Objective repayment performance in Bangladesh and its determinants: How to improve the allocation of loans by MFIs?

Transcription

1 Objectve repayment performance n Bangladesh and ts determnants: How to mprove the allocaton of loans by MFIs? Mare Godqun # Summary: The am of ths artcle s to produce a comprehensve analyss of the performance of mcrofnance nsttutons (MFIs) n terms of repayment. We use 1629 loan observatons to analyze wth a probt the determnants of the repayment performance of borrowers of the BRAC, the BRDB and the Grameen Bank. We test for endogenety of the sze and duraton of the loan n the determnaton of repayment and use nstrumental varables to correct for t. We then use a comparatve analyss of the determnants of the repayment performance and of the loan sze to gve polcy recommendatons to mprove the allocaton of loan by MFIs. Keywords: Mcrofnance, socal tes, group homogenety, socal ntermedaton, Asa, Bangladesh. I would lke to thank Sylve Lambert and Jean-Claude Berthélemy for ther helpful comments, suggestons and encouragement. I am grateful to Pr. J. Morduch for provdng me the data. # TEAM Unversté Pars I Panthéon-Sorbonne mare.godqun@unv-pars1.fr 1

2 Introducton. Mcrofnance Insttutons (MFIs) provde fnancal servces (credt and savng) to the poor so as to reduce the credt ratonng they face and help allevate poverty. Should t be proft orented or not, each MFI search the repayment performance as hgh repayment rates allow the MFI to lower the nterest rate t charges to the borrower reducng thus the fnancal cost of credt and enabelng more borrowers to have access to credt. Improvng the repayment rate could also help reduce the dependence on subsdes and help the MFI reach a better sustanablty level. It s also argued that hgh repayment rates reflect the adequacy of MFI's servces to clents needs and restrct the cross subventon 1 of the borrowers. Repayment performance also acts as an mportant postve sgnal when the MFI has to rse new funds. For all these reasons, hgher repayment rates are largely assocated wth benefts both for the MFI and the borrower 2. The MFI wll thus frstly tend to reach the frst best level of a 100% on tme repayment rate and, f such a level of repayment performance cannot be reached, t wll try to allocate hgher loans to borrowers wth lower probablty of default and reduce the delay n repayment. Understandng what the MFI should do to meet these objectves depends on the most common factors nfluencng repayment. Consderng those factors, we can dfferentate between those related to nformaton asymetres, those related to adverse shocks and those related to a low performance of nsttutons such as justce or educaton. When ganng nformaton on the characterstcs or on the behavor of the borrower s costly for the MFI, problems of adverse selecton allocaton of loans to borrowers who dd not have the ablty to take advantage of hs loan or wth a low probablty- and problems of moral hazard the borrower has not produced the requred level of effort to take advantage of hs loan or has used s loan for unproductve purposes are more lkely to occure. Adverse selecton and moral hasard wll ncrease the proporton of borrowers who do can not repay ther loans on tme next to those who experenced dfferent adverse shocks (lke llness or natural dsasters). Borrowers mght also have enough monney to remburse ther loan but stll default strategcally. Cost of strategc default mght ndeed be low f they have low collateral requrement and f the legal system gve lttle power to the MFI to enforce contracts. MFIs wll try to restrct the 1 As borrowers have dfferent probablty of default and as t s dffcult for the MFI to charge a dfferent nterest rate to each borrower relatve to hs probablty of default, borrowers who are more prone to default wll be subsdzed by lower rsk borrowers. 2 If a good repayment performance s a prerequste for fnancal sustanablty, they are not a suffcent condton of fnancal health as hgh admnstratve costs or hgh borrower turnover could be the counterpart of those hgh repayment rates. 2

3 occurrence of those stuatons and desgn approprate ncentves to repay and sometme strengthen the economc ablty of the borrower. In ths paper, we want to contrbute to the mprovement of the repayment performance of MFIs n examnng the determnants of the repayment performance wth a partcular nterest n mcrofnance nnovatons. Ths analyss wll hghlght n whch drecton the MFI should work so as to ncrease the repayment rate and as a consequence approach the frst best. Ths paper also ncludes an analyss of the current adequacy of mcrocredt supply (n terms of loan sze) based on the comparson of the determnants of the repayment performance to these of the loan sze. As the on tme repayment rate s often nferor to 100%, the MFI wll have second best strateges to ncrease the repayment. Ths work uses an objectve repayment varable,.e. a repayment varable based on the declaraton of the borrower (not the MFI's). It addresses the problem of endogenety of prncpal and duraton of the loan n the determnaton of the repayment performance. The results ndcate that mcrofnance nnovatons have a mxed explanatory power among the other nfluencng factors of the rembursement. The layout of ths artcle s as follows: after a bref presentaton of the conceptual framework (secton 1), secton 2 nvolve a bref emprcal ltterature revew and secton 3 lays down the context of the case study. Secton 4 presents the econometrc methodology. The results of the regresson model are dscussed n secton 5 and the artcle concludes wth mplcatons for polcy recommendatons and future research. 1. The conceptual framework How to get closer to the frst best level of repayment performance? Credt ratonnng and collateral requrement are the tradtonal means used to cope wth asymetres n nformaton on the credt market (Stgltz & Wess, 1981) but both of those means lead to the excluson of the poor borrowers. To explan the success of mcrofnance n provdng credt to the poor, a large lterature uses the prncpal/agent theory to demonstrate that mcrofnance contracts whch are lendng o jont-lable groups allow the lender to bypass moral hazard (Stgltz, 1990) and adverse selecton (Gatack, 1999) due to asymmetres n nformaton. It s also argued (Besley & Coates, 1995) that jont lable lendng groups help enforce repayment as socal nteractons make t more costly to default. Regular repayment 3

4 schedules (Armendarz de Aghon & Morduch, 2000) or dynamc ncentves 3 (Besley, 1995) are other approprate ncentve mechansms used by MFIs to ncrease ther repayment performance. Socal ntermedaton, n other words, the provson of non-fnancal servces nest to credt and savng servces (Edgcomb & Barton, 1998) also develops the economc ablty of the borrower to repay. The prevous mechansms are consdered to be fnancal nnovatons (Edgcomb & Barton, 1998) that make t fnancally sustanable for mcrofnance to lend to the poor. When the use of such mechansms s unsuffcent for the MFI to reach a frst best repayment rate and when borrowers are heterogeneous n ther probablty of default, the MFI should also allocate loans of dfferent sze to the borrowers so as to maxmze the proporton of outstandng debts repad on tme The second best perspectve: ncreasng the proporton of outstandng debts repad on tme. The purpose of ths secton s to understand why borrowers prefer bgger loans and why the MFI should allocate bgger loans to borrower wth lower probablty of default The context. We consder the credt relatonshp of mcrofnance nsttutons provdng credt to jont lable credt group at a unform nterest rate. The MFI faces demand for credt of borrowers heterogenous by ther localsaton, lendng group, ablty and preferences. The MFI maxmzes the global net expected return of ts borrower under a zero proft condton. As borrowers face hgh credt ratonnng, there s a large set of hghly productve projects and borrowers face an ncreasng medum productvty of captal. The expected proft for the borrower wll thus be strctly ncreasng n the loan sze for a gven duraton as shown n fgure 1 4. The borrower acqures the nformaton on the cash requrement, expected return and probablty of success of the dfferent projects he s able to manage before applyng for a loan. Ths set of projects s possbly restrcted by the local envronment as the dstance from marketable actvtes, the economc drve and the pre-exstng concurrence of smlar project n the area may nduce varaton n the set of proftable projects. The set of accessble projects 3 Dynamc ncentves refer to the threat not to refnance a borrower who defaults on debt oblgaton. The ncentve power of dynamc ncentve s enhance f the MFI allocate larger loans over tme to good repayment borrowers. 4 Fgure 1 stands for a constant margnal productvty of captal. We could also allow for an ncreasng margnal productvty of captal whch would strengten the atractveness of bgger loans for the borrowers. 4

5 may be further restrcted by hs lendng group (Madajewcz, 1999) whch may ncte the borrower to undergo smlar project n terms of sze, actvty and probablty of success so as to better montor hs actvtes and lmtate the varaton n default probablty among group members and consequent group cross subsdaton. The borrowers have no own funds to nvest The behavour of the borrower: demand for credt and repayment. The demand for credt. We consder a borrower who s gven the chance to obtan a loan from a mcrofnance nsttuton. The loan applcaton of the borrower wll correspond to the sze and duraton whch maxmzes hs expected return 5. The optmal duraton depends on the dstrbuton wth tme of the returns of the project, on the scale of the project, on the preferences of the borrower for the present consumpton and on the allocaton of a new loan condtonal to the total repayment of the prevous loan. The duraton wll however be such that the sum of the return of the project and external returns s hgher than each nstallement. Ths mnmal duraton of the loan ncreases when the returns to scale of the project are ncreasng and remans constant when they are constants. For a gven duraton, each loan sze wll correspond to a sngle project as the borrower wll undertake the project that has the hgher expected return for each loan sze. As the net return s an ncreasng functon of the sze of the loan, the borrower wll always prefer bgger loan and wll ask for the maxmum loan sze he can ask for (L max n fgure 1) gven hs set of accessble projects defned by hs characterstcs, the characterstcs of hs envronment and these of hs lendng group. The repayment. The borrower wll default on hs loan when the duraton of ths loan s nferor to the mnmum duraton. An ncrease n the duraton along wth unregular repayment schedules may also ncrease hs probablty of default. For a gven borrower and a duraton of the loan, t s argued (Fremer & Gordon, 1965) that the repayment probablty decreases wth the sze of the loan as shown n fgure 2 where P mn represent the probablty of default due to external factors such as llness or acdental destructon of the borrower s productve assets. The ncrease n the default probablty may be dfferent among borrowers reflectng the dfference n endowments and moral hasard or strategc default assocated costs. 5

6 The MFI: mprovng the proporton of outstangng debts repad on tme. If the above mentoned ncentve mechansms and non fnancal servces are mperfect, the MFI wll choose a new target probablty of default and wll respond to the borrower demand for credt only f hs default proballty s nferor to the target. If there s observable heterogenety n the repayment probablty of borrowers for a gven sze of the loan, the MFI then allocate larger loans to the safer borrowers as shown n fgure Lterature revew. Followng the dscusson on the theoretcal lterature on mcrocredt we expect that jont lablty, especally through peer selecton, peer montorng, and peer pressure should be assocated wth a better repayment performance. Group homogenety and socal tes are also expected to ncrease the repayment performance not per se but because they allow a better effcency of group dynamcs. Group homogenety as a result of effectve peer selecton (group homogenety n terms of rsks, Ghatak, 1999) and as a mean to ncrease peer montorng (group homogenety n terms of nterest, economc power..., Stgltz, 1990) should go together wth hgher repayment rate. Hgh level of socal tes should have the same mpact as they facltate peer montorng and ncrease the potental socal sancton of peer pressure (Besley & Coates, 1995). Dynamc ncentves and socal ntermedaton, whch are extra group mcrofnance fnancal nnovatons, are also expected to ncrease the repayment performance. Addressng the queston of the relatve performance of group loans compared to ndvdual loans and usng data from Zmbabwe, Bratton (1986) states that group loans perform better than ndvdual loans n years of good harvest and worse n drought years when peers are expected to default. Paxton (1996) analyzes wth a mean and covarance structural model the determnants of successful group loan repayment of 140 credt groups n Burkna Faso. She rases one s attenton on what she calls the domno effect 6 that can outwegh the postve effects of group lendng. Zeller (1998) uses nformaton on 146 credt groups n Madagascar and provdes evdence n favor of group lendng. Zeller shows ndeed that the group generates nsurance whch leads to a better repayment performance. Analyzng the potental postve effects assocated wth group dynamcs, some studes examne the mpact of dfferent levels of peer selecton, peer montorng and peer pressure. 5 We consder that the utlty of the borrower s ncreasng wth the return he gets from hs project. 6

7 Wenner (1995) presents a methodology to test whether selecton mechansm has an mpact on the repayment performance of 25 Costa Rcan credt groups and whether group members use local nformaton for the screenng of ther peers. Hs study shows that lendng groups use prvate nformaton to select ther peers and that ths selecton mechansm ncreases the group repayment performance 7. On the same pont, the above mentoned study of Zeller (1998) confrms the postve role of peer selecton (nternal rules of conducts) on repayment performance. Wydck (1999) uses data from 137 Guatemaltese credt groups to show how socal coheson affects group performance n terms of repayment rate, group nsurance and moral hazard. He found that peer montorng n urban groups and peer pressure n rural ones sgnfcantly affects group performance. Lmtng the conclusons on the mpact of group dynamcs of the prevous three artcles, Dagne, Chmombo, Smtowe & Mataya (2000), workng on data from Malaw, found that peer montorng, peer pressure and jont lablty had lttle or negatve mpact on repayment performance whereas peer selecton was found to be lmted. Socal tes and group homogenety are supposed to mprove the power of group dynamcs. Nevertheless, the studes gve mxed results. The study of Sharma & Zeller (1997), based on the analyss of repayment rates of 128 credt groups n Bangladesh, leads to a controversal negatve mpact of preexstng socal tes as well as group homogenety n terms of asset and enterprse dversty. The study of Zeller (1998) nvestgates the effects of ntragroup poolng of rsky assets or projects on repayment rates. Whle ths analyss supports the postve role of socal coheson, t also concludes that rsk dversfcaton (up to a certan level) has a sgnfcant postve effect on repayment performance. Ths could be explaned by a matchng problem (Paxton, 1996). The matchng problem arses when credt terms and condtons are no longer approprate for each member as credt s repeated agan. If ntal group homogenety and pror experence of group actvtes were assocated to better repayment performance, as tme goes the postve effect of group homogenety on peer montorng balance wth the negatve mpact of a matchng problem and the absence of rsk dversfcaton that lmt the possbltes of ntra-group nsurance. Parallel to group lendng, MFIs usually use dynamc ncentves and socal ntermedaton. MFIs use dynamc ncentves when they ncrease wth tme the amount they lend to a specfc borrower and condton new loans to pror rembursement dscplne. Mcrofnance s 6 The domno effect occurs when at least one member of a credt group default due to the default of other members. 7 Wenner challenges the postve effect of ths feature as further analyss ndcate that costs faced by the borrowers to get ths nformaton overcompensate the nduced benefts n terms of repayment performance. 7

8 sometmes referred to as socal ntermedaton (Edgcomb & Barton, 1998) as many MFI provde servces or tranng that go beyond fnancal servces. Contrary to group lendng, those two man features of MFI have been lttle documented up to now. The most mportant factor that motvates lendng groups to repay n the study of Dagne, Chmombo, Smtowe & Mataya (2000) s the relatve value they assocate to access to future credt. For Sharma & Zeller (1997), credt ratonng up to a certan level, has sgnfcant postve effect on repayment performance. Moreover, we cannot assume that these dynamc ncentves keep the same ntensty as credt s repeated over tme especally f borrowers observe that credt sn t systematcally dened to defaultng or late borrowers. In a study devoted to the Grameen Bank and not exclusvely focussed on a repayment rate analyss, Khandker, Kally & Khan (1994) found that the longer the branch operates n an area, the hgher the loan default rate. They explan ths feature by the possble decreasng margnal proftablty of new projects but as many MFI suffer today from a growth crss, we have to be cautous on ths pont. Khandker, Khally & Khan (1994) also found that membershp tranng, whch relates to socal ntermedaton, has a postve nfluence on repayment. The last set of varables to be documented n studes on the determnants of repayment rates are the characterstc of the area and of the borrower. Khandker, Khally & Khan (1994) rase the queston of whether default s random and nfluenced by erratc behavor or whether t s systematcally nfluenced by area characterstcs that determne local producton condtons or branch-level effcency. Ther emprcal test on Grameen overdue loans backs up the partal nfluence of area characterstcs. Rural electrfcaton, road wdth, prmary educatonal nfrastructure and commercal bank densty are postvely correlated wth a low default rate as well as the predcted manager's pay and member tranng. We can nfer from ths study that the economc drve of the area postvely nfluences the repayment rate. Paxton (1996) shows n the same way that access to other credt sources, market sellng actvtes and urban locaton were lnked to a better repayment performance. Questonng the mpact of the characterstcs of the borrower, Zeller (1998) showed that tradtonal prors aganst women, young borrowers or hgh famly sze could not be used as sgnals of repayment ablty. The above mentoned studes attrbute a debatable role to the explotaton of group dynamcs whereas the role of socal ntermedaton and socal tes s very lttle documented. It s thus mportant to produce further tests of the mpact of these fnancal nnovatons on the repayment performance so as to understand better the common reasons for the success of mcrofnance. Ths artcle provdes a test of the explanatory power of socal tes and group 8

9 homogenety as well as socal ntermedaton and dynamc ncentves. Impact of the man characterstcs of the loan contract and of the borrower s also taken nto account. 3. Data. The data come from a quas expermental survey ran n Bangladesh n by the BIDS (Bangladesh Insttute for Development Studes) and the World Bank. The survey covered 1798 households, comng from 87 vllages from 29 dfferent thanas (subdstrcts) of these households were elgble to MFI, whch means they were poor enough to beneft from mcrofnance servces and 905 of these households actually took part n a mcrofnance program. For the purpose of our work, we concentrate our nterest on households that had actually borrowed from one of the three MFIs. On the 2349 workable observatons, 485 corresponded to BRAC loans, 430 to BRDB loans, 1081 to Grameen Bank loans and 353 to other credt provders (see Table 1 for the descrpton of other credt provders). The exploraton of the data bears out lttle dfferences across the MFIs n terms of general characterstcs of the loan (the nterest rate s the same for all loans) or type of borrowers. Dfferences appear when we consder non fnancal servces: health, educaton, marketng and profesonnal tranng servces. These servces ncrease the value the borrower attrbutes to hs relaton wth the MFI and can be referred to as socal ntermedaton. The followng secton lays down the econometrc model that wll enable us to analyze the factors nfluencng the repayment performance. 4. The econometrc framework. Followng the dscusson on the theorcal ltterature, we descrbe the nteracton between the borrower and the MFI wth the followng model: Stage one: The borrower apples for a loan of a specfc sze and duraton whch correspond to the larger scale of hs accessble projects dfned by hs/her characterstcs, the characterstcs of hs envronment and these of hs/her lendng group. Stage two: Before allocatng a loan to the borrower, the credt offcer of the MFI computes the probablty of default of ths applcaton gven the nformaton he has on the borrower, on hs lendng group, on the envronnement and on the predcted effectveness of hs repayment ncentves for ths borrower. When the computed probablty s nferor to the acceptablty 8 These 29 thanas were randomly chosen out of the 391 thanas of Bangladesh. 9

10 treshold, he allocates the loan the borrower appled for, otherwse, he alloactes a smaller loan whch correspond to the acceptablty treshold. Stage three: The borrower remburses hs loan on tme or not gven hs/her envronment, hs/her characterstcs/ablty, the characterstcs of hs/her group and the characterstcs of the loan contract. The break down n three steps of the mcrocredt relatonshp sheds lght on the possblty of endogenety of the prncpal characterstcs of the mcrocredt contract (loan sze and duraton) n the estmaton of repayment. The determnaton of the loan sze and duraton n stage two and the determnaton of the repayment n stage three mght ndeed be based on shared omtted varables varables observed by the MFI ans the borrower but not avalable n our dataset (lke the characterstcs of the envronment). We constructed ndvdual dummy varables for repayment. We thus used a probt model to estmate the probablty for a borrower to repay hs loan at the due date 9. We used the method of Smth and Blundell (1986) to test for exogenety n such a model. Endogenety of the sze and the duraton of the loan couldn t be rejected and those varables are nstrumented. The prevous dscusson leads us to our estmaton strategy: Step one : Estmaton of the sze and duraton of the loan: P 5 5 p p p p p p = Pˆ + ε = α + β j X j + ρ j Yj + σ j Wj + γ IVp + j= 1 j= 1 j D 5 5 d d d d d d = D + ε = α + β j X j + ρ j Yj + σ j Wj + γ IVd + j= 1 j= 1 j ˆ ε (1) ε p d Where X j represents the varables of socal tes, group homogenety and sze of the group. Socal tes of the borrower wth the rest of the communty as a whole can ncrease the socal cost of peer pressure and the dentfcaton of the borrower to ther weekly meetng group gves argument n ths sense. We tred therefore not to restrct socal tes to socal tes among the group 10 as n prevous studes (Wydck 1999). We used the age of the group (.e. the number of years durng whch the borrower took part to hs present group) for nner group socal tes and proxes such as lvng n the same house as the spouse for extended socal tes. We postulate that the members of the group know each other better and develop socal tes as the age of the lendng group (AGEGP) ncreases. Ths s why we expect the ablty of the 9 Estmatons usng larger defntons of repayment (repayment done before three, sx or twelve months followng the maturty date) have also been conducted and gave sensbly the same results. For such estmatons, the sample had to be reduced to borrowers whose orgnal date of rembursement stands three to twelve months before the date of the survey. We chosed the exposed defnton of repayment so as to work wth the largest sample. 10

11 members of the group to montor and pressure each other to ncrease wth the age of the group. The varable AGEGP should thus have a postve mpact on the repayment performance. We used two varables for extra-group socal tes. The frst one, SCOHAB, takes the value of one f the borrower and hs/her spouse lve n the same house. We postulate that the socal net of the borrower s bgger when he does not lve n the same place as hs/her spouse does. That s why we expect SCOHAB to have a negatve mpact on repayment. The second varable, RESE, gves the number of months the borrower lved out of the vllage n the prevous year. The longer the borrower lves out of the vllage, the smaller hs socal tes wth other vllagers. RESE s then also expected to have a negatve mpact on repayment. Group homogenety s based on shared characterstcs (gender, age, educaton level) of the borrower and ts group leader. We expect the varables of group homogenety (SAMESEX, SAMEEDU, SAMEAGE) to have a postve mpact on repayment performance. Small groups are expected to show a better repayment performance as they can easly montor each other. Because borrowers of large groups can beneft from larger ntra-group nsurance possbltes, large group may also have a postve mpact on group performance. We use dfferent dummes (NMBG1, NMBG2, NMBG4, NMBG5) for the sze of the group to control for the possble non lnear effect of the sze of the group. Y. j descrbes the varables of socal ntermedaton (such as access to health servces) and dynamc ncentves (proxed by credt ratonng). The varables of socal ntermedaton (FACL, FACT, FACH, FACM) and the varable of dynamc ncentve (CRd) are expected to have a postve mpact on the repayment. W.j stands for the exogenous control varables. Control varables gather characterstcs of the borrower and hs household and basc nformaton on the loan (dummy for the MFI, sze and duraton of the loan, purpose of the loan, perod of redempton date n the year and number of loans attrbuted by the MFI durng the year of the loan). IVp and IVd represent the nstruments respectvely for prncpal and duraton. Because the determnaton of sze and duraton of the loan s smultaneous 11, we run smultaneous regressons for those two varables. Step two: Smth and Blundell (1986) exogenety test of prncpal and duraton n the determnaton of the repayment performance: 10 Socal tes nsde the group are expected to facltate the peer montorng and the use of peer presure. 11 The MFI uses the same varable to allocate both the prncpal and the duraton of the loan. Ths was confrmed by a Ramsey RESET test of omtted varable. 11

12 R ) ) α σ W + ηε + µε + ε 5 5 = + ωp + υd + β j X j + ρjyj + j= 1 j= 1 j j j p d (2) Where R, the latent varable of the model, s the capacty for an ndvdual to generate cash n excess of the amount (prncpal plus nterests) he has to repay before the ntal date of payng back. What we observe s the rembursement R * whch takes the value of 1 f R >0 and 0 f R <0. Exogenety s rejected f the coeffcents of the errors ( η, µ ) of the nstrumental regresson of prncpal and duraton are sgnfcant. Ths would ndeed mean that the structure of the error term s the followng: ε αε + βε = + µ p d The test of nstrument confrms that we used approprate nstrumental varables 12. Step three: Estmaton of the repayment performance: R 5 5 = + υd + ωpˆ + β j X j + ρ jyj + j= 1 j= 1 j α ˆ σ W + ε (3) j j Step four: Larger loans, for whom? After a comment on the regresson of the repayment probablty, we compare ts determnants to those of the sze of the loan. If we assume that the loan sze reflects the percepton the MFI has of the capactes of the borrower, ths allows us to analyze the adequacy of the loan allocaton by the MFI. Because, as prevously exposed, the loan sze s a result of both demand (stage one) and supply (stage two) factors, we cannot smply assume that the loan sze reflects the abltes the MFI attrbutes to a specfc borrower. We must ndeed consder separately the two followng cases: 1. If the demand of the borrower n terms of loan sze s hgher than the fnal loan sze, the loan sze reflects the percepton of the MFI. 2. If the demand s equal to the attrbuted loan sze, t could be that the MFI would have gven more f the borrower had asked for f the MFI has a hgher representaton of the abltes of the borrower. 12 We use ELECTRICITY,.e. prvate access to electrcty (whole sample, BRDB, GB, RTG3) and the PGWBL,.e. the number of weeks the borrower had to wat before recevng hs loan (for the BRAC), as nstrumental varable for loan sze. For the duraton of the loan, we use COLS,.e. sgnature or personal guaranty requred as prmary collateral (whole sample, BRAC, BRDB, RTG3) and PGWBL (GB) as nstrumental varables. 12

13 Step four wll thus be restrcted to the borrowers who were credt ratoned 13 (over 70% of the sample). Remarks: 1. We ran the regresson on both splt-sample (one regresson by MFI) and on the entre sample. Ths allows us to assess f the mpact of a specfc varable s drven by a specfc MFI or not. The estmaton on splt sample and the applcaton of the test of Smth and Blundell (1986) also proved the need of correcton for endogenety of prncpal and duraton n the estmaton of the repayment performance. We also present the regresson for the repayment probablty when a delay of three months (RTG3) s admtted for repayment. Correcton for endogenety of prncpal and duraton has also been made for ths regresson. 2. What we observe here are only margnal mpacts of mcrofnance novaton snce each of the studed MFIs uses group lendng, socal ntermedaton and dynamc ncentves n ther credt relatonshps. Ths nduces an underestmaton of the effect of these fnancal nnovatons. 5. Results and dscusson The repayment behavor of the borrower. The results of the probt estmaton of the repayment are reported n Table 6 (repayment on tme and wth a grace of three months, whole sample) and Table 7 (repayment on tme, splt sample). Socal tes among the group, proxed by the age of the group (AGEGP), had a sgnfcant negatve mpact on repayment rate. We explan ths contrastng feature n dfferent ways. At frst, we can refer to the matchng problem (Paxton, 1996): as duraton of membershp ncrease, the credt need of the members of the group dffers. A decreasng power of socal penaltes can also explan ths feature as members know each other better and are more reluctant to control and sancton themselves. Ths seems to overcompensate the benefts of ncreasng experence n consumpton smoothng and provson of nsurance wth membershp duraton. The age of the borrowng group was stll sgnfcant when we allowed for arrears of three months for the repayment (estmaton of RTG3). The varable SCOHAB 14 whch s a proxy for socal tes out of the group had a sgnfcant negatve effect as expected. If the spouse of the borrower lves elsewhere, the socal network 13 As we do not have nformaton on the loan sze demand, credt ratonment refer to the subjectve aprecaton of 13

14 of the borrower wll be larger. Group homogenety (SAMESEX, SAMEEDU, SAMEAGE) proved to have no sgnfcant mpact on repayment performance n the whole sample as n the study of Wydck (1999) 15. Group homogenety n terms of sex showed a postve mpact (regresson of RTG3 and BRDB) whereas homogenety n terms of educaton showed a negatve effect (same regressons). Homogenety n terms of age showed both a negatve (BRDB) and a postve (GB) effect on the repayment performance. Varables for socal ntermedaton proved to have lttle or negatve nfluence n the whole sample. Access to health servces (FACH) had sgnfcant negatve effect whereas professonal tranng (FACT), marketng (FACM) and lterature (FACL) servces showed no effect. Negatve mpact of these non-fnancal servces could be attrbuted to correlaton wth unobservable varables lke the level of rsk of the project of the borrower (for professonal tranng) or dosyncratc shocks (for the access to health). However, access to professonal tranng has a postve sgnfcant effect n the estmaton of RTG3. Ths questons the possblty for these servces to ncrease the borrower s ablty to succeed n ther projects. The costs ncurred by non-fnancal servces cannot be justfed by ther postve effects on repayment performance. Important dfferences among the MFIs appeared n the mpact of socal ntermedaton (non-fnancal servces) on repayment performance. Whereas, the BRAC experenced a postve nfluence of ts marketng servce and negatve nfluence of ts health servces, the BRDB experenced a negatve nfluence of both ts professonal and health servces. Access to basc lteracy had a negatve mpact on the repayment of the Grameen borrowers whle access to health servces proved to have a postve mpact. The dfference among the MFI on the mpact of non-fnancal servces gve rse to a feld for future research on non-fnancal servces: what knd of servce s offered to a specfc borrower?, why?, s there a dfference n the qualty of the provson of these servces among the dfferent MFIs? and what are the costs and benefces nduced by these servces? The splt sample regresson gves evdence of the postve mpact of dynamc ncentves on the repayment behavor of the borrowers as credt ratonng (CRd) has a postve sgnfcant mpact on the repayment performance. Small groups (NMBG1) had better results n terms of rembursement than medum groups whch confrms the postve mpact of peer montorng but bg groups (NMBG4) also proved to have a better performance probably due to a large use of ntra-group nsurance. When a the borrower who would or would not lke to borrow more at the same nterest rate. 14 SCOHAB s a dummy whch takes the value of one f the borrower and hs/her spouse lve n the same place. 15 In the study of Wydck, SAMESEX s the only varable of socal tes that affected repayment performance and t has a negatve mpact, only n urban groups. 14

15 grace perod of 3 months s gven for rembursement, the relatve performance of smaller group dsappears. We can thus surmse that development of nsurance servces could gve rse to a better explotaton of peer montorng. The characterstcs of the contract showed sgnfcant mpact on the repayment of the borrower. The nstrumented sze of the loan (PPRIN) presented a postve mpact contrary to what was found before nstrumentaton and n the study of Sharma and Zeller, Accordng to the theory of dynamc ncentves, loan sze ncreases wth the duraton n membershp. The postve mpact of loan sze could thus be explaned by a better ablty of the borrowers to succeed n ther projects wth tme. It s mportant to know whether the mpact of the loan sze s postve or negatve because a negatve mpact of loan sze could lead to embarrassng conclusons 16. Duraton (PDURATION) showed a negatve mpact n the whole sample but a postve one n two of the three splt sample regressons. The repayment s sgnfcantly hgher at the end of the year (ENDY) and durng the harvest seasons (HARVEST). Ths s stll the case when we allow a grace perod of three months (RTG3). Borrowers that have to pay back ther loan durng the lean agrcultural season seem to face permanent dffcultes n payng back, maybe because of expensve copng strateges. That s why, f the cash management allows t, MFIs should avod fxng the redempton date of the loan durng the lean agrcultural season 17. Intal fees (FEE) had a postve effect. These fees are relatvely low compared to the amount borrowed (cf. Table 3) but ths formalzaton of the membershp of the MFI seems to play a symbolc role for the borrower 18. The number of weeks the borrower had to wat before recevng hs loan (PGWBL) had no sgnfcant mpact on repayment performance and we cannot postulate an approprate screenng of the MFI wth ths nstrument. The estmaton also reveals that personal (PURPP) and agrcultural loans (PURPA) are better rembursed than busness ones. Ths must be related to the negatve effect of employment (EO) and self-employment (SELFEO) n non agrcultural actvtes. These actvtes are more rsky but have an mportant role to play n the economc development of the area. Ths can justfy the prorty gven by the Grameen Bank to self-employment and 16 Indeed, when bg loans correspond to rsker projects wth hgher level of returns they are mportant for the economc development of the area and for the long term wealth of the borrower. Ths s a rather optmstc nterpretaton and we could also thnk that bg loans do not meet the borrower needs and are not suted to the local economy. If borrowers cannot take advantage of bg loans they wll not be able to be ndependent of the MFI and to get out poverty. Lastly, we could thnk that ncentves to strategc default ncrease wth loan sze. Ths dscusson would temperate the nterest for the MFI to attrbute loans of bg sze. 17 Ths s one of the propostons of the Gunean borrowers of the Crédt Rural de Gunée (Condé, Bouju & Gentl, 2001). 18 Ths can be lnked to the postve mpact of rules of conduct n the study of Wenner (1997). 15

16 ncome generatng actvtes even accountng for the lower performance of those projects n terms of rembursement. The sze of the loan portfolo (NBYT) showed sgnfcant negatve mpact 19 whch could have dfferent meanngs. Frstly, n a fnancal perspectve, as the MFIs grow they can fnance projects wth hgher levels of rsk because they can dversfy ther rsks. Secondly, ths could be explaned by a fall n the manager's tme allocated to each group as the loan portfolo ncreases. And lastly f as frequently reported, the MFIs do not systematcally exclude bad or late borrowers, the experence of the MFI behavor wll decrease the expected weght of the MFI's sancton n case of default. Consderng that the number of borrowers steadly ncreased over the years, the negatve mpact of the sze of the MFI portfolo could be lnked to a better knowledge of the borrowers of the behavor of the MFI. Indvdual and household characterstcs lke sex, household sze, dependency rato have no sgnfcant mpact on repayment performance whch confrms the fndng of Zeller (1998) that tradtonal bas aganst female or hgh famly sze are not justfed. Contrary to the tradtonal belef, female borrowng (SEX) even had a postve mpact on BRAC and BRDB rembursement performance and on RTG3 for the whole sample. Ths could justfy the preference gven to women by the MFIs. However, the repayment performance s nfluenced by age, wealth 20 (wealth s proxed by the value of productve assets, PASSET, the log of per capta consumpton, LNPCXA, and the use of specfc garbage dsposton, GARBAGE) and ablty (proxed by the educaton level, YEDLEVEL). Ths general fndng holds n the splt sample regressons and n the estmaton of the repayment probablty, gven a grace perod of three months. The share of the varance explaned by our varables remans small whch could advocate for mssng varables. These varables mght be communty-specfc varables lke local economc condtons, degree of monetarzaton, collectve shocks 21 lke exposure of the area to natural dsasters (Zeller, 1998 ; Khandker, Khally & Khan, 1994) ; program-specfc varables lke functonng costs of the MFIs (Khandker, Khally & Khan, 1994) or borrower-specfc characterstcs lke dosyncratc shocks (llness and njures) Is the fxaton of loan sze effcent? 19 The negatve mpact of the sze of the loan portfolo can be related to sgnfcant negatve mpact of the age of the branch n Khandker, Khally & Khan (1994). 20 The varables for wealth dd not prove any sgnfcance n the regresson of RTG3. 21 Vllage fxed effects were not used because of too lttle loan observatons per vllage. 16

17 Table 8 report the results of the smultaneous estmaton of the sze of the loan ssued from the smultaneous regresson of the prncpal and duraton of the loan (step 4). We consder that MFIs gve bgger loans for project they antcpate to be successful. As duraton n membershp grows, MFIs tend to gve larger loans to ther more experenced borrowers and ths corresponds to the use of dynamc ncentves. However, as prevously exposed, the age of the group has a negatve mpact on repayment performance. Groups made of borrowers of both sex receved larger loans but the standard composton of groups s unsex. Group homogenety n terms of age had a postve mpact on loan sze whch could mean that MFIs consder group homogenety as a specfc factor nfluencng the repayment behavor. Access to each of the non-fnancal servces has a postve nfluence on the loan sze, except for the BRDB. These servces mght ndeed ncrease the borrowers' capabltes and thus ncrease ther probablty of success 22 even f they proved mxed effect on the repayment performance. The sze of the group had a non lnear mpact. Small groups and bg ones receved smaller loans compared to medum sze groups. The MFI may take nto account the overcompensaton of lack of nsurance possbltes n smaller groups (the long term repayment performance s ndeed smaller) on postve effect of peer montorng (repayment at the due date s hgher). As they receved smaller loans, bggest groups are maybe expected to be too heterogeneous to explot group dynamc effcently (ths s confrmed by the regresson of RTG3). The MFIs seem to screen borrowers wth the tme they have to wat before they receve a loan as the longer they wat, the smaller the loan they receve. Ths screenng dd not prove ts effectveness as the number of weeks the borrower had to wat before recevng hs loan dd not prove to have any effect on the repayment (contrary to what was found before nstrumentaton). The MFI attrbutes sgnfcantly smaller loans to borrowers who have to pay ntal fees but these borrowers had a better repayment performance. Ths could be due ether to an effectve nctatve power of ntal fees or to an neffectve use of ntal fees. The sze of the loan portfolo had a sgnfcant postve nfluence on the prncpal of the loan offered by the MFI whch could be due to a hgher accepted level of rsk by the MFI. As prevously seen ths varable has a negatve mpact on repayment performance. 22 The use of health, marketng, lterature or professonal tranng servces mght also be compulsory for bg loans and the avalablty of these servces could reflect the age of the branch whch s postvely correlated wth the loan sze. 17

18 Wealthy borrowers (wealth s proxed by log of per capta consumpton, prvate access to electrcty, use of tolets and the number of landed relatves) receved bgger loans. Educaton level and age of the borrower had a postve effect on the loan sze attrbuted by the MFI whch makes sense but these two varables had a negatve mpact on the repayment. Other tradtonal factors lke household sze and the number of landed relatves are used n the MFI s determnaton of the sze of the loan but wth no effect on repayment performance. Ths leads us to thnk that the MFI should not take these varables nto consderaton when determnng the loan sze. 6. Concluson. The am of ths artcle was to test the explanatory power of models that attrbute the performances of MFIs n terms of repayment rate to the use of group lendng, socal ntermedaton and dynamc ncentves. Fnancal nnovatons showed mxed results n the determnaton of the repayment performance. The age of the group, a prox for socal tes nsde the group, showed a sgnfcant negatve mpact on the rembursement whch rases the queston of the necessty of specfc ncentves nstruments for experenced borrowers. The socal tes of the borrower out of hs group have the expected postve mpact as well as the proxy for dynamc ncentves. In terms of sex, group homogenety proved a postve mpact on repayment performance 23 but we can not attrbute the same postve mpact to group homogenety n terms of age or educaton level. Non-fnancal servces dd not show a postve mpact n all the cases whereas MFIs tend to attrbute bgger loans to borrowers who have access to these servces. Negatve mpact of these non-fnancal servces could be attrbuted to correlaton wth unobservable varables lke the level of rsk of the project of the borrower (for professonal tranng) or dosyncratc shocks (for the access to health). The sze of the group showed a non-lnear effect revealng the conflct of the benefts of peer montorng and these of nsurance due to group. The development of nsurance servces could permt the explotaton of the full beneft of peer montorng. We also found that other loan characterstcs lke the purpose of the loan, the duraton and the prncpal of the loan have to be taken as control varables for an effectve analyss of determnants of the repayment performance. 23 We ddn't address the queston of the predcted postve mpact on repayment of group homogenety n terms of rsk as a result of peer selecton (Ghatak, 1999). Nevertheless, both the study of Zeller (1998) and the study of Sharma and Zeller (1997) gve evdence from Madagascar and Bangladesh that show that ths knd of homogenety has a negatve mpact on repayment performance. 18

19 We draw one s attenton to the endogenety of the prncpal and duraton of the loan n the determnaton of the repayment performance and we advse the use of nstrumental varables for future analyss of repayment rate. Our estmaton could underestmate the effects of the fnancal nnovatons as we can only estmate margnal effects as we do not have a control group and as each of the three studed MFIs use such fnancal nnovatons. Further researches on the effectve explanatory power of socal tes are needed as emprcal studes do not manage to prove any mpact of socal tes on repayment performance. Group homogenety s a prerequste n many MFIs and further researches must be undertaken to understand what type, f any, of group homogenety postvely nfluences the borrowers rembursement. The mxed mpact of non-fnancal servces and the dfference amongst MFIs n ther provson gve rse to further studes on the provson (condton of attrbuton, level of qualty, ) and mpact of these servces. In the same lne, further nformaton on the formaton before loan and tme borrowers have to wat before they receve ther loan could provde nterestng comments. Seasonalty seems to affect the borrower and hs/her vulnerablty. Further researches on ths pont and on possbltes to use the harvest cycle to ncrease the repayment performance are thus necessary. The above analyss gves however rse to comments on ways to mprove the repayment performance by a more adequate loan provson. MFIs should fnd new ncentves for more experenced lendng groups of borrowers. As engagement n non agrcultural actvtes as well as partcpatng to a professonal tranng has a negatve mpact on repayment performance, MFIs should devote a specfc attenton to the tranng of borrowers engaged n non agrcultural projects. MFIs could take advantage n payng attenton to the way they grow: devote the same attenton to borrowers along tme, pay attenton to keep credble the threat of droppng out bad borrowers and avod gvng sgnfcantly bgger loans as they extend ther borrowers portfolo. MFIs should avod fxng the redempton date of the loan durng the lean agrcultural season. MFIs should lmt the use tradtonal varables lke the educaton level, sex or sze of the famly nto consderaton when determnng the loan sze as those varables mght have an unexpected effect and go contrary to tradtonal expectatons on age, sex or educaton. 19

20 References. Armendarz de Aghon, B. & Morduch, J. 2000, Mcrofnance beyond group lendng., Economcs of transton, Vol. 8, No. 2, pp Besley, T. 1995, Nonmarket nsttutons for credt and rsk sharng n low-ncome countres, Journal of Economc Perspectves, Vol. 9, No. 3, pp Besley, T. & Coate, S. 1995, Group lendng, repayment ncentves and socal collateral., Journal of Development Economcs, Vol. 46, pp Bratton, M. 1986, Fnancng smallholder producton: A comparson of ndvdual and group credt schemes n Zmbabwe. Publc Admnstraton and Development, No. 6, pp Condé, K., Bouju, S. & Gentl D. 2001, Le Crédt Rural de Gunée vu par ses acteurs., Edtons du GRET,. Dagne, A. Chmombo, W. Smtowe, F. & Mataya, C. 2000, Desgn and sustanablty ssues of rural credt and savng programs for the poor n Malaw: An acton-orented research project. IFPRI, Washngton, D.C. Edgcomb, E. & Barton, L : Socal ntermedaton and mcrofnance programs: A lterature revew., Mcroenterprse Best Practces, USAID. Fremer, M & Gordon, M.J., 1965 Why bankers raton credt., Quarterly Journal of Economcs, Vol. 79, pp Ghatak, M. 1999, Group lendng, local nformaton and peer selecton., Journal of Development Economcs, Vol. 60, pp Goetz, AM. & Sen Gupta, R. 1996, Who Takes The Credt? Gender, Power and Control over Loan Use n Rural Credt Programmes n Bangladesh, World Development, Vol. 24, No. 1, pp Henry, C. Sharma, M. Lapenu, C. & Zeller, M. 2000, Assessng the Relatve Poverty of Mcrofnance Clents: A CGAP Operatonal Tool. CGAP, Washngton, D.C. Khandker, S.R. Khally, B. Khan, K. 1994, Grameen Bank: performance and sustanablty., World Bank Dscusson Paper, No.306. Madajewcz, M., 1999 Captal for the poor: The effect of wealth on the optmal credt contract., Columba Unversty, Draft, June. Morduch, J. 1998, Does mcrofnance really help the poor? New evdence from flagshp programs n Bangladesh., Mmeo., Department of Economcs, Harvard Unversty. Paxton, J.A. 1996: Determnants of successful group loan repayment: an applcaton to Burkna Faso., Ph.D. Dssertaton, The Oho State Unversty. Ptt, M.M. & Khandker, S. R. 1998, The mpact of group-based credt programs on the poor households n Bangladesh: does the gender of partcpants matter?, Journal of Poltcal Economy, Vol.106, No. 5, pp Ptt, M.M. 1999, Reply to Jonathan Morduch s Does mcrofnance really help the poor? New evdence from flagshp programs n Bangladesh., Mmeo, Department of Economcs, Brown Unversty. Ptt, M.M.& Khandker, S. R. 1996, Household and ntrahousehold mpact of the Grameen Bank and smlar targeted credt programs n Bangladesh., World Bank Dscusson Papers, No.320. Sharma, M. & Zeller, M. 1997, Repayment performance n group-based credt programs n Bangladesh: An emprcal analyss., World Development, Vol. 25, No.10, pp Smth, R.J. & Blundell, R.W. 1986, An exogenety test for a smulataneous equaton tobt model wth an applcaton to labor supply., Econometrca, Vol. 54, No. 3, pp Stgltz, J.E. 1990, Peer montorng and credt markets., The World Bank Economc Revew, Vol. 4, No.3., pp