Micro-finance Lifespans: A Study of Attrition and Exclusion in Self-Help Groups in India

Jean-Marie Baland

University of Namur

Rohini Somanathan

Delhi School of Economics

and

Lore Vandewalle

University of Namur

Prepared for Presentation at tbe Brookings-NCAER India Policy Forum 2007New Delhi

July 17-18, 2007

Conference DraftNot for Citation

Micro-finance Lifespans: A Study of Attrition and

Exclusion in Self-Help Groups in India

Jean- Marie Baland, Rohini Somanathan and Lore Vandewalle†

July 2007

Abstract

We study member attrition and group failure in Self-Help Groups (SHGs), the dominant

institutional form in Indian micro-finance. We use survey data on groups of women in two

rural and relatively poor areas of the country: the districts of Keonjhar and Mayurbhanj

in northern Orissa and the district of Raigarh in Chattisgarh. We find that 10% of the

1,100 SHGs created over the period 1998-2006 are no longer active and 20% of women who

joined a group are no longer part of an SHG network. We estimate duration models of

group and member survival. An important determinant of group survival is the presence of

some educated members, perhaps because they can ensure accurate accounting and provide

leadership in other aspects of group functioning. Members stay longer in groups in which

they have relatives. Although fractionalization and other measures of social cohesion do

not systematically influence group failure, they do result in more member exit. Members of

the most disadvantaged ethnic groups are particularly vulnerable to group heterogeneity.

Keywords: microcredit, collective action, social exclusion. JEL codes: I38, G21, O12,

O16∗Thanks to the PRADAN team for many useful discussions and for their support in facilitating data collection,

to Sandeep Goyal, Sanjay Prasad, Amit Kumar, Rahul Sharan and Saurabh Singhal for research assistance andto FUCID (La Fondation Universitaire de Cooperation Internationale et de Developpement) and the ARCprogram of the French speaking community in Belgium for financial support.

†CRED, University of Namur ( jean-marie.baland@fundp.ac.be), Delhi School of Economics (ro-hini@econdse.org) and CRED and FNRS, University of Namur (lore.vandewalle@fundp.ac.be )

1

1 Introduction

Micro-finance is often advocated as a solution to multiple social problems. Poor households

with access to credit can make investments that bring them out of poverty, household and

regional income and wealth disparities are reduced, and groups meetings provide forums for

collective action that improves gender relations and local governance. Over the last few years,

savings and credit groups have also helped manage some important social programs of the Indian

government, such as the distribution of food grains and school meals in state primary schools.

The number of organizations involved in the micro-finance sector is very large and contracts vary

considerably. Most loans involve some form of group lending with members sharing responsibility

for repayment. It is useful to distinguish between two principal institutional forms through

which lending takes place. In the first, micro-finance institutions organize potential lenders into

groups. Group composition may be determined by random factors, as in the case of FINCA

in Peru, or the matching preferences of members as in the Grameen Bank model.1 In all

these cases, micro-finance institutions are intimately involved in all stages of the group lending

process; group formation, rules that govern member interactions, and the provision of credit.

Bank representatives are often present at group meetings. An alternative setting is one in

which the institutions involved with group formation, group interactions and group loans are

much more loosely connected. Government and non-government agencies are involved in the

formation of groups, which determine their own rules for savings and lending. Some of these

groups subsequently borrow from commercial banks. This is the dominant model in Indian

micro-finance, in terms of both outreach and total loan disbursements.

The present structure of the micro-finance sector in India emerged in the early nineties when

the Reserve Bank of India issued guidelines to all nationalized commercial banks encouraging

them to lend to informal groups which came to be called Self Help Groups (SHGs). Since then,

such groups have been actively promoted by a number of different agencies and the National

Bank for Agriculture and Rural Development (NABARD) has provided banks with subsidized

credit for SHG lending.2 Government statistics currently report over two and a half million

groups and 32 millions households covered by the program.3

1See Karlan (2006) for a description of group operations in FiNCA and Armendariz de Aghion and Mor-

duch(2005), chapter 4 for details on Grameen Bank group lending practices.2See Reserve Bank of India (1991) and National Bank for Agriculture and Rural Development (1992) for the

original policy statements.3NABARD (2006), page 38.

2

This paper reports findings from a study of about 1,100 SHGs with a total of 16,000 women

members in selected regions of rural Orissa and Chattisgarh. The groups were formed during the

period 1998-2006 by PRADAN, a non-government organization working in these areas. We find

that 10% of all groups created over this period are no longer active and that 14% of members left

groups while it was still functioning. Some women who leave groups join other SHGs. Taking

them into account, we find that about one-fifith of members who joined a group had left the

SHG network within the 8-year period. We’re interested in the factors underlying these rates

of group survival and member attrition.

We estimate models of both group and member duration and find that factors behind group

survival are quite different from those affecting member longevity. The presence of educated

members is strongly associated with longer-lived groups, perhaps because they can facilitate

transactions and ensure that group accounts are accurate. The presence of other SHGs in the

area also has a positive effect on group duration. It may be that a dense cluster of groups allows

for the sharing of costs, provides each group with ideas for successful activities, or simply instills

in members the desire to survive, compete and be part of a larger network. Measures of frac-

tionalization and social heterogeneity that are commonin the literature do not have systematic

effects on group survival.

Both member and group characteristics matter in explaining the departure of individuals from

groups. Controlling for a large set of individual and group characteristics, we find that tribal

and scheduled caste women have shorter SHG lifespans than groups that lie higher in the social

hierarchy. In addition, higher levels of education, bigger landholdings and more relatives within

the SHG are also associated with a lower probability of exit. Group level social fractionalization

indices increase departures, especially for the Scheduled Tribes. We also find that the bulk of

the difference in member duration observed between Chattisgarh and Orissa can be attributed

to characteristics of groups in these areas; regional variations in performance are negligible once

these characteristics are incorporated in our model.

Our results suggest that it is problematic to evaluate the success of micro-finance interventions

based on conventionally reported coverage figures because they do not account for attrition.

The aggregate rates of group failure and member exit are not however not the main problem in

the evaluation of micro-credit programs in India. These rates, though sizable, are not enormous;

our main concern is their systematic relationship with measures of social disadvantage. This

makes is unlikely that women moving out of SHGs are going on to enter into individual contracts

with lending institutions. It also means that some of those in desperate need of credit cannot

obtain it from within this sector. An additional concern is that lending by commercial banks to

3

Self Help Groups is considered priority sector lending by the banking system and may therefore

crowd out other lending. Disbursements by commercial banks to SHGs were about half the

size of all direct bank credit to small farmers in 2003-2004 and SHG credit has been rapidly

rising since.4 Although the duration of membership is only one, admittedly crude, measure of

the performance of the micro-finance sector, our study suggests that survey data which follows

members and groups in this sector is critical to an assessment of Indian micro-finance.

We provide a brief institutional history of the micro-finance sector in India in Section 2. Our

survey data, some summary statistics and empirical methods are described in Sections 3 and 4

respectively. Results are presented in Section 5 and are followed by a some concluding remarks.

2 Micro-finance Institutions in India

Many detailed accounts on the characteristics of rural credit markets in India and the spread

of rural banking in the post independence period are available. A number of studies, starting

with the All India Rural Credit Survey in 1954, found that the rural poor had high levels of

indebtedness and very limited bank access.5

As part of a process of providing banking services to this population, the State Bank of India

was set up in 1955, the 14 largest commercial banks were nationalized in 1969, and the National

Bank for Agriculture and Rural Development (NABARD) was created in 1982. Each of the

nationalized banks were assigned the role of lead banks in particular parts of the country and

they were required to maintain specific ratios of urban to rural branches. A vast network

comprising thousands of cooperative banks and regional rural banks was created primarily to

provide credit to the agricultural sector. These efforts created a vast system for credit delivery

with high default rates and operating costs. 6

In the early nineties central bankers tried to revitalize the elaborate and largely inefficient

4Loan disbursements to farmers with less than 2.5 acres of land were 7953 crore rupees in 2003-4 while SHG

linked loans were Rs. 3904 crore. (Reserve Bank of India, 2006, Tables 58 and 71).5See for example, Bell(1990) for summary statistics on rural borrowing and indebtedness based on rural credit

surveys and Karmakar (1999) for recent figures on the numbers of different types of rural banking institutions.6The extent to which this system reached the poor is unclear, but there is evidence that they influenced

regional poverty rates (Burgess and Pande, 2005).

4

banking system through which rural credit came to be provided after Indian Independence. The

start of institutionalized micro-finance in India is often attributed to the circular that was issued

by the Reserve Bank to all nationalized commercial banks in 1991, announcing the objective of

linking informal groups of rural poor with the existing banking system. Some non-government

organizations at the time had organized women into informal groups that used their pooled

savings for mutual insurance and small credit needs. Based on studies of these informal groups,

it was believed that they had the “potential to bring together the formal banking structure and

the rural poor for mutual benefit”(RBI, 1991). The following year NABARD launched a pilot

project which linked 500 groups with commercial banks. The banks were offered finance from

NABARD for such lending at the rate of 6.5% per annum. It was recommended that banks either

lend directly to groups at 11.5% per annum or route their loans through voluntary agencies at

the lower rate of 8.5% in order to cover the transactions costs of these organizations (NABARD,

1992). Banks were also permitted to classify such lending under Advances to Weaker Sections,

and this category formerly accounted for a large fraction of their unprofitable loans.

Another major change came in April 1999, with the launching of the Swarnajayanti Gram

Swarozgar Yojana, popularly known as the SGSY (RBI, 1999). The program was introduced

to increase the penetration of SHGs among families below the poverty line. This reflected a

significant change in state policy : It provided direct subsidies to borrowers as only part of the

initial loan had to be repaid. It also restricted the composition of a group to families below

the poverty line. Subject to caps, rates of subsidy were 50% for borrowers from the Scheduled

Castes and Tribes and 30% for other poor households. A proper evaluation of the changes that

the SGSY brought about in the composition and performance of SHGs is yet to be undertaken.7

The NABARD pilot program of 1992 was widely regarded as successful and the number of SHGs

linked to the banking system has been rising rapidly over the last 15 years. Table 1 shows official

figures for the number of groups that have received bank loans since 1992. The total number

of groups linked is currently reported to be over 2.5 million. There is considerable variance in

the distribution of SHGs across states and an overwhelming fraction is in the 4 southern states.

Alternative models of lending have been growing in the last few years and private banks have

also entered the sector. The combined coverage of institutions in this class remains small relative

7Our own surveys indicate that the combination of restrictions of group composition and subsidies may

have been a factor causing the dissolution of some groups. Surveyed groups were asked about whether or not

they received a subsidy. Although very few of the subsidized groups failed, other groups sometimes cited their

exclusion from state subsidies as a reason for group dissolution. In some cases, a few members were excluded

from the group by the others because they were not on government poverty lists.

5

to SHG lending. This contrasts sharply with countries such as Bangladesh and Indonesia, where

each of the major specialized micro-finance institutions is larger than the combined non-SHG

sector in India. (RBI, 2005 (chapter 2), Basu, 2005).

Table 1: SHG- Commercial Bank Linkages (Cumulative), 1992-2007

Year (end-March) No. of SHGs linked Bank loans

(Rs. crore)

1992-93 255 .29

1993-94 620 .65

1994-95 2122 2

1995-96 4757 6

1996-97 8598 12

1997-98 14317 24

1998-99 32995 57

1999-00 114775 193

2000-01 263825 480

2001-02 461478 1026

2002-03 717360 2048

2003-04 1079091 3904

2004-05 1618456 6898

2005-06 a 2238565 11397

2006-07 b 2580000 14479Sources: Figures from 1992-2005 have been taken from Reserve

Bank of India (RBI) (2006) while those for 2006-2007 are from

RBI (2007).

aprovisional estimatesbupto end February 2007

It appears that the institutions that came to dominate Indian micro-finance sector resulted

from the combined presence of a vibrant non-government sector engaged in rural development

and an extensive but unprofitable network of rural banks and agricultural cooperatives that

were created with the explicit purpose of providing small loans to the rural poor.8 Policy

makers could observe the phenomenal expansions in outreach of micro-finance institutions like

the Grameen Bank in Bangladesh and other countries. The Grameen Bank alone, starting from

8Harper (2002) provides some additional reasons for why SHGs rather than Grameen type institutions are

more successful in the Indian context.

6

humble beginnings had reach almost a quarter of all Bangladeshi villages by 1991.9 The linking

of banks with SHGs seemed a natural step to take and it gave India its own particular brand

of micro-finance.

3 Data

Our data is based on surveys of SHGs created by PRADAN, one of the earliest non-government

organizations in the micro-finance sector. We surveyed a total of 1,103 SHGs created by

PRADAN in two of its field locations, one in northern Orissa and the other in central Chhatis-

garh. This included every group that has been formed since the start of the program in these

areas. In addition to group-level data, we collected information on the backgrounds and SHG

activities of 16,892 women who, at any stage, had been members of these groups. We begin this

section with a brief outline of the PRADAN SHG program. This is followed by a description of

our survey methodology and some descriptive statistics on groups and members.

3.1 The PRADAN SHG Program

The first SHG created by PRADAN was in the Alwar district of Rajasthan in 1987. In the

subsequent years most groups were formed in what now constitutes the state of Jharkhand.

More recently, new SHG projects have appeared in West Bengal, Madhya Pradesh, Orissa and

Chhatisgarh. Table 2 provides a list of PRADAN locations in each of the 7 states in which the

organization operates, together with the year of the first SHG and the total number of SHGs

in existence at the end of March 2006. 10

The groups formed by PRADAN are a small fraction of the micro-finance sector, but an impor-

tant presence in their respective areas. The SHG program operates by first targeting adminis-

trative blocks with high levels of rural poverty within particular districts and then building a

dense network of SHGs in these area over a few years. In many of the more recent PRADAN

locations, SHGs have been the first intervention in each village and group meetings were used to

9Figures for the total number of villages are from the Bangladesh Bureau of Statistics (www.bbs.gov.bd) and

those for the number covered by the Grameen Bank are available at www.grameen-info.org.10current aggregate figures for the SHG program are available at www.pradan.net

7

introduce other activities aimed at raising agricultural productivity and raising incomes. Since

it is the backward areas that are chosen for new interventions, the social composition of the

administrative blocks in which PRADAN operates are often quite different from other parts of

the state and district. PRADAN areas often have a high proportion of social groups classified

as Scheduled Tribes and lower literacy rates than the state average.

The groups themselves consist entirely of women and follow the guidelines issued by NABARD

and the Reserve Bank (RBI, 1999, NABARD, 1992). Each SHG has between 10-25 members

and is fairly homogeneous in terms of caste. Large villages often have multiple groups, one

in each hamlet. PRADAN professionals begin the process of group formation by calling a

meeting in some public space in the village. They discuss the benefits of membership and

some general principles followed by successful groups (compulsory attendance, weekly savings,

typical interest rates, bookkeeping). Interested women are enlisted, a regular meeting time is

set and the professional is usually present at meetings until membership becomes fairly stable

and all members are familiar with group practices. Each group is provided with a register for

keeping accounts and a cash box, and either designates one of the members to keep accounts or

hires an accountant. The register, cash box and keys are usually rotated across the members.

As groups mature, they get linked to other groups in the area and select representatives who

attend these cluster meetings. Smoothly functioning groups typically open a savings account

with a nearby commercial bank within a year of their inception. After this stage, PRADAN

professionals discuss possible self-employment projects with the group, some members decide on

particular projects, the total investment required is determined and the group then applies to a

commercial bank for a loan. This loan constitutes their first bank linkage. Bank funds usually

come into the group which then lends to individual members at interest rates determined by

the group. Payments are made to the group which then repays the bank on the stipulated date.

As the group gains confidence and shows that it can manage its functioning independently,

PRADAN professionals withdraw and their direct interactions with group members increasingly

take place through the larger cluster meetings and on occasions when they visit the village to

initiate other livelihood projects. Regular communication with PRADAN takes place mainly

through copies of weekly accounting transactions that are sent in to the local office. Groups are

free to determine the rules under which they operate and the stringency with which they are

implemented. These rules and interest rates vary considerably accross groups and also change

periodically within particular groups. After the start of the SGSY in 1999, some subsidies to

groups are routed through PRADAN, provided the groups satisfy the selection criteria required

by the scheme. Subsidized and unsubsidized SHGs therefore co-exist in the same area.

8

As a result of the growing independence of these groups over time, the organizations promoting

SHGs are not always informed about groups that are no longer active. Successful groups may

stop sending in accounts as they reduce their reliance on PRADAN, others may temporarily

suspend meetings because its members migrate seasonally and yet others may actually collapse.

Survey data is therefore required to correctly estimate rates of group survival.

3.2 The Survey Design

We study all PRADAN groups created in the districts of Keonjhar and Mayurbhanj in northern

Orissa and the district of Raigarh in eastern Chattisgarh. PRADAN works in a total of 8

administrative blocks in these district. Both the Orissa districts are serviced by the professionals

in Keonjhar and we henceforth refer to all groups in this area as the Keonjhar SHGs. The three

survey districts are are shown in Figure 1 and the surveyed areas within each of these districts

are indicated in Figures 2-4. Although only a small fraction of total district villages are covered

by the program, they are geographically clustered and upto 40% of the villages in some blocks

have a PRADAN SHG. These dense pockets make it easier for professionals to visit these areas

and also allow groups to benefit from frequent contact with each other. 11

The first SHGs in Keonjhar and Raigarh were started in 1998 and 1999 respectively. We survey

the census of PRADAN SHGs in these locations and therefore have group and member level

characteristics for each group formed in the area, and each member that was ever part of the

network, whether or not they were part of the group at the time of the survey.

At the group level we collected information on group rules and activities and on the timing

of various significant events. These events include the inception of the SHG, the creation of

savings accounts, the dates and sizes of bank loans, membership of SHG clusters and the date of

the last meeting in case the SHG is no longer active. Group rules include fines (for attendance

and late repayment), minimum savings requirements, interest rates and the rotation of group

office-bearers and representatives. Our data on collective activities undertaken by the group

include its involvement in village and family conflicts, government contracts to provide school

meals and various types of other activities such as the selling of forest produce or visiting a

government official.

11Some of these benefits are studied by Nair (2005)

9

Figure 1: Study Area

OrissaChattisgarh

Madhya Pradesh

Maharashtra

Andhra Pradesh

JharkhandWest Bengal

Uttar Pradesh

Legend

State Boundary

STUDY DISTRICTS

Keonjhar

Mayurbhanj

Raigarh

THE STUDY AREA MAP

10

For all present and past members, and for each accountant, we recorded the dates at which

the they entered and left the group, and the total value of loans that were taken either from

the group’s internal funds or a bank. For the members, we collected data on a standard set of

characteristics relating to their social and economic background: caste, education, age, marital

status, fertility, household landholdings and some parental information. To better understand

the determinants of group cohesion and conflict, we also created a matrix of relationship between

group members which recorded family ties between members. Since we had a census rather than

a sample of groups in each area, we were also able to record the number of other PRADAN

groups in the same revenue village.

For groups that were no longer active, we asked members the main reason for group failure and

recorded the most popular response. Similarly, we asked members who had left groups the main

reason for their departure.

Figure 2: Raigarh

11

Figure 3: Keonjhar

Figure 4: Mayurbhanj

3.3 Descriptive Statistics

Table 3 gives us the number of SHGs that were created and dissolved in our survey areas over

the period 1998-2006. The survey in Keonjhar was conducted during the summer of 2006 and

included all groups created till the end of June that year. The Raigarh survey was in January

2007 and included groups formed at any point during 2006. Of the 1,103 groups created, 10%

were no longer active by the time of the survey. A higher proportion of groups failed in Raigarh

than in Keonjhar(12% as compared with 9% ).

Table 4 contains descriptive statistics for a variety of group characteristics according to the

survival status of the group. A few of the differences between existing and inactive groups are

worth emphasizing. First, the differences in the duration of these two types of groups are not

enormous. Existing groups have been in operation for an average of about three years in both

areas, while defunct groups in had lives that were 20% shorter in Keonjhar and 45% shorter in

Raigarh. These groups do not fail soon after they are formed. Second, there are many more

homogenous groups in Keonjhar in both categories, and a these groups as a whole have higher

rates of failure. If we look more closely within homogeneous groups, we find that this pattern is

marked for the Scheduled Tribes, but not for other caste categories.Third, groups that succeed

are both more involved in village activities and in the lives of their members. Lastly, members

of these groups are, on average, more educated and have more land and more of them act as

accountants for their group. Difference in group size are negligible. The stated reasons for group

failures and member departures are summarized in Table 6. In both cases, the two principal

reasons relate to group conflicts and difficulty in either saving or repaying loans.

Tables 5-8 compares women who left SHGs with those who are either still in them, or were in

them until the time that they were active. The differences in demographics between the two

groups of members are not large. What is striking here is the importance of family and social

connections. Women with relatives in a group are more likely to stay in it and homogeneous

single caste groups are better able to retain members. We also see that even after we control for

duration in a group, members who eventually leave are given fewer loans and are less likely to

be a Chairman of the group at some point during their stay. From Table 7 it appears unlikely

that defaulting with a loan is an important reason for the exit of a member. The duration in

a group does not seem to depend on whether or not the member received a share of a loan

obtained by the group through a bank linkage.

4 Empirical Methods

Many of the more standard regression techniques that are used in economic analysis are not

appropriate to the type of data we have or the questions that we are interested in. The main

data issue is that we would like to understand group failure and member exit using a data set in

which most groups and members are still active. In other words, our data is right censored. To

see why this matters, suppose that we measure group failure (or member exit) using a binary

variable which takes the value of 1 for groups that are no longer active (or members that have

left the system) and zero otherwise. We have data on a set of covariates that we believe to be

associated with longer group (or member) lives. If groups are created at different points in time

(or members enter at different times), and older groups (and members) have more favorable

characteristics, the above method would not be able to detect the correct effect or covariates

and may even assign them the wrong sign. A similar problem would occur if we used more

continuous outcome measures such as loans taken or the number of days a group or member

remains in the system.

If we tried to get around the above problem by using only data on failed groups and departed

members, we would lose a lot of data, especially in the case of a rapidly growing program,

where a large fraction of the observations are in fact censored. What we require are methods

that allow us to use censored observations by making use of information on the censored group

or member until the time of censoring, rather than simply ignoring these observations or not

accounting for the fact that they are censored. This is precisely what the methods of survival

analysis that have been popular in the biomedical and quality control fields, are designed to do.

13

These methods are used to estimate the time until events occur; in our case, the events being

either group dissolution or member exit. We begin with some standard definitions of functions

that we are going to estimate for our data.

Denote by the nonnegative random variable X, the time until the event of interest. In our

context, this is the time until the dissolution of the group or until a member departs in the case

of the member level analysis. The distribution of X can be represented in several ways.12 The

survival function ST (t) gives the probability of surviving beyond a time t or, in other words,

the probability that the random variable T ≥ t. The hazard rate hT (t) reflects the chance

of experiencing the event in the next instant in time. The cumulative hazard rate HT (t) is

the accumulation of the hazard rates over time. For a continuous random variable T with a

probability density function f(t), we have S(t) =∫∞

tf(x)dx, the hazard function is given by

h(t) = f(t)S(t)

= −dlogS(t)dt

and the cumulative hazard function H(t) = −logS(t)

These three representations of the distribution of T can be estimated either parametric or

non-parametric methods. Non-parametric estimators are a natural choice when dealing with a

homogeneous population because of the flexibility they offer. Our population is far from homo-

geneous but we begin with these nonparametric estimates to summarize the survival behavior

of groups and members before going on to a paramteric model which allows us to easily incorpo-

rate covariates and thereby estimate the causal effects of group and member characteristics on

survival rates. A variety of different nonparametric estimators and parametric models are avail-

able. For nonparametric estimates we focus on the Nelson-Aalen estimator of the cumulative

hazard function which is shown to have desirable small sample properties and on a smoothed

hazard rate derived from this estimator which allows us a better visual examination of changes

in the instantaneous hazard. For parametric estimates we use the Weibull model for reasons

discussed below.

4.1 Nonparametric Estimates of Cumulative Hazard Rates

With right censored data, the exact lifetime is only observed if failure or exit occurs before the

time of censoring, namely the date at which the group was surveyed. In the discussion below,

we will usually refer to events as the the exit of SHG members, although the same principle

applies for group failure.

12This discussion is based on Klein and Moeschberger (2003), chapters 2 and 3.

14

Suppose that in our data, members exit groups at D distinct times t1 < t2 < ... < tD and that

at time ti there are di departures. Time, in our case, is the number of days since the member

joined the group. Let Yi represent the number of individuals who are at risk at time ti. In our

case, this is the number of members who are still part of the group at ti or who or who leave

it at ti. The ratio di/Yi estimates the conditional probability that a group or a member who

survives to time ti, experiences the event at time ti. The Nelson-Aalen estimator is then given

by:

H(t) =

{0 if t ≤ t1∑

ti≤tdi

Yiif t1 ≤ t

(1)

By smoothing the jump sizes of this estimator with a parametric kernel, estimates we can obtain

a hazard function h(t).

Observations on groups that were active at the time of the survey and on the members in them

are all right censored. Depending on the duration at that time, they are subtracted form the

relevant Yi.

To test the equality of survivor functions across groups against the global alternative that at

least one of the populations differs from the others at some time, a number of nonparametric

tests are available. These are typically based on weighted comparisons of differences between

the observed failures and those expected under the null hypothesis. We compare the survival

functions in our two regions based on the Wilcoxon-Breslow-Gehan test, which extends the

Wilcoxon rank sum test that is commonly used to compare distributions to censored data.

Although the test does not rely on any assumptions about the types of distributions being

compared, it does assume that the nature of censoring is the same in both samples.

4.2 Parametric Methods: The Weibull Model

Parametric methods of survival analysis assume particular forms for the survival function. This

additional structure, while restrictive, allows us to easily incorporate the effects of covariates

on survival time. We use the Weibull distribution to study the effects of group and member

characteristics on their survival. The Weibull survival function is given by

SX(x) = exp(−λxα),

the hazard function by

hX(x) = αλxα−1,

15

and the cumulative hazard function by

HX(x) = λxα.

By using this model we restrict hazard rates to be monotonic over time, but they can be either

increasing, decreasing or constant, depending on the value of the parameter α. If α = 1 this

model reduces to an exponential model which has a constant hazard rate, for α > 1 hazard

rates are increasing and with α < 1 they are decreasing. Notice that the natural log of the

cumulative hazard function in the Weibull model is linear as a function of the log of member

duration. Figure 1 plots these two variables for our data set (using Nelson-Aalen estimates of

H(t)). The model seems to fit the data very well except for members with very short durations

within groups.

Figure 5: Appropriateness of the Weibull Model

The Weibull model has both an accelerated failure time (AFT) representation, where the de-

pendent variable is the natural log of survival time and is a linear function of covariates, and

a proportional hazard (PH) representation. We present our results within the structure of a

proportional hazards model. Given a vector of covariates Z and corresponding coefficients β,

the hazard rate is given by

h(x|Z) = (αλxα−1) exp(β′Z).

and (αλxα−1) is referred to as the baseline hazard, h0. All our results are presented in the form

of hazard ratios corresponding to our explanatory variables. For binary variables these tell us

the factor by which the hazard function moves up or down relative to the baseline hazard. In

general, it gives us the ratio of the hazard function to the baseline hazard when for a unit change

16

in an explanatory variable. A hazard-ratio of 1 therefore suggests that the co-variate has not

effect on the risk of failure.

5 Results

5.1 Nonparametric Estimates

We first present non-parametric estimates of hazard functions separately for each geographic

area and then discuss the effects of a group and member characteristics based on the Weibull

model. Figure 6 shows Nelson-Aalen estimates for cumulative hazard functions for SHGs in

Figure 6: Nelson-Aalen Estimates of Regional Hazard Rates: SHG level

Keonjhar and Raigarh as well smoothed hazard functions for these areas. 13. The hazard

13Keonjhar refers to all groups under the Keonjhar branch office of PRADAN and therefore covers both

Keonjhar and Mayurbhanj districts

17

functions are obtained by a kernel smoothing of the hazard contributions provided by the Nelson-

Aalen cumulative hazard functions and, like all estimates obtained by kernel procedures, they

are less reliable at the end points of the the time-interval where the sample is thin.

As seen in the summary statistics in Table 3, the risk of failure is higher in Raigarh. The double-

humped hazard rate for Raigarh is interesting as it reflects two different phases in a group’s life

when it is especially vulnerable; about a year after inception, and then again after three of four

years. Hazard rates in Keonjhar change much less over a group’s life. Our summary statistics in

show that Raigarh groups are much more heterogeneous than those in Keonjhar and that group

conflict is more salient. It may be that conflict in heterogeneous groups appears periodically,

depending on the timing of events such as bank linkages or bad harvests.

Figure 7 displays hazard rates using member-level data for the two regions. The risk of member

departures in the early stages of membership are very similar, but once again, we see a second

hump in the Raigarh hazard function that does not appear in Keonjhar. The differences in the

hazard functions of members in the two areas appear less marked than the differences in group

hazard across these regions.

Figure 7: Member Level Regional Hazard Rates

18

During the survey, members who had left groups were asked for the principal reason for their

departure. Table The most frequent reason stated was difficulty in saving the amount required

by the group or in repaying loans. The second most frequent was conflict with other group

members. Figures 8 and 9 are each based on a truncated sample. For Figure 8, among the

group of members who have left groups, we include only those that left due to difficulty in

savings or repayment . Similarly, in the case of Figure 9 we keep only those that stated conflict

as their reason for leaving the group. The reversal of hazard rates across regions in these two

figures is striking. Exit due to difficulty in saving and repayment is much more important

in Keonjhar and reverses the relative position of the hazard functions, while conflict is more

important in Raigarh.

Figure 8: Hazard Due to Difficulty in Saving: Member-Level Data

5.2 Parametric Estimates

Hazard ratios from estimating the Weibull model as shown in Tables 9-12.

Of the various group characteristics that we consider, the only ones that systematically affect

19

Figure 9: Hazard Due to Member Conflict: Member Level Data

20

hazard rates are the number of other PRADAN SHGs in the area and the maximum level of

education within the group. Both these lower the risk of group failure. An additional year

of education for the most educated member of the group lowers the risk of failure by 8% and

an additional group in the same village lowers the hazard rate by 18%. It is conceivable that

the presence of an educated member facilitates interactions with banks and other officials and

ensures better book-keeping. Other groups in the village may help either through the sharing

of information or by making it more likely that PRADAN professional frequently visit the area.

We have not of course looked at these particular mechanisms directly as yet, and these remain

conjectures based on some anecdotal evidence.

Member departures are much more sensitive to variables that capture group composition. Ed-

ucated members stay longer within groups as do those who have been married and those with

children. Average age, average land owned and more initial members are all are both positively

related to survival. Fractionalization does seem to encourage exit, and this is especially true of

the groups with Scheduled Tribes.

To consider the effects of heterogeneity more carefully, we estimate the model with members

from groups with only one or two official caste categories. These results are shown in Tables

11-12. Scheduled Tribes are much more likely to exit groups if they are mixed, even in case

where groups are formed together with other tribes. Scheduled Castes do not show differential

rates of exit for homogeneous and mixed groups. For Backward Castes the probability of exit

from mixed groups is not significantly higher than for homogeneous groups except when they

are in a minority.

6 Concluding Remarks

Our analysis of group and member duration in this paper brings us towards three main conclu-

sions. First, rates of group of failure and member attrition are not insignificant. They imply

that about one-fifth of those joining an SHG network leave it within an 8-year period. These

rates in themselves are not alarming and their welfare implications are not straightforward. It

may be that participation in SHGs works best for short periods of time, after which borrowers

are able to enter into individual lending contracts with rural banks. Our study of group failure

and changes in group composition does not however support this type of an explanation and

this brings us to our other two results: Groups with educated members and and those in villages

with other SHGs are less likely to fail and it is therefore the remote, disadvantaged communities

that are most likely to be deprived of credit through these institutions. Lastly, the Scheduled

Tribes which have been widely recognized as being especially disadvantaged, are most likely to

remain in the system if they are in narrowly homogeneous groups. This poses a real challenge

to policy makers who want micro-finance institutions in India to provide disadvantaged com-

munities with access to credit: Homogeneous groups of uneducated members are likely to fail,

yet attrition rates of these members in mixed groups are high.

22

References

Armendariz de Aghion, Beatriz, and Jonathan Morduch (2005) The Economics of Microfinance

(Cambridge, MA: MIT Press)

Basu, Priya, and Pradeep Srivastava (2005) ‘Scaling up microfinance for india’s rural poor.’

World Bank Policy Research Working Paper 3646.

Bell, Clive (1990) ‘Interactions between institutional and informal credit agencies in rural india.’

World Bank Economic Review 4(3), 297–327

Burgess, Robin, and Rohini Pande (2005) ‘Can rural banks reduce poverty? evidence from the

indian social banking experimen.’ American Economic Review 95(3), 780–795

Harper, Malcolm (2002) ‘Self-help groups and grameen bank groups: What are the differ-

ences?’ In Beyond Micro-Credit: Putting Development Back into Micro-Finance, ed.

Thomas Fisher and M.S. Sriram (New Delhi: Vistar Publications)

K. G. Karmarkar, K.G. (1999) Rural Credit and Self-Help Groups: Microfinance needs and

concepts in India (New Delhi: Sage)

Karlan, Dean (2006) ‘Social connections and group banking.’ Working paper.

Klein, John P., and Melvin L. Moeschberger (2003) Survival Analysis (New York: Springer)

Nair, Ajay (2005) ‘Sustainability of microfinance self-help groups in india: Would federating

help?’ World Bank Policy Research Working Paper 3516.

National Bank for Agriculture and Rural Development (1992) Guidelines for the Pilot Project

for linking banks with Self Help Groups: Circular issued to all Commercial Banks

(www.nabard.org: Mumbai)

(2006) NABARD Annual Report, 2005-2006 (www.nabard.org: Mumbai)

Reserve Bank of India (1991) Improving access of rural poor to banking-Role of interven-

ing agencies-Self Help Groups: Circular issued to all Scheduled Commercial Banks

(www.rbi.org.in: Mumbai)

(1999) Swarnajayanti Gram Swarozgar Yojana: Circular issued to all Scheduled Commer-

cial Banks (www.rbi.org.in: Mumbai)

(2005) Internal Group to Examine Issues Relating to Rural Credit and Microfinance

(www.rbi.org.in: Mumbai)

23

(2006) Handbook of Statistics on Indian Economy (www.rbi.org.in: Mumbai)

(2007) Annual Policy Statement for the Year 2007-2008 (www.rbi.org.in: Mumbai)

24

Table 2: Number of PRADAN SHGs in India ( 31/03/06)

State Location Yeara First SHG # SHGs

Chhatisgarh Raigarh 1998 1999 532

Jharkhand Godda 1987 1989 314

Jharkhand Barhi 1992 1992 411

Jharkhand Lohardaga 1992 1995 449

Jharkhand West Singhbhum 1992 1996 363

Jharkhand Gumla 1994 1994 484

Jharkhand Dumka 1995 1989 318

Jharkhand East Singhbhum 1997 1996 392

Jharkhand Khunti 2000 1997 314

Jharkhand Koderma 2000 1992 359

Jharkhand Petarbar 2000 1998 322

Jharkhand Deogarh 2002 1989 280

Rajasthan Dausa 1999 1999 171

Rajasthan Dholpur 1999 2000 180

Madhya Pradesh Kesla 1986 1996 300

Madhya Pradesh Vidisha 2000 2000 44

Madhya Pradesh Sidhi 2002 2005 49

Madhya Pradesh Dindori 2005 2005 110

Orissa Keonjhar 1990 1998 506

Orissa Balliguda 2001 2001 201

Rajasthan Alwar 1986 1987 162

West Bengal Purulia 1987 1995 218

West Bengal Bankura 2005 2000 142

Total 6701

aThis refers to the year in which a PRADAN office was started in the area. The Deogarh and Dumka SHGs

were initially under the Godda office and the Koderma and Peterbar SHGs were managed by the Barhi office.

This is why the first SHG in these areas predates the opening of the PRADAN branch office.

25

Table 3: Year-wise formation and dissolution of SHGs: Survey Data, 1998-2006.

Year Started Dissolved Bank loan

Keonjhar Raigarh Keonjhar Raigarh Keonjhar Raigarh

1998 4 0 0 0 0 0

1999 10 18 0 0 0 0

2000 51 61 0 0 3

2001 27 36 3 5 2 7

2002 155 30 4 5 14 23

2003 89 46 11 7 100 31

2004 95 172 9 8 95 100

2005 85 160 17 23 89 140

2006 16 47 2 20 62 91

Totala 532 570 46 69 362 395

aThere are two main reasons why the totals in this table do not match those in Table 2. First, we included all

groups that were formed before the survey date, and some of these were created after March 2006. Second, Table

2 is based on administrative data that do not always account for group failures since these are not consistently

reported.

26

Table 4: Group Characteristics by Survival Status

Keonjhar Raigarh

Existing Dissolved Existing Dissolved

Number of Observations 486 46 501 69

(91) (9) (88) (12)

Average duration (days) 1103 878 1129 620

COMPOSITION

Total number of castes 87 22 61 14

Number of castes in a group 2.4 1.8 4.2 3.5

Number of caste categories (st, sc, bc, fc) 1.8 1.3 2.4 2.3

Fractionalization index 0.27 0.16 0.51 0.46

HOMOGENOUS GROUPS (%) 33 52 9 13

Scheduled Tribes (% of homogenous) 66 86 61 67

Scheduled Castes (% of homogenous) 12 10 23 33

Backward Castes (% of homogenous) 22 4 14 0

Forward Castes (% of homogenous) 0 0 2 0

GROUP ACTIVITIES LAST YEAR

Midday meals (%) 9 0 12 1

PDS (%) 3 0 4 0

Panchayat meetings (%) 34 22 56 35

Exposure trips (%) 70 41 13 6

Federation meetings (%) 12 2 2 0

Meet government officials (%) 20 7 32 16

Involvement in family or village conflicts/ help member in distress (%) 44 25 52 26

RULES

Minimum weekly saving (%) 100 100 94 96

Saving compulsory (%) 29 20 38 39

Groups with fines for absence(%) 93 67 38 26

Absence fine (Rs.) 3.1 2.6 3.8 3.2

Higher interest rates default (%) 16 13 92 91

OTHER CHARACTERISTICS

Any subsidies received (%) 13 0 5 1

Any group projects (%) 80 67 27 6

Total bank linkages 1.2 0.2 1.4 0.3

Member accountants (% accts) 68 42 59 62

MEMBERS

Average number of members 16 15 15 15

Left (%) 13 14 15 15

Literate (%) 34 14 30 25

No school (%) 58 87 64 70

Maximum education (years) 9 5 8 7

Mean education (years) 2.8 1 2 1.6

Mean land (Acres) 1.7 1.4 2 1.927

Table 5: Characteristics of Present and Past Members

Keonjhar Raigarh

present departed all present departed all

Number of women 7471 1097 8568 7048 1268 8316

(87) (13) (100) (85) (15) (100)

CASTE CATEGORY COMPOSITION

ST (%) 59 60 59 45 50 46

SC (%) 13 13 13 22 25 23

BC (%) 27 26 27 31 23 29

FC (%) 1 1 1 2 2 2

BACKGROUND

Education (number of years) 2.7 2.4 2.7 1.9 1.6 1.9

No school (%) 60 65 61 65 68 65

Read and write (%) 32 29 32 30 24 29

Father’s education (number of years) 2.1 1.4 2.0 2.1 1.3 2.0

Land (acres)

RELATION TO GROUP

Relatives within group (%)a 12 8 11 8 6 8

In homogenous groups (%) 35 33 35 10 7 10

Previous shg membership (%) 4 9 5 6 6 6

Joined other shg after leaving (%) 21 19

CHAIRMAN b

membership < 2 years (%) 0.058 0.0061 0.045 0.088 0.034 0.070

2 years < membership < 4 year (%) 0.076 0.037 0.073 0.091 0.037 0.087

4 year < membership (%) 0.083 0.027 0.082 0.088 0.051 0.085

aPercentage of members who have at least one relative in their group.bPercentage of members who have been chairman, given the duration of their membership.

28

Table 6: Stated Reasons for Group Failure and Member Exit

Keonjhar Raigarh

GROUPS

Pradan withdrew support 8 8

(17) (12)

Personal conflicts / leadership problems / accountant problems 20 27

(43) (40)

Unpaid loans / irregular savings 14 17

(31) (24)

Others 4 17

(9) (24)

Total 46 69

(100) (100)

MEMBERS

PERSONAL REASONS

Illness / dead 8.3 8.1

Left village / married / seasonal migration / going to school 17.8 12.0

RELATION TO GROUP

The family was not supportive 6.2 9.1

Could not reimburse a loan taken / difficulty in saving 29.2 17.1

Could not attend the meetings 9.8 12.8

Personal conflict with the group 15.5 20.3

Excluded by the group 4.9 1.0

OTHERS

Wanted to join another group 0.5 6.5

Othersa 7.8 13.1

aOthers includes not understanding the working of the shg, pradan official stopped visiting the

group, the group is too big and no clear reason

29

Table 7: Loans

Keonjhar Raigarh

present departed present departed

Average group loans per member years (Rs.) 1111 441 344 90

Number of members who received part of linked loan 1920 29 1596 52

if shg got only one linkage (72) (56) (85) (45)

Duration in group after first linkage if member had a share in linkage 376 341

Duration in group after first linkage if member had no share in linkage 391 390

linked loans per member years (Rs.) 4197 2564 4218 2184

(for members who were present during all linkages)

30

Table 8: Castes

Keonjhar Raigarh

Scheduled tribe 2867 2030

(57) (43)

Scheduled caste 670 1118

(13) (23)

Backward caste 1420 1525

(28) (32)

Forward caste 87 77

(2) (2)

BIGGEST CASTES

SCHEDULED TRIBE

Bhuiyans 1126 204

Kharia 15 468

Ho 444 5

Munda 533 12

Santhals 501 0

Bathundi 807 0

Gond 433 619

Ganda 372 128

Rathia 0 1602

SCHEDULED CASTE

Harijans 419 11

Chauhan 0 905

BACKWARD CASTE

Yadav 5 711

Mahanta 820 100

Kurmi 493 14

Teli 95 497

31

Table 9: Hazard Rate Estimates at SHG Level, Weibull model

(1) (2) (3)

Shape parameter 1.11 1.12 1.16

Homogenous shg, caste 1,13

(0,38)

Homogenous shg, ST 1,18 1,09

(0,41) (0,39)

Homogenous shg, SC 1,9 1,84

(0,95) (0,93)

Homogenous shg, BC 0,25 0,26

(0,26) (0,28)

Fractionalization 0,83 0,84 0,78

(0,46) (0,47) (0,44)

Average relations in group 0,9 0,84 0,91

(0,59) (0,55) (0,6)

Number of initial members 0,96 0,95* 0,96

(0,03) (0,03) (0,03)

Maximum education in group 0,92** 0,92** 0,91**

(0,02) (0,02) (0,03)

Average land (Acres) 0,95

(0,06)

Average age 0,96**

(0,02)

Average total children 1,12

(0,21)

Average separated 3,9

(3,48)

Concurrent pradan shgs 0,82** 0,82** 0,81**

(0,03) (0,03) (0,03)

Raigarh 1,61** 1,54* 1,72**

(0,37) (0,35) (0,44)

Number of observations 1061 1061 1056

Number of Departures 107 107 105(*) significant at a 10% significance level

(**) significant at a 5% significance level

32

Table 10: Hazard Rate Estimates at Member Level, Weibull model

(1) (2) (3) (3) (3) (3) (4)

ST SC BC FC

Shape parameter 0,74 0,74 0,73 0,76 0,75 0.57 0,74

Caste category, SC 1,08 1,04 0,98

(0.07) (0.06) (0.07)

Caste category, BC 0,92 0.87** 0.84**

(0.05) (0.05) (0.05)

Caste category, FC 1,15 1,03 1,04

(0.2) (0.18) (0.18)

Education (in years) 0.97** 0.97** 0.95** 1.04** 0.97** 0.84** 0.97**

(0.01) (0.01) (0.01) (0.02) (0.01) (0.05) (0.01)

Land (Acres) 0,99 1 1 0.95* 1 1,04 1,01

(0.01) (0.01) (0.01) (0.03) (0.01) (0.05) (0.01)

Age 1* 0.99** 1 1 0.99** 0.96* 1

(0.002) (0.002) (0.003) (0.01) (0.01) (0.02) (0.003)

Separated 0.83** 0.83** 0,86 0.62** 1,01 0,47 0.83**

(0.06) (0.06) (0.09) (0.12) (0.16) (0.35) (0.07)

Total children 0.89** 0.89** 0.91** 0.88** 0.89** 0.8* 0.9**

(0.01) (0.01) (0.02) (0.03) (0.03) (0.1) (0.01)

Relation 0.37** 0.09** 0.09** 0.1** 0.09** 0.001** 0.09**

(0.06) (0.02) (0.04) (0.06) (0.05) (0.003) (0.02)

Homogenous shg, castes 0,91 0,91 1,08 0,86 0,26 0.86*

(0.07) (0.09) (0.22) (0.16) (0.29) (0.08)

Homogenous shg, SC 1,07

(0.17)

Homogenous shg, BC 1,01

(0.16)

Fractionalization 1.39** 1.57** 1,53 1,08 0,61 1.35**

(0.17) (0.25) (0.42) (0.28) (0.78) (0.16)

Average relations in group 7.78** 6.41** 10.91* 5.97** 174.93** 7.56**

(2.33) (2.82) (6.99) (3.48) (426.01) (2.28)

Number of initial members 1.03** 1.04** 1,01 1.04** 1,06 1.04**

(0.01) (0.01) (0.01) (0.01) (0.06) (0.01)

Maximum education in group 1,01 1,01 1,01 0,98 1.07 1,01

(0.01) (0.01) (0.02) (0.02) (0.09) (0.01)

Average land (Acres) 0.95**

(0.02)

Average age 0.99**

(0.01)

Average total children 0,98

(0.05)

Average separated 0,92

(0.23)

Concurrent pradan shgs 1.02** 1.03** 0,99 1.03** 1,07 1.02**

(0.01) (0.01) (0.01) (0.01) (0.05) (0.01)

Raigarh 1.14** 1,03 1,05 1.31** 0.81* 1,65 1,08

(0.05) (0.05) (0.07) (0.17) (0.1) (0.82) (0.06)

Dissolved 1.36** 1.49** 1.4** 2.1** 1,3 0,92 1.47**

Number of observations 15573 15529 8064 2781 4432 252 15529

Number of Departures 2009 2004 1091 390 489 34 2004

33

Table 11: Hazard Rate Estimates at Member Level for Individual Caste Categories:Weibull

model (Restricted Sample)

ST SC BC FC

Shape parameter 0.78 0.78 0.78 1.12

One caste category 1.48** 0,8 0,92

(0,17) (0,29) (0,27)

Member of majority caste category 1.26** 1,12 1,08 0.05

(0,12) (0,21) (0,18) (0,31)

Member of minority caste category 1.35* 1,19 1.66** 0

(0,24) (0,27) (0,32) (0,04)

Concurrent pradan shgs 1,01 0,99 1.06** 0,96

(0,01) (0,02) (0,02) (0,2)

Number of Observations 6124 1590 2852 75

Number of Departures 749 195 284 7

34

Table 12: Hazard Rate Estimates at Member Level for Individual Castes/Tribes : Weibull

model (Restricted Sample)

Average Shape One caste Member of majority Member of minority Concurrent Education Number of

education parameter category caste category caste category pradan shg observations

(failures)

ST

Bhuyians 0.88 0.78 0,89 0.59* 2.86** 0.9** 0.9* 1021

(0,26) (0,17) (1.15) (0,04) (0,05) (124)

Kharia 0.95 0.91 5.96* 1,46 0,31 1.16** 0,93 327

(6,59) (1,55) (0,47) (0,04) (0,08) (53)

Ho 0.89 0.71 1,24 2.42** 1,89 1,07 0,95 375

0.61 0.95 1.32 0.05 0.08 (55)

Munda 1.10 0.88 2,02 2** 0,63 1,03 0,88 350

(0,93) (0,81) (0,51) (0,07) (0,08) (59)

Santhals 1.43 0.66 1,65 3.4** 6.84* 1.12** 1 380

(1,01) (1,59) (7,89) (0,06) (0,06) (38)

Bathundi 2.34 0.67 0,58 0,83 0,42 1,15** 0,9** 627

(0,29) (0,29) (0,44) (0,06) (0,05) (59)

Gond 2.74 0.82 0,89 1,27 2,15* 0,92* 0,98 695

(0,57) (0,4) (0,97) (0,04) (0,04) (66)

Ganda 2.98 0.92 0,61 0,76 0 0,94 0,82** 368

(0,48) (0,36) (0,001) (0,07) (0,07) (35)

SC

Harijans 3.20 0.85 2,19 0,14** 0,96 1,1 1,2** 260

(2,53) (0,08) (0,52) (0,07) (0,07) (29)

BC

Yadav 1.48 0.96 2,1 0,95 1.06 0.92 296

(1,92) (0,43) (0.05) (0.07) (27)

Mahanta 5.23 0.74 0,6 2,13** 1,28 1,13** 0,93** 675

(0,5) (0,78) (0,78) (0,04) (0,03) (63)

Kurmi 5.05 0.78 1,3E+13 0,77 0,55 1,13** 1,0 399

3,25E+16 (0,26) (0,33) (0,06) (0,04) (54)

35