P M M A W o r k i n g p a p e r 2 0 0 7-2 9 Poverty Targeting and Impat of a Governmental Miro-Credit Program in Vietnam Nguyen Viet Cuong Minh Thu Pham Nguyet Pham Minh Deember 2007 IDRC photo: N. MKee Nguyen Viet Cuong (National Eonomis University, Hanoi, Vietnam) _nguyenviet@yahoo.om Minh Thu Pham (Inst. of Labour S. and Soial Aff., Hanoi, Vietnam) mthupham@gmail.om Nguyet Pham Minh (Credent Tehnology Ltd) nguyetphamminh@yahoo.om VU Thieu (National Eonomis University, Hanoi, Vietnam)
vuthieu@fmail.vnn.vn Duong Toan (National Eonomis University, Hanoi, Vietnam) duongkhanhtoan999@yahoo.om
Abstrat It is argued that without ollateral the poor often fae binding borrowing onstraints in the formal redit market. This justifies a miro-redit program, whih is operated by the Vietnam Bank for Soial Poliies to provide the poor with preferential redit. This paper examines poverty targeting and impat of the miro-redit program. It is found that the program is not very pro-poor in terms of targeting. Among the partiipants, the non-poor aount for a larger proportion of loans. The non-poor also tend to reeive larger amounts of redit ompared to the poor. However, the program has positive impat on poverty redution of the partiipants. This positive impat is found for all the three Foster-Greer-Thorbeke poverty measures. Key words: Miro-redit, poverty, poverty targeting, impat evaluation, instrumental variables, fixed-effet model. JEL lassifiation: I32; I38; H43; H81 This work was arried out through a grant from the Poverty and Eonomi Poliy (PEP) Researh Network (www.pep-net.org), whih is finaned by the Australian Aid Ageny (AusAID) (http://www.ausaid.gov.au), the Canadian International Development Ageny (www.adi-ida.g.a), and the International Development Researh Centre (IDRC) (www.idr.a). I would like to thank the other three members of our researh team,-- Prof. Vu Thieu, Pham Minh Thu, Duong Khanh Toan, Pham Minh Nguyet -- for provision of douments, analysis of data, and sharing of their omments on the interim as well as final reports. Comments from Marrit Van den Berg, David Bigman (Wageningen University, the Netherlands), Arrar Abdelkrim, Jean-Yves Dulos, John Cokburn, Habiba Djebbari (PEP), Steve Younger (Cornell University) and two other referees were very helpful in improving the paper. Our thanks also go to Deborah Cobb-Clark and staff at the Soial Poliy Evaluation, Analysis, and Researh Centre (SPEAR), Australian National University for omments given during our seminar there. 3
1. Introdution Although Vietnam has experiened remarkable redution in poverty over the past 10 years, nearly 20 perent of the population still lives below the poverty line (Table 1). It is often argued that miro-redit is an important tool for smoothing onsumption and promoting prodution, espeially for poor households (e.g. Zeller, et al. 1997; Conning and Udry, 2005). However, without ollateral the poor an fae binding onstraints in the redit market. Thus, the Vietnamese government has set up the Vietnam Bank for Soial Poliies (VBSP) to provide the poor with preferential miro-redit sine 2003. The role of miro-redit in improving household welfare is found in many empirial studies. Miro-redit programs that are assessed are implemented in several developing ountries suh as Bangladesh, Pakistan, Thailand, et al. For example, Pitt and Khandker (1998) measured the impat of group-based lending programs in Bangladesh, and found that the programs had positive and statistially signifiant impat on household onsumption. In another paper, Khander (2003) found that miro-finane brings benefits for the poorest, thereby signifiantly reduing poverty in Bangladesh. Signifiant impats of redit on inreases for farmers in Pakistan are also found in Khander and Faruqee (2003). Burgess and Pande (2002) examined the expansion of bank branhes on household welfare, and showed that this expansion dereases poverty and inequality. Zaman (2001) found positive impat of miro-redit provided by the Bangladesh Rural Advanement Committee on poverty and vulnerability redution in Bangladesh. Other suessful stories of the role of miro-redit programs in reduing poverty an be found in a review paper of Morduh and Haley (2002). However, there are several studies that do not find signifiant impat of miro-redit on welfare improvement and poverty redution. For example, Diagne and Zeller (2001) did not find statistially signifiant impat of miro-redit on household inome in Malawi. Morduh (1998) showed that most potential impats of miro-redit from the Grameen bank in Bangladesh were on vulnerability redution instead of poverty redution. Coleman (1999) found only negligible impat of a miro-redit program in Thailand on household welfare. In Vietnam, questions on poverty targeting and impat of the VBSP program remain unanswered so far. Most of the evaluation reports simply desribe the implementation and outputs of the program, suh as how many people reeived redit from the program or how muh apital was put into the program. The Government has spent a huge amount of money to finane the VBSP program. Aording to VBSP (2005), the total outstanding loans for households were 8249 billion VND in 2004. 1 Information on the quantitative assessment of a 1 1 USD = 15 000 VND in 2004 4
program an be of interest for several reasons. Firstly, it is very helpful in determining whether the program should be expanded, terminated, or revised. A program with bad targeting and negligible impat should be onsidered for termination or modifiation. Seondly, the assessment an provide useful information for improving the program. For example, if it is found that only a small proportion of the poor in urban areas reeive redit from the program, the program seletion should be hanged to inrease the effetiveness of targeting in those areas. The main objetive of this paper is to examine how well the VBSP program reahes the poor, and to what extent the program has an impat in terms of household welfare and poverty redution. To measure impat the paper employs two methods, inluding the instrumental variables regression and the fixed-effet panel data with instrumental variables. Data used in the analysis are from Vietnam Household Living Standard Surveys that were onduted in 2002 and 2004. The paper is omposed of five setions. The first setion gives a brief literature review of miro-redit program assessments, while the seond setion introdues the data soures and examines the poverty targeting of the VBSP program. The third setion presents the methodology of impat evaluation. Empirial findings on program impat are presented in the fourth setion, with the fifth setion disussing the onlusion and study reommendations. 2. Poverty Targeting of a VBSP Program 2.1 Data Soures The study relies on data from the two VHLSSs, whih were onduted by the General Statistial Offie of Vietnam (GSO) with tehnial support from the World Bank (WB) in the years 2002 and 2004. The 2002 and 2004 VHLSSs overed 30000 and 9000 households, respetively. 2 The seletion of the samples follows a method of stratified random luster sampling so that the households are representative at the national, rural and urban, and regional levels. It is very interesting that the 2002 and 2004 VHLSSs set up a panel of 4000 households, whih are representative of the whole ountry, and regions of large populations. The surveys olleted information through household and ommunity level questionnaires. Information on households inludes basi demography, employment and labor fore partiipation, eduation, health, inome,, housing, fixed assets and durable goods, partiipation of households in poverty alleviation programs, and espeially 2 In 2002, GSO inreased the sample size to 30000 households so that the data ould be representative for some large provines. However, this large sample survey was very expensive, and the sample size of VHLSS 2004 was redued to 9000 households. 5
information on loans that households had obtained or still owed during the twelve months before the interview. Information on ommune harateristis was olleted from 2960 and 2181 ommunes in the 2002 and 2004 surveys, respetively. Data on ommune harateristis onsists of demography and the general situation of ommunes, general eonomi onditions and aid programs, non-farm employment, agriulture prodution, loal infrastruture and transportation, eduation, health, and soial affairs. Commune data an be linked with household data to assess relationship between harateristis of households and harateristis of ommunes in whih the households are loated. It is unfortunate that the ommune data in the 2004 VHLSS are only available for rural areas. This study fouses on the rural population. The main reason is that several ommune variables are used in regression analysis of the VBSP impat, and there are only data on ommune information for rural areas in the 2004 VHLSS. 2.2 Desription of the VBSP Program The poor often fae shortages of apital and assets. Without ollateral they find it more diffiult to aess redit in formal markets. Table 1 ompares inome, and main assets between the poor and non-poor in Vietnam. It shows that the poor have lower inome and per apita than the non-poor. The domesti and foreign remittanes are very limited for the poor. They also tend to have lower value of fixed and durable assets ompared to the non-poor. The government of Vietnam was aware of this fat, and had onduted poliies to provide the poor with preferential miro-redit. Between 1995 and 2002, the Vietnam Bank for the Poor (VBP) was established under the ontrol of the Bank for Agriulture and Rural Development (BARD) with the purpose of providing poor households with favorable redit. Sine the government has aimed at expanding the redit program for the poor, they losed VBSP and launhed a new bank alled the Vietnam Bank for Soial Poliies (VBSP) beginning 2003. VBSP was independent of BARD and expanded its operations rapidly. The branhes of VBSP are urrently established in all the distrits of Vietnam. The poor an borrow from a lose VBSP branh at low interest rates without ollateral. 6
Table 1: Household harateristis of the poor and non-poor for rural areas in 2004 Household harateristis Poor Non-Poor Mean Std. Err. Mean Std. Err. Inome and (VND thousands) 3 Inome per apita 2226.9 26.8 7100.3 110.6 Expenditure per apita 1599.7 9.7 5405.0 74.7 Foreign remittane 62.8 27.0 1386.0 130.2 Domesti remittane 698.6 41.6 2324.7 88.5 Household asset Value of fixed asset (VND thousands) 7286.0 454.3 31149.0 1572.9 % households having a motorbike 14.1 0.9 58.8 0.6 % households having a olor television 29.4 1.2 77.8 0.5 % household having a telephone 0.1 0.1 27.3 0.7 Housing Living areas (m2) 46.6 0.6 62.7 0.5 % households living in permanent house 4.8 0.6 24.5 0.6 % households living in semi-permanent house 55.3 1.4 59.6 0.7 % households living in temporary house 39.9 1.3 15.9 0.5 Land areas Total area of land (m2) 5614.0 1537.3 30462.2 9534.9 Area of annual rop land (m2) 2512.5 154.6 6397.5 2523.6 Perennial rop land (m2) 1764.9 718.3 2553.3 442.0 Forestry land (m2) 1077.2 877.9 20513.7 7306.4 Area of aquaulture water surfae (m2) 259.3 95.8 997.8 249.4 Soure: Estimation from VHLSS 2004 The VBSP program is designed as a group-based lending sheme. In order to borrow redit from VBSP, a household has to join a redit group in their loality. A redit group should inlude from 5 to 50 members who are loated in the same village. If the number of members in a village is lower than 5, they need to join a group in another village. Eah redit group sets up a management board, whih is responsible for borrowing and redit use of its members. Following are several riteria that a household should meet to beome a member of a redit group: - The household has a long-term residene permit at the loality in whih the redit group is loated; - The household has someone who is able to work (working fore); - The household is lassified as the poor by ommune authority; 4 and - The household has a demand for redit. The redit needs to be used in prodution, or onsumption neessary for subsistene. 5 Total loan size is not more than 7 million 3 1 USD = 15 000 VND in 2004 4 The proedure to lassify a household as poor by the loal authority is rather ompliated. Basially, it depends on the inome poverty line - whih is set by the Ministry of Labor, Invalid, and Soial Affairs - and other speifi riteria set up by eah ommune. 7
VND. A household an borrow several times, but the total outstanding loans annot be larger than 7 million VND. One a member of a redit group, a household an apply for loans with the VBSP. Firstly, they send a letter of intent to their redit group, where the household speifies the amount and purpose of the loan that they intend to take. When reeiving the appliation, the redit group will arrange a meeting of all members to onsider the relevane of the loan. The redit group determines whih household is able to borrow, as well as the amount and terms of eah loan. A list of appliants will be prepared by the redit group and sent to the People s Committee in that ommune. One the list is ratified by the People s Committee, it will be sent to a VBSP branh for final approval. Credit proessing time is quite fast; it often takes from one to four weeks to obtain redit sine households send the borrowing request to their redit groups. It is shown that VBSP s proess of lending and monitoring redit is rather stringent, whih is expeted to ensure high repayment rates. Aording to VBSP (2005), the ratio of overdue outstanding loans to the total outstanding loans is about 2.96 perent in 2005. Among the overdue loans, the amount of loans that borrowers annot return aounts for 59.9 perent. VBSP branhes try to keep their overdue outstanding loans, sine the repayment rate an affet the amount of finaning that a bank branh an reeive. The VBSP at the national level alloates fewer funds to VBSP branhes with overdue outstanding loans. On the other hand, redit groups and the People s Committee are also highly responsible for the repayment of redit group members. They tend to exlude very poor households who might not be able to repay loans (Dufhues, et al. 2002). Non-poor or even better-off households an get loans from VBSP, sine they are expeted to have higher apaity to repay the loans. 2.3 Poverty Targeting of the VBSP Program In this study, a household is lassified as poor if their per apita is below the poverty line whih is set by WB and GSO Please explain these aronyms. The poverty line is equivalent to the level that allows for nutritional needs and some essential non-food onsumption suh as lothing and housing. This poverty line was first estimated in 1993. Poverty lines in the following years are estimated by deflating the 1993 poverty line using the onsumer prie index. 6 Figure 1 presents the poverty rates over the period 1993-2004. 5 Speifially, the loan an be used for the following ativities: prodution, business, and servie provision, whih an generate inome in the future; home repair in ase of serious damage; and eduational ost for primary and seondary shool pupils. 6 Regional prie differenes and monthly prie hanges over the survey period have been taken into aount when the poverty lines were alulated. 8
Figure 1: Poverty rate over the period 1993-2004 (%) 70 66.4 60 58.1 50 45.5 40 37.4 35.6 30 24.9 28.8 25.0 20 19.5 10 9.2 6.6 3.6 0 1993 1998 2002 2004 Urban Rural Total Soure: Estimation of VHLSS in 1993, 1998, 2002, and 2004. The figure shows that the proportion of people with per apita under the poverty line dropped dramatially from 58.1 perent in 1993 to 37.4 perent in 1998. The poverty rate ontinued to derease to 28.9 perent and 19.5 perent in 2002 and 2004, respetively. 7 However the poverty rate remains rather high in rural areas, at 25 perent in 2004. The VBSP mainly targets rural areas, sine around 95 perent of the poor are loated in rural areas. As a result, about 87 perent of the VBSP partiipants in 2004 were rural people. The poverty targeting of the VBSP program is examined in table 2. The left panel of this table investigates how well the program reahes households who are defined as poor by the WB-GSO poverty line. It shows that only 12 perent of the poor households in rural areas borrowed redit from the VBSP in 2004. This means that the overage rate of the program was relatively low: nearly 88 perent of poor households did not use the favorable redit, while the overage rate for the non-poor was 6.9 perent. The poor tended to reeive smaller amounts of redit than the middle inome and the rih. The loan size per a partiipating poor household was VND 3174.6 thousands, whih was lower than the amount of VND 3714.8 thousands that a non-poor household borrowed on average. In addition, the VBSP program had very high leakage rates. Among the borrowing households, poor households aounted for only 32.5 perent of this number. In other words, a large proportion of borrowing households were non-poor. 7 The poor are lassified based on the poverty line onstruted by WB-GSO. The poverty lines in the years 1993, 1998, 2002, and 2004 are equal to 1160, 1790, 1917, and 2077 thousands VND, respetively. 9
The right panel of Table 2 examines how the program targets households who are lassified as poor by ommune authorities. As regulated by the program, only households who are lassified as poor by ommune authorities are eligible for redit borrowing. This shows that the overage of the program is a bit higher, at 17.9 perent. This is beause the ratio of poor households lassified by ommunes is lower than the ratio of poor households lassified by the WB-GSO poverty line. However, the leakage rate is also high for this lassifiation level. 75.9 perent of the program partiipants were found non-poor households. Table 2: Perentage of borrowing households, average redit amount and interest rate, overage and leakage rates of the program for rural areas using poverty lassifiation in 2004 Poor by WB-GSO Poor by ommune authorities Indiators Coverage rate: % borrowing households Amount of borrowed redit (thousands VND) Average of monthly interest rate (%) Leakage rate: distribution of borrowing households (%) Leakage rate: Distribution of borrowed redit amount (%) Poor Non- Poor Total Poor Non- Poor Total 12.0 6.9 8.0 17.9 6.8 8.0 [0.8] [0.4] [0.4] [1.4] [0.4] [0.4] 3174.6 3714.8 3537.0 3199.0 3644.1 3537.0 [117.8] [101.1] [78.2] [143.2] [91.4] [78.2] 0.30 0.28 0.29 0.34 0.27 0.29 [0.02] [0.01] [0.01] [0.02] [0.01] [0.01] 32.9 67.1 100 24.1 75.9 100 [2.1] [2.1] [1.9] [1.9] 29.5 70.5 100 21.8 78.2 100 [2.1] [2.1] [1.9] [1.9] Figures in brakets are standard errors (Standards errors are orreted for sampling weights and luster orrelation). Note: Number of observations used is 6427, from the 2004 VHLSS. Soure: Estimation from VHLSS 2004. Using poverty status of households after program implementation an result in misleading analysis of the program targeting. Households who reeived redit an inrease their inome and and rid themselves of poverty. Thus, table 3 analyses program targeting using poverty status in 2002, i.e., before the program. The estimates of the overage rates of the program do not differ signifiantly from those in table 2. However, the leakage rates are smaller. When the poor were lassified using the WB-GSO poverty line, they aounted for 45.5 perent of the program partiipants. 10
Table 3: Perentage of borrowing households, average redit amount and interest rate, overage and leakage rates of the program for rural areas using poverty lassifiation in 2002 Poor by ommune Poor by WB-GSO authorities Indiators Coverage rate: % borrowing households Amount of borrowed redit (thousands VND) Average of monthly interest rate (%) Leakage rate: distribution of borrowing households (%) Leakage rate: Distribution of redit amount (%) Poor Non- Poor Total Poor Non- Poor Total 13.0 6.4 8.1 17.0 6.8 8.1 [1.1] [0.6] [0.5] [1.9] [0.5] [0.5] 3151.2 3555.2 3371.3 3045.5 3490.8 3371.3 [152.9] [161.1] [115.4] [207.4] [135.2] [115.4] 0.30 0.31 0.31 0.34 0.29 0.31 [0.02] [0.02] [0.02] [0.03] [0.02] [0.02] 45.5 54.5 100 26.8 73.2 100 [3.3] [3.3] [2.8] [2.8] 42.6 57.4 100 24.2 75.8 100 [3.6] [3.6] [3.0] [3.0] Figures in brakets are standard errors (Standards errors are orreted for sampling weights and luster orrelation). Number of households in panel data VHLSS 2002-2004 is 2867. Soure: Estimation from VHLSS 2002 and VHLSS 2004 There are at least two reasons why the VBSP program did not reah the poor households well enough. The first is the differene in poverty definition between the WB- GSO approah and the approah employed by loal ommune authorities. In a ommune, a household is lassified as poor if their inome is below the inome poverty line onstruted by the Ministry of Labor, Invalid and Soial Affairs (MOLISA) and they meet several riteria suh as if they lak food or live in a damaged house. These riteria are set up by eah ommune, and they an be very different from one ommune to another (The poverty lassifiation proedures by ommune authorities are presented in Box A.1 of Appendix). As a result, the poverty lassifiation of ommune authorities is not onsistent aross ommunes and over time. Table 4 presents the distribution of population by the poverty lassifiation of ommune authorities and WB-GSO over the period 2002-2004. It shows that 13.1 perent of rural people were lassified as poor using the ommune approah in 2002, while this figure was 35.5 perent using the WB-GSO approah. Only 9.8 perent of rural people were lassified as poor by both approahes. Also in 2002, 25 perent of rural people were lassified as poor by ommune authorities but were onsidered non-poor aording to the WB-GSO approah. 11
Table 4: Distribution of rural population using the poverty lassifiation of ommune authorities and WB-GSO (in perent) The year 2002 The year 2004 Poor by ommune Poor by ommune Poor by Expenditure using WB-GSO poverty line Poor Non-Poor Total Poor 9.8 [0.4] 3.3 [0.2] 13.1 [0.4] authorities Non- Poor 25.7 [0.5] 61.2 [0.6] 86.9 [0.4] Total 35.5 [0.6] 88.5 [0.6] 100 Poor 7.3 [0.4] 3.5 [0.3] 10.8 [0.5] authorities Non- Poor 17.7 [0.6] 71.5 [0.7] 89.2 [0.5] Total 25.0 [0.7] 75.0 [0.7] Figures in brakets are standard errors (Standards errors are orreted for sampling weights and luster orrelation). Soure: Estimation from VHLSS 2002 and VHLSS 2004 The seond reason why the VBSP program did not effetively reah poor households is mentioned in Dufhues, et al. (2002). Credit groups and ommune heads are relutant to inlude poor households in the list of redit appliants. Non-poor an find it easier to obtain redit, sine they are expeted to be more reliable in using redit effetively and repaying redit. One important issue in examining the effetiveness of redit is the usage of redit. Table 5 tabulates loan size by stating the purpose for loans as reported by respondents. A large proportion of redit was used in prodution and investment. The poor used about 62.5 perent of the VBSO redit amount for prodution apital and apital investment, while this proportion for the non-poor was at 58.9 perent. Credit was also used for dept repayment. However, the poor and non-poor also used 29.2 perent and 33.7 perent, respetively of the redit amount for onsumption. Table 5: Distribution of redit amount by redit usage and poverty status for rural areas in 2004 Poor Non-Poor Total Prodution apital 41.6 51.9 48.9 [4.0] [2.9] [2.4] Capital investment 20.9 7.0 11.1 [3.3] [1.3] [1.5] Dept repayment 8.3 7.4 7.7 [2.3] [1.6] [1.3] Consumption 29.2 33.7 32.3 [3.8] [2.7] [2.2] Total 100.0 100.0 100.0 Figures in brakets are standard errors (Standards errors are orreted for sampling weights and luster orrelation). Soure: Estimation from VHLSS 2004 100 12
3. Methodology to Impat Evaluation 3.1 Parameters of interest The main objetive of program impat evaluation is to assess the extent to whih the program has hanged outomes of subjets. 8 Suppose that there is a program assigned to some people in population P, and denote b D as a binary variable of partiipation in the program of a person, i.e., b D equals 1 if she/he partiipates in the program, and D b equals 0 otherwise. Further, let Y denote the observed value of the outome of interest. This variable an reeive two potential values orresponding to the values of the partiipation variable, b i.e., Y = Y1 if D = 1, and Y = Y0 otherwise. 9 Then the program impat on a person i is defined as: = Y Y. (1) i i1 i0 The most popular parameter of the program impat is Average Treatment Effet on the Treated (ATT) (Hekman, et al., 1999), whih is the expeted impat of the program on the atual partiipants: 10 b b b b ATT( 0,1) = E( D = 1) = E( Y1 Y0 D = 1) = E( Y1 D = 1) E( Y0 D = 1). 11 (2) Sine the size of loans taken by a household an be regarded a ontinuous variable, one an be interested in additional impats of a program when the size of loans hanges by an amount, denoted byδ. Denote D as a ontinuous variable indiating the size of loans that a household borrows. For simpliity, denote Yi ( D ) as potential outome of person i orresponding to the value of variable D. We an measure the hange in program impat due a hange in the amount of redit from d to d + δ : i i i i = ( D = d + δ ) ( D = d ) = Y ( D = d + δ ) Y ( D d ). (3) Sine we annot estimate (5) for eah person, we are interested in its average: [ ( D = d + δ ) ( D = d )] = E[ Y( D = d + δ )] E[ Y( D = d )] E. (4) Expetation in (6) an be written for those who partiipate in the program: E [ ( D = d + δ ) ( D = d ) D > 0 ] = E[ Y( D = d + δ ) Y( D = d ) D > 0]. (5) 8 In the literature of impat evaluation, a broader term treatment instead of program/projet impat is sometimes used to refer to an intervention whose impat is evaluated. 9 Y an be a vetor of outomes, but for simpliity let us onsider a single outome of interest. 10 There are other parameters suh as average treatment effet (ATE), loal average treatment effet, marginal treatment effet, or even effet of non-treatment on non-treated whih measures what impat the program would have on the non-partiipants if they had partiipated in the program, et. 11 In some formulas, the subsript i is dropped for simpliity. 13
We an divide the right-hand side of (7) by δ to obtain a parameter alled the average treatment effet of additional redit amount on the treated: 12 [ ( D = d + δ ) Y( D = d) D > 0] E Y ATT ( d, δ ) =. (6) δ This parameter measures how the average program impat on the treated hanges due to a small hange in the amount of redit. If we onsider [ Y( D ) X,D > 0] E as a real funtion of D, and denote this funtion by fd > o ( D ), the impat parameter an be represented by the derivative of ( D ) fd > o with respet to D. 3.2 Impat evaluation methods The main problem in measuring impat of a miro-redit program is endogeneity of program partiipation. The borrowing of redit an be orrelated with unobserved harateristis of households suh as motivation for higher inome or abilities in business. By failing to ontrol for unobservable fators affeting program partiipation, the program impat estimation is no longer unbiased. Most of the studies on impat evaluation of miroredit programs are aware of the endogeneity problem of program partiipation. Sine experimental designs are diffiult to be implemented for miro-redit programs, quasiexperimental and non-experimental designs are often used in impat evaluation. Examples of evaluation of miro-redit based quasi-experiments are Coleman (1999), and Pitt and Khandker (1998). Popular methods in non-experimental designs inlude instrumental variables (Khander and Faruqee, 2003; Burgess and Pande, 2002), sample seletion (e.g., Zaman, 2001), and models based on panel data (e.g., Khander, 2003; Nguyen and Westbrook, 2006). To measure program impat on household welfare, the paper assumes welfare an be speified as follows: ln( Y ) = α + X β + D γ + ε, (7) i i i i where Y is per apita or per apita inome, X is a vetor of household and regional harateristis, and D is the program variable. The program impat is measured by parameter γ. 12 This an be alled the marginal treatment effet on the treated. However, in some papers, e.g., Hekman and Vytlail (2005), marginal treatment effet is defined as the treatment effet on the persons at the margin, i.e., those who are different between program partiipation and nonpartiipation. 14
It should be noted that when we are interested in the impat of partiipation in the program regardless of the size of the program, we an use D as a binary variable. When we are interested in the impat of additional redit amount on the partiipant, D is the loan size, whih is a ontinuous variable. In the ase of redit programs, the main problem in getting the unbiased estimator of γ is the orrelation between the variables D and ε in equation (7). For the VBSP program, there an be unobserved variables suh as business and prodution skills of households and the prevailing business environment, whih would affet both the outomes and program partiipation. As a result, the problem of endogeneity an happen, and methods that do not deal with this problem an lead to biased estimates of the program impat. This study uses two methods to estimate program impat. 13 The first method is the instrumental variables (IV) regressions. This method requires at least one instrumental variable Z, whih must be orrelated with the D variable but not orrelated with the error tem, ε, given the X variables. If instruments are found, all the oeffiients in (7) an be identified and estimated onsistently using different estimators suh as parametri two-stage least squares, generalized method of moments (GMM), and limited information maximum likelihood (LIML). 14 The seond method is the fixed-effet with IV regression using panel data from VHLSS 2002-2004. Using fixed-effet transformation, we an remove unobserved variables that are time-invariants. Then, the IV regressions are applied to solve the problem of orrelation between the D variable and the remaining time-variant error terms. 4. Impat Measurement to a VBSP program 4.1 Impat of the VBSP Program on Household s Expenditure and Inome This setion presents empirial findings of the VBSP program s impat. The first step is to selet the outome and onditioning variables. A household is expeted to use redit in prodution or onsumption. If the redit is used effetively, their inome and onsumption per apita will inrease. We measure the program s impat on onsumption per apita and inome per apita. One reason for using per apita as an outome is that is a popular welfare indiator with whih we an measure impat of VBSP on poverty redution. 13 We do not use parametri sample seletion models, sine it requires assumption on the joint distribution of errors in the outome and treatment equations. Although there are several nonparametri estimators in sample seletion methods, it is diffiult to write software programs to implement the estimation. 14 Examples of instrumental variables as well as a detailed disussion of instrument variable methods an be seen in eonometris textbooks (Wooldridge, 2001 and Greene, 2003), papers (Baum et. al., 2003, and Staiger and Stok, 1997), or literature on the review of impat evaluation (Moffitt, 1991). 15
Total per apita is olleted using very detailed questionnaires in VHLSS. Total inludes food and non-food. Food inludes purhased food, foodstuff, and self-produed produts of households. Non-food omprises on eduation, healthare, houses and ommodities, power, water supply, and garbage olletion? Total inome figures per apita are also olleted arefully. Household inome an ome from any soure. Total inome inludes inome from agriultural and non-agriultural prodution, salary, wage, pension, sholarship, inome from loan interest and house rental, remittanes, and subsidies. Inome from agriultural prodution omprises rop inome, livestok inome, aquaulture inome, and inome from other agriulture-related ativities. There an be a large number of explanatory variables in outome equations. The household variables inlude demography, household assets, housing, eduation, employment, and health status. The ommune and village variables inlude infrastruture and soioeonomi harateristis. The explanatory variables should not be affeted by the program. It should be noted that data on ommunes and villages are olleted only for rural areas. Summary statistis of explanatory variables in the 2004 VHLSS are presented in table A.1 in Appendix. The first method used to measure program impat is instrumental variables using single ross setion data from the 2004 VHLSS. The key identifiation issue is to find a valid instrument for program partiipation, i.e. redit borrowing. Suh an instrument should be orrelated with the program partiipant and exluded from the outome equation. In this study, two instrumental variables are employed. The first one is the ommune poverty rate whih is based on the poverty lassifiation of ommune authorities. It is obvious that households partiipation is orrelated with riteria of program seletion. One of the seletion riteria is the poverty status lassified by the ommune authorities. A ommune whih has a large number of poor households will have a large number of potential partiipants in the program. However, when there are many appliants for redit borrowing, redit groups and ommune heads tend to sreen the appliant list more arefully, sine they also have responsibility in ensuring the repayment rate of the borrowers. More appliants an be exluded from the borrowing list. As a result, an eligible household who lives in a ommune with a large number of poor households will fae higher ompetition when borrowing from the program. The seond instrumental variable is the distane from a village (where households are loated) to the nearest bank. The 2004 VHLSS olleted just information on the distane from a village to a branh bank. There is no information on whether the losest branh bank 16
is a VBSP one. Although the nearest bank an be any ommerial bank instead of a VBSP branh, the VBSP bank an be loated lose to the nearest bank. Households in a village whih is loser to a VBSP branh bank are more likely to obtain redit from the bank. The ondition of orrelation between the instrumental variables and redit borrowing an be investigated by running a regression of borrowing on the instrumental variables and other explanatory variables. Table A.2 in Appendix reports results of seleted regressions. The seond and third olumns show regressions of program partiipation and size of VBSP loans. Variable ommune poverty rate and distane to the nearest bank are statistially orrelated with the partiipation of households in the VBSP program. As expeted, both the instrumental variables are statistially signifiant and negatively orrelated with program partiipation. Living in an area with many eligible households or far from banks redues the hane of program partiipation. Although the instrumental variables an be statially signifiant in the regressions on the endogenous variables, i.e., the program partiipation and the redit size, they an be weakly orrelated with the endogenous variables. The problem of weak identifiation auses the traditional two-stage least square estimator to not funtion properly, whih leads to unreliability of the statistial inferene about the estimates (Stok and Yogo, 2005). In the study, this test is based on the Cragg-Donald statisti (Cragg and Donald, 1993). The test statisti in per apita and inome equations is equal to 24.74 and 29.99, respetively (Table A.3 in Appendix). As a rule of thumb, if a test is over 10, the instruments would not be weak (Staige and Stok, 1997). However, to examine whether the impat estimates are sensitive to different instrumental variable estimators, the study uses three types of parametri estimators, inluding two-stage least squares, generalized method of moments (GMM), and limited information maximum likelihood (LIML). The ondition of un-orrelation between the instrumental variables and the error term in outome equations annot be tested, sine the error term is unobserved. In this study, there are at least two reasons for the absene of the ommune poverty rate and distane to the nearest bank in the outome equation. Firstly, ommune and village variables that are most important in determining households welfare are often a funtion of infrastrutures and geographi harateristis. Infrastruture variables an inlude road, market and shool, et. Geographi variables an be dummy regional variables, geographi types of loality, and distane to the nearest town, et. Provided these variables are ontrolled for in the outome equation, the instrumental variables would be unorrelated with the unobserved variables in this outome equation. Seondly, empirial findings show that ommunes and villages do not play an important role in households welfare one household variables are ontrolled for. Table A.2 in Appendix shows that only a few variables of villages and ommunes are statistially signifiant in outome regressions. 17
Using the two instruments, we an then perform an over-identifiation test. Table A.3 in Appendix presents the Sargan-Hansen tests for estimators of 2SLS and LIML. Based on this test statisti, we annot rejet the hypothesis on over-identifiation of instrumental variables. In addition, the endogeneity of program partiipation and loan size an be tested using the instruments. Results from Durbin-Wu-Hausman tests show that the hypothesis on the exogeneity of program partiipation and loan size from the program is strongly rejeted. The seond method is fixed-effet with IV regression using panel data from VHLSS 2002-2004. In this method, there is only one instrument, whih is the poverty rate of ommunes. This is beause the 2002 VHLSS did not ollet information on the distane from villages to the nearest bank. Table 6 present the results of impat evaluation for rural areas using the instrumental variable method. In this table, only the estimates of oeffiients of program partiipation and the amount of borrowed redit from instrumental variables regression are presented. 15 The left panel of the table presents estimates from IV the regressions using single-ross setion data of the 2004 VHLSS, while the right panel reports estimates from the fixed-effet with IV regressions using panel data from the 2002 and 2004 VHLSSs. It shows that the estimates of the oeffiient of the loan size are positive and statistially signifiant at the 1 perent and 5 perent levels for log of per apita and log of inome per apita. Program partiipation also has positive and statistially signifiant oeffiients. The estimates do not differ signifiantly aross various IV estimators. The estimates from the fixed-effet with IV methods are also positive, and are statistially signifiant at the 1 perent level. Thus, ompared to the IV regressions, the estimates from the fixed-effet with IV ones have small standard errors. 15 Some of regression results using 2SLS estimators are presented in Table A.1 in the Appendix. 18
Table 6: Program impat on and inome per apita Program variable Loan size (in VND thousands) Program partiipation (dummy variable) IV estimators IV regression per apita inome per apita Fixed-effet with IV regression inome per apita per apita 2SLS 0.00019*** 0.00017** 0.00021*** 0.00029*** [0.00007] [0.00008] [0.00004] [0.00006] GMM 0.00018*** 0.00017** [0.00007] [0.00008] LIML 0.00019*** 0.00017** [0.00007] [0.00008] 2SLS 0.68611*** 0.62768** 0.70177*** 0.96788*** [0.25375] [0.27715] [0.13829] [0.17866] GMM 0.68572*** 0.62714** [0.25249] [0.27705] LIML 0.68612*** 0.62777** [0.25376] [0.27719] Number of observations in regression 6427 6427 5552 5552 Figures in brakets are standard errors (Standards errors are orreted for sampling weights and luster orrelation). 5552 is the total number of observations in the panel data of VHLSS 2002-2004. The number of households in the panel data is 2772. Soure: Estimation from VHLSS 2002 and VHLSS 2004 The differene in impat between the poor and non-poor partiipants is tested by adding the interation between the poverty status and the program variables to the IV regression. We use the poverty status in 2002, sine it is not affeted by the program. It is found that all the estimates of the interation oeffiients are not statistially signifiant (Table A.4 in Appendix). It suggests that the differene in program impat between the poor and non-poor households is not statistially signifiant. 4.2 Impat of the VBSP Program on Poverty Sine the VBSP program has a positive impat on the onsumption per apita, it is expeted that the program an also redue poverty. This study measures poverty by three Foster-Greer-Thorbeke poverty indexes whih an all be alulated using the following formula (Foster, Greer and Thorbeke, 1984): q 1 z yi Pα = n, (8) z i= 1 α where y i is a welfare indiator (onsumption per apita in this paper) for person i, z is the poverty line, n is the number of people in the sample population, q is the number of poor people, and α an be interpreted as a measure of inequality aversion. When α = 0, we have the headount index H whih measures the proportion of people below the poverty line. When α = 1 and α = 2, we have the poverty gap PG whih measures 19
the depth of poverty, and the squared poverty gap P 2 whih measures the severity of poverty, respetively. Impat of the program on poverty of the partiipants is given by: P ( D = 1,Y ) P ( D =,Y ), (9) P = 1 1 α α α 0 where the first term in the left-hand side of (9) is the measure of poverty in the presene of the VBSP program. This term is observed and an be estimated diretly from the sample data. However, the seond term in the left-hand side of (9) is the ounterfatual measure of poverty, i.e., poverty indexes of the redit reipients if they had not reeived the redit. This term is not observed diretly, and it is estimated using predited in the absene of the miro-redit program. Sine the use of instrumental variables produes statistially signifiant results, it is also utilized to estimate ounterfatual. Reall the outome equation as follows: ln( Y ) = α + X β + D γ + ε (10) i i i i Counterfatual in the absene of the program for a partiipant is: ˆ i0 ( ˆ α X ˆ β ) Y = exp +, (11) i However, we do not use this ounterfatual to estimate the poverty indies diretly. Using the ounterfatual to estimate poverty for eah households and then adding these up will lead to biased estimators of poverty indies (Hentshel, et al., 2000). Instead, we employ the idea of small area estimation by Elbers, et al. (2003). Firstly, we estimate equation (10) using the instrumental variables regression. Then, for household i, denote P i as the indiator of poverty for the household. P i is equal to 1 if per apita of the household is below the poverty line, and equal to 0 otherwise. The estimator to predit the expeted poverty of household i if they had not borrowed redit is as follows: ( ˆ α X ˆ β ) ˆ ˆ ln z ln( Yi 0 ) ln z + i P i = Φ = Φ, (12) ˆ σ ˆ σ where Φ is the umulative standard normal funtion; αˆ and βˆ are estimators of α and β, respetively; σˆ is the estimator of the standard deviation of error term ε in the outome equation, Ŷ i0 is predited per apita of household i in the absene of the VBSP redit. 20
Poverty rate for the group of the partiipants is simply the sum of expeted poverty of the partiipants. Thus the estimator of the headount index is simply as follows: N mi Pˆ 0( D = 1,Y0 ) = Pˆ i, (13) M i = 1 where m i is the size of household i; M is the total population of the partiipating group; and N is the number of households in the partiipating group. To estimate the poverty gap index PG, and the poverty severity index P 2, we employ a method proposed by Minot, et al. (2003) to estimate the umulative distribution of the per apita in the absene of the VBSP redit by hanging the poverty line from the lowest per apita to the highest per apita in the sample. The estimated umulative distribution is then used to estimate the poverty indexes PG and P 2 (in the state of no-redit from the program). To estimate standard error of estimates, the paper uses a nonparametri bootstrap tehnique with 200 repliations. Table 7: Program impat on poverty indies Atual No-redit Differene Poverty indies & IV estimators ounterfatual 2SLS Headount ratio 0.3633*** 0.4145*** -0.0512* [0.0234] [0.0238] [0.0270] Poverty gap index 0.0898*** 0.1222*** -0.0324** [0.0077] [0.0142] [0.0134] Poverty severity index 0.0319*** 0.0487*** -0.0168** [0.0035] [0.0082] [0.0078] GMM Headount ratio 0.3633*** 0.4098*** -0.0465 [0.0234] [0.0232] [0.0295] Poverty gap index 0.0898*** 0.1192*** -0.0294** [0.0077] [0.0135] [0.0139] Poverty severity index 0.0319*** 0.0470*** -0.0151** [0.0035] [0.0078] [0.0077] LIML Headount ratio 0.3633*** 0.4128*** -0.0496* [0.0234] [0.0245] [0.0297] Poverty gap index 0.0898*** 0.1219*** -0.0321** [0.0077] [0.0152] [0.0159] Poverty severity index 0.0319*** 0.0487*** -0.0168* [0.0035] [0.0090] [0.0092] ***, **, and * represent statistial signifiane at 1%, 5% and 10%, respetively. Figures in brakets are standard errors. Standard errors are alulated using bootstrap (non-parametri) with 200 repliations and are orreted for sampling weights. Soure: Estimation from VHLSS 2004 Table 7 presents an estimation of the VBSP s impat on poverty of the partiipants. The IV regressions are used to predit the per apita without the program, and 21
the program variable is the size of loans. The IV estimators inlude 2SLS, GMM, and LIML. The three estimators yield rather similar results of estimation of the program s impat on poverty. For example, based on the 2SLS estimator, the impat on redution in poverty rate is estimated at 5.1 perent with a signifiane level of 10 perent. In other words, the VBSP program helps the partiipants redue the poverty rate by 5.1 perentage points. The program also redues the poverty gap index of the partiipants by 0.032 with a statistial signifiane level of 5 perent. The poverty severity index is also redued by 0.017 due to redit from the program, while the statistial signifiane level of this estimate is 5 perent. 5 Conlusions The paper examines the VBSP s poverty targeting and the impat of its preferential redit program for the poor. The program is designed to provide the poor households with redit at low interest rates without ollateral. However, the program s targeting methods leave muh to be desired: only 12 perent of the poor households in rural areas partiipated in the program in 2004. Meanwhile, the program overed 6.4 perent of the non-poor households. The non-poor households aounted for 67.1 perent of the benefiiaries. The poor households also reeived smaller amounts of redit than the non-poor. Thus, in terms of targeting, the program is not very pro-poor. Although, the poor aess the program more proportionally than the non-poor, they aount for a smaller proportion of the program partiipants. 16. One of the main reasons for suh ineffetive targeting an be explained in Dufhues, et al. (2002). Heads of redit groups and ommunes are relutant to verify the poor households in the list of redit appliants beause of their low repayment apaity. The Government and VBSP should therefore employ measures to redue the lending program s leakage rate and inrease its overage rate at the same time, while keeping the program effetive. Further studies on the lending system and the seletion proess should be onduted in order to generate more detailed suggestions for the modifiation. Empirial results from impat evaluation show that the program has positive and statistially signifiant impat on onsumption per apita and inome per apita of the partiipating households. Sine the program has a positive impat on households, it is expeted that the program an ontribute to poverty redution. It is found that the program has positive and statistially signifiant effets on redution of the poverty rate, poverty gap and poverty severity. 16 The poor aess more proportionally than the non-poor, i.e., the program overage for the poor is higher than for the non-poor. 22
Referenes Burgess, Robin and Rohini Pande. 2002. Do Rural Banks Matter? Evidene from the Indian Soial Banking Experiment. Mimeo, Columbia University. Coleman, Brett. 1999. "The Impat of Group Lending in Northeast Thailand." Journal of Development Eonomis 60, no. 1: 105-141. Conning, Jonathan, and Christopher Udry. 2005. Rural Finanial Markets in Developing Countries, Eonomi Growth Center, Yale University. Center Disussion Paper, no. 914. Cragg, John, and Stephen Donald. 1993. Testing Identfiability and Speifiation in Instrumental Variables Models. Eonometri Theory 9, 222-240. Dehejia, Rajeev, and Sadex Wahba. 1998. Propensity Sore Mathing Methods for Non Experimental Causal Studies. NBER Working Paper no. 6829, Cambridge, Mass. Diagne, Aliou, and Manfred Zeller. 2001. Aess to Credit and Its Impat on Welfare in Malawi. Researh report no. 116, IFPRI, Washington D.C. Dufhues, Thomas; Pham Thi My Dung; Ha Thi Hanh; and Gertrud Buhenrieder. 2002. Fuzzy Information Poliy of the Vietnam Bank for the Poor in Lending to and Targeting of the Poor in Northern Vietnam. Paper presented at the International Symposium Sustaining Food Seurity and Managing Natural Resoures in Southeast Asia Challenges for the 21 st Century. January 2002, Chiang Mai, Thailand. Elbers, Chris; Lanjouw Jean; and Lanjouw Peter. 2003. Miro-level Estimation of Poverty and Inequality. Eonometria 71, no.1: 355-364. Foster, James; Joel Greer; and Erik Thorbeke. 1984. A Class of Deomposable Poverty Measures. Eonometria 52, no. 3: 761-766. Hall, Alastair, and Fernanda Peixe. 2000. A Consistent Method for the Seletion of Relevant Instruments. Eonometri Soiety World Congress 2000 Contributed Papers. Hekman Jame and Edward Vytlail. 2005. Strutural Equations, Treatment Effets and Eonometri Poliy Evaluation. NBER Working Paper no. W11259. Hekman, James; Robert Lalonde; and Jeffrey Smith. 1999. "The Eonomis and Eonometris of Ative Labor Market Programs," Handbook of Labor Eonomis 3, Ashenfelter, A. and D. Card, eds., Amsterdam: Elsevier Siene. Hekman, Jame, Hidehiko Ihimura, and Petra Todd. 1997. Mathing as an Eonometri Evaluation Estimators: Evidene from Evaluating a Job Training Programme. Review of Eonomi Studies 64, no. 4: 605-654. Hentshel, Jesko; Jean Lanjouw; Peter Lanjouw; and Javier Poggi. 2000. Combining Census and Survey Data to Trae the Spatial Dimensions of Poverty: A Case Study of Euador. World Bank Eonomi Review 14, no. 1: 147-65 Ihimura, Hidehiko, and Christopher. 2001. Propensity Sore Mathing with Instrumental Variables. Amerian Eonomi Review 91, no. 2: 119-124. Khander, Shahidur. 2003. Miro-finane and Poverty: Evidene using Panel Data from Bangladesh. World Bank Poliy Researh Working Paper no. 2945, January 2003. Khandker, Shahidur, and Rashid Faruqee. 2003. The Impat of Farm Credit in Pakistan. Agriultural Eonomis 28, no. 3: 197-213 Minot, Niolas; Bob Baulh; and Mihael Eppreht. 2003. Poverty and Inequality in Vietnam: Spatial Patterns and Geographi Determinants. Final report of projet Poverty Mapping and Market Aess in Vietnam onduted by IFPRI and IDS. 23