This article was downloaded by: [Flinders University of South Australia] On: 21 March 2013, At: 02:10 Publisher:

Routledge Informa Ltd Registered in England and Wales Registered Number: 1072954 Registered office: Mortimer House, 37-41 Mortimer Street, London W1T 3JH, UK

The Journal of Development Studies

Publication details, including instructions for authors and subscription information: http://www.tandfonline.com/loi/fjds20

Assessing the Effect of Microfinance on Vulnerability and Poverty among Low Income Households
Ranjula Bali Swain & Maria Floro
a b a b

Department of Economics, Uppsala University, Uppsala, Sweden

Department of Economics, American University, Washington, DC, USA Version of record first published: 18 Apr 2012.

To cite this article: Ranjula Bali Swain & Maria Floro (2012): Assessing the Effect of Microfinance on Vulnerability and Poverty among Low Income Households, The Journal of Development Studies, 48:5, 605-618 To link to this article: http://dx.doi.org/10.1080/00220388.2011.615917

PLEASE SCROLL DOWN FOR ARTICLE Full terms and conditions of use: http://www.tandfonline.com/page/terms-andconditions This article may be used for research, teaching, and private study purposes. Any substantial or systematic reproduction, redistribution, reselling, loan, sub-licensing, systematic supply, or distribution in any form to anyone is expressly forbidden. The publisher does not give any warranty express or implied or make any representation that the contents will be complete or accurate or up to date. The accuracy of any instructions, formulae, and drug doses should be independently verified with primary sources. The publisher shall not be liable for any loss, actions, claims, proceedings, demand, or costs or damages whatsoever or howsoever caused arising directly or indirectly in connection with or arising out of the use of this material.

Journal of Development Studies, Vol. 48, No. 5, 605618, May 2012

Assessing the Eect of Micronance on Vulnerability and Poverty among Low Income Households
Downloaded by [Flinders University of South Australia] at 02:10 21 March 2013

RANJULA BALI SWAIN* & MARIA FLORO**

*Department of Economics, Uppsala University, Uppsala, Sweden, **Department of Economics, American University, Washington DC, USA

Final version received May 2011

ABSTRACT We empirically investigate whether participation in the Indian Self Help Group (SHG) micronance programme has helped reduced poverty and household vulnerability using cross-sectional SHG rural household survey data. The potential selection bias is eliminated by propensity score matching to estimate the average treatment on treated eect using nearest neighbour matching and a local linear regression algorithm. We nd that vulnerability in SHG members is not signicantly higher than in non-SHG members, even though the SHG members have a high incidence of poverty. However, vulnerability declines signicantly for those that have been SHG members for more than one year. These results are found to be robust using sensitivity analysis and the Rosenbaum bounds method.

1. Introduction An extensive literature has examined the impact of micronance in alleviating poverty (Morduch, 1999). While several studies have shown a positive impact in reducing poverty, at least ve have challenged this view expounding that the results are more mixed (Morduch, 1999; Amin et al., 1999; Puhazhendi and Badatya, 2002; de Aghion and Morduch, 2006; Karlan, 2007).1 Exploring beyond poverty, this article investigates if micronance reduces household vulnerability. In other words, do micronance programmes reduce the households exposure to future shocks and improve their ability to cope with them? Answering this question is crucial since the goal of poverty alleviation is not just about improving economic welfare via increased incomes and consumption. It is also about devising means for preventing households from falling into poverty and enabling them to meet their survival needs including food security, to make productive investments and to avoid selling their limited resources in times of income or expenditure shocks. Static poverty measures are helpful in assessing the current poverty status of households but tend to ignore poverty dynamics over time.2 Thus even though average household incomes do not fall into poverty levels, their degree of vulnerability or the risk of being poor in the future, can still remain high. The cumulative impact of micronance programmes on the households wellbeing may therefore not be captured by standard poverty
Correspondence Address: Ranjula Bali Swain, Department of Economics, Uppsala University, Box 513, Uppsala, Sweden, 75120. Email: [email protected] An Online Appendix is available for this article which can be accessed via the online version of this journal available at http://dx.doi.org/10.1080/00220388.2011.615917 ISSN 0022-0388 Print/1743-9140 Online/12/050605-14 2012 Taylor & Francis http://dx.doi.org/10.1080/00220388.2011.615917

606 R. Bali Swain & M. Floro measures alone. A limited literature on the impact of micronance on vulnerability provides evidence that micronance tends to strengthen crisis-coping mechanisms, helps diversify incomeearning sources, and enables asset creation. In fact, a few studies suggest that it has a more signicant impact in reducing vulnerability than income-poverty (Hashemi et al., 1996; Morduch, 1999). Our objectives in this article are twofold. First, we estimate two important dimensions of wellbeing namely, poverty and ex-ante vulnerability of households in SHG and non-SHG groups using 2003 rural household survey data. Second, we empirically investigate whether micronance programmes like the Self Help Group (SHG) programme lead to a reduction in vulnerability or not. Vulnerability in our study is dened as a forward-looking, ex-ante measure of the households ability to cope with future shocks and proneness to food insecurity that can undermine the households survival and the development of its members capabilities. The empirical analysis is based on a 2003 household survey data collected on one of the largest micronance programmes in the developing world, the National Bank for Agriculture and Rural Development (NABARD) self-help group (SHG) programme in ten rural districts in India. We estimate several poverty measures as well as an ex-ante vulnerability measure using Chauduri, Jalan and Suryahadi (2002) methodology, which allows for household vulnerability estimation using cross-sectional data. We also take into account any variation in the eect of SHG participation on vulnerability due to dierences in the economic environment and the design of the SHG bank linkage. We correct for potential selection bias in the household sample using propensity score matching to obtain the average treatment on treated eect (impact) on vulnerability. Finally we test the sensitivity of the results to unobservables. Some researchers suggest that the poor are likely to be more vulnerable (Prowse, 2003; Cannon et al., 2003; Feldbru gge and von Braun, 2002). If this is the case, then the SHG members, with a higher proportion of poor households, are likely to be more vulnerable. Controlling for selection bias, our results show that SHG member households are not more vulnerable than non-member households, even though a higher proportion of them are poor. Among the more mature SHG members however, we nd a signicant reduction in vulnerability compared to the non-SHG members. These results are found to be robust using the sensitivity analysis and Rosenbaum bounds method. The article is organised as follows. Section 2 discusses the notion of vulnerability and the conceptual framework used in the estimation of vulnerability. Section 3 explores the role of micronance SHGs in reducing vulnerability. Section 4 provides an overview of the sample data used in our analysis and the methodologies used in addressing potential participation bias, in estimating vulnerability, and in assessing the eect of SHG participation. Section 5 provides the results of the propensity score matching and the resulting poverty and vulnerability estimates for SHG and non-SHG members. The results of sensitivity analyses involving the use of aected treatment on treated (ATT) eect and Rosenbaum bounds methods to test the robustness of the propensity score matching estimates are provided in section 6. Concluding remarks are presented in the nal section. 2. Understanding Vulnerability It should be noted that vulnerability as a notional concept, has been viewed dierently by researchers, thus leading to varied denitions and measures. Some see vulnerability as an aspect which can cause poverty or hinder people from escaping out of poverty (Prowse, 2003: 9). This view that poor people are generally more vulnerable is shared by Cannon and Rowell (2003) and Feldbru gge and von Braun (2002). Some have taken a dierent perspective of vulnerability whereby poverty is viewed as one element, which may contribute to an enhanced vulnerability (Cardona, 2004). Others such as Calvo (2008) treat vulnerability as a dimension of poverty itself and dene it as a threat of suering any form of poverty in the future.3 In Calvo and Dercons (2005) model, vulnerability is seen as a combination of poverty (failure to reach a minimum

Downloaded by [Flinders University of South Australia] at 02:10 21 March 2013

Assessing the Eect of Micronance 607 outcome) and risk (dispersion over states of the world) that translates into a threat of being poor in the next period. This notion of vulnerability builds upon the probability of outcomes failing to reach the minimal standard as well as on the uncertainty about how far households may fall below that threshold. This uncertainty is a source of distress and impinges directly on wellbeing. Chauduri et al. (2002) in their study of Indonesian households, dene vulnerability within the framework of poverty eradication as the ex-ante risk that a household will, if currently non-poor, fall below the poverty line, or if currently poor, will remain in poverty (p. 4). On the other hand, Ligon and Schechter (2003) take a utilitarian approach in dening vulnerability, arguing that it depends not only on the mean of household consumption but also on variation in consumption in the context of a risky environment. The risk faced by the household is decomposed into aggregate and idiosyncratic risk. A growing number of empirical studies have proposed varied measures and proxy indicators of vulnerability as well (Zimmerman and Carter, 2003; Calvo and Dercon, 2005; Glewwe and Hall, 1998; Ligon and Schechter, 2003; Carter and Barrett, 2006; Morduch, 2004). Some make use of household panel data, where available, to analyse the extent of consumption uctuations over time as households experience income uctuations (Morduch, 2004; Kamanou and Morduch, 2005). Other studies examine the impact of various forms of shocks on households consumption (Ligon and Schechter, 2003; Carter et al., 2007), or other aspects of household wellbeing, for instance, health (Dercon and Hoddinott, 2005). While there are eorts to address data issues, empirical analyses of vulnerability remain severely constrained by the paucity of panel data in many developing countries and by limited information on the idiosyncratic and covariate shocks experienced by households (Gu nther and Harttgen, 2009: 12221223). Chauduri et al. (2002) propose a method for estimating vulnerability that can be applied to cross-sectional household surveys such as the 2003 Indian rural household survey, thus avoiding the data problems mentioned. It has been adopted in vulnerability studies including Zhang and Wan (2006), Gu nther and Harttgen (2009) and Imai et al. (2010).4 A discussion of this vulnerability estimation method is presented in section 4. 3. Micronance Self-Help Groups and Household Vulnerability Very few studies have explored the eect of micronance in terms of reducing vulnerability. Evidence on Bangladeshi micronance institutions conclude that micronance access has led to consumption smoothing or a reduction in the variance in consumption by member households across time periods (Khandker, 1998; Morduch, 1999; Zaman, 2000). The Puhazhendi and Badatya (2002) study nds that micronance provides loans for both production and consumption purposes, thereby allowing consumption smoothing and enabling households to mitigate the eects of negative shocks. Building on these studies, we argue in this article that micronance SHG participation can help member households in the face of liquidity constraints and a multitude of risks, thereby reducing their vulnerability. For instance, SHG programmes provide loans to those members who face liquidity constraints in meeting investment needs as well as unexpected consumption expenses. These production and consumption loans help ease the members productivity and earnings and help their households in coping with contingencies and idiosyncratic shocks. The training of members provided by the SHG programme also can enhance their entrepreneurship skills as well as their ability to perceive and process new information, evaluate and adjust to changes, thus increasing both their productivity and self-condence. In addition, SHGs can promote or strengthen social networks that provide mutual support by facilitating the pooling of savings, regular meetings, etc. that help empower their members, especially women. Group meetings are often used to discuss communal issues leading to the improved ability of member households to manage risk and deal with shocks These nonpecuniary eects of SHGs can reduce the vulnerability of the members and by association, that of

Downloaded by [Flinders University of South Australia] at 02:10 21 March 2013

608 R. Bali Swain & M. Floro their households in ways that may not be adequately captured by changes in household earnings alone. While SHGs may help rural households deal with vulnerability to idiosyncratic shocks the protection aorded by them in dealing with covariate shocks such as epidemics, ooding or declining crop prices is likely to be weak (Zimmerman and Carter, 2003; Morduch, 2004; Dercon, 2005).5 An enabling economic environment and the presence of services and infrastructural support such as health centres and ood control systems that reduce exposure to these aggregate shocks can help enhance the eectiveness of SHGs in reducing household vulnerability.

Downloaded by [Flinders University of South Australia] at 02:10 21 March 2013

4. Data and Methodology 4.1. Data Description The NABARD SHG-Bank linkage programme in India is one of the largest and fastest-growing micronance programmes in the developing world. Initiated in 1996, the SHG programme has grown to nance 687,000 SHGs in 20062007 as compared to 198,000 SHGs in 20012002. According to NABARD (2006), about 44,000 branches of 547 banks and 4896 nongovernmental organisations (NGOs) participate in the SHG-Bank linkage programme. These micronance SHGs typically include 10 to 20 (primarily female) members in the village. In the initial months, the group members save and lend among themselves to build group nancial discipline. Once the group demonstrates stability for six months, it receives loans of up to four times the amount it has saved. The bank then disburses the loan and the group decides how to manage the loan. As savings increase through the groups life, the group accesses a larger amount of loans. The SHGs are linked to banks in several ways: SHGs that are formed and nanced by banks (model 1), those formed by NGOs but directly nanced by banks (model 2), and those that are formed by the NGOs and nanced by the banks through the NGOs (model 3). The data used for the empirical analyses in this article was collected in 2003 as part of a larger study that investigates the NABARD SHGBank linkage programme.6 The sample survey was conducted in two representative districts of the following ve states: Orissa, Andhra Pradesh, Tamil Nadu, Uttar Pradesh, and Maharashtra.7 NABARDs choice to expand the SHG programme occurs at the district level without any specic policy to target certain villages (Bali Swain and Varghese, 2009). Thus, within the states, the study selected is sampled at the district level, which is the basic administrative unit, avoiding those districts with over and under exposure of SHGs. The sampling strategy involved random selection of SHG member-households in each district. The control group (non-SHGs) was chosen to reect a comparable socio-economic group to the SHG respondents. These households were selected from villages that were similar to the SHG villages in terms of the level of economic development, socio-cultural factors and infrastructural facilities, but did not have a SHG programme. After rening the data further and dropping those with missing values, we are left with a sample of 840 households. Table 1 shows characteristics of SHG and non-SHG members and their households. In general, SHG members are younger, have higher levels of education, and have less non-land wealth compared to non-SHG respondents. They also have higher food consumption per capita per month and bigger landholding size compared to non-SHG households, although there is large variation in land quality. SHG households live in villages that are closer to public transport and primary health care centres but further away from banks, compared to non-SHG households. Using a subjective indicator based on the survey response as to whether or not their household experienced severe shortage of food and/or cash in the past three years, we nd that 39 per cent of the SHG households have experienced economic diculties, compared to 27 per cent of non-SHG households. The t-test results conrm the signicant dierence between the SHG members and non-members in terms of size of landholdings and their access to market infrastructure and services, as well as incidence of food and/or cash shortages in the past.

Assessing the Eect of Micronance 609

Table 1. Selected characteristics of survey respondents and their households (standard deviation in parentheses) All N Average real food expenditure per capita per month Average age of respondent Proportion with some (in %) Primary education Secondary education Post-secondary education Average number of children Dependency ratio Average number of workers in the household Average number of workers engaged in primary activity Mean size of owned land in 2000 (in acres) Mean value of non-land wealth years ago (in Rupees){ Distance to bank (in km) Distance to health care Distance to market Distance to paved road Distance to bus stop Lack of cash or food in 2000 840 307 (442) 35 (8.41) 18 17 3 1.5 (1.27) 0.66 (0.22) 2.48 (1.24) 2.49 (1.37) 0.85 (1.43) 64,691 (90197) 7.33 3.55 5.39 3.06 3.75 0.38 (6.87) (2.84) (4.02) (3.32) (3.55) (0.49) SHG members 789 308 (453) 35 (8.44) 18 18 3 1.5 (1.27) 0.66 (0.22) 2.46 (1.23) 2.48 (1.37) 0.87 (1.45) 63,708 (86775) 7.48 (7.02) 3.46 (2.78) 5.38 (4.07) 3.03 (3.33) 3.69 (3.59 0.39(0.49) Non-SHG{{ 51 282 (194) 36 (8.08) 24 (0.43) 12 2 1.4 (1.25) 0.62 (0.23) 2.70 (1.40) 2.55 (1.30) 0.48** (1.12) 79,891 (132625) 4.96***(3.16) 4.95*** (3.30) 5.46 (3.16) 3.59 (3.04) 4.71** (2.76) 0.27* (0.45)

Downloaded by [Flinders University of South Australia] at 02:10 21 March 2013

Notes: {Calculated with 2000 as the base year. {{T- test results for equality of means of SHG members and non-SHG members are indicated by *** if signicant at 1 per cent level, ** if signicant at 5 per cent level, and * if signicant at 10 per cent level.

4.2. Propensity Score Matching Method The decision to participate in SHGs depends on the same attributes that determine the vulnerability of the household. Self-selection bias could arise from the potentially unobservable traits of the SHG members. For instance, higher entrepreneurship, ability to recognise opportunity, and other critical aspects make the households more likely to participate in the SHG programme. However, the same characteristics could also aect their vulnerability. A number of studies on micronance have addressed the problem of selection, reverse causality and other biases using dierent approaches. To correct for selection bias created by programme selection, we use the propensity score matching (PSM) method. This technique allows us to identify the programme impact when a random experiment is not implemented, as long as there is counterfactual or control group. In contrast to other regression methods, the PSM does not depend on linearity and has a weaker assumption on the error term. The matching relies on the assumption of conditional independence of potential outcomes and treatment given observables. The data collection method meets the three conditions outlined in Heckman et al. (1997), thus allowing the use of the PSM method. First, the survey questionnaire is the same for participants and non-participants and therefore yields the same outcome measures. Second, both groups come from the same local environment or markets. Third, a rich set of observables for both outcome and participation variables are available for the performance of the PSM method. As with any impact evaluation, the main problem with identifying SHG impact is that the outcome indicator for SHG member households with and without theprogramme is not observed because by denition, all the participants are SHG members in period 1. Since we only have information on the households once they participate in the programme, there is need to identify a

610 R. Bali Swain & M. Floro control group that allows us to infer what would have happened with the SHG participant household if the SHG programme had not been in place. The PSM uses the Propensity Score or the conditional probability of participation to identify a counterfactual group of nonparticipants, given conditional independence. The probability (P(X)) of being selected is rst determined by a logit equation and then this probability (the propensity score) is used to match the households. Y1 is the outcome indicator for the SHG programme participants (T 1), and Y0 is the outcome indicator for the SHG members (T 0), then Equation (1) denotes the mean impact: D EY1 jT 1; PX EY0 jT 0; PX 1

Downloaded by [Flinders University of South Australia] at 02:10 21 March 2013

where the propensity score matching estimator is the mean dierence in the outcomes over common support, weighted by the propensity score distribution of participants. The literature proposes several propensity score matching methods to identify a comparison group.8 Since the probability of two households being exactly matched is close to zero, distance measures are used to match households. Following Smith and Todd (2005), we rst choose the neighbour to neighbour (NN) algorithm (with one person matching). This algorithm is the most straightforward and matches partners according to their propensity score. We further estimate the local linear regression (LLR) method (for bandwidths 1).9 The LLR method uses the weighted average of nearly all individuals in the control group to construct the counterfactual outcome. Bootstrapped standard errors for the LLR procedures are used (Abadie and Imbens, 2007; Heckman et al., 1997). 4.3. Estimating Poverty and Vulnerability We examine the poverty prole of the SHG and non-SHG households using standard measures of poverty such as the headcount ratio, poverty gap ratio and the squared poverty gap or Foster-Greer-Thorbecke (FGT). The head count ratio measures the proportion of population under the poverty line. The poverty gap ratio measures the depth of poverty and is the total amount that is needed to raise the poor from their present incomes to the poverty line as a proportion of the poverty line and averaged over the total population. The squared poverty gap or FGT index takes inequality among the poor into account and captures the severity of poverty. The poverty line used in our study is based on the ocial (consumption-based poverty) line for India, which assumes the minimum subsistence requirement of 2400 calories per capita per day for rural areas. The ocial poverty line estimate is derived from the household consumer expenditure data collected by National Sample Survey Organisation (NSSO) of the Ministry of Statistics and Programme Implementation, every fth year. Since the poverty line estimate is drawn from the 61st round of the NSS which covers period July 2004 to June 2005,10 we adjust the ocial poverty line using the 2003 Consumer Price Index for agricultural workers in rural areas to correspond with the survey period. Hence our estimated 2003 poverty line is Rs 356.3 per capita per month. Next, we estimate the households vulnerability using the Chauduri, Jayan and Suryahadi (2002) approach that allows the estimation of expected consumption and its variance with crosssection data. The Chauduri et al. approach is widely used in several studies on vulnerability (Jha and Dang, 2009; Zhang and Wan, 2006; Imai et al., 2010) and is considered to be one of the best estimators (Ligon and Schechter, 2004).11 It is based on the notion of vulnerability as the probability of being poor and implies accounting for the expected (mean) consumption, as well as the volatility (variance) of its future consumption stream. The stochastic process generating the consumption of the household is dependent on the household characteristics and the error term (with mean zero). It captures the idiosyncratic shocks to consumption that are identically and independently distributed over time for each household. Hence, any unobservable sources of

Assessing the Eect of Micronance 611 persistent or serially correlated shocks or unobserved household specic eects over time on household consumption are ruled out. It also assumes economic stability thereby ruling out the possibility of aggregate shocks. Thus the future consumption shocks are assumed to be idiosyncratic in nature. This does not mean however, that they are identically distributed across households. Furthermore, we assume that the variance of the idiosyncratic factors (shocks) depend upon observable household characteristics. Following the Chauduri et al. (2002) approach, we assume that the vulnerability level of a household h at time t is dened as the probability that the household nds itself to be consumption poor in period t 1. The households consumption level depends on several factors such as wealth, current income, expectation of future income (i.e. lifetime prospects), the uncertainty it faces regarding its future income and its ability to smooth consumption in the face of various income shocks. Each of these, in turn, depend on household characteristics, both observed and unobserved, the socio-economic environment in which the household is situated, and the shocks that contribute to dierential welfare outcomes for households that are otherwise observationally equivalent. Hence, the households vulnerability level in terms of its future food consumption can be expressed as a reduced form for consumption determined by a set of variables Xht: ln cht b0 Xht b1 mht 2

Downloaded by [Flinders University of South Australia] at 02:10 21 March 2013

where ln cht represents log of consumption per capita on adult equivalence scale, Xht represents selected household and community level characteristics, and mht is the unexplained part of household consumption. Since the impact of shocks on household consumption is correlated with the observed characteristics, the variance of the unexplained part of consumption mht is: s2 h F0 F1 Xht oht 3

which implies that the variance of the error term is not equal across households and depends upon Xht. The latter include the respondents educational attainment, household composition, number of workers in the household, and household wealth proxy. We also take into account the environment characteristics such as access to paved roads, markets, health care services, and public transportation. Given data limitations, we cannot identify the particular stochastic process generating b. The expected mean and variance per capita household food consumption are estimated using a simple functional form by Amemiyas (1997) three-step feasible generalised least squares (FGLS).12 Using the obtained b1 and F1 estimates, we estimate the expected log consumption and the variance of log consumption for each household. These serve as vulnerability estimates. To facilitate comparison of the vulnerability distribution among SHG and non-SHG households, we estimate additional measures using dierent thresholds in order to examine the sensitivity of our results as to the choice of vulnerability threshold. The relative vulnerability threshold uses the observed poverty rate in the population, which is approximately equal to the mean vulnerability level within a group in the absence of aggregate shocks (Chauduri et al., 2002). Thus, vulnerability levels above the observed poverty rate threshold imply that the households risk of poverty is greater than the average risk in the population, thus making it more vulnerable. We use the ocial rural poverty rate by the Planning Commission of India as the rst vulnerability threshold.13 Another vulnerability threshold is 0.50. Households with vulnerability levels between observed poverty rates and 0.50 threshold are termed relatively vulnerable whereas those above 0.50 are considered highly vulnerable. Finally, the vulnerability to poverty ratio, measures the fraction of the vulnerable population to the fraction that is poor. The higher the vulnerability to poverty ratio the more spread is the distribution of vulnerability. Whereas a lower vulnerability to poverty ratio implies greater concentration of vulnerability among a few households. Admittedly, there is some arbitrariness involved in the selection of the vulnerability thresholds

612 R. Bali Swain & M. Floro so a comparison of the vulnerability estimates using additional vulnerability thresholds shows the sensitivity of the results to the choice of vulnerability threshold. 5. Empirical Analysis This section presents the logistic and the propensity score results of matching. This is followed by a discussion on the poverty and the estimated vulnerability measures of SHG and non-SHG member households. We then present the estimated average treatment on treated (ATT) eect of SHG participation using dierent matching algorithms that take potential selection bias into account. The robustness of our results is then checked for sensitivity to unobservables.

Downloaded by [Flinders University of South Australia] at 02:10 21 March 2013

5.1. Propensity Score Matching We correct for potential selection bias using PSM method by rst estimating a parsimonious logistic equation in order to determine the probability of participating in the SHG programme.14 The variables that likely aect both the participation in SHG and the outcome variable (real food expenditure per capita per month) were chosen and these include age, age squared, sex, education dummies, lack of cash or food three years ago, owned land three years ago, distance from bank, health care centre, marketplace, and paved road.15 We obtained very similar results with both neighbour to neighbour algorithm (with one person matching) and log linear regression method (for bandwidth 1). Table A1 in the Online Appendix shows the propensity score estimation using logistic regression. It indicates that landholding size in 2000, incidence of money or food shortage (in 2000), and distance from the bank and market aect the probability of participating in the SHG. Other variables such as age, gender and education level of the respondent do not signicantly explain SHG participation. Using the derived propensity scores, we drop those SHG respondents with probabilities that cannot be matched to the propensity scores of the control group, leaving us with a sample of 742 households comprised of 691 SHG and 51 non-SHG (control group) households. Of the 691 SHG households, 532 have been members for more than one year (referred to as mature SHG members) and 159 belong to newly formed groups. Only the households on the common support are retained to assure comparability. Prior to matching, the estimated mean propensity scores (standard error) for SHG members and non-SHG member were 0.94 (0.05) and 0.89 (0.06) respectively. Figure A1 in the Online Appendix provides the histograms of the estimated propensity scores for the two groups. After the matching, there was a negligible dierence in the mean propensity scores of the two groups (0.93 (0.04) for SHG members and 0.89 (0.06) for non-SHG members). 5.2. Poverty and Vulnerability Prole for SHG and non-SHG members We construct a poverty prole of the SHGs (treatment group) and the non-SHG member (control group) in 2003 using standard measures such as the headcount index, poverty gap index and the squared poverty gap index.16 Table 2 presents the poverty prole of the SHG member and non-member households using standard poverty measures.17 Our results show that a higher proportion of the SHG members are poor (72.5% as compared to 60.8% for the non-members) although the depth of poverty is about the same between SHG and non-SHG households. Their aggregate poverty gap per household is Rs.123 compared to Rs 118 among non-SHGs. The FGT index shows that there is slightly greater inequality among the non-SHG poor (0.24) compared to the SHG poor (0.22). Following Chauduri et al. (2002) the vulnerability estimates are obtained from the FGLS estimates and are presented in Table A2 in the Online Appendix.18 The mean vulnerability level within the SHG member-household group is much lower (0.45) and statistically signicant as compared to the SHG non-members (0.62). This implies that participation in SHGs may reduce the vulnerability of the households.

Assessing the Eect of Micronance 613

Table 2. Poverty and vulnerability estimates for SHG members and non members{ (Standard deviation in parentheses) SHG members All Households N 691 Poverty Prole for SHG members and non-members Headcount ratio (%) 72.5 Aggregate poverty gap per observation 123 Poverty gap ratio (%) 35 Foster-Greer-Thorbecke (sqd poverty gap) 0.22 Vulnerability Prole for SHG members and non-members Mean 0.45 (0.39) Fraction vulnerability 0.55 Fraction relatively vulnerable 0.08 Fraction highly vulnerable 0.47 Vulnerability to poverty ratio 0.75 Non-SHG members{{ 51 60.8 118 34 0.24 0.62*** (0.39) 0.72** 0.03 0.69** 1.18

Downloaded by [Flinders University of South Australia] at 02:10 21 March 2013

Notes:{The vulnerability estimates are based on the Chauduri et al. (2002) method. {{T-test results for equality of means and proportion. ***, ** and * indicate signicance at 10 per cent, 5 per cent and 1 per cent levels respectively.

We also examine the mean vulnerability and sensitivity of the vulnerability estimate to the choice of a threshold. We use three dierent vulnerability thresholds in our study namely: (a) the observed poverty rate; (b) the vulnerability threshold of lying above the observed poverty rate but with a 50 per cent probability of falling into poverty at least once in the next year; and (c) the highly vulnerable lying above the vulnerability threshold of 0.5 for a one-year time period. We also report the ratio of the proportion of households that are vulnerable to the proportion that are poor. This is an indication of how dispersed vulnerability is in the population. The fraction of the population which is vulnerable with respect to these three thresholds is given in Table 2. Even though a higher proportion of SHG members are poor, they are relatively less vulnerable (0.55) as compared to the non-SHG (0.72). Not only are the non-SHG members more vulnerable, a larger proportion of them (0.69) are highly vulnerable. The non-members also have a higher vulnerability to poverty ratio (1.18) with a greater dispersion in incidence of vulnerability. We further examine the subset of SHG participants that have been members for more than one year. Their poverty and vulnerability prole is very similar to that of the SHG members (see Table A3 in the Online Appendix). The above results indicate that there is a large proportion of currently poor SHG members, whose vulnerability level is low enough for them to be classied as non-vulnerable. This reects the stochastic nature of the relationship between poverty and vulnerability. While poverty and vulnerability are related concepts, the characteristics of those observed to be poor at any given point in time may dier from the characteristics of those who are vulnerable to poverty. 5.3. Impact on Vulnerability Controlling for Selection Bias We now estimate the impact on our outcome variables taking the selection bias from participation into account. Heckman et al. (1997) suggest that in small samples the choice of the matching algorithm can be important, due to trade-os between bias and variances. Thus, Caliendo and Kopeinig (2008) suggest that multiple algorithms should be tried and if they give similar results, the choice may be unimportant. Using two dierent algorithms for propensity score matching to identify the comparison group, we estimate the ATT. Nearest Neighbour matching algorithm (NN) is the more intuitive of the two as it matches each treated observation to a control observation with the closest propensity score. We also employ the local linear regression (LLR) algorithm one to one person

614 R. Bali Swain & M. Floro matching (bandwidth 1), which is a generalised version of kernel matching that allows faster convergence at the boundary points.19 Table 3 presents the Average Treatment on Treated estimates (ATT) of SHG participation impact on vulnerability and average food expenditure per capita per month. The magnitude of the ATT estimates in Table 3, measures the impact of SHG participation on the outcome variables (vulnerability and food expenditure), controlling for the selection bias. Table 3, column 1 shows that the ATT point estimates (both NN and LLR) are positive but statistically insignicant for vulnerability. This indicates that after accounting for selection bias the SHG members are neither more nor less vulnerable as compared to the non-members.20 However, the SHG participants that have been members for more than a year, show a signicantly lower level of vulnerability. This suggests that the impact of micronance on vulnerability takes a longer time. By design, the SHG-Bank linkage programme provides credit to those groups that have demonstrated nancial maturity and stability during the rst six months of their existence. Thus, the more mature (older than one year) groups are credit linked and have the possibility to use micronance for reducing vulnerability whereas the newly formed SHGs do not. SHG participation on the other hand does lead to an increase in its average food expenditure per capita per month compared to that of non-SHGs using the LLR algorithm method (Table 3, column 2). A likely reason for this might be due to the provision of SHG loans that may be used for any purpose (including consumption) and thus help the households cope with economic shocks. Taking the subset of the more mature SHGs however, the results do not show any signicant increase in average food expenditure. Our results show that even though the current poverty status of SHG member households has a very high proportion of poor with a higher aggregate poverty gap, their propensity to become poor in the next period (vulnerability) is not higher. The more mature SHG participants, however, have a signicantly lower level of vulnerability. 6. Sensitivity Analysis Robustness of Results The propensity score matching hinges on the conditional independence or unconfoundedness assumption (CIA) and unobserved variables that aect the participation and the outcome variable simultaneously, that may lead to a hidden bias due to which the matching estimators may not be robust. It is not possible to directly reject the unconfoundedness assumption however. Heckman and Hotz (1989) and Rosenbaum (1987) have developed indirect ways of assessing this assumption. These methods rely on estimating a causal eect that is known to be
Table 3. Average treatment on treated estimates of SHG participation impact on vulnerability and average food expenditure per capita per month Matching algorithm All SHG members 1 NN LLR (bw 1) Mature SHG members 1NN LLR (bw1) (1) Vulnerability 0.09 (1.19) 0.11 (1.54) 70.15** (0.73) 70.11* (0.61) (2) Av. food exp. per capita per month 29.04 (0.61) 68.35* (1.89) 39.33 (42.31) 66.80 (42.55)

Downloaded by [Flinders University of South Australia] at 02:10 21 March 2013

Notes: ** Signicant at the 5 per cent level. * Signicant at the 10 per cent level. NN neighbour to neighbour, t-stats in parentheses. LLR local linear regression, p-values in parentheses standard errors created by bootstrap replications of 200. Covariates of regression same as in Table A1 in the Online Appendix.

Assessing the Eect of Micronance 615

Table 4. Simulation-based sensitivity analysis for matching estimators{ average treatment on treated eect (ATT) estimation on vulnerability with simulated confounder general multiple-imputation standard errors{{ Confounder For all SHGs Age Education For mature SHGs Education
{

(1) ATT 0.13 0.14 70.17

(2) Standard Error 0.01 0.01 0.008

(3) Outcome eect 9.01 5.2 6.830

(4) Selection eect 3.9 1.1 1.009

Downloaded by [Flinders University of South Australia] at 02:10 21 March 2013

Notes: Based on the sensitivity analysis with kernel matching algorithm with between-imputation standard error. The binary transformation of the outcome is along the 25 centile. {{Age variable ( 1 if age is less than 26 years; and 0 otherwise) and education ( 1 if no education; and zero otherwise).

equal to zero. If the test suggests that this causal eect diers from zero, the unconfoundedness assumption is considered less plausible (Imbens, 2004). Building on Rosenbaum and Rubin (1983) and Rosenbaum (1987), Ichino et al. (2007) propose a sensitivity analysis that we adopt in this article. They suggest that if the CIA is not satised given observables but is satised if one could observe an additional binary variable (confounder), then this potential confounder could be simulated in the data and used as an additional covariate in combination with the preferred matching estimator. The comparison of the estimates obtained with and without matching on the simulated confounder shows to what extent the baseline results are robust to specic sources of failure of the CIA, since the distribution of the simulated variable can be constructed to capture dierent hypotheses on the nature of potential confounding factors. To check the robustness of our ATT estimates, we use two covariates to simulate the confounder namely: young (respondents under the age of 26 years) and illiterate (with no education). These covariates are selected in order to capture the eect of unobservables like ability, entrepreneurial skills, experience and risk aversion etc., which may have an impact on the member participation in the SHG programme and on the vulnerability of the household. If the ATT estimates change dramatically with respect to the confounders, then it would imply that our results are not robust. We employ the Kernel matching algorithm with between-imputation standard error, in order to use only the variability of the simulated ATT across iterations. Since our outcome variable is continuous, the confounder is simulated on the basis of the binary transformation of the outcome along the 25th centile. The results of these two confounders21 are presented in Table 4. For both the young and no education confounders the simulated ATT estimates are very close to the baseline estimate. The outcome and selection eect on vulnerability is positive but not very large. The results indicate a robustness of the matching estimates. We further test the robustness of our results using Rosenbaums (2002) bounding approach and nd our results to be robust (see Table A4 in the Online Appendix, with discussion). 7. Concluding Remarks This article explores an important dimension of household welfare that conventional measures of poverty do not address, namely vulnerability. We examine the likely eect of Self-Help micronance groups (SHG) on the vulnerability of participating member households using an Indian household sample survey data from 2003. We argue that a households ability to mitigate risk and cope with shocks is enhanced through SHG participation by increasing household earnings through provision of micronance and training, aiding the households in the face of shocks by providing consumption loans, and enhancing their resilience by strengthening social support and improving womens empowerment.

616 R. Bali Swain & M. Floro We use propensity score matching to extricate the potential selection bias that may arise due to unobservable attributes. Additionally, we empirically examine the current poverty status of households in SHG and non-SHG groups using several poverty measures and then make inferences about whether or not these households are currently vulnerable to future poverty using the Chauduri et al. approach. After matching the treated and comparison groups on the basis of their propensity scores, we estimate the average treatment on treated eect using nearest a neighbour matching algorithm and local linear regression. The robustness is checked with help of sensitivity analysis and Rosenbaum bounds. Our main empirical results show that after we account for the selection bias, even though SHG-member households are found to be poorer than the non-SHG member (control group) households, they are not more vulnerable. Vulnerability is signicantly lower for the more mature households as compared to the non-SHG members These results are found to be robust using the sensitivity analysis and Rosenbaum bounds method. The SHGBank linkage programme is a joint liability micronance programme where the loan may be used for any purpose, be it production or consumption. Micronance in this case provides an additional resource for consumption smoothing thus reducing the variability in food consumption levels and hence vulnerability. Finally, micronance SHG can strengthen mutual support networks that help reduce the vulnerability of members and that of their households in ways that may not be adequately captured by the change in household earnings. Notes
1. The dierences in the empirical ndings arise from varying measures of poverty, dierent country contexts and types of micronance organisations being analysed, use of dierent theoretical models, survey designs and econometric techniques, and/or dierent time periods covered by the studies. 2. See Glewwe and Hall (1998); Calvo and Dercon (2005); Carter and Ikegami (2007); Ligon and Schechter (2002); Dercon and Krishnan (2000); Dercon (2005). 3. This concept is based on the notion that the future is uncertain, and the possibility of failing to reach some standard of minimal achievement in any well-being dimension is at least a disturbing background noise for some, and an everpresent, oppressing source of stress and dismay for many others (Calvo, 2008: 1011). 4. Chauduri et al. (2002) measure of vulnerability is an unpublished working article that has been adopted in several studies. Zhang and Wan (2006) explores the eect of livelihood diversication and education on household vulnerability in rural Chinese households. Gu nther and Harttgen (2009) examine the impact of idiosyncratic and covariate shocks in rural and urban households in Madagascar while the study by Imai et al. (2010) analyses the impact of taxation policies on household welfare in China. We would like to thank the reviewer of this article for bringing some of these studies to our attention. 5. Rural livelihoods in developing countries like India often exhibit high correlations between risks faced by households in the same village or area. Hence, when farm prices decline, or there is a drought or ood in the area, all households are adversely aected simultaneously. Idiosyncratic shocks are, by denition, uncorrelated across households in a given community and therefore can be mutually insured within communities. 6. The process involved discussion with statisticians, economists and practitioners at the stage of sampling design, preparing pre-coded questionnaires, translation and pilot testing with at least 20 households in each of the ve states (100 households in total). The questionnaires were then revised, printed and the data collected by local surveyors that were trained and supervised by the supervisors. The standard checks were applied both on the eld and during the data punching process. 7. These districts (in parentheses) are Orissa (Koraput and Rayagada), Andhra Pradesh (Medak and Rangareddy), Tamil Nadu (Dharmapuri and Villupuram), Uttar Pradesh (Allahabad and Rae Bareli), and Maharashtra (Gadchiroli and Chandrapur). 8. See Townsend (1995); Dercon (2005); Zimmerman and Carter (2003); and Morduch (2004). 9. Bandwidths are smoothing parameters, which control the degree of smoothing for tting the local linear regression. 10. See Poverty Estimates for 20042005, Government of India, Press Information Bureau, March 2007. 11. In a comparative study of various vulnerability estimation strategies, Ligon and Schechter (2004) nd that when the environment is stationary and consumption expenditures are measured without error, then the estimator proposed by Chauduri et al. is the best estimator of vulnerability. 12. For details on the statistical estimation refer to Chauduri et al. (2002). 13. Planning Commission estimates, as accessed on 22 September 2010 at http://www.planningcommission.gov.in/data/ datatable/Data0910/tab%2019.pdf. 14. Using saturated logit models as opposed to simple ones is debatable, as the purpose of logit equation is not only to predict SHG participation (as in selection models) but also for covariate balancing. 15. The variables were chosen through hit and miss method while keeping in mind the balance.

Downloaded by [Flinders University of South Australia] at 02:10 21 March 2013

Assessing the Eect of Micronance 617

16. The poverty gap is the average (over all individuals) gap between poor peoples living standards and the poverty line. It indicates the average extent to which individuals fall below the poverty line (if they do). It thus measures how much would have to be transferred to the poor to bring their income (or consumption) up to the poverty line. The poverty gap however does not capture the dierences in the severity of poverty among the poor and ignores inequality among the poor. To account for the inequality amongt the poor we calculate the squared poverty gap index which is dened as the average of the square relative poverty gap of the poor. The squared poverty gap index (Foster-GreerThorbecke Index) is a weighted sum of poverty gaps (as a proportion of the poverty line), where the weights are the proportionate poverty gaps themselves. Pq zyi a Pa 1 The measures are dened for a  0, where a is a measure of the sensitivity of the index to poverty. i 1 z n When a 0, we have the headcount index (the proportion of the population for whom income (or other measures of living standard) is less than the poverty line), a 1 is the poverty gap index and a 2 is the squared poverty gap index. 17. The poverty and vulnerability prole for the SHG and non-SHG member households is presented here for the sample on common support. Imposing the common support condition in the estimation of the propensity score may improve the quality of the matches used to estimate ATT (Becker and Ichino, 2002). 18. The three step feasible generalised least squares (FGLS) results are presented in Table A2 in the Online Appendix. The results show that SHG membership leads to a statistically signicant increment in the consumption. The coecients of the control variables have the expected signs. 19. We also employed NN (bandwidth 10) and LLR (bandwidth 4), both of which gave very similar results to those in Table 3. 20. The results in Table 2 showing that the SHG members have lower vulnerability as compared to the non-SHGs; do not account for the selection bias and are hence biased. 21. Both these confounders are dangerous confounders, since both the outcome and the selection eect are positive.

Downloaded by [Flinders University of South Australia] at 02:10 21 March 2013

References
Abadie, A. and Imbens, G. (2007) Bias corrected matching estimators for average treatment eects. Working Paper, Department of Economics, Harvard University. Amemiya, T. (1997) The maximum likelihood estimator and the non-linear three stage least squares estimator in the general nonlinear simultaneous equation model, Econometrica, 45, pp. 955968. Amin, S., Rai, A.S. and Topa, G. (1999) Does microcredit reach the poor and vulnerable? Evidence from northern Bangladesh.Working Paper 28, Centre for International Development, Harvard University. Bali Swain, R. and Varghese, A. (2009) Does Self Help Group participation lead to asset creation?, World Development, 37(10), pp. 16741682. Becker, S. and Ichino, A. (2002) Estimation of average treatment eects based on propensity score. The Stata Journal, 2, pp. 358377. Caliendo, M. and Kopeinig, S. (2008) Some practical guidance for the implementation of propensity score matching. Journal of Economic Surveys, 22, pp. 3172. Calvo, C. (2008) Vulnerability to multidimensional poverty: Peru, 19982002. World Development, 36(6), pp. 10111020. Calvo, C. and Dercon, S. (2005) Measuring Individual Vulnerability. Department of Economics Working Paper Series, 229 (Oxford: Oxford University). Cannon, T.T and Rowell, J. (2003) Social Vulnerability, Sustainable Livelihoods and Disasters. Report to DFID. Conict and Humanitarian Assistance Department (CHAD) and Sustainable Livelihoods Oce, London. Cardona, O.D. (2004) The need for rethinking the concepts of vulnerability and risk from a holistic perspective: a necessary review and criticism for eective risk management, in: G. Banko, G. Frerks, and D. Hilhorst (eds) Mapping Vulnerability: Disasters, Development and People (London: Earthscan Publications), pp. 3751. Carter, M. and Barrett, C. (2006) The economics of poverty traps and persistent poverty: an asset-based approach. Journal of Development Studies, 42 (2), pp. 178199. Carter, M. and Ikegami, M. (2007) Looking forward: theory-based measures of chronic poverty and vulnerability. CPRS Working Paper No. 94, University of Wisconsin, Madison. Carter, M., Little, P., Mogues, T. and Negatu, W. (2007) Poverty traps and the long term consequences of natural disasters in Ethiopia and Honduras, World Development, 35(5), pp. 835856. Chauduri, S., Jalan, J. and Suryahadi, A. (2002) Assessing household vulnerability to poverty from cross sectional data: a methodology and estimates from Indonesia. Discussion Paper 0102-52, Department of Economics, Columbia University. De Aghion, B.A. and Morduch, J. (2006) The Economics of Micronance (Cambridge, MA: MIT Press). Dercon, S. (2005) Vulnerability: a micro perspective. Mimeo (Oxford: Oxford University). Dercon, S. and Krishnan, P. (2000) Vulnerability, seasonality and poverty in Ethiopia. Journal of Development Studies, 36(6), pp. 2553. Dercon, S. and Hoddinott, J. (2005) Health, shocks and poverty persistence in: S. Dercon (ed.), Insurance Against Poverty (Oxford: Oxford University Press), pp. 124136. Feldbru gge, T. and von Braun, J. (2002) Is the world becoming a more risky place? Trends in disasters and vulnerability to them. Discussion Paper on Development Policy No. 46. Centre for Development Research, Bonn.

618 R. Bali Swain & M. Floro

Glewwe, P. and Hall, G. (1998) Are some groups more vulnerable to macroeconomic shocks? Hypothesis tests based on panel data from Peru. Journal of Development Economics, 56(1), pp. 181206. Gu nther, I. and Harttgen, K. (2009) Estimating households vulnerability to idiosyncratic and covariate shocks: a novel method applied in Madagascar. World Development, 37(7), pp. 12221234. Hashemi, S.M., Schuler, S.R. and Riley, A.P. (1996) Rural credit programs and womens empowerment in Bangladesh. World Development, 24(4), pp. 635653. Heckman, J. and Hotz, J. (1989) Alternative methods for evaluating the impact of training programmes (with discussion). Journal of the American Statistical Association, 84(804), pp. 862874. Heckman, J., Ichimura, H. and Todd, P. (1997) Matching as an econometric evaluation eestimator: evidence from evaluating a job training programme. Review of Economic Studies, 64, pp. 605654. Heitzmann, K., Canagarajah, R.S. and Siegel, P.B. (2002) Guidelines for assessing the sources of risk and vulnerability. Social Protection Discussion Paper Series. Social Protection Unit, The World Bank. Washington DC. Ichino, A., Mealli, F. and Nannicini, T. (2007) From temporary help jobs to permanent employment: what can we learn from matching estimators and their sensitivity? Journal of Applied Econometrics, 23(3), pp. 305327. Imai, K.S., Wang, X. and Kang, W. (2010) Poverty and vulnerability in rural China: eects of taxation, Journal of Chinese Economic and Business Studies, 8(4), pp. 399425. Imbens, G. (2004) Non parametric estimation of average treatment eects under exogeneity: a review. Review of Economics and Statistics, 86(1), pp. 429. Jha, R. and Dang, T. (2009) Vulnerability to poverty in select Central Asian countries. European Journal of Comparative Economics, 6(1), pp. 1750. Kamanou, G. and Morduch, J. (2005) Measuring vunerability to poverty, in: S. Dercon (ed.), Insurance Against Poverty (Oxford: Oxford University Press), chapter 8. Karlan, D. (2007) Impact evaluation for micronance: review of methodological issues. Poverty Reduction and Economic Management (PREM) Doing Impact Evaluation Discussion Paper 7, November, World Bank, Washington DC. Khandker, S.R. (1998) Fighting Poverty with Microcredit: Experience in Bangladesh (New York: Oxford University Press). Ligon, E. and Schechter, L. (2002) Measuring vulnerability: The directors cut. UN/WIDER Working Paper. Ligon, E. and Schechter, L. (2003) Measuring vulnerability. The Economic Journal, 113(486), pp. 95102. Ligon, E. and Schechter, L. (2004) Evaluating dierent approaches to estimating vulnerability. Social Protection Discussion Paper 0201, World Bank, Washington D.C. McCulloch, N. and Baulch, B. (2000) Simulating the impact of policy upon chronic and transitory poverty in rural Pakistan. Journal of Development Studies, 36(6), pp. 100130. Morduch, J. (1999) The micronance promise. Journal of Economic Literature, 37, pp. 15691614. Morduch, J. (2004) Consumption smoothing across space: testing theories of risk-sharing in the ICRISAT study region of South India, in: S. Dercon (ed.), Insurance against Poverty (Oxford: Oxford University Press). National Bank of Agriculture and Rural Development (1992) Guidelines for the Pilot Project for linking banks with Self Help Groups, NB.DPD.FS. 4631/92-A/91-92, Circular No. DPD/104, India. National Bank of Agriculture and Rural Development (2006) Progress of SHGBank Linkage in India: 200506, Working Paper, NABARD, India. Pitt, M. and Khandker, S.R. (1998) The impact of group-based credit programs on poor households in Bangladesh: does the gender of participants matter? Journal of Political Economy, 106, pp. 958996. Prowse, M. (2003) Towards a clearer understanding of vulnerability in relation to chronic poverty. CPRC Working Paper No. 24, University of Manchester. Puhazhendi, V. and Badatya, K.C. (2002) SHGbank linkage programme for rural poor an impact assessment. Paper, National Bank for Agriculture and Rural Development NABARD, Mumbai, India. Rosenbaum, P. (1987) Sensitivity analysis to certain permutation inferences in matched observational studies. Biometrika, 74(1), pp. 1326. Rosenbaum, P. (2002) Observational Studies. 2nd ed. (New York: Springer). Rosenbaum, P. and Rubin, D. (1983) Assessing sensitivity to an unobserved binary covariate in an observational study with binary outcome. Journal of the Royal Statistical Society, Series B, 45, pp. 212218. Smith, J. and Todd, P. (2005) Does matching address Lalondes critique of nonexperimental estimators? Journal of Econometrics, 125, pp. 305353. Townsend, R. (1995) Consumption insurance: an evaluation of risk bearing systems in low-income economies. Journal of Economic Perspectives, 9(3), pp. 83102. Zaman, H. (2000) Assessing the poverty and vulnerability impact of micro-credit in Bangladesh: a case study of BRAC. Working Paper Series No. 2145 for World Development Report, World Bank, Washington, DC. Zhang, Y. and Wan, G. (2006) An empirical analysis of household vulnerability in Rural China. Journal of the Asia Pacic Economy, 11(2), pp. 196211. Zimmerman, F. and Carter, M. (2003) Asset smoothing, consumption smoothing and the reproduction of inequality under risk and subsistence constraints. Journal of Development Economics, 71, pp. 233260.

Downloaded by [Flinders University of South Australia] at 02:10 21 March 2013