment information conditions

prevent repeated interaction between borrowers and lenders so that all lending transactions

are inherently one-off. In the second environment, information conditions are such that

lenders can choose to trade with the same borrower repeatedly and banking relationships

can emerge endogenously.

Our results indicate that the impact of credit reporting on repayment behavior and credit

market performance is highly dependent on the potential for relationship banking. When

bilateral relationships are not feasible the credit market essentially collapses in the absence

3

of credit reporting. As repayments are not third-party enforceable, many borrowers default

and lenders cannot profitably offer credit contracts. The introduction of a credit registry in

this environment greatly enhances the performance of the credit market. The availability

of information on past repayment behavior allows lenders to condition their offers on the

borrowers’ reputation. As borrowers with a good track record get better credit offers, all

borrowers have a strong incentive to sustain their reputation by repaying their debt. As a

consequence a well functioning credit market is established in which a large percentage of

the available gains from trade is realized.

When relationship banking is feasible, credit reporting has no such effect on market

performance. In this environment, the market participants solve the moral hazard problem

related to repayment even in the absence of a credit registry. By repeatedly interacting

with the same borrower, lenders establish long-term relationships which enable them to

condition their credit terms on the past repayments of their borrower. As only a good

reputation leads to attractive credit offers from the incumbent lender, borrowers have strong

incentives to repay. The disciplining effect of these banking relationships is sufficiently strong

so that the introduction of a credit registry only slightly improves credit market performance.

Nevertheless, even when relationship banking is feasible, a credit registry does affect market

outcome. First, the credit market is less dominated by specific borrower-lender relations, as

these are no longer necessary to enforce repayment. Second, by improving the information

available to “outside” lenders, a credit registry reduces the ability of lenders to extract rents

from relationships.

The plan of the paper is as follows: Part 2 presents our experimental design and part 3

the corresponding predictions. Part 4 presents our results. Part 5 concludes.

2 Experimental Design

Our objective is to study how the repayment behavior of borrowers in a competitive credit

market is affected by information sharing among lenders. We therefore implement a simple

lending game in which loan repayment is not third-party enforceable, and embed this game

4

in a competitive trading environment.

2.1 Experimental Credit Market

Our lending game is based on the trust game introduced by Berg et al. (1995). In this

sequential, two-player game the first-mover has an endowment which he can keep or transfer

to the second-mover. If the second-mover receives a transfer from the first-mover he earns an

income which exceeds the value of that transfer. The second-mover then decides how to share

his income with the first-mover by choosing the size of a return-transfer. In a very stylized

way the trust game captures the basic features of a credit transaction when borrowers have

riskless investment opportunities, but loan repayments are not enforceable. The extension

of credit (first-mover transfer) then generates positive gains from trade. However, lending is

risky because borrowers (second-movers) can maximize their short-term profits by keeping

all earned income to themselves. One shot and repeated trust games have been studied

intensively in the experimental literature (for an extensive review see e.g., Camerer, 2003,

p. 83 ff. and p. 446 ff.). Our experimental design implements a version of the trust game

which differs in at least two fundamental aspects from the previous literature. First, in

order to measure the impact of information sharing on repayment behavior, we implement

a repeated trust game in which we exogenously vary the lenders information on borrowers’

past repayment decisions. Second, as we want to study repayment behavior in a market

environment, we allow lenders and borrowers to endogenously choose their trading partners

in a competitive trading environment. To the best of our knowledge, both of these features

have not been experimentally investigated before.

Our experimental credit market involves 17 participants. These participants are ran-

domly assigned to the role of borrowers and lenders at the beginning of a session. Ten

subjects are in the role of lenders and seven subjects are in the role of borrowers. Each

session lasts for 20 periods and the roles of subjects are fixed for the whole session.3

3There are two reasons why we choose a fixed number of periods instead of a random stopping time: First,the fixed number of periods makes sure that we have sessions of comparable length for all our treatments.Thus, differences in behavior across treatments cannot be due to different learning opportunities or othertime-dependent effects. Second, the finite time horizon also provides us with within-subject evidence for therelevance of reputation effects. While a random stopping time provides constant reputational incentives, a

5

At the beginning of every period each lender is endowed with 50 capital units. A lender

has two opportunities to make use of his endowment. He can either invest the endowment

in an endowment-storing technology or he can use the endowment to extend credit to a

borrower. The first stage of each period is a continuous one-sided auction, in which lenders

and borrowers can seal credit contracts. The lenders are the contract makers, i.e., they

alone can make credit offers to the borrowers, who themselves can not apply for credit.

When making a credit offer the lender has to specify four items: the size of the loan (k),

the requested repayment (r), the set of market participants who can observe the offer and

finally, which borrowers are authorized to accept the offer. Lenders can freely decide how

they want to split their endowment between the endowment-storing technology and a credit

offer, i.e., the loan size k can be picked from the set {5, 10, 15, ..., 50}. The set for the

requested repayment r is given by {5, 10, 15, ..., 100}. There are two types of credit offers:

Public credit offers and private credit offers. A private credit offer is only addressed to one

specific borrower. It cannot be seen or accepted by other borrowers and is also not visible to

other lenders. A public offer is always shown to all borrowers and all other lenders. However,

even with public offers the lender must specify which borrowers are authorized to accept the

offer. Hereby the lender can choose, or exclude as many borrowers as he wants.4 During the

auction a lender can make as many public and private offers as he wants. However, each

lender can only conclude one credit contract per period. As soon as a borrower accepts an

offer of a given lender a contract is concluded and all other outstanding offers made by this

lender disappear from the market and can no longer be accepted by other borrowers. Each

borrower can accept at most one contract per period so that our credit market implements

an excess supply of credit.

Borrowers are endowed with 5 capital units in each period. At the second stage of a

period borrowers automatically earn an investment income which is twice the size of this

endowment and their borrowed capital, 2(5 + k). At the third stage of a period, borrowers

finite time horizon implies that reputational concerns are strong in the early periods but fade towards theend of the experiment. As a consequence the same subjects are exposed to varying intensities of reputationeffects.

4This implementation of public offers is designed to capture public announcements of credit conditionsby banks who can always choose not to extend credit to some clients on these terms.

6

who received a loan decide whether they want to make the repayment requested by the

lender (r = r) or not repay at all (r = 0). Partial repayments are not possible.5

At the end of each period, each lender is informed about his borrower’s repayment de-

cision, profits are calculated and all market participants get to know their own and their

partner’s payoffs for the period. Payoff functions, the number of lenders and borrowers and

the number of trading periods are common knowledge. The monetary payoffs of the market

participants per period are calculated as follows:

Payoff of lender: π = 50− k + rPayoff of borrower: v = 2(5 + k)− r

2.2 Treatments

Our goal is to study how credit reporting affects borrowers’ repayment choices and credit

market performance. In order to do so we first implement our credit market without any

opportunity for information sharing between lenders. We then implement the same credit

market with an exogenous credit reporting mechanism, which collects and disburses infor-

mation on past repayment behavior of borrowers. In the treatments with credit reporting all

lenders have free access to a credit report at the beginning of every period. The report lists,

for each borrower and all past periods, whether the borrower received a loan and whether

he repaid it.

Our credit reporting mechanism is a stylized version of a public credit registry. The main

feature of a public credit registry is that it is mandatory for (at least supervised) financial

institutions to contribute information to its records. In contrast, private credit bureaus are

based on voluntary information sharing between lenders. In our experiment all lenders must

submit information about their lending activity to the credit reporting institution in each

period. In the following we therefore refer to it as the “credit registry”.6 In reality, public

5In reality some borrowers obviously become delinquent without fully defaulting. However due to thedeterministic nature of investment earnings in our design we exclude partial repayments.

6As we are only interested in the impact of information sharing, rather than its emergence we do notconsider voluntary information sharing on which private credit bureaus are founded. For an experimentalanalysis of voluntary information sharing in a competitive credit market see Brown and Zehnder (2006b).

7

credit registries (and private credit bureaus) differ strongly in the range of lending activities

which they cover, and the range of information they collect and distribute on each credit

activity. The coverage of public credit registries, is affected by the range of financial insti-

tutions which must submit information, the types of loans which are included (commercial

credit/consumer credit) and the size of the threshold above which loans are included (Miller,

2003). The range of information collected per credit transaction can vary between default

information only (coined as “negative” or “black” information in the literature) to detailed

information on outstanding and past repaid loans (“positive” or “white” information). Fur-

ther, real world public credit registries differ in the time frame for which they distribute

this information.7 Our credit registry provides full coverage of all credit activities in our

experiment by collecting and disbursing information on each individual loan from each prior

period. The information distributed by our credit registry on each credit transaction is lim-

ited. Our credit reports tell lenders which borrowers received a loan in which period, and

whether these loans were repaid or not. Moreover, our registry provided complete historical

information on each borrower, as each lender can review the full repayment history of each

individual borrower at any time. However, the registry does not provide information on loan

sizes and interest rates (requested repayments) of the credit transactions in question. Thus,

our credit registry disburses both negative and positive information, but the detail of this

information is (like for most real life credit registries) limited.8

Table 1 provides an overview of our experimental treatments. Treatments which include

credit reporting are called “CR” treatments (for credit reporting). Treatments without

credit reporting are called “NO” treatments (for no credit reporting).

Insert Table 1 about here

7Public credit registries and private credit bureaus often restrict the historical information provided, dueto administrative costs or consumer protection laws. Brown et. al. (2007) show that in transition countries8 out of 12 public credit registries and 8 out of 9 private credit bureaus provide information on lendingactivities for longer than two years.

8We chose to exclude information on loan sizes from the credit reports in our experiment in order tosimplify the information provided to lenders. In contrast to some public credit registries our registry alsodoes not collect information on interest rates which lenders demand from borrowers. However, even wherepublic credit registries do collect information on interest rates, they do not provide to other lenders in creditreports, but rather use this information to facilitate bank supervision.

8

The impact of credit reporting on borrowers’ repayment behavior may depend on the

feasibility of alternative disciplining mechanisms in the credit market. Obviously, there will

be no impact of information sharing on repayment behavior if loan repayment is perfectly

enforceable at no cost by a third party. However, even when a third-party is absent, repay-

ment may be implicitly enforced through bilateral banking relationships (Boot and Thakor,

1994). For this reason, we are particularly interested in how the impact of credit reporting

depends on the degree of relationship banking in a credit market. In reality, the feasibility of

relationship banking in a credit market varies, depending on how mobile borrowers are and

how diverse their funding needs are, compared to the product and geographical specialization

of lenders. If borrowers are highly mobile and lenders are geographically specialized, banking

relationships will be difficult to maintain. On the other hand, if mobility of borrowers is

low or lenders are universal banks with country-wide coverage relationships are simple to

maintain.

We examine the impact of credit reporting for market conditions with varying feasibility

of relationship banking. In order to study the range of the impact which credit reporting may

have on borrower behavior we implement our CR-NO comparison for two border cases: in

one market condition relations are not feasible at all, while in the other condition borrowers

and lenders can always continue relationships if they want to. Our first condition makes

it impossible for lenders to interact repeatedly with a particular borrower by randomly

assigning identification numbers (IDs) to borrowers and lenders in each new period. This

procedure guarantees that no market participant can identify his former trading partners at

the beginning of a period and therefore intentional repeated offers by lenders to borrowers

are ruled out. We henceforth call treatments with this environment Random ID (RID)

treatments. Our second environment involves a market in which lenders and borrowers have

the opportunity to engage in long-term relationships. Repeated interaction with the same

trading partner is possible because subjects have fixed IDs for the entire experimental session.

Consequently, lenders can offer credit to the same borrower (i.e., to the same ID number)

in consecutive periods and, if the borrower accepts these offers, a long-term relationship

is established. In the following we call treatments with this environment Fixed ID (FID)

9

treatments.

2.3 Procedures

In total we conducted 20 experimental sessions, five for each of our four treatments. We had

17 subjects in each session, which makes a total of 340 participants. All experimental subjects

were volunteers. They were all participating for the first time in such an experiment, and each

participant could only participate in one session (i.e., each subject experienced only one of the

treatments). All participants were students at the University of Zurich or the Swiss Federal

Institute of Technology Zurich (ETH). The computerized experiment was programmed and

conducted with the experimental software z-Tree (Fischbacher, forthcoming). A session

lasted approximately ninety minutes. Subjects received a show-up fee of 10 Swiss francs

(CHF) and 1 additional franc for every 20 points earned during the experiment. On average

subjects earned 55 Swiss francs (1.3 CHF ≈ 1 US$ in January 2006).

To make sure that all participants fully understand the decision process and the pay-

ment structure of the game, each subject had to read a detailed set of instructions before

a session was started. An English version of our German instructions is available in Brown

and Zehnder (2006a). The experimental instructions were framed in a credit market lan-

guage. The reason why we chose a context-specific and not a neutral framing was, that the

experiment was relatively complex. In complex experiments a completely neutral language

bears the danger that subjects create their own (potentially misleading) interpretation of

the decision environment. Thus, the context-specific framing gives us control over what our

participants have in mind. In our view, this not only reduces noise but also increases the

external validity of the experiment.9 After reading the instructions participants had to pass

a test with control questions. No session started before all subjects had correctly answered

all control questions. Additionally there were two practice periods before an actual session

9Some experimenters argue, that a context-specific framing distorts incentives as it provides the subjectswith notions of how they “should behave”. However, in this experiment we are only interested in thebehavioral differences across treatments. We do not make any inferences from observed levels of variables(e.g., repayment rates or interest rates). As we do not see any reasons why the context-specific framingshould create different notions of how to behave across treatments, this problem is not relevant for ouranalysis.

10

was started in order to make the participants familiar with the bidding procedures. In both

practice periods, subjects only went through the offering stage of a period, i.e., there were

no repayment choices and subjects could not earn money in the practice periods.

3 Predictions

Under the assumption of common knowledge of rationality and selfishness of all market

participants, the predictions for each of our four treatments are straightforward. Since

repayments are not enforceable, a borrowers’ best response in the stage game is to never

repay a loan. Lenders, anticipating this behavior, will never offer credit so that the credit

market collapses in the stage game equilibrium. As our experiment lasts for a finite number

of periods, a simple backward induction argument ensures that the stage game equilibrium

is played in every period of the game. The different treatment conditions do not affect this

prediction. If lenders are certain that all borrowers are selfish, neither public information on

past repayment behavior of borrowers (RID-CR, FID-CR treatments), nor the possibility to

establish long-term relationships (FID-NO, FID-CR treatments) can overcome this inefficient

outcome.

Empirical evidence suggests, however, that not all people will simply maximize monetary

payoffs in our experiment. It has been shown that, in a wide range of economic settings,

the behavior of some people is also driven by social motives (for an overview see, e.g., Fehr

and Schmidt, 2002). Especially important for our purposes is the experimental evidence on

the “trust game” described in the previous section. In his survey of experimental evidence

for the trust game Camerer (2003) shows that even in anonymous, one-shot transactions

many second-movers do make substantial return transfers. It appears that many second

movers feel a moral obligation to repay the first mover for his initial transfer, or are willing

to reciprocate a first movers risky decision which benefits them. Recent research by Karlan

(2005) shows that the behavior of second movers in the trust game extends to their behavior

in real-life financial decisions, suggesting that the observed “social motives” are by no means

an artifact of laboratory experiments.

11

The evidence from the trust game therefore suggests that, in our experiment social mo-

tives could lead some borrowers to repay loans because they would otherwise suffer from

a bad conscience, or because they would like to reciprocate the decision of lenders to lend

to them. In the following we examine our four treatments under the assumption that the

behavior of some (non-distinguishable) borrowers display such social motives. We assume

that “social” borrowers are conditionally reciprocal: they are willing to meet their repayment

obligations (r = r) even in a one-shot situation, as long as the interest rate requested by the

lender (r−k)100k

does not exceed a personal threshold value. We assume that the remaining

share of borrowers are selfish, in the sense that they never repay loans in a one-shot situation.

3.1 Predictions for the Random ID Treatments

In the RID-NO treatment, lenders have no information on the prior behavior of any

particular borrower in the market. This treatment essentially implements a series of one-

shot interactions so that each period can be analyzed as a one-period game. In such a game,

selfish borrowers never repay their debt, while “social” borrowers repay as long as they are

offered fair financing conditions. Under these conditions the provision of credit can only be

profitable for lenders if there is a substantial fraction of social borrowers. If, in contrast, there

are only few social borrowers the credit market collapses and all lenders fully invest their

capital into the endowment-storing technology. In the Appendix we examine optimal lending

behavior in the RID-NO treatment depending on the share of social borrowers and the degree

of their social preferences. Proposition A1 of the Appendix summarizes our findings, and

suggests that under reasonable assumptions about the degree of social preferences10 at least

two-thirds of borrowers would need to be social to guarantee the existence of a functioning

credit market. Existing experimental evidence for trust games suggests that this condition

is unlikely to be satisfied (Camerer, 2003). Accordingly, we predict that the credit market

will collapse in our RID-NO treatment.

In the RID-CR treatment, lenders receive a credit report at the beginning of each

10We assume that social borrowers are never willing to repay if the requested repayment of the lenderimplies that the borrowers is left with less than half of the gains from trade.

12

period stating, for each borrower and each prior period, whether the borrower concluded

a credit contract and whether he repaid his debt. In contrast to the RID-NO treatment,

lenders in the RID-CR can therefore condition their credit offers (whether to offer credit,

the credit size, and the desired repayment) on the borrowers’ past repayment behavior. If

selfish borrowers anticipate this conditionality of loan offers they have a strong incentive to

hide their type and imitate the behavior of social borrowers. Repaying a loan is the only way

for selfish borrowers to build up a reputation as a social type and to get access to profitable

future credit offers of lenders. In the Appendix we show that this mechanism can sustain an

equilibrium in which a substantial credit volume is provided, even in cases where the share

of social borrowers is such that the credit market would collapse in the RID-NO treatment.11

Proposition A2 in the Appendix describes the following equilibrium behavior of lenders

and borrowers: In all periods lenders strictly condition their credit offers on the borrowers’

past repayment behavior, i.e., they make only credit offers to borrowers who have never

defaulted in the past. In a first phase of the RID-CR treatment this motivates all selfish

borrowers to repay loans out of reputational concerns and accordingly lenders extend the

maximal credit volume. During this “pooling” phase selfish and social borrowers behave

identically and therefore no information about the borrowers’ types is revealed. In later

periods, reputational incentives decline and repayment rates fall as selfish borrowers begin

to default with a positive probability. In this second phase, the aggregate credit volume

begins to fall as those borrowers who defaulted in prior periods receive no further loans and

those who repaid receive only loans with non-maximal credit sizes. Furthermore, competition

among lenders implies that credit offers are such that all gains from trade go to the borrowers

and lenders make zero profits throughout the experiment.

Based on the above considerations we state the following hypothesis for our Random ID

treatments:

11The assumption that there are two non-distinguishable types of borrowers implies that we analyze afinitely repeated game with incomplete information. Such games are usually characterized by a large numberof equilibria (see Fudenberg and Maskin, 1986). It is not our objective to provide a complete formal analysisof our experimental game in the Appendix. We rather prove that there are Perfect Bayesian Equilibria inwhich the reputation mechanisms intuitively described in this section ensure that a functioning credit marketexists.

13

Hypothesis Random ID Treatments: In the RID-CR treatment lenders condition

credit offers and terms on the information available in the credit registry about the prior

repayment behavior of borrowers. This creates reputation incentives in the RID-CR

treatment inducing a significantly higher repayment rate of borrowers in the RID-CR than

in the RID-NO treatment. The credit volume extended by lenders is consequently also

significantly higher than in the RID-NO treatment. The repayment rate and credit volume

in the RID-CR treatment converge, however, to that of the RID-NO treatment towards the

end of the experiment.

3.2 Predictions for the Fixed ID Treatments

In the FID-NO treatment, there is no credit registry, so that lenders do not have infor-

mation on the behavior of all borrowers in all prior periods. However, due to fixed ID’s,

lenders do have information on past behavior of those borrowers with whom they themselves

have traded in prior periods. Thus in contrast to the RID-NO treatment, lenders have the

possibility to reward known borrowers with good repayment histories with attractive con-

tract renewals. If repayment guarantees access to profitable loans from incumbent lenders,

selfish borrowers may also be motivated to repay. In the Appendix we show that there is an

equilibrium in the FID-NO treatment in which endogenously formed banking relationships

ensure the provision of a substantial credit volume even in the case where the fraction of

social borrowers is insufficient to guarantee the existence of a credit market in the RID-NO

treatment.

The equilibrium behavior of lenders and borrowers derived in Proposition A3 of the

Appendix can be described as follows: In the first period all lenders make a competitive offer

and try to conclude a contract with a borrower. Those lenders who succeed in concluding a

contract with a borrower in the first period subsequently establish a long-term relationship

with their incumbent borrower. As long as the incumbent borrower repays, they renew his

contract in every period by making him a private offer. Lenders who could not conclude

a contract in the first period invest their capital in the endowment-storing technology and

14

remain outside the credit market. The reason that they do not try to enter the market

by making competitive offers to borrowers in relationships with other lenders is that they

believe that such contract offers would only attract selfish borrowers. As outside lenders do

not contest the market, lenders who have established a relationship with a borrower can exert

a certain market power and “hold-up” their borrower. By making offers which just satisfy

the conditions under which social borrowers repay, they can skim off part of the gains from

trade in their relationship. Of course, in the first period lenders anticipate that they will

earn a rent if they manage to establish a relationship. Competition among lenders therefore

implies that they are prepared to make losses in the first period in order to get access to the

rents earned in a relationship. Within the relationships, the conditional contract renewals

of incumbent borrowers, in combination with the fact that outside lenders are not willing

to offer credit, motivates selfish borrowers to perfectly imitate the repayment behavior of

social borrowers in a first phase of the game. As lenders make profits in these periods they

maximize their income by extending maximal credit amounts. During this “pooling” phase

of the experiment, no additional information about the types of borrowers is revealed and the

lenders’ beliefs remain constant at the initial level. When the end of the game draws near,

however, lenders are only willing to renew their contracts if they get additional information

on the borrowers’ types. Therefore, in this phase, selfish borrowers start defaulting with

positive probabilities and therewith ensure that lenders can update their beliefs and remain

willing to renew their contracts. However, as defaulting borrowers no longer get credit offers

and as lenders start to lower the size of their loans, the extended credit volume decreases

towards the end of the game.

In the FID-CR treatment, the presence of a credit registry implies that lenders have

information not only on the behavior of their own prior borrowers, but on all borrowers

in the game. As a consequence, the “credit reporting” equilibrium derived for the RID-

CR treatment and described in detail in Proposition A2 of the Appendix also applies for

the FID-CR treatment. Even in the presence of a credit registry, though, the relationship

equilibrium described for the FID-NO treatment in Proposition A3 in the Appendix can

15

also be sustained in the FID-CR treatment. In the Appendix (Propositions A2 and A3)

we show that the “credit reporting” equilibrium and the “relationship banking” equilibrium

can yield identical repayment rates and practically identical credit volumes. Thus market

performance in the FID-CR treatment can be similar to that in the FID-NO, independent

of which equilibrium type arises. However, market structure and distribution of surplus will

differ between the FID-CR and FID-NO treatments if the “credit reporting” is played in

the FID-CR. As discussed above, long-term relationships are not necessary to sustain this

equilibrium and would therefore be observed less frequently than in the FID-NO. Moreover,

lenders who establish relationships earn quasi-rents in the FID-NO treatment while in the

“credit reporting” equilibrium all lenders earn zero profits in all periods. Based on the above

considerations we make the following hypothesis for our Fixed ID treatments:

Hypothesis Fixed ID Treatments: Repayment rates and credit volume are identical in

the FID-NO and FID-CR treatments: Both display high repayment rates and credit volumes

in an initial phase. Towards the final period, however, some borrowers start to default and

credit volumes decrease. In the FID-CR, the disciplining of borrowers is less reliant on

relationship lending, so that long-term relationships may be less frequent. Moreover, the

presence of a credit registry implies that in the FID-CR it will be more difficult for lenders

to extract profits from relationships than in the FID-NO.

4 Results

4.1 Random ID Treatments

In this section we examine the impact of credit reporting in a market where there is no

alternative device to motivate loan repayment. In particular, bilateral relationships are

prevented due to the random assignment of ID numbers to all lenders and borrowers in each

period. Table 2 presents summary statistics for our two corresponding treatments, the RID-

CR and RID-NO. For each treatment the table presents mean statistics across all 5 sessions,

as well as the range of results per sessions. Our results display strong differences in market

outcome between the RID-CR and RID-NO treatments. As predicted, the repayment rate

16

of borrowers is substantially higher in the RID-CR than in the RID-NO treatment. On

average 80% of all loans are repaid in the RID-CR treatment, while only 28% of loans are

repaid in the RID-NO. Moreover, looking at session level outcomes we see that in the RID-

CR the mean repayment rate exceeds 70% in all sessions, while in the RID-NO treatment

no session displays a repayment rate higher than 40%. Based on a comparison of session

averages a Mann-Whitney test suggests that the difference in repayment rates between the

two treatments is statistically significant.12

Table 2 further shows a substantial difference in lending activity between the RID-CR

and RID-NO treatments. In the RID-CR 94% of all potential lending contracts are realized,

and the average credit size is 41 (out of 50 possible) points. In contrast, in the RID-NO only

59% of potential contracts are realized, while the average credit size is only 23 points. These

results imply that 77% of the potential credit volume is realized in the RID-CR treatment,

while only 27% of the potential credit volume is realized in the RID-NO. Session level results

further show that the realized credit volume is higher in every session of the RID-CR than

in any session of the R–NO. A Mann-Whitney test based on session averages confirms that

credit volume is significantly higher in the RID-CR than in the RID-NO treatment.13


While credit market efficiency differs substantially between the RID-CR and RID-NO

treatments, the distribution of gains from trade is more similar. Table 2 shows that the

mean interest rate demanded by lenders is practically identical in the two treatments. In

our design the interest rate in % can be calculated as: (Desired repayment-Credit size)·100Credit size

. While

interest rates do vary more across sessions in the RID-CR than in the RID-NO, there does

12We conduct a Mann-Whitney Test using mean repayment rates per session as observations. The 5sessions of the RID-CR treatment display repayment rates of 87, 85, 81, 77 and 70 percent respectively. Inthe RID-NO treatment the five sessions have repayment rates of 39, 31, 29, 26 and 16 percent respectively. Aone-sided test thus cannot reject the hypothesis that repayment is more frequent in the RID-CR treatment(p = .004).

13We conduct a one-sided Mann-Whitney test using realized credit volume per session as observations.In the RID-NO treatment the five sessions display a credit volume (measured in percentage of the totalpotential volume) of 36, 29, 29, 24 and 16 percent respectively. In the RID-CR treatment, the credit volumeper session was 84, 81, 78, 76 and 66 percent respectively. A one sided Mann-Whitney test thus yields ap-coefficient of p = .004.

17

not seem to be a significant difference in the level of interest rates.14 In both treatments all

gains from trade are entirely reaped by the borrowers. Lenders earn average payoffs which

are very close to their outside option of 50 points per period. In contrast, borrowers earn

average payoffs which well exceed their outside option of 10 points per period. Looking at

differences across treatments we find that borrowers earn substantially more in the RID-CR

than RID-NO treatment, because gains from trade are higher in the RID-CR. A comparison

of session averages suggests that the difference in borrower payoffs between treatments is

highly significant.15 Lenders also earn slightly more in the RID-CR treatment, where on

average they break even, than in the RID-NO treatment where they make slight losses.

A comparison of session averages suggests that the difference in lender payoffs between

treatments is significant.16 The fact that all gains from trade are reaped by the borrowers

confirms our prediction that the excess supply of credit in our experiment would lead to a

highly competitive credit market.

The summary statistics displayed in Table 2 support our hypothesis that the presence of

the credit registry encourages loan repayment in the RID-CR treatment, because borrowers

anticipate that their future access to credit depends on their repayment history. Table

3 presents regression results confirming that borrowers’ access to credit in the RID-CR

treatment is strongly dependent on their prior repayment behavior. In our experiment

lenders can condition three aspects of credit offers on a borrowers repayment history; whether

to offer a contract at all, which credit size to offer, and which interest rate to demand. We

expect that in the RID-CR treatment borrowers with better repayment records are more

14We conduct a two-sided Mann-Whitney test using mean interest rates per session as observations. In theRID-NO treatment the five sessions display a mean interest rate of 38.4, 34.7, 32.0, 30.4, and 29.9 percentrespectively. In the RID-CR treatment, the mean interest rate per session was 48.1, 35.2, 30.1, 28.4, and24.6 percent respectively. A two sided Mann-Whitney test thus yields a p-coefficient of p = .69.

15We conduct a one-sided Mann-Whitney test using mean period profits of borrowers per session asobservations. In the RID-NO treatment the five sessions display mean borrower profits of 39.3, 34.5, 31.1,29.8, and 24.4 per period respectively. In the RID-CR treatment the five sessions display mean borrowerprofits of 52.4, 52.2, 47.1, 46.5, and 39.5 per period respectively. A one sided Mann-Whitney test thus yieldsa p-coefficient of p = .004.

16We conduct a one-sided Mann-Whitney test using mean period profits of lenders per session as observa-tions. In the RID-NO treatment the five sessions display mean lenders profits of 45.4, 45.3, 44.4, 42.9, and42.2 per period respectively. In the RID-CR treatment the five sessions display mean lenders profits of 52.7,52.4, 50.5, 49.8, and 47.7 per period respectively. A one sided Mann-Whitney test thus yields a p-coefficientof p = .004.

18

likely to receive credit, receive larger loans, and pay lower interest rates. In contrast, in

the RID-NO treatment where information conditions prevent conditional contract offers,

we should find that access to credit and the cost of funds are independent of a borrowers

past behavior. In order to test these hypotheses table 3 examines the credit conditions of

borrowers in each period of our RID-CR and RID-NO treatments. Column (1) of Table 3

reports the results of a probit regression relating a borrowers probability of sealing a credit

contract in the RID-CR treatment to his personal repayment history. The dependent variable

in this column is a dummy variable “Contract” which is 1 if the borrower seals a credit

contract in period t and 0 otherwise. We relate this dummy variable to a borrower’s “Prior

repayment rate”, which measures the share of previous loans which he repaid. We control

for time effects by including 3 dummy variables “Period 6-10”, “Period 11-15”, “Period

16-20” which are 1 only for observations within the respective phase of the experiment.

Since observations within a session may not be independent, the t-statistics reported in

parentheses in Table 3 (and in all other regressions below) are based on robust standard

errors, adjusted for clustering at the session level. The positive and significant coefficient

on “Prior repayment rate” in column (1) shows that in the RID-CR treatment borrowers

with good credit records are more likely to get credit. Column (2) and column (3) relate

the credit conditions of those borrowers who did receive credit in the RID-CR treatment to

their prior repayment behavior. In column (2) we report results for “Credit size”, while in

column (3) we report results for the “Interest rate” charged by lenders. The coefficient of

“Prior repayment rate” in both columns confirms our predictions; borrowers with good credit

histories receive larger loans and pay lower interest rates. The results reported in columns

(1-3) of Table 3 demonstrate that lenders in the RID-CR make extensive use of information

available from the credit registry in this treatment. They condition their loan offers strongly

on the prior repayment behavior of borrowers. By doing so they create strong incentives

for borrowers to repay loans at least in early phases of the experiment. Such reputation

incentives are not present in the RID-NO treatment where, the access to credit and the cost

of funds are not conditioned on prior repayment behavior. Columns (4) through (6) repeat

our regression analysis using data from the RID-NO treatment. Not surprisingly, in this

19

treatment a borrower’s prior repayment rate has no significant impact on his probability of

getting credit, the size of this loan or the interest rate.


Our results so far show that credit reporting creates strong reputation incentives for even

selfish borrowers to repay loans in the RID-CR treatment. This explains why the average

repayment rate in the RID-CR is almost three times higher than in the RID-NO treatment.

However, even if the credit registry in the RID-CR treatment does discipline borrowers to

repay loans, we expect that the repayment rate will fall towards the end of the experiment.

Remember that the value of a good credit record declines towards the end of our experiment,

due to the finite horizon of 20 periods. We therefore expect that selfish borrowers who repay

in earlier periods out of reputation concerns, will default in the final periods. Indeed, Figure

1 shows that towards the end of the RID-CR treatment loan repayments drop substantially.

While 86% of all loans are repaid in period 1 through 15, this falls to less than 50% in the

last five periods of the RID-CR treatment. In contrast, the repayment rate in the RID-NO

treatment hovers around 30% for the entire duration of the experiment. As predicted, in the

final periods of the experiment the repayment rate in the RID-CR treatment converges to

that of the RID-NO treatment. These findings support our conjecture that high repayment

rates in earlier phases of the RID-CR treatment are due to the reputation effects of the credit

registry.

Insert Figure 1 about here

A regression analysis confirms the strong impact of the credit registry on repayment rates

in the RID-CR treatment. Table 4 presents the results of a probit analysis of borrowers

repayment decisions in the RID-CR (column 1) and RID-NO (column 2), controlling for

credit conditions and the phase of the experiment in which the credit transaction takes

place. Our dependent variable is a dummy variable “Repayment” which is 1 if a borrower

repaid and 0 if he defaulted. We control for the size of loans and the interest rate by

including the variables “Credit size” and “Interest rate”. We also control for time effects

20

by including 3 dummy variables “Period 6-10”, “Period 11-15”, “Period 16-20” which are 1

only for observations within the respective phase of the experiment. Our regression analysis

identifies a negative time effect on repayment behavior in the RID-CR treatment. The

coefficients of “Period 11-15” and “Period 16-20” are both negative and significant, with

the latter more pronounced than the former. This result confirms our conjecture that the

presence of a credit registry creates reputation incentives to repay loans in earlier periods

of the RID-CR treatment. We find no corresponding time effect on repayment rates in

the RID-NO treatment. The negative and significant coefficient on “Interest rate” in both

columns suggests that the “fairness” of a credit offer does affect the probability of repayment.

Lenders who demand higher interest rates are less likely to be repaid in both treatments.

The insignificant coefficient of “Credit size” in both treatments suggests, in contrast, that

repayment behavior does not vary with the volume of credit received.


Figure 1 also shows that the fall in repayment rates in the final periods of the RID-CR

treatment is mirrored by a substantial decline in lending activity. The figure reports the

realized credit volume per period as a percentage of the maximum credit volume for the

RID-CR and RID-NO treatments.17 In the RID-CR treatment the total volume of credit

rises from 64 percent in period 1 to 92% in period 12 and remains above 80% until period

17. As predicted, credit volume then falls in the final periods of the RID-CR treatment.

Surprisingly, the RID-NO treatment starts off with a similar credit volume to that of the

RID-CR. However, in this treatment lending activity falls rapidly, declining to less than

30% of its potential from period 9 onwards. Again, confirming our predictions we find that

the lending volume in the RID-CR treatment converges to that of the RID-NO in the final

periods of the experiment.

Regression results also confirm that lenders anticipate the fall in repayment incentives

in the RID-CR treatment over time. Looking back to Table 3 we see that borrowers in the

17As the maximum loan size was 50 units and 7 loans were possible in each period, the maximum creditvolume per period in a session was 350 units.

21

RID-CR treatment are less likely to receive credit in the final 5 periods of the experiment

than beforehand, even controlling for their credit history (see negative coefficient of “Period

16-20” in column (1) of Table 3).

Our results in this section suggest that, in the absence of alternative disciplining devices,

credit reporting can strongly motivate borrowers to repay loans. Lenders use the information

available from the credit registry to condition lending terms on borrowers prior repayment

behavior. By doing so they generate strong reputation incentives for borrowers to repay

loans at least in early phases of the experiment. High repayment rates make it feasible for

lenders to extend high credit volumes, despite the fact that repayment is not third-party

enforceable. Strong competition among lenders implies that all surplus generated by credit

reporting is reaped by borrowers.

4.2 Fixed ID Treatments

In this section we examine the impact of credit reporting on repayment behavior when there

is an alternative mechanism to motivate repayment: bilateral banking relationships. Our

predictions suggest that in this environment credit reporting may not necessarily enhance

repayment incentives and credit market volume. However, it might alter the structure of

trade, by reducing the prevalence of bilateral relations, and also limit the ability of lenders

to hold up borrowers in bilateral relations.

Table 5 displays summary statistics for our two treatments in which bilateral relationships

are feasible due to fixed identities of all borrowers and lenders throughout the experiment;

the FID-CR and FID-NO treatments. The table presents treatment means as well as the

variation of results across sessions. Our results show only negligible differences in market

outcome between the two treatments. The repayment rate of borrowers is very high in

the FID-CR (79%) and the FID-NO (74%), with little variation across sessions in either

treatment. Our results also show little differences in lending activity between the FID-CR

and FID-NO treatments. In the FID-CR treatment 94% of all potential lending contracts are

realized, and the average credit size is 42, while in the RID-NO 92% of potential contracts

are realized with an average credit size of 40 points. Interest rates and the distribution of

22

gains from trade are also very similar in the FID-CR and FID-NO treatments. The mean

interest rate demanded by lenders is only slightly higher in the FID-CR (29%) than the

FID-NO (26%). Similar to our results for the random ID treatments we find that in both

the FID-CR and FID-NO treatments all gains from trade are reaped by the borrowers.

Lenders earn average payoffs which are very close to their outside option of 50. In contrast,

borrowers earn average payoffs which well exceed their outside option of 10 points per period.

This result confirms that the excess supply of credit in our experiment did induce a highly

competitive credit market. Mann-Whitney tests based on session averages suggests that the

slight differences in repayment rates, credit volume and interest rates between the FID-CR

and FID-NO treatments are not statistically significant.18


The high repayment rates presented in Table 5 suggest that the incentives for borrowers

to repay loans are high in both the FID-CR and FID-NO treatments. Indeed, our data

shows that reputation incentives were very strong in both treatments, as lenders strongly

conditioned their loan offers on past repayment behavior. Table 6 replicates our regression

analysis from table 3 relating borrowers’ access to credit to their prior repayment rate, now

using data from the FID-CR and FID-NO treatments. Again, we examine the impact of

a borrowers “Prior repayment rate” on his probability of sealing a “Contract”, the “Credit

size” he gets, and the “Interest rate” he pays. Columns (1-3) show results using data from

the FID-CR treatment, while columns (4-6) show results for the FID-NO treatment. In

all regressions the main explanatory variable is a borrower’s “Prior repayment rate”, which

18We conduct a Mann-Whitney Test using mean repayment rates per session as observations. The 5sessions of the FID-CR treatment display repayment rates of 86, 82, 78, 76 and 72 percent respectively. Inthe FID-NO treatment the five sessions have repayment rates of 79, 77, 76, 72 and 67 percent respectively.A two-sided test thus cannot reject the hypothesis that repayment equally frequent in the two treatments(p = .22). We further conduct a Mann-Whitney test using realized credit volume per session as observations.In the FID-CR treatment the five sessions display a credit volume (measured in percentage of the totalpotential volume) of 82, 82, 80, 76, and 76 percent respectively. In the FID-NO treatment, the credit volumeper session was 81, 78, 72, 69, and 69 percent respectively. A two sided Mann-Whitney test thus yields ap-coefficient of p = .158. We finally conduct a two-sided Mann-Whitney test using mean interest rates persession as observations. In the FID-CR treatment the five sessions display a mean interest rate of 37, 35, 28,24, and 23 percent respectively. In the FID-NO treatment, the mean interest rate per session was 31, 26,26, 25, and 22 percent respectively. A two sided Mann-Whitney test thus yields a p-coefficient of p = .55

23

measures the share of previous loans which he repaid. We control for time effects in all

regressions by including 3 dummy variables “Period 6-10”, “Period 11-15”, “Period 16-

20” which are 1 only for observations within the respective phase of the experiment. The

significant coefficient of “Prior repayment rate” in all columns show that in both the FID-CR

and the FID-NO treatment borrowers with good credit records are more likely to get credit,

receive larger loans and pay lower interest rates.


Even if borrowers have strong reputation incentives to repay loans in the FID-CR and

FID-NO treatments, these incentives should wear off towards the end of the experiment.

We therefore expect a significant decline in repayment rates in both treatments in the final

phase of the experiment. Figure 2 shows that this is indeed the case. The repayment rate in

both treatments hovers around 80 percent from the beginning of the experiment until period

17. In the final three periods we then see a significant fall in repayment rates to below 50

percent. This decline in repayment rates is mirrored by a substantial drop in the volume of

extended by lenders. Figure 2 shows that prior to period 18 well above 70 percent of the

potential credit volume per period was extended. The credit volume then declines rapidly

and reaches just 30 percent in the final period.

Insert Figure 2 about here

Table 7 provides a detailed analysis of the repayment behavior of borrowers in the fixed

ID treatments, replicating our analysis of the Random ID treatments in table 4. The table

presents the results of a probit analysis of borrowers repayment decisions in the FID-CR

(column 1) and the FID-NO (column 2), controlling for credit conditions and the phase of

the experiment in which the credit transaction takes place. The dependent variable is a

dummy variable “Repayment” which is 1 if a borrower repaid and 0 if he defaulted, and

we control for the size of loans and the interest rate by including the variables “Credit

size” and “Interest rate”. We again control for time effects by including 3 dummy variables

“Period 6-10”, “Period 11-15”, “Period 16-20” which are 1 only for observations within the

24

respective phase of the experiment. T-statistics reported in parentheses are again based

on robust standard errors, and adjusted for clustering at session level. We find a negative

and significant coefficient on the variable “Period 16-20” in both columns, showing that

borrowers were less likely to repay a given loan in the final phase of the experiment in both

treatments. This result supports our hypothesis that the high repayment rates in the two

treatments are due the reputation incentives created by bilateral relations (FID-NO) and

the presence of a credit registry (FID-CR).


Table 7 also displays an interesting discrepancy between repayment behavior in the FID-

CR and FID-NO treatments. In the FID-CR treatment the probability of repaying a loan was

independent of the agreed loan size. In contrast, in the FID-NO treatment our results suggest

that the repayment rate is significantly dependent on the loan size offered to the borrower.

In this treatment the coefficient on “Credit size” is positive and highly significant. These

results suggests that credit reporting does alter repayment incentives even when bilateral

relations are feasible. Without a credit registry borrowers will only repay loans if the relation

with their current lender is valuable to them. As a result the borrower will be more likely

to repay the higher his expected earnings from future loan contracts from this lender, which

depend on future loan sizes,and interest rates. If the current loan size is an indicator of

future loan sizes then it is rational for the borrower to condition his repayment choice on

the credit size. In contrast the presence of the credit registry in the FID-CR treatment

creates incentives to repay loans independent of the current credit size. The reason for

this is that the credit registry, as implemented in our experiment, only provides partial

information to future prospective lenders. The registry only reports the prior repayment

behavior of borrowers, but not the conditions of the loans which they repaid or defaulted

upon. Given that a borrowers public reputation is dependent on his repayment behavior,

there are strong incentives to repay even small loans in the FID-CR. Indeed the borrower

should be particularly concerned about his public reputation if he is not interested in pursuing

a bilateral relationship with his current lender. The reasoning above would also imply that

25

borrowers in the FID-CR treatment would also condition their repayment choice less on the

interest rate than borrowers in the FID-NO treatment. This is indeed the case. In both

treatments we find a negative and significant coefficient on “Interest rate” suggesting that

the “fairness” of a credit offer affects the probability of repayment. However, in the FID-

CR treatment the impact of the interest rate on repayment behavior is substantially weaker

than in the FID-NO treatment. This result again shows that the repayment incentives of

borrowers are altered by the presence of credit reporting.

Our predictions suggest that the presence of a credit registry in the FID-CR treatment

may lead to a different trading structure than in the FID-NO treatment. We expect that

the FID-NO treatment will be pervaded by long-term relationships as bilateral relationships

are the key to disciplining borrowers. In contrast, the existence of a credit registry in the

FID-CR implies that long-term relationships between particular borrowers and lenders are

not necessary to discipline borrowers. We thus predict that there will be fewer relationships

in the FID-CR treatment than in the FID-NO treatment. The summary statistics presented

in Table 5 show that this is the case. The final line of that table reports the ratio of contracts

which are renewed form one period to another, i.e., the share of transactions which involved

the same lender - borrower pair as in the previous period. In the FID-NO treatment this ratio

of renewed contracts is 48%. Thus roughly half of all loans made in this treatment involve

the same lender and borrower as in the previous period. The table, however, also shows that

contract renewals are also quite common even when a credit registry exists. Indeed, nearly

40% of all contracts in the FID-CR treatment are renewed from one period to another. As

predicted, the share of renewed contracts, is lower in the FID-CR than that in the FID-NO

treatment. However, due to the strong variation across sessions displayed in Table 5, this

difference is only of borderline significance.19

The fact that bilateral relations are so frequent in the FID-CR treatment is surprising.

Although lenders have access to a credit registry in this treatment it seems that they still rely

strongly on credit relationships to motivate loan repayment. This finding is less astonishing

19In the five sessions of the FID-NO, average renewal rates are 56, 56, 56, 38 and 35 percent respectively.In comparison, the five FID-CR sessions have renewal rates of 55, 43, 42, 32 and 23 percent respectively. Aone-sided Mann-Whitney test using these session averages as observations yields a p-value of p = .11.

26

when we compare the information available within a relationship to that available from a

credit registry. Within a long term relationship, lenders typically have much more informa-

tion about a borrower than they could elicit from a credit report. In our experiment this is

also the case. Our credit registry only provided information on whether a borrower repaid a

loan or not. Within a relationship, however, the lender had additional information on con-

tract terms (credit size, repayment size) which a lender had accepted and repaid. Our results

suggest that this additional information encouraged lenders to maintain relationships with

a particular borrower, although they could easily obtain the credit record of each borrower

at no cost.

We conclude our results section by investigating the impact of credit reporting on the

distribution of gains from trade in bilateral relations. Padilla and Pagano (1997) suggest

that information sharing between lenders may mitigate the hold-up problem in banking

relationships. The presence of a credit registry or credit bureau implies that competitors

are better informed about a borrowers quality, and thus limits the potential for incumbent

banks to extract all informational quasi-rents generated within a relationship. Given that

relationships arise in both our FID-CR and FID-NO treatments we can test whether the

presence of the credit registry in the FID-CR does alter the distribution of gains from

trade in relationships. Table 8 reports the results of a regression analysis in which we

relate lenders payoffs per credit transaction to the duration of the relationship in which

the credit transaction took place. The dependent variable for all regressions in the table is

“Lender Profit” which captures the payoff of a lender per credit transaction, in excess of his

outside option of 50, i.e, the rent he earns per transaction. Our first explanatory variable

“Short-term” is 1 only for all transactions which take place in relationships with a final

duration of less than three periods. Our second explanatory variable “Long-term” is 1 only

for all transactions which take place in relationships with a final duration of at least three

periods. As we include both of these dummy variables but no constant to our regression, the

two dummies identify the mean rent in short- and long-term relationships. We control for

the phase of the experiment in which a transaction takes place by including the 3 dummy

variables “Period 6-10”, “Period 11-15”, “Period 16-20” which are 1 only for observations

27

within the respective phase of the experiment. Column (1) reports estimation results for the

FID-CR treatment, while column (2) reports results for the FID-NO treatment.

The results in table 8 show that lenders earn positive rents in both treatments from

long-term relations. Our results suggest that on average lenders earn rents of 10.3 points

per period from long-term relationships in the FID-NO treatment, and 5.4 points per pe-

riod in the F-CR treatment. The fact that the coefficient of “Long-term” is lower and less

precise in column (1) for the FID-CR treatment suggests that credit reporting does reduce

the ability of lenders to extract rents from long term relationships. In order to test the

significance of this result we pool the data from the FID-CR and FID-NO treatments and

repeat our regression analysis, including the interaction term “CR*Long-term” which is 1

only for long-term interactions in the FID-CR treatment. We further include the interaction

term “CR*Short-term” which is 1 only for short-term interactions in the FID-CR treatment.

In this pooled regression the interaction term “CR*Long-term” captures the difference in

rents earned by lenders in long-term treatments between treatments. We expect a negative

coefficient on this term if the presence of a credit registry reduces lenders ability to extract

profits in bilateral relations. The results reported in column (3) show that the interaction

term does yield a significantly negative coefficient. This result confirms theoretical predic-

tions suggesting that information sharing between lenders can mitigate hold-up issues in the

credit market.


Table 8 further shows that lenders do not earn positive rents from short-term relations

in either treatment. Interestingly though, while lenders make significant losses in short-term

relations in the FID-NO treatment (9.3 points per period), this does not seem to be the case

in the FID-CR treatment. The lower and insignificant coefficient of “Short-term” in column

(1) suggests that the credit registry helps lenders to avoid substantial losses in short-term

encounters. The difference in earnings from short-term interactions between the treatments

is confirmed by the positive and significant interaction term “CR*Short-term” in column

(3) of table 8. These findings suggest a further benefit of information sharing, even when

28

relationship banking is feasible: It helps lenders to reduce losses in one-off transactions, by

avoiding encounters with borrowers who may not repay their loans.

5 Conclusions

In this paper we applied experimental methods to examine the impact of information sharing

between lenders on the repayment behavior of borrowers in a competitive credit market. Our

results suggest that the impact of credit reporting on repayment behavior and credit market

performance depends strongly on the feasibility of relationship banking as an alternative dis-

ciplining device. Credit reporting is highly valuable in markets where banking relationships

are difficult to establish, for example, due to highly mobile borrowers. In such markets,

the disciplining of borrowers to repay loans is strongly dependent on the existence of an

institutionalized information-sharing mechanism. By contrast, in markets where relation-

ship banking is prevalent, these relationships may themselves motivate repayment, so that

credit reporting has little impact on borrower behavior. However, even when the presence

of relationship banking implies that credit reporting does not substantially improve credit

market performance, it does alter market structure and the distribution of gains from trade.

The presence of an information sharing mechanism implies that the disciplining of borrowers

is less reliant on bilateral relations, so that the trading pattern is characterized by fewer long

run relationships. Moreover, information sharing implies that the information advantage of

incumbent lenders over “outside” lenders about borrowers is reduced, weakening their ability

to hold-up borrowers in relationships.

Our methodology and results suggest several avenues of future research. First, experi-

mental methods could be applied to study the endogenous emergence of information sharing.

Theoretical models (Klein, 1992; Jappelli and Pagano, 1993) suggest that private credit bu-

reaus are more likely to emerge when they are most valuable to lenders. Experimental

methods would make it possible to examine this hypothesis by studying the emergence of

credit bureaus in a variety of market environments. Experimental methods could also be

applied to study alternative designs of credit bureaus and credit registries. As suggested

29

by theoretical work (Padilla and Pagano, 2000; Vercammen, 1995), the type of information

recorded by a credit registry, the history of credit records provided, and also the incentive

mechanisms related to providing and retrieving information may affect the functioning and

impact of a credit registry. It would be valuable to study these effects in a controlled manner

through carefully designed experiments.

30

References

Barron, John M., and Michael Staten. (2003). “The Value of Comprehensive Credit Re-

ports: Lessons from the U.S. Experience.” In Credit Reporting Systems and the Inter-

national Economy, edited by Margaret J. Miller, pp. 273-310. Boston: MIT Press.

Berg, Joyce, John W. Dickhaut, and Kevin A. McCabe. (1995). “Trust, Reciprocity, and

Social History.” Games and Economic Behavior 10(1), 122-142.

Boot, Arnoud W. A., and Anjan V. Thakor. (1994). “Moral Hazard and Secured Lending

in an Infinitely Repeated Credit Market Game.” International Economic Review 35(4),

899-920.

Brown, Martin, Armin Falk, and Ernst Fehr. (2004). “Relational Contracts and the Nature

of Market Interactions.” Econometrica 72(4), 747-780.

Brown, Martin, Tullio Jappelli, and Marco Pagano (2007). “Information Sharing and

Credit: Firm-level Evidence from Transition Countries.” Mimeo.

Brown, Martin, and Christian Zehnder. (2006a). “Credit Reporting, Relationship Banking,

and Loan Repayment.” Swiss National Bank Working Papers 2006-3.

Brown, Martin, and Christian Zehnder. (2006b). “The Emergence of Information Sharing

in Credit Markets.” Mimeo.

Camerer, Colin. F. (2003). Behavioral Game Theory: Experiments in Strategic Interaction.

Princeton: Princeton University Press.

Diamond, Douglas W. (1989). “Reputation Acquisition in Debt Markets.” Journal of Po-

litical Economy 97(4), 828-862.

Djankov, Simeon, Caralee McLiesh and Andrei Shleifer. (forthcoming). “Private Credit in

129 Countries.” Journal of Financial Economics.

Elsas, Ralf, and Jan Pieter Krahnen. (1998). “Is Relationship Lending Special? Evidence

from Credit-File Data in Germany.” Journal of Banking and Finance 22, 1283-1316.

Fehr, Ernst, and Klaus Schmidt. (2002). “Theories of Fairness and Reciprocity - Evidence

and Economic Applications.” In Advances in Economics and Econometrics - 8th World

Congress, edited by Mathias Dewatripont, Lars P. Hansen and Stephen J. Turnovsky,

Econometric Society Monographs, pp. 208-257. Cambridge: Cambridge University

Press.

31

Fehr, Ernst, and Christian Zehnder. (2005). “Reputation and Credit Market Formation.”

NCCR Finrisk Working Paper No. 179, University of Zurich.

Fischbacher, Urs. (forthcoming). “z-Tree: Zurich Toolbox for Readymade Economic Ex-

periments.” Experimental Economics.

Fudenberg, Drew, and Eric Maskin. (1986). “The Folk Theorem for Repeated Games with

Discounting and Incomplete Information.” Econometrica 54, 533-54.

Galindo, Arturo, and Margaret J. Miller. (2001). “Can Credit Registries Reduce Credit

Constraints? Empirical Evidence on the Role of Credit Registries in Firm Investment

Decisions.” Paper prepared for the Annual Meetings of the Inter-American Develop-

ment Bank, Santiago Chile.

Hunt, Robert M. (2005). “A Century of Consumer Credit Reporting in America.” Federal

Reserve Bank of Philadelphia Working Paper 05-13.

Jappelli, Tullio, and Marco Pagano (2002). “Information Sharing, Lending and Defaults:

Cross-Country Evidence.” Journal of Banking and Finance 26, 2017-2045.

Jappelli, Tullio, and Marco Pagano, and Magda Bianco (2005). “Courts and Banks: effects

of Judicial Enforcement on Credit Markets.” Journal of Money, Credit, and Banking

37, 223-244.

Kallberg, Jarl G., and Gregory F. Udell (2003). “The Value of Private Sector Credit

Information.” Journal of Banking and Finance 27, 449-469.

Karlan, Dean S. (2005). “Using Experimental Economics to Measure Social Capital and

Predict Financial Decisions.” American Economic Review 95, 1688-1699.

Klein, Daniel B. (1992). “Promise Keeping in the Great Society: A Model of Credit

Information Sharing.” Economics and Politics 4(2), 117-136.

Kreps, David M., and Robert Wilson. (1982). “Reputation and Imperfect Information.”

Journal of Economic Theory 27, 253-279.

Levine, Ross. (1998). “The Legal Environment, Banks, and Long-Run Economic Growth.”

Journal of Money, Credit, and Banking 30, 596-613.

Love, Inessa, and Nataliya Mylenko. (2003). “Credit Reporting and Financing Con-

straints.” World Bank Policy Research Working Paper 3142.

Luoto, Jill, Craig McIntosh, and Bruce Wydick. (2004). “Credit Information Systems in

Less-Developed Countries: Recent History and a Test.” Working Paper University of

California at San Diego, University of San Francisco.

32

Miller, Margaret J. (2003). “Credit Reporting around the Globe.” In Credit Reporting

Systems and the International Economy, edited by Margaret J. Miller, Boston: MIT

Press.

Padilla, A. Jorge, and Marco Pagano. (1997). “Endogenous Communication among

Lenders and Entrepreneurial Incentives.” Review of Financial Studies 10, 205-236.

Padilla, A. Jorge, and Marco Pagano. (2000). “Sharing Default Information as a Borrower

Discipline Device.” European Economic Review 44(10), 1951-1980.

Pagano, Marco and Tullio Jappelli. (1993). “Information Sharing in Credit Markets.” The

Journal of Finance 43(5), 1693-1718.

Petersen, Mitchell A., and Raghuram G. Rajan. (1994). “The Benefits of Lending Rela-

tionships: Evidence from Small Business Data.” The Journal of Finance 49(1), 3-37.

Powell, Andrew, Nataliya Mylenko, Margaret J. Miller, and Giovanni Majnoni. (2004).

“Improving Credit Information, Bank Regulation and Supervision: On the Role and

Design of Public Credit Registries.” World Bank Policy Research Working Paper 3443.

Stiglitz, Joseph E., and Andrew Weiss. (1981). “Credit Rationing in Markets with Imper-

fect Information.” American Economic Review 71(3), 392-410.

Vercammen, James. A. (1995). “Credit Bureau Policy and Sustainable Reputation Effects

in Credit Markets.” Economica 62, 461-478.

World Bank (2006). Doing Business in 2007: How to Reform. Oxford: Oxford University

Press.

33

A Appendix

A.1 Model and Assumptions

There are m lenders and n < m borrowers in a game which lasts for T > 1 periods. In each period

t each lender has k units of capital to lend. Capital has an opportunity cost of 1 (repayment plus

interest) per unit. The lender can lend any part of this capital kt ∈[0, k

]in each period to any one

borrower. In order to do so, the first stage of each period is such that all lenders can simultaneously

submit a credit offer to any subset of borrowers.20 A credit offer [kt, rt] consists of a loan size kt

and a desired repayment rt (principal plus interest). At the second stage of each period borrowers

choose in random order from the available offers. In each period borrowers are free to accept one

of the available loans or not to borrow at all. The repayment of a loan cannot be enforced by

the lender. At stage three of each period the borrower can choose to either make the requested

repayment rt = rt, or not repay at all rt = 0.

The period payoff of a lender πt is calculated as follows:

πt = k − kt + rt

Each borrower has a fixed return a from self-financed projects in each period. Additionally, the

borrower can invest any capital kt borrowed in a project which yields a safe return of bkt, where

b > 1.The period payoff of a borrower vt is therefore given by:

vt(kt, rt, rt) = a + bkt − rt

There are 2 types of (non-distinguishable)borrowers: A share p are social types who suffer

mental costs g(rt, kt) if they don’t repay in cases where they perceive the financing conditions as

“fair” (i.e., in cases where the desired repayment does not exceed a certain reference value rt ≤ φk).

The rest of the borrowers are purely selfish profit-maximizers. Thus, the period utility of borrower

can be written as:

ut(kt, rt, rt) = vt − g(rt, kt)

where the mental costs g(rt, kt) are defined as follows:

g(rt, kt)social ={ ∞ if rt < rt ≤ φk

0 if rt ≥ rt or rt > φk

g(rt, kt)selfish = 0

20Since continuous auctions have defied a fully rigorous analysis so far we make this assumptionon the trading mechanism for tractability reasons.

34

As a consequence social borrowers always repay their loans as long as they have received “fair”

financing conditions in a given period. With respect to the reference repayment of social borrowers,

we assume that φ ∈ (1, (b+1)/2]. This means that the reference repayment is somewhere in between

the repayment where all the gains from trade go to the borrower and the repayment where the gains

from trade are equally split between lender and borrower.The total material surplus per trade is given by

πt(kt, rt)− k + ut(kt, rt, rt)− a = (b− 1) kt

Since we assume that b > 1, it is value maximizing if n transactions place per period, each with

the maximum credit volume k.

In the following we derive equilibria of this game for different credit market conditions. In

section A.2 we analyze a credit market, in which lenders and borrowers interact only in one-off

situations, as in our Random ID treatments. In section A.3 we investigate a credit market where

the market participants have the possibility to engage in repeat interactions, as in our Fixed ID

treatments. In each environment we examine both the case without information sharing between

banks and the case of a public credit registry with mandatory reporting.

The incomplete information nature of the game (i.e., uncertainty about the players’ types)

implies that it should be analyzed as reputation game analogous to the one described in Kreps et

al. (1982). The equilibrium concept is perfect Bayesian Equilibrium. Such games are generally

characterized by a large number of equilibria (see Fudenberg and Maskin, 1986). Our aim is not

to provide a complete formal analysis of the game, but rather to show, that there are equilibria in

the RID-CR, FID-NO and FID-CR treatments in which the reputational mechanisms intuitively

described in section 3 ensure that a substantial credit volume can be sustained.

A.2 Lending without Relationship Banking

In each period all market participants receive freshly assigned identification numbers. Lenders

therefore cannot recognize any of the borrowers, even if they have financed them before.

A.2.1 Lending without a Credit Registry

Lenders do not receive any information on the prior behavior of any borrower in the market. Each

period t can therefore be viewed as a one-period game. In the following we consequently drop the

time index t and analyze the one-period outcome. Proposition A1 establishes that in the one-off

situation lenders are only willing to extend credit if the fraction of social borrowers p is sufficiently

large.

35

Proposition A1: Consider a game of T = 1 period. If p ≥ 1φ there exists a perfect Bayesian

equilibrium in which all borrowers receive maximum credit k. If p < 1φ no credit is extended in

equilibrium.

Proof of Proposition A1: Lenders anticipate that social borrowers will repay a loan k only if

r (k) ≤ φk, while selfish borrowers will never repay a loan. The expected profit of a lender is thus:

Eπ (k, r(k)) ={

k − k + pr(k) if r(k) ≤ φkk − k if r(k) > φk

Thus only if there exists a k > 0 for which pφk > k will any lender offer any credit. This requires

p ≥ 1φ . If this condition is satisfied lenders can profitably extend credit. Due to competition among

lenders, these will earn zero profits so that pr∗(k) = k. Social borrowers thus earn u(k∗, r∗) =

a + bk − r∗(k) = a + (b− 1p)k∗. Borrowers prefer the highest possible credit level as our parameter

assumptions imply bp > 1. We have therefore established that in a one-period game the equilibrium

contract offer of lenders will be

[k∗, r∗] =

{ [k, k

p

]if p ≥ 1

φ

[0, 0] if p < 1φ

A.2.2 Lending with a Credit Registry

At the end of each period lenders are exogenously forced (legal obligation) to submit information on

the repayment behavior of their current borrower to a public credit registry. In return they receive

a credit report which states for each borrower and each past period whether the borrower got a

loan and whether he repaid his loan or not. The provision of information and access to the credit

registry information has no cost for lenders. Proposition A2 shows that even if the share of social

borrowers would lead to a market collapse in a one-shot transaction (i.e., if p < 1φ (Proposition

A1)), a public credit registry can sustain substantial credit volumes.

Proposition A2: Consider a game of T ≥ 2 periods, in which the ID numbers of borrowers and

lenders are freshly assigned in each period, but a credit registry provides a complete repayment

history of all borrowers to all lenders in each period t > 1. If 1φT < p < 1

φ the following strategies

and beliefs form a perfect bayesian equilibrium.

• In all periods all lenders offer a contract [k∗t , r∗t ] to all borrowers who always repaid in the

past. No lender offers any credit to a borrower who defaulted in any previous period j < t.

The credit size and the requested repayment in each period t are defined as follows:

36

[k∗t , r∗t ] =

[k, k

]if t < T − s[

k, kpφs

]if t = T − s[

kpblφs−l+1 , k

pblφs−l

]if t = T − s + l for all l ∈ {1, 2, ..., s}

where s is the smallest integer that satisfies p ≥ 1φs+1 .

• Social borrowers accept the contract [k∗t , r∗t ] in all periods t and repay the loan in each period.

• Selfish borrowers accept the contract [k∗t , r∗t ] in all periods t. Their repayment probability γ∗tis given by

γ∗t =

1 if t < T − s(φs−1)p

1−p if t = T − sφs−l−1

φs+1−l−1if t = T − s + l for all l ∈ {1, 2, ..., s− 1}

0 if t = T

• Lenders believe that any borrower who defaults on a loan in periods t < T − s is selfish.

Proof of Proposition A2: Proof is by construction and is established in 4 steps:

Step 1 (repayment by social borrowers): Social borrowers will repay in each period as long as their

financing conditions are fair; i.e., r∗t ≤ φk∗t . Given the strategies of lenders this condition is

satisfied in every period.

Step 2 (repayment by selfish borrowers): In period T selfish borrowers will always default. In non-

final periods t < T selfish borrowers will repay with a positive probability if their following

incentive constraint is met:−rt + V Rt+1 ≥ V D

t+1, where V Rt+1 and V D

t+1 represent the future ex-

pected utilities of a selfish borrower at the beginning of period t+1 after repaying respectively

defaulting in period t. We first consider a selfish borrower’s incentives in the next to last pe-

riod T−1: Given the lenders’ strategies above we have V RT = a+bk∗T = a+ k

pbs−1φand V D

T = a.

As r∗T−1 = kpbs−1φ

the incentive constraint is met with equality in period T−1. It is therefore a

best strategy for the selfish borrower to repay with any probability γ∗T−1 ∈ [0, 1]. Concerning

the decision in T − 2 we have V RT−1 = 2a + bk∗T−1 − r∗T−1 + bk∗T = 2a + k

pbs−2φ2 and V DT = 2a.

As r∗T−2 = kpbs−2φ2 the incentive constraint is again met with equality in period T −2 and it is

therefore a best strategy for the selfish borrower to repay with any probability γ∗T−2 ∈ [0, 1].

The same argument can be made for all periods t > T − s. In period T − s the lenders’

strategies imply that V RT−s+1 = sa+

∑sl=1 bk∗T−s+l−

∑s−1l=1 r∗T−s+l = sa+ bk∗T−s+1 = sa+ k

pφs

and V DT = sa. Since r∗T−s = k

pφs also the incentive constraint in period T − s is satisfied

with equality such that any repayment probability γ∗T−s ∈ [0, 1] is a best reply for a selfish

37

borrower. In periods t < T − s all lenders offer the contract [k∗t , r∗t ] =[k, k

]with certainty.

As r∗t = k < bk the incentive constraint is met with inequality in these periods. It is therefore

a best strategy for selfish borrowers to repay with probability γ∗t = 1 in all periods t < T −s.

Step 3 (beliefs of lenders): In equilibrium all borrowers of the same type play identical strategies.

As all lenders have access to the credit registry they have identical information concerning

borrowers’ types in each period. At the beginning of each period all lenders form a belief

about the probability of each borrower being social based on the information retrieved from

the credit registry. In equilibrium the lenders’ contract offers are such that social borrowers

always have an incentive to repay. In periods t < T − s also selfish borrowers repay with

certainty in equilibrium. As defaulting is off the equilibrium path in these periods, Bayes’

Rule does not apply and it must be specified how lenders update their beliefs in case of default:

We assume that lenders have the out-of-equilibrium belief that any borrower without a clean

record is selfish. In periods T − s ≤ t ≤ T only selfish borrowers default with a positive

probability in equilibrium such that lenders rationally believe that not repaying borrowers

are selfish. Bayesian updating implies that the belief about borrowers who have always repaid

in the past is calculated as pet =

pet−1

pet−1+(1−pe

t−1)γ∗t−1.

Step 4 (contract offers of lenders): The fact that in equilibrium all lenders have identical informa-

tion concerning the borrowers’ types implies a competitive market for clients and therefore

lenders earn zero profits in each period. The desired repayment r∗t yields zero profits if

[pet + (1− pe

t )γ∗t ] r∗t = k∗t . In the final period T selfish borrowers always default (γ∗T = 0). In

equilibrium lenders’ desired repayment in T is r∗t = φk∗t . Thus, lenders are only willing to

extend credit to borrowers for which their belief is at least peT = 1

φ . Selfish borrowers must

therefore choose their repayment probability in period T − 1 so that this necessary belief in

T is achieved: γ∗T−1 =pe

T−1(r(k)−k)

k(1−peT−1)

. However, also in period T − 1 lenders are only willing

to extend credit if they earn at least zero-profits. This implies that the repayment proba-

bility of selfish borrowers γ∗T−1 must not be too small. This is only possible if the belief in

period T − 1 is already large enough: peT−1 ≥ 1

φ2 . Exactly the same arguments apply for the

preceding periods such that we can calculate the minimally necessary belief for each period

T − j for all j ≤ s: pminT−j = 1

φj+1 . By definition the period T − s is the last period in which

the initial population fraction of social borrowers is above the minimal belief necessary for

lenders to extend credit: 1φs ≥ p ≥ 1

φs+1 . It is therefore in this period that borrowers start to

partly default, such that the minimal belief can be sustained in all subsequent periods. The

equilibrium repayment probabilities of borrowers in all periods t ≥ T − j for 0 < j ≤ s stated

in Proposition A2 are obtained by solving the following equation for γ∗t : pet

pet+(1−pe

t )γ∗t

= pmint+1 ,

where peT−s = p and pe

T−j = pminT−j .

In the early periods of the game t < T −s all borrowers repay with probability γ∗t = 1. Thus,

competition drives repayment requests down to r∗t = k∗t . In period T − s the lenders’ belief

38

may be strictly higher than the minimally necessary belief: p > 1φs+1 . Competition implies

that also in this case the requested repayment is set such that the zero-profit condition is

satisfied with equality: r∗T−s =k∗T−s

pφs . In all later periods T − s < t ≤ T the lenders’ belief is

always exactly at the threshold and accordingly the zero-profit condition can only be satisfied

if the desired repayments are set at the maximal possible level: r∗t = φk∗t . As each individual

lender earns zero-profits in every period any credit size kt ∈ [0, 1] is a best-response of a

lender to the equilibrium strategies of borrowers and other lenders in every period.

A.3 Lending with Relationship Banking

We now assume that borrowers have fixed ID numbers as in our FID-CR and FID-NO treatments,

so that lenders can identify those borrowers whom they have traded with in the past.

A.3.1 Market without a Credit Registry

Proposition A3 shows that the possibility for relationship banking allows to sustain substantial

credit volumes even in the case where there is no credit registry and the share of social borrowers

would lead to a market collapse in a one-shot transaction.

Proposition A3: Consider a game of T ≥ 2 periods, in which the ID numbers of borrowers

and lenders are fixed for all periods, but there is no credit registry. If 1φT < p < 1

φ the following

strategies and beliefs form a perfect bayesian equilibrium.

• In period 1 all lenders offer the contract [k∗t , r∗t ] =[k, max{0, k − ρ}] to all borrowers, where

ρ = (T − s− 2 + pφs)(φ− 1)k.

• In all periods 2 ≤ t ≤ T all lenders who concluded a contract in the previous period offer the

contract [k∗t , r∗t ] to their incumbent borrower if this borrower repaid in the previous period.

If the incumbent borrower of a lender defaulted in the previous period or if the lender didn’t

conclude a contract in the previous period, the lender does not offer any credit at all. The

credit size and the requested repayment in each period t are defined as follows:

[k∗t , r∗t ] =

[k, φk

]if 2 ≤ t ≤ T − s[(

φb

)lk, φ

(φb

)lk

]if t = T − s + l for all l ∈ {1, 2, ..., s}

where s is the smallest integer that satisfies p ≥ 1φs+1 .

• Social borrowers accept the contract [k∗t , r∗t ] in all periods t and repay the loan in each period.

• Selfish borrowers accept the contract [k∗t , r∗t ] in all periods t. Their repayment probability γ∗tis given by

39

γ∗t =

1 if t < T − s(φs−1)p

1−p if t = T − sφs−l−1

φs+1−l−1if t = T − s + l for all l ∈ {1, 2, ..., s− 1}

0 if t = T

• All lenders believe that any borrower who defaulted on a loan in any prior period is selfish.

Furthermore, outside lenders believe that borrowers who switch lenders in any period t > 1

are selfish.

Proof of Proposition A3: Proof is by construction and is established in 4 steps:

Step 1 (repayment by social borrowers): Social borrowers will repay in each period as long as their

financing conditions are fair; i.e., r∗t ≤ φk∗t . Given the strategies of lenders this condition is

satisfied in every period.

Step 2 (repayment by selfish borrowers): The argument is very similar to the one in Step 2 of

the Proof of Proposition A2. In period T selfish borrowers will always default. The incentive

constraint for selfish borrowers in non-final periods t < T is again :−rt + V Rt+1 ≥ V D

t+1. A

selfish borrower’s incentives in the next to last period T − 1 are as follows: The lenders’

strategies above imply that V RT = a + bk∗T = a + b

(φb

)sk = a + φsk

bs−1 and V DT = a. As

r∗T−1 = φskbs−1 the incentive constraint is met with equality in period T − 1. It is therefore a

best strategy for the selfish borrower to repay with any probability γ∗T−1 ∈ [0, 1]. Similar

to Step 2 of the Proof of Proposition A2 the same argument can be made for all periods

t ≥ T − s. In periods t < T − s all lenders offer the contract [k∗t , r∗t ] =[k, φk

]with certainty.

As r∗t = φk < bk the incentive constraint is met with inequality in these periods. It is

therefore a best strategy for selfish borrowers to repay with probability γ∗t = 1 in all periods

t < T − s.

Step 3 (beliefs of lenders): In equilibrium all borrowers of the same type play identical strategies.

However, lenders only observe the repayment behavior of borrowers with whom they directly

interact. The lenders strategies imply that they either form a bilateral relationship with a

borrowers from period 1 on, or they do not enter the credit market at all. At the beginning

of each period all lenders who are part of an ongoing relationship form a new belief about the

probability of their incumbent borrower being social based on his repayment behavior in the

last period. In equilibrium the lenders’ contract offers are such that social borrowers always

have an incentive to repay. In periods t < T − s selfish borrowers also repay with certainty

in equilibrium. As defaulting is off the equilibrium path in these periods, Bayes’ Rule does

not apply and it must be specified how lenders update their beliefs in case of default: We

assume that lenders have the out-of-equilibrium belief that any borrower who ever defaulted

40

in the past is selfish. In periods T − s ≤ t ≤ T only selfish borrowers default with a positive

probability in equilibrium such that lenders rationally believe that not repaying borrowers

are selfish. Bayesian updating implies that the belief about an incumbent borrower who has

always repaid in the past is calculated as pet =

pet−1

pet−1+(1−pe

t−1)γ∗t−1. We further assume that

lenders who are not in a relationship with a borrower believe that borrowers who switch

lenders are selfish. In equilibrium borrowers do not switch lenders such that this is an out-

of-equilibrium belief.

Step 4 (contracts of lenders): The derivation of the repayment behavior of the borrowers in

periods T − s ≤ t ≤ T and the best-response offers of lenders in periods T − s < t ≤ T is

identical as in the Proof of Proposition A2 and therefore omitted here.

Since the end-game situation for lenders is identical as in Proposition A2 lenders make zero

profits in all periods T − s < t ≤ T . However, in periods 2 ≤ t ≤ T − s those lenders who are

in a relationship with a borrower earn rents. The reason is that outside lenders have the out-

of-equilibrium belief that switching borrowers are selfish and consequently do not make any

credit offers. In these periods all borrowers repay with certainty and as lenders do not have

to fear competition from other lenders they extend the maximal credit volume k∗t = k and

ask for the highest possible repayment r∗t = φk∗t . As a consequence the lenders make positive

profits. In contrast, in period 1 those lenders anticipate that only lenders who conclude a

contract in the first period succeed in establishing a relationship. Thus, the excess supply of

lenders implies that lenders compete for borrowers in the first period. The rent from periods

2 ≤ t ≤ T −s for lenders who establish a relationship is equal to ρ = (T −s−2+pφs)(φ−1)k.

Competition implies that lenders are willing to make losses in the first period as long as these

losses are not larger than the rent from a relationship. Thus, lenders offer contracts of the

following form in the first period: [k∗t , r∗t ] =[k, max{0, k − ρ}].

A.4 Applicability of Propositions A2 and A3 to Experiment

In our experiment lenders could only choose credit sizes from the set {0, 5, 10, ..., 50}. The

limited number of possible credit sizes may render it impossible that lenders choose the credit

size such that borrowers are always indifferent in the end-game. However, alternatively we

could also write down equilibria in which lenders do not lower the credit size in the end-

game but instead offer the maximal credit only with a certain probability to borrowers who

have always repaid in the past (or a combination of randomization and decreasing credit

volumes). As probabilities can always be chosen continuously the discrete set of repayments

would no longer be a problem. However, as we believe that the reduction of credit volumes

is more realistic and the qualitative predictions of the equilibria remain the same we decided

to present these more intuitive equilibria.

41

Table 1: Experimental Treatments

Relationship Banking

Not Possible Possible(Random ID) (Fixed ID)

Credit ReportingYes RID-CR FID-CR

No RID-NO FID-NO

42

Table 2: Summary statistics: Random ID treatments

Treatment RID-CR RID-NOmean max min mean max min

Repayment rate 0.80 0.87 0.70 0.28 0.39 0.16

Realized contracts (%) 93.7 96.4 90.7 58.7 78.6 40.0Credit size 41.0 44.1 36.3 22.8 27.2 18.3Credit volume (%) 76.8 83.7 65.9 26.6 36.4 15.6

Interest rate (%) 33.3 48.1 24.5 33.1 38.4 29.9Payoff lender 50.6 52.7 47.7 44.0 45.4 42.2Payoff borrower 47.6 52.4 39.5 31.8 39.3 24.4

Notes: The table presents summary statistics for the RID-CR and RID-NO treat-ments. For each statistic the table presents the mean value across the 5 sessions ofeach treatment. In addition the table presents the highest mean value (max) andlowest mean value (min) of an individual session. Repayment rate is the share ofcredit transactions in which the borrower made the desired repayment of the lender.Realized contracts is the percentage of potential contracts (7 per session and period)which were realized. Credit size is the mean size of credit extended in realized credittransactions. Credit volume is the percentage of the potential credit volume (7 x50 = 350 per session and period) which was realized. Interest rate is the mean ofthe calculated interest rate per transaction: (desired repayment- credit size)

credit size · 100. Pay-off lender is the payoff per lender and period. Payoff borrower is the payoff perborrower and period.

43

Tab

le3:

Rep

aym

ent

his

tory

and

cred

itac

cess

:R

andom

IDtr

eatm

ents

RID

-CR

RID

-NO

(1)

(2)

(3)

(4)

(5)

(6)

Con

trac

tC

redit

size

Inte

rest

rate

Con

trac

tC

redit

size

Inte

rest

rate

Pri

orre

pay

men

tra

te1.

256*

*20

.474

**-2

0.18

*-0

.083

0.91

4-3

.009

(3.5

4)(4

.61)

(4.4

3)(0

.23)

(0.6

0)(0

.93)

Per

iod

6-10

0.45

54.

768*

-2.0

75-0

.414

*-5

.463

2.74

4(1

.62)

(4.5

7)(1

.34)

(2.0

5)(1

.24)

(0.6

4)Per

iod

11-1

50.

415

5.95

8**

-5.0

23*

-0.8

74**

-8.0

45-2

.338

(1.2

3)(5

.69)

(2.9

4)(5

.85)

(2.7

5)(0

.42)

Per

iod

16-2

0-0

.984

**1.

000

18.5

82-1

.303

**-7

.502

*-1

0.79

8(9

.34)

(0.6

1)(0

.81)

(5.0

5)(3

.03)

(1.9

1)C

onst

ant

0.80

420

.166

*48

.9**

0.89

2**

26.2

87**

35.3

28**

(1.7

0)(3

.42)

(15.

43)

(2.9

5)(1

2.68

)(9

.40)

Obse

rvat

ions

663

621

621

665

376

376

R-s

quar

ed0.

180.

020.

070.

02

Not

es:

The

tabl

ere

port

sre

gres

sion

esti

mat

esus

ing

cred

itco

ntra

ctda

tape

rbo

rrow

eran

dpe

riod

from

the

RID

-CR

and

RID

-N

Otr

eatm

ents

.C

olum

n(1

)an

d(4

)re

port

prob

ites

tim

ates

forC

ontrac

t(P

roba

bilit

yof

seal

ing

aco

ntra

ct).

Col

umn

(2)an

d(5

)re

port

OLS

esti

mat

esfo

rC

redi

tsi

ze(S

ize

ofcr

edit

rece

ived

)us

ing

data

only

for

borr

ower

sw

hodi

dse

ala

cred

itco

ntra

ctin

apa

rtic

ular

peri

od.

Col

umn

(3)

and

(6)

repo

rtO

LS

esti

mat

esfo

rIn

tere

stra

te(d

esi

red

repaym

ent

-C

redit

size)

Cre

dit

size

·100

,ag

ain

usin

gda

taon

lyfo

rbo

rrow

ers

who

did

seal

acr

edit

cont

ract

ina

part

icul

arpe

riod

.A

llsi

xre

gres

sion

ses

tim

ate

coeffi

cien

tsfo

rth

efo

llow

ing

expl

anat

ory

vari

able

s:Pri

orre

paym

ent

rate

isth

era

tio

ofpr

ior

cred

ittr

ansa

ctio

nsin

whi

chth

ebo

rrow

erm

ade

the

desi

red

repa

ymen

tof

the

lend

er.

Per

iod

6-10

isa

dum

my

vari

able

for

allt

rans

acti

ons

whi

chto

okpl

ace

inpe

riod

s6-

10.

Per

iod

11-1

5is

adu

mm

yva

riab

lefo

ral

ltr

ansa

ctio

nsw

hich

took

plac

ein

peri

ods

11-1

5.Per

iod

16-2

0is

adu

mm

yva

riab

lefo

ral

ltra

nsac

tion

sw

hich

took

plac

ein

peri

ods

16-2

0.In

allr

egre

ssio

nsth

eT

-sta

tist

ics

repo

rted

inpa

rent

hese

sar

eba

sed

onst

anda

rder

rors

adju

sted

for

clus

teri

ngat

sess

ion

leve

l.*

indi

cate

ssi

gnifi

canc

eat

the

5pe

rcen

tle

vel;

**in

dica

tes

sign

ifica

nce

atth

e1

perc

ent

leve

l.

44

Table 4: Repayment behavior: Random ID treatments

(1) (2)RID-CR RID-NO

Interest rate -0.046** -0.013**(2.95) (5.61)

Credit size -0.011 -0.008(1.06) (0.91)

Period 6-10 -0.021 0.06(0.12) (0.41)

Period 11-15 -0.459** -0.34(2.83) (1.45)

Period 16-20 -1.296** -0.199(6.48) (0.60)

Constant 3.226** 0.126(3.56) (0.48)

Observations 656 411

Notes: The table reports probit estimates of the repayment choice per borrowerand period for the RID-CR and RID-NO treatments. The dependant variable inboth columns is a dummy variable which is 1 only if the borrower chooses to makethe desired repayment of the lender. Both regressions estimate coefficients for thefollowing explanatory variables: Interest rate: (desired repayment - Credit size)

Credit size · 100.Credit size: Size of credit received. Period 6-10 is a dummy variable for alltransactions which took place in periods 6-10. Period 11-15 is a dummy variablefor all transactions which took place in periods 11-15. Period 16-20 is a dummyvariable for all transactions which took place in periods 16-20. In both regressionsthe T-statistics reported in parentheses are based on standard errors adjustedfor clustering at session level. * indicates significance at the 5 percent level; **indicates significance at the 1 percent level.

45

Table 5: Summary statistics: Fixed ID treatments

Treatment FID-CR FID-NOmean max min mean max min

Repayment rate 0.79 0.86 0.72 0.74 0.79 0.67

Realized contracts (%) 93.9 98.6 89.3 91.6 97.9 85.0Credit size 42.2 43.3 41.5 40.3 42.6 37.8Credit volume (%) 79.3 82.4 75.6 73.9 81.4 68.6

Interest rate (%) 29.2 36.7 23.1 25.9 31.4 21.5Payoff lender 49.9 52.4 47.8 48.5 49.2 47.7Payoff borrower 49.8 51.7 47.7 49.1 51.9 46.7

Contract renewal ratio 0.39 0.55 0.23 0.48 0.56 0.35

Notes: The table presents summary statistics for the FID-CR and FID-NO treat-ments. For each statistic the table presents the mean value across the 5 sessions ofeach treatment. In addition the table presents the highest mean value (max) andlowest mean value (min) of an individual session. Repayment rate is the share ofcredit transactions in which the borrower made the desired repayment of the lender.Realized contracts is the percentage of potential contracts (7 per session and period)which were realized. Credit size is the mean size of credit extended in realized credittransactions. Credit volume is the percentage of the potential credit volume (7 x50 = 350 per session and period) which was realized. Interest rate is the mean ofthe calculated interest rate per transaction: (desired repayment- credit size)

credit size · 100. Pay-off lender is the payoff per lender and period. Payoff borrower is the payoff perborrower and period.

46

Tab

le6:

Rep

aym

ent

his

tory

and

cred

itac

cess

:Fix

edID

trea

tmen

ts

FID

-CR

FID

-NO

(1)

(2)

(3)

(4)

(5)

(6)

Con

trac

tC

redit

size

Inte

rest

rate

Con

trac

tC

redit

size

Inte

rest

rate

Pri

orre

pay

men

tra

te2.

671*

*24

.507

**-3

2.49

5*3.

378*

*20

.914

**-2

3.36

3*(9

.33)

(10.

51)

(4.0

3)(7

.55)

(6.0

0)(2

.93)

Per

iod

6-10

-0.6

252.

422*

-2.8

61-0

.456

4.17

8**

1.07

4(1

.56)

(2.9

6)(2

.16)

(1.0

6)(4

.98)

(0.8

4)Per

iod

11-1

5-0

.651

*4.

608*

*-5

.186

**-0

.982

*4.

333*

*0.

432

(2.1

4)(7

.42)

(4.8

9)(2

.20)

(7.3

3)(0

.18)

Per

iod

16-2

0-1

.821

**3.

213

-5.2

07-1

.964

**3.

684*

0.02

9(4

.83)

(1.6

5)(1

.97)

(4.2

8)(3

.92)

(0.0

1)C

onst

ant

0.68

219

.624

**59

.489

**0.

364

21.2

94**

43.7

36**

(1.6

5)(7

.00)

(17.

04)

(1.5

1)(7

.18)

(7.8

0)

Obse

rvat

ions

665

622

622

665

606

606

R-s

quar

ed0.

270.

130.

210.

11

Not

es:

The

tabl

ere

port

sre

gres

sion

esti

mat

esus

ing

cred

itac

cess

data

per

borr

ower

and

peri

odfr

omth

eFID

-CR

and

FID

-N

Otr

eatm

ents

.C

olum

n(1

)an

d(4

)re

port

prob

ites

tim

ates

forC

ontrac

t(P

roba

bilit

yof

seal

ing

aco

ntra

ct).

Col

umn

(2)an

d(5

)re

port

OLS

esti

mat

esfo

rC

redi

tsi

ze(S

ize

ofcr

edit

rece

ived

)us

ing

data

only

for

borr

ower

sw

hodi

dse

ala

cred

itco

ntra

ctin

apa

rtic

ular

peri

od.

Col

umn

(3)

and

(6)

repo

rtO

LS

esti

mat

esfo

rIn

tere

stra

te(d

esi

red

repaym

ent

-C

redit

size)

Cre

dit

size

·100

,ag

ain

usin

gda

taon

lyfo

rbo

rrow

ers

who

did

seal

acr

edit

cont

ract

ina

part

icul

arpe

riod

.A

llsi

xre

gres

sion

ses

tim

ate

coeffi

cien

tsfo

rth

efo

llow

ing

expl

anat

ory

vari

able

s:Pri

orre

paym

ent

rate

isth

era

tio

ofpr

ior

cred

ittr

ansa

ctio

nsin

whi

chth

ebo

rrow

erm

ade

the

desi

red

repa

ymen

tof

the

lend

er.

Per

iod

6-10

isa

dum

my

vari

able

for

allt

rans

acti

ons

whi

chto

okpl

ace

inpe

riod

s6-

10.

Per

iod

11-1

5is

adu

mm

yva

riab

lefo

ral

ltr

ansa

ctio

nsw

hich

took

plac

ein

peri

ods

11-1

5.Per

iod

16-2

0is

adu

mm

yva

riab

lefo

ral

ltra

nsac

tion

sw

hich

took

plac

ein

peri

ods

16-2

0.In

allr

egre

ssio

nsth

eT

-sta

tist

ics

repo

rted

inpa

rent

hese

sar

eba

sed

onst

anda

rder

rors

adju

sted

for

clus

teri

ngat

sess

ion

leve

l.*

indi

cate

ssi

gnifi

canc

eat

the

5pe

rcen

tle

vel;

**in

dica

tes

sign

ifica

nce

atth

e1

perc

ent

leve

l.

47

Table 7: Repayment behavior: Fixed ID treatments

(1) (2)FID-CR FID-NO

Interest rate -0.02** -0.032**(4.04) (4.98)

Credit size -0.006 0.019**(0.85) (2.81)

Period 6-10 0.085 0.03(0.82) (0.18)

Period 11-15 -0.042 0.152(0.35) (0.63)

Period 16-20 -0.415* -0.427*(2.12) (2.54)

Constant 1.79** 0.85(3.53) (1.92)

Observations 657 641

Notes: The table reports probit estimates of the repayment choice per borrowerand period for the FID-CR and FID-NO treatments. The dependant variable inboth columns is a dummy variable which is 1 only if the borrower chooses to makethe desired repayment of the lender. Both regressions estimate coefficients for thefollowing explanatory variables: Interest rate: (desired repayment - Credit size)

Credit size · 100.Credit size: Size of credit received. Period 6-10 is a dummy variable for alltransactions which took place in periods 6-10. Period 11-15 is a dummy variablefor all transactions which took place in periods 11-15. Period 16-20 is a dummyvariable for all transactions which took place in periods 16-20. In both regressionsthe T-statistics reported in parentheses are based on standard errors adjustedfor clustering at session level. * indicates significance at the 5 percent level; **indicates significance at the 1 percent level.

48

Table 8: Relationships and lenders’ rents in Fixed ID treatments

(1) (2) (3)FID-CR FID-NO FID-CR and FID-NO

Short-term -3.039 -9.31** -10.295**(1.8) (5.21) (5.5)

Long-term 5.391* 10.25** 9.106**(3.17) (6.21) (8.52)

Period 6-10 1.875 -2.763 -0.262(2.07) (1.23) (0.2)

Period 11-15 -0.402 -2.282 -1.18(0.33) (1.22) (1.14)

Period 16-20 -7.157 -7.644** -7.250**(2.21) (4.88) (4.14)

CR*Short-term 7.967*(2.89)

CR*Long-term -2.886*(2.28)

Observations 657 641 1298R-squared 0.04 0.20 0.12

Notes: The table reports OLS estimates using payoff data per lender and period fromthe FID-CR and FID-NO treatments. The dependent variable in all three columns islender profit per period (lender profit = payoff - 50). Column (1) uses data from theFID-CR treatment, Column (2) uses data from the FID-NO treatment and Column(3) uses data from both treatments. Columns (1) and (2) estimate coefficients for thefollowing explanatory variables: Short-term is a dummy variable which is 1 only fortransactions where the final duration of the relation between a lender and his currentborrower is less than 3 periods. Long-term is a dummy variable which is 1 only fortransactions where the final duration of the relation between a lender and his currentborrower is at least 3 periods. Period 6-10 is a dummy variable for all transactionswhich took place in periods 6-10. Period 11-15 is a dummy variable for all transactionswhich took place in periods 11-15. Period 16-20 is a dummy variable for all transactionswhich took place in periods 16-20. Column (3) additionally estimates coefficients forthe following explanatory variables: CR*Short-term is a dummy variable which is 1only for all short-term transactions which took place in FID-CR treatment. CR*Long-term is a dummy variable which is 1 only for all long-term transactions which tookplace in FID-CR treatment. All three regressions are estimated without constants,and the T-statistics reported in parentheses are based on standard errors adjusted forclustering at session level. * indicates significance at the 5 percent level; ** indicatessignificance at the 1 percent level.

49

Figure 1: Repayment rate and credit volume: Random ID treatments

0

10

20

30

40

50

60

70

80

90

100

0 5 10 15 20 0 5 10 15 20

Repayment Rate Credit Volume

RID−NO RID−CR

Per

cent

Period Period

50

Figure 2: Repayment rate and credit volume: Fixed ID treatments

0

10

20

30

40

50

60

70

80

90

100

0 5 10 15 20 0 5 10 15 20

Repayment Rate Credit Volume

FID−NO FID−CR

Per

cent

Period Period

51

Credit Reporting, Relationship Banking, and Loan Repayment

Documents