(c) Approach to sampling

We preferred to employ the stratified random sampling methodology to encompass all the
spheres of FI in India. Efforts were made to cover all the different categories of FI approaches (i.e.,
through microfinance methods as well as no frill account approach of banks) and also in terms of their
age, gender, social backwardness, and other related traits. For the purpose of analysis and in order to
comprehend the impact on the socially backward groups, broadly three categories were made viz.,
SC/ST including scheduled castes 2 and scheduled tribes, 3 OBC—including other backward classes 4 and
GEN including social groups other than SC/ST and OBC groups. Preferably, self-help groups under
microfinance who were involved in FI involving members of the age group of 30–50 years were
preferred. As NABARD provides the details of the progress of microfinance region wise in India by
classifying the entire country into six regions viz., northern region, northeastern region, eastern region,
central region, western region and southern region, efforts were made to cover the sample of
beneficiaries across different regions duly involving the priority states 5 under the SHG-Bank linkage
program.

The sample states are considered based on the level of progress in terms of bank-linkage,
poverty level, HDI, and GDI (refer Table 1). About 31% of our sample comes from priority states.
However, FI both quantitatively and qualitatively is skewed in favor of the southern region in view of the
active involvement of non governmental organizations, banks, and various government departments
and organizations. As mentioned earlier, FI is also rigorously pursued by RBI by exhorting the banks to
open no frill accounts to the poorer sections of the society by offering collateral free loans, overdraft
facilities, remittances, and payments under government sponsored schemes, etc. To capture this impact,
covering all the regions as explained above, a sample frame was constructed. About 44% of the sample
comes from the priority states.

The sample was categorized (similar to that of SHG based FI) in terms of SC/ST, OBC, and GEN
categories to maintain homogeneity of classification and analysis. Further, a master sample frame (Table
2) is developed by merging both the sample frame of SHG based FI and no-frill account based FI. Since
no-frill account based FI took off in a rigorous manner from 2006 onward about 7% our sample comes
from no-frill account based FI and the remaining is from the well matured approach of SHG based FI. The
sample represents active, women members all bank linked (with a bank loan) before March 2007.

The size of the sample for the entire study is 1052 including 526 control group beneficiaries and 263
mencomparison group beneficiaries. Three percent of our sample comes from northern region.
Similarly, 3% is from northeastern region, 13% from eastern region, 8% from central region, 5% from
western region, and the remaining 68% comes from the southern region. Further, 30% of the sample
comes from scheduled castes/scheduled tribes category, 40% comes from other backward castes, and
the remaining 30% comes from the general category. The sample size and proportionate samples from
different regions are estimated in proportion to the prevalence of financial inclusion activity of those
regions. Stratified random sampling approach was employed to choose the villages in the selected
districts, which were again selected randomly. Randomization has been growing in popularity because if
it can be implemented properly, randomization can give a robust indication of program impact
(Khandker, Koolwal, & Samad, 2010). While selecting the households too stratified random sampling
was employed to include all the categories of sampling such as scheduled castes/scheduled tribes
category, other backward castes, and general category. While Questionnaire-1.A designed with about 26
parameters was employed to collect the quantified data, Questionnaire-1.B was used to collect the
qualitative data from the program participants and non-participants during the field visits.
Questionnaire 2 was developed to deduce the expert opinion of microfinance practitioners [NGOs].
Questionnaire 3 was used to bring forth the expert opinion of the linkage bankers. In the interest of
space, the questionnaires are not provided with this paper, but would be made available on request.

(d) Use of control group for establishing “attribution”

The impact of developmental interventions is often coupled with the impacts due to other observable
and unobservable interventions and therefore is difficult to measure and identify. It is, as such, difficult
to isolate the actual effect, which may be the result of many different external factors. It is also difficult
to identify the causality. The selection of control groups is important for the issue of
causality/“attribution” of impact. While collecting “before and after data” exclusive reliance is on survey
data to establish a very good counterfactual. In impact evaluation we need to consider the treated
person, household, village, or association to have the same characteristics as the person, household,
village, or association and the only difference is the participation in the program. Because if there is no
other reason for the differences between the two groups, if we see differences in the two groups, then it
is due to stated intervention. Accordingly, standards similar to the ones in selecting the sample were
adopted in selecting the control group based on information about income level, neighborhood factors,
age, social class/caste, region, and religion. Nearest neighbor matching method instead of the
propensity score is employed for selecting the most comparable control group/ comparison group due
to its advantages. Selection of comparable units is made based on matching as many relevant
parameters (like income level, asset holding, social class, education levels, occupation, family size, and
others) as possible and the size of the control group is equivalent to that of the sample.

Subsets of the treatment sample were dropped if similar comparison units were not available (Ravallion,
2008). Heckman, Ichimura, and Todd (1997) encourage dropping treatment observations with weak
common support as only in the area of common support can inferences be made about causality. It is
ensured that the units in the treatment group are similar to those in the control group by observing that
the expected value of the dependent variable is statistically identical BEFORE exposure to the treatment
variable. Further, it was also ensured that not only the “level” but also the “trend” of income of the
treatment groups were similar to that of the control groups. In the case of women, the trend of annual
income was of the order of (CAGR) 2.36% and (CAGR) 2.79% in the case of men. With this approach, any
baseline differences in the dependent variable between the treatment and control groups BEFORE
exposure to the treatment have been eliminated. The upshot is that any differences in the mean level of
the dependent variable between treatment and control groups AFTER exposure to treatment could be
precisely attributed to the treatment variable. This eliminates the possibility that an unobserved factor
correlated with the units in the treatment group, and also correlated ex ante with a higher level of the
dependent variable, is driving any difference in the average outcome between both groups observed
AFTER exposure to treatment.

(e) Attrition issues

In order to do away with the problem of possible differential attrition between treatment and
comparison owing to differences between study participants (both treatment and comparison) due to
those who were present at the pretest and absent at the post-test, it was decided to consider only those
participants who have been in the program of FI all through the study period. The threat of differential
attrition for treatment and comparison is nullified by carefully selecting such beneficiaries and non
beneficiaries of the program in designing the sample and the control group.

(f) Comparison of impact with that of men

To make the analysis more illustrative and comparative in terms of gender effects, it was decided to
consider an equal size sample of men under FI. In selecting this men sample, standards similar to the
ones in selecting the sample were adopted based on information about income level, neighborhood
factors, age, social class/caste, region, and religion. Nearest neighbor matching method instead of the
propensity score is employed for selecting the most comparable comparison group due to its
advantages.

(g) Difference-in-differences estimator

It has been claimed that controlling for selection on observable characteristics may not be sufficient
since remaining unobservable differences may still lead to a biased estimation of treatment effects.
These differences may arise from differences in the benefits which individuals expect from participation
in a treatment, which might influence their decision to participate. These features are unobservable to a
researcher and might cause a selection bias. To account for selection on unobservables, Heckman,
Lalonde, and Smith (1999) suggest econometric selection models and difference-in-differences (DID)
estimators. The DID estimator requires access to longitudinal data and can be seen as an extension to
the classical BAE. Whereas the BAE compares the outcomes of participants after they participate in the
program with their outcomes before they participate, the DID estimator eliminates common time trends
by subtracting the before-after change in nonparticipant outcomes from the before-after change for
participant outcomes. The simplest application of the method does not condition on X and forms simple
averages over the group of participants and non-participants. Changes in the outcome variable Y for the
treated individuals are contrasted with the corresponding changes for non-treated individuals
(Heckman, Ichimura, & Todd, 1998):

DDID 1⁄4 1⁄2Y1t Y 0t0 jD 1⁄4 1Þ ðY0t Y 0 t0 jD 1⁄4 0Þ: ð1Þ

The DeltaID estimator is based on the assumption of time invariant linear selection effects. The critical
identifying assumption of this method is that the biases are the same, on average, in different time
periods before and after the period of participation in the program, so that differencing the differences
between participants and non-participants eliminates the bias (Heckman et al., 1998). Allowing for the
determinants of income of a participant to be factored in while estimating the impact, let us consider for
the ith person in the sample, which is given by:

Yi 1⁄4 a þ bPi þ cXi þ ei; ð2Þ

where a, b, and c are parameters, X stands for the control variables, such as production levels, asset
creation, expenses toward food security, and living standards of the person, while is a residual that
includes other determinants of income and measurement errors. The estimated value of b gives you the
impact of the FI program on the beneficiary. We need to note here that if the ith person participates in
the FI program then P = 1 and so its impact on her income would be a + b + cXi + ei. If it does not
participate, then P = 0 and so its impact on her income would be a + cXi + ei. The difference between the
two is the net impact of the program, which is just b. Now, the ith household in the equation for Yia is
the same household as the ith in the equation for Yib. We can then take the difference between the
“after” equation and the “before” equation:

Y ia Y ib 1⁄4 bPi þ cðXia XibÞ þ lia lib: ð3Þ…………………………………….3

The above Eqn. (3) is an OLS and gives an unbiased estimate of the program’s impact. The unobservables
the ones correlated with program participation have been eliminated. The regression coefficient on the
interaction effect between the participation dummy variable and time in Eqn. (3) identifies the impact.
Further, we need to notice that Eqn. (3) does not require a balanced panel. There can also be situations
in which the changes over time in the outcome indicator are influenced by the initial conditions. Then
one will also want to control for differences in initial conditions. This can be done by simply adding Xa
and Xb in the regression separately, so that the regression takes the form:

Y ia Y ib 1⁄4 bPi þ cXia þ cXib þ lia lib: ð4Þ

With this equation, one can allow X to affect changes over time in the duration of program participation
by the ith person. Binswanger, Khandker, and Rosenzweig (1993) used this method to estimate the
impacts of rural infrastructure on agricultural productivity in India, using district-level data. Duflo (2001)
estimated the impact on schooling and earnings in Indonesia of building schools. Frankenberg, Suriastini,
and Thomas (2005) use a similar method to assess the impacts of providing basic health care services
through midwives on children’s nutritional status (height-for-age), also in Indonesia.

Galiani, Gertler, and Schargrodsky (2005) used a DD design to study the impact of privatizing water
services on child mortality in Argentina, exploiting the joint geographic (across municipalities) and
Intertemporal variation in both child mortality and ownership of water services. Jacoby (2002) used a DD
design to test whether intra household resource allocation shifted in response to a school-feeding
program, to neutralize the latter’s effect on child nutrition. Another example can be found in Pitt and
Khandker (1998) who assessed the impact of participation in Bangladesh’s Gramin Bank (GB) on various
indicators relevant to current and future living standards.

Since the extreme values (outliers) may distort the effects and reduce precision, the outliers are handled
in accordance with the most established methods in the fields of economic research and econometric
methods. Accordingly, a careful assessment of outliers is made for each analysis and each instrument
and employing the extreme data exclusion method, such values are removed from the data employed
for analysis after having assessed whether this has any effect on the results.

(h) Endogeneity of FI program participation

Studies examining impact of program participation routinely suffer from possible bias due to
endogeneity of decision to participate in the program and the unobserved household characteristics.
The unobserved heterogeneity between the members of the FI program beneficiary group and control
group members includes the unobserved attitudes and characteristics of the husbands, wives, and other
family members, including preexisting women’s autonomy. It seems quite probable that independent
women are more likely to be able to join a FI program but these women are also more likely to spend
their time in activities that are better remunerated and of higher status. If this unobserved
heterogeneity is not statistically accounted for then their effects will be captured by the variable
measuring program participation and cause exaggeration of its statistical effects. For gaining traction on
the problem, another one way of addressing the potential for endogeneity bias is to use instrumental
variables. This is an alternative to attempting to identify and control for all possible factors that might be
correlated with both the treatment variable and the outcome when random assignment was not used to
allocate units to the treatment group. The logic of an instrumental variable is that it is not correlated
with these alternative factors whatsoever, and in fact, it must only be correlated with the independent
variable to qualify as an instrumental variable. The instrumental variable works exclusively through the
independent variable to affect the dependent variable which is called the exclusion restriction.
Accordingly, we estimate the impact duly addressing the issue of endogeneity as explained above.

Selection of relevant variables is carefully made considering their prior use and reliability demonstrated
for all of the measures. The key indicators (based on ex ante evaluations of the program) are: Change in
income levels (“Income level” indicates the income in the hands of the beneficiary remaining for
household consumption and further investment), Impact on food security (expenses toward “food
security” indicates the expenses toward all those household expenses toward meeting all the food
related needs), Impact on standard of living (expenses toward “standard of living” mean all those non-

food expenses of the household toward meeting the consumption needs in maintaining the standard of
living such as; health, education, day-to-day expenses of home making and quality of life, etc.), Impact
on production level (overall economic production level of the household from all the income generating
economic activities out of which household income is generated), and Impact on asset creation (“asset
creation level” means investments toward creation of income generating or other related assets for the
household apart from expenses toward food security and standard of living).

5. RESULTS AND DISCUSSION ON FINDINGS

Employing the above described econometric approach and using the above explained variables of study,
the econometric specification of the study is as under:
 INia INib 1⁄4 bPi þ cðFSia FSibÞ þ cðLSia LSibÞ
þ cðPLia PLibÞ þ cðACia ACibÞ þ lia lib; ð5Þ

where INia INib represent the change in income levels, FSia - FSib represent the change in expenses
toward food security, LSia LSib represent the change in expenses toward standards of living (non-food
expenses), PLia PLib indicates the change in economic production levels and ACia ACib indicates the
change in economic asset creation for the house holds. Pi represents the dummy for participation in the
FI program and lia lib stands for the error term. We have run eight regression models that are detailed
with technical specifications as below:

Reg 5.1 and Reg 5.3 are run with Least Squares (NLS and ARMA) with Newey West Heteroskedasticity
consistent coefficient covariance with variance as the weight with maximum 500 iterations and
estimated with HAC standard errors & covariance (Bartlett kernel, Newey-West fixed bandwidth = 5.0).
Reg 5.2 and 5.4 are estimated with Generalized Methods of Moments using standard errors &
covariance computed using HAC weighting matrix (Bartlett kernel, Newey West fixed bandwidth = 6.0)
with maximum 500 iterations.

(a) Robustness checks

To offer the evidence of structural validity robustness tests are conducted by carrying out different
iterations under regression analyses to arrive at the acceptable levels of goodness of fit indicators for
the models tried out in the analyses. The conditional exogeneity relation is key to determining the
critical core variables, i.e., those variables whose coefficients should make economic sense and be
robust. The regression coefficients were consistently estimated by ordinary or generalized least squares
under standard conditions. The efficiency of the estimators used in the robustness check was improved
by using generalized least squares (GLS) instead of OLS. Durbin Watson statistic was calculated for
different regressions detailed above to ensure absence of autocorrelation. The variance inflation factors
are ascertained to assure ourselves about the absence of multi collinearity in the case of OLS
regressions.

(b) Analysis of impact on income levels

Table 3 presents the pooled regression results of estimating the impact of FI programs in the case of
women as well as men participants. On the expected lines the increase in production levels has
positively and significantly (at 1% level of significance) contributed for the increase in income levels of
the women participants. However, it is interesting to note that this is not significant in the case of men
beneficiaries. Further, mostimportant outcome of the analysis is that participation in the FI programs
has impacted very significantly (at 1% level of significance) and positively in increasing the income levels
of the participants. The coefficient of participation is 4830 in the case of women and 4018 in the case of
men which denotes the change due to participation in the FI programs. The goodness of fit indicators in
the different regression estimations are found to be at acceptable levels. We notice that there has been
a significant change in the economic parameters of the participants (Table 4). All the figures are
reported net of inflation effect so as to capture the real impact (Figure 1).
In order to quantify the actual impact essentially due to the participation in FI programs is presented for
easy comprehension in Table 5. While the change in income level has been to the extent of INR
11,523.55 during the period of participation, the impact particularly due to participation in FI programs
is to the extent of INR 10,430.22 in the case of women participants. However, in the case of men
participants, the change was INR 12,860.28 during the period of participation and the impact mainly due
to participation in FI programs is to the extent of INR 8787.55. We have provided the relevant figures
net of inflation figures for comprehension of the impact.

Our results suggest that FI programs have greater impacton women than on men even though the
income levels ofmen are quite higher than that of women. The impact among women is to the extent of
9.91% as against 4.81% in the case of women. Even though the percentage total change during the
period was 66.68% (CAGR 6 10.76%) for women and 38.69% (CAGR 6.76%) for men, the percentage of
change due to impact of FI programs for men was 60.36% (CAGR 8.40%) and 26.46% (CAGR 3.97%) for
men. The results are comparable to the results of Miriam and Love (2009), who in their World Bank
study employed difference in difference strategy to examine the effects of providing financial services

to low income clients and found that average income went up by about 7% as the employment
increased by 1.4%. AIMS study by Chen and Snodgrass (1999) too found that average expenditure on
food in SEWA Bank borrower households was 21% higher than in control households. However, Mont
gomery and Weiss (2011) in their Pakistan study employed difference in difference with controls,
exploiting the staggered introduction of bank services across villages covering 1454 Khushhali Bank
clients with an equal number of randomly selected non-clients from the same villages or settlements
observed that there was no evidence of effects on income growth. Similarly, Roodman and Morduch
(2009) in their Bangladesh study employed difference in difference method with controls and two-stage
least squares regression and stated that they do not find a positive effect of women or men’s borrowing.
On the other hand, Pitt et al. (2006) in their retrospective panel study in Bangladesh covering 2074
households observed that credit to women significantly increases women’s purchasing power and role
in finance and borrowing, ability to oversee household projects.

(c) Income analysis of SC/ST categories

Participation of SC/ST women in FI programs has significantly contributed to the increase of net
household incomes. The analysis of net household incomes for both pre-FI and post FI stages, revealed a
significant growth of (CAGR) 12.01% and an average INR 7771 increase in net household income. When
compared to the control groups, there has been an increase of INR 5349 in the case of women
beneficiaries under FI programs. However, among the control groups of the same category the increase
was to the extent of (CAGR) 4.52% with an average INR 2498 increase in net household increase. In the
case of SC/ST men participating in FI programs, the increase in income was to the extent of (CAGR)
7.82% and the increasing difference in household income in the case of men beneficiaries as against
women beneficiaries was on an average of INR 4384.