Policy Research Working Paper                 9467




Poverty Alleviation and Interhousehold Transfers
 Evidence from BRAC’s Graduation Program in Bangladesh

                              Selim Gulesci




Development Economics
Knowledge and Strategy Team
November 2020
Policy Research Working Paper 9467


  Abstract
 Poor households often rely on transfers from their social                          more likely to both give and receive transfers to/from
 networks for consumption smoothing, yet there is limited                           wealthier households within their communities; and less
 evidence on how antipoverty programs affect informal                               likely to receive transfers from their employers. As a result,
 transfers. This paper exploits the randomized roll-out of                          the reciprocity of their within-village transfers increases.
 BRAC’s ultra-poor graduation program in Bangladesh and                             The findings imply that, within rural communities, there
 panel data covering over 21,000 households over seven years                        is positive assortative matching by socio-economic status. A
 to study the program’s effects on interhousehold transfers.                        reduction in poverty enables households to engage more in
 The program crowds out informal transfers received by the                          reciprocal transfer arrangements and lowers the interlinkage
 program’s beneficiaries, but this is driven mainly by out-                         of their labor with informal insurance.
 side-village transfers. Treated ultra-poor households become




 This paper is a product of the Knowledge and Strategy Team, Development Economics. 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://www.worldbank.org/prwp. The author may
 be contacted at gulescis@tcd.ie.




         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
        Poverty Alleviation and Interhousehold Transfers:
Evidence from BRAC’s Graduation Program in Bangladesh*

                                          SELIM GULESCI




1     Introduction
In rural parts of developing countries, individuals are faced with substantial risks
and often rely on transfers from their social networks for consumption-smoothing
(Rosenzweig (1988), Udry (1994)); yet there is limited evidence on how large-scale
antipoverty programs affect interhousehold transfers. This paper provides evidence
on the causal effect of a large-scale poverty alleviation program on informal transfers
within rural Bangladeshi communities. Implemented by the NGO BRAC, “Targeting
the Ultra Poor” program (henceforth “the program”) targets the poorest households
in rural communities in Bangladesh and provides them with a transfer of productive
assets, training and other support services. Bandiera et al. (2017) show that the pro-
gram leads to long-run improvements in the socio-economic conditions of beneficiary
households. Similar programs targeting ultra-poor households have been evaluated
in other contexts and, by and large, have found positive effects on socio-economic con-
ditions of the targeted households (Banerjee et al. (2015, 2016, 2018), Blattman et al.
(2016), Collins and Ligon (2017)); but to date, we have limited evidence on these pro-
grams’ effects on informal transfers. In this paper, I exploit the randomized roll-out
of the program and panel data covering over 21,000 households over seven years to
study the program’s effects on interhousehold transfers.
   * Selim Gulesci is Associate Professor at Trinity College Dublin, his email address is: gulescis@tcd.ie.
The author thanks the editor (Eric Edmonds) and three anonymous referees for comments that im-
proved the paper; Oriana Bandiera, Tim Besley, Robin Burgess, Maitreesh Ghatak, Imran Matin, Imran
Rasul and Munshi Sulaiman for their support and advice throughout the project; Yann Bramoullé, Greg
Fischer, Eliana La Ferrara, Rohini Pande, Jean-Phillipe Platteau for helpful discussions; seminar partic-
ipants at Aix-Marseille, Bocconi, EBRD, Edinburgh, IFS, George Washington, Koc, LSE-UCL, Namur,
Sabanci, Uppsala, Vanderbilt and Wharton for their comments. The usual disclaimer applies. This
document is an output from research funding by the UK Department for International Development
(DFID) as part of the iiG, a research programme to study how to improve institutions for pro-poor
growth in Africa and South-Asia. The views expressed are not necessarily those of DFID. Support from
the Bagri Fellowship is gratefully acknowledged. Previous versions of this paper were circulated with
the titles “Labor-tying and Poverty in a Rural Economy” and “Poverty and Interlinkage in Rural Labor
Markets”.


                                                    1
In order to evaluate the effects of the program, potential beneficiary households (hence-
forth referred as the “ultra-poor”) were identified by the program in 40 BRAC branch
offices covering 1,309 villages. Within every village, BRAC conducted participatory
wealth appraisals whereby community members assigned a wealth class for each house-
hold: poor, middle or upper class. Following this community appraisal, BRAC pro-
gram officers identified eligible “ultra-poor” households among the poor.1 A baseline
survey was conducted on all of the ultra-poor households, as well as a representa-
tive sample of the rest of the community. Importantly, the survey included questions
on informal transfer networks of respondents’ households. Using the baseline village
census, the identity of any household with whom the respondent’s household inter-
acted within the village was recorded. After the baseline survey, the program was
randomly introduced in 20 branch offices in 2007, and the other 20 remained as con-
trols until 2011. The same sample of households were re-surveyed in 2009 and 2011,
after which the control communities were also treated. An additional followup survey
was conducted in 2014, that is seven years after the baseline and three years after the
control branches had also been treated, and I will use this data to examine long-run
trends of outcomes for early versus late treated ultra-poor households.
I find that the program crowded-out informal transfers received by the ultra-poor
households in treatment villages, but this was driven mainly by outside-village trans-
fers. The ultra-poor in treatment villages were 13 percentage points (ppt) less likely to
have received informal transfers (during the 12 months preceding the survey) relative
to the control group, while the effect on the intensive margin is imprecisely estimated.
When I disaggregate the transfers by location, I find that they were 4ppt less likely
to have received transfers from within their village but the value of transfers received
were 29% higher relative to the control group. Moreover, they were more likely to
report receiving food transfers from other households within the village. On the other
hand, outside village transfers were crowded out by 12ppt on the extensive margin
and 23% in value.
Examining who the ultra-poor households exchange transfers with, I find that the pro-
gram increased their likelihood to receive as well as give transfers from/to wealthier
households within their village. Here, as a proxy for wealth, I use the socio-economic
wealth rank assigned to each household at baseline, during the community wealth ap-
praisal. I find that, after the program, the likelihood that the ultra-poor in treatment
villages give transfers to other poor, middle and upper class households increased.
Moreover, middle and upper class households were more likely to report giving trans-
fers to the ultra-poor in their communities, and this effect is maintained seven years af-
   1 Households   who were in the poor class according to the community wealth appraisal, but not se-
lected as potential beneficiaries of the program by BRAC are referred to as the “other poor” households
throughout the paper.


                                                  2
ter the baseline. As a result of these changes, the reciprocity of ultra-poor households’
transfer arrangements with higher socio-economic classes increased significantly. In
particular, the reciprocity of ultra-poor households’ transfer links with middle class
households increased by 9ppt (19% relative to control) and by 12ppt (44%) with upper
class households.
Finally, I find that the program reduced the likelihood that ultra-poor households re-
ceive transfers from their employers. Ultra-poor households in treatment communi-
ties were 4ppt less likely to report an employer as a source of transfers - a large effect
considering that in the control group only 9% of ultra-poor households reported an
employer as a source of transfers. To understand whether this effect was driven by a
general shift away from wage-labor I restrict the sample to ultra-poor households who
were engaged in wage-employment at baseline as well as at the two followup surveys;
and I find that they were also less likely to receive transfers from their employers.
While this analysis has a caveat (since it is based on a selected subsample), it never-
theless suggests that the effect is not only due to a shift away from wage employment.
Overall, the findings suggest that there may be a causal relationship between poverty
and alternative forms of reciprocity in informal insurance networks: poor households
may reciprocate transfers from wealthier members of the community by selling their
labor cheaply to them; while less poor households may reciprocate with transfers of
their own.
The paper contributes to the literature on informal insurance in village economies in
a number of ways. First, it is related to the literature on risk-sharing through recipro-
cal transfers (Townsend 1994, Udry 1994). This literature has demonstrated that risk-
sharing networks are likely to be concentrated along dimensions of proximity such
as kinship, caste or geographic proximity.2 What has been less studied is whether
poverty has a causal effect on who agents are matched with in risk-sharing networks.
Previous empirical studies have exploited cross-sectional variation or lab experiments
to shed light on this.3 This paper contributes to the literature by showing that an ex-

   2 Theoretical   reasons for this have been studied by La Ferrara (2003), Genicot and Ray (2003),
Bramoullé and Kranton (2007), Bloch, Genicot and Ray (2008) and Ambrus, Mobius and Szeidl (2014).
Empirically, this point has been demonstrated by Fafchamps and Lund (2003), La Ferrara (2003), Der-
con et al. (2006), Fafchamps and Gubert (2007), Attanasio et al. (2012), Barr et al. (2012), Munshi and
Rosenzweig (2016), Angelucci et al. (forthcoming), Chandrasekhar et al. (forthcoming).
    3 Jalan and Ravallion (1999) show that poorer households in rural China are less well insured against

fluctuations in their income. Fafchamps and Gubert (2007) show that the wealth difference between two
households is negatively correlated with having a risk-sharing connection in the Phillipines, while De
Weerdt (2004) finds the opposite relationship in Tanzania. Schechter and Yuskavage (2011a, 2011b) high-
light that the failure to distinguish between reciprocated and unreciprocated links could be one reason
for these conflicting findings. They show, using survey data form Paraguay, that reciprocated links are
more likely to exist between households who have similar wealth or live in close proximity, while un-
reciprocated transfers links are more likely between households of different wealth or education levels.
Attanasio et al. (2012) show experimentally that among close friends and relatives, individuals with


                                                   3
ogenous improvement in the socio-economic statuses of the poorest households in a
village (caused by an antipoverty program) increases their engagement in reciprocated
transfers with wealthier households within the community. To the best of my knowl-
edge, this is the first study to demonstrate the causal effect of socio-economic status
on matching in reciprocal transfer links.
The paper also contributes to the literature on crowding-out of informal transfers by
public transfers. Early empirical studies have demonstrated a modest negative cor-
relation between household income and transfers received (Cox and Jakubson (1995),
Altonji et al. (1997)) – although the relationship is likely to be heterogenous and non-
linear (Cox (1987), Schoeni (1997), Cox et al. (2004)). 4 The program I study leads to
sustained long-term improvements in the socio-economic conditions of targeted poor
households through a one-off transfer of assets and skills. I show that while the pro-
gram crowds-out informal transfers received by the targeted poor households, the ef-
fect is relatively small and varies based on geographic proximity. While transfers from
outside the village are crowded out, transfers from households living in the same vil-
lage as the targeted poor are not. This suggests that transfers within and outside the
local community are driven by different motives. In particular, our findings are in line
with within-community transfers being driven by incentive-related motives (such as
reciprocal exchange), as opposed to preference-related motives (Leider et al. (2009),
Ligon and Schechter (2012)).
Finally, the paper is related to the literature on interlinkage in rural labor markets.5
Previous empirical work has studied interlinkages between trade and financial mar-
kets in developing countries (McMillan and Woodruff (1999), Casaburi and Reeds
(2015), Casaburi and Machiavello (2016)), but empirical work on interlinkage in the
similar risk attitudes group together for risk-sharing. Ligon and Schechter (2012) also use an exper-
imental approach and show that transfers within the experiment are more likely to be motivated by
reciprocity (as opposed to other-regarding preferences) when individuals are connected (and exchange
transfers) outside of the experiment. Ado and Kurosaki (2014) replicate the same methodology in a
different context and find similarly that incentive-related motives (reciprocity and sanction aversion)
are more likely to be associated with sharing in the real world.
    4 Jensen (2003) and Juarez (2009) find large crowding out effects of increases in public transfers

targeted to the elderly in South Africa and Mexico respectively. Albarran and Attanasio (2002) find
that beneficiaries of a conditional cash transfer program in Mexico (Progresa) were less likely to receive
informal transfers, and conditional on receiving any, they received lower amounts of transfers. On the
other hand, Angelucci and de Giorgi (2009) provides evidence that households eligible for Mexico’s
Progresa program increased transfers to ineligible households residing in treatment villages. Walker
(2017) shows that a one-time unconditional cash transfer in Kenya did not lead to any change in transfer
giving to family and friends among recipient households in the short-run.
    5 The literature on village economies has shown that employers may provide assistance to their

poorer workers in times of need, through the provision of consumption loans or transfers (Platteau
and Abraham (1987), Platteau (1995a, 1995b)). Theoretically, workers with limited outside options may
prefer to maintain a long-term relationship with an employer in order to have smoother earnings across
the agricultural seasons, even though this may end up providing them with a lower wage relative to
causal labor contracts available on the spot labor market (Bardhan (1983), Mukherjee and Ray (1995),
Caselli (1997)).


                                                   4
rural labor market has been largely descriptive (Bardhan and Rudra (1978, 1981),
Richards (1979), Anderson (1990)). This paper contributes to the literature by showing
that poverty level of the worker may have a causal effect on interlinkage of contracts
in the rural labor market with informal insurance arrangements. Importantly, due to
the nature of the particular antipoverty program I study (it is designed to increase
self-employment among the targeted poor households) the effects may in part be due
to the fall in wage-labor caused by the program. It is unclear whether a reduction
in poverty caused by another type of intervention would have a similar effect on the
interlinkage of employment and transfers.

2        “Targeting the Ultra Poor” Program
The TUP program is a multi-faceted intervention with the central aim of lifting house-
holds out of extreme poverty. The main beneficiary of the program is an adult woman,
and to improve their socio-economic conditions, the program combines the transfer of
productive assets with complementary training, a cash stipend and other supporting
services. This section describes the selection procedure of the beneficiaries and the
details of the program.

2.1           Targeting
In order to identify the potential beneficiaries of the program, BRAC proceeds in four
steps. First, BRAC officials in Dhaka select the BRAC branch offices where the pro-
gram would be implemented. Then, BRAC officials at the branches identify which
communities would be targeted. In the third step, program officials organize a par-
ticipatory wealth ranking in each community selected for the program. This exercise
places all households into one of several wealth bins corresponding to the poor, middle
and upper classes.6 Finally, BRAC officers verify which households among the poor
satisfy pre-determined inclusion and exclusion criteria. Based on these criteria, they
subdivide the poor (as determined by the community appraisal) into “ultra-poor”,
who are eligible for the TUP program, and the rest (“other poor”).7 Eligibility is deter-
mined at baseline, and was not reevaluated within treatment villages over time. The
share of treated households relative to the average village population was 6%.8
       6 This procedure is similar to community appraisal methods studied by Alatas et al (2012) in Indone-

sia.
       7 In
         particular, the program has three exclusion criteria, all of which are binding. Households who
are borrowing from an NGO providing microfinance, who are recipients of any government benefits
or who do not have any able-to-work adult female members are excluded from the program. To be
selected, a household has to satisfy three of the following five inclusion criteria: (i) total land owned
including homestead is no more than 10 decimals (approximately 0.1 acre); (ii) there is no adult male
income earner in the household; (iii) adult women in the household work outside the homestead; (iv)
school-going-aged children are working in an income-generating activity; and (v) the household has no
productive assets.
    8 Eligibility is determined at the household, rather than at the individual level. The program targets

the leading woman (female head of household) in eligible households as the person in charge of the


                                                      5
2.2    Program Components
The first component of the program is asset transfer. Assets are chosen by the bene-
ficiaries from a menu offered by BRAC. The menu includes assets related to various
income-generating activities (such as livestock rearing, vegetable cultivation, setting
up small retail shops, production of small crafts) but almost all eligible women in our
sample opted for livestock rearing. For this, beneficiaries had the option to choose be-
tween six different livestock packages containing either one or two animal types (e.g.
only cows or a cow and five goats), and all packages were on average of similar value
of BDT 9500 (approximately USD 514 in PPP terms) in 2007. BRAC encouraged pro-
gram recipients to commit to retain the assets for two years, although this commitment
was not strictly enforceable. After two years, beneficiaries were under no obligation
or no encouragement to retain the livestock asset.
The second key component of the program is skills transfer. In this, the program aims
to teach the beneficiaries skills that are complementary to the transferred livestock as-
set, such as maintaining livestock health, best-practices related to feeding the animals,
insemination to produce offspring and milk, rearing calves, bringing outputs to the
market etc. The initial skills transfer was conducted through a combination of class-
room training at BRAC regional offices. This was followed with regular assistance and
coaching by a livestock specialist and program officers, as well as refresher trainings
that took place periodically to strengthen the initial technical training. During these
refresher trainings and visits, further knowledge to participants is delivered, their per-
formance is monitored and various challenges are discussed followed by troubleshoot-
ing. As part of this, every beneficiary was visited by a livestock specialist every one to
two months for the first year, and by BRAC program officers weekly for the first two
years of the program.
In addition to asset and skills transfer, the program entails a number of supporting
services. Program officers organize weekly meetings which bring program benefi-
ciaries from the same village together with one another, and they conduct life-skills
awareness sessions. The objective of these sessions is to sensitize participants and to
make them more aware of their legal and political rights. Some of the topics covered
during these visits include health issues, such as waterborne diseases, de-worming,
food, nutrition and anaemia, family planning, immunization; social issues such as
child marriage, dowry, domestic violence, marriage registration and the importance
of children’s education.

assets (livestock) transferred and to receive the training. If there are multiple female members in the
household, then it is up to the household to decide which female member should be the main point of
contact with the program officers. This is not to say that other household members do not benefit from
the program indirectly, but in terms of the program’s operations, the main beneficiary is the female
head of the household.


                                                  6
During their weekly visits, program officers also distribute weekly cash stipends to
the program beneficiaries. This is meant to alleviate any short-run fall in earnings due
to the occupational change induced by the program and to provide consumption sup-
port for ultra-poor households. The exact amount of the stipend varies across regions,
but it is meant to be equivalent to half a day’s wage for casual laborers (BRAC (2016)).
These cash transfers typically last for the first 40 weeks after the asset transfer. To en-
sure proper utilization of this stipend, beneficiaries are also provided with diet charts
which indicate affordable and nutrient-rich food items available in their localities.
Another component of the program is to promote savings behavior. To achieve this,
program beneficiaries are encouraged to save a portion of their monthly income in
BRAC’s savings account. The idea behind this is to help the ultra-poor households
start building financial security for further enterprise expansion and to lower their
vulnerability to shocks. The program officers collect savings during the weekly meet-
ings they organize for the program members.
The program also provides health support through BRAC’s community-based health
workers. Beneficiaries receive guidance about preventative care, treatment, antenatal
and prenatal care, contraception and childcare, as well as financial assistance for sur-
gical cases. In some areas, BRAC also implements public health interventions such as
latrine and tube-well installation.
Finally, in order to strengthen the social networks of the ultra-poor households, BRAC
establishes local committees called Gram Daridro Bimochon (Village Poverty Allevia-
tion) Committees (henceforth ‘GDBC’). Each committee comprises of representatives
of the program beneficiaries (i.e. ultra-poor women) and community leaders such as
wealthier landlords, teachers, priests, local politicians. The GDBC meets monthly and
discusses any issue raised by the representatives of the program beneficiaries with the
aim of helping them build social networks and leverage community ties. The commit-
tee is meant to help participants protect their assets, facilitate access to government
services, and offer support in times of need by convening local community support.
The implementation of the components described above last for a maximum period
of 24 months. After this period, the beneficiary households are “graduated” from the
program. Upon graduation, they are invited to participate in BRAC’s microfinance
program, with the idea that access to microcredit can enable them to maintain and
expand their business activities; but they are in no obligation to borrow from BRAC
microfinance.
Figure 1 provides a visual summary of the main components of the program and the
timeline of its implementation.



                                            7
3     Conceptual Framework
We are interested in understanding whether and if so how a multi-faceted antipoverty
program, such as the TUP program, may affect the informal transfer arrangements of
beneficiary households. The program entails many components and there are multi-
ple ways through which it may impact both the supply of informal transfers to the
targeted ultra-poor households, as well as the demand for such transfers by them. In
this section, I discuss how the various components of the program may affect both
the level of transfers received/given by ultra-poor households, as well as who they
interact with in transfer networks.
First, the asset and cash transfer components of the program may affect interhousehold
transfers received/given by the ultra-poor households. A large literature in public
economics discusses alternative mechanisms through which formal transfer programs
may crowd out private (informal) transfers, depending on the motivations driving in-
formal transfers. If they are driven by pure altruism (Becker (1974)), formal transfers
would unambiguously crowd out informal transfers (Barro (1974)). Alternatively, in-
formal transfers may take place in exchange for a service provided by the recipient
(Bernheim et al. (1985), Cox (1987)) or in anticipation of reciprocal future transfers
(Kranton (1996)). In that case, the effect of an increase in the amount of formal trans-
fers on the amount of informal transfers received is ambiguous.9 If informal transfers
are part of informal insurance arrangements between agents, an increase in formal
transfers received by an agent will partly be undone by transfers to insurance partners
of the agent.10
Second, the program enables the poor to develop new sources of income (mainly live-
stock rearing) and lowers their dependence on seasonal casual labor.11 This may im-

    9 Cox  et al. (1998) show that if transfers occur as a result of bargaining between the two parties
where the recipient provides some services to the donor in exchange of the transfer he receives, then
conditional on receiving a positive transfer, an increase in the recipient’s income would lead to an
increase in his outside option and thus may result in an increase in the informal transfer he receives.
Cox et al. (2004) show that, the combination of altruism and exchange motives may result in a non-
linear response to an increase in the income of the recipient.
   10 Under perfect risk-sharing (Townsend (1994)), any increase in the resources available to an agent

will enter the resource pool shared with his/her insurance partners and will increase the informal trans-
fers given by the agent. Therefore, if the agent was a net recipient of informal transfers ex-ante, the
increase in formal transfers s/he receives would lead to a decrease in the amount of informal transfers
s/he receives. Generalizing the perfect risk-sharing model to allow for imperfect insurance due to, for
instance, imperfect enforceability (Coate and Ravallion (1993), Ligon et al. (2002)) or asymmetric in-
formation (Ligon (1998)) would yield similar predictions, although the mechanism at work may differ.
Attanasio and Rios-Rull (2000) show that, under imperfect enforceability, introduction of unconditional
formal transfers in a situation where agents have very high marginal utility of consumption will induce
a reduction in the amount of equilibrium risk-sharing, which implies a lower level of informal transfers
for any given income shock.
   11 See Bandiera et al. (2017) for a thorough discussion of the types of jobs available to the ultra poor

women in this context and the seasonality associated with them.


                                                    8
prove their resilience to shocks (both by increasing the level of income and lowering
seasonality of earnings) and lower their need for informal insurance.
Third, the program provides health support to the beneficiary households. This may
partially lower the need for informal insurance (and lower transfers), both by lowering
the risk of health shocks and by enabling treated households to rely on BRAC’s health
services rather than receiving help from the social networks when faces with health
shocks.
Third, the program encourages and facilitates saving behavior among the targeted
ultra-poor households. These savings may act as an alternative source of insurance,
thus lowering the poor’s demand for informal insurance and crowd out their inter-
household transfers (Chandrasekhar et al. (2011), Flory (2018)).
Fifth, a unique aspect of the program is the creation of committees that bring together
the village elite with representative of the ultra-poor households (GDBCs). One of the
central aims of GDBCs is to encourage the local elites to help vulnerable ultra-poor
households in times of need. This may lead to more transfers being funneled from the
elites towards the ultra poor households, thus lowering any crowding out effect that
may occur due to the other components of the program.
Sixth, once the initial two years of the program are over and the ultra poor households
graduate, they are invited to participate in BRAC’s mainstream microfinance program.
Participation in microcredit has ambiguous effect on participation in informal trans-
fer networks. On one hand, participation in microfinance group meetings may lead
to creation of new social networks, increasing risk-sharing and transfers among the
poor (Feigenberg et al. (2013)). On the other hand, standard microcredit contract that
BRAC offers is rather restrictive, with regular loan repayments (Battaglia et al. (2019))
that may help shield the targeted poor from kinship taxes in the form of demands for
transfers from their social networks (Baland et al. (2011)).
The program may also affect who the ultra-poor households receive transfers from
or give transfers to. Previous work has shown that the program leads to long-term
improvements in socio-economic statuses of ultra-poor households (Bandiera et al.
(2017)). This may affect who ultra-poor households are matched with in transfer net-
works (Genicot (2006), Munshi and Rosenzweig (2016)). Moreover, an important as-
pect of the program is to create community networks and community support for
ultra-poor households, through (a) increasing the interactions among ultra-poor house-
holds (b) the establishment of GDBCs that bring together the village elite with repre-
sentative of the ultra-poor. Both of these features of the program may lead to the
creation of new transfer networks and/or weakening of existing ties.
An alternative source of transfers for the rural poor may be their employers. An em-

                                           9
inent literature in development economics studies interlinked labor and patronage
relationships in village economies (Bardhan (1983), Mukherjee and Ray (1995), Plat-
teau (1995a, 1995b)). When faced with a negative income shock, poor workers may
need a transfer or a loan from their employers. This demand for informal insurance
may affect the terms of the labor contract, generating an interlinkage between the la-
bor and insurance markets.12 The program, by improving the outside option(s) of the
poorest workers may enable them to transition from interlinked labor to casual labor
and/or self-employment. Thus, the program may reduce their likelihood to be receiv-
ing transfers from their employers.

4     Data Description
The dataset I use was collected in order to evaluate BRAC’s TUP program in Bangladesh.
The data covers 1409 communities, with the average community consisting of 90 house-
holds (387 individuals).13 At baseline, an initial census of all households was carried
out in every community, covering 126,810 households. This census collected informa-
tion on the identity as well as wealth, occupation, education and demographic char-
acteristics of all the households living in these communities at baseline.
Following the census, a detailed household questionnaire was carried out on a rep-
resentative sample of households. For the sampling, census data was combined with
information on households’ socio-economic statuses from the community wealth ap-
praisal BRAC conducted as part of their procedure of identifying the ultra-poor house-
holds (see Section 2). The sample for the household survey included all (ultra and
other) poor households and a random sample of the rest of the community. This
corresponds to 7953 ultra-poor households and an additional 19,012 non-ultra-poor
households. Households in this sample were surveyed at baseline (between April and
December 2007), at midline (2009) and endline (2011). As detailed in Section 2, the
implementation of the program lasts at most two years – less for some of its compo-
nents, such as the subsistence allowance which lasts for a maximum of 40 weeks. This
implies that by the time the midline survey was conducted, the program implemen-
tation was nearly completed in many communities, and by the endline survey it had
been completed in all communities for at least 1 year. An additional followup survey
was conducted in 2014 and I will use this data to examine the long-run effects of the
program for early versus late-treated ultra-poor households.
The household survey measured a number of individual outcomes, including occupa-

    12 More generally, the idea that a risk-neutral employer may provide a risk-averse worker with in-
surance against income fluctuations is not only limited to the rural labor market (Knight (1921), Baily
(1974), Azariadis (1975)).
    13 The districts and communities in which the data was collected were determined by BRAC, as part

of their aim to identify the poorest parts of the country. See Bandiera et al. (2017) for details.


                                                 10
tional choices, labor supply, income, social and economic networks of the household.
The main survey modules were directed towards the main female in the household,
who is the intended beneficiary in ultra-poor households. In cases where the main fe-
male was different from the household head, the head of the household was also sur-
veyed for the business activities and land modules. To capture the social and economic
networks of the household, respondents were asked to list households they interacted
with for each of the surveyed activities. For example, in the business activities mod-
ule, the respondent reported her main employer(s) for all income-generating activities
she was involved in. If they reported a household within the same community as their
own, the identity of this household was recorded (using the census listing).
To capture transfers exchanged between households, respondents were asked to re-
port any transfers their household received and any transfers their household gave
to others during the 12 months preceding each survey wave. For each transaction,
they were asked the value of the transfer, whether it was in cash or in kind, the loca-
tion of the sender/recipient,14 and the identity of the sender/recipient. If the transfer
source/recipient was a household residing in the same village as the respondent, their
identity was recorded using the census listing. During the piloting of the baseline sur-
vey, it was observed that many respondents reported major transfer transactions, but
left out smaller, in-kind transfers (which would only come up upon prompting by enu-
merators) that typically consisted of food items. In order to capture such transactions,
respondents were asked if their household ever received food from other households
and if so who the main sources/recipients of such transfers were. For every respon-
dent, up to three sources and three recipients of food transfers were recorded.15
Attrition: Over the four years from baseline to endline, 13% of ultra-poor households
and 15% of non-ultra-poor households attrited from the original sample. Table S1.1
in Appendix S1 tests for differential attrition among the ultra-poor in treatment and
control communities. Two findings are of note. First, attrition rates are the same in
treatment and control communities. Second, attrition is not differentially correlated
with the interaction of treatment status and the baseline levels of key outcomes of in-
terest: transfers received/given (neither on the extensive nor the intensive margin)
or the likelihood to have received transfers from employers. Therefore, to ease com-

   14 Whether the sender/recipient of the transfer resided in the same village (as the respondent’s house-

hold), different village/town in the same district, different district or outside Bangladesh
   15 One distinction between the food transfer questions and the larger transfer transactions is the for-

mer captures not only the realization of transfers in a given period but also the most relevant potential
transfer links. Therefore, by construction, the information on food transfers only captures the identity
while the latter includes information on both the value of the transfers and the identity of the transfer
sender/recipient. Whenever I estimate the effects on realized transfer transactions and their values, I
use only the information from the latter. On the other hand, whenever I estimate the effects on network
characteristics (such as the wealth status of transfer sources/recipients, or the reciprocity of connec-
tions), I use both types of transfer questions.


                                                   11
parability across different specifications, I will restrict the sample to households that
appear in all three waves throughout.

4.1   Poverty and Informal Transfers at Baseline
Table 1 presents baseline descriptive statistics on characteristics of households in our
sample. I use the wealth classes determined by the community appraisal at baseline
to divide the sample into three groups: poor, middle and upper class households. I
further divide the poor households into those who were deemed (by BRAC) to be el-
igible for the program (the ultra-poor) and those who were not found to be eligible
(the other poor). The first panel of the table shows that this classification (based on
the community-determined wealth classes and BRAC’s selection criteria) is associated
with significant differences in terms of the household’s economic status and the main
(female) respondent’s human capital endowment. As indicators of economic status,
I use the value of physical assets (land, livestock, other productive assets and house-
hold durables) and per capita household consumption. As proxies for human capital,
I use respondents’ (the primary female in the household) literacy and height. Table
1 shows that ultra-poor households have lower assets and consumption compared to
other wealth classes – including the other poor households. They also have lower lit-
eracy rates and the respondents from ultra poor households are shorter than women
from other wealth classes. The differences between the different wealth classes are sta-
tistically significant at conventional levels.16 This shows that the program’s selection
criteria was successful in identifying the poorest households in these communities.
Panel B of Table 1 shows summary statistics related to informal transfers received. On
average 22% of ultra-poor households reported receiving any transfers in the year
preceding the baseline survey. The corresponding rates were similar in other wealth
classes – except for the middle class where 16% of households had received any trans-
fers during the past year. However, on the intensive margin (i.e. in terms of the value
of transfers received), ultra-poor households had received significantly less compared
to other wealth classes. Relative to the ultra-poor households, the value of informal
transfers received by middle class households was 90% higher and the upper classes
had received nearly 12-times more. When asked whether they ever received food from
other households, 92% of ultra-poor households responded positively and they re-
ported, on average, 2 households as their most common sources of food transfers. The
corresponding figures are similar for the other-poor households but lower for wealth-
ier households. Only 43% of the wealthiest households in the community reported
ever receiving food transfers and on average they reported 1 household as source of
food transfers. The final row in Panel B shows the fraction of transfer sources (food
or otherwise) who were living within the same community as the respondent’s house-
  16 Table   S1.2 in Appendix S1 reports p-values of pairwise comparisons across the wealth classes.


                                                   12
hold. For the poor and middle classes, 88% of transfer sources were from the commu-
nity, while the corresponding figure was lower (70%) for the upper class households.
This suggests that the richest households in these communities are likely to have social
networks that expand beyond their locality and send remittances.17
Panel C of Table 1 summarizes the pattern of informal transfers given by the sampled
households. Only 2% of ultra-poor households made any transfers in the past year,
and only 44% reported ever giving food to other households. The corresponding rates
are higher for wealthier households. Among the richest households in the community,
21% had given some transfers to others in the past year and 81% reported ever giving
food transfers. The majority of these transfers were given to households within the
same community, albeit with some variation across the wealth classes.
Comparison of rates and levels of transfers received versus transfers given (Panel B
versus C) suggest that, especially for poor households in these communities, trans-
fers received are not always reciprocated with transfers given. In Panel D, I report the
fraction of transfer sources who are also reported as transfer recipients. Note that I
can only do this for within-community transfer sources, since the network-mapping
allows us to identify the identity of network members within (but not outside) the
community. Nevertheless, since the majority of the transfers happen within the com-
munity, especially for the poorer households, this should capture a significant share
of the transfer network. On average, only 42% of transfer sources of the ultra-poor
households were reciprocated (i.e. reported also as recipients of transfers), while the
corresponding rate was 50% for other poor, 67% for middle class and 80% for the rich-
est households in the community.18
The final row in Table 1 shows that 10% of the ultra-poor respondents reported an
employer (of the primary female or the male head of the household) as a source of
transfers. The corresponding rate was 8% among the other poor and 3% for middle
class households. This suggests that one potential source, especially for the poorest
households, may be their members’ employers. The literature on interlinked labor
contracts suggests that there may be a trade-off between earnings versus risk-sharing
motives. Bardhan (1983) shows that workers with limited outside options may end up
selecting jobs that provide them with a smoother income across the year but a lower
wage rate. Consistent with this, I find that receiving transfers from an employer is
correlated with having a lower hourly wage and less volatile annual wage earnings at
baseline.19
  17 Munshi   and Rosenzweig (2016) find a similar pattern in India. Their model attributes this pattern
to greater demand to share resources making wealthier households more likely to send migrants outside
the network (their focus is on the subcaste network).
   18 This pattern is in line with Schechter and Yuskavage (2011) who find that in Paraguay, unrecipro-

cated transfer links are more likely when one household is wealthier or more educated than the other.
   19 To estimate the correlation between the terms of the labor contract and receiving informal trans-



                                                  13
Table 2 presents an overview of how households from different wealth classes were
matched in transfer networks at baseline. Panel A shows percentages of households
receiving transfers from different socio-economic classes at baseline. 18% of ultra poor
reported another ultra poor household as a transfer source, while the corresponding
rates are lower for other wealth classes (10% for other poor, 5% for middle and 1% for
the upper class households). As we move up along the wealth classes likelihood of
receiving transfers from a given class rises. Interestingly, within every wealth class,
the likelihood to receive transfers from one’s own wealth class is the highest. For ex-
ample, 26% of respondents from the top wealth class reported other rich households
in their village as sources of transfers, while the corresponding rate is 21% for the ul-
tra poor. Similarly, 69% of the middle class respondents reported other middle class
households in their village as sources of transfers, while the corresponding rate is
66% for ultra-poor households (the difference is significant at 99% confidence level).
The lower panel shows the likelihood that a household reported giving transfers to
another, by wealth class of the respondent and the transfer recipient. The wealthier
households are more likely to give transfers to middle or upper class households com-
pared to the ultra poor: only 12% of the upper class respondents reported giving trans-
fers to an ultra poor household, while 52% and 29% reported middle or upper class
households. Similarly, among the middle class households, only 7% reported an ultra
poor as a transfer recipient, while 60% reported other middle class households. While
these patterns suggest that households are more likely to exchange transfers with sim-
ilar socio-economic classes as their own, they do not necessarily prove causality. Socio-
economic classes are likely to be endogenous to the structure of social networks within
a community. I will return to this in section 5.3 where I estimate the causal effect of
poverty on the way households are matched in transfer connections.

4.2    Randomization and Balancing at Baseline
To evaluate the TUP program, the timing of the program’s roll-out was randomized at
BRAC branch office level. BRAC determined 40 branch offices that would implement
the program. Standard procedures to identify who would be the beneficiaries of the
program were carried out (by BRAC program officers) in all of these branches in the
same way. Following the identification of potential beneficiary households, 20 branch
offices were randomly selected to receive the program in 2007, and the rest in 2011.

fers from the employer, I regress the worker’s wage rate on an indicator variable for having reported
an employer’s household as a source of transfers at baseline, controlling for observable chacteristics of
the worker (wealth, literacy, height, age, age-squared). Table S1.3 in the Appendix shows the results.
Workers who reported their employers as source of transfers for their household received significantly
lower hourly wages. Transfers from employers were associated with a 5% lower wage for female and
7% lower wage for male workers relative to the sample average. Moreover, workers receiving transfers
also had less volatile earnings from wage-labor but this correlation is imprecisely estimated at conven-
tional levels.


                                                   14
The randomization was stratified at the subdistrict level – within each subdistrict, one
branch was randomly allocated to treatment and one to the control group. All of the
selected communities in treatment branches were treated in 2007 while the control
communities were not treated until after the endline survey in 2011.
Table S1.4 reports baseline balance tests which compare the characteristics of ultra-
poor households in treatment and control groups at baseline. For every variable, the
table reports the mean and standard deviation in treatment and control communities,
as well as the p-value on a test of equality of means (column 7) and the normalized
difference of means (column 8). Overall, the samples are well balanced: only 2 out of
26 tests for the equality of the means are rejected; and all of the normalized differences
are below the .25 threshold recommended by Imbens and Wooldrige (2009). Thus we
can conclude that the randomization was successful and the control sample provides
a valid counterfactual for the treatment group.

5     Empirical Analysis
5.1     Estimation
In order to estimate the effects of the TUP program, I pool observations from the two
followup surveys and estimate an ANCOVA model. In particular, I estimate:

                           yidt = α + λ Ti + βyid0 + γSt=2 + δd +         idt ,                      (1)

where yidt is the outcome of interest for household i from subdistrict d at survey wave t
with time periods referring to 2007 baseline (t=0), 2009 midline (t=1) and 2011 endline
(t=2); Ti is an indicator variable equal to 1 if household i lived in a treatment branch
and 0 otherwise; yid0 is the baseline level of the outcome variable for household i;
and δd are subdistrict (randomization strata) fixed effects. The parameter of interest
is λ, the difference between treatment and control observations. The standard errors
are clustered at the BRAC branch office level (the unit of randomization) in all the
regressions. Under the identifying assumption that the control branches represent a
valid counterfactual for the treated branches in the absence of the program, namely
that trends in all outcomes of interest are the same in treatment and control branches,
λ identifies the causal effect of the TUP program on yidt . I will estimate (1) on the full
sample of ultra-poor households, thus estimating the intention to treat (ITT) effect.20
In order to test for the differences in treatment effects at midline and endline surveys,

    20 Of the households identified as ultra-poor in treatment branches, 86% eventually received an asset.

The other 14% either ceased to meet the eligibility criteria when transfers were implemented, or chose
not to take-up the program. As such, the ITT estimates reported in the paper are close to the average
treatment effect on the treated (ATT).


                                                   15
I estimate the following specification:

                    yidt = α + λ Ti + θ Ti · St=2 + βyid0 + γSt=2 + δd +            idt ,               (2)

where the main point of departure from specification 1 is the inclusion of the interac-
tion term Ti · St=2 . As such, in specification 2, λ corresponds to the treatment effect
at midline and θ estimates the difference between the treatment effect at endline com-
pared to midline. The results are reported in Appendix S2. In general, I find that the
estimates for θ are statistically and economically insignificant for our outcomes of in-
terest. In the few cases where this is not the case, I draw the reader’s attention to the
difference between midline and endline treatment effects.21

5.2       Effects on the Level of Informal Transfers
Table 3 presents findings on the impact of the program on transfers received/provided
by ultra-poor households during the 1past year. “Treatment” shows the estimate for λ
in specification (1). The first column of the table shows that the program reduced the
likelihood of receiving transfers by 13 percentage points (ppt), which corresponds to
a 26% reduction relative to the control group. On the intensive margin (column 2), the
value of transfers received by the ultra-poor was lower by BDT 105 (13% relative to
control) but this effect is imprecisely estimated. Columns 3 and 4 show that the pro-
gram did not have a significant impact on the likelihood or level of transfers provided
by ultra-poor households to others – the point estimates are positive, but imprecisely
estimated. As a result, the effect on net trasfers (columns 5 and 6) was a reduction.
Ultra-poor households in treatment branches were 13 ppt less likely to be net transfer
recipients and received, in net, BDT 151 less in transfers (24% relative to control).
The lower panels of Table 3 break down the transfers by the location of the households
giving or receiving transfers to/from the ultra-poor households. Panel B shows the
treatment effects on transfers exchanged with households within the same community,
and Panel C with those outside the village. I find that the ultra-poor in treatment
branches were 4ppt less likely to receive transfers from within their communities. On
the intensive margin, they received more transfers from their neighbors. In particular,
the value of transfers they received from households in their community was higher by
BDT 41, which corresponds to a 29% increase relative to the control group. There was
no significant treatment effect on transfers the ultra-poor households made to other
households within their communities (columns 3 and 4), and as such the net effect
(columns 5 and 6) was similar to the effect on transfers received. In contrast, Panel C
shows that transfers received from outside the village were significantly lower, both
   21 I
      also conducted a robustness check that controls for the interaction of the baseline level of the
outcome, yid0 , with the survey wave fixed effects, St=2 , in both specifications (1) and (2). The results are
robust to this additional control – available from the author upon request.


                                                    16
on the extensive (by 12ppt) and on the intensive margin (by BDT 144, or 23% relative
to control).22
Next, I analyze the effects on ultra-poor households’ participation in food transfers.
Table 4 presents the findings. On average, 94% of the control group reported that their
household received food transfers. The program led to a small (1.3ppt) increase in this
rate. Moreover, ultra-poor in treatment communities reported 0.10 more households
as sources of food transfers, a 4% increase relative to the control group. Looking at
the corresponding effect on food transfers given by ultra-poor households, columns 3
and 4 of the table show that ultra-poor were 7ppt more likely to give food transfers to
others and on average they transferred food to .23 more households (18% more relative
to control).23 The rest of the Table shows that these changes are driven by changes in
food transfers within the village (Panel B), while food transfers sources from outside
the village are slightly lower (Panel C).24
To summarize, the findings in this section suggest that the program crowded out in-
formal transfers received by ultra-poor households, but this effect is less pronounced
for within-village transfers. If anything, it led to a modest increase in the value of
transfers received and the number of food transfer sources within the community.

   22 In Appendix S2, I compare the treatment effects measured at midline and endline surveys. Table
S2.1 shows that the treatment effects at midline are not statistically different from those estimated at
endline. However, on the intensive margin, the large point estimates for θ suggest the crowding out
effect (on the intensive margin) is not diminishing, and possibly getting larger, over time. It is possible
to conduct the same analysis on informal loans exchanged with other households. Appendix Table S1.8
presents the treatment effects on interhousehold loans received/given by ultra poor households. The
program lowers the value of informal loans received by ultra poor households, while it increases both
the likelihood of giving loans and the value of loans given by ultra poor to other households. As a
result, in net, value of loans received by ultra poor household goes down, by nearly 100% relative to the
control group. This is true for both loans within and outside the village. These findings complement
the results reported in Bandiera et al. (2017) where I analyzed the treatment effects on likelihood of
receiving and giving loans (see Table IV). I showed that the program increased the likelihood of both
receiving and giving loans by the ultra-poor. The key difference between the results in Bandiera et al.
(2017) and those reported here is the fact that loans received in Bandiera et al. (2017) include formal
loans, whereas the results in Table S1.8 are limited to interhousehold loans alone.
    23 For food transfers, data on the intensive margin (value of transfers) was not collected. However, at

midline and endline surveys, respondents were asked to report whether their households’ frequency of
borrowing (lending) from (to) households reported as sources (recipients) of food transfers at baseline
had increased, decreased or remained the same. In Appendix Table S1.5 I use this information to test if
the program affected the frequency of borrowing or lending food to baseline network partners. Overall,
I fail to find a consistent effect. Ultra-poor in treated communities were (relative to control) less likely
to say their borrowing from baseline sources of food transfers had increased, but they were no more
likely to say it had decreased; and there was no significant change in their frequency of giving food to
baseline recipients either. Based on this, I conclude that the program did not crowd out food transfers
the ultra-poor exchanged with their baseline network members. On the other hand it led to a modest
increase (on the extensive margin) in food transfers the ultra-poor exchanged with other households
within their communities.
    24 Table S2.2 in Appendix S2 tests whether the treatment effects at midline differ from those at end-

line. The effects typically get larger in magnitude by the endline, the differences are never precisely
estimated.


                                                    17
5.3    Matching in Transfer Networks
In this section, I examine whether the program led to any changes in terms of who the
ultra-poor households exchanged transfers with. In particular, I estimate the effects of
the program on the likelihood that the ultra-poor exchange (receive or give) transfers
with households from different wealth classes in their communities. The first panel of
Table 5 shows that the program increased ultra-poor households’ likelihood to receive
transfers from other ultra-poor households by 6ppt (44% relative to control) and from
other poor households by 4ppt (10%). There was no significant effect on the likelihood
of receiving transfers from middle or upper class households. Panel B shows that the
program led to an increase on the likelihood of giving transfers to all wealth classes.
In particular, ultra-poor in treated communities were 5ppt (48%) more likely to give
transfers to households within their own wealth class, 3ppt (13%) more to other poor,
5ppt (14%) to middle and 2ppt (21%) more likely to provide transfers to upper class
households in their communities. 25
As an alternative test of effects on matching and to control for any self-reporting bias,
I limit the sample to non-ultra-poor households and estimate the treatment effects on
their likelihood to exchange transfers with ultra-poor and other wealth classes within
the community. For this, I first estimate specification (1) on the sample of non-ultra
households. The estimate for λ corresponds to the average treatment effect on all non-
ultra-poor households. Secondly, I estimate:

               yidt = ∑3       c    3            c
                       c=1 αc Ci + ∑c=1 λc Ti × Ci + β yid0 + γ St=2 + δd +            idt ,          (3)

where Ci is baseline wealth class of household i (other poor, middle or upper class).
In this specification, λc gives the treatment effect on wealth class c. Table 6 presents
the results. Panel A1 shows that non-ultra-poor households in treated communities
were on average 2ppt more likely to report receiving transfers from ultra-poor house-
holds within their communities. This corresponds to a large (37%) increase relative
to the control group. On the other hand, there was no significant impact on transfers
received from other wealth classes (columns 2-4). Panel A2 breaks down the effects by
the respondent’s wealth class, i.e. the estimates for λc in specification (3). Column 1
   25 Note  that at baseline, the ultra-poor households in the treatment group were less likely to re-
ceive/give transfers to upper class households compared to the ultra poor in the control group and
this difference was statistically significant (see Table S1.4). As such, the treatment effect on transfers
received/given to upper class households should be interpreted with caution. Having said that, the
direction of the baseline difference suggests that the treatment effects I estimate are likely to be lower
bound estimates of the true treatment effect. In Appendix Table S2.3 I assess the difference between
midline and endline treatment effects on non-ultra-poors’ likelihood to received transfers from dif-
ferent wealth classes. The results show no significant differences between the two surveys for most
outcomes, except for the likelihood to give transfers to the upper class. The treatment effect at endline
is significantly higher than at the midline. This suggest that it takes time for the ultra-poor households
to start providing transfers to the highest wealth classes.


                                                   18
shows that other poor households were 4ppt more likely to report receiving transfers
from ultra-poor in treated communities, corresponding to a 46% increase relative to
the control group. The middle classes were 2.5ppt (52%) more likely to receive trans-
fers from ultra-poor, while the corresponding effect was 1ppt (71%) for the wealthiest
class. The remaining columns in Panel A2 shows no significant effects for other wealth
classes. This implies that the increase in likelihood of transfers from ultra-poor house-
holds observed in column 1 did not crowd out non-ultra-poor households’ likelihood
of receiving transfers from other wealth classes26
The lower panel of Table 6 shows the effects on transfers given by non-ultra-poor
households to different wealth classes. On average, the non-ultra-poor households
were 3ppt more likely to report giving transfers to ultra-poor in their communities
(a 35% increase relative to the control group). Columns 2 to 4 show that the treat-
ment effects on likelihood of transfers to other wealth classes are negative but small
in magnitude and imprecisely estimated. Panel B2 breaks down the effects by wealth
class of the respondent’s household. Respondents from other poor and middle class
households were 3ppt more likely to report giving transfers to ultra-poor households.
While the corresponding effect for upper class households is also positive, it is impre-
cisely estimated.27 The rest of the table shows that there were no significant spillover
effects of the program on non-ultra-poor households’ likelihood to give transfers to
households from other non-ultra-poor households.28
In order to assess the long-run effects of the program, I use data from the third follow-
up survey that was conducted in 2014, that is seven years after the baseline. At that
point, 49% of control communities had also been treated by the program. Moreover,
only the female respondents were surveyed and detailed network information was not
collected. However, respondents from all wealth classes were asked whether they pro-
vided informal transfers to any of the ultra-poor households within their communities,

   26 Note that the point estimate of the “Treatment effect for upper class” in column 4 is large (-0.04) and

border-line significant (p-value=0.109), suggesting that the program may have crowded out transfers
received by upper class households from their own wealth class. In Appendix Table S2.4 I assess the
differences between midline and endline treatment effects on non-ultra-poors’ likelihood of receiving
transfers from different wealth classes. Two findings are of note: Based on the pooled specification
(Panel A1), the midline treatment effect on non-ultra-poor’s likelihood to receive transfers from ultra-
poor was 2ppt and significant, while this effect had increased to 3ppt by the endline (the difference
is marginally significant at 10% level). This suggests that the effects are increasing over time as the
ultra-poor are becoming more likely to provide transfers to the non-ultra-poor in their communities.
Secondly, at midline, I find that upper class households were 6ppt less likely to receive transfers from
other upper class households, but this effect had dissipated by the endline survey. This implies that in
the short run, the program crowded out likelihood that upper class households receive transfers from
their own wealth class.
    27 I cannot reject the null hypotheses that the treatment effects on middle and upper classes, or on

other poor and upper classes are equal (p-values are 0.448 and 0.287 respectively).
    28 Appendix Table S2.4 presents the results of comparing midline and endline treatment effects. Over-

all, I do not find significant differences between the treatment effects measured at midline and endline.


                                                    19
and if so to which ones. They were also asked whether they used to provide transfers
to any of the ultra-poor households seven years ago, and if so to which ones they used
to provide transfers to. I use these two pieces of information to test whether house-
holds in the early-treatment communities (i.e. treated in 2007) are differentially more
likely to provide transfers to ultra-poor households in their communities, and how
the effects vary by the wealth class of the respondent’s household. For this, I estimate
specification (3), using as yid0 the retrospective question on pre-treatment transfers.
Table 7 presents the results. The first column shows that non-ultra-poor households in
early-treatment communities were significantly more likely to report giving transfers
to ultra-poor households in their communities. In particular, the long-run treatment
effect for other poor households is 1.3ppt, for middle classes 2ppt and for the up-
per classes it is 3ppt. These effects are both statistically and economically significant
(very few non-ultra-poor households in the control group reported giving transfers to
the ultra-poor, as demonstrated by the means reported in the lower panel). The sec-
ond column shows that the percentage of ultra-poor households reported as receiving
transfers from middle and upper class households were higher by 0.5 and 1.2 ppt re-
spectively – corresponding to about 100% increase relative to the respective control
group means. This implies that the program led to a long-run increase in the likeli-
hood that ultra-poor households receive informal transfers from higher wealth classes
within their communities.
To sum up, I find that the program increased ultra-poor households’ likelihood to
give transfers, not only to other ultra-poor households but to households in wealthier
classes within their communities. One possible explanation for this could be that the
program, by improving ultra-poor’s economic conditions, may have enabled them to
reciprocate past transfers they used to receive from others. In line with this, I show
in Appendix Table S1.6 that the increase in ultra-poor’s transfers are largely driven by
an increase in their likelihood to provide transfers to households who at baseline used
to be their transfer sources. Households who used to provide transfers to ultra-poor
at baseline experience a larger increase in their likelihood to receive transfers from
ultra-poor in their communities. Nevertheless, there is also some significant, although
weaker, effects for other poor and middle class households who were not providing
transfers to the ultra-poor at baseline. Moreover, I showed above that the ultra-poor
also become more likely to receive transfers from wealthier classes within their com-
munities. This suggests that the effects are not only driven by ultra-poor reciprocating
for past transfers. Instead, some of the impact must be driven by a change in terms
of who the ultra-poor households are exchanging transfers with. In the following
section, I will test directly whether the program affected the reciprocity of informal
transfer links of the ultra-poor.


                                          20
5.4    Effects on Reciprocity of Informal Transfers
In order to estimate the treatment effect on reciprocity of transfers, I estimate specifica-
tion (1) using as the outcome variable the fraction of ultra-poor’s transfer sources who
are also reported as receivers of transfers. Table 8 presents the results. In the control
group, 51% of ultra-poor’s transfer sources were also reported as recipients of trans-
fers by the ultra-poor. The program led to an 8ppt increase (15%) in the reciprocity of
ultra-poor’s transfer connections. The following columns of the table break down this
measure according to the baseline wealth class of the connection. The results show
that the reciprocity of ultra-poor’s transfers increased with all wealth classes, and es-
pecially with households who were ranked in higher wealth classes at baseline. In par-
ticular, reciprocity of ultra-poor’s transfers increased by 9ppt (17%) with other poor,
by 9ppt (19%) with middle class and by 12ppt (44%) with upper class households. In
contrast, the increase in reciprocity of transfers with other ultra-poor was 4.4ppt (7%)
and marginally significant.29
An alternative source of transfers for ultra-poor households at baseline may be their
employer(s). Baseline correlations presented in section 4.1 suggested that such trans-
fers may be reciprocated by working for a lower wage rate, lowering the costs of labor
for the employers.30 In Table 9, I test whether the program affected ultra-poor house-
holds’ likelihood to receive transfers from their members’ employers. The dependent
variable in column 1 is an indicator for having received any transfers from an em-
ployer of either the main female respondent or the male head of the household. On
average, 9% of ultra-poor households in the control group reported an employer as a
source of transfers for their household. The program led to a fall in the likelihood of
receiving transfers from employers by 4 ppt. This effect is both statistically and eco-
nomically significant (43% relative to the control group). The next two columns show
the effect for the female and the male respondent(s) separately. Ultra-poor households
were 4ppt less likely to receive transfers from an employer of the main female re-
spondent, while the corresponding effect for the male respondents were 1.3ppt. While
these effects suggest that the improvement in socio-economic conditions of ultra-poor
households may have reduced their engagement in interlinked labor contracts, they
may also be driven by the fall in wage-labor caused by the program. In fact, Table S1.7
in Appendix S1 shows that the program significantly reduced labor supply of both fe-
male and male respondents into wage-labor (although the effect on the extensive mar-
gin is insignificant for the latter). As such, it is possible that the fall in likelihood of
   29 Appendix   Table S2.5 tests whether the treatments effects on reciprocity of ultra-poor’s transfer
links differ between the midline and endline surveys. The results show that the treatment effects were
significantly higher at endline relative to midline. This implies that the increase in reciprocity happened
gradually over time, and was only fully realized by the endline survey. In other words, it took time for
the ultra-poor households to establish reciprocal transfer links.
   30 See Table S1.3 discussed in footnote 19 above.



                                                   21
receiving transfers from employers may simply capture a fall in having an employer.
In order to assess whether the effect is driven by this, in Panel B of Table 9 I restrict the
sample to ultra-poor households who are engaged in wage-employment at baseline as
well as at the two followup surveys. Similarly, I find that households within this sub-
sample are less likely to receive transfers from their employers. Note that this analysis
has a caveat (since it is based on a selected sample), as such, it should be interpreted
as suggestive evidence. 31
Overall, the findings in this section imply that the program led to a change in the
mechanisms through which ultra-poor households reciprocate transfers from others.
The reduction in their poverty level enabled the ultra-poor to have reciprocal transfer
connections with other households in their communities, and reduced their likelihood
to receive transfers from their employers. The latter suggests that there was a reduc-
tion in the interlinkage of labor and insurance arrangements of the ultra-poor.

6      Discussion
6.1         Outside-village Transfers
One of the emerging findings is that the program crowded out outside-village trans-
fers. Unfortunately, I have limited information on the identify of outside village trans-
fer sources/recipients. One important piece of information about outside-village trans-
fers: I know if the transfers are given/received to/from the respondent’s first-degree
family members.32 Moreover, at baseline, respondents were asked to classify the
wealth status of their first-degree family members’ households relative to their own
(whether the family members are better-off, same or worse-off in terms of their wealth).
I exploit this information to shed some light on the changes in transfers outside the
village. First, note that a large share of the outside-village transfers consists of trans-
fers from first-degree family members. In particular, 63% (81%) of the outside-village
transfers received (given) by ultra-poor in the control group at midline/endline were
from (to) their first-degree family members living outside the village. Second, I an-
alyze the effects on outside-village transfers separately for family versus non-family
transfers, Appendix Table S1.9 displays the results. On the extensive margin, ultra-
poor households were significantly less likely to be net receivers of transfers from
outside the village and this seems to be driven mainly by non-family transfers. On
the intensive margin, I find that the effects on within-family transfers display a sim-
ilar pattern as outside-family transfers. Treated ultra-poor households received BDT
74 less in net transfers from their first-degree family members who live outside the
    31 In
        Appendix Table S2.6 I assess the differences in treatment effects on probability of having trans-
fers from employers at endline relative to midline. None of the differences are precisely estimated and
the treatment effects were similar at the two followup surveys.
   32 Parents, children, siblings and siblings-in-law.



                                                   22
village; and they received BDT 124 less from others. While the former is imprecisely
estimated, the two effects are statistically not different from one another.
Next, I distinguish the intra-family transfers by the relative wealth of the family mem-
bers at baseline. Appendix Table S1.10 displays the effects on transfers from/to family
members who were reported to be wealthier than the respondent’s household at base-
line in Panel A; and the transfers from/to family members whose wealth conditions
were same or worse-off as the respondent’s household at baseline. I find that treated
households give significantly more transfers (in value) to their wealthier family mem-
bers – the value of transfers given to wealthier family members is higher by BDT 6.3,
which is a nearly 2-fold increase relative to the mean in the control group (BDT 2.6).
This is in line with the effects on within-village transfer networks, where I showed
that the ultra-poor were more likely to give transfers to wealthier households in the
village. So, even though I do not have more detailed data to shed light on the full net-
work outside the village, the effects within the first-degree family network are largely
in line with what I find in terms of matching in within-village transfers.

6.2   Mechanisms
As described in Section 2, the program entails a number of components which may
affect ultra-poors’ informal transfers and who they interact with in risk-sharing net-
works within the community. This paper evaluates the effects of the entire program
package because the research design does not allow to isolate and contrast the impacts
of specific program components. Disentangling these mechanisms is beyond the scope
of the current paper. Nevertheless, I can exploit the variation in participation rates in
various program components in order to shed some light on the role of various pro-
gram components in crowding out of interhousehold transfers. Following Gelbach
(2016), I decompose the overall ITT impacts into components explained by different
potential mediators. The analysis provides useful suggestive evidence on which chan-
nels might contribute more significantly to the overall effects on transfers.
During the first 2 years of the program 86% of the ultra poor households participated
in the enterprise support (which includes asset transfer, training and cash transfer) and
all of them also received the encouragement to save. Hence, there is no independent
variation I can explore to disentangle the mediating effects of these four components.
On the other hand, 54% of the ultra poor received a sanitary latrine and/or tubewell,
34% received health support, 7% received help in paying schools fees and 30% re-
ceived some other type of support (e.g. assistance with paying costs to transport live-
stock to/from the nearby market). After the first 2 years of the program were over and
the treated ultra-poor households ‘graduated’ from the program, they were invited
to participate in BRAC’s mainstream microfinance program. 33% of the ultra-poor
households in the treatment group ended up participating in BRAC microfinance. In

                                           23
addition to these components, which are directly implemented by BRAC, 15% of the
ultra-poor households received some sort of support from the GDBC that were estab-
lished in treatment villages.
Appendix Table S1.11 displays the results of the mediation analysis. The first row in
each panel replicates the baseline ITT estimates (specification 1). The second row es-
timates the same ITT specification but also controls for the mediators. The difference
between these estimates corresponds to the total mediated effect, shown in the third
row. The remaining rows then show how much each mediator contributes to explain-
ing this mediated effect (assuming no complementarity between mediators). Three
findings emerge: First, having received enterprise support from BRAC (which com-
bines asset transfer, training, cash transfers and encouragement to save – since there
is no independent variation in participation in these components) is a significant me-
diator for crowding-out of informal transfers, on both the extensive and the intensive
margins. Second, having received health support from BRAC seems to crowd in infor-
mal transfers, at least on the intensive margin. Third, having microfinance loans from
BRAC is another significant mediator that is associated with crowding out of infor-
mal transfers. The findings that enterprise support and microcredit access crowd out
informal transfers is in line with the hypothesized effects of these components in Sec-
tion 3. While the positive association between health support of the program and the
crowding in of transfers may seem surprising, one explanation for this could be that
households who need health support from BRAC possibly need and are more likely to
receive support from their social networks as well. Healthcare costs can be rather high
in this context and health expenses are likely to be prioritized over other types of needs
by network members. While these results are intriguing, they are merely suggestive
due to endogeneity of participation in the program components. Further research that
opens up the black box of the program is needed to fully understand which compo-
nents are driving the effects of this multi-faceted program on interhousehold transfers
and to pin down the mechanisms behind these effects.33

6.3    Comparison of Magnitudes
Previous studies evaluating the effects of public transfer programs on informal trans-
fers have found relatively large crowding-out effects. Studying the effects of an in-
crease in old-age pensions in South Africa, Jensen (2003) finds that a unit increase in
the pension amount leads to a .25 to .30 reduction in private transfers. Juarez (2009)
finds that an exogenous increase in the income of elderly in Mexico City (through a
transfer program) led to a .33 reduction (per unit of public transfer) in private trans-

  33 See,for example, Banerjee et al. (2018) for evidence from Ghana on how disentangling some of
the components that are bundled together in BRAC’s approach may affect the program’s impact ultra-
poor’s socio-economic status.


                                               24
fers.34 Albarran and Attanasio (2002) evaluate the effects of a conditional cash transfer
(CCT) program that was implemented in rural communities in Mexico (“Progresa”).
They find that the grant, which had an average value of 250 pesos per household 35
reduced net transfers by 140 pesos per beneficiary household.36 Angelucci et al. (2012)
show that a CCT program implemented in urban Mexico (“Oppurtinades”) had a very
limited crowding-out effect on monetary transfer received, but reduced the likelihood
of in-kind transfers significantly (they don’t have data on the value of in-kind trans-
fers). Relative to these estimates, the crowding-out effect I find is smaller. I find a
fall of 13ppt in the likelihood of receiving transfers and BDT 152 decrease in the value
of net informal transfers received by the average household that was eligible for the
TUP program. This is low, relative to the value of assets transferred by the program
(BDT 9500). It is perhaps more relevant to benchmark the crowding-out effect on infor-
mal transfers relative to the effect of the program on annual household consumption.
When I estimate the effect of the program on consumption, using specification (1), I
find that the program led to an increase of BDT 1136 in annual per capita consump-
tion (p-value=0.000). The reduction in informal transfers is therefore small relative
to the improvement in economic conditions of ultra-poor households caused by the
program.
There are a number of possible explanations for the relatively small crowding-out ef-
fect I find on informal transfers. A key distinction between the TUP program and the
programs in the aforementioned studies is the multi-faceted nature of the TUP pro-
gram. As discussed in Section 3, many of these components may have opposing effects
on interhousehold transfers, limiting the extent of crowding-out. Another key feature
of the TUP program is that it is a one-off, big push intervention; while the transfer pro-
grams studied in previous work entail small but regular cash transfers. It is possible
that access to continuing transfers are perceived differently by social networks of the
beneficiaries, compared to a one-shot transfer, and therefore their crowding out effects
are different. Moreover, the program I study targeted some of the poorest households
in rural Bangladesh where out-migration is limited, especially for the poor (Bryan et
al. (2014)). In comparison, the households studied in previous work from South Africa
or Mexico had greater access to migrant networks and remittances. Furthermore, the
crowding out effect of CCT programs may be different due to the fact that these pro-
grams are conditional on investments in child education and health, whereas the TUP
  34 Juarez (2009) finds a larger, almost one-to-one, crowding-out effect at the individual level.
                                                                                           The data
does not allow me to distinguish transfers received by different members of the same household, thus
I cannot estimate the effects at the individual level.
    35 The amount of the transfer varied across beneficiary households, depending on the number of

children in the household.
    36 Albarran and Attanasio (2003) show that the crowding-out effect was larger in villages where

households’ earnings were less volatile (i.e. average variance of household income is lower), which
is in line with theoretical models of risk-sharing under imperfect enforcement.


                                                   25
program did not entail such conditionalities. In light of this, a more relevant com-
parison may be with unconditional cash transfer (UCT) programs, but there is limited
evidence on the long-term impacts of cash transfer programs in general (Haushofer
and Shapiro (2018)), and on their effects on private transfers in particular. Evaluations
of TUP and similar programs find improvements in the socio-economic statuses of
targeted household which are sustained in the long-run (Bandiera et al. (2017), Baner-
jee et al. (2016)). To the extent that the long-term effects of UCT or CCT programs
on the socio-economic statuses of targeted households differ from those of TUP-type
programs, their impacts on private transfers and social networks of the beneficiaries
are also likely to differ. Future research contrasting the effects of alternative transfer
programs on private transfers is needed to shed light on this issue.

7     Conclusion
Faced with highly volatile earnings and limited access to credit and insurance markets,
the poorest households in rural economies often rely on transfers from their social net-
works. It is important to understand how large-scale antipoverty programs targeting
poor households within a village affect their access to informal transfers and the struc-
ture of transfer networks within targeted communities.37
In this paper, I showed that a program that targeted ultra-poor households in rural
communities in Bangladesh led to a small reduction in the level of private transfers
they received. This crowding-out effect was driven mainly by a reduction in transfers
they received from outside the village, while there was a small increase in transfers
they received from their neighbors. The program enabled the ultra-poor households
to exchange transfers with households from other, wealthier socio-economic classes
within their communities. They were more likely to receive and give transfers from/to
households who, at baseline, were in higher wealth classes relative to the ultra-poor.
Finally, the program improved the reciprocity of ultra-poor households’ transfer con-
nections while reducing the likelihood that they receive transfers from their members’
employers.
The findings show that poverty has a causal effect on households’ participation in
transfer networks and who they are matched with in reciprocal transfer arrangements.
Moreover, the finding that the program reduces the likelihood of transfers from em-
ployers demonstrates an interlinkage between insurance and labor contracts, which is
particularly relevant for poor workers with limited outside options. This implies that
policies affecting either the labor or the insurance market are likely to have impact(s)
on the other one. Evaluations of labor and insurance policies in village economies
should take such interlinkages into account.
    37 Comolaand Prina (2017) show that accounting for changes in network structure is important for
estimating the treatment effects of development programs in rural communities.


                                                26
References
 [1] Alatas, V., A. V. Banerjee, R. Hanna, B. Olken and J. Tobias (2012). “Targeting the
     Poor: Evidence from a Field Experiment in Indonesia,” American Economic Review,
     102(4): 1206-40.

 [2] Ado, A. and T. Kurosaki (2014), “Motives for Sharing in Developing Countries:
     Experimental Evidence from Jakarta”, working paper.

 [3] Albarran, P. and O. Attanasio (2002), “Do Public Transfers Crowd Out Private
     Transfers? Evidence from a randomised experiment in Mexico”, in Insurance
     Against Poverty (S. Dercon, ed.), Oxford University Press.

 [4] Albarran, P. and O. Attanasio (2003), “Limited Commitment and Crowding out of
     Private Transfers: Evidence from a Randomised Experiment”, Economic Journal,
     113: C77-C85.

 [5] Altonji, J.G., F. Hayashi and L. J. Kotlikoff (1997), “Parental Altruism and Inter
     Vivos Transfers: Theory and Evidence”, Journal of Political Economy 105: 1121-66.

 [6] Ambrus, A., M. Mobius and A. Szeidl (2014), “Consumption Risk-Sharing in So-
     cial Networks", American Economic Review, 104(1): 149-182.

 [7] Anderson, J. (1990), “Rural Credit and The Mix Between Permanent and Tempo-
     rary Wage Labor Contracts in Pernumbuco, Brazil", American Journal of Agricul-
     tural Economics, 72(5): 1139-1145.

 [8] Angelucci. M. and G. de Giorgi (2009). “Indirect Effects of an Aid Program: How
     Do Liquidity Injections Affect Non-Eligibles’ Consumption?", American Economic
     Review 99: 486-508.

 [9] Angelucci, M., O. Attanasio and V. Di Maro (2012), “The Impact of Oportunidades
     on Consumption, Savings and Transfers”, Fiscal Studies, 33 (3): 305-334.

[10] Angelucci, M., G. de Giorgi and I. Rasul (forthcoming), “Consumption and In-
     vestment in Resource Pooling Family Networks”, Economic Journal, in press.

[11] Attanasio, O.P. and J.V. Rios Rull (2000), “Consumption Smoothing in Island
     Economies: Can Public Insurance Reduce Welfare?” European Economic Review,
     44(1): 225-258.

[12] Azariadis, C. (1975), “Implicit Contracts and Underemployment Equilibria", Jour-
     nal of Political Economy, 83(4): 1183-1202.



                                          27
[13] Baland, J.-M., C. Guirkinger, and C. Mali (2011), “Pretending to Be Poor: Borrow-
     ing to Escape Forced Solidarity in Cameroon.” Economic Development and Cultural
     Change, 60(1), 1-16.

[14] Bandiera, O., R. Burgess, N. Das, S. Gulesci, I. Rasul and M. Sulaiman (2017) “La-
     bor Markets and Poverty in Village Economies”, Quarterly Journal of Economics,
     132(2): 811-870.

[15] Banerjee, A. V., E. Duflo, N. Goldberg, D. Karlan, R. Osei, W. Pariente, J. Shapiro,
     B. Thuysbaert and C. Udry (2015), “A Multi-faceted Program Causes Lasting
     Progress for the Very Poor: Evidence from Six Countries,” Science 348: Issue 6236.

[16] Banerjee, A. V., E. Duflo, R. Chattopadhyay and J. Shapiro (2016), “The Long Term
     Impacts of a Graduation Program: Evidence from West Bengal,” Working Paper.

[17] Banerjee, A. V., D. Karlan, R. Osei, H. Trachtman and C. Udry (2018), “Unpacking
     a Multi-Faceted Program to Build Sustainable Income for the Very Poor,” Working
     Paper.

[18] Bardhan, P. (1983), “Labor-Tying in a Poor Agrarian Economy: A Theoretical and
     Empirical Analysis", The Quarterly Journal of Economics, 98(3): 501-514.

[19] Bardhan, P. and A. Rudra (1978), “Interlinkage of Land, Labor, and Credit Rela-
     tions: An Analysis of Village Survey Data in East India", Economic and Political
     Weekly, 13: 367-84.

[20] Bardhan, P. and A. Rudra (1981), “Types of Labour Attachment in Agriculture:
     Results of a Survey in West bengal, 1979", Oxford Bulletin of Economics and Statis-
     tics, 43: 89-111.

[21] Barr, A., M. Dekker, and M. Fafchamps (2012), “Who Shares Risk with Whom
     under Different Enforcement Mechanisms", Economic Development and Cultural
     Change, 60(4): 677-704.

[22] Barr, A. and G. Genicot (2007), “Risk Sharing, Commitment and Information: An
     Experimental Analysis", Journal of the European Economic Association, 6(6): 1151-
     1185.

[23] Barro, R. (1974), “Are Government Bonds Net Wealth?”, Journal of Political Econ-
     omy, 82(6): 1095-1117.

[24] Battaglia, M., S. Gulesci and A. Madestam (2019), “Repayment Flexibility and
     Risk Taking: Experimental Evidence from Credit Contracts", working paper.



                                          28
[25] BRAC (2016), “Targeting the Ultra Poor Programme Brief 2016: Challeng-
     ing the Frontiers of Poverty", http://www.brac.net/images/index/tup/brac_
     TUP-briefNote-Jun17.pdf.

[26] Becker, G. (1974), “A Theory of Social Interactions”, Journal of Political Economy,
     82(6): 1063-1093.

[27] Bernheim, B. D., A. Shleifer and L. H. Summers (1985), “The Strategic Bequest
     Motive”, Journal of Political Economy, 93(6): 1045 -1076.

[28] Blattman, C., E. P. Green, J. Jamison, M. C. Lehmann and J. Annan (2016), “The
     Returns to Microenterprise Support among the Ultrapoor: A Field Experiment in
     Postwar Uganda", American Economic Journal: Applied Economics, 8(2): 35-64.

[29] Bloch, F., G. Genicot, and D. Ray (2008), “Informal Insurance in Social Networks",
     Journal of Economic Theory, 143(1): 36-58.

[30] Bramoullé, Y. and R. Kranton (2007), “Risk-Sharing Networks", Journal of Eco-
     nomic Behaviour and Organization, 64(3-4): 275-94.

[31] Bryan, G., Chowdhury, S. and A. M. Mobarak (2014), “Underinvestment in a Prof-
     itable Technology: The Case of Seasonal Migration in Bangladesh”, Econometrica,
     82(5): 1671-1748.

[32] Casaburi, L. and R. Macchiavello (2016). “Firm and Market Response to Saving
     Constraints: Evidence from the Kenyan Dairy Industry”, working paper.

[33] Casaburi, L. and T. Reed (2016), “Interlinked Transactions and Pass-Through: Ex-
     perimental Evidence from Sierra Leone”, working paper.

[34] Caselli, F. (1997), “Rural Labor and Credit Markets”, Journal of Development Eco-
     nomics, 54(2): 235-260.

[35] Chandrasekhar, A., C. Kinnan and H. Larreguy (forthcoming) “Social Networks
     as Contract Enforcement: Evidence from a Lab Experiment in the Field”, American
     Economic Journal: Applied Economics, in press.

[36] Chandrasekhar, A., C. Kinnan and H. Larreguy (2011) “Information, networks
     and informal insurance: Evidence from a lab experiment in the field”, working
     paper.

[37] Collins, E. and E. Ligon (2017), “Asset Transfers and Household Neediness,”
     Working Paper.



                                          29
[38] Coate, S. and M. Ravallion (1993), “Reciprocity without Commitment: Character-
     izationf and Performance of Informal Insurance Arrangements”, Journal of Devel-
     opment Economics, 40: 1-24.

[39] Comola, M. and S. Prina (2017), “Treatment Effect Accounting for Network
     Changes”, working paper.

[40] Cox, D. (1987), “Motives for Private Income Transfers”, Journal of Political Econ-
     omy, 95(3): 508-546.

[41] Cox, D. and G. Jakubson (1995), “The Connection between Public Transfers and
     Private Interfamily Transfers”, Journal of Public Economics, 57: 129-167.

[42] Cox, D., Z. Eser and E. Jimenez (1998), “Motives for Private Transfers over the Life
     Cycle: An Analytical Framework and Evidence for Peru”, Journal of Development
     Economics, 55(1): 57-80.

[43] Cox, D., B. E. Hansen and E. Jimenez (2004), “How Responsive are Private Trans-
     fers to Income? Evidence from a Laissez-faire Economy”, Journal of Public Eco-
     nomics, 88: 2193-2219.

[44] Dercon, S., J. De Weerdt, T. Bold and A. Pankhurst (2006), “Group-Based Funeral
     Insurance in Ethiopia and Tanzania", World Development, 34(4): 685-703.

[45] De Weerdt, J., S. Dercon (2006), “Risk-sharing Networks and Insurance Against
     Illness", Journal of Development Economics, 81(2): 337-356.

[46] Fafchamps, M., F. Gubert (2007), “The formation of risk sharing networks", Jour-
     nal of Development Economics, 83: 326-350.

[47] Fafchamps, M., S. Lund (2003), “Risk-sharing Networks in Rural Phillipines",
     Journal of Development Economics, 71(2): 261-287.

[48] Feigenberg, B., E. Field and R. Pande (2013), “The Economic Returns to Social
     Interaction: Evidence from a Randomized Experiment with Microentrepeneurs",
     The Review of Economic Studies, 80(4): 1459-1483.

[49] Flory, J. (2018) ?Formal Finance and Informal Safety Nets of the Poor: Evidence
     from a Savings Field Experiment,” Journal of Development Economics, 135: 517-533.

[50] Gelbach, J. B. (2016), “When Do Covariates Matter? And Which Ones, and How
     Much.” Journal of Labor Economics, 34(2): 509-43.

[51] Genicot, G. (2006), “Does Wealth Inequality Help Informal Insurance?", working
     paper.

                                           30
[52] Genicot, G. and D. Ray (2003), “Group Formation in Risk-Sharing Arrangements",
     Review of Economic Studies, 70(1): 87-113.

[53] Imbens, G. W. and J. M. Wooldridge (2009), “Recent Developments in the Econo-
     metrics of Program Evaluation", Journal of Economic Literature, 47(1): 5-86

[54] Jalan, M. and M. Ravallion (1999), “Are the Poor Less Well Insured? Evidence on
     Vulnerability to Income Risk in Rural China", Journal of Development Economics,
     58: 61-81.

[55] Jensen, R. T. (2003), “Do private transfers ‘displace’ the benefits of public trans-
     fers? Evidence from South Africa”, Journal of Public Economics, 88: 89-112.

[56] Juarez, L. (2006), “Crowding out of Private Support to the Elderly: Evidence from
     a Demogrant in Mexico”, Journal of Public Economics, 93(3-4): 454-463.

[57] Haushofer, J. and J. Shapiro (2018), “The Long-term Impact of Unconditional
     Cash Transfers: Experimental Evidence from Kenya”, working paper.

[58] Knight, F. H. (1921), “Risk, Uncertainty and Profit”. New York: Houghton Mifflin.

[59] Kranton, R. (1996), “Reciprocal Exchange: A Self-Sustaining System”, American
     Economic Review, 86(4): 830 - 851.

[60] Kocherlakota, N. (1996), “Implications of Efficient Risk-Sharing without Commit-
     ment”, Review of Economic Studies, 63(4): 595-609.

[61] La Ferrara, E. (2003), “Kin Groups and Reciprocity: A Model of Credit Transac-
     tions in Ghana", American Economic Review, 93(5): 1730 -1751.

[62] Leider, S., M. Mobius, Rosenblat, T., Do, Q.-A. (2017) “Directed Altruism and
     Enforced Reciprocity in Social Networks”, Quarterly Journal of Economics, 124(4):
     1815-1851.

[63] Ligon, E., J. P. Thomas and T. Worrall (2002), “Informal Insurance Arrangements
     with limited Commitment: Theory and Evidence from village Economies”, Re-
     view of Economic Studies, 69: 209-244.

[64] Ligon, E. and L. Schechter (2012), “Motives for Sharing in Social Networks”, Jour-
     nal of Development Economics, 99(1): 13-26.

[65] McMillan and Woodruff (1999), “Interfirm Relationships and Informal Credit in
     Vietnam", Quarterly Journal of Economics, 114: 1285-1320.

[66] Mukherjee, A. and D. Ray (1995), “Labor tying", Journal of Development Economics,
     47: 207-239.

                                          31
[67] Munshi, K. and M. Rosenzweig (2016), “Networks and Misallocation: Insurance,
     Migration, and the Rural-Urban Wage Gap", American Economic Review, 106(1):
     46-98.

[68] Platteau, J.P. (1995a), “A Framework for the Analysis of Evolving Patron-Client
     Ties in Agrarian Economies,” World Development, 23(5): 767-786.

[69] Platteau, J.P. (1995b), “An Indian Model of Aristocratic Patronage,” Oxford Eco-
     nomic Papers, 47(4): 636-662.

[70] Platteau, J.P. and Abraham, A. (1987), “An inquiry into quasi-credit contracts: The
     role of Reciprocal Credit and Interlinked Deals in Small-scale fishing communi-
     ties”, Journal of Development Studies, 23: 461?90.

[71] Rosenzweig, M. (1988), “Risk, Implicit Contracts and the Family in Rural Areas
     of Low-income Countries", Economic Journal, 98: 1148-1170.

[72] Schechter, L. and A. Yuskavage (2011a), “Inequality, Reciprocity, and Credit in
     Social Networks,” American Journal of Agricultural Economics, 94(2): 402-410.

[73] Schechter, L. and A. Yuskavage (2011b), “Reciprocated Versus Unreciprocated
     Sharing in Social Networks,” working paper.

[74] Schoeni, R. F. (1996), “Does Aid to Families with Dependent Children Displace
     Familial Assistance?”, Rand Working Paper, DRU-1453-RC.

[75] Townsend, R. (1994), “Risk and Insurance in Village India,” Econometrica, 62(3):
     539-91.

[76] Udry, C. (1994), “Risk and Insurance in a Rural Credit Market: An Empirical
     Investigation in Northern Nigeria,” Review of Economic Studies, 61(3): 495-526.

[77] Walker, M. (2017), “Informal Taxation and Cash Transfers: Experimental Evi-
     dence from Kenya", Working Paper.




                                          32
8    List of variables
    • ‘Wealth’ – sum of the monetary values of land, livestock and durable assets owned
      by the respondent’s household.

    • ‘Consumption’ – total household consumption expenditure over the previous year
      divided by adult equivalent household size. The adult equivalence scale gives
      weight .5 to each child younger than 10. The expenditure items covered are:
      food, fuel, cosmetics, entertainment, transportation, utilities, clothing, footwear,
      utensils, textiles, dowries, education, charity and legal expenses.

    • ‘Literacy’ – a dummy variable =1 if the female respondent reported that she can
      read and write a letter.

    • ‘Height’ – gives the female respondent’s height in centimeters.

    • ‘Received any transfers during last year’ – a dummy variable equal to 1 if the re-
      spondent’s household received any transfers in cash or in kind during the 12
      months preceding the survey.

    • ‘Value of transfers received’ – monetary value of transfers received (in cash or in
      kind) by the respondent’s household during the 12 months preceding the survey.

    • ‘Gave any transfers during last year’ – a dummy variable equal to 1 if the respon-
      dent’s household gave any transfers in cash or in kind during the 12 months
      preceding the survey.

    • ‘Value of transfers given’ – monetary value of transfers made (in cash or in kind)
      by the respondent’s household during the 12 months preceding the survey.

    • ‘Ever receives food’ – a dummy variable equal to 1 if the respondent’s household
      ever receives food transfers from other households.

    • ‘Number of food transfer sources’ – number of households (up to a maximum of 3)
      who are reported as the main sources of food transfers received by the respon-
      dent’s household.

    • ‘Ever gives food’ – a dummy variable equal to 1 if the respondent’s household ever
      gives food transfers to other households.

    • ‘Number of food transfer recipients’ – number of households (up to a maximum
      of 3) who are reported as the main recipients of food transfers received by the
      respondent’s household.



                                            33
• ‘Fraction of transfer sources within village’ – fraction of households who are re-
  ported as sources of transfers (food or other) received by the respondent’s house-
  hold, who live within the same community as the respondent’s household.

• ‘Fraction of transfer recipients within village’ – fraction of households who are re-
  ported as recipients of transfers (food or other) given by the respondent’s house-
  hold, who live within the same community as the respondent’s household.

• ‘Reciprocity of transfers within-village’ – fraction of transfer sources within the vil-
  lage who are also reported as recipients of transfers (food or other).

• ‘Reciprocity of transfers with ... ’ – reciprocity of transfers with different wealth
  classes gives the fraction of transfer sources of a given wealth class within the
  village (conditional on having any) who are also reported as recipients of trans-
  fers.

• ‘Received transfers from any employer’ – a dummy variable equal to 1 if either the
  primary female respondent or the male head of the household is working for an
  employer who lives in the same community and is also reported as a source of
  transfers for the respondent’s household.




                                          34
                  Figure 1: BRAC’s Ultra-poor Graduation Program – Timeline

                                                                                                              Graduation
                                                                     Community Mobilization
                                                                       Healthcare Support
                                                                           Coaching
                                                                            Savings
                                 Consumption Allowance
                            Asset Transfer
                                                         Refresher                    Refresher   Refresher
                  Technical Skills Transfer
                                                          Training                     Training    Training
      Targeting
    0 Months                                             12 Months                                           24 Months
   Implementation start                                                                               Implementation end


Source: Adapted from http://www.brac.net/images/index/tup/brac_TUP-briefNote-Jun17.pdf
Notes: The figure shows the typical timeline of BRAC’s graduation program in Bangladesh. The specific
timeline may vary depending on the local conditions, as well as beneficiary requirements.




                                                               35
         Table 1: C OMPARISON OF S OCIO -E CONOMIC C LASSES AT B ASELINE
                                                     Ultra poor     Other poor      Middle class     Upper class
                                                         (1)           (2)              (3)             (4)
A. Socio-economic Status:
Wealth (BDT)                                            5850.98       14544.02        152293.69        841129.07
                                                      (30655.60)     (72425.83)      (315650.26)      (964643.22)
Consumption (BDT)                                       9829.35       10127.50         12205.72         20664.38
                                                       (4518.74)      (4771.70)       (7112.22)        (36441.19)
Literacy                                                  0.07           0.17            0.27             0.51
                                                         (0.26)         (0.37)          (0.44)           (0.50)
Height (cm)                                              148.71         149.13          149.82           150.13
                                                         (5.37)         (5.32)          (5.14)           (5.07)
B. Informal Transfers Received:
Received any transfers during last year                  0.22            0.22            0.16             0.22
                                                        (0.41)          (0.41)          (0.37)           (0.41)
Value of transfers received (BDT)                       270.08          328.30          513.29          3097.03
                                                      (1798.85)       (2708.87)       (4081.52)       (29555.99)
Ever receives food                                       0.92            0.92            0.83             0.42
                                                        (0.27)          (0.26)          (0.38)           (0.49)
Number of food transfer sources                          2.17            2.24            1.94             0.90
                                                        (0.95)          (0.95)          (1.11)           (1.17)
Fraction of transfer sources within village              0.88            0.88            0.88             0.70
                                                        (0.24)          (0.24)          (0.26)           (0.41)
C. Informal Transfers Given:
Gave any transfers during last year                      0.02            0.03            0.09             0.21
                                                        (0.14)          (0.18)          (0.28)           (0.41)
Value of transfers given (BDT)                          45.20           62.69           251.40           691.21
                                                      (1044.62)       (1323.43)       (2993.95)        (4273.22)
Ever gives food                                          0.44            0.54            0.73             0.81
                                                        (0.50)          (0.50)          (0.45)           (0.39)
Number of food transfer recipients                       1.00            1.27            1.70             1.91
                                                        (1.24)          (1.30)          (1.21)           (1.13)
Fraction of transfer recipients within village           0.95            0.94            0.92             0.87
                                                        (0.18)          (0.19)          (0.21)           (0.26)
D. Reciprocity:
Reciprocity of transfers within-village                   0.42           0.50            0.67             0.80
                                                         (0.44)         (0.44)          (0.42)           (0.35)
Received transfers from an employer                       0.10           0.08            0.03             0.00
                                                         (0.30)         (0.28)          (0.16)           (0.00)
Observations                                             6,732          7,340           6,742            2,215
Source: Author’s analysis based on original survey data.
Notes: The sample includes observations from the baseline survey. The sample is restricted to ultra-poor house-
holds in column 1, other poor households in column 2, middle class households in column 3 and upper class house-
holds in column 4. For variable definitions, see section 8. All monetary values are in Bangladeshi TAKAs. In 2007,
1USD=69BDT nominal and 1USD=18.5BDT at PPP. Appendix Table S1.2 reports p-values for tests of equality of
means across the four wealth classes.




                                                    36
           Table 2: M ATCHING IN T RANSFER N ETWORKS AT B ASELINE

                        Ultra poor Other poor Middle class Upper class
                            (1)           (2)            (3)              (4)
Panel A:                  Percentage of households receiving transfers from ...
      Ultra poor           17.60            9.70              4.41             1.26
      Other poor           32.58            39.51            17.49             6.00
      Middle class         65.57            66.08            68.78             23.52
      Upper class          21.33            23.50            21.51             26.14
Panel B:                      Percentage of households giving transfers to ...
      Ultra poor           14.05            8.83              7.25             11.51
      Other poor           21.60            31.43            21.76             28.89
      Middle class         28.30            36.09            60.25             51.74
      Upper class           3.12            5.26             11.23             28.76
Observations               6,732            7,340            6,742             2,215
Source: Author’s analysis based on original survey data.
Notes: The sample includes observations from the baseline survey. In column 1, the
sample is restricted to ultra-poor households; in column 2, to other poor households;
in column 3, to middle class households; and in column 4 to upper class households.
Panel A gives the percentage of households receiving transfers from the relevant group;
Panel B gives the percentage of households giving transfers to the relevant group. The
four rows within Panel A (B) show the percentage of respondents whose household
received (gave) any transfers to an ultra-poor (other poor, middle class or upper class)
in their village. Equality of means across the four wealth classes are all rejected at 99%
confidence (based on one-sided t-tests).




                                           37
   Table 3: E FFECTS ON I NFORMAL T RANSFERS OF U LTRA -P OOR H OUSEHOLDS

                                   Transfers received                Transfers given     Net transfers received
                                    in past 12 months              in past 12 months         in past 12 months
                                   (Yes=1)      (BDT)              (Yes=1) (BDT)          (Yes=1)        (BDT)
                                      (1)         (2)                 (3)        (4)          (5)          (6)
Treatment                         -0.129*** -104.664                 0.012     40.877 -0.131*** -151.721**
                                    (0.020)    (78.648)             (0.007) (45.704) (0.019)            (74.478)
Mean in control                      0.497     783.055               0.059    143.978       0.484       639.076
Adjusted R-squared                   0.126       0.027               0.030      0.001       0.121         0.019
Observations                         13464      13464                13464     13464        13464        13464
Panel B:                                                          Within-village Transfers
Treatment                         -0.037***         41.246*         -0.000     -4.120 -0.036***         45.558*
                                   (0.013)          (23.040)        (0.002) (4.078)        (0.013)      (23.656)
Mean in control                     0.148           142.403          0.015     11.865       0.147       130.538
Adjusted R-squared                  0.041             0.004          0.013      0.001       0.041         0.003
Observations                        13464            13464           13464     13464        13464        13464
Panel C:                                                          Outside-village Transfers
Treatment                         -0.123***        -144.864*        0.011*     45.021 -0.123*** -197.782***
                                   (0.019)          (72.655)        (0.006) (43.650) (0.018)            (65.197)
Mean in control                     0.430           640.652          0.048    132.113       0.420       508.539
Adjusted R-squared                  0.117             0.027          0.023      0.001       0.112         0.018
Observations                        13464            13464           13464     13464        13464        13464
Source: Author’s analysis based on original survey data.
Notes: The table reports ITT estimates based on specification (1) estimated with OLS. The sample includes observations from
ultra-poor households surveyed at the midline and endline surveys. All specifications control for the baseline level of the out-
come, survey wave and subdistrict (strata) fixed effects. Standard errors are clustered at the BRAC branch office level (unit of
randomization). All monetary values are measured in Bangladeshi Takas, deflated to 2007 prices using the annual CPI index
published by the Bank of Bangladesh. In 2007, 1USD=69BDT nominal and 1USD=18.5BDT at PPP. *** (**) (*) indicates significance
at the 1% (5%) (10%) level. The dependent variable in Panel A, column 1 is a dummy variable (DV) equal to 1 if the respondent’s
household received any transfers (in cash or in kind) from another household during the 12 months preceding the survey; in col-
umn 2 it is the monetary value of transfers received during the past 12 months; in column 3 it is a DV equal to 1 if the respondent’s
household gave any transfers to another household during the past 12 months; in column 4 it is the monetary value of all transfers
given by the respondent’s household during the last 12 months; in column 5 it is a DV equal to 1 if the respondent’s household
was a net transfer receiver in the past 12 months; in column 6 it is the monetary value of transfers received minus transfers made
by the respondent’s household in the past 12 months. All dependent variables in Panel B (C) are identical to those in Panel A,
except they refer to within (outside) village transfers.




                                                            38
Table 4: E FFECTS ON F OOD T RANSFERS OF U LTRA -P OOR H OUSEHOLDS

                   Transfers received        Transfers given
                   (Yes=1)     (No.)     (Yes=1)        (No.)
                      (1)       (2)         (3)           (4)
Treatment          0.013**    0.097*     0.069**      0.229***
                    (0.005)   (0.050)     (0.026)      (0.065)
Mean in control      0.936     2.340       0.560         1.284
Adjusted R-squared   0.069     0.169       0.260         0.314
Observations         13464     13464       13453        13464
Panel B:                     Within-village Transfers
Treatment          0.020*** 0.136** 0.068**           0.227***
                    (0.006)   (0.051)     (0.026)      (0.063)
Mean in control      0.907     2.199       0.547         1.236
Adjusted R-squared   0.078     0.193       0.259         0.318
Observations         13464     13464       13464        13464
Panel C:                     Outside-village Transfers
Treatment          -0.027** -0.034** -0.003             -0.001
                    (0.012)   (0.015)     (0.004)      (0.006)
Mean in control      0.107     0.141       0.037         0.047
Adjusted R-squared   0.065     0.074       0.037         0.036
Observations         13464     13464       13464        13464
Source: Author’s analysis based on original survey data.
Notes: The table reports ITT estimates based on specification (1) estimated with OLS. The sam-
ple includes observations from ultra-poor households surveyed at the midline and endline sur-
veys. All specifications control for the baseline level of the outcome, survey wave and subdistrict
(strata) fixed effects. Standard errors are clustered at the BRAC branch office level (unit of ran-
domization). *** (**) (*) indicates significance at the 1% (5%) (10%) level. The dependent variable
in column 1 is a dummy variable (DV) equal to 1 if the respondent’s household ever receives any
food transfers from other households The dependent variable in Panel A, column 1 is a dummy
variable (DV) equal to 1 if the respondent’s household ever receives any food transfers; in col-
umn 2 it is the number of households (capped at 3) that the respondent’s household receives
food transfers from; in column 3 it is a DV equal to 1 if the respondent’s household ever gives
food transfers to other households; in column 4 it is the number of households (capped at 3) that
the respondent’s households gives food transfers to. All dependent variables in Panel B (C) are
identical to those in Panel A, except they refer to within (outside) village food transfers.




                                               39
              Table 5: E FFECTS ON M ATCHING IN T RANSFER N ETWORKS

                    (1)           (2)            (3)            (4)
Panel A:          Transfers received by Ultra-Poor Households from ...
                ultra poor other poor middle class upper class
Treatment        0.056***      0.036***         0.018         -0.008
                  (0.007)       (0.012)        (0.012)       (0.008)
Mean in control    0.128         0.345          0.701          0.260
Adj. R-squared     0.439         0.415          0.327          0.407
Observations       13464         13464          13464         13464
Panel B:             Transfers given by Ultra-Poor Households to ...
                ultra poor other poor middle class upper class
Treatment        0.052***       0.032**       0.054***      0.016***
                  (0.008)       (0.013)        (0.019)       (0.005)
Mean in control    0.108         0.252          0.394          0.078
Adj. R-squared     0.368         0.306          0.254          0.145
Observations       13464         13464          13464         13464
Source: Author’s analysis based on original survey data.
Notes: The table reports ITT estimates based on specification (1) estimated with OLS. The sample includes
observations from ultra-poor households. All specifications control for the baseline level of the outcome,
survey wave and subdistrict (strata) fixed effects. Standard errors are clustered at the BRAC branch office
level (unit of randomization). *** (**) (*) indicates significance at the 1% (5%) (10%) level. In Panel A, the
dependent variables in columns 1, 2, 3 and 4 are dummy variables equal to 1 if the respondent reported
that her household received any transfers from any ultra-poor, other poor, middle class or upper class
households (within her community) respectively. In Panel B, the dependent variables in columns 1, 2, 3
and 4 are dummy variables equal to 1 if the respondent reported that her household gave any transfers to
any ultra-poor, other poor, middle class or upper class households (within her community) respectively.




                                                    40
Table 6: E FFECTS ON M ATCHING                       IN   T RANSFER N ETWORKS , R EPORTED                           BY    N ON -
U LTRA -P OOR H OUSEHOLDS
                                                                         (1)           (2)          (3)            (4)
   Transfers received by Non-Ultra-Poor from ...                        ultra         other       middle          upper
                                                                        poor          poor         class          class
   Panel A1: Pooled
   Treatment                                                         0.023*** 0.012                 0.001          -0.008
                                                                      (0.004) (0.008)              (0.013)        (0.007)
   Panel A2: By Wealth Class
   Treatment                                                         0.039*** 0.008                  0.003         0.007
                                                                      (0.007) (0.012)              (0.012)        (0.009)
   Treatment effect for middle class                                   0.025   0.013                 0.005         -0.001
                                                                      (0.005) (0.008)              (0.015)        (0.007)
   Treatment effect for upper class                                    0.012   0.003                -0.026         -0.040
                                                                      (0.006) (0.012)              (0.030)        (0.025)
   Control mean                                                        0.062   0.319                 0.691         0.279
   Mean for other poor                                                 0.085   0.448                 0.711         0.272
   Mean for middle class                                               0.048   0.224                 0.770         0.261
   Mean for upper class                                                0.017   0.128                 0.385         0.360
   Observations                                                        32594   32594                32594          32594
   Transfers given by Non-Ultra-Poor to ...                             ultra         other       middle          upper
                                                                        poor          poor         class          class
   Panel B1: Pooled
   Treatment                                                         0.025*** -0.011                -0.011         -0.002
                                                                      (0.005) (0.010)              (0.020)        (0.007)
   Panel B2: By Wealth Class
   Treatment                                                         0.031*** -0.002                -0.012         -0.006
                                                                      (0.007) (0.016)              (0.019)        (0.009)
   Treatment effect for middle class                                   0.027   -0.013               -0.019         -0.001
                                                                      (0.006) (0.011)              (0.022)        (0.008)
   Treatment effect for upper class                                    0.017   -0.027               -0.000         0.006
                                                                      (0.012) (0.018)              (0.022)        (0.016)
   Control mean                                                        0.078    0.338                0.600         0.177
   Mean for other poor                                                 0.084    0.397                0.507         0.119
   Mean for middle class                                               0.065    0.268                0.721         0.182
   Mean for upper class                                                0.091    0.329                0.584         0.373
   Observations                                                        32594   32594                32594          32594
   Source: Author’s analysis based on original survey data.
   Notes: The table reports ITT estimates based on specification (3) estimated with OLS. The sample includes observations
   from non-ultra-poor households. All specifications control for the baseline level of the outcome, survey wave and sub-
   district (strata) fixed effects. Standard errors are clustered at the BRAC branch office level (unit of randomization). ***
   (**) (*) indicates significance at the 1% (5%) (10%) level. In Panel A, the dependent variables in columns 1, 2, 3 and 4 are
   dummy variables equal to 1 if the respondent reported that her household received any transfers from any ultra-poor,
   other poor, middle class or upper class households (within her community) respectively. In Panel B, the dependent vari-
   ables in columns 1, 2, 3 and 4 are dummy variables equal to 1 if the respondent reported that her household gave any
   transfers to any ultra-poor, other poor, middle class or upper class households (within her community) respectively. In
   the pooled regressions (Panels A1 and B1) each observation is weighted using sampling weights calculated as the fraction
   of households surveyed from each wealth class (lower, middle and upper) relative to the number of households from the
   relevant wealth class in the community census.



                                                              41
          Table 7: L ONG -R UN E FFECTS ON M ATCHING IN T RANSFERS

                                   Whether gave                       Fraction of
                                  transfers to any                    ultra-poor
                                     ultra-poor                    households who
                                     household                    were given transfers
                                         (1)                               (2)
Treatment                              -0.004                            -0.057
                                       (0.005)                          (0.130)
Treatment × other poor                 0.017*                            0.322
                                       (0.010)                          (0.240)
Treatment × middle class              0.019***                          0.566**
                                       (0.007)                          (0.230)
Treatment × upper class               0.034***                          1.206**
                                       (0.011)                          (0.532)
Treatment effect for other poor         0.013                            0.265
                                       (0.006)                          (0.163)
Treatment effect for middle class       0.015                            0.509
                                       (0.004)                          (0.176)
Treatment effect for upper class        0.030                            1.149
                                       (0.009)                          (0.494)
Control mean                            0.012                            0.441
... for ultra poor                      0.008                            0.280
... for other poor                      0.007                            0.313
... for middle class                    0.012                            0.510
... for upper class                     0.034                            1.066
Observations                            21547                            21547
Source: Author’s analysis based on original survey data.
Notes: The table reports ITT estimates based on specification (1) estimated with OLS. The
sample includes observations from the long-run followup survey (2014). All specifications
control for subdistrict (strata) fixed effects, and the pre-treatment level of the outcome as
reported (retrospectively) in the 2014 survey. Standard errors are clustered at the BRAC
branch office level (unit of randomization). *** (**) (*) indicates significance at the 1% (5%)
(10%) level. The dependent variable in column 1 is a dummy variable equal to 1 if the
respondent’s household gave any transfers or loans to any ultra-poor household in her vil-
lage. The dependent variable in column 2 is the fraction of ultra-poor households living in
the respondent’s village who received transfers or loans from the respondent’s household,
multiplied by 100 (so that the effects correspond to percentage points).




                                             42
          Table 8: E FFECTS ON R ECIPROCITY OF U LTRA -P OOR ’ S T RANSFERS

                   Reciprocity       Reciprocity of transfers with
                   of transfer  ultra    other     middle       upper
                      links     poor      poor       class       class
                        (1)      (2)       (3)        (4)         (5)
Treatment           0.076***   0.044* 0.094*** 0.086***        0.117***
                     (0.018)   (0.024) (0.020)     (0.020)      (0.026)
Mean in control       0.510     0.627    0.560      0.449        0.267
Adjusted R-squared    0.217     0.189    0.208      0.235        0.192
Observations          12774     2792      4823       9474        3135
Source: Author’s analysis based on original survey data.
Notes: The table reports ITT estimates based on specification (1) estimated with OLS. The sample includes obser-
vations from ultra-poor households surveyed at the midline and endline surveys. All specifications control for the
baseline level of the outcome, survey wave and subdistrict (strata) fixed effects. Standard errors are clustered at
the BRAC branch office level (unit of randomization). *** (**) (*) indicates significance at the 1% (5%) (10%) level.
The dependent variable in column 1 is the fraction of transfer sources within the village who are also reported as
recipients of transfers (of last 12 months’ or food transfers ever) from the respondent’s household. The dependent
variables in columns 2 - 5 are the fraction of ultra-poor, other poor, middle class or upper class households (re-
spectivey) who are reported as sources of transfers (food, cash or other) and also reported as recipients of transfers
given by the respondent’s household.




                                                        43
             Table 9: E FFECTS ON T RANSFERS FROM E MPLOYERS

                     Household received transfers from...
                       any          female             male
                   employer(s) respondent’s respondent’s
                                 employer(s) employer(s)
                        (1)            (2)              (3)
Treatment           -0.039***     -0.035***         -0.013***
                     (0.008)        (0.008)          (0.004)
Mean in control       0.091          0.081            0.032
Adjusted R-squared    0.068          0.059            0.039
Observations          13464          13464             7778
Panel B:             Sample of wage-workers in all 3 surveys
Treatment           -0.040***        -0.026           -0.019
                     (0.013)        (0.016)          (0.015)
Mean in control       0.133          0.138            0.073
Adjusted R-squared    0.079          0.071            0.077
Observations           6336           4488             1332
Source: Author’s analysis based on original survey data.
Notes: The table reports ITT estimates based on specification (1) estimated with OLS. The sam-
ple includes observations from ultra-poor households surveyed at the midline and endline sur-
veys. All specifications control for the baseline level of the outcome, survey wave and subdistrict
(strata) fixed effects. Standard errors are clustered at the BRAC branch office level (unit of ran-
domization). *** (**) (*) indicates significance at the 1% (5%) (10%) level. The dependent variable
in column 1 is a dummy variable (DV) equal to 1 if either the main female or the male head of the
household had worked (in the last 12 months) for an employer who lives in the same village and
is reported as a source of transfers for the household. The dependent variable in column 2 (3) is a
DV equal to 1 if the main female respondent (male head of the household) had worked (in the last
12 months) for an employer who lives in the same village and is reported as a source of transfers
for the household. In Panel B, the sample is restricted to ultra-poor households where either the
main female respondent or the male head of the household had worked for an employer in the 12
months preceding the survey.




                                               44
                                                     A PPENDIX :
   P OVERTY A LLEVIATION AND I NTERHOUSEHOLD T RANSFERS :
E VIDENCE FROM BRAC’ S G RADUATION P ROGRAM IN B ANGLADESH

                                                    Selim Gulesci




S1: Additional Tables
                                               Table S1.1: ATTRITION
                                                                    (1)          (2)        (3)        (4)        (5)          (6)
     Treatment                                                    -0.015       -0.014     -0.014     -0.013     -0.012       -0.011
                                                                  (0.013)     (0.012)     (0.011)    (0.011)    (0.011)      (0.013)
     Received any transfers                                        0.023                                                      0.012
                                                                  (0.014)                                                    (0.018)
     Treatment × Received any transfers                            0.010                                                      0.009
                                                                  (0.022)                                                    (0.030)
     Value of transfers received                                             0.032***                                        0.026**
                                                                              (0.009)                                        (0.012)
     Treatment × Value of transfers received                                   -0.006                                        -0.007
                                                                              (0.014)                                        (0.018)
     Gave any transfers                                                                   -0.045                            -0.311**
                                                                                          (0.040)                            (0.116)
     Treatment × Gave any transfers                                                       -0.024                             -0.036
                                                                                          (0.050)                            (0.137)
     Value of transfers given                                                                        -0.266                  1.497*
                                                                                                     (0.346)                 (0.749)
     Treatment × Value of transfers given                                                             0.468                   0.565
                                                                                                     (0.461)                 (0.995)
     Received any transfer from an employer                                                                     -0.001       -0.003
                                                                                                                (0.020)      (0.020)
     Treatment × Received transfer from employer                                                                -0.024       -0.027
                                                                                                                (0.025)      (0.025)
     Attrition in control                                          0.162       0.162       0.162      0.162      0.162        0.162
     F-test (p-value)                                                                                                         0.842
     Adjusted R-squared                                            0.004       0.005       0.003      0.003      0.003        0.007
     Observations                                                  7953        7953        7953       7953       7953         7953
     Source: Author’s analysis based on original survey data.
     Notes: The dependent variable is a dummy variable equal to 1 if the respondent has not been surveyed at midline and/or
     endline survey waves. The sample includes all ultra-poor households. All specifications control for subdistrict fixed effects,
     a dummy variable (DV) equal to 1 if the control variable is missing and the interaction of this DV with the treatment
     dummy. Standard errors are clustered at the BRAC branch office level (unit of randomization). ‘Received any transfers’ is
     a DV equal to 1 if the respondent’s household received any transfers in cash or in kind during the 12 months preceding the
     survey. ‘Gave any transfers’ is a DV equal to 1 if the respondent’s household gave any transfers in cash or in kind during the
     past 12 months. ‘Value of transfers received (given)’ is the monetary value of transfers received (given) by the respondents’
     household from (to) other households during the past 12 months; expressed in 1000s of Bangladeshi TAKAs, top 1% of
     outliers have been coded to missing. In 2007, 1USD=69BDT. ‘Received transfers from any employer’ is a dummy variable
     equal to 1 if either the female or the male respondent is working for an employer who lives in the same community and is
     also reported as a source of transfers for the respondent’s household. In column (6), F-test gives the p-value associated with
     the null hypothesis that all interaction terms are jointly equal to zero. *** (**) (*) indicates significance at the 1% (5%) (10%)
     level.




                                                                    1
                                                                Table S1.2: C OMPARISON OF S OCIO -E CONOMIC C LASSES
                                                          Ultra poor (UP) Other poor (OP) Middle class (M)                  Upper class (R)      UP vs. OP      UP vs. M     UP vs. R     OP vs. M OP vs. R          M vs. R
                                                                 (1)            (2)             (3)                              (4)                (5)           (6)          (7)          (8)      (9)              (10)
    A. Socio-economic Status:
    Wealth (BDT)                                               5850.98              14544.02              152293.69            841129.07             .000          .000         .000          .000         .000        .000
                                                             (30655.60)            (72425.83)            (315650.26)          (964643.22)
    Consumption (BDT)                                          9829.35              10127.50               12205.72             20664.38             .000          .000         .000          .000         .000        .000
                                                              (4518.74)             (4771.70)             (7112.22)            (36441.19)
    Literacy                                                     0.07                  0.17                  0.27                 0.51               .000          .000         .000          .000         .000        .000
                                                                (0.26)                (0.37)                (0.44)               (0.50)
    Height (cm)                                                 148.71                149.13                149.82               150.13              .000          .000         .000          .000         .000        .018
                                                                (5.37)                (5.32)                (5.14)               (5.07)
    B. Informal Transfers Received:
    Received any transfers during last year                      0.22                  0.22                  0.16                  0.22              .945          .000         .723          .000         .684        .000
                                                                (0.41)                (0.41)                (0.37)                (0.41)
    Value of transfers received (BDT)                           270.08                328.30               513.29                3097.03             .137          .000         .000          .001         .000        .000
                                                              (1798.85)             (2708.87)             (4081.52)            (29555.99)
    Ever receives food                                           0.92                  0.92                  0.83                  0.42              .389          .000         .000          .000         .000        .000
                                                                (0.27)                (0.26)                (0.38)                (0.49)
    Number of food transfer sources                              2.17                  2.24                  1.94                  0.90              .000          .000         .000          .000         .000        .000
2




                                                                (0.95)                (0.95)                (1.11)                (1.17)
    Fraction of transfer sources within village                  0.88                  0.88                  0.88                  0.70              .769          .441         .000          .608         .000        .000
                                                                (0.24)                (0.24)                (0.26)                (0.41)
    C. Informal Transfers Given:
    Gave any transfers during last year                          0.02                  0.03                  0.09                  0.21              .000          .000         .000          .000         .000        .000
                                                                (0.14)                (0.18)                (0.28)                (0.41)
    Value of transfers given (BDT)                              45.20                 62.69                251.40                 691.21             .387          .000         .000          .000         .000        .000
                                                              (1044.62)             (1323.43)             (2993.95)             (4273.22)
    Ever gives food                                              0.44                  0.54                  0.73                  0.81              .000          .000         .000          .000         .000        .000
                                                                (0.52)                (0.50)                (0.45)                (0.39)
    Number of food transfer recipients                           1.00                  1.27                  1.70                  1.91              .000          .000         .000          .000         .000        .000
                                                                (1.24)                (1.30)                (1.21)                (1.13)
    Fraction of transfer recipients within village               0.95                  0.94                  0.92                  0.87              .054          .000         .000          .000         .000        .000
                                                                (0.18)                (0.19)                (0.67)                (0.26)
    D. Reciprocity:
    Reciprocity of transfers within-village                      0.42                  0.50                  0.67                  0.80              .000          .000         .000          .000         .000        .000
                                                                (0.44)                (0.44)                (0.42)                (0.35)
    Received transfer from an employer                           0.10                  0.08                  0.03                  0.00              .001          .000         .000          .000         .000        .000
                                                                (0.30)                (0.28)                (0.16)                (0.00)
    Observations                                                6,732                 7,340                 6,742                 2,215
    Source: Author’s analysis based on original survey data.
    Notes: The sample includes observations from the baseline survey. The sample is restricted to ultra-poor households in column 1, other poor households in column 2, middle class households in column 3 and upper class
    households in column 4. For variable definitions, see section 8. All monetary values are in Bangladeshi TAKAs. In 2007, 1USD=69BDT. Columns 5-10 report p-values for the equality of means across the four wealth classes
    (based on one-sided t-tests).
Table S1.3: C ORRELATIONS BETWEEN H AVING T RANSFERS                                 FROM     E MPLOYERS        AND
T ERMS OF L ABOR C ONTRACTS
                                        Women’s labor contracts:                          Men’s labor contracts:
                                Hourly     Range of          Range /             Hourly    Range of           Range /
                                 wage     wage income hourly wage                 wage    wage income hourly wage
                                  (1)          (2)               (3)               (4)         (5)                 (6)
 Interlinked contract          -0.277***    -88.121*          -8.370            -0.705***   -101.017             -0.947
                                (0.099)     (47.645)          (8.051)            (0.154)    (74.337)           (8.341)
 Observations                    6358         6358             6358               7988        7984                7984
 Mean level of outcome           5.743     1218.879          221.179              9.643     2271.537           252.602
 Adjusted R-squared              0.217        0.113            0.067              0.268       0.074              0.073
 Source: Author’s analysis based on original survey data.
 Notes: The table reports baseline correlations between having received transfers from employers and contract terms (wage per
 hour and volatility (range) of monthly wage earnings). ‘Received transfers from an employer’ is a dummy variable equal to 1
 if respondent is working for an employer who lives in the same community and who is also reported as a source of transfers
 for the respondent’s household. Sample includes baseline observations from ultra-poor and non-ultra-poor households. Each
 observation is weighted using sampling weights calculated as the fraction of households surveyed from each wealth class (lower,
 middle and upper) relative to the number of households from the relevant wealth class in the community census. Columns 1
 - 3 refer to terms of labor contracts of primary female in the household and columns 4 - 6 refer to terms of labor contracts of
 the male household head. All regressions control for the following characteristics of the relevant individual [primary female in
 columns 1 - 3 and male household head in columns 3 - 6] : whether the individual is literate, individual’s age (in years) and age
 squared, total household wealth (in BDT) and female respondent’s height (in centimeters) – male respondent’s height was not
 available, thus we control for female respondent’s height as a proxy. In columns 1 and 4, the dependent variable is wage per
 hour (in BDT), which is the sum of earnings from wage labor divided by hours spent in wage-labor during the year preceding the
 survey. In columns 2 and 5, the dependent variable is the range of monthly wage earnings, which is the difference between the
 maximum and minimum monthly earnings of the respondent from wage employment during the year preceding the survey. In
 columns 3 and 6, Ithe dependent variable is the range of monthly wage earnings divided by the wage per hour (in BDT). In 2007,
 1USD=69BDT. *** (**) (*) indicates significance at the 1% (5%) (10%) level.




                                                          3
                                                              Table S1.4: B ASELINE B ALANCING
                                                                                       Treatment Group                            Control Group                            Difference
                                                                                   Mean       SD        N                    Mean       SD       N                   t-test Normalized
                                                                                    (1)       (2)      (3)                    (4)       (5)     (6)                    (7)         (8)
    A. Socio-economic Status:
    Wealth (BDT)                                                                  5373.04       20145.37 4,045             6570.46        41750.82 2,687 0.449                     -0.026
    Consumption (BDT)                                                             9921.14       4411.00 3,822              9687.54         4677.66 2,474 0.552                     0.036
    Literacy                                                                       0.07           0.26   4,023              0.07            0.25   2,675 0.624                     0.017
    Height (cm)                                                                   148.72          5.36   3,773             148.70           5.39   2,456 0.941                     0.002
    B. Informal Transfers Received:
    Received any transfers during last year                                        0.20           0.40          4,045        0.25           0.43          2,687 0.251              -0.090
    Value of transfers received (BDT)                                             223.97         1189.77        4,045       339.49         2443.29        2,687 0.260              -0.043
    Ever receives food                                                             0.93           0.26          4,043        0.91           0.28          2,687 0.659              0.028
    Number of food transfer sources                                                2.11           0.93          4,045        2.27           0.96          2,687 0.159              -0.126
    Fraction of transfer sources within village                                    0.90           0.23          3,798        0.86           0.26          2,523 0.141              0.102
    C. Informal Transfers Given:
    Gave any transfers during last year                                             0.02           0.14         4,045        0.02           0.14          2,687 0.716              -0.010
    Value of transfers given (BDT)                                                 34.72          734.69        4,045        60.97         1386.18        2,687 0.465              -0.017
    Ever gives food                                                                0.46            0.50         4,043        0.41           0.49          2,687 0.471              0.064
4




    Number of food transfer recipients                                             1.02            1.22         4,045        0.98           1.26          2,687 0.822              0.021
    Fraction of transfer recipients within village                                 0.96            0.17         1,867        0.94           0.20          1,123 0.247              0.061
    D. Reciprocity:
    Reciprocity of transfers within-village                                         0.43            0.44        3,711         0.41           0.44         2,442      0.677         0.033
    Reciprocity with ultra-poor HH’s                                                0.65            0.47         882          0.57           0.49          303       0.213         0.119
    Reciprocity with other poor HH’s                                                0.54            0.49        1,271         0.51           0.49          922       0.537         0.052
    Reciprocity with middle class HH’s                                              0.36            0.46        2,584         0.36           0.46         1,831      0.985         -0.002
    Reciprocity with upper class HH’s                                               0.12            0.32         774          0.18           0.37          662       0.195         -0.109
    Received transfer from an employer                                              0.10            0.30        4,045         0.10           0.30         2,687      0.920         0.004
    E. Likelihood of transfers from/to other wealth classes:
    Received any transfers from...
    ... other poor                                                                  0.31            0.46        4,045         0.34           0.47         2,687 0.389              -0.043
    ... middle class                                                                0.64            0.48        4,045         0.68           0.47         2,687 0.204              -0.063
    ... upper class                                                                 0.19            0.39        4,045         0.25           0.43         2,687 0.015              -0.095
    Gave any transfers to...
    ... other poor                                                                  0.22            0.41        4,045         0.22           0.41         2,687 0.976              -0.002
    ... middle class                                                                0.28            0.45        4,045         0.29           0.45         2,687 0.807              -0.019
    ... upper class                                                                 0.02            0.15        4,045         0.04           0.20         2,687 0.053              -0.079
    Source: Author’s analysis based on original survey data.
    Notes: The sample includes observations from the baseline survey and is limited to ultra-poor households. Columns 1,2,3 (4,5,6) give the mean, strandard deviation and number of
    observations respectively for the relevant variable in treatment (control) communities. Column 7 reports the p-value for the test of equality of treatment and control samples’ means with
    the standard errors clustered at the branch office level (unit of randomization); column 8 reports normalized differences computed as the difference in means in treatment and control
    observations divided by the square root of the sum of the variances. For variable definitions, see section 8. All monetary values are in Bangladeshi TAKAs. In 2007, 1USD=69BDT.
                 Table S1.5: U LTRA -P OOR ’ S F OOD T RANSFERS WITH T HEIR B ASELINE N ETWORK
                           (1)           (2)             (3)                                       (4)            (5)            (6)
                          Fraction of baseline food transfer                                      Fraction of baseline food transfer
                           sources from whom transfers...                                          recipients to whom transfers...
                       decreased unchanged increased                                           decreased unchanged increased
    Treatment            -0.003         0.034         -0.031**                                   -0.053          0.065         -0.012
                        (0.036)        (0.035)         (0.014)                                  (0.050)         (0.050)       (0.008)
    Mean in control       0.398         0.514           0.088                                     0.506          0.457         0.038
    Adjusted R-squared    0.048         0.053           0.053                                     0.131          0.140         0.085
5




    Observations         12180          12180           12180                                     3077           3077           3077
    Source: Author’s analysis based on original survey data.
    Notes: The table reports ITT estimates based on specification (1) estimated with OLS. The sample includes observations from ultra-poor
    households surveyed at the midline and endline surveys. All specifications control for survey wave and subdistrict (strata) fixed effects.
    Standard errors are clustered at the BRAC branch office level (unit of randomization). The dependent variable in column 1 (2) is a dummy
    variable equal to 1 if the amount of food transfers received from the baseline food transfer sources decreased (increased) at midline/endline.
    The dependent variable in column 3 (4) is a dummy variable equal to 1 if the amount of food transfers given to baseline food transfer recipients
    decreased (increased) at midline/endline.
          Table S1.6: E FFECTS ON M ATCHING , H ETEROGENEITY BY B ASELINE C ONNECTIONS

    Reported by :                       Ultra-Poor Households              Non-Ultra-Poor Households
                                  (1)        (2)       (3)        (4)        (5)       (6)          (7)
    Panel A:                          Sample of Households Giving Transfers to Ultra-Poor at Baseline
                                 ultra      other   middle upper           other     middle       upper
                                 poor       poor      class      class     poor       class        class
    Treatment                  0.107*** 0.097*** 0.079*** 0.114*** 0.121*** 0.089***             0.074***
                                (0.026)    (0.024)   (0.023)   (0.027)    (0.031)    (0.023)      (0.025)
    Mean in control              0.497      0.509     0.501     0.241      0.558      0.384        0.126
    Adj. R-squared               0.232      0.215     0.241     0.198      0.334      0.427        0.223
    Observations                 2370       4386      8828       2872      1296        978          510
    Panel B:                       Sample of Households Not Giving Transfers to Ultra-Poor at Baseline
                                 ultra      other   middle upper           other     middle       upper
6




                                 poor       poor      class      class     poor       class        class
    Treatment                  0.026***     0.007     0.024     -0.001 0.028*** 0.020***           0.004
                                (0.006)    (0.011)   (0.019)   (0.003)    (0.006)    (0.004)      (0.003)
    Mean in control              0.059      0.117     0.164     0.024      0.048      0.026        0.005
    Adj. R-squared               0.102      0.104     0.058     0.019      0.269      0.130        0.080
    Observations                 11094      9078      4636      10592      13384      12506        3920
    Source: Author’s analysis based on original survey data.
    Notes: The table reports ITT estimates based on specification (1) estimated with OLS. In columns 1-4, the sample includes observations
    from ultra-poor households; while in columns 5-7 non-ultra-poor households. In particular, the sample in column 5 includes other
    poor, in column 6 middle class and in column 7 upper class households. All specifications control for the baseline level of the outcome,
    survey wave and subdistrict (strata) fixed effects. Standard errors are clustered at the BRAC branch office level (unit of randomization).
    *** (**) (*) indicates significance at the 1% (5%) (10%) level. The dependent variables in columns 1, 2, 3 and 4 are dummy variables
    equal to 1 if the respondent reported that her household gave any transfers to any ultra-poor, other poor, middle class or upper class
    households (within her community) respectively; in columns 4, 5 and 6 (where the sample is limited to non-ultra-poor households)
    the dependent variables are dummies equal to 1 if the respondent’s household received any transfers from any ultra poor households
    (within her community). In Panel A, the sample is restricted to households who had at baseline reported giving transfers to at least
    one ultra-poor household within their community; in Panel B the sample includes households who did not report any ultra-poor
    household as a recipient of transfers.
       Table S1.7: E FFECTS ON L ABOR S UPPLY IN U LTRA -P OOR H OUSEHOLDS
                      Either respondent                                     Women’s labor                            Men’s labor
                   works for     hours in                              works for   hours in                  works for      hours in
                    a wage      wage-labor                              a wage   wage-labor                   a wage      wage-labor
                      (1)           (2)                                   (3)          (4)                       (5)            (6)
Treatment          -0.082*** -433.971***                               -0.101*** -235.044***                   -0.017       -67.471**
                    (0.015)      (44.818)                               (0.013)     (24.954)                  (0.013)        (30.639)
Mean in control      0.702       1614.405                                0.589      764.414                    0.430         692.373
Adjusted R-squared   0.170         0.195                                 0.262        0.257                    0.140          0.135
Observations         13464        13464                                  13464       13464                      7778           7778
Source: Author’s analysis based on original survey data.
Notes: The table reports ITT estimates based on specification (1) estimated with OLS. *** (**) (*) indicates significance at the 1% (5%) (10%)
level. The dependent variable in column 1 is a dummy variable (DV) equal to 1 if either the main female respondent, or the male head of the
household was engaged in income generating activity where s/he was employed by others during the 12 months preceding the survey. ‘Hours
in wage-labor’ gives the total annual hours the respondent spent in wage employment, where annual hours are computed by multiplying the
number of hours worked in a typical day by the number of days worked in a year for each wage labor activity and then summing across
all wage labor activities. The dependent variable in column 2 is the sum of the hours spent by the main female and the male head of the
household in wage-labor. Outcome variables in columns 3-4 (5-6) refer to the main female (male head of the household).




                                                              7
                      Table S1.8: E FFECTS ON I NTERHOUSEHOLD L OANS


                      Credit received        Credit given     Net credit received
                   (Yes=1)      (BDT)   (Yes=1)       (BDT)  (Yes=1)      (BDT)
                      (1)         (2)      (3)          (4)     (5)         (6)
Treatment           -0.020 -378.141* 0.046*** 553.392*** -0.027 -954.619***
                    (0.019) (207.860) (0.006)       (80.279)  (0.018)   (263.890)
Mean in control      0.224    1130.131    0.028      224.281   0.218     905.850
Adjusted R-squared 0.096         0.034    0.021        0.012   0.088       0.012
Observations         13464      13464     13464       13464    13464      13464
                        Panel B: Loans within the village
Treatment           -0.013     -90.552* 0.017*** 192.035*** -0.014 -281.826***
                    (0.008) (50.925)     (0.003)    (32.899)  (0.008)    (65.304)
Mean in control      0.068     237.230    0.013       74.133   0.068     163.096
Adjusted R-squared 0.026         0.012    0.010        0.008   0.025       0.007
Observations         13464      13464     13464       13464    13464      13464
                       Panel C: Loans outside the village
Treatment           -0.011 -289.012* 0.031*** 362.863*** -0.016 -675.220***
                    (0.015) (163.245) (0.004)       (52.772)  (0.014)   (203.455)
Mean in control      0.179     892.901    0.018      150.147   0.177     742.753
Adjusted R-squared 0.081         0.026    0.013        0.008   0.077       0.009
Observations         13464      13464     13464       13464    13464      13464
Source: Author’s analysis based on original survey data.
Notes: The table reports ITT estimates based on specification (1) estimated with OLS. The sample includes observations from
ultra-poor households surveyed at the midline and endline surveys. All specifications control for the baseline level of the out-
come, survey wave and subdistrict (strata) fixed effects. Standard errors are clustered at the BRAC branch office level (unit of
randomization). All monetary values are measured in Bangladeshi Takas, deflated to 2007 prices using the annual CPI index
published by the Bank of Bangladesh. In 2007, 1USD=69BDT nominal and 1USD=18.5BDT at PPP. *** (**) (*) indicates significance
at the 1% (5%) (10%) level. The dependent variable in Panel A, column 1 is a dummy variable (DV) equal to 1 if the respondent’s
household received any loans from another household during the 12 months preceding the survey; in column 2 it is the monetary
value of loans received during the past 12 months; in column 3 it is a DV equal to 1 if the respondent’s household gave any
loans to another household during the past 12 months; in column 4 it is the monetary value of all loans given by the respondent’s
household during the last 12 months; in column 5 it is a DV equal to 1 if the respondent’s household was a net loan receiver in
the past 12 months; in column 6 it is the monetary value of loans received minus loans given by the respondent’s household in
the past 12 months. All dependent variables in Panel B (C) are identical to those in Panel A, except they refer to within (outside)
village informal loans.




                                                           8
Table S1.9:       E FFECTS      ON    O UTSIDE -V ILLAGE T RANSFERS                    OF   U LTRA -P OOR H OUSE -
HOLDS

                                           Transfers given
                                     Transfers received              Net transfers
                                         in past 12 months
                                     in past 12 months            in past 12 months
                                         (Yes=1)
                                    (Yes=1)          (BDT)
                                                 (BDT)         (Yes=1)        (BDT)
                                       (1)  (3)    (2) (4)         (5)          (6)
 Panel A:                               Within-Family Transfers
 Treatment            -0.010    -63.842  0.013***     8.436      -0.011      -74.894
                     (0.007)   (54.182)   (0.004) (39.976) (0.007)           (49.939)
 Mean in control       0.157   401.939     0.024    106.703       0.154      295.236
 Adjusted R-squared    0.046     0.023     0.012      0.000       0.044        0.014
 Observations         13464      13464     13464     13464       13464        13464
 Panel B:                               Outside-Family Transfers
 Treatment          -0.117*** -86.216*** -0.001 36.538** -0.116*** -123.587***
                     (0.019)   (30.892)   (0.004) (16.085) (0.019)           (33.826)
 Mean in control       0.317   238.713     0.027     25.410       0.313      213.303
 Adjusted R-squared    0.113     0.006     0.021      0.000       0.112        0.005
 Observations         13464      13464     13464     13464       13464        13464
 Source: Author’s analysis based on original survey data.
 Notes: The table reports ITT estimates based on specification (1) estimated with OLS. The sample includes observations from
 ultra-poor households surveyed at the midline and endline surveys. All specifications control for the baseline level of the out-
 come, survey wave and subdistrict (strata) fixed effects. Standard errors are clustered at the BRAC branch office level (unit of
 randomization). All monetary values are measured in Bangladeshi Takas, deflated to 2007 prices using the annual CPI index pub-
 lished by the Bank of Bangladesh. In 2007, 1USD=69BDT nominal and 1USD=18.5BDT at PPP. *** (**) (*) indicates significance at
 the 1% (5%) (10%) level. The dependent variables in Panel A refer to transfers received from or given to households who live
 outside the village but are part of the first-degree family (parents, siblings, children, parents-in-law or siblings-in-law) net-
 work of the respondent. In contrast, the dependent variables in Panel B refer to transfers received from or given to households
 who live outside the village and are not part of the first-degree family network of the respondent’s household. In particular,
 the outcome variable in column 1 is a dummy variable (DV) equal to 1 if the respondent’s household received any transfers (in
 cash or in kind) during the 12 months preceding the survey; in column 2 it is the monetary value of transfers received during the
 past 12 months; in column 3 it is a DV equal to 1 if the respondent’s household gave any transfers to another household during
 the past 12 months; in column 4 it is the monetary value of all transfers given by the respondent’s household during the last 12
 months; in column 5 it is a DV equal to 1 if the respondent’s household was a net transfer receiver in the past 12 months; in column
 6 it is the monetary value of transfers received minus transfers made by the respondent’s household in the past 12 months.




                                                             9
  Table S1.10: E FFECTS ON FAMILY T RANSFERS OF U LTRA -P OOR H OUSEHOLDS

                   Transfers received      Transfers given    Net transfers received
                   in past 12 months in past 12 months           in past 12 months
                   (Yes=1)     (BDT)      (Yes=1) (BDT) (Yes=1)              (BDT)
                      (1)        (2)        (3)        (4)      (5)            (6)
Panel A:                       Transfers From/to Wealthier Family Members
Treatment           -0.007      3.812      0.002    6.269**   -0.007         -2.457
                    (0.006) (18.703) (0.002)        (2.462) (0.006)        (19.293)
Mean in control      0.087    137.473      0.007      2.636    0.087        134.837
Adjusted R-squared 0.066        0.032      0.006     -0.000    0.066          0.027
Observations         13464     13464       13464     13464    13464          13464
Panel B:                     Transfers From/to Non-wealthy Family Members
Treatment            0.001    -65.533 0.012***        2.983    0.000        -69.394
                    (0.005) (56.269) (0.004) (39.763) (0.005)              (49.065)
Mean in control      0.078    264.466      0.017    104.067    0.076        160.399
Adjusted R-squared 0.014        0.015      0.009      0.000    0.013          0.009
Observations         13464     13464       13464     13464    13464          13464
Source: Author’s analysis based on original survey data.
Notes: The table reports ITT estimates based on specification (1) estimated with OLS. The sample includes observations from
ultra-poor households surveyed at the midline and endline surveys. All specifications control for the baseline level of the out-
come, survey wave and subdistrict (strata) fixed effects. Standard errors are clustered at the BRAC branch office level (unit of
randomization). All monetary values are measured in Bangladeshi Takas, deflated to 2007 prices using the annual CPI index
published by the Bank of Bangladesh. In 2007, 1USD=69BDT nominal and 1USD=18.5BDT at PPP. *** (**) (*) indicates signifi-
cance at the 1% (5%) (10%) level. The dependent variables in Panel A refer to transfers received from or given to wealthier
households who live outside the village but are part of the first-degree family (parents, siblings, children, parents-in-law or
siblings-in-law) network of the respondent. The relative wealth of the network members is self-declared by the respondent
at baseline. In contrast, the dependent variables in Panel B refer to transfers received from or given to poorer households who
live outside the village and are not part of the first-degree family network of the respondent’s household. In particular, the
outcome variable in column 1 is a dummy variable (DV) equal to 1 if the respondent’s household received any transfers (in cash
or in kind) during the 12 months preceding the survey; in column 2 it is the monetary value of transfers received during the past
12 months; in column 3 it is a DV equal to 1 if the respondent’s household gave any transfers to another household during the past
12 months; in column 4 it is the monetary value of all transfers given by the respondent’s household during the last 12 months; in
column 5 it is a DV equal to 1 if the respondent’s household was a net transfer receiver in the past 12 months; in column 6 it is the
monetary value of transfers received minus transfers made by the respondent’s household in the past 12 months.




                                                           10
          Table S1.11: M EDIATION A NALYSIS ON N ET T RANSFERS R ECEIVED
                                                                          Extensive                       Intensive
                                                                           margin                           margin
                                                                           (Yes=1)                           (BDT)
                                                                               (1)                              (2)
1. Baseline ITT impact                                                    -0.131***                       -151.721**
                                                                            (0.019)                         (74.478)
2. Unrestricted Estimate (includes mediators as controls)                 -0.104***                         167.599
                                                                            (0.024)                        (184.197)
3. Total Mediated Effect (difference between ITT and unrestricted effect) -0.027**                         -319.319*
                                                                            (0.013)                        (188.647)
4. Mediator: received enterprise support from BRAC                         -0.043**                       -425.316**
                                                                            (0.016)                        (211.314)
5. Mediator: received sanitary latrine or tubewell from BRAC                 0.008                            4.063
                                                                            (0.009)                         (56.664)
6. Mediator: received health support from BRAC                               -0.001                       136.057**
                                                                            (0.006)                         (55.779)
7. Mediator: received education support from BRAC                            -0.001                           4.104
                                                                            (0.001)                          (4.715)
8. Mediator: received other support from BRAC                                0.007                           25.580
                                                                            (0.006)                         (37.423)
9. Mediator: received help from GDBC committee                               0.005                            -3.210
                                                                            (0.003)                         (17.737)
10. Mediator: received microcredit from BRAC                                 -0.002                        -60.597**
                                                                            (0.004)                         (29.875)
Source: Author’s analysis based on original survey data.
Notes: The table shows the results from mediation analysis following Gelbach (2016). In each panel, the first row shows
the ITT impact based on specification (1) estimated with OLS, corresponding to the estimates in Table 3 (in the paper)
columns 5 and 6. The second row shows the ITT impact once the mediators are controlled for?these are having received
enterprise support (asset, training and cash transfer) from BRAC, having received any other type of support from BRAC
(e.g. health or education support, sanitary latrine, tubewell), having received sanitary latrinne or tubewell from BRAC,
having received health support from BRAC (for medical care of any household member), having received any help with
school expenses from BRAC, having received any help with transportation costs from BRAC, having received any support
from the GDBC (Village Poverty Alleviation) committee and having received any microcredit from BRAC. Row 3 shows
the total mediated effect, and rows 4 - 10 show the contribution of each mediator. The sample includes observations from
ultra-poor households surveyed at the midline and endline surveys. All specifications control for the baseline level of the
outcome, survey wave and subdistrict (strata) fixed effects. Standard errors are clustered at the BRAC branch office level
(unit of randomization). All monetary values are measured in Bangladeshi Takas, deflated to 2007 prices using the annual
CPI index published by the Bank of Bangladesh. In 2007, 1USD=69BDT nominal and 1USD=18.5BDT at PPP. *** (**) (*)
indicates significance at the 1% (5%) (10%) level. The dependent variable in column 1 is a dummy variable (DV) equal to
1 if the respondent’s household was a net transfer receiver in the past 12 months; in column 2 it is the monetary value of
transfers received minus transfers made by the respondent’s household in the past 12 months.




                                                       11
S2: Comparison of Midline vs. Endline Impacts

  Table S2.1: E FFECTS ON I NFORMAL T RANSFERS OF U LTRA -P OOR H OUSEHOLDS

                                      Transfers received       Transfers given                         Net transfers received
                                       in past 12 months     in past 12 months                             in past 12 months
                                      (Yes=1)       (BDT)   (Yes=1)      (BDT)                          (Yes=1)         (BDT)
                                          (1)         (2)       (3)        (4)                              (5)           (6)
 Treatment                           -0.127*** -39.352         0.003    20.821                         -0.129***       -66.353
                                       (0.037) (106.079) (0.016) (78.595)                                (0.035)      (95.392)
 Treatment × Endline                    -0.004    -130.624     0.017    40.111                           -0.004       -170.735
                                       (0.063) (245.323) (0.028) (116.845)                               (0.057)     (192.236)
 Mean in control                         0.497     783.055     0.059    143.978                           0.484       639.076
 Adjusted R-squared                      0.126       0.027     0.030      0.001                           0.120          0.019
 Observations                           13464       13464     13464      13464                            13464         13464
                                         Panel B: Transfers within the village
 Treatment                            -0.049** 55.780**       -0.004     -7.351                         -0.048**           63.323**
                                       (0.019)    (27.359) (0.006)      (7.839)                          (0.019)           (29.878)
 Treatment × Endline                     0.025     -29.068     0.007      6.462                           0.022             -35.529
                                       (0.026)    (49.401) (0.009)      (9.203)                          (0.025)           (51.095)
 Mean in control                         0.148     142.403     0.015    11.865                            0.147            130.538
 Adjusted R-squared                      0.041       0.004     0.014      0.001                           0.041              0.003
 Observations                           13464       13464     13464      13464                            13464              13464
                                        Panel C: Transfers outside the village
 Treatment                           -0.119*** -94.086         0.004    28.196                         -0.119***           -130.179
                                       (0.034)    (92.347) (0.013) (77.689)                             (0.032)            (84.062)
 Treatment × Endline                    -0.008    -101.556     0.014    33.650                          -0.007             -135.206
                                       (0.057) (216.646) (0.022) (118.138)                              (0.052)           (164.083)
 Mean in control                         0.430     640.652     0.048    132.113                          0.420             508.539
 Adjusted R-squared                      0.117       0.027     0.023      0.000                          0.112               0.018
 Observations                           13464       13464     13464      13464                           13464              13464
 Source: Author’s analysis based on original survey data.
 Notes: The table reports ITT estimates based on specification (2) estimated with OLS. The sample includes observations from
 ultra-poor households surveyed at the midline and endline surveys. All specifications control for the baseline level of the out-
 come, survey wave and subdistrict (strata) fixed effects. Standard errors are clustered at the BRAC branch office level (unit of
 randomization). All monetary values are measured in Bangladeshi Takas, deflated to 2007 prices using the annual CPI index
 published by the Bank of Bangladesh. In 2007, 1USD=69BDT. *** (**) (*) indicates significance at the 1% (5%) (10%) level. The de-
 pendent variable in Panel A, column 1 is a dummy variable (DV) equal to 1 if the respondent’s household received any transfers
 (in cash or in kind) from another household during the 12 months preceding the survey; in column 2 it is the monetary value of
 transfers received during the past 12 months; in column 3 it is a DV equal to 1 if the respondent’s household gave any transfers
 to another household during the past 12 months; in column 4 it is the monetary value of all transfers given by the respondent’s
 household during the last 12 months; in column 6 it is a DV equal to 1 if the respondent’s household was a net transfer receiver in
 the past 12 months; in column 6 it is the monetary value of transfers received minus transfers made by the respondent’s household
 in the past 12 months. All dependent variables in Panel B (C) are identical to those in Panel A, except they refer to within (outside)
 village transfers.




                                                              12
Table S2.2: E FFECTS ON F OOD T RANSFERS OF U LTRA -P OOR H OUSEHOLDS

                        Transfers received       Transfers given
                        (Yes=1)     (No.)    (Yes=1)        (No.)
                           (1)       (2)        (3)          (4)
    Treatment             0.013     0.037     0.048*       0.152*
                         (0.010)   (0.065)    (0.027)      (0.078)
    Treatment × Endline   0.000     0.120      0.041        0.153
                         (0.017)   (0.098)    (0.040)      (0.126)
    Mean in control       0.936     2.340      0.560        1.284
    Adjusted R-squared    0.069     0.170      0.260        0.315
    Observations          13464     13464      13453        13464
    Panel B:                     Within-village Transfers
    Treatment             0.015     0.059      0.045       0.147*
                         (0.011)   (0.063)    (0.027)      (0.075)
    Treatment × Endline   0.010     0.155      0.046        0.160
                         (0.021)   (0.099)    (0.040)      (0.123)
    Mean in control       0.907     2.199      0.547        1.236
    Adjusted R-squared    0.078     0.194      0.259        0.318
    Observations          13464     13464      13464        13464
    Panel C:                     Outside-village Transfers
    Treatment            -0.012    -0.017     -0.001        0.003
                         (0.014)   (0.019)    (0.006)      (0.008)
    Treatment × Endline -0.031     -0.034     -0.004       -0.007
                         (0.022)   (0.031)    (0.010)      (0.014)
    Mean in control       0.107     0.141      0.037        0.047
    Adjusted R-squared    0.066     0.074      0.037        0.036
    Observations          13464     13464      13464        13464
    Source: Author’s analysis based on original survey data.
    Notes: The table reports ITT estimates based on specification (2) estimated with OLS. The sam-
    ple includes observations from ultra-poor households surveyed at the midline and endline sur-
    veys. All specifications control for the baseline level of the outcome, survey wave and subdistrict
    (strata) fixed effects. Standard errors are clustered at the BRAC branch office level (unit of ran-
    domization). *** (**) (*) indicates significance at the 1% (5%) (10%) level. The dependent variable
    in column 1 is a dummy variable (DV) equal to 1 if the respondent’s household ever receives any
    food transfers from other households The dependent variable in Panel A, column 1 is a dummy
    variable (DV) equal to 1 if the respondent’s household ever receives any food transfers; in col-
    umn 2 it is the number of households (capped at 3) that the respondent’s household receives
    food transfers from; in column 3 it is a DV equal to 1 if the respondent’s household ever gives
    food transfers to other households; in column 4 it is the number of households (capped at 3) that
    the respondent’s households gives food transfers to. All dependent variables in Panel B (C) are
    identical to those in Panel A, except they refer to within (outside) village food transfers.




                                                  13
             Table S2.3: E FFECTS ON M ATCHING IN T RANSFER N ETWORKS

                        (1)           (2)           (3)             (4)
Panel A:              Transfers received by Ultra-Poor Households from ...
                    ultra poor other poor middle class upper class
Treatment            0.048***        0.022         0.008          -0.009
                      (0.008)       (0.014)       (0.015)        (0.010)
Treatment × Endline    0.018         0.027         0.019           0.002
                      (0.013)       (0.018)       (0.023)        (0.013)
Mean in control        0.128         0.345         0.701           0.260
Adj. R-squared         0.439         0.415         0.327           0.407
Observations           13464         13464         13464          13464
Panel B:                 Transfers given by Ultra-Poor Households to ...
                    ultra poor other poor middle class upper class
Treatment            0.047***        0.021         0.030           0.005
                      (0.011)       (0.014)       (0.020)        (0.007)
Treatment × Endline    0.010         0.022         0.048         0.023*
                      (0.015)       (0.026)       (0.031)        (0.012)
Mean in control        0.108         0.252         0.394           0.078
Adj. R-squared         0.368         0.306         0.255           0.145
Observations           13464         13464         13464          13464
Source: Author’s analysis based on original survey data.
Notes: The table reports ITT estimates based on specification (1) estimated with OLS. The sample includes ob-
servations from ultra-poor households. All specifications control for the baseline level of the outcome, survey
wave and subdistrict (strata) fixed effects. Standard errors are clustered at the BRAC branch office level (unit of
randomization). *** (**) (*) indicates significance at the 1% (5%) (10%) level. In Panel A, the dependent variables
in columns 1, 2, 3 and 4 are dummy variables equal to 1 if the respondent reported that her household received
any transfers from any ultra-poor, other poor, middle class or upper class households (within her community)
respectively. In Panel B, the dependent variables in columns 1, 2, 3 and 4 are dummy variables equal to 1 if the
respondent reported that her household gave any transfers to any ultra-poor, other poor, middle class or upper
class households (within her community) respectively.




                                                      14
Table S2.4: E FFECTS ON M ATCHING , R EPORTED BY N ON -U LTRA -P OOR H OUSEHOLDS
                                                                                       (1)            (2)          (3)            (4)
           Transfers received by Non-Ultra-Poor from ...                              ultra          other       middle          upper
                                                                                      poor           poor         class          class
           Panel A1: Pooled
           Treatment                                                                0.018*** 0.009                 -0.010         -0.010
                                                                                     (0.004) (0.009)              (0.012)        (0.008)
           Treatment × Endline                                                       0.009*   0.005                 0.020          0.005
                                                                                     (0.005) (0.011)              (0.018)        (0.012)
           Panel A2: By Wealth Class
           Treatment                                       0.038***                                   0.004         0.003        0.010
                                                            (0.007)                                 (0.014)       (0.013)       (0.011)
           Treatment × Endline                               0.001                                    0.006         0.000        -0.005
                                                            (0.008)                                 (0.018)       (0.016)       (0.013)
           Differential effect by endline for middle class   0.012                                    0.004         0.020         0.006
                                                            (0.008)                                 (0.016)       (0.019)       (0.016)
           Differential effect by endline for uppper class   0.016                                    0.002         0.027       0.047**
                                                            (0.010)                                 (0.019)       (0.030)       (0.022)
           Treatment effect for middle class                 0.019                                    0.008        -0.005        -0.001
                                                            (0.006)                                 (0.009)       (0.014)       (0.009)
           Treatment effect for upper class                  0.004                                   -0.001        -0.039        -0.062
                                                            (0.005)                                 (0.011)       (0.029)       (0.024)
           Control mean                                      0.062                                    0.319         0.691        0.279
           Mean for other poor                               0.085                                    0.448         0.711        0.272
           Mean for middle class                             0.048                                    0.224         0.770        0.261
           Mean for upper class                              0.017                                    0.128         0.385        0.360
           Observations                                      32594                                   32594         32594         32594
           Transfers given by Non-Ultra-Poor to ...                                   ultra          other       middle          upper
                                                                                      poor           poor         class          class
           Panel B1: Pooled
           Treatment                                                                0.025*** -0.014                -0.019         -0.016
                                                                                     (0.007) (0.013)              (0.020)        (0.009)
           Treatment × Endline                                                        -0.001  0.006                 0.016        0.028*
                                                                                     (0.008) (0.016)              (0.025)        (0.014)
           Panel B2: By Wealth Class
           Treatment                                       0.030***                                  -0.008        -0.013         -0.013
                                                            (0.009)                                 (0.017)       (0.022)        (0.010)
           Treatment × Endline                               0.002                                    0.012         0.003          0.015
                                                            (0.009)                                 (0.024)       (0.026)        (0.011)
           Differential effect by endline for middle class   -0.001                                  -0.009         0.012          0.011
                                                            (0.011)                                 (0.021)       (0.026)        (0.015)
           Differential effect by endline for uppper class   0.001                                   -0.006         0.017          0.028
                                                            (0.015)                                 (0.026)       (0.031)        (0.024)
           Treatment effect for middle class                 0.026                                   -0.014        -0.026         -0.014
                                                            (0.008)                                 (0.013)       (0.024)        (0.010)
           Treatment effect for upper class                  0.015                                   -0.030        -0.010         -0.016
                                                            (0.015)                                 (0.021)       (0.026)        (0.018)
           Control mean                                      0.078                                    0.338         0.600         0.177
           Mean for other poor                               0.084                                    0.397         0.507         0.119
           Mean for middle class                             0.065                                    0.268         0.721         0.182
           Mean for upper class                              0.091                                    0.329         0.584         0.373
           Observations                                      32594                                   32594         32594          32594
           Source: Author’s analysis based on original survey data.
           Notes: The table reports ITT estimates based on specification (2), with the addition of indicators for wealth classes and their
           interactions with treatment and endline indicators, estimated with OLS. The sample includes observations from non-ultra-
           poor households. All specifications control for the baseline level of the outcome, survey wave and subdistrict (strata) fixed
           effects. Standard errors are clustered at the BRAC branch office level (unit of randomization). *** (**) (*) indicates significance
           at the 1% (5%) (10%) level. In Panel A, the dependent variables in columns 1, 2, 3 and 4 are dummy variables equal to 1 if the
           respondent reported that her household received any transfers from any ultra-poor, other poor, middle class or upper class
           households (within her community) respectively. In Panel B, the dependent variables in columns 1, 2, 3 and 4 are dummy
           variables equal to 1 if the respondent reported that her household gave any transfers to any ultra-poor, other poor, middle
           class or upper class households (within her community) respectively. In the pooled regressions (Panels A1 and B1) each
           observation is weighted using sampling weights calculated as the fraction of households surveyed from each wealth class
           (lower, middle and upper) relative to the number of households from the relevant wealth class in the community census.


                                                                        15
       Table S2.5: E FFECTS ON R ECIPROCITY OF T RANSFERS , BY W EALTH C LASS

                    Reciprocity       Reciprocity of transfers with
                    of transfer  ultra     other    middle      upper
                       links     poor      poor       class      class
                         (1)      (2)       (3)        (4)        (5)
Treatment              0.032     0.005 0.069*** 0.040            0.049
                      (0.021)   (0.033) (0.024) (0.025)         (0.031)
Treatment × Endline   0.089**   0.079*     0.051    0.094**    0.142***
                      (0.035)   (0.041) (0.044) (0.039)         (0.048)
Mean in control        0.510     0.627     0.560      0.449      0.267
Adjusted R-squared     0.219     0.190     0.208      0.237      0.198
Observations           12774     2792      4823       9474       3135
Source: Author’s analysis based on original survey data.
Notes: The table reports ITT estimates based on specification (2) estimated with OLS. The sample includes obser-
vations from ultra-poor households surveyed at the midline and endline surveys. All specifications control for the
baseline level of the outcome, survey wave and subdistrict (strata) fixed effects. Standard errors are clustered at
the BRAC branch office level (unit of randomization). *** (**) (*) indicates significance at the 1% (5%) (10%) level.
The dependent variable in column 1 is the fraction of transfer sources within the village who are also reported as
recipients of transfers (of last 12 months’ or food transfers ever) from the respondent’s household. The dependent
variables in columns 2 - 5 are the fraction of ultra-poor, other poor, middle class or upper class households (re-
spectivey) who are reported as sources of transfers (food, cash or other) and also reported as recipients of transfers
given by the respondent’s household.

                                   Table S2.6: T RANSFERS FROM E MPLOYERS
                                Household received Main female respondent                                                     Male head of HH
                                    transfers from         received transfers                                                 received transfers
                                    any employer           from an employer                                                   from an employer
                                           (1)                       (2)                                                              (3)
Treatment                              -0.044***                -0.040***                                                          -0.015**
                                        (0.011)                   (0.010)                                                           (0.007)
Treatment effect after 4 years           0.011                     0.010                                                             0.004
                                        (0.017)                   (0.015)                                                           (0.012)
Mean in control                          0.091                     0.081                                                             0.032
Adjusted R-squared                       0.068                     0.059                                                             0.039
Observations                             13464                     13464                                                             7778
                              Panel B: Sample of wage-workers in all 3 surveys
Treatment                              -0.052***                 -0.044**                                                              -0.010
                                        (0.018)                   (0.021)                                                              (0.020)
Treatment effect after 4 years           0.023                     0.036                                                               -0.019
                                        (0.026)                   (0.025)                                                              (0.038)
Mean in control                          0.133                     0.138                                                                0.073
Adjusted R-squared                       0.079                     0.071                                                                0.077
Observations                              6336                      4488                                                                1332
Source: Author’s analysis based on original survey data.
Notes: The table reports ITT estimates based on specification (2) estimated with OLS. The sample includes observations from ultra-poor households
surveyed at the midline and endline surveys. All specifications control for the baseline level of the outcome, survey wave and subdistrict (strata) fixed
effects. Standard errors are clustered at the BRAC branch office level (unit of randomization). *** (**) (*) indicates significance at the 1% (5%) (10%) level.
The dependent variable in column 1 is a dummy variable (DV) equal to 1 if either the main female or the male head of the household had worked (in the
last 12 months) for an employer who lives in the same village and is reported as a source of transfers for the household. The dependent variable in column
2 (3) is a DV equal to 1 if the main female respondent (male head of the household) had worked (in the last 12 months) for an employer who lives in the
same village and is reported as a source of transfers for the household. In Panel B, the sample is restricted to ultra-poor households where either the main
female respondent or the male head of the household had worked for an employer in the 12 months preceding the survey.



                                                                      16