WPS6399


Policy Research Working Paper                        6399




  Measuring the Effect of a Community-level
Program on Women’s Empowerment Outcomes
                         Evidence from India

                                Eeshani Kandpal
                                  Kathy Baylis
                              Mary Arends-Kuenning




The World Bank
Development Research Group
Poverty and Inequality Team
April 2013
Policy Research Working Paper 6399


  Abstract
  This paper uses primary data from rural north India                               variables and truncation-corrected matching on primary
  to show that participation in a community-level                                   data to disentangle the program’s mechanisms, separately
  female empowerment program significantly increases                                considering its effect on women who work, and those
  access to employment, physical mobility, and political                            who do not work but whose reservation wage is increased
  participation. The program provides support groups,                               by participation. The analysis also finds significant
  literacy camps, adult education classes, and vocational                           spillover effects on non-participants relative to women
  training for rural women in several states of India; the                          in untreated districts. It finds consistent estimates for
  data are from Uttarakhand. The paper uses instrumental                            average treatment and intent to treat effects.




  This paper is a product of the Poverty and Inequality Team, Development Research Group. It is part of a larger effort by
  the World Bank to provide open access to its research and make a contribution to development policy discussions around
  the world. Policy Research Working Papers are also posted on the Web at http://econ.worldbank.org. The author may be
  contacted at ekandpal@worldbank.org.




         The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development
         issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the
         names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those
         of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and
         its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent.


                                                       Produced by the Research Support Team
Measuring the Effect of a Community-level Program on
Women’s Empowerment Outcomes: Evidence from India
          Eeshani Kandpal, Kathy Baylis and Mary Arends-Kuenning




   JEL Codes: D13; I24; J16; O15; O17
   Keywords: Women’s empowerment; rural community-level interventions; impact evaluation; India
   Sector: GEN, ARD, POV


    Kandpal: World Bank, 1818 H Street NW, Washington DC 20433 (e-mail: ekandpal@worldbank.org).
Baylis: Agricultural and Consumer Economics, University of Illinois at Urbana-Champaign, 1301 W. Gre-
gory Drive, MC-710, Urbana, Illinois 61801 (e-mail: baylis@illinois.edu). Arends-Kuenning: Agricultural
and Consumer Economics, University of Illinois at Urbana-Champaign, 1301 W. Gregory Drive, MC-710,
Urbana, Illinois 61801 (e-mail: marends@illinois.edu). We are grateful to Yusuke Kuwayama, Jed Friedman,
Nolan Miller, Alex Winter-Nelson, Don Fullerton, Craig McIntosh, Alain de Janvry, Thomas Walker, Chris
Barrett, Phil Garcia, Laura Schechter, Carl Nelson, Catalina Londo˜ no, Sahan Dissanayake, Ben Wood, and
participants of the NEUDC 2010, PAA 2011, AAEA 2012, and AEA 2012 conferences. We thank Sumita
Kandpal, Geeta Gairola, Basanti Pathak, Preeti Thapliyal, Ravi Mehta and the entire Mahila Samakhya
Uttarakhand family for their support. Kandpal acknowledges with gratitude the �nancial support provided
by the Goodman Fellowship and the Due Ferber International Research Award of the WGGP, the College
of ACES AYRE Fellowship, and the Survey Research Laboratory’s Seymour Sudman Dissertation Award.
Baylis thanks the University of Illinois Research Board for the Arnold O. Beckman award.

                                                   1
1    Motivation
Empowerment allows individuals to reach their full potential, to improve their political and social
participation, and to believe in their own capabilities. Gender empowerment also has important
rami�cations for the rest of the household; empowered women have fewer children and higher child
survival rates (Rosenzweig and Schultz, 1982; Dyson and Moore, 1983), healthier and better-fed
children (Lundberg, Pollak and Wales, 1997; Kanbur and Haddad, 1994), and a generally greater
allocation of resources to children (Thomas, 1990; Handa, 1996). Development programs aim to
empower women by increasing their control over contraceptive choices, by providing them access
to credit, and through education.
    Women’s empowerment is particularly hard to achieve within a generation because it is
driven not only by information about choices, but also by the acceptability of these choices, and
what Kabeer (1994) calls the “power within�, i.e. a woman’s intrinsic belief in her ability to control
resources and to make decisions. Communities are often governed by strict social norms, which
can both be driven by and drive the choices traditionally made by women in the village. If the
social stigma associated with working outside the home or using contraceptives is prohibitive, then
mere access to education or birth control may not change empowerment outcomes. Instead, giving
a woman access to others who have made different choices can expand her information set and
demonstrate the outcomes associated with these choices. As an alternative to targeting individual
women, empowerment for women may be affected by combining learning and influence through
community action and peer networks.
    In this paper, we use primary data from rural north India to examine the impact of a program
called Mahila Samakhya on female empowerment outcomes. Mahila Samakhya aims to empower
women by educating them. The program provides literacy camps, adult education classes, and
vocational training. The program also creates support groups on issues of social importance, such
as domestic violence and alcoholism. These support groups strengthen networks, which plays an
important role in determining empowerment. We measure empowerment using (1) the ownership
of identi�cation cards for the national government’s rural employment guarantee scheme, which
proxies for access to outside employment, (2) the ability to leave the household without permission,
which reflects physical mobility, and (3) participation in weekly village council meetings, which
measures political participation. The literature identi�es access to outside employment, physical
mobility, and political participation as three important components of gender empowerment. These
variables represent a wide variety of domains in which a program like Mahila Samakhya can empower
women: economically, domestically, and socially.
    Mahila Samakhya is an innovative approach to improving female empowerment. While a number
of programs aim to improve female empowerment through education, Mahila Samakhya combines
education with support groups, and has the explicit objective of increasing gender empowerment.
We posit that this program affects female bargaining power in two ways. First, education provided
by the program directly improves job prospects and increases the reservation wage. By increasing
a woman’s earning potential, the program helps improve her bargaining position, allowing her to
control a greater share of the household’s resources and to become a more active participant in
her community. Further, the program may have an indirect effect through improved information
flows that may change social norms. These social spillovers also empower participants who do not
have access to outside employment and thus do not bene�t from the direct employment aspect of
Mahila Samakhya (Montgomery and Casterline, 1996). As a result, even unemployed participants
and non-participants may be empowered by Mahila Samakhya.


                                                  2
    In establishing whether Mahila Samakhya has a signi�cant impact on female empowerment, we
account for two potential sources of endogeneity: (1) the program’s choice of communities in which
to operate, and (2) the individual’s choice of whether to participate. We conduct this analysis in
�ve stages. First, we estimate an intent-to-treat (ITT) effect of the program by comparing the
bargaining power outcomes of treated and untreated women. Second, we consider the potential of
differences among women in communities that receive the program and those in communities with-
out. We match non-participants in treated districts (“non-participants�) and women in untreated
districts (“the untreated�), to examine whether they are signi�cantly different from each other. We
also compare census data on indicators of female bargaining power in the blocks where the program
was placed to the blocks where it was not placed from before the Mahila Samakhya intervention to
�nd no signi�cant differences between treated and untreated blocks. All of our results suggest that
the program was not targeted in its placement.
    Third, we test whether program participants are signi�cantly more empowered than similar
women from untreated districts to determine whether the program has a signi�cant treatment effect.
Using instrumental variables and matching, we compare pre-determined levels of empowerment and
empowerment outcomes (similar to a differences-in-differences approach) to �nd that even after
accounting for the pre-determined level of empowerment, the program signi�cantly improves access
to employment, physical mobility, and political participation. Our instrument relies on the roll-out
of the program to control for selection in the participation decision. Using both approaches, we
�nd a positive, signi�cant treatment effect of the program on women’s empowerment outcomes;
participants are more likely to have access to outside employment, are able to leave the house
without permission and are more likely to attend village council meetings, although this last effect
is not signi�cant in all speci�cations. The marginal effects from the IV approach emphasize the
potential for large numbers of women to bene�t from interventions like Mahila Samakhya. These
results are qualitatively similar to the ITT effects estimated, and suggest that our instrument
adequately controls for sample selection.
    Fourth, to explore the program’s mechanism, we focus on participants who do not work, com-
paring them to untreated women who also do not work. Using both instrumental variables and
matching, we �nd that even participants who do not bene�t from the enhanced employability from
participation are signi�cantly more likely to leave the house without permission. We also consider
those women who participate in MS but do not have access to outside work. Results suggest that
participants who do not have access to outside employment are more likely than non-participants
without access to outside employment to leave the house without permission and to participate in
the village council. Thus, whether through increasing her reservation wage or by changing social
norms, MS appears to improve outcomes even for those women not able to bene�t directly from
the intended increase in own income.
    Fifth, we we attempt to isolate the influence effect of MS by comparing non-participants to
untreated women to test for the presence of spillover effects in treated areas. We are interested in
knowing whether MS can affect non-participants in a village by changing social norms. We �nd
that non-participants in treated villages have greater access to outside employment, greater physical
mobility and higher attendance of village council meetings than untreated women, which point to
the positive spillover effects of Mahila Samakhya.
    Most studies of program impact analyze interventions targeted at the individual. Only a small
number of papers examine community-level interventions because these programs often aim to
change outcomes that are difficult to measure and use methods that combine direct individual in-
tervention (education) with the process of the intervention (community meetings). Thus, evaluating


                                                 3
community-level programs often poses the dual difficulty of imprecisely-measured outcomes and a
“treatment� that is hard to identify. In this paper, we use changes in outcomes and an instrument
suggested by detailed �eld tests to disentangle the mechanisms of the program and identify the
effect of the program on female empowerment.
    Few other papers study Mahila Samakhya ; one of these papers focuses on the program’s effect
on building village-level social capital and trust rather than studying its intended impact on female
empowerment (Janssens, 2010). In other work, Kandpal and Baylis (2013) explore whether Mahila
Samakhya affects the diversity of participants’ peer group, while Kandpal and Baylis (2011) study
whether peers’ participation improves female bargaining power and child welfare outcomes; neither
of these studies explicitly estimates treatment effects of the program’s intended impact, which is to
empower women through participation.
    This paper contributes to the literature in several ways. It is the �rst to evaluate the impact of
Mahila Samakhya on empowerment outcomes and provides robust estimates of the various effects
of this program. We decompose the effect of community-level interventions like Mahila Samakhya
into its three components: (1) a direct effect that works through employment opportunities outside
the family farm, (2) a direct effect that works through higher reservation wages, and (3) an indirect
effect that works through the channels of social influence and learning. As a result, this paper
may provide valuable evidence on the effectiveness of community-level interventions in changing
ingrained social outcomes like women’s bargaining power. Our results establish that the Mahila
Samakhya program directly and indirectly increases female empowerment.


2     Background on Uttarakhand and Mahila Samakhya
Uttarakhand is a a small rural state in the Indian Himalayas, comprising less than one percent
of the Indian population. Only �ve cities in the state contain more than 100,000 people. On the
surface, Uttarakhandi women may appear to be more empowered than the average. These women
led the Chipko movement to prevent deforestation1 as well as the demand for a separate state.
However, while the state has a literacy rate of 72 percent, the Census reports that only 60 percent
of all women are literate.2 A more detailed measure of literacy from a nationally representative
household survey �nds 43 percent of Uttarakhandi women cannot read at all, while an additional 5
percent can only read parts of a sentence (International Institute for Population Studies and Macro
International, 2007). Therefore, the effective literacy rate for females may be closer to 50 percent.
    Although 43 percent of all Uttarakhandi women work, nearly two-thirds of these women (64
percent) are not paid for their work, and over 70 percent work in agriculture. These women are
likely to work on their family’s farmland, which does little to empower them. In addition, 23 percent
of Uttarakhandi women have no say over household �nancial decisions, and almost 43 percent do
not have the �nal say on their own healthcare. Over half (55 percent) did not have the �nal say
on large purchases made by their household (International Institute for Population Studies and
Macro International, 2007). Hence, Uttarakhandi women can have little say in the household or
community.
    In 1988, Mahila Samakhya was launched in three states of India to empower women through
formal, informal, and vocational education. In theory, the community-level program was placed
   1 The Hindi word Chipko means “to stick�. In the Chipko movement of the seventies, Uttarakhandi villagers, and

women in particular, literally hugged trees to prevent deforestation.
   2 The national literacy rate is 65 percent, and female literacy of 54 percent. The most literate state is Kerala,

with a 91 percent overall literacy rate and 88 percent female literacy rate.


                                                         4
in districts identi�ed by (1) low rates of female education, (2) low school attendance by girls, (3)
remoteness, and (4) lack of development and restricted access to infrastructure. In practice, as our
results will highlight, the program does not appear to be targeted in any way, and the selection
of districts into the program appears random. Participation in the program is voluntary, and no
monetary incentives are offered.3 The program entered Uttarakhand in 1995 and covers 2,416
villages in six of 13 Uttarakhandi districts. More than 42,000 women participate in this program,
and over 2,500 girls have been educated in its centers.
    The DFID-funded Mahila Samakhya conducts biweekly literacy camps and provides continuing
education to women and girls. The camps and continuing education classes are provided to a
cluster of three to �ve villages, depending on the size of the class and the proximity of the villages.
The program also offers weekly vocational training to enable participants to earn an income. Such
training is intended to improve the participant’s employability, giving her access to job opportunities
off the family farm, and hence improving her level of empowerment in the household and the
community. Indeed, in pre-tests, we frequently observed that participants used this training to
become midwives, herbal medicine manufacturers, bakers, grocers, candle makers, and tailors. In
addition, the program provides special education on resolving domestic disputes and conflicts within
the community. All classes are capped at 25 women, with an average size of 17 women per class or
camp.
    The program also encourages women to participate in village politics as a means of self-
empowerment. In �eld tests, we observed participants hearing about the success women have
had in the labor force and the important roles women can play in Indian society. They were also
told about the bene�ts of having a daughter and of not discriminating against her. Groups of par-
ticipants that meet on a weekly basis provide support on issues like domestic violence, alcoholism,
dowry, and female infanticide. These groups vary in size from �ve to 15 women, and foster ties
between participants. These interventions have the potential to generate signi�cant spillover effects
wherein even participants who cannot work outside the home or family farm, and therefore do not
bene�t directly from improved employability, can be empowered by their participation.
    Mahila Samakhya enters a village through program workers called sahayoginis. The worker
�rst conducts several rounds of talks with local women to determine what their needs are, and
what they would like from the program. This process can take up to several weeks, but as a
result, the program’s activities are tailored to each village. The program often starts with literacy
or education camps because these are the most frequently-voiced concerns. Initially, only a few
women may participate, but as others see the bene�ts of participation, they muster up the courage
to participate despite family opposition.
    The program can meet with resistance from the men in the village, who may see the program
as subversive and be unwilling to let their wives participate. In such cases, workers stress the
educational rather than empowerment component of the the intervention. Once the men observe the
bene�ts of participation, generally in the form of earnings, they reduce their opposition. Sometimes,
as the women become more mobile, men might again oppose participation, but usually the women
are sufficiently empowered at this point that the opposition no longer restricts their involvement.
   3 When participants travel to district-meetings, they are housed and fed at the program headquarters, and their

travel expenses are reimbursed.




                                                        5
3    Literature Review
 Kabeer (1999, p. 1) de�nes gender empowerment as “the process by which those who have been
denied the ability to make strategic life choices acquire such an ability.� She then decomposes
empowerment into (1) resources, representing access, both present and future, to human and social
resources; (2) agency, including decision-making and negotiating, where the latter can include
deception or manipulation; and (3) achievements or outcomes. We posit that by providing women
the access to outside employment, the Mahila Samakhya program improves the availability of human
resources to women. In addition, by expanding and diversifying their social networks, as well as
expanding their information sets, the program improves access to social resources. The program
improves agency by informing women about the opportunities available to them and by teaching
them how to negotiate, both inside and outside the household, and better equipping them to avail
of the opportunities for outside employment made available to them by the program. Through
this combined impact on resources and agency, the program improves women’s achievements or
outcomes.
    A large body of literature considers the determinants of and proxies for gender empowerment.
Gender empowerment is measured by a woman’s ability to make household decisions, relative to
her husband’s ability to make household decisions. Since this ability cannot be explicitly measured,
economists study whether variables such as education, contraceptive use, and asset-ownership are
correlated with high female empowerment. These self-reported variables reflect the wide variety
of choices and decisions at stake in the household bargain: employment, fertility, and resource
allocation. Empowerment itself has been measured by a woman’s relative physical mobility, eco-
nomic security, decision-making ability, freedom from domestic violence, and political awareness
and participation.
    Another section of the literature �nds that a woman’s access to employment outside the house
increases her household bargaining power (for a study in Bangladesh, see Anderson and Eswaran,
2009; for a study in India, see Rahman and Rao, 2004). The ownership of assets, in particular,
is one important way through which access to employment helps empower women in developing
countries (for example, see Agarwal, 2001, for evidence from India). In addition, several analyses
have found that access to credit programs— micro-�nance organizations or rotating savings and
credit associations (ROSCA)— has a positive effect on female empowerment (for a study in Kenya,
see Anderson and Baland, 2002; for a study in Bangladesh, see Hashemi, Schuler and Riley, 1996).
    Past literature has highlighted the importance of the source of income in terms of its effect on
female bargaining power. For instance, Lilja et al. (1996) distinguish between income from private
plots versus communal household plots in terms of their effect on female intrahousehold bargaining
power. The authors suggest that tradeoffs may exist between the amount earned from private plots
and the amount of wages received from the communal plot. Further, Jones (1983) �nds that if
a woman is not paid or paid too little for her work, as is typical for work done on the household
�eld, she may see no bene�ts to such work for the household bargain, and may even refuse to work
on the communal �eld the following year. Hence, as Lilja et al. (1996) point out, if household
members can influence their per unit labor payment on the communal �eld, then women’s implicit
wages would increase as their labor contribution rises.
    Studies have also found a positive link between empowerment and contraceptive use (for a
study in Bangladesh, see Schuler and Hashemi, 1994), as well as between the woman’s influence
on resource allocation and her family’s social status (for a study in Bangladesh, see Quisumbing
and de la Bri` ere, 2000). In particular, the more educated she and her father are relative to her


                                                 6
husband, the more empowered she is. Hashemi, Schuler and Riley, 1996 �nd that physical mobility
improves the degree of control over microcredit loans, since it reflects a woman’s access to outside
employment opportunities. A study of the determinants of female autonomy in India �nds that
a better-educated woman has greater bargaining power, as measured by physical mobility and
say in household resource allocation, through the channel of increased information (Rahman and
Rao, 2004). The same study also �nds culture, as measured by state �xed-effects, to signi�cantly
increase bargaining power despite controlling for religion and caste. Further evidence from India
shows strong positive correlations between female education as a proxy for bargaining power, and
freedom of movement and better maternal health as bargaining outcomes (see Malhotra, Pande and
Grown, 2003 for a review of this literature).
    The literature further agrees that the clearing of marriage markets depends on the number of
men and women in the market (for theoretical models, see Becker, 1973a,b; Neelakantan and Tertilt,
2008). As a result, the local sex ratio works through the spousal age ratio to influence marriage
markets and therefore household bargaining power. Scholars have found that, particularly in the
Indian context, women have less bargaining power if their husbands are signi�cantly older (for
evidence from India, see Caldwell, Reddy and Caldwell, 1983; Kantor, 2003).
    Since empowerment is an unobservable latent variable, economists use its observable character-
istics as proxies for empowerment. Women with a more equal spousal age ratio, greater access to
outside employment or a high level of political participation are also likely to have greater bargaining
power. Thus, the indicators of a high level of empowerment include (1) access to outside employ-
ment, (2) physical mobility, and (3) political participation (Anderson and Eswaran, 2009; Hashemi,
Schuler and Riley, 1996; Rahman and Rao, 2004). The corresponding dependent variables we use to
reflect high levels of female autonomy are (1) the ownership of identi�cation cards for the national
government’s rural employment guarantee scheme, which proxies for access to outside employment,
(2) the ability to leave the household without permission, which reflects physical mobility, and (3)
participation in weekly village council meetings, which measures political participation. We choose
these variables because they represent a diverse set of ways in which the Mahila Samakhya program
can potentially empower women.
    Employment and access to credit do not always improve empowerment. As Faulkner and
Lawson (1991) demonstrate, if women’s employment is concentrated in inferior positions, then
access to employment alone may in fact make them worse off than before. However, as discussed
above, the Mahila Samakhya not only provides access to employment, it also increases agency and
access to social resources. The study by Rowlands (1997) of two similarly-intentioned Honduran
development programs with very different implementation and outcomes provides some clues as to
what program characteristics might lead to true empowerment. The �rst was a health promotion
program, which trained local women in preventive health care and basic treatment, with the goal
of having these women provide these services for payment in the future. This program had about a
50 percent dropout rate, and outcomes depended heavily on the individual health promoter. Only
some of the women who successively completed the training program reported feeling individually
empowered, although they all now had access to outside income. The second program grew out
of the informal support group-like meetings of a group of local women, in which they discussed
their experiences and problems. The expanded, formalized program worked in much the same way
that Mahila Samakhya does, through an iterative, grassroots implementation process. Almost all
participants reported increased levels of self-con�dence and self-esteem as a result of participation,
and Rowlands (1997) reports that these changes were very noticeable.
    With the exception of analyses of credit extension mechanisms, the studies discussed above focus


                                                   7
on interventions targeted at the individual. Only a small number of papers look at community-level
interventions. For instance, Imai and Eklund (2008) use survey data on a women’s community-
based organization in rural Papua New Guinea to assess the effectiveness of autonomous women’s
groups compared to those that receive external support. Their analysis— using a Heckman Selection
Model as well as Propensity Score Matching— shows that the autonomous groups are more effective
in improving child welfare. Thus, community-level interventions targeted at women can generate
signi�cant bene�ts to children’s well-being.
    One of the few analyses of Mahila Samakhya uses data from the state of Bihar to evaluate
the program’s effect on community-level trust and social capital (Janssens, 2010). The paper uses
Propensity Score Matching to calculate Intent-to-Treat estimates of the program. Matching women
from treated villages to those from untreated villages, results suggest that the program signi�cantly
increases trust and engenders social capital. Participants are more likely to contribute to local
educational and infrastructural community projects. Signi�cant spillovers also exist with non-
participants; non-participant households in program villages exhibit higher levels of trust and are
more likely to engage in community building activities than households in non-program villages.
    Other work examines allied aspects of the program without explicitly evaluating it: Kandpal and
Baylis (2013) ask whether participation affects the diversity of participants’ peer group, while Kand-
pal and Baylis (2011) examine whether friends’ participation improves female bargaining power and
child welfare outcomes. None of these studies considers the causal model behind the mechanisms
of community-level interventions. Next, we present a model that decomposes the effect of program
participation in a community-level intervention.


4     Decomposing the Effect of Program Participation
Participation in Mahila Samakhya can have a direct effect by improving a woman’s job prospects,
and an indirect effect that changes perceived social norms through information spillovers. We
attempt to decompose the effect of participation in Mahila Samakhya into these two effects. Par-
ticipation in the program directly increases the woman’s educational attainment, which in turn
improves her job prospects (Phipps and Burton, 1998), potentially leading her to �nd employment
outside her home. Further, even a participant who does not work may bene�t directly from the ed-
ucation because when bargaining with her husband over household resources, knowing about better
job opportunities and having more marketable skills increase her disagreement utility.4 The direct
effect of the program thus works through employability, skills, and reservation wages, which in turn
affect bargaining power through outside options. The education gained through this program is
therefore expected to raise bargaining power. By contrast, the indirect effect works by expanding
the woman’s information set about alternatives, and therefore affecting her perceived social norms.
    Following Montgomery and Casterline (1996), we can think of a woman facing an optimization
problem in which she chooses a course of action c in order to maximize her expected utility. Let
the set {c1 , . . . , cN } represent the universe of choices that a woman could potentially make, and let
{c1 , . . . , cK }, where K ≤ N , represent the subset of choices that the woman is aware of as being
available to her. The choice that the woman makes as a result of her household bargain leads to one
of several possible outcomes, indexed by the variable j , where each outcome is associated with a
   4 The disagreement utility is each spouse’s intertemporal utility if they remained single or if they were non-

cooperating in marriage, and depends on the spouse’s own earning potential and the partner’s earning potential as
well as on the non-cooperative equilibrium outcome of investment in children.




                                                       8
vector of the “determinants of the woman’s well-being�, Yj (Montgomery and Casterline, 1996, p.
157). Let P (Yj |c, I) represent the probability of the woman experiencing the vector of determinants
Yj , which is conditional on the choice made by her, c, and her information set, I. Then, the vector
of determinants Y1 occurs with probability P (Y1 |c, I), Y2 occurs with probability P (Y2 |c, I), and
so on. The vector Yj thus represents the outcome of the household bargain, given the components
of the woman’s information set. While we do not explicitly model the household bargain here, the
formulation of the optimization problem presented below is fully compatible with a standard Nash
bargaining framework.
    Consider a woman who gets vocational training and then chooses to become a candle maker, cj .
Her decision to engage in this outside employment affects the amount of her household income she
controls, Yj , which is an outcome of her household bargain. The effect of the bargaining outcome
on her individual utility is denoted by U (Yj ); thus, Yj is the outcome of the household bargain,
facilitated by choice cj , and is a determinant of the woman’s utility. Then, the woman’s expected
utility maximization problem is given by:

                                       max               U (Yj )P (Yj |c, I).                            (1)
                                   c∈{c1 ,...,cK }
                                                     j

The woman’s information set I consists of:

                           I = (pt , qt , E(pt+1 ), E(qt+1 ), E(Yt+1 |c), Σ2 , Z),                       (2)

where pt and qt are known current prices and quantities of the goods and services consumed by
the woman. Expected future prices and quantities, E(pt+1 ) and E(qt+1 ), and the expected future
vector of private determinants of well-being, E(Yt+1 |c), have associated variances Σ2 . Vector Z
represents all remaining constraints and costs.
    By educating a woman, Mahila Samakhya improves her household bargaining position and
increases expected lifetime outcomes.5 We can think of this education as adding a choice cK +1 ,
where K + 1 ≤ N , to the woman’s existing choice set. For example, choice cK +1 may be getting
a job that would not have been feasible without the education. We assume that this choice can
only be added by Mahila Samakhya because there is no other opportunity for formal or vocational
education.
    Some women may not select the newly available choice cK +1 . However, the education provided
by Mahila Samakhya gives these women the potential to choose cK +1 , which increases their reser-
vation wage and leads them to directly bene�tting from the program. For example, the addition of
cK +1 to the womans choice set can lead to an increase in P (Yn |cm , I), where cm is a choice that
was available before the educational component and Yn is an improved outcome for the woman.
    The program’s indirect effect works through spillovers from social learning and social influence,
as well as social norms. Participants learn about new opportunities and new behaviors from each
other, which expands each woman’s choice set and can improve her expected utility. So, while par-
ticipation expands choice sets, it also changes expectations of future employment and empowerment
outcomes, given the choices. Changing expectations to reduce the variance of outcomes associated
with previously unknown choices can improve individuals’ expected utility. Assuming that these
women are risk-averse, even if certain choices were available to them, the unknown distribution
of outcomes associated with these choices might discourage them from making these choices. For
   5 Even in cases where education is an irrelevant alternative, the woman is not worse off than before, hence

participation leads to a weak improvement to her household bargaining position.


                                                         9
example, in our �eldwork, we encountered several women who lived in a village where all women
only worked on the farm. As a result, for any individual woman considering working off the farm,
the variance in possible outcomes associated with choosing to work off the family farm may be
large. The earliest Mahila Samakhya participants from these villages told us that when they �rst
considered off-farm employment, they were worried that their families would ostracize them or block
access to their children. However, at Mahila Samakhya meetings they were introduced to women
who do work outside the home, which gave them a more realistic picture of the outcomes from
working outside the home. These women thus realized that the probability of some of the worst
outcomes that they had feared was considerably lower than had previously thought. The fact that
these women worked off the family farm was usually accepted by the families; none of them had
been ostracized, nor had they lost access to their children.
    By influencing the behavior of participants, the program also indirectly affects non-participant
friends of participants. We can think of this indirect effect as working through the non-participant’s
information set. Exposure to participant friends may change the non-participant’s expectations in
two ways: (1) by expanding the set of choices known to her through her network, and (2) by
changing her expectation of future distribution of outcomes, E(Yt+1 |c) as well as the associated
subjective variance, Σ2 , by showing her what happens if she makes a certain choice. Over time,
as more and more participants change their behavior (by becoming educated, getting a job, having
greater bargaining power in the household), the social norm also changes to become less restrictive
on all women, regardless of their participation status. As a result, the program’s indirect effect may
be substantial as the spillovers from the program increase over time.
    Now consider the marginal effect of participation in Mahila Samakhya. If participation changes
the utility-maximizing choice available to a woman from cK to cK +1 , then the marginal effect of
participation on her utility is as follows:

                          Σj U (Yj )P (Yj |cK +1 , I) − Σj U (Yj )P (Yj |cK , I)                 (3)
We assume that having more choices must make participants at least weakly better off, therefore
implying that P (Yj |cK +1 , I) �rst-order stochastically dominates P (Yj |cK , I). Corresponding to
the marginal effect in equation 3, if spillovers change the information set from I to I , then the
marginal effect of participation on utility is as follows:

                             Σj U (Yj )P (Yj |c, I ) − Σj U (Yj )P (Yj |c, I)                    (4)

Since Mahila Samakhya may expose women to information that certain outcomes are in fact sig-
ni�cantly worse than previously thought, �rst-order stochastic dominance of P (Yj |c, I ) does not
always follow. However, assuming that women are risk-averse, the variance-reducing effect of Mahila
Samakhya on the information set, by itself, would increase participants’ utility, even if having more
choices might make participants worse off by making them aware of their subordinate status without
giving them the ability to change it.
    In summary, via the two direct and one indirect effects discussed above, Mahila Samakhya
can expand the woman’s choice set {c1 , ..., cK }, and the newly available choice cK +1 increases the
probability of the woman obtaining a more favorable value of Yj (and decreases the probability of
obtaining an unfavorable Yj ). Further, information can adjust expectations and perceived variance
so as to also adjust the subjective probability distribution of outcomes.




                                                   10
5     Data
5.1     Survey Description
We use primary data on the participation in Mahila Samakhya, female empowerment, child welfare,
and social networks of 487 Uttarakhandi women. The survey and sampling strategy, described in
detail in Kandpal and Baylis (2011), covers 69 villages in six Uttarakhand districts, four with the
program and two without. The villages in our sample were randomly chosen from the six districts.
The sample size is 487 women. The survey was designed to trace self-reported networks, and
hence was implemented using restricted snowball sampling. We �rst randomly sampled a woman
in each village, asked her about �ve people she is in regular contact with, then conducted follow-up
interviews with two randomly-chosen friends from these �ve. We repeated this process once, for
each network. As a result, starting with one randomly chosen woman, we ended with networks of
seven women. The top code of �ve for listing friends was introduced due to budgetary restrictions,
but after considerable pre-testing. With and without a top code (as well as with larger top codes),
we found that most women were not able to name more than �ve people outside the immediate
family who they were in contact with on a regular basis.
    The survey instrument includes the following key questions to help identify the effect of par-
ticipation in the Mahila Samakhya intervention on an individual’s level of empowerment. Female
Empowerment Dependent Variables: (1) Access to Outside Employment: whether the respondent
has an identi�cation card for the National Rural Employment Guarantee Scheme (NREGS). (2)
Physical Mobility: whether the woman can leave the house without permission.6 (3) Political
Participation: whether the woman participates in the local village council. (4) Proxy for Initial
Bargaining Power: the woman’s age relative to that of her husband.
    Participation: (1) Whether the woman participates in the Mahila Samakhya intervention. (2)
How long the program has been in her village. (3) Exposure to the program: the interaction of the
number of years the program has been in her village with the woman’s age minus sixteen.7 Other
Socioeconomic Characteristics: (1) Literacy and educational attainment.8 (2) The number of male
and female children born to the woman and their ages. (3) The woman’s caste: since outcomes like
physical mobility and political participation depend on social class, we �nd it important to control
for the woman’s caste, and wealth, as proxied by (4) the number of rooms in the house and the
primary source of lighting. We also control for (5) the amount of time, in minutes, a woman spends
collecting �rewood each day, which reflects the amount of free time she has to participate in Mahila
Samakhya.9
    The National Rural Employment Guarantee Scheme, NREGS, guarantees at least a hundred
   6 Since this variable is difficult to verify, it might suffer from reporting bias: participants know the “correct answer�

to this question is that they do not need permission to leave the house, and thus might be systematically more likely
to overstate their physical mobility than non-participants. However, in �eld tests, we observed that participants were
signi�cantly more sensitive to their lack of household bargaining power and were likely to underreport the amount
of say they had in the household because the program had made them aware of the entire feasible set of outcomes
for women. Therefore, if we were to expect a sizable reporting bias by participants, it would be in the downward
direction, i.e. participants would be likely to underreport their physical mobility.
   7 We subtract sixteen because women younger than sixteen cannot participate.
   8 We asked participants about literacy and educational attainment prior to participation. Because the program

is not viewed as a school, many participants distinguished between their educational attainment and learning via
Mahila Samakhya even without our asking.
   9 Note that Uttarkhand is a exogamous, patrilocal society, and women typically move to another village after

marriage, and live in their in-laws’ joint family home. Thus, they have little or no say in how far they need to travel
to collect �rewood.



                                                           11
days of paid work to the rural poor. Having an identi�cation card (or having their name listed on
the household card) gives the women access to outside employment. However, program supervisors
sometimes deny women these cards or refuse to add their names to the existing household card
because the work generated by NREGS is of a manual nature, and is thus considered “unsuitable�
for a woman. Mahila Samakhya officers encourage participants to demand the cards, and where
necessary, to report the supervisor to the local administrative officer. Kabeer (2000) raises several
pertinent questions about the true meaning of choice in such a context, including does access
alone represent choice? And, who or what influenced the decision to get access? Not all program
participants choose to get this card, and indeed the process of getting one can be onerous on the
woman, possibly involving several trips to the local and district NREGS offices, and making their
case forcefully. Thus, owning an NREGS card represents a woman’s choice to gain access to outside
employment. While the decision to get a card is undoubtedly based on a confluence of economic
needs and influence by program officers, women also learn about the feasibility and acceptability of
women, in the form of other Mahila Samakhya participants owning these cards. The card not only
represents access to employment, but also access to monetary income, the ability to work off the
family farm, and increased physical mobility to women. The Mahila Samakhya program informs
women of their right to claim these cards and advocates for them to get the cards, because NREGS
officials typically tend to exclude women. However, the program does not guarantee access to
the cards, nor does participation in Mahila Samakhya necessarily mean getting an NREGS cards.
Hence, owning these cards represents access to the information that women are entitled to get
NREGS cards, and the con�dence to demand something not easily obtained, which also correlates
with empowerment. Thus, participation in Mahila Samakhya can give women what Kabeer (2000)
calls the “power to choose� to access outside employment.
    We distinguish between pre-determined empowerment characteristics, like the spousal age ratio,
and characteristics that might be affected by participation, such as owning an NREGS identi�cation
card or participation in village council meetings. Since Mahila Samakhya targets married women,
and none of the women in the sample participated in the program before marriage, the spousal
age gap is not likely to be affected by program participation. Program officials of the Mahila
Samakhya intervention in Uttarakhand told us that women married to much older men have little
say in the household, because often the age gap arises from a second marriage for the man, or some
“undesirable� quality in the woman or her background. Hence, we treat a woman’s spousal age gap
as a proxy for her pre-participation level of empowerment. We cannot rule out the possibility that
an older relative of the woman, say her mother, is a Mahila Samakhya participant and that therefore
the respondent’s age at marriage was not completely unaffected by participation. However, program
participants tend to be young women, and the program only came into the region in 1995, so the
influence of the participation of an older relative on a later participant’s marriage decision is likely
to be minimal.
    The difference between matched pre-determined empowerment characteristics of participants
and untreated women thus provides a baseline level of empowerment for participants. After es-
tablishing that matched participants and untreated women do not have signi�cantly different pre-
determined levels of empowerment, we use the spousal age ratio to control for differences in initial
bargaining power when estimating the effect of participation on characteristics like owning an
NREGS identi�cation card.




                                                  12
5.2     Summary Statistics
Table 1 shows that the average woman in the sample is 32 years old, while her husband is 38 years
old. She married at age 19 and has 9 years of education, one less than her husband. Her sons
are, on average, eight years old, while her daughters are six. Only twenty women reported not
having any children at all; the average number of children is 1.15, with an average age of 7.42. The
average woman’s house has three rooms and electricity. Sons and daughters have, on average, equal
amounts of education; about three years.
    Table 2 indicates that participants are signi�cantly more empowered than non-participants.
While on average, 61 percent of the women in the sample said they had NREGS cards, only
49 percent of non-participants did. In contrast, over 68 percent of participants had these cards.
Similarly, while 71 percent of the sample said they did not need permission to leave the house,
this was true for only 59 percent of non-participants while 78 percent of participants did not
need permission. Finally, while only 14 percent of non-participants reported attending village
council meetings, almost half of all participants did. In summary, whether in the form of access
to employment, physical mobility, or political participation, women who participate in Mahila
Samakhya have higher levels of empowerment. However, these statistics do not tell us whether
more empowered women self-select into the program, or whether participation actually improves
female autonomy.
    Table 3 shows us balance across key characteristics of the four treated and two untreated districts
in the sample. The only two signi�cant differences are in the number of sons and the time spent on
collecting �rewood. On average, participants have 0.27 sons more than non-participants and spend
signi�cantly more time collecting �rewood. The magnitude of the difference in the number sons
suggests the economic impact, if any, is small. However, the difference in time to collect �rewood is
large, making it important to control for the differences in time constraints. We discuss this variable
in further detail in the next section. Overall, table 3 suggests that the program does not appear
to be targeted in placement because there are no other signi�cant differences between women in
treated and untreated districts. These data illustrate that the women we sample in treated and
untreated districts are largely similar in covariates.
    Another source of endogeneity could arise from the potential placement of the program in villages
where women have a relatively high level of bargaining power and thus are more likely to respond
favorably to the treatment. We use administrative block-level 10 data from the Indian censuses of
1991 and 2001 on village-level female bargaining power. Matching the year that Mahila Samakhya
entered a block to the most recent census before that year, we compare the levels of bargaining
power in blocks blocks which later received Mahila Samakhya with those that did not. Results
for t-tests of equality, presented in table 4, show that there are no signi�cant differences in any of
the block-level measures of female bargaining power: sex ratio for the entire population and for
children up to the age of six, the scheduled caste or scheduled tribe sex ratio, and the literacy
ratio are all statistically similar across treatment assignment. Similarly, the sex ratio of workers
and non-workers also does not vary between treated and untreated blocks. Thus, these t-tests also
fail to yield evidence that would cause us to believe that the program was systematically targeted
at villages where women were either less empowered or already more empowered, and hence more
likely to respond to the treatment.
  10 Blocks are small groups of villages and are the administrative unit directly above the village council. Within each

district, the placement decision of Mahila Samakhya is made by block, rather than by village, hence these blocks are
the appropriate unit of analysis when examining whether the program is targeted in any way. For blocks without
the program, we use data from the 2001 census.


                                                          13
    It may still be that untreated districts in our sample are not representative of statewide trends
and that women in these districts may be more or less empowered than average, implying that
program placement may be targeted. However, the nationally-representative NFHS-3 (International
Institute for Population Studies and Macro International, 2007) and DLHS-3 (Ministry of Health
and Family Welfare and International Institute for Population Studies, 2010) show that the women
in untreated districts in our sample do not differ signi�cantly from the rest of the state. For instance,
the average age at at marriage for Uttarkhandi women is 20.6, while in our untreated sample, it
is 19.8; 43 percent of all Uttarkhandi women work while 45 percent of the untreated women in
our sample do. The total fertility rate in the state is 2.6, which corresponds closely to the average
family size of one boy and one girl in our untreated sample. Finally, while 84 percent of the state
has access to electricity, 90 percent of our untreated sample does. This lack of signi�cant differences
suggests that the program is not targeted at districts by levels of female empowerment.
    The next concern with identifying the effect of the program is self-selection. Table 5 indicates
the presence of self-selection into Mahila Samakhya. The average participant is three percentage
points closer in age to her husband than the average non-participant in treated districts, which
suggests that women with greater initial bargaining power may self-select into the program. Further,
participants tend to have older and more sons than non-participants, although the differences are
not signi�cantly different from zero. Participants are signi�cantly more likely to be Brahmin than
non-participants. Participants are less likely to live with their husbands; the difference of 19 percent
is highly signi�cant. However, in our pre-tests, we found that even women who do not live with
their husbands live with other male relatives, including fathers, fathers- or brothers-in law, uncles,
and sons. In all our �eldwork, we only encountered seventeen women who lived alone or without any
older male relative. Of these seventeen women, all but six lived with their mothers- or sisters-in-law.
The presence of these relatives in the house represents restrictions on the woman’s empowerment,
even if she lives without her husband.
    Several other characteristics, such as the number and age of daughters, the spousal education
ratio, and the woman’s time to collect water, are not statistically different for participants and non-
participants. Further, none of the wealth indicators, including number of rooms, electri�cation,
improved toilet facilities, materials used in floor and wall construction, are different for these two
groups, suggesting that poorer participants neither select into the program nor are they targeted
based on indicators of wealth (number of rooms, electri�cation, access to improved toilet facilities,
and nature of the construction materials used for the floor and walls of the house). Nonetheless, this
table highlights the importance of controlling for selection in to the Mahila Samakhya program. In
addition to the variables listed above in the Survey Description, we also control for the imbalanced
characteristics discussed here.


6     Empirical Analysis
6.1    Methodology
We use several methods to identify the effect of the program in the presence of self-selection. We �rst
match women by their pre-existing bargaining outcomes and other characteristics across treated
and untreated districts. We then control for potential issues of truncation in the matching. Next,
we use an instrumental variables approach, where our instrument is the interaction between the
time the program has been in the village and the years over 16 of the woman to capture the exposure
of the woman to the program. We separately control for the age of the woman in both �rst and


                                                  14
second stage regressions. Last, we estimate the intent-to-treat effect, where we match all women
regardless of participation in treated districts to those in untreated districts to estimate the effect
of the program.
    We estimate three sets of treatment effects. The �rst simply estimates whether women living
in treated districts, regardless of participation status, are more empowered than women living in
untreated districts. The second treatment effect examines whether non-participants in treated and
untreated districts are signi�cantly different in terms of female empowerment outcomes. The third
estimates the impact of the program on participants relative to untreated women with similar
characteristics to account for any issues of self-selection. Although table 3 suggests the lack of any
substantial differences between treated and untreated districts on observables, the �rst estimate
more formally tests this assumption.
Intent-to-Treat: We estimate intent-to-treat effects where we treat all women living in villages
with Mahila Samakhya — regardless of the woman’s participation status– as covered by the pro-
gram. If, even after ignoring the participation decision, we �nd that women living in treated villages
(regardless of their participation status) are signi�cantly more empowered than the untreated, we
can conclude that the program is effective.
Instrumental Variables: Propensity Score Matching only accounts for selection or targeting
on observables. While we have argued above that Mahila Samakhya may not be targeted in its
placement, women still chose whether to participate, and PSM may not fully control for all the
unobservable factors governing a woman’s participation decision. To prevent contamination from
unobservable characteristics influencing participation, we use two-stage least squares (2SLS) and
instrument for participation using exposure to the program. Our instrument for participation in
Mahila Samakhya is the number of years the program has been in a village interacted with the
woman’s age minus sixteen. The youngest participant we encountered in our �eld tests or data
collection was sixteen; we subtract sixteen from the age of the woman to accurately reflect the
number of years she could have participated in the program. The instrument tells us the years
of exposure to the program, and any effect of this variable on female empowerment likely works
through participation in the program, rather than directly. This variable is driven by the year
the program started in the village as there is little migration among married women in the region.
Since women often migrate at the time of marriage, and we do not know whether the woman’s natal
village had the program, migration at the time of marriage might lead to measurement error, which
in turn would bias results downwards. However, unmarried women tend not to participate in the
program, so exposure would have to be indirect, and thus the resultant bias would be small. We
control separately for the woman’s age so as to account for any trends in the dependent variable
that vary by age.
    A potential concern about our instrument may be that the instrument fails to account for
village-speci�c trends in the dependent variable that vary in the woman’s age and are correlated
with the empowerment outcomes. If this concern were correct, then our instrument would not
adequately control for the participation decision. Thus, by distinguishing between participants
and non-participants in treated villages, we would be inflating our estimates. However, the ITT
discussed above gets around this problem by treating participants and non-participants in treated
districts alike and only distinguishing between treated and untreated districts. If the ITT is con-
sistent with the results from the IV estimation, we will be able to conclude that the instrument
adequately controls for such bias.
Propensity Score Matching: To account for potential targeted program placement or selection
in participation, we use Propensity Score Matching (PSM) and instrumental variables to control for


                                                 15
self-selection. When treatment assignment or participation is not random but determined by observ-
ables, PSM allows us to compare treated individuals to untreated individuals (or non-participants
to the untreated) using observables such as demographic and economic characteristics to construct
the control group. Each individual in the dataset is assigned a propensity score that tells us the
likelihood of an individual being treated. That propensity score is a conditional probability measure
of treatment participation, given observable characteristics, x, and is expressed as follows:


                                     Pi (x) = P [Di = 1|X = x],                                  (5)

  We conduct this analysis maintaining the unconfoundedness assumption (Imbens and
Wooldridge, 2009):


                                     Di ⊥ (Yi (1), Yi (0))|Pi (x)                                (6)

where ⊥ signi�es independence, given that the balancing condition is satis�ed (Cameron and
Trivedi, 2005). The unconfoundedness assumption implies treatment assignment, Di is independent
of empowerment outcomes, Yi (corresponding to the Yi in section, after controlling for propensity
scores. In other words, we assume there are no unobservables that affect empowerment outcomes
and the probability of treatment.
    Treated and untreated individuals are matched based on proximity of their propensity scores,
Pi (), thus creating a control group. We then estimate treatment effects by comparing the outcome
of interest for the treated and control groups. PSM eliminates selection bias if controlling for
x eliminates selection bias from endogenous placement. Because treated and untreated districts
do not differ signi�cantly with respect to observable characteristics (table 3), it is a reasonable
assumption that the distribution of individual unobservable characteristics is similar across treated
and untreated districts. Because the program appears to have been distributed randomly across
districts, and individual selection into the program does not differ by district, a PSM approach will
give an unbiased measure of the program impact.
    For the treatment effect comparing non-participants and the untreated, each non-participant
is matched with replacement to an untreated woman based on the closeness of the propensity
score. For the treatment effect comparing participants to untreated individuals, each participant
is matched with replacement to an individual from an untreated district. We use kernel matching
in which all treated observations are matched with a weighted average of the propensity score for
all control observations. Weights are inversely proportional to the distance between the propensity
scores of treated and control observations (Becker and Ichino, 2002).
    Truncation may be a concern here because we are matching the full distribution of women in
untreated districts to a subset in treated districts that has chosen to be treated. Untreated women
represent the full distribution of outcomes, while participants represent a left-truncated sample
of this full distribution. If uncorrected, this truncation could bias our treatment results upward.
Hence, after matching the full sample of participants to untreated women, we truncate the sample
of untreated to only include women whose propensity to participation is no lower than the lowest
percentile of the participation propensity for participants. Similarly, we also re-estimate the match
between non-participants and the untreated only including untreated women whose propensity
score is no greater than the highest propensity score for non-participants. Note that all matches
are conducted on the region of common support.


                                                 16
Dependent Variables: The dependent variables measuring empowerment are of two kinds: (1)
pre-determined characteristics (proxied for by spousal age ratio) which cannot be affected by Mahila
Samakhya, and (2) characteristics that can be affected by participation, like having an NREGS
identi�cation card, leaving the house without permission, and attending village council meetings.
The independent variables on which we conduct the match and the 2SLS regression include observed
factors that likely affect both program participation and female empowerment: (1) spousal age ratio,
de�ned as the respondent’s age over her husband’s age, (2) the number and age of her children,11 (3)
her years of education, whether she is literate (in the case of participants, whether she was literate
prior to participation), and whether she has less than four years of education (and is thus likely to
need the education provided by the program). Time constraints may play an important factor in
determining participation and bargaining outcomes, so we also include (4) the time spent each day
by the respondent on collecting �rewood (reflecting time constraints),12 and (5) whether she lives
with in-laws and the number of sisters-in-law living with her (reflecting whether she can leave her
children in someone’s care while participating in the program), (6) whether she is a Brahmin, (7)
the number of rooms in her house, and (8) whether her house has electricity. Village �xed effects
are also included.


7     Results
Table 6 summarizes the estimated effects of the impact of Mahila Samakhya on proxies of female
empowerment, using an intent-to-treat estimate, 2SLS, PSM, and truncation-corrected PSM. Com-
paring estimates across methods highlights the robustness of our �ndings: all four methods suggest
that participation in Mahila Samakhya increases ownership of NREGS cards as well as a woman’s
ability to go out without permission. However, while the PSM results do not �nd a signi�cant
effect on participation in village council meetings, 2SLS estimates suggest otherwise. Overall, these
results indicate that participation in the program empowers women by giving them greater access
to employment and by increasing their physical mobility. In the rest of this section, we discuss
these results in detail.

7.1     Intent-to-treat
Although we instrument for selection into the program using years of exposure to Mahila Samakhya,
our instrument may not account for trends in the dependent variable that both vary in the woman’s
age and that are speci�c to treated villages. Thus, �rst we estimate a simple intent-to-treat effect
where we treat all women living in villages with Mahila Samakhya — regardless of the woman’s
participation status– as covered by the program. These intent-to-treat estimates, presented in
   11 To address concerns over whether the spousal age ratio and the age and number of children are truly exogenous,

we re-estimated all the results presented below without these three variables. The corresponding results are stronger
in signi�cance than the results including these variables, but the signs and magnitudes are similar. If the spousal
age ratio and age and number of children are not influenced by participation, but indicate pre-existing levels of
empowerment, and if we have self-selection into Mahila Samakhya, our estimates of the effect of participation would
be biased upward if we excluded these variables.
   12 The variable time spent collecting �rewood reflects an exogenous time constraint on the woman because Uttarak-

hand is exogamous and patrilocal. Sons tend to live with their parents, and the location of a married woman’s house
is not chosen by her and is therefore exogenous to the amount of time she spends collecting �rewood. The more time
a woman spends each day on �rewood collection the less time she has to participate in the program. However, it is
also possible that participants who spend a large amount of time collecting �rewood may feel isolated and may thus
be more interested in the social capital building activities of the program.


                                                         17
table 6 show a signi�cant and positive impact of participation on owning an NREGS card and on
being able to leave the house without permission. The results do not, however, show a signi�cant
impact on political participation. Nonetheless, we �nd evidence that the Mahila Samakhya program
was successful in empowering women in two of the three ways considered.

7.2     Instrumental Variables Results
The robust �rst stage results presented in table 7 show that the program exposure instrument is
highly signi�cant and positively correlated with participation in the program, validating the use of
this instrument to predict participation. The �rst-stage results also tell us that Brahmin women are
signi�cantly more likely to participate, as are women with greater �rewood collection times (sug-
gesting that the increased interest in community building activities dominates the time constraint),
while women who live with their parents-in-law are less likely to participate. The F-statistic for
this �rst stage is 12.38 for the physical mobility regression, 11.90 for political participation, and
13.55 for the access to outside employment regression. All three values are greater than the rule-
of-thumb cutoff of 10 proposed by Staiger and Stock (1997). To test the validity of the instrument
in the exactly-identi�ed regression, we used the procedure outlined in Nichols (2007), and added a
non-linear transformation of the instrument (non-logged exposure to the program), which allowed
us to use Sargan’s test for misspeci�cation. Sargan’s test results suggest that the instruments are
valid.
    Robust second-stage results (table 8) show that Mahila Samakhya participants are signi�cantly
more empowered than non-participants in all three ways: participants are more likely to leave
the house without permission, to participate in village council meetings, and have NREGS cards.
The increase in signi�cance of the physical mobility and political participation dependent variables
highlights the importance of correcting for sample selection by instrumenting for participation.
Other than program participation, the empowerment outcomes seem to be the result of different
data generating processes, with little overlap in signi�cance of explanatory variables across the
three regressions. Results suggest that older women are more likely to participate in village council
meetings, but that the number and age of children are important determinants of whether the
woman has an NREGS card— women with more young children are less likely to have NREGS
cards, perhaps because the time constraints imposed by raising children do not permit them to
work outside the house. Women who live with their sisters-in-law are more likely to leave without
permission and to participate in the village council, which may be because having the additional
help around the house enables women to leave more easily. Brahmins are less likely to have NREGS
cards, while women whose houses have electricity are more likely to go out without permission. 13
    The predicted outcomes based on the highly signi�cant 2SLS estimates are presented in �gures 2
and 3. These outcomes tell us that while 67.0 percent of non-participants can go out without
permission, only 49.4 percent of the untreated can do so. Similarly, 71.3 percent of non-participants
  13 To address concerns over the robustness of our standard errors, we also clustered by village and network— the

signi�cance of results does not change in either case. We also used randomization inference (RI) (Rosenbaum, 2002)
to estimate the z-scores for the three empowerment outcomes. Rather than drawing repeated samples of observations
from the known full population, RI assumes that the population is restricted to the observed sample. The treatment
assignment is assumed to be the only random variable. All observed outcomes and covariates are assumed �xed.
Using predicted participation from the �rst-stage of the 2SLS regression as the continuous treatment for RI, we �nd
participants to be more empowered than non-participants for all three outcomes. The z-scores are 6.53 for owning
an NREGS card, 1.56 for going out without permission, and 4.73 for participating in the village council. These
results suggest that, with the exception of physical mobility, participants are signi�cantly more empowered than
non-participants.


                                                        18
have NREGS cards, only 26.6 percent of the untreated do. Predicted political participation is low
for both groups; six percent of non-participants attend village council meetings, while 3.9 percent
of the untreated do.14 The fact that all predicted outcomes are higher for non-participants than for
the untreated suggests that Mahila Samakhya generates sizable spillover effects for non-participants
living in treated districts. Figure 3 tells us that 78.2 percent of participants can leave the house
without permission, 53.7 percent participate in the village council, and 79.7 percent have NREGS
cards. Compared to the predicted values for the untreated, these outcomes represent signi�cantly
higher levels of empowerment for participants. Participants are 28.8 percent more likely to go out
with permission, 49.8 percent more likely to participate in the village council, and 53.1 percent
more likely to have access to NREGS cards. The 2SLS estimates are similar in magnitude and
signi�cance to the ITTs and suggest that any omitted age-variable trends do not substantially
change our �ndings.
    Since our sampling strategy relied on networks, we do not have a randomly-selected sample.
Statewide data15 on NREGS show that in 2010, women in untreated districts used an average of 35
percent of the person-days of work generated by NREGS, while in treated districts, they used an
average of 41 percent, which is consistent with Mahila Samakhya having a positive effect on access
to outside employment and suggests that our estimates reflect statewide trends.

7.3     PSM Results
Table 11 presents the two sets of treatment effects discussed above estimated using PSM: the �rst
comparing non-participants to the untreated, and the other comparing participants to the untreated.
The upper panel of the table shows the results comparing non-participants to the untreated. These
results tell us that a non-participant is not signi�cantly more empowered by simply living in a
treated district. Without matching, only the NREGS cards variable is signi�cantly different, with
non-participants being signi�cantly more likely to own NREGS cards. The decrease in signi�cance
in NREGS card ownership after matching highlights the importance of controlling for selection in
to the program. Indeed, given that treated and untreated districts are very similar, these estimates
tell us that Mahila Samakhya does not target districts with particularly low (or high) levels of
empowerment.
    The lower panel of table 11 presents treatment effects of the program on participants. These
results show that participants and untreated women have statistically equal spousal age ratios,
suggesting that individuals do not choose to participate based on initial bargaining power. Hence,
any differences in the other measures of empowerment likely stem from the effect of the program.
Evidence suggests that the program signi�cantly increases access to outside employment, as 80.9
percent of participants own NREGS identi�cation cards, compared to only 14.4 percent of untreated,
which translates to a difference of 66.5 percentage points. Participants are also signi�cantly more
likely to leave the house without permission. However, according to the matched results, partici-
pants are not signi�cantly more likely to participate in village council meetings. A woman’s ability
to participate in village-level politics may depend on a high-stakes bargain with her husband and
in-laws, while the decision to get an NREGS card may be the result of a lower-stakes bargain be-
cause outside employment will earn the household extra income. Political participation may depend
  14 Thirty-six percent of all women in our sample attend village council meetings, but only 8 percent of untreated

women do. Fourteen percent of non-participants attend village council meetings, so the predicted outcome of 8
percent is signi�cantly lower. However, this discrepancy may simply highlight the importance of spillover effects
generated by Mahila Samakhya.
  15 District-level data on women’s access to NREGS identi�cation cards are not available for the entire state.




                                                        19
not only on program participation but also on the behavior of peers and support from them; this
link is studied in greater detail in Kandpal and Baylis (2011).
    A potential concern related to the use of PSM in comparing participants to women in untreated
districts is that our treated sample is truncated by only including those women who participate.
The true propensity of having access to Mahila Samakhya for women living in the treated districts
is one, so the best control-group matches would be women with high propensity scores in the
untreated districts. Therefore, by de�nition, women in villages without the program represent
the full distribution of outcomes, while treated women represent a left-truncated sample of this
full distribution. However, in the above matching process, we are comparing the full distribution
of women in untreated districts to a subset in treated districts that has chosen to be treated,
and therefore likely has higher propensity scores. If uncorrected, this truncation could bias our
treatment results upward.
    Thus, we may have a right-truncated distribution of non-participants and a left-truncated dis-
tribution of participants in treated districts. The distributions of propensity scores for program
participation of treated and untreated women highlight the problem of truncation (�gure 1). The
distribution of the untreated is bimodal, so we should not compare individuals in or around the lower
mode of the untreated distribution to participants. Similarly, comparing non-participants to indi-
viduals around the higher mode of the propensity scores for the untreated would also be misleading.
We correct for truncation by re-estimating the treatment effects comparing non-participants to the
untreated (presented in the upper panel table 11) with the sample of the untreated limited to those
whose propensity scores are below the lowest percentile of propensity scores for participants. For
the comparison between participants and the untreated, we limit the sample of untreated to women
with propensity scores for program participation greater than the lowest percentile of propensity
scores for participants. Results presented in table 14 indicate that truncation does not influence
the outcomes presented in table 11: the sign, size, and signi�cance of the estimates do not change
for either comparison set (non-participants versus untreated, and participants versus untreated).

7.4    Comparing PSM and 2SLS Estimates
The marginal effect estimates obtained from 2SLS cannot be interpreted as average treatment
effects, and thus cannot be compared directly to the estimates obtained from PSM. Multiplying
2SLS marginal effect estimates with individual propensity scores for participation gives us the
distribution of treatment effects. The average of this interaction is, then, the average treatment
effect. In table 10, we present the 2SLS estimates for average treatment effects of participation for
the entire sample, participants, non-participants, and the untreated.
    Marginal effect estimates point to the signi�cant bene�ts from participation. In the untreated
districts in our sample, 57.5 percent of all women (69 of 120 women) can go out without permission.
If these districts were to be covered by Mahila Samakhya, our results suggest that 75 percent (89
of 120 women) would be able to do so. Similarly, only 19 percent of all untreated women (25 of 131
women) participate in village council meetings, whereas if the program covered these districts, 41
percent of these women (55 of 131 women) would participate in village council meetings. Finally,
only 19 percent of the women had NREGS cards (27 of 140 women), but if they were to receive the
program, 88 percent (122 of 140 women) would have access to outside employment.




                                                 20
7.5     Disentangling Direct and Indirect Effects
Social learning, strategic interactions, and modeling information flows can tell us a lot about the
mechanisms underlying observed behavior (Maertens and Barrett, 2012). Therefore, in this section,
we attempt to disentangle the direct effect of participation in the program from its indirect effect
that may work through learning from peers, new information, and social interactions. Table 11
also shows that program participation increases the likelihood of a woman working, compared to
untreated women; the associated t-statistic is 1.52, making this difference short of statistically
signi�cant at the ninety percent level. In addition to an effect of the program on empowerment
through increased employability, there may also be a sizable effect even on participants who do
not work. Since participation in Mahila Samakhya does not affect the woman’s employment, it
must instead work either by increasing the woman’s disagreement utility and therefore affecting
the household bargain, or through the peer network effects of social learning and social influence.
Table 12 presents treatment effects of the program on these women by matching them to the
untreated. Participants who do not work are still more likely to have an NREGS card and to
participate in village council meetings. They are not, however, more likely to leave the house
without permission, perhaps because without working and earning an income, they do not have
adequate intrahousehold bargaining power. The fact that women who don’t work still own NREGS
cards maybe because NREGS only generates a hundred days of employment; hence participants
may not have been working at the time of the interview, but still had access to the NREGS program.
    However, simply observing improved outcomes for women who do not work does not let us
disentangle the effect of any changes in opportunity cost from an indirect, information-related
effect. To do so, we would need to isolate the effect of the program on women whose opportunity
cost does not change due to the program. Women who do not own NREGS cards are least likely to
have had their opportunity cost changed by the program: they do not work, and also do not have
access to outside employment, as represented by NREGS cards. Table 13 presents these treatment
effects, and shows that women who do not have NREGS cards are still more likely to participate in
village council meetings and to leave the house without permission. Thus, even women who do not
face increased opportunity costs from participation are more empowered than non-participants.16
We tried various metrics for the matching process; results are robust.
    The only exception is that non-participants are signi�cantly more likely to own an NREGS
identi�cation card than the untreated, after controlling for truncation. In our �eld tests, respondents
reported not having known that working outside the home was a possibility for them simply because
they had never seen anyone in the village do so. As a result of having participant friends, these
women may realize that working outside the home is in fact part of their choice set. The increase
in signi�cance in the truncation-corrected estimation thus suggests that having participant friends
may increase a woman’s choice set by showing her that working on NREGS-generated projects is
a possibility for her. The access to employment off the family farm and not involving household
chores likely increases the woman’s intrahousehold bargaining power.
    Table 9 presents robust second-stage results for treatment effects on women who do not have
NREGS cards or those do not work using 2SLS. These estimates are similar to those obtained
from matching: for women who do not work, we �nd that participants are signi�cantly more likely
to have NREGS cards than non-participants, although 2SLS does not yield a signi�cant effect of
  16 Further restricting the sample to women who do not work and do not have NREGS cards shrinks the sample

to 60 observations for the physical mobility estimation; the resultant treatment effect of 0.409 has an associated
standard error of 0.311 (t-statistic of 1.31). The political participation estimation, with 67 observations, yields a
treatment effect of 0.391, and a standard error of 0.235 (t-statistic of 1.66).


                                                         21
participation on attending village council meetings. For women without NREGS cards, we �nd that
participation in Mahila Samakhya signi�cantly increases the ability to go out without permission.
The positive effect on attending village council meetings is also signi�cant at the 90 percent level.
We thus �nd that Mahila Samakhya improves the bargaining power of not only women who work,
but also of those who do not have access to outside employment. This effect may come from a
combination of increasing the woman’s reservation wage and increasing her information set about
alternative choices. However, all but one of the �rst-stage F-statistics for these regressions are
lower than 10, ranging from 7.27 to 8.54. The only exception is the physical mobility regression for
women without NREGS cards, where the F-statistic is 11.75. While we might be concerned about
weak instruments in this case, the Sargan test indicate that the instruments are exogenous.
    Restricting the sample further to those women who do not work and do not have NREGS cards,
we have 73 observations in the physical mobility regression and 78 in the political participation re-
gression. The �rst-stage F-statistic for the physical mobility regression is 8.23, but is only 6.33 for
the political participation regression; as a result, we do not report the results from this regression.
We �nd that program participation increases a woman’s ability to leave the house without permis-
sion; the associated t-statistic is 1.83, meaning that the estimate is signi�cant at the 90 percent
level. Although the sample is small and weak instruments are a concern, these results suggest that
the indirect effect of Mahila Samakhya, working through increased reservation wages and expanded
information sets, can improve women’s physical mobility, and thereby empower them.
    In summary, using PSM, 2SLS, and ITTs, we �nd that the Mahila Samakhya program empowers
women in a variety of ways, via both direct and indirect routes. In addition to directly increasing
access to outside employment, political participation, and physical mobility, our results suggest the
program generates signi�cant spillovers that allow non-participants to bene�t via greater access to
outside employment. The program also empowers participants who do not work (and therefore only
bene�t from an increased reservation wage) to participate in village council meetings. Finally, we
�nd that participants who do not have access to off-farm work still have greater physical mobility
and political participation than untreated women.


8    Conclusion
This paper uses primary data from the north Indian state of Uttarakhand to study the impact of a
community-level rural women’s empowerment program called Mahila Samakhya. Mahila Samakhya
aims to empower women through education and information, taking a grassroots approach to its
implementation. We conceptually disentangle the effect of the program into a direct component
that works through access to outside employment and an increased reservation wage, as well as
an indirect component that works through information spillovers. This distinction between direct
and indirect effects extends beyond the Mahila Samakhya program to other programs operating at
the community-level that may have spillover effects. We �nd evience that Mahila Samakhya both
directly and indirectly succeeds in helping empower rural women.
    Our empirical approach addresses two sources of potential bias. First, the program may be
placed in villages and districts that are inherently different from those without the program, which
might bias our estimates. Second, the characteristics of women who participate in the program may
be different from those who choose not to participate, again, biasing our results. We �rst consider
program placement, and �nd that the characteristics of the women are similar across treated and
untreated districts. We further formally estimate bargaining power outcomes for non-participants
in treated districts and their matched counterparts in untreated villages and �nd no signi�cant

                                                  22
differences. To address the concern of self-selection into the program, we use PSM to compare the
outcomes of participants to those of untreated women. We also use 2SLS to instrument for the
decision to participate using the roll-out of the program as well as estimating the intent to treat,
and thus avoiding the problem of self-selection. By comparing participants to untreated women and
using 2SLS to control for self-selection in the participation decision, we provide accurate estimates
of the impact of the Mahila Samakhya program on women’s economic, social, and intrahousehold
empowerment. We test the validity of our instrument by estimating an intent to treat effect and
�nd similar results. Had we directly compared participants to non-participants within the same
district or failed to control for selection, we would have misestimated the true effect of participation.
    We �nd that the program has resulted in signi�cant increases in women’s access to outside
employment, ability to go out without permission, and political participation, all of which are
associated with higher levels of empowerment. We also �nd that these bene�ts were not restricted
to those women who found outside work after the vocational training. Our results suggest that
participants who do not work are still more able to leave the house without permission and have more
access to outside employment. After correcting for truncation in the distribution of participants,
we �nd that non-participants are signi�cantly more likely to own NREGS cards than untreated
women. This result may be evidence of positive spillover effects of the program on non-participating
neighbors.
    The marginal effects of the impact of Mahila Samakhya on empowerment translate to signi�cant
increases in the number of women with higher empowerment outcomes. If the untreated districts in
our sample were to be covered by Mahila Samakhya, 17 percent more women (an increase of 20 out
of 120 women) would be able to go out without permission, 33 percent more (an increase of 30 out
of 131 women) would participate in village council meetings, and 58 percent more (an increase of
95 out of 140 women) would have NREGS cards. These numbers highlight the potential of Mahila
Samakhya in effecting signi�cant social change.
    Results also show that even participants without access to outside employment are more em-
powered than untreated women. One may criticize the program for spending scarce resources on
individuals who do not then use their new-found skills to �nd employment, but we show that the
increased reservation wage is bene�cial in and of itself. Further, we �nd that Mahila Samakhya
had an indirect effect on the empowerment of non-participants relative to the untreated, indicating
that the program has a spillover effect. This indirect effect likely works through either increasing
access to information such as with a demonstration effect, or by changing social norms within the
village. Understanding these indirect effects calls for further research.
    The Mahila Samakhya intervention adopts a slow and careful grassroots approach to rolling
out its activities. Thus, our results cannot be generalized to programs following a faster, more
individual-focused, or a top-down approach. Further, these results should be interpreted with some
caution if selection on unobservables is a serious concern. That said, any bias from unaccounted-for
program placement would likely be in the downward direction because the program would target
women with low levels of empowerment. One might worry that areas with relatively empowered
women are more effective in lobbying for the program to be implemented in their district, however
we show that the characteristics of empowerment that are not likely to be affected by participation
are not signi�cantly different across treated and untreated districts and that other empowerment
outcomes do not signi�cantly differ over treated and untreated blocks before the introduction of
the program, suggesting that empowered women are not lobbying for the program.
    The Mahila Samakhya program is unique, but it may be fruitfully replicated elsewhere in the
developing world because it attempts to harness local peer networks to empower women. Similar


                                                  23
programs, including one in Honduras (Rowlands, 1997), indicate the potential of replicating a pro-
gram like the Mahila Samakhya elsewhere. The success of this program has encouraging implications
not just for female empowerment goals, but also for the other factors affected by empowerment,
such as child welfare. By empowering women to have greater say in their households and commu-
nities and to engage in income-generating activities, the program may generate signi�cant bene�ts
to the rest of the participant’s household.




                                               24
References
Agarwal, B. 2001. “Gender Inequality, Cooperation, and Environmental Sustainability.� In Eco-
 nomic Inequality, Collective Action, and Environmental Sustainability. , ed. P. Bardhan, S.
 Bowles and J.M. Baland. Princeton University Press.
Anderson, S., and J.-M. Baland. 2002. “The Economics of Roscas and Intrahousehold Resource
 Allocation.� The Quarterly Journal of Economics, 117(3): 963–995.

Anderson, S., and M. Eswaran. 2009. “What Determines Female Autonomy? Evidence from
 Bangladesh.� Journal of Development Economics, 90(2): 179 – 191.
Becker, G. 1973a. “A Theory of Marriage: Part I.� The Journal of Political Economy, 81(4): 813–
 846.

Becker, G. 1973b. “A Theory of Marriage: Part II.� The Journal of Political Economy, 82(2): S11–
 S26.
Becker, S., and A. Ichino. 2002. “Estimation of Average Treatment Effects Based on Propensity
 Scores.� The Stata Journal, 2(4): 358–377.
Caldwell, J., P. Reddy, and P. Caldwell. 1983. “The Causes of Marriage Change in South
 India.� Population Studies, 37(3): 343–361.
Cameron, A. C., and P. Trivedi. 2005. Microeconometrics. Cambridge University Press.
Dyson, T., and M. Moore. 1983. “On Kinship Structure, Female Autonomy, and Demographic
 Behavior in India.� Population and Development Review, 9(1): 35–60.
Faulkner, A. H., and V.A. Lawson. 1991. “Employment versus empowerment: A case study
  of the nature of women’s work in Ecuador.� Journal of Development Studies, 27(4): 16–47.
Handa, S. 1996. “Expenditure behavior and children’s welfare: An analysis of female headed
 households in Jamaica.� Journal of Development Economics, 50(1): 165 – 187.
Hashemi, S., S. Schuler, and A. Riley. 1996. “Rural Credit Programs and Women’s Empow-
 erment in Bangladesh.� World Development, 24(4): 635 – 653.
Imai, K., and P. Eklund. 2008. “Women’s Organizations and Social Capital to Reduce Prevalence
  of Child Malnutrition in Papua New Guinea.� Oxford Development Studies, 36(2): 209–233.
Imbens, G., and J. Wooldridge. 2009. “Recent Developments in the Econometrics of Program
  Evaluation.� Journal of Economic Literature, 47(1): 5–86.
International Institute for Population Studies and Macro International. 2007. National
  Family and Health Survey (NFHS-3), 2005-06: India: Volume 1. Mumbai: IIPS.
Janssens, W. 2010. “Women’s Empowerment and the Creation of Social Capital in Indian Vil-
  lages.� World Development, 38(7): 974–988.
Jones, C. 1983. “The Mobilization of Women’s Labor for Cash Crop Production: A Game Theo-
  retic Approach.� American Journal of Agricultural Economics, 65(5): 142–47.


                                              25
Kabeer, Naila. 1994. “Reversed Realities: Gender Hierarchies in Development Thought.� , ed.
 Naila Kabeer, Chapter Empowerment from Below: Learning from the Grassroots, 223–263. Verso.
Kabeer, Naila. 1999. “Resources, Agency, Achievements: Reflections on the Measurement of
 Women’s Empowerment.� Development and Change, 30(3): 435–464.
Kabeer, Naila. 2000. The Power to Choose: Bangladeshi Women and Labour Market Decisions
 in London and Dhaka. Verso.
Kanbur, R., and L. Haddad. 1994. “Are better off households more equal or less equal?� Oxford
 Economic Papers, 46(3): 445–58.
Kandpal, E., and K. Baylis. 2011. “Standing Together: Peer Networks, Female Empowerment,
 and Child Welfare.� Working Paper.
Kandpal, E., and K. Baylis. 2013. “Expanding Horizons: Can Women’s Support Groups Diver-
 sify Peer Networks in Rural India?� American Journal of Agricultural Economics, 95(2): 360–367.
Kantor, P. 2003. “Women’s Empowerment Through Home-based Work: Evidence from India.�
 Development and Change, 34(3): 425–445.
Lilja, N., J. Sanders, C. Durham, H. De Groote, and I. Demb´    e. 1996. “Factors influencing
                                                              el´
  the payments to women in Malian agriculture.� American Journal of Agricultural Economics,
  78(5): 1340–1345.
Lundberg, S., R. Pollak, and T. Wales. 1997. “Do husbands and wives pool their resources?
 Evidence from the United Kingdom child bene�t.� Journal of Human Resources, 32(3): 463–480.
Maertens, Annemie, and Christopher B. Barrett. 2012. “Measuring Social Networks’ Effects
 on Agricultural Technology Adoption.� American Journal of Agricultural Economics, 95(2): 353–
 359.
Malhotra, A., R. Pande, and C. Grown. 2003. “Impacts of Investments in Female Education
 on Gender Equality.� International Center for Research on Women.
Ministry of Health and Family Welfare and International Institute for Population
 Studies. 2010. District Level Household and Facility Survey, 2007-2008.
Montgomery, M., and J. Casterline. 1996. “Social Learning, Social Influence, and New Models
 of Fertility.� Population and Development Review, 22: 151–175.
Neelakantan, U., and M. Tertilt. 2008. “A note on marriage market clearing.� Economics
 Letters, 101(2): 103 – 105.
Nichols, A. 2007. “Causal Inference with Observational Data.� The Stata Journal, 7(4): 507–541.
Phipps, S., and P. Burton. 1998. “What’s Mine is Yours? The Influence of Male and Female
 Incomes on Patterns of Household Expenditure.� Economica, 65(260): 599–613.
Quisumbing, A., and B. de la Bri`ere. 2000. “Women’s Assets And Intrahousehold Allocation In
 Rural Bangladesh: Testing Measures Of Bargaining Power.� International Food Policy Research
 Institute (IFPRI) FCND Discussion Paper 86.

                                              26
Rahman, L., and V. Rao. 2004. “The Determinants of Gender Equity in India: Examining Dyson
 and Moore’s Thesis with New Data.� Population and Development Review, 30(2): 239–268.
Rosenbaum, P. 2002. Observational Studies. Springer Series in Statistics.
Rosenzweig, M., and T. P. Schultz. 1982. “Market Opportunities, Genetic Endowments, and
 Intrafamily Resource Distribution: Child Survival in Rural India.� The American Economic
 Review, 72(4): 803–815.
Rowlands, Jo. 1997. Questioning Empowerment: Working with Women in Honduras. Oxfam.
Schuler, S., and S. Hashemi. 1994. “Credit Programs, Women’s Empowerment, and Contra-
  ceptive use in Rural Bangladesh.� Studies in Family Planning, 25(2): 65–76.

Staiger, D., and J. Stock. 1997. “Instrumental Variables Regression with Weak Instruments.�
  Econometrica, 65(3): 557–586.
Thomas, D. 1990. “Intra-Household Resource Allocation: An Inferential Approach.� The Journal
 of Human Resources, 25(4): 635–664.




                                              27
9   Tables and Figures




             Figure 1: Propensity Scores for Program Participation




                                      28
Figure 2: Predicted Outcomes from 2SLS: Non-participants versus Untreated




  Figure 3: Predicted Outcomes from 2SLS: Participants versus Untreated



                                   29
                         Table 1: Descriptive Statistics
Variables                          Observations    Mean      Std. Dev    Min      Max
Respondent’s Age                        472         32.18       8.11         20   65
Husband’s Age                           437         37.89       9.25         23   80
Respondent’s Age at Marriage            463         19.25       3.34          1   30
Average age of sons                     487          8.09       7.79          0   36
Average age of daughters                487          6.20       6.70          0   30
Respondent’s Years of Education         483          7.19       5.02         0    17
Husband’s Years of Education            415         10.11       3.71         0    17
Sons’ Years of Education                487         3.766       4.45          0   17
Daughters’ Years of Education           487          3.11       4.29          0   17
Number of Rooms                         487          3.33       2.12          0   19
Electri�cation                          487          0.89       0.31          0    1




                         Table 2: Dependent Variables
      Dependent Variables                         Percent Yes    Observations
      Has NREGS ID Card
      All                                            60.62             485
      Non-participants                               48.94             188
      Participants                                   68.02             297
      Can Leave House Without Permission
      All                                            70.89             454
      Non-participants                               58.82             170
      Participants                                   78.17             284
      Participates in Village Council Meetings
      All                                            36.36             473
      Non-participants                               14.20             176
      Participants                                   49.49             297




                                       30
                         Table 3: Treated and Untreated Districts
Variables                        Untreated    Treated      Difference   t-test    Observations
Demographics
Spousal Age Ratio                   0.85           0.85      -0.03     -0.17         487
                                   (0.01)         (0.01)    (0.02)
Age at Marriage                    19.76          18.69       1.08      1.33         487
                                   (0.05)         (0.54)    (0.81)
Age of Sons                         6.96           9.03      -2.07     -1.66         487
                                   (0.84)         (0.76)    (1.25)
Age of Daughters                    5.45           6.98      -1.52     -1.78         487
                                   (0.46)         (0.84)    (1.29)
Number of Sons                      1.09           1.38      -0.29     -2.27∗        487
                                   (0.04)         (0.08)    (0.13)
Number of Daughters                 0.99           1.13      -0.14     -1.29         487
                                   (0.05)         (0.07)    (0.11)
Own-to-husband’s education          0.65           0.61      0.03       0.35         487
                                   (0.12)         (0.04)    (0.09)
Lives with Husband‡                 0.83           0.76       0.07      0.42         449
                                   (0.09)         (0.09)    (0.04)
Lives with In-laws‡                 0.56           0.45       0.11      1.18         487
                                   (0.11)         (0.04)    (0.09)
Works‡                              0.45           0.65      -0.08     -1.11         454
                                   (0.07)         (0.12)    (0.18)
Brahmin‡                            0.21           0.14       0.07      0.45         487
                                   (0.21)         (0.06)    (0.16)
LN(Firewood Collection Time)        3.65           4.66      -1.02     -5.3∗∗∗       487
                                   (3.43)         (3.18)    (5.63)
Wealth Indicators
Number of Rooms                      3.58          3.07      0.51       0.96         487
                                    (0.49)        (0.29)    (0.53)
House has Electricity‡               0.90          0.88      0.02       0.21         487
                                   (0.004)        (0.05)    (0.08)
Improved Toilet‡                     0.18          0.21      0.04      -0.30         487
                                    (0.07)        (0.07)    (0.05)
Floor†                               1.41          1.55      -0.14     -0.43         487
                                    (0.41)        (0.14)    (0.33)
Walls†                               1.39          1.76      -0.37     -1.85         487
                                    (0.18)        (0.11)    (0.19)
‡ No=0; Yes=1. † Impermeable=1; semi-permeable=2; permeable=3
Standard errors in parentheses
∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001




                                             31
Table 4: Block-level Data from Indian Censuses of 1991 and 2001 on Female Bargaining Power
 Variables                                   Untreated      Treated    Difference      t-test   Observations

 Sex Ratio (M/F)                                 1.02         0.99         -0.02      -0.52         47
                                                (0.03)       (0.02)       (0.04)
 Sex Ratio 0-6 (M/F)                             1.07         1.05          0.02      1.76          47
                                                (0.01)       (0.01)       (0.01)
 Ratio of Scheduled Caste Pop (M/F)              1.06         1.04         0.02       0.61          47
                                                (0.03)       (0.01)       (0.03)
 Ratio of Scheduled Tribe Pop (M/F)†             0.34         0.32         0.01       0.08          47
                                                (0.14)       (0.09)       (0.17)
 Literacy Ratio (M/F)†                           1.05         1.08         -0.03      -1.86         47
                                                (0.01)       (0.01)       (0.02)
 Ratio of Total Workers (M/F)†                   1.05         1.01         0.04       1.75          47
                                                (0.02)       (0.01)       (0.03)
 Ratio of Main Workers (M/F)†                    1.09         1.04         0.05       1.73          47
                                                (0.03)       (0.01)       (0.03)
 Ratio of Non-workers (M/F)                      0.87         0.95         -0.09      -1.57         47
                                                (0.05)       (0.02)       (0.05)
 †The distributions of the underlying variables for this ratio were signi�cantly different from normal; they
 were thus logged and the ratio of the resultant variables was used here.
 Standard errors in parentheses
 ∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001




                                                     32
   Table 5: Participants and Non-participants in Districts with Mahila Samakhya
Variables                        Non-part.    Part.    Difference    t-test    Observations
Demographics
Spousal Age Ratio                   0.84       0.86      -0.03     -2.53∗∗        303
                                   (0.01)     (0.01)    (0.01)
Age at Marriage                    18.48      19.17      -0.69      1.63          327
                                   (0.38)     (0.21)    (0.42)
Age of Sons                         7.26       8.97      -1.71      -1.81         345
                                   (0.77)     (0.50)    (0.95)
Age of Daughters                    6.33       6.54      -0.21      -0.25         345
                                   (0.73)     (0.44)    (0.84)
Number of Sons                      1.16       1.37      -0.21      -1.87         345
                                   (0.09)     (0.06)    (0.11)
Number of Daughters                 0.98       1.14      -0.16      -1.31         345
                                   (0.08)     (0.07)    (0.12)
Own-to-husband’s education          0.66       0.58       0.07      1.38          345
                                   (0.05)     (0.03)    (0.05)
Low Education‡                      0.29       0.31      -0.14      -0.26         345
                                   (0.05)     (0.03)    (0.06)
Lives with Husband‡                 0.85       0.67       0.19     3.10∗∗         312
                                   (0.04)     (0.03)    (0.06)
Lives with In-laws‡                 0.55       0.44       0.12      1.90          345
                                   (0.05)     (0.03)    (0.06)
Works‡                              0.52       0.59      -0.06      -1.02         336
                                   (0.05)     (0.03)    (0.06)
Brahmin‡                            0.05       0.21      -0.16     -3.51∗∗∗       347
                                   (0.02)     (0.03)    (0.04)
LN(Firewood Collection Time)        3.91       4.91      -0.99     -4.39∗∗∗       347
                                   (0.22)     (0.11)    (0.23)
Wealth Indicators
Number of Rooms                     3.09       3.30      -0.21      -0.81         345
                                   (0.21)     (0.13)    (0.26)
                                   ‡
House Has Electricity                0.89      0.89      0.00       0.01          345
                                   (0.03)     (0.02)    (0.04)
Improved Toilet‡                    0.28       0.26       0.02      0.35          345
                                   (0.03)     (0.05)    (0.06)
Floor†                              1.63       1.86      -0.23      -1.71         345
                                   (0.08)     (0.11)    (0.14)
Walls†                              1.77       1.81      -0.03      -0.25         345
                                   (0.08)     (0.10)    (0.13)
‡ No=0; Yes=1. † Impermeable=1; semi-permeable=2; permeable=3
Standard errors in parentheses.
∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001




                                             33
Table 6: Does Participation in Mahila Samakhya Empower Women?
          Method                Has NREGS      Can Go Out      Council
                                   Card       W/o Permission   Meetings

Intent-to-Treat                     0.313          0.488          -0.07
                                   (0.13)∗∗       (0.15)∗       (-0.049)
2SLS Estimate                       1.475          0.367          0.487
                                  (0.26)∗∗∗       (0.15)∗      (0.487)∗∗∗
PSM                                 0.665         0.433          0.088
                                  (0.14)∗∗∗      (0.16)∗∗        (0.14)
Truncation-corrected PSM            0.678         0.429          0.098
(women who do not work)           (0.13)∗∗∗      (0.16)∗∗        (0.14)
Standard errors in parentheses.
∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001




                                         34
           Table 7: Robust First Stage Estimates
                                                    (1)
                                            Program Participation
Participation Instrument                             0.033∗∗∗
                                                    (0.00)
Own Age                                              0.001
                                                    (0.00)
Spousal Age Ratio                                    0.171
                                                    (0.27)
Literate                                             0.054
                                                    (0.07)
Less Than Four Years of Education                    0.010
                                                    (0.08)
Own Years of Education                             −0.003
                                                  (−0.01)
Number of Children                                   0.066
                                                    (0.04)
Age of Children                                    −0.006
                                                    (0.01)
Brahmin                                              0.309∗∗∗
                                                    (0.05)
Lives with In-laws                                 −0.103∗
                                                    (0.05)
Lives with Sister-in-law                           −0.524
                                                    (0.42)
LN(Firewood Collection Time)                         0.040∗∗∗
                                                    (0.01)
Number of Rooms                                      0.004
                                                    (0.01)
House Has Electricity                              −0.076
                                                    (0.06)
Constant                                             0.022
                                                    (0.26)
Standard errors statistics in parentheses
∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001




                                35
Table 8: Does Participation in Mahila Samakhya Empower Women? (Robust 2SLS Estimates)
                                                  (1)             (2)               (3)
                                              Has NREGS       Can Go Out      Village Council
                                                 Card        W/o Permission    Participation
     Participation                                1.475∗∗∗         0.367∗           0.487∗∗∗
                                                 (0.26)           (0.15)           (0.14)
     Own Age                                    −0.005           −0.002             0.006∗
                                                (0.01)            (0.00)           (0.00)
     Spousal Age Ratio                          −0.026           −0.499           −0.224
                                                (0.41)            (0.28)           (0.26)
     Less than Four Years of Education            0.002          −0.008           −0.002
                                                 (0.01)           (0.01)           (0.01)
     Literate                                   −0.116           −0.092             0.005
                                                (0.12)            (0.06)           (0.07)
     Own Years of Education                       0.001            0.007          −0.002
                                                 (0.02)           (0.01)           (0.01)
     Number of Children                         −0.122             0.008          −0.028
                                                (0.07)            (0.04)           (0.04)
     Age of Children                              0.016            0.004            0.003
                                                 (0.01)           (0.00)           (0.00)
     Brahmin                                    −0.588∗∗∗        −0.032           −0.065
                                                (0.12)            (0.07)           (0.07)
     Lives with In-laws                           0.172          −0.044             0.015
                                                 (0.09)           (0.06)           (0.06)
     Lives with Sister-in-law                     0.902            1.012∗           0.905
                                                 (0.76)           (0.44)           (0.47)
     LN(Firewood Collection Time)               −0.034           −0.015             0.008
                                                (0.02)            (0.01)           (0.01)
     Number of Rooms                              0.007          −0.018             0.001
                                                 (0.02)           (0.01)           (0.01)
     House Has Electricity                        0.147            0.239∗∗∗         0.077
                                                 (0.11)           (0.07)           (0.07)
     Constant                                   −0.099             1.068∗∗∗       −0.082
                                                (0.42)            (0.27)           (0.25)
     Observations                               421              391              411
     First-stage F-stat                          13.55            12.38            11.90
     p-value for Sargan’s test                    0.579            0.242            0.803
     Standard errors in parentheses.
     ∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001




                                                 36
Table 9: Does Participation in Mahila Samakhya Increase Reservation Wages? (Robust 2SLS
Estimates)
                                      Women Who Do Not Work                 Women Without NREGS Cards
                                  (1)           (2)             (3)             (4)             (5)
                              Can Go Out  Village Council  Has NREGS        Can Go Out    Village Council
                              W/o Perm.        Part.           Card          W/o Perm.         Part.
                              Has NREGS    Can Go Out     Village Council   Can Go Out    Village Council
                                 Card       W/o Perm.          Part.         W/o Perm.         Part.
  Participation                   2.220∗∗         0.034         −0.082           1.134∗∗         0.638∗
                                 (0.87)          (0.40)          (0.35)         (0.47)          (0.38)
  Own age                       −0.012          −0.000           0.018∗∗∗      −0.008            0.002
                                (0.01)          (0.01)          (0.01)         (0.01)           (0.00)
  Spousal Age Ratio             −0.430          −0.642          −0.731∗        −1.011          −0.231
                                (0.92)          (0.48)           (0.44)        (0.65)           (0.43)
  <4 Years of Ed.                 0.397           0.059          0.011         −0.150            0.087
                                 (0.41)          (0.17)         (0.17)         (0.22)           (0.18)
  Literate                      −0.210          −0.098           0.070           0.169         −0.147
                                (0.35)          (0.11)          (0.14)          (0.20)          (0.15)
  Own Years of Ed.                0.040           0.000         −0.006         −0.015            0.004
                                 (0.03)          (0.02)          (0.01)        (0.01)           (0.01)
  Number of Children            −0.175            0.046         −0.016         −0.034          −0.073
                                (0.15)           (0.08)          (0.07)        (0.06)           (0.06)
  Age of Children                 0.006           0.003          0.019∗∗         0.009           0.014∗
                                 (0.02)          (0.01)         (0.01)          (0.01)          (0.01)
  Brahmin                       −1.215∗∗∗       −0.032           0.196         −0.371∗         −0.227
                                (0.42)          (0.21)          (0.18)         (0.22)           (0.15)
  Lives with In-laws              0.393∗∗       −0.057           0.081           0.195           0.136
                                 (0.20)         (0.09)          (0.09)          (0.15)          (0.11)
  Lives with Sis.-in-law          2.166           1.504∗∗        0.442           0.925           0.407
                                 (1.78)          (0.75)         (0.86)          (0.85)          (0.64)
  LN(Firewood Time)             −0.064          −0.008           0.020         −0.069∗∗          0.002
                                (0.06)          (0.03)          (0.02)         (0.03)           (0.02)
  Number of Rooms                 0.062∗        −0.012          −0.016         −0.044          −0.028
                                 (0.04)         (0.02)           (0.02)        (0.04)           (0.02)
  House Has Electricity           0.119           0.290∗∗        0.138           0.480∗∗∗        0.087
                                 (0.22)          (0.12)         (0.10)          (0.18)          (0.13)
  Constant                      −0.473            1.107∗∗        0.072           1.376∗          0.029
                                (0.98)           (0.44)         (0.41)          (0.72)          (0.39)
  Observations                  178            163             175             144             157
  First-stage F-stat              8.54           8.03            8.54           11.75            7.27
  p-value for Sargan’s test       0.391          0.818           0.908           0.511           0.569
  Standard errors in parentheses
  ∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001




                                                      37
Table 10: Average Treatment Effects from 2SLS Estimates
               Average Treatment Effect
     NREGS                               0.818
     Can Go Out W/O Permission           0.204
     Village Council                     0.269
       Average Treatment Effect on Participants
     NREGS                               0.893
     Can Go Out W/O Permission           0.222
     Village Council                     0.295
     Average Treatment Effect on Non-participants
     NREGS                               0.643
     Can Go Out W/O Permission           0.159
     Village Council                     0.212
         Average Treatment Effect on Untreated
     NREGS                               0.621
     Can Go Out W/O Permission           0.155
     Village Council                     0.205




                          38
Table 11: Does Participation in Mahila Samakhya Empower Women? (PSM Results)
                                           Non-Participants vs. the Untreated
Unmatched             Spousal       Has NREGS            Can Go Out      Council     Respondent
                     Age Ratio         Card             W/o Permission   Meetings      Works
Non-participants        0.835           0.719                0.671        0.152         0.641
Untreated               0.849           0.256                0.578       0.0986         0.523
Difference               -0.015          0.463                0.093        0.053         0.118
                       (0.014)       (0.071)∗∗∗             (0.081)      (0.055)       (0.097)

Matched               Spousal       Has NREGS            Can Go Out      Council     Respondent
                     Age Ratio         Card             W/o Permission   Meetings      Works
Non-participants        0.835             0.719               0.671       0.152         0.641
Untreated               0.839             0.634               0.747       0.089         0.297
Difference               -0.004            0.085              -0.076       0.063         0.344
                       (0.033)           (0.213)            (0.221)      (0.148)       (0.031)

Observations             160              160                143           150          108
                                             Participants vs. the Untreated
Unmatched             Spousal       Has NREGS            Can Go Out      Council     Respondent
                     Age Ratio         Card             W/o Permission   Meetings      Works
Participants            0.862           0.809                0.793         0.502       0.581
Untreated               0.850           0.177                0.654         0.205       0.440
Difference              0.0119           0.632                0.139         0.297       0.141
                       (0.009)       (0.044)∗∗∗            (0.052)∗∗     (0.06)∗∗∗    (0.06)∗∗

Matched               Spousal       Has NREGS            Can Go Out      Council     Respondent
                     Age Ratio         Card             W/o Permission   Meetings      Works
Participants            0.862           0.809                0.793        0.502         0.581
Untreated               0.848           0.144                0.361        0.414         0.321
Difference               0.014           0.665                0.433        0.088         0.261
                       (0.026)       (0.142)∗∗∗            (0.158)∗∗     (0.142)       (0.172)

Observations             341              339                312           332          315
Standard errors in parentheses.
∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001




                                                   39
Table 12: Does Participation in Mahila Samakhya Empower Women Who Do Not Work? (PSM
Results)
                                          Participants vs. the Untreated
                  Unmatched        Has NREGS            Can Go Out         Council
                                      Card             W/o Permission      Meetings
                  Participants         0.730                0.667           0.416
                  Untreated            0.179                 0.708          0.315
                  Difference            0.552                -0.042          0.101
                                    (0.072)∗∗∗             (0.085)         (0.084)

                  Matched          Has NREGS            Can Go Out         Council
                                      Card             W/o Permission      Meetings
                  Participants         0.730                0.667           0.416
                  Untreated            0.101                0.500           0.146
                  Difference            0.629                0.167           0.269
                                    (0.118)∗∗∗             (0.184)         (0.132)∗

                  Observations          145                 132                143
                  Standard errors in parentheses.
                  ∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001




Table 13: Does Participation in Mahila Samakhya Empower Women Who Do Not Have NREGS
Cards? (PSM Results)
                                          Participants vs. the Untreated
                         Unmatched         Can Go Out             Council
                                          W/o Permission          Meetings
                         Participants             0.821              0.525
                         Untreated                0.675              0.189
                         Difference                0.146              0.336
                                                 (0.087)          (0.081)∗∗∗

                         Matched           Can Go Out             Council
                                          W/o Permission          Meetings
                         Participants           0.821                0.525
                         Untreated              0.308                0.025
                         Difference              0.513                0.500
                                               (0.263)∗           (0.144)∗∗∗

                         Observations             122                135
                         Standard errors in parentheses.
                         ∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001




                                                  40
Table 14: Does Participation in Mahila Samakhya Empower Women Who Do Not Work?
(Truncation-corrected PSM Results)
                                                 Non-Participants vs. the Untreated
     Unmatched             Spousal       Has NREGS              Can Go Out       Council     Respondent
                          Age Ratio         Card               W/o Permission    Meetings      Works
     Non-participants        0.822           0.667                  0.722          0.130        0.579
     Untreated               0.824           0.147                  0.629          0.100        0.440
     Difference               -0.003          0.519                  0.093          0.030        0.139
                            (0.029)       (0.110)∗∗∗               (0.146)        (0.089)    (0.060)∗∗∗

     Matched               Spousal       Has NREGS              Can Go Out       Council     Respondent
                          Age Ratio         Card               W/o Permission    Meetings      Works
     Non-participants        0.822              0.667               0.722          0.130        0.579
     Untreated               0.729              0.125               0.611          0.000        0.341
     Difference               0.093               0.54               0.111          0.130        0.238
                            (0.075)           (0.254)∗∗            (0.213)        (0.072)      (0.168)

     Observations              58                58                  45                53        44
                                                      Participants vs. the Untreated
     Unmatched             Spousal       Has NREGS              Can Go Out       Council     Respondent
                          Age Ratio         Card               W/o Permission    Meetings      Works
     Participants            0.861           0.808                  0.797          0.505        0.582
     Untreated               0.851           0.177                  0.654          0.205        0.427
     Difference               0.01            0.631                  0.143          0.299        0.155
                            (0.009)       (0.044)∗∗∗              (0.052)∗∗      (0.06)∗∗∗    (0.061)∗∗

     Matched               Spousal       Has NREGS              Can Go Out       Council     Respondent
                          Age Ratio         Card               W/o Permission    Meetings      Works
     Participants            0.861           0.808                  0.797          0.505        0.582
     Untreated               0.846           0.131                  0.367          0.407         0.647
     Difference               0.015           0.678                  0.429          0.098        -0.066
                            (0.025)       (0.132)∗∗∗              (0.156)∗∗       (0.142)      (0.192)

     Observations             340               338                 311            331          314
     Standard errors in parentheses.
     ∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001




                                                          41