NO LONGER TRAPPED? PROMOTING
ENTREPRENEURSHIP THROUGH CASH
TRANSFERS TO ULTRA-POOR WOMEN IN
NORTHERN KENYA
VILAS J. GOBIN, PAULO SANTOS, AND RUSSELL TOTH
We examine the short-to-medium-run impacts of the Rural Entrepreneur Access Program, a poverty
graduation program that promotes entrepreneurship among ultra-poor women in arid and semi-arid
northern Kenya, a context prone to poverty traps. The program relies on cash transfers (rather than asset
transfers) in addition to business skills training, business mentoring, and savings. Participation in each of
the program’s three rounds was randomly determined through a public lottery. In the short-to-medium-
run, we find that the program has a positive and significant impact on income, savings, and asset accumu-
lation, similar to more traditional poverty graduation programs that rely on asset transfers.
Key words: Cash transfers, entrepreneurship, field experiment, microenterprise, poverty graduation,
ultra-poor.
JEL codes: C93, D13, J24, O12, O13, Q12.
Microenterprises are the source of employ-ment for more than half of the labor force indeveloping countries (de Mel, McKenzie, andWoodruff 2008; Gindling and Newhouse 2014)and can be potential engines of economic de-velopment by raising the income of owners,creating a demand for labor and raising wages,and increasing market competition to generatelower prices for consumers (Bruhn 2011; WorldBank 2012). Despite these potential benefits,some policymakers are concerned that some ofthe poorest people, sometimes referred to asthe ultra-poor, are constrained from establish-ing such businesses or from participating in
many popular approaches aimed at stimulatingmicroenterprise formation. Yet results of re-cent evaluations of interventions meant to tar-get individual constraints, particularly access tofinance through microcredit, and access tohuman capital through microenterprise trainingprograms, have been mixed (Banerjee, Karlan,and Zinman 2015; Karlan and Zinman 2011).
The apparent lack of success of these “one-constraint-at-a-time” approaches suggests thatthe ultra-poor might need a multifaceted “bigpush” that simultaneously addresses multipleconstraints. One influential approach, pio-neered by the non-governmental organization(NGO) Building Resources AcrossCommunities (BRAC), is Challenging theFrontiers of Poverty Reduction – Targeting theUltra-Poor (CFPR/TUP). During a limitedperiod (two years), its participants benefit froma set of interventions, including initial con-sumption support and an asset transfer (such aslivestock), together with savings services, skillstraining, and regular follow-up visits (Matin,Sulaiman, and Rabbani 2008; Goldberg andSalomon 2011). Several recent impact evalua-tions of such “graduation” programs providesupport for this approach across a diverse setof developing countries. In Bangladesh the
Vilas J. Gobin is a Ph.D. candidate in Economics and PauloSantos is a senior lecturer, both in Economics at MonashUniversity, Melbourne, Australia. Russell Toth is a lecturer inEconomics at the University of Sydney, Sydney, Australia. Theauthors thank Gaurav Datt, Hee-Seung Yang, Asadul Islam,Andreas Leibbrandt, and Dean Karlan for comments on earlierdrafts of this paper. The data used in this study were supplied bythe BOMA Project. The authors thank staff of the BOMAProject, especially Kathleen Colson, Ahmed Omar, MeshackOmarre, Fredrick Learapo, Sabdio Doti, Bernadette Njoroge,Nate Barker, and Alex Villec for their assistance in data collec-tion, as well as in facilitating the first author’s stay in Kenya. Allanalyses, interpretations, or conclusions based on these data aresolely that of the authors. The BOMA Project disclaims responsi-bility for any such analyses, interpretations, or conclusions.Correspondence to be sent to: [email protected].
Amer. J. Agr. Econ. 99(5): 1362–1383; doi: 10.1093/ajae/aax037Published online July 25, 2017
VC The Authors 2017. Published by Oxford University Press on behalf of the Agricultural and Applied EconomicsAssociation. All rights reserved. For Permissions, please email: [email protected]
Downloaded from https://academic.oup.com/ajae/article-abstract/99/5/1362/4036204by George Mason University useron 11 July 2018
program enabled ultra-poor women to engagein microentrepreneurial activities resulting in a38% increase in earnings that persisted twoyears after the program (Bandiera et al. 2016).Likewise, Banerjee et al.’s (2015) study acrosssites in six countries documents similar impactson consumption, productive assets, income andrevenue, and stability in impacts up to at leastone year after the program.
While these results are promising, more re-search is necessary. Banerjee et al. (2015)find the weakest impacts in Peru, where theintervention sites are particularly remote,and Honduras, where negative impacts onassets are attributable to a disease outbreakamong the in-kind asset (chickens). On onehand, this raises questions about the impactsof such programs in particularly remote set-tings, where program delivery is more chal-lenging and market access might be morelimited. On the other hand, it raises the ques-tion of whether cash transfers, which mightreduce program costs and reduce the risks ofworking with live assets, could replicate theseimpacts. Concerns about external validity arealso present in Bauchet, Morduch, and Ravi’s(2015) study, which evaluates a similar inter-vention in Andhra Pradesh, India, and findsno net impact on consumption, income, orasset accumulation. The authors argue thatthese results reflect mistargeting of individu-als with strong labor market opportunitieswho quickly opted out of the program.Furthermore, evidence suggests that recipi-ents liquidated transferred assets to pay downdebt, another source of targeting risk.
This paper presents a randomized evaluationof the Rural Entrepreneur Access Project(REAP), a variation of the CFPR/TUP gradu-ation approach, implemented in arid and semi-arid northern Kenya, a region where more than80% of the population is estimated to be livingbelow the national poverty line (KenyaNational Bureau of Statistics and Society forInternational Development 2013). The REAPcomprises an initial package of interventions,including a U.S. $100 cash transfer to set up amicroenterprise, business skills training, andbusiness mentoring, which are followed, sixmonths later, by an additional $50 cash transfer(conditional on having an active enterprise),and training on the importance of savings aswell as an introduction to savings groups.1 Thissequence of interventions targets ultra-poorwomen and intends to enable them to gain theassets and skills necessary to graduate frompoverty, a motivation similar to the one behind
the CFPR/TUP (BRAC 2013). While we areable to evaluate the impact of this bundle ofinterventions, the design of the program pre-vents us from disentangling the impacts of itsindividual components, a shortcoming sharedwith much of the prior literature.
This program, while similar in spirit to otherultra-poor programs, also has a number of not-able differences. First, contrary to most suchprograms, REAP relies entirely on the transferof cash rather than of a physical asset.Although cash transfers provide flexibility thatmay provide beneficiaries with greater remu-nerative options, these transfers have played aminor role in these programs given concernsabout possible misuse (e.g., for consumption orpayment of existing debt). The purpose of cashtransfers in ultra-poor programs is typically forconsumption support, intended to prevent ben-eficiaries from “eating” their assets (sometimesliterally, in the case of livestock transfers). Thisconcern is potentially more important in thecase of REAP, given that there was no provi-sion of initial consumption support.
Second, the program focuses explicitly onenterprises, with the requirement that womenform three-person groups to run the enterprisejointly. This may provide additional social sup-port and accountability around the use of grantfunds, but may also introduce additional chal-lenges in jointly running a business.
Finally, and not necessarily less important,REAP targets women in the Arid and Semi-Arid Lands (ASALs) of east Africa, a regionaleconomy based on one main activity (livestock)and with very limited market access. Studies ofpoverty dynamics in this area offer the mostpersuasive empirical evidence in support of thethreshold-based poverty traps hypothesis(Kraay and McKenzie 2014). Hence, it pro-vides a setting that is arguably more extremethan those that have been the subject of previ-ous studies. In the context of this literature,REAP could be interpreted as a “cargo net”policy intervention (Barrett, Carter, andIkegami 2008) designed to enable ultra-poorwomen to escape persistent poverty.
While the program differs from otherultra-poor programs in these importantaspects, the findings are qualitatively similar.After one year, this program has a positiveand statistically significant impact on income
1 The program is implemented through an NGO, the BOMAProject. See http://bomaproject.org/the-rural-entrepreneur-access-project/ for a complete description of REAP.
Gobin, Santos, and Toth Promoting Entrepreneurship Through Cash Transfers 1363
Downloaded from https://academic.oup.com/ajae/article-abstract/99/5/1362/4036204by George Mason University useron 11 July 2018
(31%), savings (131%), asset accumulation(35%), and, less clearly, on livestock (12%).The primary channel for impact is throughthe setup of new petty trade enterprises,which deal primarily in food items, with timeuse data showing a corresponding reductionin leisure time and household activity. Asseveral other studies document, thereappears to be a weak impact on consumptionand expenditure, which is particularly evidentfollowing the introduction and promotion ofnew savings mechanisms. This suggests thatin the medium-run, asset accumulation andsavings are absorbing the increase in income.In line with the relatively low cost of distrib-uting cash, the program is highly cost-effective: in just over one year, the averageincrease in household income covers the costof delivering the program. Hence, the evi-dence demonstrates that substituting signifi-cant cash transfers for in-kind asset transfersdoes not mute program impact, and that thegraduation approach can be effective in rela-tively remote and challenging settings.
The program we evaluate is perhaps mostsimilar to the Women’s Income GeneratingSupport (WINGS) program, studied byBlattman et al. (2016), which also focuses on en-terprise development through cash transfers.2
The authors concluded that the program led toan important increase in microenterprise own-ership and income. This, in turn, suggests that,even with no consumption support and in a con-text of little accountability around the use ofgrant funds, recipients were remarkably compli-ant in directing the funds to enterprise forma-tion (rather than immediate consumption, asfeared). The authors also determined that thepromotion of self-help groups led to a doublingof the reported earnings, attributing this toincreased informal finance and economic co-operation. This suggests the need for deeper fi-nancial services (insurance, in the case ofWINGS; savings, in the case of REAP) to re-inforce such interventions. Taken together,these results suggest that cash transfers can bean effective way of promoting business develop-ment among ultra-poor women regardless ofthe program’s set-up.3
Overview of the Intervention
The REAP implementation occurred in four-teen locations in the southern and centralparts of Marsabit County, in the ASALs ofnorthern Kenya.4 In this region, more than80% of the population live below the nationalpoverty line (Kenya National Bureau ofStatistics and Society for InternationalDevelopment 2013).5 The main livelihoodoption in these locations is pastoralism, withlivestock serving both as a source of incomeand food for herders and their families.
Pastoralism is highly susceptible to weatherand other shocks, and repeated droughts fre-quently have devastating impacts on house-holds’ livelihoods (Silvestri et al. 2012). Thishas resulted in many households losing theirability to meet their basic needs due to the lossof herds, which is hard to recover from. Aprominent body of literature provides signifi-cant evidence that a non-linear threshold gov-erns asset (livestock) dynamics in this region.Below the threshold, livestock holding sharplyconverges toward a low-level steady state ofalmost no livestock holding (Lybbertet al. 2004; Barrett et al. 2006; Santosand Barrett 2011; Barrett andSantos 2014; Toth 2015). Households experi-encing such a steady state must resort to beg-ging, unskilled wage labor, various forms ofpetty trade, and become reliant on food aid tomeet dietary needs.6 This raises importantquestions about what type of “cargo net” poli-cies (if any) can help such destitute householdsescape persistent poverty.7
2 However, there were significant differences between the twoprograms: WINGS distributed $150 per individual, instead of$150 per group of three beneficiaries, and provided more busi-ness training but less mentoring than REAP. Further, WINGSalso operated in a substantially different context: post-warUganda, with substantially lower levels of baseline income andlower levels of business activity.
3 Finally, the recently evaluated GiveDirectly program(Haushofer and Shapiro 2016), an unconditional cash transfertargeted at poor households in Kenya that differs substantiallyfrom the poverty graduation approach, reaches similar conclu-sions: in the short-run, these transfers lead to increased invest-ment in, and revenue from, livestock and small businesses, evenin the absence of additional interventions such as training.
4 See figure A.1 in appendix A of the supplementary materialonline for a map displaying these locations.
5 In 2005/06, the poverty line was estimated at Kenya Shillings(Ksh) 1,562 Purchase Power Parity (PPP) $77.07 at 2014 pricesper adult equivalent per month for rural households (KenyaNational Bureau of Statistics 2007). In 2009, it was estimated thatnationally, 45.2% of the population lived below the poverty line(Kenya National Bureau of Statistics and Society forInternational Development 2013).
6 Little et al. (2008) examine different proxies for poverty andwelfare in northern Kenya. These authors identify poverty asbeing most prevalent among sedentary households that are nolonger directly involved in pastoral production or are in the pro-cess of exiting pastoralism.
7 Barrett, Carter, and Ikegami (2008) distinguish between pol-icies that increase the assets of the poor and allow them to escapepoverty (“cargo nets”) from more traditional “safety net”
1364 October 2017 Amer. J. Agr. Econ.
Downloaded from https://academic.oup.com/ajae/article-abstract/99/5/1362/4036204by George Mason University useron 11 July 2018
Opportunities to engage in non-pastoralactivities are further constrained by thesecommunities’ lack of attention by nationaldevelopment processes. Additionally, theseareas have low population densities and lim-ited access to markets or other infrastructure(Elliot and Fowler 2012). By targeting thepoorest women in these communities, REAPaims to provide households with a pathwayout of poverty through the alleviation of fi-nancial and human capital constraints,thereby enabling these women to obtain asustainable livelihood without directly engag-ing in the livestock economy.
Structure and Timing of the Intervention
The REAP’s main aim is to graduate ultra-poor women from poverty through a set ofinterventions that include the development ofbusiness plans and mentoring, grants, and ac-cess to savings. The sequence of these interven-tions appears in figure 1, and an elaboration ofeach intervention follows below.
Participant selection. The REAP estab-lished guidelines for the formation of localcommittees that determine eligibility for pro-spective participants.8 These committeesidentified women who were among the poor-est in the community. The committees priori-tized women with no other sources ofincome, and who demonstrated a likelihoodof responsible entrepreneurship and willing-ness to run a business with two other wom-en.9 Trained business mentors ensured thatthe local committees followed these criteriawhen selecting participants.10 After the par-ticipant selection and acceptance process, thebusiness mentor proceeded to form businessgroups of three women.
Business planning and business skills train-ing. In the month leading up to program
enrollment, business mentors met with bene-ficiaries to assist with the development of abusiness proposal. The meetings’ purposewas to allow mentors to get a better under-standing of the group members’ abilities andprevious business experience before goingthrough the basics of setting up a business.On the day of program enrollment, all partic-ipants had to attend a short business skillstraining session that mentors delivered underthe supervision of REAP field officers.11
Over the course of the program, participantsbenefited from approximately seventeenhours of training.12
First grant and business mentoring. At theend of the business skills training session,REAP provided each group with a cash grantof $100 (Purchase Power Parity [PPP]$237.97 at 2014 prices) for the purpose ofestablishing their business, an amountequivalent to approximately 7.5 months of ex-penditure per capita.13 The REAP requiredeach group to invest the entire grant in thebusiness, but members had the flexibility touse the cash as they saw fit, including by mod-ifying their initial business proposal.14
After the distribution of the initial grant,mentors regularly met with the businessgroup (at least once a month) to monitortheir progress and offer advice and training.The role of the mentor was to help start upthe business (e.g., by providing informationregarding where to source goods or marketconditions). Additionally, the mentor wasalso responsible for assisting the group withrecord-keeping, and if necessary, in managingconflicts within the group. Over the course ofthe program, each business was expected tobenefit from approximately thirty hours ofmentoring.
policies that seek to prevent these families from falling into thepoverty trap in the first place.
8 The committees generally comprise ten persons, with equalrepresentation of clans and ethnic groups in the community, andwith at least half of them being women.
9 In addition, and recognizing the importance of inter-ethnicrivalries in northern Kenya, selection committees were asked toselect participants to ensure equal representation from variousclans and ethnic groups and appropriate representation of per-sons from the town center and more distant villages. Finally, im-mediate relatives of any BOMA Project staff were consideredineligible.
10 Mentors are employed at the location level. Mentors par-ticipated in a training of trainers program, which lasted for fivedays and occurred prior to the recruitment of participants. Eachlocation comprises many sub-locations that are formed bysmaller villages, known as manyattas.
11 These two sessions took approximately four hours to com-plete and covered the following content: accounting, financialplanning, product ideas, marketing, pricing and costing, inven-tory management, customer service, business investment andgrowth strategies, employee management, savings, and debt.
12 This included a half day training on savings that took placesix months after the business training, and nine one-hour trainingsessions that took place during savings group meetings.
13 From here on, all monetary values reported in the articleare in PPP terms at 2014 prices unless otherwise stated. We usethe following PPP exchange rates to convert Kenya Shillings toUSD PPP: $36.83 (2012), $38.38 (2013), and $40.35 (2014). Thesevalues are then converted to 2014 prices by multiplying the ratioof the 2014 U.S. Consumer Price Index (CPI) to the U.S. CPI forthe relevant year.
14 The investment of the grant was ensured by the mentor,who met with the businesses soon after disbursement.Additionally, the conditionality of the second grant, as describedbelow, likely created incentives for groups to invest the first grantin a business.
Gobin, Santos, and Toth Promoting Entrepreneurship Through Cash Transfers 1365
Downloaded from https://academic.oup.com/ajae/article-abstract/99/5/1362/4036204by George Mason University useron 11 July 2018
Second grant, savings training, and savingsgroup formation. Six months after the start ofthe business, groups were eligible for an add-itional grant of $50 (PPP $118.98) conditionalon meeting the following criteria: two ormore original members remained involved inthe business, the group held assets collect-ively, and the business value (defined as thesum of cash on hand, business savings andcredit outstanding, and business stock andassets) was equal to or greater than the valueof the initial grant. Participants were awareof these conditions from the start of the pro-gram. Another requirement mandated that
participants attend a short training session onsavings to develop a basic understanding ofthe formation and operation of savingsgroups, including their rules, record-keeping,and issuance of loans.
After the savings training and the distribu-tion of the second grant, mentors encouragedparticipants to form a savings group (SG) orjoin existing ones. The decision to join agroup was both voluntary and individual (i.e.,it was not a business group decision). TheSGs most closely resembled Village Savingsand Loans Associations (VSLAs), alsoknown as Accumulating Savings and Credit
Figure 1. Description of the intervention, study sample, and experimental design
1366 October 2017 Amer. J. Agr. Econ.
Downloaded from https://academic.oup.com/ajae/article-abstract/99/5/1362/4036204by George Mason University useron 11 July 2018
Associations (ASCAs), described inAllen (2006). The groups are self-managedand allow members to save money and accessloans that are paid back with interest.
Research Design
Several factors allow us to identify the pro-gram impacts in a relatively straightforwardway. These include the sequential roll-out ofthe program, the randomized allocation toeach cycle, the perfect compliance of obser-vations to treatment and control groups, andthe low attrition rate. In this section, we pro-vide details of the random allocation of par-ticipants to treatment and control groups. Wealso report on tests of the assumptions under-lying the identification strategy and discussspillover and anticipation effects.
Randomization of Program Assignment
In November 2012, the local selection com-mittees across fourteen locations in northernKenya identified 1,755 women eligible forREAP. Capacity constraints resulted in thedivision of eligible women into three equal-sized groups that successively enrolled inREAP over the next three funding cycles(March/April 2013, September/October 2013,or March/April 2014, hereafter referred to asgroups A, B, and C, respectively).15 In orderto ensure actual and apparent transparencyand fairness, a random public lottery in eachlocation determined assignment to each ofthe three funding cycles.16 None of the eli-gible participants declined to take part in theprogram, which did not permit them to par-ticipate outside of their randomly allocatedgroup.
Eligible women had interviews at the base-line (November 2012) with an additional twofollow-ups at six-month intervals, timed tooccur before each new funding cycle (see fig-ure 1). Survey attrition was very low in bothfollow-up survey rounds, with less than 4% ofwomen not re-interviewing at the first follow-
up (midline) and less than 6% of womendoing so at the second follow-up (endline; seetable B.1 in appendix B of the online supple-mentary appendix). Taken together, less than2% of women never re-interviewed and arelost. An analysis of the correlates of attrition,presented in table B.2 in the online supple-mentary appendix, shows that although low,attrition at endline is correlated with assign-ment to group A, but it is not correlated withassignment to group B in any of the surveys.Additionally, no baseline characteristics sig-nificantly predict attrition at either midline orendline at the 5% significance level. Theseresults indicate that, upon disbursement ofboth grants (as happens with group A at end-line), respondents seemed to feel less moti-vated to answer the surveys, but this behaviordid not manifest itself along observabledimensions of heterogeneity. The estimatedimpacts of REAP are largely robust to thisattrition, as discussed below.
Checking the Integrity of the RandomizedDesign
We test the assumption that baseline charac-teristics are uncorrelated with treatment sta-tus by comparing the distribution of thebaseline characteristics of participants. Wemake several comparisons that take into ac-count the changing composition of the treat-ment and control groups throughout theprogram’s progressive roll-out. We presentthe results in table 1.
In panel A, we present summary statistics(mean and standard deviations) of variablesthe program may impact (expenditure, in-come, savings, and asset ownership) or thatmay mediate the program’s impact (house-hold size, previous business experience, andeducation). The baseline characteristics ofthe participants (and their households) aresimilar to those of other ultra-poor house-holds in other regions of northern Kenya(see Merttens et al. 2013), suggesting that thefindings of this study may be generalizable toultra-poor women across northern Kenya.Average monthly expenditure per capita isapproximately PPP $33.96, which is wellbelow the national poverty line.Approximately 70% of this expenditure is onfood. Households are relatively large andhave approximately 3.8 children on average,with less than 50% of children enrolled inschool. Many households are food insecure,with children going to bed hungry at least
15 The capacity of the BOMA Project to reach participantsdetermined sample size. We conduct ex post power calculationsto determine if there is sufficient power, given the predeterminedsample size, to reliably estimate program impacts, and find thatin most cases the minimum detectable effect size is as low as15%.
16 The subsequent disqualification of three of the initial 1,755women led to the assignment of 585 women to the first and se-cond cycles and 582 women to the final funding cycle.
Gobin, Santos, and Toth Promoting Entrepreneurship Through Cash Transfers 1367
Downloaded from https://academic.oup.com/ajae/article-abstract/99/5/1362/4036204by George Mason University useron 11 July 2018
Ta
ble
1.
Su
mm
ary
Sta
tist
ics
an
dB
ala
nce
Ch
eck
sfo
rth
eT
rea
tme
nt
an
dC
on
tro
lG
rou
ps
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
(9)
(10
)(1
1)
(12
)(1
3)
(14
)(1
5)
(16
)(1
7)
Va
ria
ble
:
Mo
nth
lyin
com
ep
er
cap
ita
Mo
nth
lye
xp
en
dit
ure
pe
rca
pit
a
Mo
nth
lyfo
od
ex
pe
nd
itu
rep
er
cap
ita
Mo
nth
lyn
on
-fo
od
ex
pe
nd
itu
rep
er
cap
ita
Sa
vin
gs
pe
rca
pit
aT
LU
pe
rca
pit
a
Du
rab
lea
sse
tin
de
x
Me
als
pe
rd
ay
Nig
hts
chil
dh
un
gry
inp
ast
we
ek
Pro
po
rtio
no
fch
ild
ren
insc
ho
ol
Ho
use
ho
ldS
ize
#ch
ild
ren
Ma
rrie
dY
ear
so
fe
du
cati
on
Bu
sin
ess
ex
pe
rie
nce
Be
ne
fitt
ing
fro
mH
SN
P
Pa
rtic
ipa
tin
gin
CA
RE
VS
LA
Pa
ne
lA
:Me
an
sa
nd
sta
nd
ard
de
via
tio
ns
of
va
ria
ble
sa
tb
ase
lin
e� X
A2
1.7
70
34
.56
22
4.1
82
10
.38
03
.77
20
.63
8�
0.2
34
1.9
41
0.5
49
0.4
35
5.7
78
3.8
75
0.8
00
0.3
28
0.5
76
0.1
06
0.0
89
(sta
nd
ard
de
via
tio
n)
(22
.36
5)
(36
.672
)(2
8.7
39)
(18
.06
9)
(8.3
14
)(0
.72
3)
(4.0
84)
(0.3
94)
(0.6
51)
(0.2
89
)(1
.90
8)
(1.7
29
)(0
.40
0)
(1.4
54)
(0.4
95
)(0
.30
8)(0
.28
5)
Ob
serv
ati
on
s5
85
58
55
85
58
55
85
58
55
85
58
15
78
58
35
85
58
55
85
58
55
85
58
55
85
� XB
22
.31
93
4.4
80
23
.86
21
0.6
17
3.9
20
0.6
40
0.1
13
1.9
50
0.5
76
0.4
42
5.6
92
3.7
37
0.8
31
0.4
70
0.5
62
0.1
03
0.1
06
(sta
nd
ard
de
via
tio
n)
(22
.57
4)
(33
.920
)(2
5.9
91)
(18
.61
2)
(7.9
32
)(0
.89
8)
(4.5
74)
(0.3
87)
(0.7
06)
(0.2
88
)(1
.82
3)
(1.6
96
)(0
.37
5)
(1.7
30)
(0.4
97
)(0
.30
4)(0
.30
8)
Ob
serv
ati
on
s5
85
58
55
85
58
55
85
58
55
85
58
55
79
57
95
85
58
55
85
58
55
85
58
55
85
� XC
22
.44
93
2.8
25
22
.49
41
0.3
31
5.1
23
0.6
84
0.1
24
1.9
33
0.5
76
0.4
12
5.5
96
3.7
11
0.7
73
0.4
14
0.5
38
0.1
13
0.1
08
(sta
nd
ard
de
via
tio
n)
(24
.00
8)
(29
.320
)(2
1.0
90)
(15
.62
2)
(14
.42
0)
(0.8
14
)(4
.31
2)(0
.33
9)(0
.70
1)(0
.27
2)
(1.8
61
)(1
.69
1)
(0.4
19
)(1
.68
3)(0
.49
9)
(0.3
17)
(0.3
11
)O
bse
rva
tio
ns
58
25
82
58
25
82
58
25
82
58
25
80
57
25
79
58
25
82
58
25
82
58
25
82
58
2
Pa
ne
lB
:tte
stco
mp
ari
son
of
me
an
so
fb
ase
lin
ech
ara
cte
rist
ics
H0
:� X
A¼
� XBþ
C
(p-v
alu
es)
0.5
93
0.6
10
0.4
66
0.9
16
0.1
23
0.5
40
0.0
99
*0
.99
40
.43
00
.58
80
.16
30
.08
3*
0.9
19
0.1
45
0.3
02
0.8
99
0.2
20
H0
:� X
B¼
� XC
(p-v
alu
es)
0.9
27
0.3
73
0.3
23
0.7
76
0.0
78
*0
.37
90
.96
70
.42
60
.99
10
.07
5*
0.3
73
0.7
98
0.0
14
**
0.5
75
0.4
11
0.5
51
0.9
01
H0
:� X
A¼
� XC
(p-v
alu
es)
0.6
17
0.3
72
0.2
53
0.9
61
0.0
51
*0
.30
70
.14
60
.71
10
.49
60
.17
10
.10
00
.10
20
.26
40
.35
10
.19
20
.68
50
.26
8
Pa
ne
lC
:F-t
est
fro
mre
gre
ssio
no
ftr
ea
tme
nt
on
va
ria
ble
sa
bo
ve
a
Tre
atm
en
tg
rou
pC
on
tro
lg
rou
pF
-Sta
tp
-va
lue
AB
an
dC
0.7
60
.72
3B
C1
.18
0.2
83
AC
1.1
50
.30
8
No
te:
All
mo
ne
tary
va
lue
sa
rere
po
rte
din
20
14
US
D,
PP
Pte
rms.
Su
pe
rscr
ipt
ain
dic
ate
sth
isre
gre
ssio
nin
clu
de
sm
on
thly
exp
end
itu
rep
erca
pit
aa
nd
ex
clu
de
sm
on
thly
foo
dex
pen
dit
ure
per
cap
ita
an
dm
on
thly
no
n-f
oo
dex
pen
dit
ure
per
cap
ita.
Ast
eri
sks
*,*
*,
an
d*
**
sta
nd
for
sig
nifi
can
cea
tth
e1
0%
,5%
,a
nd
1%
lev
els
,re
spe
ctiv
ely
.
1368 October 2017 Amer. J. Agr. Econ.
Downloaded from https://academic.oup.com/ajae/article-abstract/99/5/1362/4036204by George Mason University useron 11 July 2018
two times per month. Households also ownvery little livestock: less than one TropicalLivestock Unit (TLU) per capita, which iswell below the self-sufficiency threshold formobile pastoralists in East African ASALs(McPeak and Barrett 2001).17 Hence, thesehouseholds appear to be overwhelminglydrawn from the cohorts that have fallen outof direct participation in the livestock econ-omy (Little et al. 2008), with asset holdingsbelow the poverty trap threshold (Barrettet al. 2006; Toth 2015). In line with this, wefind that more than half of the participants re-port having some business experience, typic-ally petty trade or selling livestock andlivestock products.
In panel B, we present the t-tests of thenull hypothesis of equality of means at base-line. These results indicate that randomiza-tion was successful in creating groups ofindividuals that are statistically balanced: inonly one case can we reject the null hypoth-esis of balance at the conventional 5% level.The results of an F-test of the joint effect ofthese variables on treatment status, asreported in panel C, reinforce thisconclusion.
Spillover Effects and Program Anticipation
The geographical proximity of individuals inthe treatment and control groups may lead tocontrol households benefiting from the prod-ucts and services of the businesses establishedunder REAP.18 Such spillovers could poten-tially bias our estimated treatment effects.We investigate three possible pathways forsuch influence: lower prices to consumers andlower profits of non-REAP businesses (bothdue to increased competition from new busi-nesses), and easier access to loans (due to theincrease in savings groups).
Given that more than 95% of the busi-nesses established under REAP are in pettytrade (primarily food items), the main impactof increased competition among businessesmay be a consequent reduction in market pri-ces of food. We do not have data on the mar-ket price for foodstuffs these businessesusually sell, hence we cannot directly test
whether this is the case. However, we haveinformation about the number of REAPbusinesses in an individual’s manyatta atbaseline, and include it as a control variablewhen estimating the impact of the program asa proxy for the effect of competition onprices.19
The increased competition from new busi-nesses established under REAP may alsoaffect the welfare of non-participant house-holds by reducing any income they may earnfrom pre-existing petty trade businesses. Theonly proxy for this increase in competition inour data is the number of individuals perbusiness at location level for which we havedata at the time of the baseline. We can esti-mate the evolution of this measure of marketsize if we are willing to assume that the onlymeaningful source of change in businessnumbers is the program and that there is noimportant change in population size. Bothassumptions seem reasonable given the con-text and the short time horizon. We can thenuse this new variable to examine the effect ofchanges in market size on income from pettytrade from pre-existing businesses. We pre-sent these estimates in table 2.20 There is noevidence of an effect of changes in marketsize on the income from petty trade earnedby control group participants.21 These resultsare consistent with recent findings from amarket-level experiment in Kenya that lever-ages a cluster randomization design to showthat business training that increases the in-come of treated women does not lead to ad-verse spillovers on non-treated women in thesame markets (McKenzie and Puerto 2017).
These tests suggesting an absence of spill-overs assume that the spillover effect is linearin our measure of market size. However, onemight be concerned that the effect thatchanges in market size have on income frompetty trade differs depending on how muchincome households earn from petty trade orby the size of the location, for example.We investigate these two non-linearities
17 Tropical Livestock Unit (TLU) is a standardized unit,designed to measure the size of a mixed livestock herd: one TLUis equivalent to one head of cattle, 0.7 camels, ten sheep/goats, ortwo donkeys.
18 We would also expect such benefits to extend to the widercommunity, but a lack of data prohibits us from examining thebenefits of REAP to households outside of our study sample.
19 Overall, there were 1,932 businesses before the program.The program funded 195 businesses (approximately 10% of thepre-existing number) in each funding cycle. See table C.1, in ap-pendix C of the supplementary material online, for furtherdetails.
20 Note that at baseline there are no statistically significant dif-ferences in per capita income from non-REAP petty trade be-tween groups A, B, and C.
21 Restricting the sample to only those households that wereengaged in petty trade at baseline does not change this conclu-sion (see table C.2 in appendix C of the supplementary materialonline).
Gobin, Santos, and Toth Promoting Entrepreneurship Through Cash Transfers 1369
Downloaded from https://academic.oup.com/ajae/article-abstract/99/5/1362/4036204by George Mason University useron 11 July 2018
separately by including interaction terms inour original specification (see tables C.3 andC.4 in appendix C of the online supplemen-tary appendix). In both cases, the interactionterms are not statistically significant.
Another potential source of spillovereffects might be easier access to loans.Although only REAP participants can ac-tively participate in all SG activities, loanscan be (and typically are) available to othermembers of the community so that they candeal with shocks and emergencies (usuallyhealth-, school-, and food-related expendi-tures). We do not have baseline informationon borrowing from REAP SGs, but we dofind that at midline, 7% of participants in thecontrol group borrowed from a REAP SG,and this value increases to 20% at endline.This change may result in our estimatedtreatment effects being downward biased if,for example, the control group is now con-suming more than they would otherwise inthe absence of REAP SGs.
Finally, bias could arise from participantschanging their behavior in anticipation ofreceiving the program.22 It is unclear in which
direction such an anticipation effect may biasour results. For example, if participants delayinvestments in anticipation of receiving thegrants, then estimated treatment effectswould be upward-biased. On the other hand,treatment effects may be downward-biased ifparticipants awaiting the grant are more will-ing to invest given the certainty of receivingthe grant, which may act as a form of insur-ance (Bianchi and Bobba 2013).
The design of the study does not allow fora straightforward way to test for anticipationeffects. However, if anticipation results in ei-ther behavior, then we would expect to finddifferences between individuals that enroll inthe program in the second and third fundingcycles, given that one group would anticipatereceiving funding six months sooner thanthe other. If this intuition is correct, thenthese differences would be ascertainable dur-ing the midline survey (when group B wouldimmediately receive the first grant, whilegroup C would still be six months away fromparticipating in the program). We check fordifferences in our outcome variables: monthlyincome per capita, monthly expenditure percapita, monthly consumption per capita, sav-ings per capita, TLU per capita, durable assetindex, and the number of nights that a childhas gone to bed hungry in the last week.23 Wesee no statistically significant differences be-tween groups B and C, as shown in columns(1) to (7) of table 3.
Additionally we find no statistically signifi-cant differences in income earned from non-REAP businesses (table 3, column [8]), whichsuggests that participants awaiting the grantdid not alter their pre-existing business prac-tices. We also do not find any statistically sig-nificant differences in how these two groupsof participants allocate their time (see tableF.1 in appendix F of the online supplemen-tary material) or in the proportion of womenthat have ever taken a loan at midline.24 Lessthan 2% of the women in groups B and Cwho took loans used them for investment in abusiness or livestock. The limited use of loans
Table 2. Spillover Effect of REAP onIncome from Non-REAP Petty Trade
(1)Monthly incomefrom non-REAP
petty trade per capita
Population per business 0.047(0.037)
Midline 1.189**(0.588)
Endline 0.631(0.661)
Observations 2851R-squared 0.029
Note: The sample is restricted to participants in groups B (baseline and mid-
line) and C (all surveys) who received funding in Sept./Oct. 2013 and
March/April 2014, respectively. The dependent variable is monthly income
from non-REAP petty trade per capita. Population per business is derived by
dividing the number of individuals in a location by the number of businesses
in that location (using figures reported in table C.1 of the online supplemen-
tary appendix). Additional controls include location fixed effects. Robust
standard errors, clustered at the sub-location level, are shown in parenthe-
ses. All monetary values are reported in 2014 USD, PPP terms. Asterisks *,
**, and *** indicate significance at the 10%, 5%, and 1% levels,
respectively.
22 Given that we are dealing with ultra-poor women, it is diffi-cult to conjecture how behavior would change in anticipation ofthis program. Individuals might try to observe other businessesand how they operate, or business groups might meet to discusswhat will happen when they are enrolled in the program, but
both lack of access to capital and human capital constraints arelikely to prevent them from taking any action that would affectmeasured outcomes.
23 We define these outcome variables in appendix D and ap-pendix E of the supplementary material online.
24 Approximately 24.2% (24.3%) of group B (C) participantshave ever taken a loan from banks, Microfinance institutions(MFIs), moneylenders, savings and self-help groups, or family.More than 36% of participants in group A have accessed loans andthis is statistically significantly higher compared to groups B and C.
1370 October 2017 Amer. J. Agr. Econ.
Downloaded from https://academic.oup.com/ajae/article-abstract/99/5/1362/4036204by George Mason University useron 11 July 2018
for business investment is attributable to thelimited access to capital in this region asOsterloh and Barrett (2007) demonstrate insimilar locations in northern Kenya. Thoughnot definitive for disproving anticipationeffects, these results suggest that anticipationeffects should be of limited importance,if any.
Main Results
The random assignment of treatment statusallows us to obtain unbiased estimates ofREAP’s impact by estimating the followingregression for each outcome of interest:
ð1Þ YiðtÞ ¼ hþ bTijðtÞ þ dYið0Þ þ sMi
þ uXið0Þ þ eij¼ f1; 2g; t ¼ fmidline; endlineg
where Y i(t) is the outcome of interest forhousehold i at time t, Y i(0) is the baselinevalue of the outcome variable for householdi, Mi is a set of location dummy variables, andXi(0) is the number of REAP businesses inan individual’s manyatta at baseline.25
Finally, Tij is the treatment status of individ-ual i. Given the structure of the program, wecan consider two sets of interventions,indexed by j: business training, a cash grantof $100, and mentoring that the programintroduces first and that we label Ti1, fol-lowed by savings training, an additional cashgrant of $50, and continued mentoring, whichwe label Ti2.
Simplifying the notation by dropping theith individual subscript, it is clear from thedescription of the program (and from figure 1)that we can observe T1 at both midline andendline (T1(midline) and T1(endline)), andthe joint effect of the two sets of interven-tions at endline ðT1 þ T2ðendlineÞÞ. To esti-mate the impact of T1 at midline, we comparegroup A to a combined control group com-prised of those benefiting from the programin the second and third cycles (i.e., groups Band C). We can similarly estimate the impactof T1 on group B at endline by comparinggroup B with the control group C. We canuse these two estimates to test the hypothesis
Ta
ble
3.
Te
stin
gfo
rA
nti
cip
ati
on
Eff
ect
s
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
Vari
ab
le:
Mo
nth
lyin
com
ep
er
cap
ita
Mo
nth
lyexp
en
dit
ure
per
cap
ita
Mo
nth
lyco
nsu
mp
tio
np
er
cap
ita
Savin
gs
per
cap
ita
TL
U per
cap
ita
Du
rab
leass
et
ind
ex
Nig
hts
that
chil
dh
as
go
ne
tob
ed
hu
ngry
inp
ast
week
Mo
nth
lyin
com
efr
om
no
n-R
EA
Pp
ett
ytr
ad
ep
er
cap
ita
¼1
ifin
gro
up
B2.7
44
�2.0
82
�1.2
38
0.6
10
�0.0
72
�0.1
25
�0.0
84
0.3
04
(2.8
24)
(2.6
66)
(2.2
83)
(0.6
53)
(0.0
96)
(0.4
07)
(0.0
54)
(0.4
38)
Ob
serv
ati
on
s1133
1133
1106
1133
1133
1133
1077
1133
R-s
qu
are
d0.0
68
0.2
93
0.1
61
0.0
62
0.2
60
0.2
57
0.1
83
0.0
66
Gro
up
Cm
ean
23.4
37
51.7
03
49.8
28
3.1
78
1.1
19
2.0
86
0.5
65
1.5
51
No
te:G
rou
ps
Ba
nd
Cre
fer
toR
EA
Pp
art
icip
an
tsw
ho
rece
ive
dfu
nd
ing
inS
ep
t./O
ct.
20
13
an
dM
arc
h/A
pri
l2
01
4,r
esp
ect
ive
ly.
Est
ima
tes
are
fro
mO
rdin
ary
lea
stsq
ua
res
(OL
S)
reg
ress
ion
su
sin
gm
idli
ne
da
taco
lle
cte
dp
rio
rto
Gro
up
B
pa
rtic
ipa
nts
rece
ivin
gth
efi
rst
gra
nt.
Th
esa
mp
leis
rest
rict
ed
top
art
icip
an
tsin
gro
up
sB
an
dC
an
dth
ere
fere
nce
cate
go
ryin
the
reg
ress
ion
sis
gro
up
C.
Re
gre
ssio
ns
incl
ud
eco
ntr
ols
for
the
nu
mb
er
of
RE
AP
bu
sin
ess
es
ina
ma
ny
att
aa
tb
ase
-
lin
e,
loca
tio
nfi
xe
de
ffe
cts,
an
db
ase
lin
ele
ve
lso
fth
eo
utc
om
ev
ari
ab
le(w
ith
the
ex
cep
tio
no
fm
on
thly
con
sum
pti
on
per
cap
ita
for
wh
ich
ba
seli
ne
lev
els
are
no
ta
va
ila
ble
).R
ob
ust
sta
nd
ard
err
ors
,cl
ust
ere
da
tth
esu
b-l
oca
tio
nle
ve
l,a
resh
ow
n
inp
are
nth
ese
s.A
llm
on
eta
ryv
alu
es
are
rep
ort
ed
in2
01
4U
SD
,P
PP
term
s.A
ste
risk
s*
,*
*,a
nd
**
*in
dic
ate
sig
nifi
can
cea
tth
e1
0%
,5%
,a
nd
1%
lev
els
,re
spe
ctiv
ely
.
25 When we exclude the number of REAP businesses in anindividual’s manyatta at baseline from this model, we find no dif-ference in the statistical significance or size of the estimatedimpacts of the program.
Gobin, Santos, and Toth Promoting Entrepreneurship Through Cash Transfers 1371
Downloaded from https://academic.oup.com/ajae/article-abstract/99/5/1362/4036204by George Mason University useron 11 July 2018
that the impact of T1 is constant throughoutthe period
ð2Þ H0 : b½T1ðmidlineÞ� ¼ b½T1ðendlineÞ�:
Failure to reject equation (2) would sug-gest that the impact of this subset of interven-tions is stable. This, in turn, provides furthersupport to our assumption that there were noadverse effects from late entry into treatment(due, e.g., to increased market competition).
It is important to note that failure to rejectequation (2) is not enough to plausibly iden-tify the impact of T2 in isolation given that, atmidline, beneficiaries of T1 will potentially bedifferent from the same individuals at base-line, both in ways that are easy to control(e.g., asset ownership) and in ways that arenot easy to observe (e.g., experience in man-aging a business as part of a group).Therefore, without further assumptionsregarding how such variables influence theoutcomes we analyze, we are limited in ourability to identify the effect of T2 conditionalon previously benefiting from T1.
The Six-Month Impact of REAP
Panels A and B of table 4 provide the esti-mates of the impact of T1 at midline and end-line, respectively. We cluster standard errors(presented in parentheses) at the sub-locationlevel, as this is the level at which stratificationoccurred.26 We also present (within squarebrackets) adjusted p-values or q-values that ac-count for testing the program’s impact on sev-eral outcomes.27 We obtained the q–values byimplementing the step-up method to controlfor the false discovery rate (FDR) asBenjamini and Hochberg (1995) proposed.28
After accounting for the possibility of mul-tiple inference (by adjusting p-values), andsearching for consistent impacts across bothperiods, we see that beneficiaries have highermonthly income per capita after six months ofbenefiting from REAP (column [1]). These
changes are economically significant in bothperiods, and represent an improvement of44.6% over the control group mean (or 0.255Standard deviation (SDs)) at midline, and34.7% over the control group mean (or 0.250SDs) at endline.
Somewhat surprisingly, these changes donot seem to translate into changes in monthlyexpenditure per capita or monthly consump-tion per capita (reported in columns [2] and[3]), which, although positive, are much lessprecisely estimated. This is especially trueduring endline, when we can reject the equal-ity between increases in income and expend-iture (p – value ¼ 0.087) and between incomeand consumption (p – value ¼ 0.014), al-though it is also clear at midline when com-paring increases in income and consumption(p – value ¼ 0.008).
One explanation for this discrepancy is theallocation of additional income to asset accu-mulation. Our data offer some support to thisexplanation, in particular at endline when weobserve increases in savings (column [4]) anddurable assets (column [6]) that, like thechanges in income, are economically signifi-cant (34.6% or 0.203SDs, and 26.4% or 0.112SDs above the respective control groupmean). The effects on livestock (column [5])and on the number of nights a child has goneto bed hungry (column [7]) are less preciselyestimated, and marginally insignificant afteraccounting for multiple inference. Despite thisapparent difference in the impact of T1 be-tween periods, with the effects being appar-ently more positive in the second period, wecan never reject the null hypothesis of equalityof impact across periods (equation [2]).29
The One-Year Impact of REAP
Table 4, panel C, provides estimates of thecombined impact of T1 and T2 after one yearof participation in REAP (i.e.,ðT1 þ T2ðendlineÞÞ). These estimates alignwith those in panels A and B (i.e., the impactof T1), with treated participants reporting sig-nificantly higher income per capita, savingsper capita, and durable asset ownership.After one year of participation in REAP, in-come per capita is 30.8% (0.222 SDs) highercompared to the control group mean, andsavings per capita is 131.4% (0.769 SDs)
26 As previously noted, locations are comprised of sub-locations that are made up of smaller villages known as manyat-tas. There are sub-locations.
27 Because we estimate the impacts of REAP on several out-comes, we increase the probability of type 1 errors by testingmultiple simultaneous hypotheses at set p-values. For example,by performing seven independent tests, the probability of a type1 error is no longer 0.05, but rather 0.302. The adjusted p-valuesshould be interpreted as the smallest significance level at whichthe null hypothesis is rejected.
28 We make use of the procedure outlined byAnderson (2008) to obtain the q–values.
29 Depending on outcome, the q-values are between 0.588 and0.680 (see table 4 panel D).
1372 October 2017 Amer. J. Agr. Econ.
Downloaded from https://academic.oup.com/ajae/article-abstract/99/5/1362/4036204by George Mason University useron 11 July 2018
Ta
ble
4.
Th
eIm
pa
cts
of
RE
AP
on
Ho
use
ho
ldO
utc
om
es
(1)
(2)
(3)
(4)
(5)
(6)
(7)
Ou
tco
me:
Mo
nth
lyin
com
ep
er
cap
ita
Mo
nth
lyexp
en
dit
ure
per
cap
ita
Mo
nth
lyco
nsu
mp
tio
np
er
cap
ita
Savin
gs
per
cap
ita
TL
Up
er
cap
ita
Du
rab
leass
et
ind
ex
Nig
hts
that
chil
dh
asgo
ne
tob
ed
hu
ngry
Pan
el
A:
Th
eim
pact
sat
six
mo
nth
sm
eas
ure
dat
mid
lin
eT
reatm
ent
11.0
84*
**
5.6
68
2.7
73
1.0
77
�0.0
11
0.5
26
�0.0
70
(2.6
39)
(3.9
33)
(2.3
35)
(0.7
14)
(0.0
44)
(0.3
92)
(0.0
60)
[0.0
01]
[0.2
91]
[0.2
91]
[0.2
91]
[0.8
09]
[0.1
05]
[0.2
91]
Ou
tco
me
at
base
lin
e0.1
46*
*0.2
57***
0.0
68***
0.4
17***
0.5
03***
0.0
08
(0.0
66
)(0
.061)
(0.0
24)
(0.1
07)
(0.0
51)
(0.0
93)
RE
AP
bu
sin
esse
sin
man
yatt
a�
0.4
98*
�0.1
20
�0.2
24
�0.0
55
�0.0
14
0.1
26*
*�
0.0
05
(0.2
72)
(0.5
45)
(0.3
10)
(0.0
68)
(0.0
17)
(0.0
53)
(0.0
08)
Ob
serv
ati
on
s1682
1682
1646
1682
1682
1682
1597
R-s
qu
are
d0.0
67
0.1
98
0.1
42
0.0
78
0.2
70
0.2
31
0.1
89
Co
ntr
ol
gro
up
mean
24.8
49
50.8
05
49.2
25
3.4
30
1.0
75
2.0
50
0.5
14
Pan
el
B:
Th
eim
pact
sat
six
mo
nth
sm
easu
red
at
en
dli
ne
Tre
atm
ent
8.7
44**
2.4
46
0.0
38
1.5
36**
0.1
80
0.7
51*
�0.1
53
(3.2
09)
(3.9
63)
(2.1
34)
(0.6
07)
(0.0
99)
(0.4
01)
(0.0
83)
[0.0
49]
[0.6
29]
[0.9
86]
[0.0
49]
[0.1
05]
[0.0
89]
[0.1
05]
Ou
tco
me
at
base
lin
e0.1
11
0.2
86***
0.0
58
0.3
52**
0.5
09***
0.0
14
(0.0
90
)(0
.083)
(0.0
42)
(0.1
43)
(0.0
68)
(0.0
93)
RE
AP
bu
sin
esse
sin
man
yatt
a0.7
28
0.0
91
�0.2
32
�0.0
88
�0.0
10
0.1
28*
*0.0
03
(0.5
06)
(0.3
83)
(0.2
58)
(0.0
74)
(0.0
18)
(0.0
60)
(0.0
08)
Ob
serv
ati
on
s1117
1117
1117
1117
1117
1117
1089
R-s
qu
are
d0.0
54
0.0
33
0.2
16
0.0
29
0.1
29
0.2
74
0.0
33
Co
ntr
ol
gro
up
mean
25.2
32
57.3
94
46.2
45
4.4
40
1.3
03
2.8
43
0.7
89
Pan
el
C:
Th
eim
pact
sat
on
eyear
measu
red
at
en
dli
ne
Tre
atm
ent
7.7
69***
�1.5
65
�2.1
95
5.8
36***
0.1
58
0.9
94*
*�
0.1
08
(2.0
67)
(3.4
44)
(2.2
02)
(0.7
65)
(0.0
86)
(0.3
72)
(0.0
87)
[0.0
01]
[0.6
51]
[0.3
76]
[0.0
01]
[0.1
21]
[0.0
21]
[0.3
04]
Ou
tco
me
at
base
lin
e0.1
59***
0.2
91***
0.0
40
0.7
63***
0.5
71***
0.3
27*
**
(0.0
60)
(0.0
74)
(0.0
69)
(0.2
62)
(0.0
65)
(0.0
98)
RE
AP
bu
sin
esse
sin
man
yatt
a0.0
33
0.3
13
�0.0
10
�0.1
21*
�0.0
09
0.1
57*
*�
0.0
08
(0.3
94)
(0.3
56)
(0.3
39)
(0.0
72)
(0.0
13)
(0.0
75)
(0.0
10)
Ob
serv
ati
on
s1095
1095
1095
1095
1095
1095
1068
R-s
qu
are
d0.0
72
0.0
54
0.1
66
0.0
97
0.1
88
0.2
75
0.0
34
Co
ntr
ol
gro
up
mean
25.2
32
57.3
94
46.2
45
4.4
40
1.3
03
2.8
43
0.7
89
Co
nti
nu
ed
Gobin, Santos, and Toth Promoting Entrepreneurship Through Cash Transfers 1373
Downloaded from https://academic.oup.com/ajae/article-abstract/99/5/1362/4036204by George Mason University useron 11 July 2018
higher compared to the control group mean,with both increases being statistically signifi-cant at the 1% level. Additionally, durableasset ownership is 35.0% (0.149 SDs) highercompared to the control group mean, a differ-ence that is significant at the 5% level.
As before, we find that the increase inhousehold income does not translate to an in-crease in expenditure or consumption, whichin fact decrease by 2.7% (0.027 SDs) and4.7% (0.064 SDs), respectively, althoughthese decreases are not statistically signifi-cant. As before, the impact on livestock, al-though positive (an increase of 12.1% [0.099SDs] over the control group mean), is notstatistically significant at conventionallevels once we account for multiple inference(q– value ¼ 0.121).
Robustness Tests
We check the robustness of our results to at-trition, to different specifications of equation(1), and to the effect of outliers. Detailedresults of these tests are presented in tablesG.1, G.2, and G.3 in appendix G of the sup-plementary appendix online.
Attrition, although low (recall that lessthan 2% of women are never re-interviewed),correlates with assignment to group A at end-line. We check for the robustness of ourresults to attrition by estimating differenttypes of attrition bounds. We considerbounds that either impute missing observa-tions using observations for those participantswho did not attrit, or bounds that trim obser-vations from the treatment arm that suffersfrom less attrition.30 The results are largelyrobust to attrition, though impacts on live-stock and assets show more sensitivity to thedegree of conservatism in selecting bounds(see appendix G in the supplementary onlinematerial).
In appendix G we also present results fromother regression specifications. We re-estimate equation (1) using difference-in-difference (DiD) estimates to check thatour results are not being driven by any ob-servable or unobservable baseline imbalancebetween our treatment and control group.We also re-estimate equation (1) with no con-trol variables to check that our results are notbeing driven by these variables. In both cases
Ta
ble
4.
con
tin
ue
d
(1)
(2)
(3)
(4)
(5)
(6)
(7)
Ou
tco
me:
Mo
nth
lyin
com
ep
er
cap
ita
Mo
nth
lyexp
en
dit
ure
per
cap
ita
Mo
nth
lyco
nsu
mp
tio
np
er
cap
ita
Savin
gs
per
cap
ita
TL
Up
er
cap
ita
Du
rab
leass
et
ind
ex
Nig
hts
that
chil
dh
as
go
ne
tob
ed
hu
ngry
Pan
el
D:P
-valu
es
an
dad
just
ed
p-v
alu
es
(in
bra
ckets
)fr
om
Wald
test
so
fth
eeq
uiv
ale
nce
of
treatm
en
teff
ect
s6
mo
nth
(mid
lin
e)¼
6m
on
th(e
nd
lin
e)
0.5
83
0.5
80
0.4
28
0.6
04
0.0
84
0.6
80
0.4
28
[0.6
80]
[0.6
80]
[0.6
80]
[0.6
80]
[0.5
88]
[0.6
80]
[0.6
80]
6m
on
th(m
idli
ne)¼
1year
(en
dli
ne)
0.3
69
0.1
72
0.0
66
0.0
00
0.0
63
0.1
94
0.7
22
[0.4
31]
[0.2
72]
[0.1
54]
[0.0
01]
[0.1
54]
[0.2
72]
[0.7
22]
6m
on
th(e
nd
lin
e)¼
1year
(en
dli
ne)
0.7
37
0.3
04
0.2
73
0.0
00
0.8
30
0.5
31
0.5
74
[0.8
30]
[0.7
10]
[0.7
10]
[0.0
01]
[0.8
30]
[0.8
04]
[0.8
04]
No
te:
InP
an
el
A,th
etr
ea
tme
nt
gro
up
isg
rou
pA
an
dth
eco
ntr
ol
gro
up
isg
rou
ps
Ba
nd
Cco
mb
ine
d.
InP
an
el
B,
the
tre
atm
en
tg
rou
pis
gro
up
Ba
nd
con
tro
lg
rou
pis
gro
up
C.
InP
an
el
C,
the
tre
atm
en
tg
rou
pis
gro
up
Aa
nd
con
tro
lg
rou
pis
gro
up
C.
Gro
up
sA
,B
,a
nd
Cre
fer
toR
EA
Pp
art
icip
an
tsw
ho
rece
ive
dfu
nd
ing
inM
arc
h/A
pri
l2
01
3,
Se
pt.
/Oct
.2
01
3,
an
dM
arc
h/A
pri
l2
01
4,
resp
ect
ivel
y.
Re
gre
ssio
ns
incl
ud
eco
ntr
ols
for
the
nu
mb
er
of
RE
AP
bu
sin
ess
es
ina
ma
ny
att
aa
t
ba
seli
ne
,lo
cati
on
fix
ed
eff
ect
sa
nd
ba
seli
ne
lev
els
of
the
ou
tco
me
va
ria
ble
(wit
hth
ee
xce
pti
on
of
mo
nth
lyco
nsu
mp
tio
np
ercc
ap
ita
for
wh
ich
ba
seli
ne
lev
els
are
no
ta
va
ila
ble
).R
ob
ust
sta
nd
ard
err
ors
,cl
ust
ere
da
tth
esu
b-l
oca
tio
nle
ve
l,a
re
sho
wn
inp
are
nth
ese
s,w
hil
eq
-va
lue
s,u
sin
gth
eB
en
jam
ini-
Ho
chb
erg
ste
p-u
pm
eth
od
,a
resh
ow
nin
bra
cke
ts.A
llm
on
eta
ryv
alu
es
are
rep
ort
ed
in2
01
4U
SD
,PP
Pte
rms.
Ast
eri
sks
*,
**
,a
nd
**
*in
dic
ate
sig
nifi
can
cea
tth
e1
0%
,5%
,an
d1
%
lev
els
,re
spe
ctiv
ely
(ba
sed
on
ad
just
ed
p-v
alu
es)
.
30 These bounds are described in detail in appendix G of thesupplementary material online.
1374 October 2017 Amer. J. Agr. Econ.
Downloaded from https://academic.oup.com/ajae/article-abstract/99/5/1362/4036204by George Mason University useron 11 July 2018
our previous conclusions are robust to thesealternative specifications.
Finally, we test for the robustness of ourresults to outliers by replacing outcomesabove the 99th percentile with values at the99th percentile. We find that our previousconclusions are robust to outliers.
Discussion
Income. The REAP significantly increasedthe income that participants earned in the
short-to-medium-run (i.e., six months andone year after participation in the program).Microenterprise formation clearly seems tobe the mechanism that led to this positiveoutcome.
The results appearing in table 5, where wedisaggregate income changes by source, sup-port this conclusion. The data show that theoverall increase in income is the result ofchanges in income from non-agriculturaltrade (column [4]), which includes incomefrom the REAP microenterprise, measured
Table 5. The Impacts of REAP on Sources of Income
(1) (2) (3) (4) (5) (6)Variable: Monthly
totalincome per
capita
Monthlyincome from
livestockper capita
Monthlyincome
from otheragricultureper capita
Monthlyincome
fromnon-agritrade
per capita
Monthlyincome
from laborper capita
Monthlyincome
fromtransfers
per capita
Panel A: The impacts at six months measured at midlineTreatment 11.084*** 1.146 �0.003 10.041*** �0.268 0.084
(2.639) (1.960) (0.076) (1.732) (0.216) (0.165)[0.765] [0.970] [0.001] [0.545] [0.765]
Observations 1682 1682 1682 1682 1682 1682R-squared 0.067 0.050 0.036 0.195 0.059 0.086Control group mean 24.849 19.817 0.117 3.031 1.205 0.678
Panel B: The impacts at six months measured at endlineTreatment 8.744** 3.938 0.108 4.458*** 0.334 �0.064
(3.208) (2.563) (0.110) (0.633) (1.239) (0.283)[0.323] [0.550] [0.001] [0.821] [0.821]
Observations 1117 1117 1117 1117 1117 1117R-squared 0.054 0.066 0.042 0.092 0.023 0.036Control group mean 25.232 19.811 0.172 1.534 2.911 0.803
Panel C: The impacts at one year measured at endlineTreatment 7.769*** 1.177 0.039 5.683*** 0.971 �0.132
(2.067) (1.780) (0.141) (0.668) (0.879) (0.248)[0.745] [0.781] [0.001] [0.683] [0.745]
Observations 1095 1095 1095 1095 1095 1095R-squared 0.072 0.073 0.031 0.115 0.029 0.036Control group mean 25.232 19.811 0.172 1.534 2.911 0.803
Panel D: p-values and adjusted p-values (in brackets) from Wald tests of the equivalence of treatmenteffects
6 month (midline) ¼6 month (endline)
0.372 0.380 0.005 0.637 0.613
[0.634] [0.634] [0.025] [0.637] [0.637]6 month (midline) ¼
1 year (endline)0.990 0.807 0.025 0.163 0.434
[0.990] [0.990] [0.125] [0.408] [0.724]6 month (endline) ¼
1 year (endline)0.274 0.575 0.049 0.596 0.769
[0.685] [0.745] [0.245] [0.745] [0.769]
Note: In Panel A, the treatment group is group A and control group is groups B and C combined. In Panel B, the treatment group is group B and control group is
group C. In Panel C, the treatment group is group A and control group is group C. Groups A, B, and C refer to REAP participants who received funding in March/
April 2013, Sept./Oct. 2013, and March/April 2014, respectively. Column (1) repeats the values from table 4, column (1). Regressions include controls for the num-
ber of REAP businesses in a manyatta at baseline, location fixed effects and baseline levels of the outcome variable. Robust standard errors, clustered at the sub-lo-
cation level, are shown in parentheses, while q-values, using the Benjamini-Hochberg step-up method, are shown in brackets. All monetary values are reported in
2014 USD, PPP terms. Asterisks *, **, and *** indicate significance at the 10%, 5%, and 1% levels, respectively (based on adjusted p-values).
Gobin, Santos, and Toth Promoting Entrepreneurship Through Cash Transfers 1375
Downloaded from https://academic.oup.com/ajae/article-abstract/99/5/1362/4036204by George Mason University useron 11 July 2018
as the profits that the participant earned fromthe business. The increase in income fromnon-agricultural trade is statistically signifi-cant at the 1% level of significance, and thiseffect persists for up to one year after REAPenrollment. When we examine how partici-pants allocate their time at endline, we findthat those who benefited from REAP spendan average of approximately 6% of their dayon REAP-related activities. To achieve thisincreased activity, participants decreased theaverage time spent on leisure, householdactivities, and other productive activities inroughly equal amounts (see table F.2 in ap-pendix F of the supplementary onlinematerial).31
The six-month impact of the program interms of income from non-agricultural tradeis significantly lower at endline compared tomidline. There are two potentially comple-mentary explanations for this difference:competition and shocks.32 As noted in ourprevious discussion of spillover effects, wefind no evidence of a relationship betweenchanges in our estimate of market size (thatreflects increased competition) and incomefrom non-REAP businesses (mostly non-agricultural trade) in our control group (seetable 2). Given that both REAP and non-REAP businesses operate in the same sector,we discount the importance of competition inexplaining the apparent differences in incomebetween midline and endline.
Turning to shocks, we note that our follow-up surveys occurred during the long dry sea-son of 2013 (midline) and the short dry sea-son of 2014 (endline). The short dry season of2014 followed a short 2013 rainfall season,with levels of rainfall below average.33 In theASALs of east Africa, rainfall shocks such asthis lead to less pasture and reductions in
income from pastoralism, which is the maineconomic activity in the area. This naturallyreduces demand. Although this coincidenceis not enough to dismiss alternative explana-tions, it is likely that it played a role in thedifferences we observe.34
Expenditure and consumption. Increases inincome due to REAP do not translate intoincreases in wellbeing as measurable throughconsumption or expenditure. The programseems to lead to a short-term increase and amedium-term reduction in consumption andexpenditure, but these effects are never pre-cisely estimated (q-values between 0.291 and0.986, depending on the outcome, the treat-ment, and the time horizon). This result isperhaps surprising given the low level of ini-tial consumption, but it is not unique, as wediscuss below. It also warrants considerationin the context of the program’s overall func-tioning and the emphasis placed on savings,as well as investment decisions in durableassets and livestock, which we discuss next.
Savings, livestock, and other assets. Afterthe training on savings (including the func-tioning of savings groups) that occurs aftersix months of participation in the program,more than 95% of participants join a savingsgroup, a decision that is both voluntary andindividual (while at baseline only 10% weremembers of pre-existing SGs). It is thereforenot surprising that after one year of participa-tion in REAP, participants have higher sav-ings per capita.
What might be surprising is that before thistraining, participants also saved more percapita, suggesting a shift in savings behaviorthat occurs even before the formal introduc-tion of savings groups. If we look moreclosely at the savings mechanisms thesewomen use (table 6), we see that after sixmonths (measured at endline), REAP partici-pants are saving more at home (column [1])compared to the control group.
Given the economic and social importanceof livestock among participants, one wouldexpect an investment of some of the
31 It is possible that other labor, either from children or othermembers of the household, is being mobilized to compensate forthe reduction in women’s time in some non-REAP activities. Wedo not have data on other household members time use thatwould allow us to test for this possibility. However, if child laboris increasing as a result of the new businesses, this does not resultin a decrease in school enrollment.
32 A third possible explanation, learning, also seems not to bein effect here given that income from non-agricultural trade atendline is lower than at midline.
33 The climate in the study site in characterized by two wetand two dry seasons. From January to March there is a short dryspell that is followed by long rains from April to June. From Julyto October there is a long dry period and this is followed by shortrains in November and December. See http://reliefweb.int/sites/reliefweb.int/files/resources/Marsabit-January-2014.pdf for an as-sessment of the conditions in Marsabit County in January 2014,one month before the endline survey was conducted.
34 Despite this shock, there is evidence of continued capacityof these businesses to operate: only two businesses (formed dur-ing the first funding cycle) may have disappeared (see table B.1of the supplementary material online) and the total value of thebusiness (i.e., the sum of cash on hand, business savings andcredit outstanding, and business stock and assets) is significantlyhigher at endline (for both sets of participants) compared to thebusiness value at midline (PPP $374.61 and PPP $451.55 forgroup B and group A at endline, respectively, compared to PPP$305.50 for group A at midline).
1376 October 2017 Amer. J. Agr. Econ.
Downloaded from https://academic.oup.com/ajae/article-abstract/99/5/1362/4036204by George Mason University useron 11 July 2018
increased income from entrepreneurial activ-ities in the acquisition of livestock. We findincreased livestock ownership among REAPparticipants, which aligns with these expecta-tions; however, the estimates are not statistic-ally significant once we account for multipleinference.35 Treated households also investmore in durable assets such as blankets andmobile phones, which may improve their liv-ing conditions.36
Graduation from poverty. The central aimof this program is to graduate participantsfrom poverty, which we equate with beingabove the Kenyan rural poverty line asreported by the Kenya National Bureau ofStatistics (2007). In table 7 we provide esti-mates of REAP’s impact on the probabilityof being non-poor at six months and one yearafter the start of the program, when the defin-ing characteristics of poverty lines are incomeor expenditure.
We find that beneficiaries are more likelyto have incomes above the poverty line bothafter six months and one year of participationin REAP, and these effects are statisticallysignificant at the 1% level (column [1]). Atmidline (endline) we find that T1 increases
Table 6. The Impacts of REAP on Savings Mechanism
(1) (2) (3)
Variable:Personal savings
per capita
Per capita savingsin non-REAPsavings group
Per capita savingsin REAP
savings group
Panel A: The impact at six months measured at midlineTreatment 1.014 0.054 –
(0.611) (0.226) –[0.202] [0.813] –
Observations 1682 1682 –R-squared 0.085 0.054 –Control group mean 3.030 0.400 0Panel B: The impact at six months measured at endlineTreatment 1.540*** �0.034 –
(0.534) (0.211) –[0.010] [0.871] –
Observations 1117 1117 –R-squared 0.026 0.033 –Control group mean 4.082 0.357 0Panel C: The impact at one year measured at endlineTreatment 1.445** 0.207 4.170***
(0.558) (0.333) (0.326)[0.018] [0.536] [0.001]
Observations 1095 1095 1095R-squared 0.028 0.015 0.556Control group mean 4.082 0.357 0Panel D: p-values and adjusted p-values (in brackets) from Wald tests of the equivalence of treatment
effects6 month (midline) ¼ 6 month (endline) 0.501 0.770
[0.770] [0.770]6 month (midline) ¼ 1 year (endline) 0.560 0.690
[0.690] [0.690]6 month (endline) ¼ 1 year (endline) 0.888 0.390
[0.888] [0.780]
Note: In Panel A, the treatment group is group A and control group is groups B and C combined. In Panel B, the treatment group is group B and control
group is group C. In Panel C, the treatment group is group A and control group is group C. Groups A, B, and C refer to REAP participants who received
funding in March/April 2013, Sept./Oct. 2013, and March/April 2014, respectively. Regressions include controls for the number of REAP businesses in a man-
yatta at baseline, location fixed effects and baseline levels of the outcome variable. Robust standard errors, clustered at the sub-location level, are shown in
parentheses, while q-values, using the Benjamini-Hochberg step-up method, are shown in brackets. All monetary values are reported in 2014 USD, PPP
terms. Personal savings includes savings kept at home and savings kept at a formal financial institution, including mobile service providers. Asterisks *, **,
and *** indicate significance at the 10%, 5%, and 1% levels, respectively (based on adjusted p-values).
35 See table H.1 in appendix H of the supplementary materialonline for a breakdown of the impact of REAP on componentsof TLU.
36 See table I.1 in appendix I of the supplementary materialonline for a breakdown of the impact of REAP on componentsof the durable asset index.
Gobin, Santos, and Toth Promoting Entrepreneurship Through Cash Transfers 1377
Downloaded from https://academic.oup.com/ajae/article-abstract/99/5/1362/4036204by George Mason University useron 11 July 2018
the probability that beneficiaries areabove the poverty line by 13.8% (6.5%), aneffect that represents a 81.7% (38.7%) in-crease over the control group probability ofbeing above the poverty line. The effects aresimilar at one year, with beneficiaries being13.2% more likely to have incomes above thepoverty line (a 78.6% increase over the controlgroup). Hence, the results are consistent withthe idea that REAP can serve as a “cargo net”(Barrett, Carter, and Ikegami 2008) for thosewho have fallen out of the direct livestockeconomy by allowing beneficiaries to establisha different livelihood, through microenterprisesoutside pastoralism.
Examining the impact that REAP has onthe probability that a beneficiary has expend-iture or consumption above the poverty line(columns [2] and [3], respectively) reveals asmall increase in this probability in thetreated group at midline and negligiblechanges at endline. However, none of theseimpacts are statistically significant at
conventional levels, as expected, given theearlier findings on expenditure andconsumption.
Impact heterogeneity. We next consider theevidence for differentiated impacts of REAPacross the distribution of outcomes. In table 8we present quantile regression estimates atthe 10th, 25th, 50th, 75th, and 90th percen-tiles of the distribution of outcomes atsix months (panels A and B) and one year(panel C).37 In figure J.1 in appendix J of thesupplementary material online, we
Table 7. The Impacts of REAP on the Probability of Living above the Poverty Line
(1) (2) (3)Variable: Probability that
income per adultequivalent is above
the poverty line
Probability thatexpenditure peradult equivalent
is above thepoverty line
Probability thatconsumption peradult equivalent
is abovethe poverty line
Panel A: The impact at six months measured at midlineTreatment 0.138*** 0.033 0.037
(0.034) (0.032) (0.025)Observations 1682 1682 1646R-squared 0.124 0.244 0.270Control group mean 0.169 0.500 0.606
Panel B: The impact at six months measured at midlineTreatment 0.065** �0.001 0.012
(0.029) (0.042) (0.030)Observations 1117 1117 1117R-squared 0.043 0.026 0.193Control group mean 0.168 0.581 0.526
Panel C: The impact at one year measured at endlineTreatment 0.132*** 0.017 �0.013
(0.025) (0.038) (0.032)Observations 1095 1095 1095R-squared 0.074 0.022 0.199Control group mean 0.168 0.581 0.526
Note: Estimates from a linear probability model. In Panel A, the treatment group is group A and control group is groups B and C combined. In Panel B, the
treatment group is group B and control group is group C. In Panel C, the treatment group is group A and control group is group C. Groups A, B, and C refer
to REAP participants who received funding in March/April 2013, Sept./Oct. 2013, and March/April 2014, respectively. Regressions include controls for the
number of REAP businesses in a manyatta at baseline, location fixed effects and baseline levels of the outcome variable (with the exception of probability
that consumption per adult equivalent is above the poverty line for which baseline levels are not available). Robust standard errors, clustered at the sub-loca-
tion level, are shown in parentheses. The Kenyan rural poverty line used is as defined by the Kenya National Bureau of Statistics (2007) which, after conver-
sion, is estimated to be $77.069 per month and per adult equivalent in PPP 2014 terms. Asterisks *, **, and *** indicate significance at the 10%, 5%, and 1%
levels, respectively.
37 Quantile regressions were estimated with the user-writtencommand -qreg2- which allows for standard errors that are ro-bust to intra-cluster correlation (Parente and Santos Silva 2016).These regressions also include the same control variables asthose in equation (1). We were unable to reliably estimate quan-tile regressions for the outcome number of nights that a child hasgone to bed hungry in the past week, as this variable does nothave a well-behaved density. We were also unable to estimatequantile regressions on savings per capita for some quantiles atsix months due to a large proportion of individuals with zerosavings.
1378 October 2017 Amer. J. Agr. Econ.
Downloaded from https://academic.oup.com/ajae/article-abstract/99/5/1362/4036204by George Mason University useron 11 July 2018
Tab
le8.
Th
eQ
uan
tile
Tre
atm
en
tE
ffect
so
fR
EA
P
(1)
(2)
(3)
(4)
(5)
(6)
Ou
tco
me
OL
SE
stim
ate
s10th
perc
en
tile
25th
perc
en
tile
50th
perc
en
tile
75th
perc
en
tile
90th
perc
en
tile
Pan
el
A:
Tre
atm
en
teff
ect
sat
six
mo
nth
s(a
tm
idli
ne)
Mo
nth
lyin
com
ep
erca
pit
a11.0
84***
3.1
01***
4.4
43***
7.3
06***
9.8
30***
13.4
82**
(2.6
39)
(0.7
62)
(1.3
54)
(1.7
76)
(2.9
60)
(5.3
33)
Mo
nth
lyex
pen
dit
ure
per
cap
ita
5.6
68
1.9
18*
1.6
18
0.5
84
�0.0
18
7.0
56
(3.9
33)
(1.0
51)
(1.2
73)
(1.9
55)
(3.4
86)
(9.4
00)
Mo
nth
lyco
nsu
mp
tio
np
erca
pit
a2.7
73
0.3
80
0.5
29
0.9
08
3.6
22*
6.4
03*
(2.3
35)
(0.9
16)
(1.0
60)
(1.0
99)
(2.1
35)
(3.8
40)
Savin
gs
per
cap
ita
1.0
77
––
–0.6
62
1.8
42
(0.7
14)
––
–(0
.637)
(1.2
54)
TL
Up
erca
pit
a�
0.0
11
0.0
02
0.0
16
0.0
13
�0.0
02
0.0
09
(0.0
44)
(0.0
25)
(0.0
36)
(0.0
41)
(0.0
63)
(0.0
68)
Du
rab
leass
etin
dex
0.5
26
�0.0
33
0.5
64*
0.6
55*
1.0
74*
0.5
55
(0.3
92)
(0.4
32)
(0.2
93)
(0.3
79)
(0.5
96)
(0.6
81)
Pan
el
B:T
reatm
en
teff
ect
sat
six
mo
nth
s(a
ten
dli
ne)
Mo
nth
lyin
com
ep
erca
pit
a8.7
44***
1.8
52**
2.3
72**
4.4
39***
5.8
77*
13.1
45*
(3.2
09)
(0.7
68)
(1.0
86)
(1.5
75)
(3.4
65)
(7.3
60)
Mo
nth
lyex
pen
dit
ure
per
cap
ita
2.4
46
1.5
95
�0.6
19
�1.5
36
�5.0
98
4.0
03
(3.9
63)
(1.4
77)
(2.0
98)
(3.1
30)
(3.9
89)
(10.0
86)
Mo
nth
lyco
nsu
mp
tio
np
erca
pit
a0.0
38
1.1
10
0.2
75
0.9
91
0.7
08
�3.5
89
(2.1
34)
(0.9
75)
(0.9
87)
(1.3
48)
(2.0
88)
(5.3
84)
Savin
gs
per
cap
ita
1.5
36**
––
0.3
19
1.2
44**
4.3
24***
(0.6
07)
––
(0.4
14)
(0.5
72)
(1.2
93)
TL
Up
erca
pit
a0.1
80*
0.0
55
0.1
12***
0.0
59
0.2
19*
0.5
05***
(0.0
99)
(0.0
35)
(0.0
42)
(0.0
64)
(0.1
17)
(0.1
55)
Du
rab
leass
etin
dex
0.7
51*
0.2
03
0.3
26
0.4
87
1.4
63**
1.2
64*
(0.4
01)
(0.3
37)
(0.2
99)
(0.4
13)
(0.6
76)
(0.7
03)
Pan
el
C:T
reatm
en
teff
ect
sat
on
eyear
(at
en
dli
ne)
Mo
nth
lyin
com
ep
erca
pit
a7.7
69***
2.7
85***
3.8
24***
6.4
81***
10.4
78***
12.3
61**
(2.0
67)
(0.7
51)
(0.9
91)
(1.3
67)
(2.6
21)
(5.8
59)
Mo
nth
lyex
pen
dit
ure
per
cap
ita
�1.5
65
0.8
73
�0.4
23
�1.3
27
�3.0
02
0.8
81
(3.4
44)
(1.3
97)
(1.9
13)
(2.7
79)
(3.6
45)
(11.0
22)
Mo
nth
lyco
nsu
mp
tio
np
erca
pit
a�
2.1
95
�0.3
28
�1.5
66
�2.2
46*
�1.6
41
2.0
94
(2.2
02)
(1.0
63)
(0.9
77)
(1.2
84)
(3.1
95)
(4.5
47)
Savin
gs
per
cap
ita
5.8
36***
2.4
78***
3.4
08***
4.9
28***
6.7
62***
8.0
25***
(0.7
65)
(0.3
36)
(0.4
35)
(0.5
95)
(1.1
17)
(1.1
74)
TL
Up
erca
pit
a0.1
58*
0.0
94***
0.1
34***
0.1
04*
0.1
29*
0.1
42
(0.0
86)
(0.0
34)
(0.0
40)
(0.0
56)
(0.0
78)
(0.2
79)
Du
rab
leass
etin
dex
0.9
94***
0.4
18
0.5
16*
0.8
81*
1.2
55*
1.9
95**
(0.3
72)
(0.3
06)
(0.2
99)
(0.4
59)
(0.6
43)
(0.7
87)
No
te:R
eg
ress
ion
sin
clu
de
con
tro
lsfo
rn
um
be
ro
fR
EA
Pb
usi
ne
sse
sin
am
an
ya
tta
,lo
cati
on
fix
ed
eff
ect
s,a
nd
ba
seli
ne
lev
els
of
the
ou
tco
me
va
ria
ble
(wit
hth
ee
xce
pti
on
of
mo
nth
lyco
nsu
mp
tio
np
erca
pit
afo
rw
hic
hb
ase
lin
ele
ve
lsa
ren
ot
av
ail
ab
le).
Ro
bu
stst
an
da
rde
rro
rs,c
lust
ere
da
tth
esu
b-l
oca
tio
nle
ve
l,a
resh
ow
nin
pa
ren
the
ses.
All
mo
ne
tary
va
lue
sa
rere
po
rte
din
20
14
US
D,P
PP
term
s.A
tsi
xm
on
ths,
we
are
un
ab
leto
est
ima
teq
ua
nti
letr
ea
tme
nt
eff
ect
so
fR
EA
Po
nsa
v-
ing
sp
erca
pit
afo
rso
me
qu
an
tile
sd
ue
toa
larg
ep
rop
ort
ion
of
ind
ivid
ua
lsw
ith
zero
sav
ing
s.A
ste
risk
s*
,*
*,
an
d*
**
ind
ica
tesi
gn
ifica
nce
at
the
10
%,5
%,
an
d1
%le
ve
ls,r
esp
ect
ive
ly.
Gobin, Santos, and Toth Promoting Entrepreneurship Through Cash Transfers 1379
Downloaded from https://academic.oup.com/ajae/article-abstract/99/5/1362/4036204by George Mason University useron 11 July 2018
graphically present the quantile regressionestimates for each of the 99 percentiles of thedistribution of outcomes, again distinguishingfor the duration of participation in the pro-gram (six months vs. one year) and the twoperiods of data collection. These results sug-gest several conclusions.
The first conclusion is that the effects on in-come are positive and statistically significantat each of the five quantiles reported in table8, and these effects are increasing with thequantile of the distribution. This is true forboth time periods and irrespective of thelength of participation in the program.Hence, the evidence strongly suggests thatREAP was particularly effective in terms ofincreases in income and in the short-to-medium-run, for those who were (relativelyspeaking) better-off, as predicted by ourmodel. The effect of the program estimatedat the 90th percentile is more than four timesthe effect at the 10th percentile. If the motiv-ation of the poverty graduation approach isto include the ultra-poor, we can then con-clude that this approach may take longer (orrequire modifications) for those who are atthe bottom of the distribution.
The second conclusion is that there are morepronounced effects among individuals in theupper quantiles of the other outcome distribu-tions. These patterns are clearly illustrated infigure J.1 in the supplementary material online.There, we see larger treatment effects for thosein the upper quantiles of the savings, livestock,and durable asset distributions, particularlywhen these effects are measured at endline.
The third conclusion is that the timing ofmeasurement of the program’s impact (mid-line vs. endline) seems to matter more interms of shaping the effect of the programthan the length of exposure to the program(six months vs. one year), likely reflecting thefactors discussed previously. The exceptionto this conclusion is clearly savings, for whichwe find evidence suggesting that the lack ofaccess to savings institutions (or lack ofawareness about their functioning) may haveprevented individuals from keeping liquidsavings. After the promotion of savingsgroups, we find significant treatment effectsacross the entire distribution, though theeffects are stronger for the upper quantiles.38
The similarity of the patterns exhibited intable 8 and figure J.1 could suggest that thoseindividuals with higher incomes (who gainmost from REAP in terms of income) wouldalso be the ones who would show highereffects of participating in the program interms of other outcomes, such as savings orconsumption of durable assets. To determineif this is true, we check whether individualsoccupy similar quantile positions in the con-ditional distribution of income and of otheroutcome variables. In table J.1 in appendix Jof the supplementary material online, we pre-sent the proportion of individuals who are inthe 90th percentile of different combinationsof outcome variables. The obvious conclusionis that for most pairs of outcome variables,less than 25% of individuals are in similarplaces in the distribution of outcomes. Thisresult suggests that beneficiaries may employdifferent strategies, with some choosing to in-vest more in productive assets such as live-stock, some opting for durable assets orliquid savings, and others choosing toconsume.
Such fundamental heterogeneity is reminis-cent of the distinction between subsistenceand transformative entrepreneurship(Schoar 2009), though we leave a deeper ana-lysis of these differences for future research.Nevertheless, we note that this conclusionseems to be reinforced by an analysis of theeffect of baseline heterogeneity on the effectof the program, where we interact the treat-ment indicator with baseline indicators ofcapital (see tables J.2, J.3 and J.4 in appendixJ of the supplementary material online). Atsix months (evaluated at endline) householdsat the higher end of the baseline livestock dis-tribution are saving less compared to thosewith no livestock at baseline, and at one year,those households who had some savings atbaseline own less livestock. Hence, our datasuggests that there may be some specializa-tion in terms of future activities, although wewould require further follow-up surveys tounderstand whether such a pattern of behav-ior persists.
Comparison of our findings to other studies.Finally, it is important to note that our esti-mates of the program’s impact are of a simi-lar order of magnitude to previous studies,namely Banerjee et al. (2015) and Bandieraet al. (2016). After one year, there is a 30.8%increase in income compared to the controlgroup, which is similar to the increases in in-come that can be estimated from the results
38 Note that before the introduction of savings groups, we onlyobserve significant effects on savings in the upper quantiles (75thand 90th at endline) of the savings distribution.
1380 October 2017 Amer. J. Agr. Econ.
Downloaded from https://academic.oup.com/ajae/article-abstract/99/5/1362/4036204by George Mason University useron 11 July 2018
presented in Banerjee et al. (2015), who findan average increase of 25.7% (22.8%) aftertwo years (three years), and Bandieraet al. (2016), who find a 38% increase in in-come after four years. Like our work, thesestudies do not find an important effect onconsumption: Bandiera et al. (2016) do notfind statistically significant impacts on con-sumption after two years, while Banerjeeet al. (2015) find a relatively small impact(around 5%).
The estimate of the program’s impact onsavings (131.4% increase) is also similar tothose Banerjee et al. (2015) estimated(155.5% increase after two years, and 95.7%increase after three years). Finally, we findthat REAP increases the probability of hav-ing an income above the poverty line by13.2%, a result similar to the 11% shift inwomen out of extreme poverty that Bandieraet al. (2016) estimated.
As with the other ultra-poor povertygraduation programs, our findings are moreconservative than those of Blattmanet al. (2016), which is the most similar toREAP. These differences may be attributableto both the post-war setting that they study(notably, the low levels of initial business ac-tivity and the much lower baseline income),and differences in the set-up of the program(namely, a significantly larger initial grantthan in the case of REAP).
Turning to the cost-benefit analysis of thisprogram, we estimate that the cost for oneadditional woman to be enrolled in REAP in2015 for two years was approximately $300(or PPP $713.91 at 2014 prices). This figure iswell below the direct costs of the six pro-grams that Banerjee et al. (2015) evaluated,as well as the program that Blattmanet al. (2016) evaluated. Assuming that theprogram implementation cost in 2015 was thesame as in 2013, and ignoring discounting andinflation, the gains in income (which we esti-mate to be the average of the one year andsix month impacts) would have to persist forone additional month to cover the cost of theprogram.
Conclusion
In this paper we study a multifaceted ap-proach to poverty alleviation that is gainingrecognition for its ability to set ultra-poorhouseholds on a sustainable pathway
out of extreme poverty (Banerjeeet al. 2015; Bandiera et al. 2016). We showthat a variation of the poverty graduation ap-proach, the Rural Entrepreneur AccessProject (REAP), which provides disadvan-taged women with capital, skills, and ongoingmentorship in enterprise and savings but thatexcludes consumption support, replaces assettransfers with cash transfers, and targets indi-viduals who are required to form groups ofthree, enables beneficiaries to run microen-terprises that lead to improved householdincomes. The short-to-medium-run impactsare economically significant and allowwomen to meet current household needs(through increased investment in durableassets) and plan for future shocks (throughthe accumulation of liquid savings). The path-way of change is quite clear, with a tightly-estimated shift of time use from leisure andhousehold activity into non-farm enterpriseactivity, with 95% of enterprises involved inpetty trade of consumer goods. Hence,REAP provides women with greater agencyover wealth transfers through cash, and theexact type of business to set up.Simultaneously, these factors do not appearto lead to misuse of funds or suboptimalchoices about the type of earning activity topursue, as designers of other graduation pro-grams have feared.
The estimates of REAP’s impact arelargely in line with other evaluations of simi-lar programs (Banerjee et al. 2015; Bandieraet al. 2016; Blattman et al. 2016). And al-though the existing data do not allow us toexamine the sustainability of the impactsonce participants stop receiving support, thesimilarity in results between our analysis andprior trials raises the plausible prospect thatthese impacts should be stable over time.39 Asimple cost-benefit analysis shows that, in thislikely eventuality, the program would covercosts within a reasonable time horizon(thirteen months).
We are also able to demonstrate the poten-tial applicability of this approach in a differ-ent, arguably more extreme context to thosealready studied. The REAP implementationoccurred in locations with very low averagepopulation densities, highly variable weather
39 Banerjee et al. (2015) examine two-year and three-yearimpacts and find no evidence of mean reversion of the impacts.Bandiera et al. (2016) look at two-year and four-year impactsand find more pronounced effects across many outcomes afterfour years compared to after two years.
Gobin, Santos, and Toth Promoting Entrepreneurship Through Cash Transfers 1381
Downloaded from https://academic.oup.com/ajae/article-abstract/99/5/1362/4036204by George Mason University useron 11 July 2018
conditions, low infrastructure, and limited ac-cess to markets, settings which have been ro-bustly shown to be prone to asset-basedpoverty traps. Yet, women were able to makeuse of the capital and skills that REAP deliv-ered to establish and run successful enter-prises, and obtain a sustainable livelihoodoutside of the livestock herding economy. Thisconsistency of results provides important sup-port for the robustness of the poverty gradu-ation approach, and further corroborates theexternal validity of other studies, while sug-gesting further opportunities for experimenta-tion in the design and implementation of suchprograms. For example, graduation programdesigners could consider experimenting withgroup-based approaches, transferring cash ra-ther than an asset (which can greatly reduceimplementation costs), or reducing costs byminimizing initial consumption support.
Supplementary Material
Supplementary material is available atAmerican Journal of Agricultural Economicsonline.
References
Allen, H. 2006. Village Savings and LoansAssociations: Sustainable and Cost-effective Rural Finance. Small EnterpriseDevelopment 17 (1): 61–8.
Anderson, M.L. 2008. Multiple Inference andGender Differences in the Effects ofEarly Intervention: A Reevaluation ofthe Abecedarian, Perry Preschool, andEarly Training Projects. Journal of theAmerican Statistical Association 103(484): 1481–95.
Bandiera, O., R. Burgess, N.C. Das, S.Gulesci, I. Rasul, and M. Sulaiman. 2016.Labor Markets and Poverty in VillageEconomies. Economic Organisation andPublic Policy Discussion Papers Series,LSE, STICERD, London, UK.
Banerjee, A.V., E. Duflo, N. Goldberg, D.Karlan, R. Osei, W. Pariente, J. Shapiro,B. Thuysbaert, and C. Udry. 2015. AMultifaceted Program Causes LastingProgress for the Very Poor: Evidencefrom Six Countries. Science 348 (6236):1260799.
Banerjee, A.V., D. Karlan, and J. Zinman.2015. Six Randomized Evaluations of
Microcredit: Introduction and FurtherSteps. American Economic Journal:Applied Economics 7 (1): 1–21.
Barrett, C.B., M.R. Carter, and M. Ikegami.2008. Poverty Traps and SocialProtection. SP Discussion Paper No.0804, World Bank, Washington DC.
Barrett, C.B., P. Marenya, J. McPeak, B.Minten, F. Murithi, W. Oluoch-Kosura,F. Place, J.C. Randrianarisoa, J.Rasambainarivo, and J. Wangila. 2006.Welfare Dynamics in Rural Kenya andMadagascar. Journal of DevelopmentStudies 42 (2): 248–77.
Barrett, C.B., and P. Santos. 2014. TheImpact of Changing Rainfall Variabilityon Resource-Dependent WealthDynamics. Ecological Economics 105:48–54.
Bauchet, J., J. Morduch, and S. Ravi. 2015.Failure vs. Displacement: Why anInnovative Anti-Poverty ProgramShowed No Net Impact in South India.Journal of Development Economics 116:1–16.
Benjamini, Y., and Y. Hochberg. 1995.Controlling the False Discovery Rate: APractical and Powerful Approach toMultiple Testing. Journal of the RoyalStatistical Society, Series B(Methodological) 57 (1): 289–300.
Bianchi, M., and M. Bobba. 2013. Liquidity,Risk, and Occupational Choices. Reviewof Economic Studies 80 (2): 491–511.
Blattman, C., E. Green, J. Jamison, C.Lehmann, and J. Annan. 2016. TheReturns to MicroenterpriseDevelopment among the Ultra-Poor: AField Experiment in Post-War Uganda.American Economic Journal: AppliedEconomics 8 (2): 35–64.
Bruhn, M. 2011. License to Sell: The Effectof Business Registration Reform onEntrepreneurial Activity in Mexico. TheReview of Economics and Statistics 93(1): 382–6.
Building Resources Across Communities.2013. An End in Sight for ExtremePoverty: Scaling Up BRAC’s GraduationModel for the Ultra-Poor. Briefing Note# 1: Ending Extreme Poverty. BRACUSA, New York, NY.
de Mel, S., D. McKenzie, and C. Woodruff.2008. Returns to Capital inMicroenterprises: Evidence from a FieldExperiment. Quarterly Journal ofEconomics 123 (4): 1329–72.
1382 October 2017 Amer. J. Agr. Econ.
Downloaded from https://academic.oup.com/ajae/article-abstract/99/5/1362/4036204by George Mason University useron 11 July 2018
Elliot, H., and B. Fowler. 2012. Markets andPoverty in Northern Kenya: Towards aFinancial Graduation Model. Workingpaper, Financial Sector Deepening,Nairobi, Kenya.
Gindling, T.H., and D. Newhouse. 2014. Self-Employment in the Developing World.World Development 56: 313–31.
Goldberg, N., and A. Salomon. 2011. UltraPoor Graduation Pilots: Spanning theGap between Charity and Microfinance.Commissioned workshop paper atGlobal Microcredit Summit, Valladolid,Spain, 14–17 November.
Haushofer, J., and J. Shapiro. 2016. TheShort-Term Impact of UnconditionalCash Transfers to the Poor:Experimental Evidence from Kenya.Quarterly Journal of Economics 131 (4):1973–2042.
Karlan, D., and J. Zinman. 2011. Microcreditin Theory and Practice: UsingRandomized Credit Scoring for ImpactEvaluation. Science 332 (6035): 1278–84.
Kenya National Bureau of Statistics. 2007.Basic Report on Well-Being: Based onKenya Integrated Household BudgetSurvey 2005/06. Kenya National Bureauof Statistics, Nairobi, Kenya.
Kenya National Bureau of Statistics andSociety for International Development.2013. Exploring Kenya’s Inequality:Pulling Apart or Pooling Together?Kenya National Bureau of Statistics,Nairobi, Kenya.
Kraay, A., and D. McKenzie. 2014. DoPoverty Traps Exist? Assessing theEvidence. Journal of EconomicPerspectives 28 (3): 127–48.
Little, P.D., J.G. McPeak, C.B. Barrett, andP. Kristjanson. 2008. ChallengingOrthodoxies: Understanding Poverty inPastoral Areas of East Africa.Development and Change 39 (4):587–611.
Lybbert, T.J., C.B. Barrett, S. Desta, andD.L. Coppock. 2004. Stochastic WealthDynamics and Risk Management amonga Poor Population. Economic Journal114 (498): 750–77.
Matin, I., M. Sulaiman, and M. Rabbani.2008. Crafting a Graduation Pathway forthe Ultra Poor: Lessons and Evidence
from a BRAC Programme. WorkingPaper No. 109, BRAC Research andEvaluation Division, Dhaka, Bangladesh.
McKenzie, D., and S. Puerto. 2017. GrowingMarkets Through Business Training forFemale Entrepreneurs: A Market-LevelRandomized Experiment in Kenya.Policy Research Working Paper No.7993, World Bank, Washington DC.
McPeak, J.G., and C.B. Barrett. 2001.Differential Risk Exposure andStochastic Poverty Traps among EastAfrican Pastoralists. American Journal ofAgricultural Economics 83 (3): 674–9.
Merttens, F., A. Hurrell, M. Marzi, R. Attah,M. Farhat, A. Kardan, and I. MacAuslan.2013. Kenya Hunger Safety NetProgramme Monitoring and EvaluationComponent: Impact Evaluation FinalReport 2009 to 2012. Oxford PolicyManagement, Oxford, United Kingdom.
Osterloh, S., and C.B. Barrett. 2007. TheUnfulfilled Promise of Microfinance inKenya: The KDA Experience. InDecentralisation and the SocialEconomics of Development: Lessonsfrom Kenya, ed. C.B. Barrett, A. Mude,and J. Omiti, 1–46. Wallingford, UK:CABI.
Parente, P.M., and J. Santos Silva. 2016.Quantile Regression with ClusteredData. Journal of Econometric Methods 5(1): 1–15.
Santos, P., and C.B. Barrett. 2011. PersistentPoverty and Informal Credit. Journal ofDevelopment Economics 96 (2): 337–47.
Schoar, A. 2009. The Divide betweenSubsistence and TransformationalEntrepreneurship. Unpublished manu-script. MIT, Cambridge, MA.
Silvestri, S., E. Bryan, C. Ringler, M.Herrero, and B. Okoba. 2012. ClimateChange Perception and Adaptation ofAgro-Pastoral Communities in Kenya.Regional Environmental Change 12 (4):791–802.
Toth, R. 2015. Traps and Thresholds inPastoralist Mobility. American Journal ofAgricultural Economics 97 (1): 315–32.
World Bank. 2012. World DevelopmentReport: Jobs. Washington DC: WorldBank.
Gobin, Santos, and Toth Promoting Entrepreneurship Through Cash Transfers 1383
Downloaded from https://academic.oup.com/ajae/article-abstract/99/5/1362/4036204by George Mason University useron 11 July 2018