Moral Hazard and Peer Monitoring in a Laboratory Microfinance Experiment *

Transcription

1 Moral Hazard and Peer Montorng n a Laboratory Mcrofnance Experment * Tmothy N. Cason, Lata Gangadharan and Pushkar Matra February 2007 Abstract Most problems wth formal sector credt lendng to the poor n developng countres can be attrbuted to the lack of nformaton and nadequate collateral. One common feature of successful credt mechansms s group-lendng, where the loan s advanced to an ndvdual f he/she s a part of a group and members of the borrowng group can montor each other. Snce group members have better nformaton about each other compared to lenders, peer montorng s less expensve than lender montorng. Theoretcally ths leads to greater montorng and greater rates of loan repayments. Ths paper reports the results from a laboratory experment of group lendng n the presence of moral hazard and (costly) peer montorng. We compare peer montorng treatments when credt s provded to members of the group sequentally and smultaneously, and a case of ndvdual lendng wth bank montorng. Our results suggest that peer montorng results n hgher loan frequences, hgher montorng and hgher repayment rates compared to bank montorng. Although the dynamc ncentves provded by sequental leadng generate the greatest equlbrum effcency, smultaneous group leadng provdes equvalent emprcal performance. JEL Classfcaton: G21, C92, O2. Key words: Group Lendng, Montorng, Moral Hazard, Laboratory Experment, Loans, Development * We have benefted from extensve dscussons wth Dyut Banerjee and from comments by Chkako Yamauch, partcpants at the Expermental Economcs Workshop at the Unversty of Melbourne, Behavoral Economcs workng group at Monash Unversty, ESA Internatonal Meetngs n Atlanta, ESA European Meetngs n Nottngham, the NEUDC Conference at Cornell Unversty and semnar partcpants at Jadavpur Unversty, Kolkata and at the Australan Natonal Unversty, Canberra. We would lke to thank Gautam Gupta for hs assstance n organzng the sessons at Jadavpur Unversty, Kolkata. Tana Dey, Smon Hone, Vnod Mshra and Roman Sheremeta provded excellent research assstance. The usual caveat apples. Tmothy Cason, Department of Economcs, Krannert School of Management, Purdue Unversty, West Lafayette, IN , USA. Emal: cason@purdue.edu Lata Gangadharan, Department of Economcs, Unversty of Melbourne, VIC 3010, Australa. Emal: latag@unmelb.edu.au Pushkar Matra, Department of Economcs, Monash Unversty, Clayton Campus, VIC 3800, Australa. Emal: Pushkar.Matra@Buseco.monash.edu.au

2 1. Introducton There now exsts a sgnfcant body of research that examnes the falure of formal sector credt lendng programs amed at the poor n developng countres. Evdence of ths falure s shown n the nablty to reach target groups and low overall repayment rates. Ths falure s attrbuted prmarly to asymmetrc nformaton (both adverse selecton and moral hazard) and nadequate enforcement. 1 The last few decades have, however, wtnessed the development of nnovatve and hghly successful mechansms for the provson of credt to the poor. The most common of these s group-lendng. Rather than the bank (or the lender) makng a loan to an ndvdual who s solely responsble for ts repayment, the bank makes a loan to an ndvdual who s a member of a group and the group s jontly lable for each member s loans. In partcular, f the group as a whole s unable to repay the loan because some members default on ther repayment, all members of the group are nelgble for future credt. The Grameen Bank n Bangladesh s possbly the most well known of such group lendng programs. The repayment rate n ths lendng scheme s around 92 percent, and less than 5 percent of loan recpents were outsde the target group (Morduch, 1999). The success of the Grameen Bank has led polcy makers and NGO s around the world to ntroduce smlar schemes. Around 100 mllon people are estmated to have partcpated n some form of a mcrofnance project (see Gne, Jakela, Karlan and Morduch, 2005). The 2006 Nobel Prze for Peace to mcrofnance poneer Muhammed Yunus has also put the success of mcrofnance n the world spotlght. Mcro-lendng s also movng from non-proft towards a proft-makng enterprse, wth bg banks such as Ctgroup now backng such loans (Bellman, 2006). 2 1 For example, t has been argued that the percentage of nelgble benefcares n the Integrated Rural Development Program (IRDP) n Inda, one of the largest programs of provson of formal sector credt to the poor n the world, was between 15 and 26 percent, wth the hghest reported beng 50 percent. The repayment rate for IRDP loans was only about 40 percent for the whole country (see Pulley, 1989). 2 Whle mcrofnance programs are most wdespread n less developed countres they are by no means confned to them. Mcrofnance programs have been ntroduced n transton economes lke Bosna and Russa and even

3 The success of these group lendng programs arses, n part, because they are better able to address the enforcement and nformatonal problems that generally plague formal sector credt n developng countres. Group lendng programs help solve the enforcement problem through peer montorng. Stgltz (1990) and Varan (1990) argue that snce group members are lkely to have better nformaton compared to an outsder (the bank), peer montorng s relatvely cheaper compared to bank montorng, leadng to greater montorng and hence greater repayment. Banerjee, Besley and Gunnane (1994) argue that explanatons based on peer montorng do a better job of explanng the success of group lendng programs than other explanatons. Ghatak and Gunnane (1999) develop a model of moral hazard and montorng and fnd that f the socal sanctons are effectve enough or f montorng costs are low enough, the jont lablty provded by group lendng mproves repayment rates. Chowdhury (2005), by contrast, s less optmstc. He fnds that n the absence of sequental fnancng or lender montorng, group lendng programs wll typcally nvolve undermontorng wth the borrowers nvestng n undesrable projects. 3 The emprcal evdence on these ssues, unfortunately, s rather lmted. The theoretcal propostons and results are often supported by anecdotal evdence but these results have not been establshed as emprcal regulartes. In recent years researchers have called for well desgned economc experments to help examne the roles of varous mechansms that drve performance n mcrofnance programs (Morduch, 1999, Armendarz de Aghon and Morduch, 2005). The am of ths paper s to understand specfc aspects of group lendng schemes, usng controlled expermental methods. We report the results from a laboratory experment of n developed countres lke Australa, Canada and the US. See for example Conln (1999), Armendarz de Aghon and Morduch (2000, 2005), and Fry, Mhajlo, Russell and Brooks (2006). 3 How group lendng solves the problem of adverse selecton s analysed by Ghatak (1999, 2000), Van Tassel (1999), Armendarz de Aghon and Goller (2000). The argument s based on endogenous group formaton (and postve assortatve matchng): group lendng wth jont lablty wll result n self selecton wth safe borrowers clubbng together and screenng out rsky borrowers. 2

4 group lendng n the presence of moral hazard and (costly) peer montorng. 4 We compare treatments when credt s provded to members of the group sequentally and smultaneously, as well as a benchmark treatment when loans are gven to ndvduals and montored by lenders. Our work complements the growng body of research that can broadly be characterzed as feld experments n mcrofnance (see for example Gne, Jakela, Karlan and Morduch, 2006; Gne and Karlan, 2006, Kono, 2006, Cassar, Crowley and Wydck, 2007). The laboratory approach that we use n ths paper can address ssues n dfferent ways compared to feld experments. It s dffcult to vary specfc propertes of nsttutons n controlled experments n the feld due to problems of replcablty, data accessblty and comparablty (see for example Bolnk, 1988 and Hulme, 2000). Furthermore some relevant varables, such as actual montorng costs, reman unobserved. The laboratory approach on the other hand can help us control for specfc parameters and observe behavor n smulated mcrofnance nsttutons. In our case t can help n solatng and clarfyng the mpact of dfferent desgn features on repayment rates and project choce, by mplementng an envronment that s carefully algned wth the theoretcal models relatng to moral hazard and peer montorng n mcrofnance programs. Of course, the laboratory approach has some drawbacks. For example, whle the laboratory experment ncluded human subject behavor, the subjects are unversty students makng decsons for relatvely low stakes. 5 In feld experments, by contrast, partcpants are often the actual borrowers who are nvolved n mcrofnance programs. Ths advantage of feld experments comes at the cost of some loss of expermental control, however. For example, spllover effects could exst from one vllage to another or from the treatment group to the control group, creatng more nose n the data. 4 In ths paper, we focus on nformatonal asymmetres due to moral hazard and not due to adverse selecton. In partcular we restrct ourselves to exogenously formed groups (wth random re-matchng) and leave the ssue of endogenous group formaton (postve assortatve matchng) for future research. 5 We do, however, employ subjects both from a developed (Australa) and a developng (Inda) country to measure possble subject pool effects, and fnd vrtually none. 3

5 Laboratory experments that examne the mpact of specfc desgn features on performance of mcrofnance models are rare. Abbnk, Irlenbusch and Renner (2006) and Seddk and Ayed (2005) both examne the role of group selecton n the context of group lendng. Both experments are desgned as nvestment games where each group member nvests n an ndvdual rsky project whose outcome s known only to the ndvdual, and both fnd that self-selected groups have a greater wllngness to contrbute. Nether of these papers analyze the role of peer montorng. Our experment examnes several aspects of group lendng programs. The frst s the argument that sequental lendng s crucal to the success of group lendng schemes. The Grameen Bank, for example, adopts ths knd of a lendng polcy: groups have fve members each and loans are ntally gven to two randomly chosen members, to be repad n regular nstallments over a perod of one year. If they pay ther ntal nstallments, then two more borrowers n the group receve the loan and so on. Ray (1998) argues that ths knd of sequental lendng mnmzes the contagon effect assocated wth ndvdual default. Sequental lendng can also mnmze the potental of coordnaton falure. Anket (2006) and Chowdhury (2005) argue that n a smultaneous group lendng scheme wth jont lablty and costly montorng, peer montorng by borrowers alone s nsuffcent and that sequental lendng that ncorporates dynamc ncentves s requred. 6 Our experment examnes the emprcal valdty of these predctons by comparng the performance of sequental lendng and smultaneous lendng n the presence of moral hazard and costly peer montorng The second ssue s whether peer montorng ndeed does better than actve lender montorng. The bank or the lender n general s an outsder who often has less nformaton about the borrowers. Therefore montorng ndvdual borrowers could be very expensve for 6 Dynamc ncentves mean that banks make future loan accessblty contngent on full repayment of the current loan to prevent strategc default. 4

6 the lender. Specfcally, we study whether n the presence of moral hazard, group lendng wth peer montorng does better than ndvdual lendng wth bank montorng. 7 The thrd ssue s the relatve benefts of ndvdual and group lendng. Over the years there has been a dscernble shft from group lendng to ndvdual lendng n mcrofnance programs and there are a number of theoretcal reasons that have been advanced to explan ths shft. 8 Frst, clents often dslke tensons caused by group lendng. Second, low qualty clents can free-rde off hgh qualty clents leadng to an ncrease n default rates. Thrd, group lendng can be more costly for the clents as they often end up repayng the loans of ther peers. Theoretcally the results are mxed. 9 Our laboratory experment s able to address each of these ssues. Our results show that group lendng (wth peer montorng) performs better compared to ndvdual lendng (wth actve lender montorng), reflected n hgher loan frequences and repayment rates. Ths occurs even though repayment rates wth ndvdual lendng consderably exceed the theoretcal predcton, whch may reflect socal preferences such as recprocty. These results dffer from those observed n the feld by Gne and Karlan (2006) and Kono (2006), who fnd hgh performance for ndvdual lendng. Ther explanaton s based on the argument of Gref (1994), who argues that a more ndvdualstc socety requres less nformaton among players and s thus able to grow faster. 10 We also fnd that wthn group lendng t does not 7 Peer montorng and peer enforcement have been observed to deter free rdng n several experments relatng to other socal dlemma stuatons, such as common pool resource envronments and the voluntary provson of publc goods. See Fehr and Gaechter (2000), Barr (2001), Masclet, Noussar, Tucker and Vlleval (2003), Walker and Halloran (2004), and Carpenter, Bowles and Gnts (2006) for expermental evdence. 8 The terms ndvdual and group lendng as defned n ths paper essentally correspond to the terms ndvdual and group (jont) lablty. We use the term group lendng to descrbe the stuaton where ndvduals are both borrowers and smultaneously guarantors of ther partners loans. 9 Armendarz de Aghon and Morduch (2000, 2005) argue that group lendng (jont lablty) s just one element n successful mcrofnance schemes. Chowdhury (2005) argues that mere jont lablty does not work and he emphaszes the role of dynamc ncentves: n hs model a combnaton of jont lablty and dynamc ncentves work best n terms of project choce and repayment. Che (2002) argues that jont lablty schemes create problems of free-rdng and worsen repayment rates, but when projects are repeated multple tmes, group lendng domnates ndvdual lendng. Ra and Sjostrom (2004) emphasze the mportance of cross-reportng n achevng effcency n group lendng. 10 We would lke to pont out a specfc desgn feature of the feld experment conducted by Gne and Karlan (2006), whch could at least partly explan the dfferent results. In that experment, the exstng feld centres wth 5

7 matter whether loans are made smultaneously or sequentally. Although the dynamc ncentves provded by sequental lendng can mprove effcency relatve to smultaneous group lendng, performance s equvalently hgh n the two group lendng treatments because agents tend to play the effcent equlbrum n the smultaneous case. 2. Theoretcal Framework Consder a scenaro where two borrowers requre one unt of captal (say $1) each for nvestng n a partcular project. The bank, whch provdes ths captal n the form of a loan, can ether make the loan to an ndvdual (ndvdual lendng) or t can loan to the borrowers as a group (group lendng). Jont lablty for the repayment of the loan exsts n the case of group lendng. Borrowers can nvest n two dfferent types of projects: one project has a large verfable ncome and no non-verfable prvate beneft, whle the other has a large nonverfable prvate beneft and no verfable ncome. The bank prefers the frst project, where t can recoup ts nvestment, but the borrowers prefer the second one. In the absence of montorng, the borrowers wll choose to nvest n the second project and the bank, knowng ths, wll choose not to make the loan. Let us elaborate on the model, whch follows Chowdhury (2005) and Ghatak and Gunnane (1999). Suppose that there are two borrowers: B 1 and B 2. Two projects are avalable to each borrower: project S (verfable) and project R (non-verfable). If Project S s chosen, the return s H (verfable by montorng) and f project R s chosen, then the return s b (not verfable) wth b < H. The 1 dollar cost of each project s fnanced by a loan from the bank (or a lender) snce the borrowers do not have any funds of ther own. When the two borrowers ( B 1 and B 2 ) borrow together as a group, each borrower receves 1 dollar from the group lablty loans were converted to ndvdual lablty loans. Lenders therefore had pror nformaton about the borrowers characterstcs from the group lendng feld sessons and ths could be used n the ndvdual lendng sessons at no extra cost. As a result the montorng costs dd not necessarly change as they moved from group lendng to ndvdual lendng. Furthermore, partcpants had some experence wth group lendng before branchng off on ther own n the ndvdual lendng schemes. These reasons could have contrbuted towards the hgher performance outcomes n the ndvdual lablty programs n ther study. 6

8 lender. The amount to be repad s r ( > 1) n the case of ndvdual lendng or 2r n the case of group lendng. We assume that ths r s fxed exogenously. In the case of the ndvdual lendng, f the borrower chooses project S the return to the bank s r ; otherwse t s 0. The return to the borrower s H r f the borrower chooses project S, and s b f the borrower chooses project R. We assume that H r < b so that borrowers prefer project R. Banks on the other hand prefer project S. In the case of group lendng, f both borrowers choose project S, the return to each borrower s H r and the return to the bank s 2r. If both borrowers choose project R, the return to each borrower s b and the return to the bank s 0. Fnally f one borrower chooses project R and the other chooses project S, then due to jont lablty the return to the borrower choosng project S s 0 whle that of the borrower choosng project R s b and the return to the bank s H. We assume that H 2r. In the case of group lendng t s therefore n the nterest of both the bank and the borrowers to ensure that the other member of the group chooses project S. An nformatonal asymmetry arses because each borrower knows the type of hs own project, but the lender or the other borrower n the group (the partner) can fnd out the borrower s project choce only wth costly montorng. The montorng process works as follows: Borrower can, by spendng an amount c( m ) n montorng costs, obtan nformaton about the project chosen by the other borrower n hs group wth probablty m [ 0,1]. Ths nformaton can be used by borrower to ensure that the other borrower n the group chooses project S. The bank (lender) can also acqure ths same nformaton by spendng an amount λ c( m). We assume that λ 1 n order to capture the noton that peer montorng s less expensve than montorng by the bank. We assume a quadratc montorng cost functon so that ( ) 2 m c m =. Montorng level m costs the bank 2 λ m 2. If the 2 7

9 borrower chooses montorng level m, then wth probablty m he can force the other borrower n the group to choose project S. 11 Indvdual Lendng Frst consder ndvdual lendng (wth bank montorng). There are three stages to the game. Stage 1: Bank chooses whether or not to lend $1 to the borrower. Stage 2: Bank chooses the level of montorng, condtonal on decdng to lend. Stage 3: Borrower chooses ether project R or project S. It s straghtforward to solve for the sub game perfect Nash equlbrum of the game by backward nducton. If the bank lends, t chooses m to maxmze λm mr 2 2 1, whch gves m * = r. Therefore the expected return to the bank s λ 2 r 1, so the bank wll provde 2λ the loan f and only f 2 r 1 0 2λ >.e. f r 2 > 2λ. Ths mples that ndvdual lendng s 2 feasble only f the costs of montorng relatve to the return are suffcently low: λ < r. 2 Group Lendng: Smultaneous The sequence of events n group lendng s as follows: Stage 1: Bank chooses whether or not to lend $2 to the group. There s jont lablty, so that f one borrower fals to meet hs oblgatons, then f the other borrower has verfable ncome he must pay back the bank for both borrowers. Stage 2: The borrowers smultaneously choose the level of peer montorng, Stage 3: Both borrowers choose ether project R or project S. m. 11 We could thnk of dfferent ways n whch montorng works n practce: nformaton acqured by the borrowers about each other s project choce may be passed on to the lender who then uses ths nformaton to force the borrowers to choose project S. Alternatvely, the borrowers can use some form of socal sanctons or peer punshment to ensure that the other borrower chooses project S. 8

10 Note that here both montorng and lendng s smultaneous and we call ths smultaneous group lendng. Agan the sub game perfect Nash equlbrum s solved by backward nducton. Borrower wll choose montorng m to maxmze m m mj( H r) ( 1 mj) b ( 1 m) mj*0 ( 1 mj) b The frst order condton s: m ( H r) m 0 j =. Lkewse the frst order condton for 2 borrower j s: ( ) 0 m H r m =. Clearly j m = m = 0 s a Nash equlbrum. We call ths * * j the neffcent (zero-montorng/zero-lendng) equlbrum. In ths case there s a strategc complementarty between the montorng levels of the two borrowers. A partcular borrower knows that f the other borrower montors and he does not, then he wll end up wth a payoff of 0. If however the other borrower does not montor then he has no ncentve to montor as well. Mere jont lablty and peer montorng does not solve the moral hazard problem. Remember however that m [ 0,1]. Now consder the reacton functon j ( ) m = m H r of borrower wth respect to that of borrower j. Snce H r > 1 (the return on project S exceeds the amount that must be repad), then there exsts a m = m < 1 (say) such that the best response s m = 1 for mj mj wth respect to that of borrower j can be wrtten as:. So the reacton functon of borrower j j m ( ) for 0, ) m H r m m = 1 for mj ( mj,1 j j j In ths case m = m = 1 s also a Nash equlbrum. We can call ths the effcent ** ** j (montorng/lendng) equlbrum. Fgure 1 presents the reacton functons for H r = It s mportant to note that the reacton functons are upward slopng. We wll return to ths ssue n our estmaton results, below. 9

11 The bank s payoffs n these two montorng game equlbra determne whether t wll lend. For the neffcent (0,0) case, the expected payoff to the bank s 2< 0 and group lendng s not feasble. The payoff to both borrowers n ths case s 0. On the other hand, for the effcent m ** ** 1 m2 1 = = case, the payoff to the bank s 2r > 0 and the payoff to both borrowers s H r 1. Clearly 2 m ** ** 1 m2 1 = = s the payoff-domnant equlbrum. Although ths also makes t a focal pont equlbrum (Schellng, 1980, p. 291), prevous expermental evdence ndcates that ths s not a suffcent condton for behavoral equlbrum selecton (e.g., Van Huyck, Battalo and Bel, 1990). Group Lendng: Sequental An alternatve to smultaneous lendng s to lend sequentally to group members wth the order chosen randomly. Here ntally only one (randomly chosen) member of the group receves a loan. Dependng on whether ths loan s repad, the bank decdes whether or not to lend to the other member of the group. Ths ncorporates dynamc ncentves, whch have become ncreasngly popular among researchers and practtoners n mcrofnance. The sequence of events s as follows: Stage 1: Bank chooses whether or not to lend $1 to one of the members of the group. It puts the other dollar nto alternatve use, whch yelds r. Stage 2: The borrowers smultaneously choose ther levels of montorng m. Stage 3: One of the borrowers s chosen at random (wth probablty 0.5) to receve the loan, f the bank lends. Ths borrower B decdes whether to nvest n R or S. If B nvests n project R, then he gets b and nether B j nor the bank receves anythng. The game stops here. If B nvests n project S the game contnues to round 2. In ths case borrower B receves H r and the bank receves r. Ths amount H r s 10

12 nvested and gves ( H r) r < 1 Of course f, so that self fnancng among the borrowers s not possble. B j s successful n her montorng, then B has to nvest n project S. Stage 4: The game moves to round 2 only f B (the randomly chosen frst borrower) nvests n project S n round 1. The bank lends $1 to (of course f B was successful n her montorng, then B j who nvests n ether project R or project S B j has to nvest n project S). If B nvests n project R then the bank gets ( H ) j r r and B j gets b. If B j nvests n project S, then the bank gets r. Total surplus s ( H r)( 1 r) + and ths s allocated among the two borrowers. B j gets ( H r)( 1 r) α + and B gets ( 1 α )( H r)( 1 r) +. There s now postve level of montorng rrespectve of α and group lendng (wth jont lablty) s a feasble and unque equlbrum outcome. Even though both borrowers stll have an ncentve to choose the non-verfable project R, the sequental lendng ncreases ther ncentve to montor. In ths case the reacton functons for the two borrowers are symmetrc and are gven by Solvng out we get The expected payoff to the bank s { α ( )( 1 ) } α ( )( 1 ) 1 mk = b+ ml H r + r b 2r 1 ml = b+ mk{ H r + r b} 2r b mk = ml m b+ 2r α H r 1+ r = = ( )( ) mmr k l mk + ml 1 r r 2 r Therefore a unque and postve level of montorng exsts, rrespectve of the value of α. Ths postve level of montorng occurs because even f borrower B j does not montor, 11

13 B has an ncentve to montor. To see ths, suppose that B j receves the loan n round 1 (remember that the order of recevng the loan s determned randomly). If B does not montor, B j wll nvest n project R and then B wll receve a payoff of 0. By choosng a postve level of montorng, B can ncrease the probablty that B j nvests n project S n whch case the game contnues onto the second round and B gets the loan. Moreover gven that B s gong to montor, B j has an even greater ncentve to montor due to the strategc complementarty of montorng. So the sequental nature of the lendng scheme and the smultaneous choce of the level of montorng (before a borrower knows whether he s the frst or the second borrower) leads to an effcent (montorng/lendng) equlbrum, as long as the equlbrum montorng levels are suffcent to provde postve returns to the lender. 3. Expermental Desgn We conducted 8 sessons n each of the three treatments wth 12 subjects n each sesson. The 288 subjects were graduate and undergraduate students at Monash Unversty and Unversty of Melbourne, Australa and Jadavpur Unversty, Kolkata, Inda. We conducted sessons n two countres to examne whether subjects n Inda, who are perhaps more exposed to ssues relatng to mcrofnance and who share more cultural smlartes to targeted borrowers, would exhbt behavoral dfferences from the subjects n Australa. 12 All subjects were nexperenced n that they had not partcpated n a smlar experment. Compared to the Australan sample, the Indan sample had a lower proporton of females, a greater proporton of Busness/Economcs/Commerce majors, and a hgher proporton of subjects that lved n a major metropols when they were aged 15. The z-tree software (see Fschbacher, 2007) was used to conduct the experment. Each sesson lasted approxmately 2 hours, ncludng 12 Followng Muhammad Yunus beng awarded the Nobel Peace Prze n 2006, mcrofnance and Grameen Bank have receved consderable meda attenton n Inda and n partcular n Kolkata, whch has cultural and lngustc smlartes to Bangladesh. 12

14 nstructon tme. Subjects earned AUD or ts purchasng power equvalent on average. 13 The nstructons (ncluded for the smultaneous lendng treatment 2 n the appendx) used the borrowng and lendng termnology employed n ths descrpton. We employed three treatments to examne the equlbrum predctons descrbed n Secton 2. In treatment 1 the 12 subjects were randomly dvded nto groups of two wth each group consstng of one borrower and one lender. In treatments 2 and 3 the 12 subjects were randomly dvded nto groups of three wth each group consstng of two borrowers and one lender. The role of each subject (as a borrower or as a lender) was determned randomly and remaned the same throughout each sesson, whch ran for 40 perods. At the end of every perod partcpants were randomly re-matched. After readng the nstructons and before the actual sesson began, the partcpants answered a set of questons relatng to the nstructons and they were pad n cash (at the end of the experment n addton to ther earnngs from the actual experment) $0.50 or Rs 5.00 for each correct answer. The two projects avalable to borrowers, S and R, each cost $1, to be fnanced by a loan from the lender. In the ndvdual lendng treatment, the lender chose whether or not to nvest $1 nto ths loan. In the group lendng (smultaneous and sequental) treatments, the lender chose whether or not to nvest $2 nto the loan ($1 to each borrower). The lender could choose to make the loan to both borrowers or to nether. She could not make a loan to only one borrower n the group. If the lender chose not to make the loan, she earned $1.50 (or $0.75 n the ndvdual lendng treatment) for the perod. In the group lendng treatments, f the borrower receved the loan, he could montor the project choce of the other borrower n the group by choosng to pay a montorng cost (C). Both borrowers could montor each other. If borrower X ncurred a cost C on montorng, there was a chance of M that the other borrower Y would automatcally be requred to choose project S. Otherwse the other 13 At the tme of the experment, 4 Australan dollars were worth about 3 U.S. dollars. 13

15 borrower could choose ether project R or project S. In the sequental lendng treatment, the borrowers were randomly determned to be the frst or the second borrower n the group to receve the loan. In ths case f the frst (randomly chosen) borrower s actual project choce was R, then the lender s second dollar was automatcally allocated to her savngs account where she earned $0.75 for ths dollar. All montorng decsons were made smultaneously. We used the strategy method to elct decsons from the borrowers. 14 The use of ths method mples that the borrowers and lenders made decsons smultaneously and borrowers made ther decson before they knew whether or not they had receved the loan. In the case of sequental lendng, the borrowers made montorng decsons before they knew whether they were the frst or the second borrower n ther group to receve a loan. They dd, however, know whether they were the frst or the second borrower to receve the loan at the tme of makng ther project choce. Panel A of Table 1 presents the relatonshp between C and M for treatments 2 and 3. Table 2 presents the earnngs of the borrowers and the lender under the dfferent project choce scenaros. Note that the even dvson for two S projects n the sequental lendng (Panel B) ndcates our choce of α = 0.5. In the ndvdual lendng treatment, f the lender decded to nvest $1 n a perod (make the loan), she could montor the project choce of the borrower by choosng to pay a montorng cost (C). Panel B of Table 1 presents the relatonshp between C and M n ths case, whch s based on λ = 4.5. Lenders pad ther selected montorng costs whenever they made the loan, regardless of whether or not the montorng was successful. If unsuccessful, the borrower could choose ether project S or project R. All decsons were revealed to all members of the two- or three- person group at the end of each perod. 4. Hypotheses to be Tested 14 The strategy method smultaneously asks all players for strateges (decsons at every nformaton set) rather than observng each player s choces only at those nformaton sets that arse n the course of a play of a game. Ths allows us to observe subjects entre strateges, rather than just the moves that occur n the game, and could help gve nsghts nto ther motvaton. 14

16 The experments were desgned to test the followng theoretcal hypotheses: Hypothess 1: Compared to the ndvdual lendng (treatment 1), the lendng rate, the average level of montorng and the average repayment rate are all hgher n the group lendng treatments (treatments 2 and 3). Hypothess 2: Compared to the sequental group lendng treatment (treatment 3), the lendng rate, the average level of montorng and the average repayment rate are not hgher n the smultaneous group lendng treatment (treatment 2). Note that the weak nequaltes ndcated n Hypothess 2 follow from the theoretcal predctons that the effcent (lend/montor) equlbrum s unque n the sequental lendng treatment, but both effcent and neffcent (no loan) equlbra exst n the smultaneous lendng case. 5. Results We present our results n the next four subsectons, wth each subsecton dealng wth a specfc aspect of the program: lendng, montorng, repayment (and project choce) and fnally effcency. In each case we present conservatve non-parametrc tests for treatment dfferences whch requre mnmal statstcal assumptons and are based on only one ndependent summary statstc value per sesson. We also report estmates from multvarate parametrc regresson models whch can solate the contrbuton of dfferent factors on lender and borrower behavor. 5.1: Lendng Fgure 2 presents the average proporton of lenders makng loans n the dfferent perods, by treatment. Clearly the average proporton of lenders makng loans s substantally lower at every perod for treatment 1 (ndvdual lendng) but there s very lttle dfference between treatments 2 and 3 (group lendng), provdng ntal support for both hypothess 1 and 2. These results are supported by non-parametrc Mann-Whtney rank sum tests wth the sesson 15

17 average as the unt of observaton (see Table 3, Panel A). Compared to the ntal 5 perods, the average proporton makng loans s also sgnfcantly lower n the later 5 perods n treatment 1 (Wlcoxon sgned rank test S = 0 for n = 8; p value < 0.01), but s not sgnfcantly dfferent between early and late perods n treatment 2 ( S = 12.5 for n = 8; not sgnfcant) and treatment 3 ( S 10.5 for n 8; not sgnfcant) n n n = =. Subjects partcpated n the experment over 40 perods, allowng us to examne ther behavor over tme more systematcally usng panel regressons. Table 4 presents the random effect probt estmaton of the lender makng the loan. These panel regressons ncorporate a random effects error structure, wth the subject (lender) representng the random effect. The dependent varable s the propensty to lend. The explanatory varables that we nclude are: dummes for smultaneous and sequental lendng treatment (the reference category s ndvdual lendng); a dummy varable ndcatng whether the lender made a loss n the prevous perod; a dummy for the sesson beng run at Jadavpur Unversty n Kolkata (to capture subject dfferences across the two countres); and two varables that capture the effect of tme on the propensty to make the loan n a partcular perod: 1 t and the nteracton term ( 1 ) ( GROUP) t, where GROUP s a dummy varable that takes the value of 1 f group lendng treatment (smultaneous and sequental lendng) and 0 f ndvdual lendng treatment. 15 We also nclude a set of demographc controls. 15 Notce from Fgure 2 that the tme trend appears qute smlar for the two group lendng treatments but s very dfferent for the ndvdual lendng treatment. The non-nteracted term ( 1 t ) n ths case captures the effect of tme on the propensty of the lender to make a loan n the ndvdual lendng treatment whle the nteracton term captures the dfferental effect of tme on the propensty of the lender to make a loan n the group lendng treatment. To get the total effect of tme n the group lendng treatments we need to add the coeffcent estmates of ( 1 t ) and ( 1 ( ) ( GROUP) ). t 16

18 Table 4 presents the coeffcent estmates and standard errors. Lendng decreased over tme n treatment 1, but ncreased over tme n the two group lendng treatments. 16 The two treatment dummes are both postve and statstcally sgnfcant ndcatng that the probablty of lendng s sgnfcantly hgher n group lendng compared to ndvdual lendng. Ths provdes further support for Hypothess 1. The dfference n the probablty of lendng n the two group lendng treatments s however statstcally sgnfcant at the 10% level (usng the test of equalty across the two group lendng treatment dummes), weakly contradctng Hypothess 2. The probablty of lendng n perod t s sgnfcantly lower f the lender receved negatve earnngs n perod t 1. The Jadavpur Unversty dummy s not statstcally sgnfcant mplyng that that there s no dfference n the probablty of lendng across the two samples. Most of the demographc control varables are not statstcally sgnfcant. 5.2: Montorng Fgure 3 presents the average level of montorng across perods. Wth the excepton of the frst few perods, average montorng s hghest n the sequental lendng treatment and lowest n the ndvdual lendng treatment. Agan usng a rank sum test wth the sesson average as the unt of observaton, the dfference n montorng rate between treatments 2 and 3 s not statstcally sgnfcant (Table 3, Panel B). The average montorng rate s always sgnfcantly lower for treatment 1, however. Ths provdes support for Hypotheses 1 and 2. In treatment 1 (ndvdual lendng) the montorng rate s lower n the last 5 perods than n the frst 5 perods (Wlcoxon sgned rank test S = 2 for n = 8; p value < 0.05), ndcatng a sgnfcant declne n montorng rates. By contrast, montorng rates ncrease between the frst 5 and last 5 perods for treatment 2 (Wlcoxon S = 4 for n = 8; p value < 0.05) and treatment 3 (Wlcoxon S = 0 for n = 8; p value < 0.01). n n n ( ) 16 The coeffcent estmates of ( 1 t ) and ( 1 ) ( GROUP) t are jontly sgnfcant. 17

19 The montorng decson s made by dfferent agents n the ndvdual and group lendng treatments. Hence we analyze the level of montorng chosen n the ndvdual and group lendng treatments separately. 17 The level of montorng chosen s restrcted n the range ( 0,1 ) and s estmated usng a tobt model. Consder frst the level of montorng chosen (by the lender) n the ndvdual lendng treatment. The explanatory varables are the same as those n Table 4, except that here, by defnton, we do not nclude the nteracton term ( 1 ) ( GROUP) t. Table 5, Panel A, presents the tobt regresson results wth player fxed effects and the Hausman-Taylor estmates for error component models. 18 The level of montorng n perod t 1 has a postve and statstcally sgnfcant mpact on the level of montorng n perod t. The Jadavpur Unversty dummy s not statstcally sgnfcant, ndcatng no locatonal dfferences. As mentoned above n the case of group lendng (wth peer montorng) the payoff for an ndvdual borrower depends both on her level of montorng and also on the level of montorng of her partner. Subjects could construct expectatons for the level of montorng of the other member of the group n dfferent ways. Here we consder the followng two smple alternatves: (1) Cournot expectatons: each subject expects the montorng level of the other member of the group to be the same as that n the prevous perod (Lagged Other Montorng); (2) Fcttous play: each subject expects the montorng level of the other member of the group to be the average over all the prevous perods (Lagged Average Other 17 The propensty to make the loan s sgnfcantly lower n the ndvdual lendng treatment, mplyng that the data on the level of montorng s often not observed n the ndvdual lendng treatment. The panel n ths case s therefore unbalanced: the observed number of montorng choces vares from 2 (n only 2 of the possble 40 cases, dd they choose to make the loan) to The tobt regresson results presented n column (1) fal to account for the possblty that the lagged dependent varable (lagged level of montorng) can be correlated wth the tme nvarant component of the error term (the unobserved ndvdual level random effect). Ignorng ths could result n based estmates. One way of obtanng unbased estmates would be to use nstrumental varables estmaton (see Hausman and Taylor, 1981). It s assumed that none of the covarates s correlated wth the dosyncratc error term. The results for the Hausman-Taylor estmaton for error component models are presented n Table 5, Panel A, column (2). Qualtatvely the results are very smlar to the tobt regresson results presented n column (1): n partcular, the greater the level of montorng n perod t 1, the greater the level of montorng n perod t and the level of montorng falls over tme. 18

20 Montorng). Hence each subject s assumed to have a long memory as opposed to the Cournot expectatons case where each subject has a short memory. Table 5, Panel B, presents the random effects tobt and the Hausman-Taylor estmaton for error component models for both specfcatons of expectaton formaton n the group lendng treatment. We fnd that montorng ncreased over tme. The postve and sgnfcant coeffcent estmate of the other borrower s lagged montorng level (n the Cournot expectatons verson) or ts counterpart lagged average other borrower s montorng (n the fcttous play verson) s consstent wth the upwardly-sloped reacton functons of the theoretcal model. Note that the coeffcent estmate on a borrower s own montorng n the prevous perod s also postve, and s substantally larger than the reacton to the other borrower s montorng level. The sequental lendng treatment dummy s never statstcally sgnfcant, consstent wth Hypothess 2. The Jadavpur Unversty dummy s always negatve but s statstcally sgnfcant only n the tobt regressons mplyng that subjects n Inda choose a lower level of peer montorng. 19 Fnally turnng to the demographc controls, we fnd that females choose a sgnfcantly lower level of montorng, as do subjects wth no prevous partcpaton experence. 5.3: Repayment Rate Fgure 4 presents the average repayment rates n the three treatments by perod. Note that ths repayment rate s not a choce varable but s the result of a combnaton of the ex ante project choce by the borrower, the level of montorng chosen by the borrower, and the success of the montorng process. Repayment occurs f the borrower chooses project S or f the borrower chooses project R and montorng s successful. It s clear from Fgure 4 that average repayment rates are lowest n the ndvdual lendng treatment. The results are agan supported by a rank sum test wth the sesson average as the unt of observaton (Table 3, 19 It s nterestng to note that f we restrct the sample to those born n South Asa (whether resdng n Australa or Inda), the Jadavpur Unversty dummy s no longer statstcally sgnfcant. 19

21 Panel C). The hgh repayment rates n the group lendng treatments are essentally drven by the hgh montorng rates and not by the borrowers ex ante choosng project S. Panel D of Table 3 presents the average proporton of borrowers choosng project R ex ante. The average proporton of subjects (ex ante) choosng project R s sgnfcantly lower n the ndvdual lendng treatment. Note also that the proporton of borrowers choosng project R s not dfferent across treatments n the ntal perods but becomes sgnfcantly dfferent across all treatments by the later perods. Recall that the earnngs of the borrower are greater f he chooses project R, but the earnngs of the lender are lower f the borrowers choose project R. In treatments 2 and 3, the borrowers are more lkely to choose project R, but they are also more lkely to choose a hgh level of montorng to be able to swtch the other borrower s project choce to S. In consequence the actual project choces are lkely to be project S and the earnngs of the lenders are postve and we move to an effcent (montorng/lendng) equlbrum. On the other hand n treatment 1 montorng rates are lower (montorng s also more costly ( λ > 1) ) and even though borrowers are more lkely to choose project S, lenders choose not to make the loan. We end up at the neffcent (low montorng/no lendng) equlbrum. The results therefore mply that group lendng wth peer montorng works better compared to ndvdual lendng wth lender montorng. Table 6 presents random effect probt regresson results for repayment and choce of project R. The explanatory varables are the same as n Table 4. The repayment rates (Table 6, column 1) are sgnfcantly hgher n the group lendng treatments compared to the ndvdual lendng treatment, consstent wth Hypothess 1. The sequental lendng treatment has sgnfcantly hgher repayment rates compared to the smultaneous lendng treatment (the test of equalty of treatment effects s sgnfcant, ndcatng that the two group lendng treatments do not have smlar effects on repayment), consstent wth Hypothess 2. The 20

22 Jadavpur Unversty dummy s not statstcally sgnfcant and fnally repayment rates are sgnfcantly lower for subjects who had a hgher proporton of correct answers n the quz. Column 2 presents the random effects probt regresson results for choce of project R. The probablty of choosng project R decreased over tme n treatment 1, whle the probablty of choosng project R ncreased over tme n the two group lendng treatments. 20 The two treatment dummes are both postve and statstcally sgnfcant, ndcatng that the probablty of choosng project R s sgnfcantly hgher n group lendng compared to ndvdual lendng. The probablty of choosng project R s no dfferent n the two group lendng treatments. In all three treatments, subjects who seek to maxmze ther own current perod monetary earnngs should always choose project R. We observe, however, that borrowers have a lower probablty of choosng R n the ndvdual lendng treatment. One possble explanaton s that recprocal motvatons are trggered more n a two person game than a three person game. 5.4: Effcency Fnally we turn to effcency, defned as the total surplus attaned by the borrowers and the lenders as a proporton of the maxmum possble surplus. Here effcency s used as a measure of the overall performance of the market. Maxmum surplus s attaned when the lender provdes the loan; the level of montorng chosen s 0 and the borrowers choose project S. Panel E of Table 3 ndcates that average effcency s lowest n the ndvdual lendng treatment. The hgher montorng rates n the sequental lendng treatment could lead to lower effcency than the smultaneous group lendng treatment, but the dfferences are only sgnfcant n the early perods. 6. Concludng Comments ( ) 20 Agan the coeffcent estmates of ( 1 t ) and ( 1 ) ( GROUP) t are jontly sgnfcant. 21

23 Ths paper reports the results from a laboratory experment of group lendng n the presence of moral hazard and costly lender or peer montorng. We compare treatments when credt s provded to members of the same group sequentally and smultaneously, and when loans are gven to ndvduals and montored by lenders. Our results suggest that n the presence of moral hazard, peer montorng results n hgher loan frequences, hgher montorng and hgher repayment rates compared to bank montorng. The dfferences are mnor between smultaneous and sequental lendng. Results from ths experment can help shed lght on the mportant polcy ssue of the optmal desgn for mcrocredt programs. Much of the success of mcrocredt programs has been attrbuted to self-selected groups and socal tes n rural communtes. However successful applcaton of these programs n other scenaros and economes requres more than strong socal tes. In urban contexts of developng and transtonal economes, for example, t mght be more dffcult to form self-selected borrowng groups compared to the more closely knt rural communtes. For ths reason several authors and polcy makers suggest that optmal desgn of mcrocredt programs look beyond the ssue of self-selecton and even look beyond group lendng. In ths experment, we focus on nformatonal asymmetres due to moral hazard and we restrct ourselves to exogenously formed groups. Our results show that n the presence of moral hazard group lendng performs better compared to ndvdual lendng, even wth no self-selecton n group formaton. Introducng dynamc ncentves (wthn group lendng) helps, but not sgnfcantly. What s mportant s peer montorng, whch works much better than actve lender montorng. 21 Optmal desgn of mcrocredt programs needs to take advantage of the fact that t s less costly for group members to montor each other, whch can result n better project choces and hgher repayment rates. 21 It has been observed that n the absence of peer montorng the success of such programs s qute lmted. See Bhatt and Tang (2002) for evdence usng data from mcrocredt programs n the US.. 22

24 References: Abbnk, K., B. Irlenbusch and E. Renner (2006): Group Sze and Socal Tes n Mcrofnance Insttutons, Economc Inqury, 44(4), Anket, K. (2006): Sequental Group Lendng wth Moral Hazard, Mmeo London School of Economcs. Armendarz de Aghon, B. and C. Goller (2000): Peer Group Formaton n an Adverse Selecton Model, Economc Journal, 110, Armendarz de Aghon, B. and J. Morduch (2000): Mcrofnance beyond Group Lendng Economcs of Transton, 8, Armendarz de Aghon, B. and J. Morduch (2005): The Economcs of Mcrofnance, MIT Press. Banerjee, A. V., T. Besley and T. W. Gunnane (1994): Thy Neghbor s Keeper: The Desgn of a Credt Cooperatve wth Theory and a Test, Quarterly Journal of Economcs, 109, Barr, A (2001): Socal Dlemmas and Shame-Based Sanctons: Expermental Results from Rural Zmbabwe, Mmeo, Centre for the Study of Afrcan Economes, Unversty of Oxford. Bellman, E. (2006): Invsble Hand: Entrepreneur Gets Bg Banks to Back Very Small Loans; Mcrolendng-for-Proft Effort In Inda Draws Busness From Ctgroup, HSBC; Ms. Dobbala's Baby Buffalo, Wall Street Journal, Eastern Edton, 15/5/2006, A1. Bhatt, N. and S-Y. Tang (2002): Determnants of Repayment n Mcrocredt: Evdence from Programs n the Unted States, Internatonal Journal of Urban and Regonal Research, 26(2), Bolnk, B. R. (1988): Evaluatng Loan Collecton Performance: An Indonesan Example, World Development, 16, Carpenter, J., S. Bowles and H. Gnts (2006): Mutual Montorng n Teams: Theory and Expermental Evdence on the Importance of Recprocty, IZA Dscusson Paper # Cassar, A., L. Crowley and B. Wydck (2007): The Effect of Socal Captal on Group Loan Repayment: Evdence from Feld Experments, Economc Journal, forthcomng. Che Y-K. (2002): Jont Lablty and Peer Montorng under Group Lendng, Contrbutons to Theoretcal Economcs, 2(1). Chowdhury, P. R. (2005): Group-lendng: Sequental Fnancng, Lender Montorng and Jont Lablty, Journal of Development Economcs, 77, Conyabuguma, M., T. Page and L. Putterman (2005): Cooperaton under the Threat of Expulson n a Publc Goods Experment, Journal of Publc Economcs, 89(8), Conln, M. (1999): Peer Group Mcro-Lendng Programs n Canada and the Unted States, Journal of Development Economcs, 60, Fehr, E. and S. Gaechter (2000): Cooperaton and Punshment n Publc Goods Experments, Amercan Economc Revew, 90(4), Fschbacher, U. (2007): z-tree-zurch Toolbox for Readymade Economc Experments, Expermental Economcs, forthcomng. Fry, T. R. L., S. Mhajlo, R. Russell and R. Brooks (2006): The Factors Influencng Savng n a Matched Savngs Program: The Case of the Australan Saver Plus Program, Mmeo, RMIT Unversty, Melbourne. Ghatak, M. and T. W. Gunnane (1999): The Economcs of Lendng wth Jont Lablty, Journal of Development Economcs, 60, Ghatak, M. (2000): Screenng by the Company You Keep: Jont Lablty Lendng and the Peer Selecton Effect, Economc Journal, 110,