Economics Discussion Paper

Transcription

1 Economcs Dscusson Paper EDP-0625 Does the Mcrofnance Reduce Poverty n Inda? Propensty Score Matchng based on a Natonallevel Household Data Thankom Arun, Katsush Ima and Frances Snha September 2006 Correspondence emal: Katsush.Ima@manchester.ac.uk School of Socal Scences, The Unversty of Manchester Oxford Road Combnng the strengths of UMIST and The Vctora Unversty of Manchester

2 Manchester M13 9PL Unted Kngdom Combnng the strengths of UMIST and The Vctora Unversty of Manchester

3 Does the Mcrofnance Reduce Poverty n Inda? Propensty Score Matchng based on a Natonal-level Household Data Authors: Thankom Arun Insttute for Development and Polcy Management, School of Envronment and Development, Unversty of Manchester, UK Katsush Ima Economcs, School of Socal Scence, Unversty of Manchester, UK Frances Snha EDA Rural Systems, Inda Abstract: Drawng upon the data n 2001, the present study analyses the effect of Mcro Fnance Insttutons (MFIs) on poverty of households n Inda. The propensty score matchng s employed to estmate the poverty-reducng effects of the access to MFIs and the loans used for productve purposes to take account of the endogenety problem. Sgnfcantly postve effects on the multdmensonal poverty ndcator suggest the role of MFIs n poverty reducton. We also show that households n rural areas need loans from MFIs for productve purposes to reduce poverty, whle smply accessng to MFIs s suffcent for urban households to reduce t. Key Words: Mcrofnance, Poverty, Evaluaton, Inda, Propensty Score Matchng JEL Classfcaton: C21, I30, I38, O16, R51 Acknowledgements: Ths study s based on the natonal-level household data n Inda provded by the EDA research team n Inda ( who coordnated and undertook a natonal level mcrofnance mpact study for the SIDBI Foundaton for Mcro Credt. The authors are grateful to comments from Raghav Gaha, Davd Hulme, and partcpants n semnars at Unversty of Manchester. The vews expressed are those of the authors and they bear full responsblty for any defcences that reman. Contact: Dr Katsush Ima Economcs, School of Socal Scences Oxford Road, Manchester M13 9 PL, UK Phone:+44 (0) Fax +44 (0) Emal : Katsush.Ima@manchester.ac.uk

4 Does the Mcrofnance Reduce Poverty n Inda? Propensty Score Matchng based on a Natonal-level Household Data I. Introducton The relatonshp between mcrofnance and poverty s stll an unsettled queston, manly due to the dsproportonate focus on fnancal sustanablty of mcrofnance nsttutons. Ths vew supports the argument that mcrofnance nsttutons should am for sustanable fnancal servces to low ncome people, and, reasonably contradcts the potental role of ths nsttutonal nnovaton n poverty reducton and empowerment. Irrespectve of these approaches, many organzatons have used these nsttutons to develop a range of servces on a contnung bass to address the requrements of the poor people. However, there s an argument that mcrofnance supports nformal actvtes whch have a low market demand and the aggregate poverty mpact of mcrofnance s modest or non exstent n a slow growth economy. However, Khandker (2003) addresses ths ssue usng household level panel data n the context of Bangladesh and confrms that mcrofnance programmes have a sustaned mpact n reducng poverty among the partcpants and a postve spll over effect at the vllage level. The fndngs of the study ndcate that mcrofnance programmes are helpng the poor beyond ncome redstrbuton wth contrbuton to local economc growth. The development of the mcrofnance sector s based on the concept that the poor possess the capacty to mplement ncome generatng actvtes and s lmted by lack of access and nadequate provson of savngs, credt and nsurance facltes. However, there are apprehensons on the role of mcrofnance nsttutons to provde servces and products to the poorest of the poor category (Hulme and Mosley, 1996). The real challenge n servng poorest of the poor s to dentfy the benefcares from varous categores such as fnancal servces alone, 2

5 non fnancal servces along wth fnance, and non fnancal servces before partcpatng n market orented fnance (Meyer,2002). The recent lterature suggest the need for mcrofnance nsttutons to move away from the perceptons of a products based organzaton to reflect the heterogenety of demand structure for fnancal servces/products by the poor people and the complex lvelhood need of the poor (Arun and Hulme, 2003). To capture the mult-dmensonal dmensons of poverty, such as basc needs, capabltes, socal captal or vulnerablty, ths study used Index Based Rankng (IBR) Indcators based on a natonal-level household survey to examne the role of mcrofnance n poverty reducton n Inda. In Inda, despte the recent economc growth at natonal level 1, poverty remans stll a serous problem for polcy-makers because () the growth s manly drven by growth n a few sectors n urban area, such as ndustry and servce sectors 2 and s unevenly dstrbuted wthn the urban, and between urban and rural area, () trckle down of the benefts from economc growth to the poor s lmted even n the area wth hgh economc growth due to socal factors, and () the rural-urban lnkages are weak due manly to the geographcal factors. Based on the US1$ a day poverty lne, poverty head count rato was 35.3 per cent n 2000, whch had been reduced from 42.3 per cent n Poverty reducton remans the country s major challenge. Recently, the Government of Inda has ntated some targeted nterventons such as Natonal Rural Employment Guarantee Scheme has been proposed to extend the Employment Guarantee Scheme (EGS) n Maharashtra to the poorest 150 dstrcts. 3 However, the scheme s not avalable to those who are old or are physcally weak. The EGS potentally has some effects to help the partcpants cope wth varous shocks (e.g. drought), but these rsk-reducng benefts are lmted only to agrcultural workers wth hgh wages because the poor cannot afford the hgh costs of partcpatng n the scheme (Scandzzo et al., 2005). Thus, the potental of ths type of 3

6 nterventon to reduce poverty and vulnerablty of households s lmted f a negatve shock occurs frequently and often n the large magntude as n Inda. The recent studes on vulnerablty n Inda suggest that a large secton of rural households n Inda s not only poor but also vulnerable to both dosyncratc and aggregate shocks (e.g. Gaha and Ima, 2004, 2006; Lgon, 2005) 4. Mcrofnance schemes play potentally mportant roles not only n reducng poverty but also n reducng vulnerablty wth the schemes approprately desgned and nstrumented to cover the poorest households. In Inda, untl the early 1990s, the mcro fnancal servces were provded through a varety of state sponsored nsttutons, whch resulted n mpressve achevements n expandng access to credt partcularly among the rural poor. Although many of these bank branches n rural areas were unproftable, they dd play a postve role n fnancal savngs and reducng poverty whch s evdent n the fact durng the perods; the share of total fnancal nsttutons n rural house hold debt has ncreased from 8.8 per cent to 53.3 percent, and the role of money lenders has declned sgnfcantly durng ths perod. However, despte the vast network of bankng and cooperatve fnance nsttutons and srong mcro components n varous programmes, the performance of formal fnancal sector s stll far behnd n reachng out to reflect and respond the requrements of the poor. Ths s true to an extent that most of the ntatves n the formal sector are based on the preconceved noton that rural people are not bankable and not reflectng the heterogeneous needs of the poor whch led to the growth of several subsdy-lnked credt programs n Inda where the state had a crucal role n dentfyng the borrowers and n the allocaton of credt and subsdy. In the 1990s, the mcrofnance schemes or mcrofnance nsttutons (MFIs) have become ncreasngly mportant n Inda manly due to the potental advantages of these schemes such as 4

7 better access to local knowledge and nformaton at communty level and use of peer group montorng. For example, Mcrofnance based on SHGs (Self-Help Groups) has become ncreasngly mportant n Inda through ts flexble nature (Mosley and Arun, 2003). Ths dstnctve mcrofnance programme s based on the exstng bankng network n delverng fnancal servces to the poor. These groups are buldng on the tradtonal nsttuton of ROSCA, and provde access to both savngs and credt for the asset less poor. The rapd expanson of SHG groups has even generated dscussons on formal fnancal nsttutons and mcrofnance nsttutons, and approprate polcy decsons based on t. A recent study n Pune dstrct n Maharashtra shows that whle the targetng performance of mcrofnance through SHGs was unsatsfactory n terms of ncome, t was satsfactory n caste, landlessness and llteracy and thus facltates empowerment of women (Gaha and Nandh, 2005). Ths study also fnds that loans were used largely for health and educaton of chldren and argues aganst confnng the mpact assessment of mcrofnance to conventonal economc crtera alone. Despte ts ncreasngly mportant roles of MFIs n poverty reducton n Inda, there have been relatvely few studes that evaluate the effect of MFIs on poverty n Inda. The present study ams to provde emprcal evdence on the relatonshp between MFIs and poverty usng the large-scale household data set n Inda whch was collected by one of the authors wth the am to assess the mpact of mcrofnance. In our study, poverty s defned by the IBR (Indexed Based Rankng) Indcator, a composte ndcator that captures varous aspects of wellbeng, ncludng land holdngs, salared ncome sources, lvestock, transport assets, housng, and santaton faclty 5. Our research queston s smple -whether the access to MFI reduces poverty. Smple comparson of the average of IBR ndcator of households wth access to MFI and that wthout s not approprate because MFI s not randomly dstrbuted due to sample selecton 5

8 where MFI targets poor households or poor households tend to take loans from or save at MFIs (EDA Rural Systems, 2005). The commonly used econometrc method s the Instrumental Varable (IV) estmaton, but the survey data do not necessarly contan data for the vald nstruments or the results are not relable f the nstrument s not sgnfcant. As an alternatve to IV, the present study employs propensty score matchng estmators to dentfy the polcy effects as average treatment effects to take nto account the sample selecton bas. The rest of the paper s organzed as follows. Secton II summarses the survey desgn and data. Secton III descrbes econometrc methodology we use to estmate average treatment effects based on propensty score. Secton IV provdes econometrc results and man fndngs, followed by the conclusons n the fnal secton. II. Survey Desgn and Data 6 Detals of Survey The orgnal survey was carred out by EDA Systems for the SIDBI (Small Industres Development Bank of Inda) n 2001 as a part of SIDBI s mpact assessment study of ts mcro fnance programme. The two-stage longtudnal soco-economc research was undertaken to assess on a natonal scale, the development mpact of MFI programmes. The study covered a sample of 20 partner Mcro Fnance Insttutons (MFIs) of SIDBI and 5327 households, ncludng both clents and non-clents (EDA Systems, 2005; SIDBI, 2005). Our study s based on the cross-sectonal data set for these households. The hypothess of the study s: Mcrofnance s an effectve strategy for extendng fnancal servces to the poor and other dsadvantaged groups by the formal sector fnance, whch was generally supported by the reports (SIDBI, 2005; EDA Systems, 2002, 2005). 6

9 Suppostons whch have also been corroborated by the survey nclude: (1) MFI servces reach those who have not yet accessed formal sector fnance; (2) MFIs outreach s generally focused on poorer areas; (3) MFIs serve all castes and communtes; and (4) majorty of clents are women and mcro fnance has supported ncreased non-farm employment (SIDBI, 2005). The study confrmed that (a) mcro credt has sgnfcantly promoted the ncdence of borrowng for nvestment (60% of clent households versus 38% of non-clent households have borrowed for nvestment); (b) mcro fnance has supported ncreased non-farm employment, especally for poor/borderlne clents; (c) MFI nvolvement has enabled clent households to dversfy ther ncome sources and reduce dependency; (d) MFIs nvolvement postvely affects chldren s educaton, although the clent- non-clent dfference s not too large; (e) MFI nvolvement helps clents protect aganst rsk through lvelhood dversfcaton, asset buldng and savngs and; (f) access to Mcro fnance has had a sgnfcant postve mpact on women s economc and socal empowerment (SIDBI, 2005). Fve MFIs were selected as representatve of 31 MFIs n SFMC s lst of current partners representng dfferent regons, models of mcrofnance (Self Help Group -SHG, Grameen, Indvdual Bankng and sector/enterprse specfc cooperatves), age, and outreach to members and range of servces. At each MFI, two to four sample areas (vllages or urban wards) were selected purposefully to represent a typcal area of the MFI n terms of soco-economc context and range of MFI programmes. Wthn each sample area, a stratfed random sample of clents, non-clents and dropouts was drawn usng wealth rankng as a bass for stratfcaton (EDA Systems, 2002, 2005). Coverage of a mcrofnance programme s computed on the bass of number of clent households compared to the total number of households n sample vllages and urban slums. In 7

10 rural areas, outreach of the sample s to 25% households n smaller vllages (<500 households) and 13% n larger vllages. The data set has 3320 clent households and 1226 non-clent households, the latter of whch s used as a control. The man objectve of ths study s to statstcally compare poverty defned by the wellbeng rankng for both groups of households usng the propensty-score rankng. Index Based Rankng (IRR) Indcators Index Based Rankng Indcators were created to overcome the ncome or consumpton based poverty measures and capture non-ncome or mult-dmensonal dmensons of poverty, such as basc needs, capabltes, socal captal or vulnerablty (Shna 2002). A score ndex, such as IBR s useful to capture varous dmensons of poverty because of ts hgher practcalty (e.g. less costly than those for expendture surveys; based on less-senstve /obtrusve and smple questons) and hgher relablty due to lower rsk of falsfcaton or error. Respondents are asked about qualty of lfe n several dmensons, ncludng (a) ncome (e.g. regularty of ncome, type of employment), (b) productve assets (e.g. land-holdng and number and qualty of lvestock, only for rural area), (c) basc needs (e.g. food securty, use of health care servces, santaton), (d) capabltes /human and socal captal (e.g. chldren s schoolng, major expenses on health treatment), (e) housng (e.g. materal or sze, ownershp) and (f) household assets (ownershp of bcycle or 4-wheeler, TV etc.). Then IBR ndcators are created as a weghted sum of scores (as a maxmum score of 60) for each category (ranged from 0-6) usng dfferent weghts dfferent for rural and urban areas n four categores, namely very poor, poor, borderlne non-poor were created (Snha, 2002). However, to keep contnuty of the varable, we use the IBR ndcators as they are n the followng sectons. 8

11 III. Methodology Our man hypothess s the access of mcrofnance nsttutons (MFIs) reduces poverty defned by the IBR ndcators. Because we have only cross-sectonal data, we can compare IBR ndcators of households wth access to MFIs and those wthout, as long as MFIs are randomly dstrbuted across the sample or there s no sample selecton bas. The methodology wdely used n the lterature s the IV estmaton or the Heckman sample selecton model where the access to MFIs s estmated n the frst stage and the effect of access to MFI on poverty s estmated n the second. Whle useful, the estmaton results are only robust f they are estmated wth vald nstruments whch affect the access to MFIs, but not poverty. In general, t s not necessarly easy to fnd such vald nstruments. Also, f the lnear regressons are used, the lnear relatonshp s assumed between dependent varables and explanatory varables and dstrbutonal assumptons have to be mposed on the explanatory varables (Foster, 2003). Because our dependent varable s the IBR ndcator whch has already ncorporated varous aspects of wellbeng, t s not easy to fnd a vald nstrument n our context. Hence, we use statstcal matchng whch has been wdely used n the medcal study where dose response of patents s analysed. Ths frst specfes a functon matchng the proxmty of one household to another n terms of household characterstcs and then households are grouped to mnmze the dstance between matched cases (Foster, 2003). Merts of usng statstcal matchng over the IV estmaton ncludes; the former does not assume lnearty; t s vald even though dstrbutons of explanatory varables of treatment and control groups overlap relatvely lttle, and t does not requre a vald nstrument. Rosenbaum and Rubn (1983) proposed statstcal matchng usng the propensty score, the predcted probablty that an ndvdual receves the treatment of nterest 9

12 (e.g. fnancal servces, such as loans or savngs n our case) to make comparsons between ndvduals wth the treatment and those wthout. Methodologcal ssues and programs for propensty score matchng estmaton are dscussed n detals, for example, by Becker and Ichno (2002), Deheja (2005), Deheja and Wahba (2002), and Smth and Todd (2005). We wll summarse estmaton methods for the propensty score matchng based on these studes. The propensty score s the condtonal probablty of recevng a treatment (or of havng access to MFI) gven pre-treatment characterstcs, X (or household characterstcs). where { 0, 1} { D = X} E{ D X } p ( X ) = Pr 1 = (1) D = s the bnary varable on whether a household has access to MFIs (1) or not (0) and X s the multdmensonal vector of pre-treatment characterstcs or tme-nvarant or relatvely stable household characterstcs n our context. It was shown by Rosenbaun and Rubn (1983) that f the exposure to MFI s random wthn cells defned by X, t s also random wthn cells defned by p(x) or the propensty score. The polcy effect of MFI can be estmated n the same way as n Becker and Ichno (2002) as: τ = = E E E { W W D = 1} 1 { E{ W1 W0 D = 1, p( X ) } { E{ W D = 1, p( X )} E{ W D = 0, p( X )} D = 1} (2) where denotes the -th household, W 1 s the potental outcome (as wellbeng or poverty status captured by IBR ndcator) n the two counterfactual stuatons wth access to MFI and wthout. So the frst lne of the equaton states that the polcy effect s defned as the expectaton of the dfference of the IBR ndcator of the -th household wth access to MFI and that for the same 10

13 household n the counterfactual stuaton where t would not have had access to MFI. The second lne s same as the frst lne except that the expected polcy effect s defned over the dstrbuton of the propensty score. The last lne s the polcy effect as an expected dfference of the expected IBR score for the -th household wth access to MFI gven the dstrbuton of the probablty of accessng MFI and that for the same household wthout MFI gven the same dstrbuton. Formally, the followng two hypotheses are needed to derve (2) gven (1). Lemma 1 Balancng Hypothess (Balancng of pre-treatment varables gven the propensty score) If p (X) s the propensty score, then D X p( X ) Ths mples that gven a specfc probablty of havng access to MFI, a vector of household characterstcs, X s orthogonal to (or uncorrelated to) the access to MFI. In other words, for a specfc propensty score, the MFI s randomly dstrbuted and thus on average households wth MFI and those wthout are observatonally dentcal (gven a propensty score). Otherwse, one cannot statstcally match households of dfferent categores. Lemma 2 Unconfoundedness gven the propensty score If treatment (or whether a household has access to MFI) s unconfounded,.e. W, W 1 2 D X Then, assgnment to treatment s unconfounded gven the propensty score,.e. W, W 1 2 D p( X ) The latter mples that gven a propensty score the IBR ndcator s uncorrelated to the access to a MFI. If the above lemmas are satsfed, the polcy effect can be estmated by the procedures 11

14 descrbed n Becker and Ichno (2002) and Smth and Todd (2005). Each procedure nvolves estmatng logt (or probt) model: Pr { D = 1 X } = Φ (h(x )) where Φ denotes the logstc (or normal) cumulatve dstrbuton functon (cdf) and h(x ) s a startng specfcaton. One possble procedure for statstcal matchng s Stratfcaton Matchng whereby the sample s splt n k equally spaced ntervals of the propensty score to ensure that wthn each nterval test the average propensty scores of treated and control households do not dffer. We dd not use Stratfcaton Matchng as observatons are dscarded when ether treated or control unts are absent. Instead, we use other varants n matchng estmators of the average effect of treatment on the treated, namely, Nearest Neghbour Matchng and Kernel Matchng. 7 Nearest Neghbour Matchng s the method to take each treated unt and search for the control unt wth the closest propensty score, whle wth Kernel Matchng all treated are matched wth a weghted average of all controls wth weghts that are nversely proportonal to the dstance between the propensty scores of treated and controls (see the Appendx 1 for detals). IV. Results Ths secton provdes the results for matchng estmators to nvestgate the effects of the access to MFIs on poverty. Because of the fundamental dfferences of envronment, ndustral structures, household characterstcs or actvtes between urban and rural areas, we wll frst derve the estmatons for total households and then apply the same specfcaton for urban areas and rural areas separately. 12

15 Only havng access to MFIs does not necessarly affect household wellbeng partcularly f one does not have a long transacton hstory or has only a zero account. Therefore the present study uses the followng two dfferent defntons of the access to MFIs: a) whether a household s a clent of any MFI ( mf_status ) or not, and b) whether a household has taken a loan from MFI for a productve actvty ( mf_productve ). 8 The frst defnton s used to see the effect of smply accessng MFI on poverty and that of utlzng mcrofnance for any productve purposes Table 1 provdes descrptve statstcs used for the estmaton. The frst two panels show that about three quarters (rrespectve of beng n rural or n urban areas) n the sample households have access to MFIs, whle about a half of them have taken loans for productve purposes. On the latter, the share s hgher n rural areas than n urban areas. The hgher share of beng headed by a female member mples that MFIs are targeted to those households and controls are selected by a smlar crteron. Educaton of household head of most of the households n rural areas s ether llterate or completed prmary school, whle all completed only prmary school n urban areas. Ths reflects only poor households wth low levels of educaton are targeted by MFIs. Household sze s about fve. About 30% of the sample households are scheduled caste or trbe reflectng that the mcrofnance targets manly poor households. The average IBR ndex of households n rural areas s lower than that n urban areas, mplyng that poverty s severer n rural areas. (Table 1 to be nserted around here) Estmaton results of logt model n Table 2 are generally ntutve n the case for the entre households where dependent varable s MFI_status (.e. Case A-1) (e.g. a household wth older 13

16 household head s more lkely to be a clent; a household wth female household head s more lkely to be a clent, whch reflects the fact that mcrofnance focuses on women; educaton has a postve and sgnfcant mpact; dependency rato has a negatve and sgnfcant effect). However, n Case B-1 where MFI_productve s estmated for the entre households, a few changes are observed. The coeffcent of Age ceases to be sgnfcant, and that for Female s negatve and sgnfcant, that s, a household wth male household head s more lkely to take a loan for productve purposes. Ths reflects the fact that whle mcrofnance focuses on women, male-headed households tend to be n a more advantageous poston to take loans for productve purposes. The coeffcent of Educaton s negatve and sgnfcant, and that of Caste_dum s postve and sgnfcant. These results mply that relatvely poor households whose household head has a lower level of educaton or s llterate, or whch are n scheduled caste or scheduled trbe tend to take a loan for productve purposes. Cases A-2 and A-3 and Cases B-2 and B-3 n Table 2 show the results of logt model appled separately for urban areas and rural areas. In Cases A-2 and A-3, n urban areas, the coeffcent of Age s not sgnfcant, whle n rural areas t s postve and sgnfcant at 10% level, whle other results look generally smlar to those n Case A-1. The results for rural areas n Case B-3 are generally consstent wth those n Case B-1 for total households except that the coeffcent of Educaton s postve and sgnfcant. (Table 2 to be nserted around here) Based on the results of logt model n Table 2, we derved the propensty scores for each case. 9 Wth the same specfcaton appled to all the cases, the balancng hypotheses are satsfed n all the cases except case B-1. Only n ths case dd we try a dfferent specfcaton where the 14

17 square of age s dropped and then t has been found that the balancng hypothess s satsfed. 10 It s assumed as n Becker and Ichno (2002) that unconfoundedness s satsfed. Table 3 and Table 4 provde the results for matchng estmators whch are based on the equatons (3) and (4). Table 3 shows the results whch are based on Cases A-1 to A-3 (for MFI_status ) and Table 4 based on Cases B-1 to B-3 n Table 2 (for MFI_productve ). All the results use bootstrapped standard errors. The columns we are nterested n those labeled as Average Poverty Reducng Effect and t value. Both Table 3 and Table 4 generally confrm that the access to MFIs has a sgnfcant effect on ncrease of the IBR score,.e., reducton of poverty, because the IBR ndcators of households wth access to MFIs s much hgher than those of households wth the same propensty score (estmated by household characterstcs) wthout. There s only one excepton where a polcy effect s postve and not sgnfcant, that s, the case for Nearest Neghbour Matchng appled to Rural Areas n Table 3. (Table 3 & 4 to be nserted around here) Appendx 2 shows some detals of propensty score matchng. It s noted that even after enforcng common support only a few observatons are dropped through the comparson of observatons for those MFI supported and for those not supported. 3 observatons are dropped n Case A-1; none n Case A-2; 3 n Case A-3; 1 observaton s dropped n Case B-1; 3 n Case B-2; 1 n Case B-3. In each case, we frst dentfed the optmal number of blocks whch ensures that the mean propensty score s not dfferent for treated and controls n each block. 11 Then, we test the balancng property of the propensty score for each block and for all covarates of the logt model. It should be noted that the balancng property s satsfed for all covarates for all blocks, whch valdates our choce of the covarates and the results of propensty score matchng. Then, 15

18 we apply Nearest Neghbour Matchng and Kernel Matchng based on our estmates for propensty score. Absolute values of the Average Poverty-Reducng Effects show that the extent to whch the access to MFIs ncreased an IBR ndcator,.e., reduced poverty. Table 3 shows that smply beng able to have access to MFIs has a larger poverty reducng effect n urban areas than n rural areas. That s, t s not suffcent for poor households n rural areas to have access to MFIs to reduce poverty. However, Table 4 ndcates that the Average Poverty-Reducng Effect n rural areas expected from takng loans from MFIs s larger than that n correspondng cases n Table 3 (.e., n Case of Nearest Neghbour Matchng and n Case of Kernel Matchng). These results mply that unless the poor households are able to take loans from MFIs from productve purposes, they cannot substantally escape from poverty, whch s consstent wth Gaha and Nandh s (2005) results based on a survey n Maharashtra. In urban areas, havng loans from MFIs for productve purposes has a poverty-reducng effect that s not much dfferent from smply havng access to MFIs. Whle propensty score matchng s a potentally useful method to control for endogenety, the results have to be nterpreted wth cauton as shown n the recent dscourse between Smth and Todd (2005) and Deheja (2005) partcularly wth matchng based on cross-sectonal data. Frst, unmeasured characterstcs or tme effects cannot be controlled for by cross-sectonal data. Second, bas assocated wth cross-sectonal matchng estmators may be large wthout a good set of covarates or f treated and control households are not strctly comparable, for example, located n dfferent markets (Smth and Todd, 2005). Thus, senstvty tests based on small changes n the specfcaton are mportant (Deheja 2005). We tred dfferent specfcatons by 16

19 ncludng squares of some of the rght-hand-sde varables (e.g. household sze or dependency rato) n the logt model and have got smlar results. V. Conclusons Drawng upon a natonal-level cross-sectonal household data set n Inda n 2001, the present study analyses the effect of Mcro Fnance Insttutons on poverty of households, whch s defned by the ndexed based rankng (IBR) ndcator, a measure reflectng mult-dmensonal aspects of poverty, such as basc needs, capabltes, socal captal or vulnerablty. To take account of sample selecton bas problems arsng from the household access to MFIs, the propensty score matchng s employed to estmate the poverty-reducng effects of the access to MFIs and the loans used for productve purposes. Sgnfcantly postve effects are observed on the IBR ndcator, whch suggests that MFIs play an mportant role n poverty reducton. If we decompose the results nto rural and urban areas, some nterestng results emerged. Frst, for households n rural areas, sgnfcant results are observed only n the case where the access to MFIs s defned as household takng loans from MFIs for productve purposes and not n the case of smply havng access to MFIs. The result mples that montorng the use of loans as well as ncreasng the productvtes s partcularly mportant n helpng the poor escapng from poverty and protectng them from varous shocks. In urban areas, such sgnfcant poverty-reducng effects are observed n both cases. As the large secton of poor households s not only poor bus also vulnerable, a stronger emphass should be placed on mcrofnance schemes as a polcy means of poverty reducton n both urban and rural areas n Inda. 17

20 Table 1 Descrptve Statstcs and Defntons of the Varables Varable Defnton Obs Mean Std. Dev. MFI_status Whether any member n the household has access to MFI (Total) (Urban) (Rural) MFI_productve Whether a hh has taken a loan for productve purposes (Total) (Urban) (Rural) Age Age of household head (Total) (Urban) (Rural) Female Whether a hh head s female (Total) (Urban) (Rural) Educaton Educaton of the household head (0= llterate, 1= completed prmary school (5th), 2= completed hgher (12th)) (Total) (Urban) (Rural) Hhsze Household sze: number of household members (Total) (Urban) (Rural) Dependency Dependency Rato (Rato of hh members under 15 or over 60 to the total) (Total) (Urban) (Rural) Caste_dum Whether a household s scheduled caste or trbe (=0) or not (=1) not (Total) (Urban) (Rural) Urban_dum Whether a household s n urban areas or not (Total) IBR ndcator Indexed Based Rankng of a household's wellbeng (Total) (Urban) (Rural)

21 Table 2 Results of logt Model on the Determnants of Access to Mcrofnance Case A: Dep Varable: whether a household has access to a MFI ( MFI_Status ) Case A-1: Total Case A-2: Urban Case A-3: Rural Coef. Z value Age (1.79) Age_square (-3.24) Female (3.68) Educaton (2.80) Hhsze (6.35) Dependency (-7.44) 1) + Coef. Z value Coef. Z value (0.24) (1.84) (-0.86) (-3.14) (2.02) (4.91) (-4.90) * 2) (3.08) (2.71) (4.57) (-5.88) Caste_dum (0.13) (-1.25) (0.85) Constant (3.61) (2.40) (3.08) No. of Obs Jont Sgnfcance LR Ch 2 (7)= LR Ch 2 (6)=48.87 LR Ch 2 (7)=94.64 Log lkelhood Pseudo R Case B: Dep Varable: whether a household has taken a loan for productve purposes ( MFI_productve ) Case B-1: Total Case B-2: Urban Case B-3: Rural Coef. Z value Coef. Z value Coef. Z value Age (0.72) (0.19) (0.23) Age_square (-1.87) Female (-2.88) Educaton (-3.98) Hhsze (10.28) Dependency (-8.76) Caste_dum (4.91) (-0.64) (-1.56) (1.09) (-2.81) (4.50) (-7.67) (4.23) 2) (1.72) (8.58) (-5.72) (4.45) Constant (-1.58) (-1.14) (-1.01) No. of Obs Jont Sgnfcance LR Ch 2 (7)= LR Ch 2 (6)=90.33 LR Ch 2 (7)= Log lkelhood Pseudo R Notes: 1) = sgnfcant at 1% level. * = sgnfcant at 5% level. + = sgnfcant at 10% level. 2) Educaton s dropped n case of urban areas as there s no varaton n the varable. + 19

22 Table 3: Results of Propensty Score Matchng (based on Case A n Table 2 where dependent varable s whether a household has access to a MFI ): Effects of MFIs n Reducng Poverty (estmaton usng Bootstrapped Standard Errors, 100 Rps.) Whether a household s a clent of any MFI ( MFI_status )or not Households Average Poverty- Households wthout MFIs wth MFIs Reducng Effect S.E. t value Nearest Neghbour Matchng Total (Case A-1) Urban(Case A-2) Rural (Case A-3) Kernel Matchng Total (Case A-1) Urban (Case A-2) Rural (Case A-3) * Notes: 1) = sgnfcant at 1% level. * = sgnfcant at 5% level. + = sgnfcant at 10% level. 20

23 Table 4: Results of Propensty Score Matchng (based on Case B n Table 2 where dependent varable s whether a household has taken a loan from MFI or the group for a productve actvty): Effects of MFIs n Reducng Poverty (estmaton usng Bootstrapped Standard Errors, 100 Rps.) Whether a household has taken a loan from MFI or from the group for a productve actvty (MFI_productve) Households wth Average Poverty- Households wthout MFIs MFIs Reducng Effect S.E. t value Nearest Neghbour Matchng Total (Case B-1) Urban (Case B-2) Rural (Case B-3) Kernel Matchng Total (Case B-1) Urban (Case B-2) Rural (Case B-3) Notes: 1) = sgnfcant at 1% level. * = sgnfcant at 5% level. + = sgnfcant at 10% level. 21

24 Appendx 1: Techncal Detals of Nearest Neghbour Matchng and Kernel Matchng Nearest Neghbour Matchng Nearest Neghbour Matchng s the method to take each treated unt and search for the control unt wth the closest propensty score. Let T be the set of treated unts (.e. households wth access to MFIs) and C be the set of control unts (.e. households wthout access to MFIs), and T W and C W j be the observed outcomes (.e. the IBR ndcator) of treated and control unts. C( ) denotes the set of control unts matched to the treated unts wth an estmated value of the propensty score of p. In Nearest Neghbour Matchng, C( ) = mn j p p j Denotng the number of controls matched wth observaton T by C N and defne the 1 weghts ω = f j C( ) and ω = 0 otherwse. The number of unts n the treated group j C N T s N. Then the formula for a matchng estmator s: j τ = 1 N T T W T ω W j j C( ) C j 1 T = W ωjw T N T T j C( ) C j = 1 N T T W T 1 N T j C ω W j C j (3) where ω = ω. j j Kernel Matchng 22

25 23 Wth Kernel Matchng all treated are matched wth a weghted average of all controls wth weghts that are nversely proportonal to the dstance between the propensty scores of treated and controls (Becker and Ichno, 2002). Then the formula for a matchng estmator s: ( ) ( ) = T C k h p p C j h p p C j T T n k n j G G W W N 1 τ (4) where.) ( G s a kernel functon and n h s a bandwdth parameter. We use a Kernel Matchng Estmator because the results are not subject to a specfc radus or a number of stratfcaton.

26 Appendx 2: Detals of Computaton of Propensty Score Matchng Case A-1 Total Regon for common support Mn. Max. Obs. No. of blocks Balancng Hypothess : Satsfed for all varables for all blocks. Inferor of block of Wth access Wth no access propensty score to MFI to MFI Total Total Case A-2 Urban Regon for common support Mn. Max. Obs. No. of blocks Balancng Hypothess : Satsfed for all varables for all blocks. Inferor of block of Wth access Wth no access propensty score to MFI to MFI Total! #"%$ & & '! ' ( )+* ", * %-." 95"5 :; <.= "> < / % 1032%& 54 "" 687 %-."? < * %-@ AB5499"<" AB54!!9< "<" & & "% 2C" 1';D? 1';D? 24

27 ! #"%$ & & '! ' ( )+* ", * %-."? * / % 1032%& 54 "" 95"5 :; <.= "E < 687 %-." < %-@ AB5499"<" AB54!!9< "<" & & "% 2C" 1';D? 1';D?! #"%$ & & '! ' ( )+* ", * %-."? * / % 1032%& 54 "" 95"5 :; <.= "E < 687 %-." < %-@ AB5499"<" AB54!!9< "<" & & "% 2C" 1';D? 1';D?! #"%$ & & '! ' ( )+* ", * %-." 25

28 ? * / % 1032%& 54 "" 95"5 :; <.= "E < 687 %-." < %-@ AB5499"<" AB54!!9< "<" & & "% 2C" 1';D? 1';D? 26

29 Endnotes: 1 For example, real GDP grew by 6.9 percent n 2004/ The average annual output growth rates n ndustry and servces sectors n the perod are 5.6% and 8.2% respectvely, whle that n agrcultural sector s 2.0% (based on World Bank Data n 2005 taken from 3 See Gaha and Ima (2005). 4 The relatve sze of dosyncratc rsks s larger n Gaha and Ima (2006) than n Lgon (2005) due to the dfference n specfcatons and ways to correct measurement errors. 5 See Snha (2003) for the conceptual framework of IBR ndcator. 6 Ths secton s based on EDA Systems (2002, 2005), SIDBI (2005) and Snha (2003). 7 We dd not use Radus Matchng as the results are senstve to the predetermned radus. 8 An extenson would be to create a bnary varable based on a partcular number of transacton years wth MFIs to capture the effects of transacton hstory on poverty. However, households wth a long hstory of transacton are a sub-set of those wth any transacton at the tme of survey (as we do not have any data on leavng the MFIs n the past), ths s unlkely to change our man conclusons. 9 Detals of the dstrbutons of propensty scores wll be provded on request. 10 The result of probt where the square of age s dropped s smlar to that of Case B See Becker and Ichno (2002) for detals of the computaton procedure. 27

30 References Arun, T.G and Hulme, D. (2003) Balancng Supply and Demand: The Emergng Agenda for Mcro Fnance Insttutons. Journal of Mcro Fnance, 5(2), pp.1-6. Deheja, R. (2005) Practcal propensty score matchng: a reply to Smth and Todd. Journal of Econometrcs 125, pp Deheja, R., and Wahba, S. (2002) Propensty score matchng methods for nonexpermental causal studes. Revew of Economcs and Statstcs 84 (1), pp Becker, S. and Ichno, A. (2002) Estmaton of average treatment effects based on propensty scores. The Stata Journal 2 (4), pp EDA Rural Systems (2002) Impact Assessment of Mcrofnance for SIDBI Foundaton for Mcro Credt (SFMC) Phase 1 Report July March 2002, Gurgaon, Inda, March 2002 ( EDA Rural Systems (2005) Impact Assessment of Mcrofnance for SIDBI Foundaton for Mcro Credt (SFMC) Phase 3 Report. Gurgaon, Inda. ( Foster, M. (2003) Propensty Score Matchng: An llustratve Analyss of Dose Response. Medcal Care, 41(10), pp Gaha, R. and Ima, K. (2004) Vulnerablty, Persstence of Poverty and Shocks-Estmates for Sem-Ard Rural Inda. Oxford Development Studes, 32(2), pp Gaha, R. and Ima, K. (2005) A Revew of the Employment Guarantee Scheme n Inda Economcs Dscusson Paper. EDP-0513, Manchester, Unversty of Manchester Gaha, R. and Ima, K. (2006) Vulnerablty and Poverty n Rural Inda-Estmates for Rural South Inda. Economcs Dscusson Paper, EDP-0602, Manchester, Unversty of Manchester. 28

31 Gaha, R. and Nandh, M. (2005) Mcrofnace, Self-Help Groups, and Empowerment n Maharashtra. A draft, Rome, the Internatonal Fund for Agrcultural Development. Hulme, D. and Mosley, P. (1996) Fnance aganst Poverty (London: Routledge). Khandker, S. R (2003) Mcro-Fnance and Poverty. Polcy Research Workng Paper No. 2945, World Bank, Wasngton D.C.. Lgon, E. (2005) Targetng and Informal Insurance. n S. Dercon (ed.) Insurance Aganst Poverty, Oxford: Oxford Unversty Press. Meyer, R.L. (2002) Track Record of Fnancal Insttutons n Assstng the Poor n Asa, Asan Development Bank Insttute Research Paper 49, ADB Insttute, December, Mosley, P. and Arun, T. (2003) Improvng Access to Rural Fnance n Inda: Supply Sde Constrants. Background paper to the Economc and Sector Work study on access to fnance, South Asa Fnance and Prvate Sector Development Unt, World Bank. Rosenbaum, P. R. and Rubn, D. B. (1983) The Central Role of the Propensty Score n Observatonal Studes for Causal Effects. Bometrca 70(1), pp Scandzzo, P., Gaha, R. and Ima, K. (2005) Opton Values, Swtches and Wages - An Analyss of the Employment Guarantee Scheme n Inda. presented at the conference Socal Protecton for Chronc Poverty Rsk, Needs, and Rghts: Protectng What? How? held by IDPM, Unversty of Manchester, February SIDBI (Small Industres Development Bank of Inda) (2005) Impact Assessment. (web-report) Snha, F (2003) Understandng and Assessng Poverty: Mult-dmensonal Assessment versus standard poverty lnes. Presented at the EDIAIS Conference, New Drectons n Impact 29

32 Assessment for Development: Methods and Practce at Unversty of Manchester, November Smth, J. A., and Todd, P. E. (2005) Does matchng overcome LaLonde s crtque of nonexpermental estmators? Journal of Econometrcs 125, pp