aidcrowd-out: theeffectofngoson government ... · aidcrowd-out: theeffectofngoson...
TRANSCRIPT
Aid Crowd-Out: The Effect of NGOs on
Government-Provided Public Services∗
(Preliminary Draft – Do not cite or circulate
without the authors’ permission)
Erika Deserranno†, Aisha Nansamba‡, and Nancy Qian§
September 5, 2019
Abstract
We document that NGO entry in rural communities with a pre-existing gov-ernment health worker reduces the number of government workers, governmenthealth delivery, as well as total health service delivery (from the government orthe NGO). This is consistent with the NGO offering stronger monetary incentivesfor commercial activities, which induces some government workers to move to theNGO and allocate their time away from health service provision.
Keywords: Foreign Aid, Aid Effectiveness.JEL: O01, O02.
∗We thank Ameet Morjaria, Jakob Svensson, Edoardo Teso for his insights; and participants ofthe UCSD Development Workshop, University of Wisconsin Development Workshop, the CEPR Con-ference on Development, and the Development Conference for Chicago Area Economists for usefulcomments.†Northwestern Kellogg MEDS, [email protected]‡NGO Uganda§Northwestern Kellogg MEDS, BREAD, NBER, CEPR, [email protected]
1 Introduction
The efficacy of foreign aid is a central and one of the most controversial questions for
researchers and policy makers today. Well-known arguments have been made both for
and against the effectiveness of traditional aid (e.g., Burnside and Dollar, 2000; East-
erly, 2006). In response to the criticisms, particularly those regarding aid being tied to
the strategic objectives of donor countries (e.g., Alesina and Dollar, 2000; Kuziemko and
Werker, 2006), aid policy makers have given increasing support to non-government or-
ganizations (NGOs) (e.g., Buthe et al., 2012; Dreher et al., 2007; Nancy and Yontcheva,
2006).1 For example, Faye and Niehaus (2012) documents that donors use bilateral aid
to influence elections in developing countries, but that this strategic behavior is not
present for aid administered by NGOs. In the past twenty years, the number of NGOs
and overall aid from major donors such as USAID and the European bilateral agencies
channeled through NGOs have more than quadrupled (Pfeiffer et al., 2008). To re-
duce reliance on large donors, several of the largest NGOs have also created innovative
methods of generating revenues that we discuss later.
The ultimate goal of foreign aid is to create self-sustainable development in coun-
tries that are otherwise poor. One of the key challenges for many poor developing
nations that NGOs can help address is the deficit in public services relative to what the
population needs. NGOs fill this gap as the aid-recipient government builds capacity to
provide these services in the long-run. The success of NGOs in achieving this objective
is far from clear. Critics point out that NGOs are largely unregulated and, for the most
part, do not coordinate with each other or recipient-country governments.2 This could1Nancy and Yontcheva (2006) studies NGO aid allocation across countries and finds that it is
unrelated to donor strategic objectives and strongly associated with poverty in recipient countries. SeeButhe et al. (2012) for a discussion of the motivations of USAID that is administered by NGOs andDreher et al. (2007) for a discussion of the motivations of Swedish aid that is administered by NGOs.
2For a few examples, see Bromideh (2011) for a discussion of NGO performance in Iran, Rahman(2003) for a discussion of NGOs in South Asian countries and Pfeiffer et al. (2008) for a discussion ofNGOs in sub-Saharan Africa.
1
create inefficiencies by replicating services and even harming the development of public
service delivery by competing with the recipient government for resources and reduce
the government’s capacity to provide services and, in turn, increase the country’s future
reliance on aid. A second concern stems from the fact that to achieve financial indepen-
dence, some NGOs ask field workers to perform parallel tasks: deliver free services to
needy individuals and sell households products using an parallel-task business model.
The latter has the advantage of providing income to field workers and the NGO, and
the disadvantage that monetarily incentivized activities can crowd out service delivery.3
In the context of health services, concerns over the efficacy of NGOs have been
articulated by medical doctors and public health experts, who call for NGOs to “Limit
hiring of public systems”, “Limit pay inequity between the public and private sectors”
and “Commit to joint planning [with the recipient government]” (Pfeiffer et al., 2008).
Farmer (2008) stated that “The NGOs that fight for the right to health care by serving
the African poor directly frequently do so at the expense of the public sector. Their
efforts too often create a local brain drain by luring nurses, doctors, and other profes-
sionals from the public hospitals to “NGO land,” where salaries are better.” NGOs are
aware of the potential challenges. For example, in discussing its parallel-task model,
BRAC, the world’s largest NGO, states: “Clearly there is a potential programmatic
trade-off here between increasing her sales and monthly income, while still ensuring
that the preventative and health education aspects of the program are being sufficiently
addressed” (Reichenbach and Shimul, 2011).
This study aims to make progress on understanding how NGOs can hinder (or help)
the development of public services in poor countries with rigorous empirical evidence
from a highly relevant context. We focus on basic health services in rural Uganda, where
the population is poor, where access to modern medical care is limited, and where infant3We provide a more detailed discussion in the Background Section.
2
mortality rates are amongst the highest in the world.4 This is also a context where the
government has been attempting to expand public provision of basic health services,
and where an NGO with similar aims entered soon after the government program was
rolled out. These features make it an ideal context for our research question.
Around 2009 and 2010, government workers began to provide basic health services
and products free of charge to their local communities. Like many public-sector workers,
they were volunteers.5 Approximately one year later, a large NGO randomly rolled out
similar health services in the same region. At this time, nearly half of the villages that
received an NGO worker had a pre-existing government worker. NGO health workers
provide similar services, also free of charge, as the government. In addition, NGO
workers sell products to households. These include the medicines that government
workers give for free, as well as other household products such as soap, oil or gloves.
The profits from sales partly go towards paying NGO village health workers, and partly
go towards funding the NGO more generally.
The main analysis investigates two dimensions of crowd-out: government health
workers (labor) and government-provided services. Our baseline equation estimates
the interaction effect of the pre-existence of a government worker and the entry of
the NGO on the number of government health workers and on the delivery of health
services. Since NGO entry is randomly assigned, the two variables are independent and
the interaction can be interpreted as causal. Note that we could equivalently restrict
our attention to a subsample of villages with a pre-existing government program and
simply estimate the uninteracted effect of randomly assigned NGO entry. We prefer the
interaction specification because the larger sample of villages provides more statistical4Infant morality in our sample is 69 per 1,000 live births. In the 2011 Ugandan Demographic and
Health Survey (DHS), they are 66 per 1,000.5In both poor and rich countries, local public services are often delivered by volunteers. For example,
Khan et al. (2015) studies volunteer tax collectors in Pakistan. In the United States, election pollworkers are often volunteers. Most of the traditional health worker or agriculture extension programslack explicit monetary incentives (Bhutta et al., 2010; Gilmore and McAuliffe, 2013).
3
precision.
The results for labor show that the arrival of the NGO in villages with a government
worker reduces the number of government workers on average: by one health worker
in every other village (i.e., half of the villages). The findings are consistent with NGO
entry crowding out the infant public sector, and the convention wisdom that higher
pay from the NGO induces volunteer government workers to switch to working for the
government. In the other half of the villages (with a pre-existing government worker),
the NGO is able to recruit a new person, increasing the total number of health workers
in the village by one.
The results on service delivery show that households in villages that have pre-
existing government workers and then receive the NGO are approximately 25.9 percentage-
points less likely to see a government health worker and 12.7 percentage-points less
likely to see any health worker. The former shows that NGO entry crowds out government-
provided services and is consistent with the shift in labor from the government to the
NGO. The latter shows that NGO entry reduces total service delivery, which is consis-
tent with time allocation data which show that NGO workers spend most of their time
on commercial activities and very little time providing health services.6
The main results support the concern that NGO entry crowds out government-
provided health services, and shows that the labor supply of health workers is an im-
portant channel. To assess the importance of the observed crowding out on health
outcomes, we examine infant mortality as an outcome and are able to provide useful
suggestive evidence. We find that in villages with a pre-existing government worker,
NGO entry has a positive, albeit imprecise effect on infant mortality. However, when we
decompose the villages into those where the NGO could not recruit a new worker (i.e.,
the government worker shifted to work for the NGO, the total number of workers was6See the Background section for a discussion about time allocation.
4
unchanged) and those where the NGO recruited a new person (i.e., the total number
of workers increased from one to two), we find highly suggestive evidence that NGO
entry increased infant mortality in the former and had no effect in the latter.
Our study focuses on crowding out, and documents the adverse effects of NGO entry
in villages that had a government worker. However, it is important to note that we also
find that in villages where there was no pre-existing government, NGO entry increases
the supply of health workers from zero to one, increase health services, and reduces
infant mortality.
The main caveat for our identification strategy is the concern of external validity –
the presence of a government worker, which is non-random, could be correlated with
factors that cause her to prefer to work for the NGO. We address this by carefully
controlling for a large number of potentially important base year characteristics (e.g.,
proxies for village-level or household-level socio-economic status, access to other medical
assistance, mortality), each interacted with the introduction of the NGO. Our results
are very robust. We also explore and rule out some obvious alternative mechanisms
(e.g., changes in the prices of drugs, relative demand for health care and household
products). See the paper for details.
The results of our paper should be cautiously interpreted as specific to the context
of our study. However, the finding that NGOs can distort the local labor market
is generalizable. Our results have several policy implications and motivate several
questions for future research, which we discuss in the Conclusion.
Our study contributes to the large literature on foreign aid. Several recent studies
have provided rigorous empirical evidence on the heterogeneous effect of aid across
contexts (e.g., Crost et al., 2014; Nunn and Qian, 2014).7 In focusing on NGOs, we add
to the recent study of Faye and Niehaus (2012), which is referenced earlier. Our results7See Easterly (2009) and Qian (2015) for literature overviews.
5
complement this earlier work by pointing to coordination with recipient governments as
an important dimension for improving NGO-administered aid. Studies that argue for
better coordination typically focus on donor coordination (e.g., Bigsten and Tengstam,
2015). Recent studies have also provided theoretical evidence that higher wages from
foreign aid can distort local labor markets (e.g., Knack and Rahman, 2004; Koch and
Schulpen, 2018).
In documenting how private spending can crowd out government spending in a poor
country, we contribute to the public economics literature, which has mainly focused on
developed countries and on how government spending crowds out private spending
(e.g., Gruber and Hungerman, 2007; Kingma, 1989; Payne, 1998).8 The labor supply
mechanism we study supports the recent paper by Wagnerly et al. (2019), which finds
that community health workers in Uganda who are randomly instructed to sell health
products (rather than distribute them free of charge) provide health services to fewer
households. We are also closely connected to a study by Baldwin et al. (2019) which find
that a participatory development program launched by an NGO caused improvements
in leadership by pre-existing institutions, but decreased investments in local public
goods through these institutions. A large number of studies have produced mixed
findings when evaluating the effect of NGO provided health services on mortality.9 The
heterogeneous effects we estimate – that NGOs reduce mortality in villages with no
government worker complements this literature.
The paper is organized as follows. Section 2 describes the context. Section 3 de-
scribes the data and descriptive statistics. Section 4 presents the main results on labor
supply and health service delivery. Section 5 presents the results on mortality. Section8In two well-known earlier studies, Kingma (1989) finds that U.S. government contributions to
National Public Radio crowds out private donations, while Payne (1998) finds little evidence thatgovernment spending crowds out private donations to U.S. homeless shelters. More recently, Gruberand Hungerman (2007) finds that U.S. government spending crowded out church spending on socialservices during the Great Depression.
9See Scott et al. (2018) for a review of the evidence.
6
6 concludes.
2 Background
2.1 Government Health Service Provision
Rural Uganda is one of the poorest regions in the world, where average per capita
income today is $522. In the 2011 Ugandan DHS data, infant mortality rates in rural
Uganda were estimated at 111, 66 and 30 per 1,000 live births, respectively; among the
highest in the world DHS (2011). According to the World Health Organization, the
main contributors of under-5 mortality in this context are neonatal mortality (60%),
diarrhea (14%), malaria (9%), pneumonia (6%) and other causes (11%). For neonatal
mortality, the main causes are birth asphyxia/trauma (30%), premature births (28%),
sepsis (18%), congenital anomalies (12%), acute respiratory infections (6%) and other
causes (6%) (WHO, 2012).
To respond to the scarcity of public health services and the shortage of health
workers, the Ugandan government founded the Village Health Team program which
hires community health workers to provide health services to their own community.
While the program was founded in 2001, it was not implemented for many years in many
areas in Uganda because of the lack of funding. In the North and Central Ugandan
regions that we study, government workers were hired around mid-2009.10
The main job of government workers is to provide the following health services during
home visits to members of their community: (i) health education (e.g., benefits of a
hospital delivery, methods of disease prevention), (ii) post-natal check ups, (iii) basic
medical advice and referrals to health clinics that are usually located in more urban10See the “Ugandan Annual Health Sector Performance Report 2008/2009” and the “Village Health
Team, Strategy and Operational Guidelines” (MoH, 2010). A survey of 150 government workers inNorth Uganda indicate that 87% of them were hired between 2009 and 2010 (Kimbugwe et al., 2014).
7
areas. The latter include helping patients decide when it is useful to travel to urban
areas for medical attention, as well as coordinating the visit to ensure that professional
medical staff is on site during the visit. The latter is important because of the high level
of absenteeism in Ugandan public health facilities. For example, Nyamweya et al. (2017)
documents a 48% average rate of staff absenteeism in Ugandan public health facilities,
which creates excessively long lines, which for pregnant women, increases the risk of
having to deliver in the health facilities without assistance from a health professional.
Government workers also provide basic medicines, such as ACT (artemisinin com-
bination therapy for malaria), ORS (oral rehydration solution), Zinc, antibiotics, and
deworming tablets free of charge, as well as distribute free bed nets during national
malaria campaigns.
The government recruits workers locally. To be eligible for the community health
worker position, a person must satisfy three conditions: village residence, age above
eighteen, and be able to read and write. Among the eligible candidates, the government
prioritizes those who are good community mobilizers and communicators, trustworthy
and willing to work for the community, and who have experience in the health sector
(Turinawe et al., 2015).11 Once hired, government workers are given five days of basic
training, a uniform that makes them easily identifiable (e.g., a t-shirt with the official
logo), and free medical products to disperse to the community. They are not paid or
given any incentives and are therefore generally considered to be motivated by altruism
(e.g., Kasteng et al., 2015; Deserranno, 2019).
The government program was rolled out in all villages. It aimed to employ two work-
ers per village. However, in our study area, only 57% of the villages had a government11The exact selection procedure varies from one village to another. In some villages, the worker
is appointed by a government official. In some others, the worker is appointed by the communitymembers, either by the village chief (LC1) or through a popular votes (Turinawe et al., 2015; Kimbugweet al., 2014). In North Uganda, 78% of the hired government workers had prior experience working as avolunteer in the health sector as a Community Drug Distributor or a Condom Distributor (Kimbugweet al., 2014).
8
worker in 2010, one year after the government program was rolled out; and no village
had more than one worker. In the other villages, the government was either unable
to recruit or retain a health worker (i.e., the recruited worker had stopped delivering
health services by 2010). To understand the recruitment and staffing of government
worker, we conducted in-depth interviews with government workers in our study areas
at all levels of the bureaucracy. The presence of a government worker in 2010 is not
random. According to higher-level government officials as well as community health
workers, the key limitation is the labor supply of those who are both qualified and
willing to work as volunteer workers.
Each government worker is affiliated with a nearby health facility: she refills her
stock of health products, attends occasional meetings, and reports to the person “in-
charge” at the health facility. District-level health officials that we interviewed stated
that each health facility is responsible for keeping track of resignations of affiliated
community health workers and finding a replacement, but are severely under-staffed,
and in practice, neither keep track of community health workers nor replace those that
drop out of the program. In other words, there is no accurate record of government
workers at a given point in time. This creates an important difficulty for NGOs, when
the latter attempts to avoid hiring government workers. We discuss this more in the
next section.
2.2 The NGO
The NGO we study has the same aims and provides the same services as the government
program. It entered our study area in June of 2010 by rolling out its program in a sub-
sample of 66 randomly selected villages and is the only other source of local healthcare
services. At this time, the government program had already been in place for at least
six months.
9
NGO workers, like government ones, are recruited locally and tasked to provide
similar free basic health services to the community. They mainly differ from govern-
ment ones in that they also sell household products from which they receive piece-rate
(whereas government workers earn no income). As we discussed in the introduction,
the motivation of the “parallel-task” model is to provide financial sustainability for the
NGO and increase its independence from donors. This model is used by several of the
largest NGOs today, including Living Goods, Grameen, BRAC, VisionSpring, Solar-
Sister, HealthStore Foundation and HoneyCare Africa. NGO workers, like government
workers, are easily identifiable from wearing their NGO uniforms.
Specifically, NGO workers buy products from the NGO at a price that is slightly
above the wholesale price, and then sell to households at a retail price that is set by the
NGO to be equivalent to the market price in that location. The difference between the
wholesale price and the buying price for the health worker goes towards the revenues
of the NGO at large. The difference between the buying price and the retail price
constitutes the income of the health worker.
The NGO and government programs differ in two other ways. First, the NGO
provides more training: workers receive ten days of initial training (rather than five)
and attend a monthly one day refresher training session, where they receive further
training and discuss the gaps in coverage and the quality of care. This could ostensibly
result in the NGO providing higher quality care than the government. Second, the
government workers distribute drugs for free, while the NGO workers sell similar drugs
for a small fee. Past studies have noted that NGOs may provide higher quality products
(Bjorkman-Nyqvist et al., 2016, Forthcoming).
In our study area, the NGO was able to successfully recruit in all the villages it
entered, whereas the government was only able to do so in half of the villages. Since
the NGO looks for individuals with the same skills and follow the same hiring criteria
10
as the government, government workers who apply to work for the NGO are typically
more competitive than other applicants. Based on the interviews that we conducted
with NGO recruiters, the NGO attempts to avoid hiring government workers. But this
is difficult to implement because of the lack of information on government workers and
the incentives for applicants to hide their role as government worker. There are also
some cases where the NGO knows that the applicant works for the government, but
still hires her because she is the only able and willing candidate.
2.3 Time Allocation of Government vs. NGO workers
Both government workers and the NGO workers are part-time employees who typically
maintain other daily occupations such as farming or small shopkeeping. Government
workers work on average ten hours per week delivering health services, including the
dispersal of free drugs (Mays et al., 2017). To the best of our knowledge, there are no
disaggregated time allocation data for government health workers.
For NGO workers, we have such data from two sources. First, Deserranno (2019)
finds that NGO workers work approximately fourteen hours a week, half of which is
devoted to health services. This roughly implies that NGO workers provide 30% fewer
hours of health services than government workers. The second source is Reichenbach
and Shimul (2011), who interviewed 660 NGO workers. Table 1 shows that over one
month, NGO workers supply a total of 49 hours, where 37% is spent on providing
health services (which includes attending refresher training) and 63% is spent on selling
medicines or health commodities. Because the time allocated for refresher training
includes visits to the branch office to resupply the products they sell, 37% is the upper
bound of health-related activities. The lower bound can be obtained if we attribute the
time attending refresher training to market activities. When we do this, we find that
NGO workers spend 21% of their total effort on health related activities and 79% of
11
their time on market activities.
The time allocation data should be interpreted as merely suggestive and can under
or over-state the true time spent providing health services. On the one hand, we assume
that the time spent by the NGO worker selling similar drugs and other products is not
a part of health care provision, where as, for government workers, we assume that
distributing drugs is part of health services provision.12 On the other hand, the self-
reported NGO allocation of time to health services is probably an upper bound of
actual service delivery because patient visits are also used to sell products and monthly
training visits are also used to refill stocks.
Health services provided by government workers can be crowded in or crowded out
by the introduction of the NGO. On the one hand, the presence of the NGOs can com-
plement existing government-provided services if there are fixed costs in provision, such
as recruiting and training workers, or if there are positive externalities. For example,
increasing general awareness of the benefits of basic health services could increase the
supply of individuals who are willing to work to deliver these services and the demand
for these services for both the government and the NGO. This could, in turn, lead to
an increase in the efficacy of health care for both organizations – e.g., a mother who is
aware of the benefits of pre-natal check ups is more likely to take up an offered check
up from a government health workers as well as the NGO.
On the other hand, the NGO can crowd out health care delivery in several ways. The
NGO distorts the local labor market by providing the monetarily incentivized activity
of selling goods to health workers, which can crowd out more altruistically motivated
health service provision. NGOs are aware of this problem. For example, BRAC, the
world’s largest NGO, which uses the parallel-task model said in an evaluation of the
community health program in 2011, “There is a perception among the NGO staff that12We make this assumption because we lack disaggregated data for time allocation for government
health workers.
12
women are more commercial-minded and very much motivated by financial incentives
as opposed to non-financial incentives” (Reichenbach and Shimul, 2011).
The potentially negative effects of this distortion on the supply of health services can
be exacerbated by the fact that total compensation is much higher for the NGO than
the government such that government workers may switch to working for the NGO. If
the government is unable to recruit a replacement, then the introduction of the NGO
can reduce the total supply of health services unless if the NGO worker sufficiently
increases the total supply of hours on the extensive margin to offset the share of hours
spent away from health service delivery on the intensive margin.
The discussion so far focuses on time allocation that is unadjusted for quality. For
example, the NGO worker may be more efficient at identifying and examining patients,
such that she can provide more services within a given time frame than the government
worker. Similarly, if the drugs sold by the NGO (and are given for free by the govern-
ment) are higher quality, the effect of the increase in prices on health may be offset by
the increase in quality. We will return to this when we present the results. We will
discuss quality later when relevant for our main results.
3 Data
3.1 Data Sources
The main analysis uses survey data that cover 127 villages in twelve geographical areas
of North/Center Uganda.13 Data on mortality and access to health services were col-
lected in May, 2010, approximately six months after the government program was rolled
out and one month before the NGO program rolled out. Data were collected again in13The areas are: Kajjansi, Kisaasi, Abaita Ababiri, Ggaba, Kangulumira (Central Uganda) and
Kitgum, Padibe East, Orom, Palabek Kal, Patongo, Pader, Rackoko (North Uganda).
13
December, 2012. The paper will sometimes refer to 2010 as the base year and 2012 as
the end year. In each wave, there is a village-level and a household-level survey. The
former is answered by the village head. The latter is answered by a randomly selected
20% of the households. The survey questions change slightly over time. We will discuss
this when relevant. We supplement the survey data with census data collected during
February to April, 2010. These include variables such as household size, mortality and
occupation for all households in each village.14 We will discuss other data sources when
they become relevant.
3.2 Descriptive Statistics
Figure 1 documents the retail price (what households pay) and the profit margin for the
NGO worker for the products they sell. The medicines that distributed free of charge
by government workers (oral rehydration salts, pain reliever, zinc, antimalarials, cold
capsules, deworming tablets) are sold at very low retail prices and provide negligible
profits to the NGO agent. This suggests that the increase in the price of drugs is likely
to play a small role for understanding the tradeoffs of government and NGO workers.
The products that provide the highest profits to the NGO workers are, on average,
less related to the most concerning health outcomes: fortified oil, cotton, soap, fortified
flour and toothpaste. This means that the NGO worker is mostly monetarily incen-
tivized to sell products that have less of a direct impact on health than those provided
by the government worker.
The data provide several other interesting statistics (not presented in tables). Con-
sistent with our sample being rural and poor, 76% of the households surveyed in 2010
report that medical services are too expensive and 51% report that health facilities
are too far. Interestingly, 31% of households report staff absenteeism in public health14These census data were used by the NGO for sampling in the household-level survey.
14
facilities as another constraint to the access of health services.
In addition to the surveys discussed earlier, we have access to an internal survey
conducted by the NGO for its community health workers in January, 2012, eighteen
months after the NGO rolled out.15 Several interesting facts emerge from the self-
reported data from health workers (not presented in tables). First, 21% of all NGO
workers used to work for the government as health workers. None are from villages
with no pre-existing government program. In contrast, in villages with pre-existing
government programs, 40% had switched from the government to the NGO. Even more
strikingly, if we examine villages with a government worker in 2010, but no government
worker in 2012, 82% of the pre-existing government workers switched to the NGO
after the latter entered the community. This is consistent with the concern that NGOs
compete with the government for labor and workers are likely to move to the NGOs that
pays a higher income. Second, NGO workers report earning 14.2 USD per month, or
38% of the average monthly income in Uganda in 2013, which is a considerable income
for a part-time job. Finally, the fraction of households that has purchased commodities
(soap, oil, salt) from the NGO, 28%, is higher than that which has purchased medical
products (antimalarials, oral rehydration salt) from the NGO worker (9% and 3%,
respectively).
4 Results
4.1 Labor Supply
We begin our analysis by investigating whether the introduction of the NGO crowds
out government-provided public health services. The first outcome we examine is the15The survey includes NGO workers in 2012, and does not include any government workers in 2012.
It will include government workers from 2010 only if they switched to work for the NGO.
15
number of government health workers. To illustrate the empirical strategy and under-
lying variation, Table 2 first presents a simple four-by-four difference-in-differences grid
view of the raw data. We divide villages according to whether there was a government
worker before the NGO roll out, and whether the NGO randomly selected the village
to be part of its program. In each cell, we present the number of government, NGO
and total (the sum of the government and NGO) workers in 2012. Column (1) shows
that in villages with no government worker in 2010, there were no government workers
in 2012, regardless of NGO entry. Column (2) shows that in villages with a government
worker in 2010, there were on average 0.95 government workers in villages that were not
assigned to be part of the NGO program, and 0.53 government workers in villages that
were assigned to be part of the NGO program. Column (3) is the difference between
columns (1) and (2). The bottom row of column (3) is the difference-in-differences. It
shows that NGO entry reduced the number of government workers by -0.58, i.e., approx-
imately one per every two villages. For brevity, we do not discuss the other statistics
in the table, which have analogous interpretations to that of government workers.
This difference-in-differences variation can be characterized in the following regres-
sion:
yi = α + β(Govi ×NGOi) + δGovi + γNGOi + ΓXi + λa + εi. (1)
The number of health workers in village i in 2012, yi, is a function of: the interaction
of a dummy that takes a value of 1 if the village receives a government health worker
in 2009, Govi, and a dummy that takes a value of 1 if it is designated to participate in
the NGO program in 2010, NGOi; the uninteracted dummy variables; and area fixed
effects, λa.16
The main benefit of estimating this regression relative to examining the raw data16We include area fixed effects because the level of our main outcome variables (labor supply, mor-
tality, etc) varies substantially across areas and because the NGO random assignment was stratifiedby area.
16
is that it allows us to estimate the precision of the relationships and also to control for
potentially confounding factors. β is analogous to the difference-in-differences shown
in Table 2. It captures the differential effect of the NGO in villages with vs. without a
government worker. Govi is not randomly assigned. But NGOi is randomly assigned,
which means that the two variables are independent. Thus, the interaction term can
be interpreted as causal. We are mainly interested in β and β + γ. The latter term
is the effect of NGO entry in villages with a pre-existing government worker and can
be interpreted causally (see the following discussion for the intuition). If NGO entry
crowds out government health workers, then β + γ < 0.
Note that instead of estimating the interaction specification with all villages, we
could alternatively estimate the uninteracted effect of random NGO entry on whether
there is a government worker in 2012 using a sample of villages with a pre-existing
government program. Conceptually, this uninteracted NGO coefficient is similar to
β + γ from our baseline specification. It has a causal interpretation because NGO is
randomly assigned. The main advantages of our interaction specification relative to
the simpler alternative is that it maximizes statistical power by pooling the villages
with and without pre-existing government programs.17 At the same time, the earlier
differences-in-difference grid. Table 2 shows clearly that the alternative strategy would
produce similar results – column (2) shows that the number of government workers
declines by -0.58 after NGO entry in villages with a government program.
The main caveat for interpretating the our baseline results regards external validity
– there may be omitted factors which vary by the presence of the government program
which confound our estimates. We will discuss this in detail after we present the main
results.17Note that the uninteracted estimates using a subsample of villages with a government program is
statistically significant for examining labor supply and the delivery of health services. These resultsare available upon request. However, for the estimates on mortality, the pooled sample is required forstatistical precision.
17
Table 3 presents the estimates. In column (1), we first examine the number of NGO
workers as the outcome variable. The uninteracted coefficient for NGO show that NGO
entry into a village with no government workers increase the number of health workers
by almost exactly one worker (0.987). The estimate is statistically significant at the
1% level. The interaction coefficient is equal to zero, which means that the effect of
NGO entry on the number of NGO workers is the same in villages with and without a
pre-existing government worker. Note that these estimates are similar to the patterns
shown in Table 2, but the exact magnitudes will differ slight because the regression
controls for area fixed effects.
Column (2) tests crowding out and examines the number of government health
workers as the outcome. The uninteracted government coefficient is 0.791, which means
that the villages with the government program but no NGO had almost one government
worker in 2012. The interaction effect is -0.446, which means that NGO entry reduces
the number of government workers by approximately “half” a worker relative to villages
that had a pre-existing government worker. Both estimates are statistically significant
at the 1% level. The uninteracted NGO coefficient is small and statistically insignificant.
Thus, the sum of the uninteracted NGO coefficient and the interaction coefficient is -
0.466. This is shown at the bottom of the table in row [1]. The p-value in row [2] shows
that it is statistically significant at the 1% level. This is consistent with NGO entry
crowding out government health workers.
In column (3), we examine the total number of health workers from either the NGO
or the government. This variable can take the values of zero, one or two in our sample.
The uninteracted government and NGO coefficients indicate that each program, without
the presence of the other, results in almost one additional health worker relative to a
village with neither program.18 The interaction effect shows that villages with both18The fact that the estimates are less than one reflects attrition: 21% of government workers and
2% of NGO workers dropped out between the two waves of the survey, 2010 and 2012. This suggests
18
programs have on average half a worker less than villages with just one program. All
three coefficients are statistically significant at the 1% level. The sum of the interaction
coefficient and the uninteracted NGO coefficient is positive. The p-value at the bottom
of the table shows that the joint estimates is statistically significant at the 1% level. The
estimates in column (3) imply that NGO entry on average increases the total number
of health workers in all villages, but the increase is smaller by almost “half” a worker
in villages with a pre-existing government health worker.
Taken together, the estimates in columns (1)-(3) show that in villages with a govern-
ment worker, the introduction of the NGO has an effect on the intensive and extensive
margins of the labor supply of health workers. On the intensive margin, the NGO
induces government workers to shift to working for the NGO. Recall from the NGO in-
ternal survey discussed earlier that 82% of NGO worker in villages with a pre-existing
government program report as being former government workers. On the extensive
margin, the entry of the NGO increases the total number of health workers by approx-
imately half a person per village, or one person per every two villages. In other words,
when the NGO arrives at two villages with government workers, in one village, govern-
ment workers will work for the NGO, and in the second village, the NGO will hire an
additional worker while government workers will continue to work for the government.
To check our assumption that the government and NGO are the only providers
of modern health services – i.e., there are no other substitutes in villages, we also
examine the presence of a traditional healer and drugstore as outcomes. Traditional
healers provide non-Western medical treatments (e.g., herbalist, spiritual healer) and
drug stores sell health products and also provide basic medical advice. The worker in
the latter is typically also the owner. The mean dependent variables at the top of the
that the NGOs monetary incentives were successful not only in increasing the rate at which vacanciesare filled (as in Deserranno (2019) and Dal Bó et al. (2013)) but also in increasing the retention rates(as in Deserranno (2019)).
19
table show that there is approximately one traditional healer per every five villages,
and one drug store per every two villages. Because the skills of traditional healers and
drugstore owners are so different from what the NGO looks for in health workers, it
is unlikely that the arrival of the NGO will affect the presence of these other entities.
The regression estimates in columns (4) and (5) show that this is indeed the case.
4.2 Service Delivery
We use two proxies for the delivery of health services. The first is a survey question
for whether a household has sought medical advice from a community health worker in
the past year and the identity of the service provider. Table 4 presents the results. The
number of observations is much larger than the earlier analysis since this analysis uses
a household level survey. To address the possibility that the error terms are correlated
with villages, we cluster the standard errors at the village level for all household-level
regressions.
Column (1) examines whether households have obtained services from the NGO.
The estimates show that NGO entry increases the probability of getting health advice
from the NGO by 31.5 percentage-points. This estimate is statistically significant at
the 1% level. The interaction effect is small and insignificant. Thus, the effect of NGO
entry on services from the NGO does not vary depending on the pre-existence of the
government program.
Column (2) examines services from the government worker. The uninteracted gov-
ernment coefficient shows that having the government program with no NGO entry is
associated with 42.8 percentage-points higher probability of obtaining help from the
government worker. The interaction term shows that this probability declines by 24
percentage-points when the NGO enters. More importantly, the sum of the interaction
coefficient and the uninteracted NGO coefficient is negative, -0.259, and statistically
20
significant at the 1% level. This means that NGO entry reduces the probability that
households obtain medical advice from government health workers by 25.9 percentage-
points in villages that already had a government health worker relative to a village with
no government worker.
That the NGO crowds out government-provided services is not surprising after find-
ing that it reduces the number of government health workers. We next examine the
effect of NGO entry on total health service delivery (from any provider) in villages with
a pre-existing government worker. This is ‘ambiguous ex ante. On the one hand, the
increase in the total number of health workers on average (Table 3 column 3) could
increase service delivery. If NGO workers are better trained, they may also be more
efficient and be able to provide more services within a given time frame. On the other
hand, if NGO workers spend a large part of their time selling goods instead of pro-
viding health services, then the shift of workers out of the government program into
the NGO program could reduce service delivery. Our empirical estimates will capture
the net of these positive and negative forces on service delivery. Alternatively, one
can ask whether the increase in services from NGO workers (Table 4 column 1) offsets
the decline in service from government workers (Table 4 column 2) in villages with a
pre-existing government worker.
Table 4 column (3) shows that having a government worker or an NGO worker
increases the probability of receiving medical attention by 36.1 and 28.8 percentage-
points respectively. However, the interaction coefficient indicates that the introduction
of the NGO in a village with a government worker reduces the probability that a
household will receive medical attention by 41.5 percentage-points relative to a village
without a government worker. Moreover, the sum of the uninteracted NGO coefficient
and the interaction term is negative, -0.127, and statistically significant at the 1% level,
which means that in villages with a government health worker, NGO entry reduces the
21
probability that households obtain any health worker advice by 12.7 percentage-points.
This is strong evidence that NGO entry crowded out all services. It implies that
the reduction in time allocated to providing health services caused by selling products
is not sufficiently offset by the increase in the total number of health workers or the
possible improvement in the quality of NGO health worker.
Columns (4)-(7) show that we find no effect on health advice from other sources
in the village (traditional healers and drug stores), or sources in more urban areas
(government hospital or health centers, and private clinics that administer Western
medicine).
The second measure of service delivery we examine pertains to post-natal care. This
is motivated by the fact that the core mission of both government and NGO workers
is to reduce infant mortality by improving these services. The mortality data are only
available for households which have given birth during the interim between the two
waves of surveys (2010, 2012), for whom such care is most relevant. Since giving birth
is potentially an outcome of the provision of NGO and government health workers, we
first need to establish that fertility is not endogenous to the presence of government
and/or NGO workers.19 Table 5 column (1) examines the presence in the household of a
woman who delivered in the past year as the outcome variable. We find no relationship
between this proxy for fertility and our main right-hand-side variables. Therefore, in
columns (2)-(5), we will restrict our sample to households with at least one woman who
delivered in the past year. This results in the sample size decreasing from 2,747 to 407
households.
Columns (2)-(4) examine three variables which may be affected by government work-
ers and NGO workers: whether the woman delivered in a hospital, whether the delivery19We can alternatively estimate a multinomial logit model with the full sample. See Appendix Table
A.2. The results are similar.
22
was assisted by a health professional, and whether the mother received post-natal care
within two days of birth.20 We find no effect on having a hospital delivery, but statisti-
cally significant effects on whether the delivery was assisted by a health profession (e.g.,
hospital doctor or nurse). The comparison between the estimates for hospital delivery
and a hospital delivery supervised by a doctor or a nurse is consistent with the high
rate of absenteeism from hospitals and the fact that one of the tasks of the community
health workers is to coordinate for the delivery to take place when the nurse and doctor
is present.21
In column (4), we examine the probability of having had a follow up visit within two
days of birth for women who delivered in the past year, which is important for reducing
infant mortality and is used as a key measure of access to health services by the WHO
and UNICEF (Jones et al., 2003).22 The estimates show a negative and statistically
significant interaction effect of −0.216.
In column (5), we examine whether mothers breastfeeding. Ex ante, we do not
expect any effect because almost all mothers already breastfeed in this context. Our
results are consistent with expectations. However, we do find an effect on whether
the child is breastfed immediately after birth, a less common behavior which the WHO
recommends. The estimates show a negative and statistically significant interaction
effect of −0.188 (column 6).
Importantly, row [1] of columns (3)-(4) and (6) show the sum of the interaction
coefficient and the uninteracted NGO coefficient. These are all negative, which imply
that in villages with a government worker, the introduction of the NGO reduced the
probability that a child was delivered by a health professional, received a post-natal20These numbers are in line with data from the 2011 Uganda DHS, which indicate that among
women who delivered in rural Uganda, 53% of them were assisted by a health professional, and 30%received post-natal check-ups within two days of births.
21Recall the discussion from the Background section.22See WHO (2014, 2016).
23
visit within two days or was immediately breastfed. However, the joint statistic is not
statistically significant.
Interestingly, Table 5 also shows that the government health program on its own
is positively associated with post-natal services. Households located in villages with
a government health worker and no NGO have a substantially higher rate of being
delivered by a health professional (column 3) and receiving a post-natal visit within
two days of birth relative to villages with no health worker. These large positive effects
are consistent with two studies of the effect of the government health worker program
in rural Uganda (Ayiasi et al., 2016; Waiswa et al., 2015).23
Note that NGO and government health workers also provide pre-natal services.
However, the pre-natal service questions were only posed to women who were pregnant
at the time of the survey (276). Thus, we do not examine these outcomes.
4.3 Robustness
4.3.1 Correlates of Government Presence
The main concern for our preferred interpretation is one of external validity – villages
with government programs may differ from other villages in ways that interact with
NGO entry to affect the outcomes of interest. One way to investigate this concern is to
examine the correlation between characteristics of the village and households measured
in 2010, before the NGO rolled out, and the interaction term. In other words, we
estimate our baseline equation, equation (1), with these characteristics as outcome
variables. We focus on variables that reflect access to health services, the demand for
health services, and socio economic variables. This addresses the concern, for example,
that government health workers were more likely to be present in poorer villages, where23Ayiasi et al. (2016) document that the presence of a government health worker in a village increases
the likelihood that a woman receives three or more antenatal care visits by 19.7%. Waiswa et al. (2015)show similarly large effects on postnatal visits and institutional deliveries.
24
the health worker would be more eager to take up the opportunity for additional income
from the NGO.
Table 6 Panel A presents the estimates when the outcome variables are the presence
of a government health worker, traditional healer, drug store, government hospital
within ten kilometers, government health center within ten kilometers, and a private
clinic within ten kilometers. The interaction coefficient in column (6) is only significant
at conventional levels for the presence of a private clinic. It shows that, in 2010, villages
with both the government program and the NGO are further away from private clinics
than villages with neither program.
Panel B presents the estimates when the outcomes are the number of households in
the village, the number of children under five years of age per households, the number
of infants per household, the number of children under five years of age who died in the
past year, the age of the female household head, the share of households who conduct
non-agricultural business as their main activity, the share of household heads who
completed primary education, and two measures of wealth (assets and food security).24
The interaction coefficient in column (6) is only significant at conventional levels for the
number of households in the village and the share of household heads who completed
primary education. It shows that, in 2010, villages with only the government program
has fewer households and fewer educated household heads than villages with neither
program, while villages with both programs have a similar number of households and
educated household heads as villages with neither program.
We examine whether our results are robust to controlling for the presence of a private
health clinic and the number of households per village, as well as their interaction with
NGO entry in the next section.25
24Assets is measured as the number of assets owned by a household, out of a list of 11 commonhousehold assets. Food security takes value 1 to 4, where 1 indicates “deficit of food the whole year,”2 indicates “occasional deficit,” 3 indicates “neither deficit nor surplus” and 4 indicates “food surplus.”
25We also examine other potential correlates in a similar manner. None are statistically significantly
25
4.3.2 Additional Controls
To alleviate the identification concern raised in the previous section and to be conser-
vative, we will control for a large number of potentially confounding factors in addition
to those which the previous section shows are positively associated with the interaction
term. We categorize the additional controls into four groups. The first group comprises
of variables that can influence the effectiveness of a community health worker: distance
to nearest government hospital, distance to the nearest government health center, dis-
tance to the nearest private clinic, the presence of a drug store and the presence of a
traditional healer. Access to these resources could help or hinder the effectiveness of
the NGO. For example, on the one hand, proximity to a government clinics could allow
referral made by the NGO to be more effective (i.e., the patients are more likely to go
to a nearby clinic than a far away one). On the other hand, the presence of a traditional
healer may be correlated with community members not believing in Western medicine,
which lowers take up of the NGO’s services. Note that these factors can influence the
effects on service delivery and mortality, but it is hard to see how they would affect the
labor supply of health workers.
Second, we consider demographic variables: the number of households in the com-
munity, the number of children under five per household and the number of infants
per household. These variables could affect the labor supply of and demand for health
workers. For example, it may be easier to find additional health workers in larger vil-
lages. Alternatively, the demand for services may be larger in more populous villages
or villages with more young children.
Third, we consider base year morbidity: the number of children under five who died
in the past year and the number of children under one who died in the past year. The
variables could be correlated to the demand for health workers, and reflect underlying
correlated with the interaction term. These are not reported and available upon request.
26
well-being (e.g., income) of the population. For example, a village with higher child
mortality rate may face a higher disease burden (e.g., in places with more rainfall
and mosquitoes), where it is more difficult to reduce mortality. Or, a village with a
higher mortality rate may be poorer, where there is a lower supply of community health
workers or where government workers is more motivated to work for higher pay with
the NGO.
Finally, we consider household-level demographic variables: the age of the female
household head and the percent of households involved in farming as the main activity,
the share of household heads who completed primary education, assets and food secu-
rity. These factors may affect the availability of health workers for government workers
and NGO. The age of the female household head may also affect beliefs in Western
medicine.
All variables are measured in the base year of 2010. When we include controls for
the interactions of the NGO indicator variable, the uninteracted NGO coefficient is no
longer meaningful. Thus, Table 7 presents the estimates evaluated at the mean values
of all of the controls. They show that the main results are very robust and are unlikely
to be driven omitted variables.
4.3.3 Non-Linear Estimation
The main results are estimated with a Linear Probability Model. We can alternatively
use a Logistic Model for the dummy variable outcomes of the presence of health workers
or whether a household received pre-natal care, and a Poisson model to estimate the
effect on the total number of workers. The estimates are shown in Appendix Tables
A.1. Similarly, one may be concerned that for the examination of post-natal care in
Table 5, we restrict the sample based on a potential outcome. Although we document
that this not the case, we also alternatively using a Multinomial Logistics Model. See
27
Appendix Table A.2. All of the non-linear estimates are consistent with those from the
main regressions.
4.4 Alternative Mechanisms
We interpret the main results on labor supply and the reduction of total health service
delivery as being consistent with NGO entry crowding out government workers because
it provides better pay and more incentives to workers to allocate time away from pro-
viding health services. The most obvious alternative explanation is one of demand: the
demand for health services relative to the products sold by the NGO is lower in villages
with a government program. The robustness exercises discussed earlier addresses this
by controlling for variables that are likely to be correlated with the demand for health
services, such as child mortality rates and access to health services in the base year.
Here, we address this concern by dividing the sample into villages that had a drug store
in 2010 with those that did not. Drug stores sell many of the same products as the
NGO. Thus, if the main results are driven by differential demand for goods (relative to
health services), then we should find a smaller effect in villages that had a drug store.
Appendix Table A.3 presents the results. We examine the total number of health
workers and total health care delivery as outcome variables, and estimate the triple
interaction effect of the presence of the government program in 2010, NGO entry and
the presence of a drug store in 2010. This specification controls for the full set of
double interactions, uninteracted coefficients and area fixed effects. We find that the
triple interaction is statistically zero, which means that the presence of a drug store
does not affect our main results. This goes against differential demand as an alternative
hypothesis.
28
5 Mortality
To understand whether the crowding-out has real-world consequences, we examine
under-one and under-five child mortality, the reduction of which is a focal point for
both government workers and the NGO workers. Ex ante, the effect of the shift to the
NGO on community health outcomes is ambiguous. On the one hand, the reduction
in labor supply and service delivery could increase mortality.26 On the other hand,
the NGO may provide higher quality care, which can offset the negative effects. Our
empirical strategy estimates the net effect of the reduction in supply and any change
in the quality.
Table 8 columns (1)-(3) examine infant mortality using different measures. Column
(1) uses the household-level survey to examine a dummy that equals one if at least
one infant died within the household since 2010. The results in column (1) show that
the presence of government workers in a village with no NGO reduces the probability
of infant mortality by 2.6 percentage-points, relative to villages with no health worker
of any type.27 The introduction of the NGO in a village with no government worker
reduces mortality by 2.3 percentage-points. In contrast, the introduction of the NGO
in a village with a government worker does not reduce mortality. If anything, the sum
of the uninteracted NGO coefficient and the interaction coefficient implies that in those
villages the NGO increases mortality by 0.6 percentage-points (−0.023+0.029 = 0.006).
However, the p-value at the bottom of the table shows that the joint estimate is not
statistically significant. The lack of precision is not altogether surprising given that
mortality is an extreme and relatively rare event. We conservatively interpret the26Appendix Table A.4 documents the negative correlation between receiving health services provided
by government and NGO health workers and infant mortality, as well as the lack of a relationshipbetween receiving services from other health care providers and infant mortality.
27In our sample, the mean in villages with neither program is 4.6 percentage-points. Thus, theestimated reduction in mortality is large in magnitude and in line with Brenner et al. (2011) whichdocument a 53.2% decline in mortality as a result of the same government program in an experimentthat takes place in 116 villages of Southwest Uganda.
29
results as showing that the NGO, at best, brings no valued added to villages with a
government worker in terms of reducing infant mortality.
We observe a similar pattern in columns (2) and (3), when we normalize the outcome
variable by the number of births. In column (2), we calculate the household-level infant
mortality ratio as the number of children who were born between the two waves of
surveys and who died below age of one during this period divided by the number
of children who were born during this period.28 In column (3), we follow a similar
approach as column (2) but aggregate all numbers at the village level rather than
at the household level and then multiply the denominator by 1,000. One caveat of
the approach in columns (2) and (3) is that any child born before 2012 but who dies
before the age of one after 2012 increases the infant mortality ratio denominator by one
without increasing the nominator, thus causing us to underestimate infant mortality.
To address this problem, in columns (4) and (5) we exclude from the analysis any child
who is less than one year old in 2012, in addition to any child who would have been
less than one in 2012 had he/she not died.29
As with the estimates in column (1), we always find that the introduction of the
NGO increases mortality for villages that had a pre-existing government program (the
joint coefficient at the bottom of the table is positive), but the joint coefficient is not
statistically significant.
Appendix Table A.5 includes additional controls similar to Table 7. The results are28The latter is measured as the sum of: (a) the number of children who were born during this period
and who died below age of one during this period, and (b) the number of children who were bornduring this period and who did not die during the trial period (i.e., children less than three years atin 2012).
29This further reduces the sample size and exacerbates the sample selection problem. An alternativemethod to estimating infant mortality rate without reducing the sample size would be to express theinfant-mortality ratio in terms of the “infant-years of exposure to the risk of dying under the age ofone” (e.g., a child who is three months old in 2012 is exposed 1/4 [3/4] of a year while a child bornafter 2010 and who is more than one year old in 2012 is exposed a full year). This approach, which isused in Bjorkman-Nyqvist et al. (Forthcoming), cannot be used in this paper as we only observe childage in completed years but not in months. This data limitation also restricts our analysis to studyinginfant (<1 year) mortality but not neonatal (<1 month) mortality.
30
very robust.
5.1 Decomposition
From the previous section, we see that although each regression coefficient is large in
magnitude and statistically significant, the joint coefficients are small in magnitude and
imprecise. As we discussed earlier, this is consistent with the fact that mortality is a
relatively rare event. It may also be due to the fact that the baseline specification does
not distinguish between villages where the government worker switched to working for
the NGO (i.e., one total worker in 2012) and where the NGO recruited an additional
worker (i.e., two total workers in 2012). Recall from Section 4.1 that each type of villages
comprise approximately half of the villages that the NGO entered. If the increase in
mortality is due to a decline in health services, then we would expect the decline in
health services and increase in mortality to be more prominent in villages where the
government worker switched to working for the NGO.
We examine this by decomposing our estimates according to whether there was one
or two health workers in 2012. Since there is never more than one government worker
and never more than one NGO worker, this is very similar to decomposing according to
whether the government worker switched to working for the NGO or if the NGO was
able to hire a new person.30 Also note that in villages with the both programs and only
one worker in 2012, the worker is always an NGO worker. Since the number of workers
is an outcome variable, this exercise should be cautiously interpreted as suggestive.
Table 9 presents the results. We estimate an equation similar to the baseline, except
that we replace the interaction of the presence of the government program in 2010 and
NGO entry with two triple interactions: the government program, NGO entry and30In villages with both programs in 2012, there are very few situations with one worker who is not
the original government worker (i.e., there are very few cases where the government resigns as a healthworker and the NGO hires a new person).
31
the presence of one worker in 2012; and the government program, NGO entry and the
presence of two workers in 2012. Columns (1) - (3) examine health service delivery by
the NGO worker, government worker, or any worker. Column (1) shows that the effect
of NGO entry on medical advice from NGO workers does not vary depending on the
pre-existence of the government worker or the number of total health workers in 2012.
Column (2) shows that the reduction in services from the government health worker is
driven by villages where there is only one health worker in 2012 – i.e., the NGO was
unable to recruit a new person. The triple interaction with one worker is -0.506 and
statistically significant at the 1% level, while the triple interaction with two workers is
small and insignificant.
In column (3), the triple interaction with one worker is -0.545 and the triple interac-
tion with two workers -0.275. Both are statistically significant at the 1% level. At the
bottom of the table, we present the sum of the uninteracted NGO coefficient and the
triple interaction with one worker (-0.255), the sum of the uninteracted NGO coefficient
and the triple interaction with two workers (0.015), and their p-values. The first joint
estimate is statistically significant at the 1% level. The latter is insignificant. This
implies that in villages where there was a government program, NGO entry reduced
the total service delivery if it could not recruit an additional worker, and had no effect
if it successfully recruited an additional worker.
Columns (4)-(6) examine infant mortality outcomes. A comparison of the two triple
interaction terms show that the increase in infant mortality is more prominent in vil-
lages with a government program and where the NGO could not recruit a new worker.
The triple interaction with one worker is positive and statistically significant and the
triple interaction with two workers small and insignificant for all mortality outcomes.
Moreover, the joint statistics at the bottom of the table show that the sum of the un-
interacted NGO coefficient and the triple interaction with one worker is positive and
32
statistically significant at the 10% level for all mortality outcomes. This implies that
the increase in mortality in villages with a pre-existing government program is driven
by villages where the NGO could not recruit a new worker.
Taken together, these results are consistent with our interpretation that in village
with a government worker, NGO entry produced heterogeneous effects depending on
their ability to recruit a new worker. In places where they successfully recruited a
worker, causing there to be two health workers by 2012, there was no change in total
health service delivery or mortality. In places where they could not and employed
the government worker, total service delivery declined and infant mortality increased.
Recall that this decomposition exercise should be interpreted as suggestive rather than
causal.
5.2 Drug Prices
Since the government community health worker gives out the drugs for free, while the
NGO sells the same drugs at a low price, the crowding out of the government worker
can also increase drug prices, and through this potentially increase infant mortality
if households cannot afford this higher price. We examine drug prices for the most
relevant diseases as dependent variables and find no evidence to support this alternative
mechanism (see Appendix Table A.6).
5.3 Other Health Outcomes
In our investigation of the possible health outcomes of the crowding out of labor and
health services, we focus on infant mortality, which is the focal point of both the NGO
and government programs. However, health workers also provide advice and medicines
for many other health conditions (e.g., malaria, basic hygiene, etc.). We investigate
whether NGO workers in village with government program improved these other out-
33
comes. We examine a large number of variables that capture better preventive health
behavior (e.g., children washing their hands, sleeping under bednets, couples using con-
traceptives) and disease incidence (cough, diarrhea, worms, tuberculosis, malaria) as
the outcome in our baseline specification. Appendix Table A.7 shows that we find no
effect.
The main interventions that reduce infant mortality (education, pre- and post-natal
checkups, coordination with urban health facilities) are very time consuming. In con-
trast, there are other activities that would benefit the health of community members
which are less time consuming and perhaps more complementary to selling goods (e.g.,
advising couples to use/buy condoms). The results in Appendix Table A.7 show that
the NGO entry does not have any impact on these other activities, whether the village
has a pre-existing government worker or not.
6 Conclusion
This paper presents novel empirical evidence that NGOs can crowd out government-
provided community health services and reduce total health services. When the NGO
arrives to villages with volunteer government health workers, government workers shift
to work for the NGO in half of the cases. In the other half of the cases, the NGO is
able to recruit a new person. In the former scenario, the NGO entry reduces the total
supply of health services and the evidence is highly suggestive that it increases infant
mortality. It is important to keep in mind that our results do not say that NGOs are bad
for developing countries. Recall that we show that in villages without any government
health worker, NGO entry increases service delivery and reduces infant mortality. Taken
together, the empirical findings portray a nuanced picture of heterogeneous effects:
NGO entry provides benefits in places with no government-provided services, but can
34
hinder its own goals in places where the government services have already begun to be
established.
The generalizable insight that NGO entry can distort the local labor market (which
may, in turn, adversely affect the local population) is a concern that has been voiced
by researchers in public health and aid workers in the field for many years.31 We are
the first to provide rigorous evidence.
The empirical estimates should be cautiously interpreted as specific to the context
of this study. The two key features that are particularly important to keep in mind are
the novelty of the government program and the limited supply of able and willing health
workers. One could imagine that the entry of the NGO into a location with a mature
program may attract fewer workers from the government because there is more worker
loyalty. Time may also allow the government to build up the supply of potential health
workers (e.g., because there is a better understanding about the importance of the job,
and more people have knowledge about good health practices). For understanding the
magnitude of the findings, it is important to keep in mind that government workers are
all volunteers and do not receive any pay. It is very likely that if the income gap between
government and NGO workers is smaller, the magnitude of crowding out would also be
smaller. At the same time, note that the context that we study is very relevant for policy
makers. Hindering the development of public services in their infancy can impede the
process towards sustainable development and the practice of using volunteers to provide
local public services is common to developing (and developed) countries (e.g., Cowley
and Smith, 2014; Finan et al., 2015; Banuri and Keefer, 2016; Deserranno, 2019).
For understanding the external validity of our results, note that the dual-task model
is popular amongst NGOs today. It is mostly used in the context of community based
agents: e.g., community health workers provide health services and sell health products,31See the Introduction for references.
35
or agricultural extension workers train farmers and sell fertilizer and seeds. The NGO
incentivizes sales and not services because the former is easier to monitor. As we
discussed in the Introduction, this business model provides revenues that help reduce
the reliance of NGOs on large donors. Our findings illustrates the challenges this
business model poses for the NGO.
For policymakers, our results have several implications. The clearest one is that
aid-recipient governments and NGOs should coordinate their efforts. Towards this end,
the recipient government should improve record keeping of the location of government-
provided workers and share these data with NGOs. NGOs should experiment with
alternative compensation schemes to minimize the crowding out of altruistic activities
by money-making activities. Some NGOs that use the parallel-task model has recently
started to do this. For example, BRAC has begun to provide monetary incentives for
health services.
This study suggests several avenues for future research. First, problems highlighted
by our results are directly related to tradeoffs illustrated in the seminal model of multi-
tasking by Holmstrom and Milgrom (1991). Our findings suggest that NGO-delivered
aid is a natural context to test this model. Second, future studies should explore whether
aid crowds out or crowd in the development of native capacity in other contexts. For
example, policymakers and aid workers have expressed much concern that food aid,
if not thoughtfully targeted, can reduce farmgate food prices and negatively affect
farmers in recipient countries (Janzen, 2015; Levinsohn and McMillan, 2005). Finally,
the results emphasize the importance of heterogeneous effects for understanding the
effect of foreign aid and NGO effectiveness. Specifically, our findings suggest that
understanding the determinants of local supply is an important question for future
study.
36
ReferencesAlesina, Alberto and David Dollar, “Who Gives Aid to Whom and Why?,” Journalof Economic Growth, 2000, 5 (1), 33–63.
Ayiasi, Richard Mangwi, Patrick Kolsteren, Vincent Batwala, Bart Criel,and Christopher Garimoi Orach, “Effect of village health team home visitsand mobile phone consultations on maternal and newborn care practices in Masindiand Kiryandongo, Uganda: a community-intervention trial,” PloS one, 2016, 11 (4),e0153051.
Baldwin, Kate, Dean Karlan, Christopher Udry, and Ernest Appiah, “Par-ticipatory Development and Institutional Crowd Out,” 2019.
Banuri, Sheheryar and Philip Keefer, “Pro-social motivation, effort and the callto public service,” European Economic Review, 2016, 83, 139–164.
Bhutta, Zulfiqar A, Zohra S Lassi, George Pariyo, and Luis Huicho, “Globalexperience of community health workers for delivery of health related millenniumdevelopment goals: a systematic review, country case studies, and recommendationsfor integration into national health systems,” Global health workforce Alliance, 2010,1 (249), 61.
Bigsten, Arne and Sven Tengstam, “International Coordination and the Effective-ness of Aid,” World Development, 2015, 69, 75 – 85. Aid Policy and the Macroeco-nomic Management of Aid.
Bjorkman-Nyqvist, Martina, Andrea Guarison, Jakob Svensson, and DavidYanagizawa-Drott, “Can good products drive out bad? evidence from local marketsfor (fake?) antimalarial medicine in Uganda,” 2016.
, Jakob Svensson, and David Yanagizawa-Drott, “Reducing Child Mortality inthe Last Mile: Experimental Evidence on Community Health Promoters in Uganda,”American Economic Journal: Applied Economics, Forthcoming.
Bó, Ernesto Dal, Frederico Finan, and Martín A Rossi, “Strengthening state ca-pabilities: The role of financial incentives in the call to public service,” The QuarterlyJournal of Economics, 2013, 128 (3), 1169–1218.
Brenner, Jennifer L, Jerome Kabakyenga, Teddy Kyomuhangi, Kathryn AWotton, Carolyn Pim, Moses Ntaro, Fred Norman Bagenda,Ndaruhutse Ruzazaaza Gad, John Godel, James Kayizzi et al., “Can volun-teer community health workers decrease child morbidity and mortality in southwest-ern Uganda? An impact evaluation,” PloS one, 2011, 6 (12), e27997.
Bromideh, Ali Akbar, “The widespread challenges of NGOs in developing countries:Case studies from Iran,” International NGO Journal, 2011, 6 (9), 197–202.
37
Burnside, Craig and David Dollar, “Aid, Policies, and Growth,” American Eco-nomic Review, 2000, 90 (4), 847–868.
Buthe, Tim, Solomon Major, and Andre de Mello e Souza, “The Politics ofPrivate Foreign Aid: Humanitarian Principles, Economic Development Objectives,and Organizational Interests in NGO Private Aid Allocation,” International Organi-zation, 2012, 66 (4), 571 – 607.
Cowley, Edd and Sarah Smith, “Motivation and mission in the public sector: evi-dence from the World Values Survey,” Theory and decision, 2014, 76 (2), 241–263.
Crost, Benjamin, Joseph Felter, and Patrick Johnston, “Aid under Fire: De-velopment Projects and Civil Conflict,” American Economic Review, June 2014, 104(6), 1833–56.
Deserranno, Erika, “Financial incentives as signals: experimental evidence from therecruitment of village promoters in Uganda,” American Economic Journal: AppliedEconomics, 2019, 11 (1), 277–317.
DHS, “Uganda Demographic and Health Survey 2011, Uganda Bureau of Statistic(UBOS),” Technical Report 2011.
Dreher, Axel, Florian Molders, and Peter Nunnenkamp, “Are NGOs the BetterDonors? A Case Study of Aid Allocation for Sweden,” Kiel Working Paper 1383, KielInstitute for the World Economy 2007.
Easterly, William, The White Man’s Burden: Why the West’s Efforts to Aid the RestHave Done So Much Ill and So Little Good, New York: Penguin Press, 2006.
, “Can the West Save Africa,” Journal of Economic Literature, 2009, 47, 373–447.
Farmer, Paul, “Challenging orthodoxies: the road ahead for health and human rights,”Health and Human Rights, 2008, pp. 5–19.
Faye, Michael and Paul Niehaus, “Political Aid Cycles,” American Economic Re-view, December 2012, 102 (7), 3516–30.
Finan, Frederico, Benjamin A Olken, and Rohini Pande, “The personnel eco-nomics of the state,” Technical Report, National Bureau of Economic Research 2015.
Gilmore, Brynne and Eilish McAuliffe, “Effectiveness of community health workersdelivering preventive interventions for maternal and child health in low-and middle-income countries: a systematic review,” BMC public health, 2013, 13 (1), 847.
Gruber, Jonathan and Daniel M. Hungerman, “Faith-based charity and crowd-out during the great depression,” Journal of Public Economics, June 2007, 91 (5-6),1043–1069.
38
Holmstrom, Bengt and Paul Milgrom, “Multitask principal-agent analyses: In-centive contracts, asset ownership, and job design,” JL Econ. & Org., 1991, 7, 24.
Janzen, Joseph P., “The Effect of U.S. International Food Aid on Prices for Dry Peasand Lentils,” IOAE, 2015.
Jones, Gareth, Richard W Steketee, Robert E Black, Zulfiqar A Bhutta,Saul S Morris, Bellagio Child Survival Study Group et al., “How many childdeaths can we prevent this year?,” The lancet, 2003, 362 (9377), 65–71.
Kasteng, Frida, Stella Settumba, Karin Kallander, Anna Vassall, and in-SCALE Study Group, “Valuing the work of unpaid community health workersand exploring the incentives to volunteering in rural Africa,” Health policy and plan-ning, 2015, 31 (2), 205–216.
Khan, Adnan Q, Asim I Khwaja, and Benjamin A Olken, “Tax farming redux:Experimental evidence on performance pay for tax collectors,” The Quarterly Journalof Economics, 2015, 131 (1), 219–271.
Kimbugwe, Geofrey, Maghanga Mshilla, Denis Oluka, Olivia Nalikka,Joseph Kyangwa, Stella Zalwango, Uthuman Kilizza, Munanura Turyasi-ima, Louis Ntambazi, Fred Walugembe et al., “Challenges faced by villagehealth teams (VHTs) in Amuru, Gulu and Pader districts in Northern Uganda,”Open journal of preventive medicine, 2014, 4 (9), 740.
Kingma, Bruce Robert, “An Accurate Measurement of the Crowd-out Effect, IncomeEffect, and Price Effect for Charitable Contributions,” Journal of Political Economy,1989, 97 (5), 1197–1207.
Knack, Stephen and Aminur Rahman, Donor fragmentation and bureaucraticquality in aid recipients, The World Bank, 2004.
Koch, Dirk-Jan and Lau Schulpen, “An exploration of individual-level wage effectsof foreign aid in developing countries,” Evaluation and Program Planning, 2018, 68,233 – 242.
Kuziemko, Ilyana and Eric Werker, “How Much is a Seat on the Security Coun-cil Worth? Foreign Aid and Bribery at the United Nations,” Journal of PoliticalEconomy, 2006, 114 (5), 905–930.
Levinsohn, James and Margaret McMillan, “Does Food Aid Harm the Poor?Household Evidence from Ethiopia,” Working Paper 11048, National Bureau of Eco-nomic Research January 2005.
Mays, Daniel C, Edward J O’Neil, Edison A Mworozi, Benjamin J Lough,Zachary J Tabb, Ashlyn E Whitlock, Edward M Mutimba, and Zohray M
39
Talib, “Supporting and retaining Village Health Teams: an assessment of a com-munity health worker program in two Ugandan districts,” International journal forequity in health, 2017, 16 (1), 129.
MoH, VHT-Village Health Team, Strategy and Operational Guidelines, Ministry ofHealth, 2010.
Nancy, Gilles and Boriana Yontcheva, “Does NGO Aid Go to the Poor? EmpiricalEvidence from Europe,” IMF Working Paper 39, International Monetary Fund 2006.
Nunn, Nathan and Nancy Qian, “US Food Aid and Civil Conflict,” AmericanEconomic Review, June 2014, 104 (6), 1630–66.
Nyamweya, Nancy Nyasuguta, Peter Yekka, Ronny Drasi Mubutu,Keneth Iceland Kasozi, and Jane Muhindo, “Staff absenteeism in public healthfacilities of Uganda: A study in Bushenyi district on contributing factors,” OpenJournal of Nursing, 2017, 7 (10), 1115.
Payne, A. Abigail, “Does the government crowd-out private donations? New evidencefrom a sample of non-profit firms,” Journal of Public Economics, September 1998, 69(3), 323–345.
Pfeiffer, James, Wendy Johnson, Meredith Fort, Aaron Shakow, AmyHagopian, Steve Gloyd, and Kenneth Gimbel-Sherr, “Strengthening healthsystems in poor countries: a code of conduct for nongovernmental organization,”American Journal of Public Health, 2008, 98 (12), 2134–40.
Qian, Nancy, “Making Progress on Foreign Aid,” Annual Review of Economics, 2015,7 (1), 277–308.
Rahman, Muhhamed, “Management of NGOs: A Study in SAARC Countries,”Ph.D. Dissertation, University of Karachi 2003.
Reichenbach, Laura and Shafiun Nahin Shimul, “Sustaining Health: The Role ofthe NGO’s Community Health Volunteers in Bangladesh, Afghanistan and Uganda,”NGO Research Monograph Series 49 2011.
Scott, Kerry, SW Beckham, Margaret Gross, George Pariyo, Krishna D Rao,Giorgio Cometto, and Henry B Perry, “What do we know about community-based health worker programs? A systematic review of existing reviews on communityhealth workers,” Human resources for health, 2018, 16 (1), 1–17.
Turinawe, Emmanueil Benon, Jude T Rwemisisi, Laban K Musinguzi, Mar-ije de Groot, Denis Muhangi, Daniel H de Vries, David K Mafigiri, andRobert Pool, “Selection and performance of village health teams (VHTs) in Uganda:lessons from the natural helper model of health promotion,” Human resources forhealth, 2015, 13 (1), 73.
40
Wagnerly, Zachary, John Bosco Asiimwe, and David I. Levine, “When Finan-cial Incentives Backfire: Evidence From a Community Health Worker Experiment inUganda,” Working Paper, 2019.
Waiswa, Peter, George Pariyo, Karin Kallander, Joseph Akuze, GertrudeNamazzi, Elizabeth Ekirapa-Kiracho, Kate Kerber, Hanifah Sengendo,Patrick Aliganyira, Joy E Lawn et al., “Effect of the Uganda Newborn Study oncare-seeking and care practices: a cluster-randomised controlled trial,” Global healthaction, 2015, 8 (1), 24584.
WHO, “Global Health Observatory Data Repository, Causes of Death Among Chil-dren, Cause-Specific Mortality and Morbidly, World Health Organization,” TechnicalReport 2012.
, WHO recommendations on postnatal care of the mother and newborn, World HealthOrganization, 2014.
, WHO recommendations on antenatal care for a positive pregnancy experience, WorldHealth Organization, 2016.
41
Tab
le1:
NGO
Tim
eAllo
cation
AcrossTa
sks
Act
ivity
% h
ealth
wor
kers
w
ho p
rovi
ded
Tim
e (m
inut
es)
per a
ctiv
ity
# of
activ
ities
pr
ovid
ed
Tota
l tim
e pe
r ac
tivity
(2) x
(3)
Tota
l tim
e pe
r act
ivity
as
% o
f tot
al ti
me
per
mon
th (4
)/29
20 x
100
(1)
(2)
(3)
(4)
(5)
Tota
l34
911
129
2010
0%Pr
egna
ncy
iden
tific
atio
n98
%25
717
56%
Atte
ndin
g de
liver
y an
d pr
ovid
ing
new
born
care
97%
287
196
7%Re
ferr
al to
clin
ic o
r hos
pita
ls88
%21
1123
18%
Atte
ndin
g re
fres
her t
rain
ing
94%
231
246
216
%Se
lling
med
icin
es99
%23
4610
5836
%Se
lling
hea
lth co
mm
oditi
es97
%21
3879
827
%
Not
es:
Sam
ple
= 66
0 N
GO
com
mun
ity h
ealth
wor
kers
inte
rvie
wed
by
the
NG
O in
200
9. W
orke
rs w
ere
aske
d (a
) whi
ch a
ctiv
ities
they
per
form
ed a
s a
heal
th w
orke
r in
the
last
mon
th, (
b) h
ow m
uch
time
was
ded
icat
ed to
eac
h of
thes
e ac
tiviti
es. R
efre
sher
trai
ning
s are
org
aniz
ed o
n a
mon
thly
bas
is in
th
e N
GO
bra
nch
and
are
aim
ed to
"ref
resh
" wor
kers
' hea
lth k
now
ledg
e an
d to
allo
w w
orke
rs to
repl
enis
h th
eir s
tock
of h
ealth
pro
duct
s.
42
Tab
le2:
Allo
cation
ofGovernm
entan
dNGO
Workers
AcrossVillag
esin
2012
No
Gov
in 2
010
Gov
in 2
010
Diff
eren
ce (2
) - (1
) (1
)(2
)(3
)N
o N
GO
0, 0
, 0 [2
4]0.
95, 0
, 0.
95 [3
7]0.
95, 0
, 0,9
5N
GO
0 , 1
, 1
[30]
0.53
, 1, 1
.53
[36]
0.53
, 0, 0
.53
Diff
eren
ce
0, 1
, 1-0
.58,
1, 0
.62
-0.5
8, 0
, -0.
38
Notes: E
ach
cell
repo
rts t
he #
Gov
hea
lth w
orke
rs, #
NG
O h
ealth
wor
kers
, Tot
al #
(NG
O +
G
ov) h
ealth
wor
kers
in 2
012.
The
num
ber o
f vill
ages
is st
ated
in b
rack
ets.
"No
Gov
" and
"G
ov" i
n th
e co
lum
n tit
les r
efer
s to
whe
ther
a v
illag
e ha
d a
gove
rnm
ent h
ealth
wor
ker i
n 20
10. "
NG
O" a
nd "N
o N
GO
" in
the
row
title
s ref
er to
whe
ther
a v
illag
e w
as ra
ndom
ly
assi
gned
to b
e pa
rt o
f the
NG
O p
rogr
am.
43
Tab
le3:
Results
onLa
borSu
pply
ofCom
mun
ityHealthWorkers
NG
O =
{0, 1
}G
ov =
{0, 1
}#
Tota
l (G
ov o
r NG
O)
= {0
, 1, 2
}Tr
aditi
onal
hea
ler
= {0
, 1}
Dru
g st
ore
= {0
, 1}
(1)
(2)
(3)
(4)
(5)
Mea
n D
ep.V
ar.
0.51
20.
425
0.94
50.
236
0.51
2
Gov
-0.0
480.
791*
**0.
791*
**-0
.003
-0.0
02(0
.046
)(0
.074
)(0
.074
)(0
.135
)(0
.101
)
NG
O
0.98
7***
-0.0
200.
980*
**0.
027
-0.0
11(0
.016
)(0
.027
)(0
.027
)(0
.122
)(0
.096
)
Gov
× N
GO
-0
.016
-0.4
46**
*-0
.446
***
-0.1
200.
013
(0.0
20)
(0.0
99)
(0.0
99)
(0.1
58)
(0.1
22)
Obs
erva
tions
127
127
127
127
127
R-sq
uare
d0.
973
0.70
90.
798
0.21
00.
663
[1] G
ov ×
NG
O +
NG
O c
oef
0.97
1-0
.466
0.53
4-0
.093
0.00
2[2
] Gov
× N
GO
+ N
GO
p-v
alue
<0.0
01<0
.001
<0.0
010.
339
0.97
5
Dep
ende
nt V
aria
bles
: Pre
senc
e of
hea
lth p
rovi
ders
in th
e vi
llage
in 2
012
Not
es: O
bser
vatio
ns a
re a
t the
vill
age
leve
l. A
ll re
gres
sion
s inc
lude
are
a fix
ed e
ffect
s. Ro
bust
New
ey-W
est s
tand
ard
erro
rs a
re re
port
ed in
the
pare
nthe
ses.
*** p
<0.0
1, **
p<0
.05,
* p<
0.1
44
Tab
le4:
Results
onHealthServiceAccess
(1)
(2)
(3)
(4)
(5)
(6)
(7)
NG
OG
ovG
ov o
r NG
OTr
aditi
onal
he
aler
Dru
g st
ore
Gov
hos
pita
l/
heal
th ce
nter
Priv
ate
clin
ic
Mea
n D
ep.V
ar.
0.23
50.
313
0.45
70.
014
0.14
30.
384
0.33
7
Gov
-0
.093
**0.
428*
**0.
361*
**-0
.002
-0.0
10-0
.042
0.00
1(0
.038
)(0
.049
)(0
.057
)(0
.006
)(0
.038
)(0
.047
)(0
.037
)
NG
O
0.31
5***
-0.0
190.
288*
**0.
003
-0.0
13-0
.022
-0.0
02(0
.043
)(0
.020
)(0
.046
)(0
.003
)(0
.025
)(0
.043
)(0
.033
)
Gov
× N
GO
-0
.016
-0.2
40**
*-0
.415
***
0.01
0-0
.006
-0.0
220.
014
(0.0
55)
(0.0
65)
(0.0
69)
(0.0
10)
(0.0
38)
(0.0
53)
(0.0
49)
Obs
erva
tions
2,74
72,
747
2,74
72,
747
2,74
72,
747
2,74
7R-
squa
red
0.19
30.
418
0.27
00.
031
0.15
90.
252
0.20
6
[1] G
ov ×
NG
O +
NG
O c
oef
0.29
9-0
.259
-0.1
270.
013
-0.0
19-0
.044
0.01
2[2
] Gov
× N
GO
+ N
GO
p-v
alue
<0.0
01<0
.001
0.00
70.
180
0.50
10.
133
0.74
4
Notes: O
bser
vatio
ns a
re a
t the
hou
seho
ld le
vel.
All
regr
essi
ons i
nclu
de a
rea
fixed
effe
cts.
Stan
dard
err
ors a
re cl
uste
red
at th
e vi
llage
leve
l. **
* p<0
.01,
**
p<0.
05, *
p<0
.1
Dep
ende
nt V
aria
ble:
Hou
seho
ld S
ough
t Med
ical
Adv
ice
from
the
Follo
win
g in
the
Past
Yea
r (20
12) =
{0, 1
}
45
Tab
le5:
Results
onNatal
andPost-Natal
Services
All (1)
(2)
(3)
(4)
(5)
(6)
Pres
ence
of a
w
oman
who
de
liver
ed in
the
past
yea
r D
eliv
ered
in a
ho
spita
l
Del
iver
y w
as
assi
sted
by
a he
alth
pr
ofes
sion
al
Rece
ived
pos
t-na
tal v
isit
with
in
two
days
of b
irth
Chi
ld b
reas
tfed
Chi
ld b
reas
tfed
imm
edia
tely
afte
r bi
rth
Mea
n D
ep.V
ar.
0.14
80.
744
0.53
30.
265
0.99
50.
428
Gov
0.
028
-0.0
530.
355*
**0.
224*
**-0
.014
0.03
2(0
.026
)(0
.067
)(0
.106
)(0
.080
)(0
.012
)(0
.094
)
NG
O
0.01
40.
045
0.20
0*0.
128*
*-0
.018
-0.0
10(0
.017
)(0
.062
)(0
.105
)(0
.063
)(0
.019
)(0
.095
)
Gov
× N
GO
-0
.053
-0.0
49-0
.257
**-0
.216
*0.
027
-0.1
88*
(0.0
34)
(0.0
88)
(0.1
16)
(0.1
10)
(0.0
21)
(0.1
09)
Obs
erva
tions
2,74
740
740
740
740
740
7R-
squa
red
0.09
80.
143
0.10
50.
181
0.04
40.
161
# of
Clu
ster
s12
710
710
710
710
710
7
[1] G
ov ×
NG
O +
NG
O c
oef
-0.0
39-0
.004
-0.0
57-0
.088
0.00
9-0
.198
[2] G
ov ×
NG
O +
NG
O p
-val
ue0.
175
0.94
40.
300
0.30
40.
287
<0.0
01
Dep
ende
nt V
aria
ble:
Hea
lth B
ehav
ior a
nd P
ost-N
atal
Car
e =
{0, 1
}H
ouse
hold
s with
a w
oman
who
del
iver
ed in
the
past
yea
r+
Notes: O
bser
vatio
ns a
re a
t the
hou
seho
ld le
vel.
Sam
ple
rest
rictio
ns a
re st
ated
in th
e co
lum
n he
adin
gs. + In
0.2
% o
f the
hou
seho
lds,
mor
e th
an o
ne
wom
an d
eliv
ered
in th
e pa
st y
ear i
n w
hich
case
we
colla
pse
wom
an-le
vel d
ata
at th
e ho
useh
old
leve
l. A
ll re
gres
sion
s inc
lude
are
a fix
ed e
ffect
s. St
anda
rd e
rror
s are
clus
tere
d at
the
villa
ge le
vel a
nd a
re p
rese
nted
in p
aren
thes
es. *
** p
<0.0
1, **
p<0
.05,
* p<
0.1
46
Tab
le6:
Villag
e-Le
velC
orrelates
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
(9)
Dep
ende
nt V
aria
bles
Mea
n D
ep.V
ar.
Coe
f.St
d. E
rr.
Coe
f.St
d. E
rr.
Coe
f.St
d. E
rr.
Obs
.R2
A. P
rese
nce o
f a H
ealth
Car
e Pro
vide
r in
2010
= {0
, 1}
Gov
Hea
lth W
orke
r0.
575
-0.0
85(0
.072
)12
70.
479
Trad
ition
al H
eale
r 0.
480
-0.0
82(0
.155
)-0
.076
(0.1
32)
0.10
0(0
.186
)12
70.
217
Dru
g St
ore
0.67
7-0
.064
(0.0
53)
-0.0
22(0
.020
)0.
089
(0.0
67)
127
0.78
6G
ov H
ospi
tal w
ithin
10
km
0.55
9-0
.051
(0.1
41)
-0.0
44(0
.108
)-0
.023
(0.1
58)
127
0.36
4G
ov H
ealth
Cen
ter w
ithin
10k
m0.
717
-0.1
08(0
.127
)-0
.042
(0.1
17)
0.14
2(0
.145
)12
70.
416
Priv
ate
Clin
ic w
ithin
10k
m0.
835
-0.0
66(0
.046
)-0
.041
(0.0
56)
0.15
4*(0
.089
)12
70.
557
B. H
ouse
hold
Com
posit
ion
and
Mor
talit
y in
201
0
# H
Hs i
n th
e vi
llage
182.
071
-98.
536*
*(4
3.47
3)-7
3.31
4(4
4.87
4)97
.596
*(5
0.99
3)12
70.
374
# in
fant
s per
HH
0.29
1-0
.007
(0.0
19)
0.01
3(0
.018
)-0
.035
(0.0
28)
127
0.38
7#
infa
nts p
er H
H w
ho d
ied
in p
ast y
ear
0.04
1-0
.010
(0.0
10)
-0.0
07(0
.007
)0.
003
(0.0
17)
127
0.42
4A
ge o
f fem
ale
HH
hea
d31
.160
0.13
3(0
.411
)0.
441
(0.3
14)
-0.4
85(0
.549
)12
70.
760
% H
Hs i
nvol
ved
in fa
rmin
g as
mai
n ac
tivity
0.56
80.
054
(0.0
40)
0.05
9*(0
.035
)-0
.045
(0.0
51)
127
0.90
6%
HH
hea
ds w
ho co
mpl
eted
prim
ary
educ
atio
n0.
381
-0.1
00**
*(0
.036
)-0
.060
(0.0
38)
0.09
0*(0
.050
)12
70.
726
Num
ber o
f ass
ets o
wne
d (o
ut o
f 11)
5.58
30.
011
(0.1
77)
-0.0
78(0
.162
)0.
116
(0.2
68)
127
0.87
2Fo
od se
curit
y (1
to 4
)2.
217
-0.0
21(0
.050
)0.
026
(0.0
51)
0.00
1(0
.066
)12
70.
934
Gov
NG
O
Gov
× N
GO
Not
es: O
bser
vatio
ns a
re a
t the
vill
age
and
year
leve
l. Ea
ch ro
w is
one
regr
essi
on. A
ll re
gres
sion
s inc
lude
are
a fix
ed e
ffect
s. Ro
bust
New
ey-W
est s
tand
ard
erro
rs a
re re
port
ed in
pa
rent
hese
s. **
* p<0
.01,
** p
<0.0
5, *
p<0.
1
47
Tab
le7:
Rob
ustnessto
Con
trols
(1)
(2)
(3)
(4)
# To
tal h
ealth
wor
kers
(G
ov o
r NG
O) =
{0, 1
, 2}
HH
soug
ht m
edica
l adv
ice
from
any
hea
lth w
orke
r in
the
past
yea
r = {0
, 1}
Del
iver
y w
as a
ssist
ed b
y a
heal
th p
rofe
ssio
nal
= {0
, 1}
Rece
ived
pos
t-nat
al v
isit
with
in tw
o da
ys o
f birt
h =
{0, 1
}
Mea
n D
ep.V
ar.
0.94
50.
457
0.53
30.
265
Gov
0.70
9***
0.43
0***
0.43
2***
0.42
8***
(0.1
29)
(0.0
72)
(0.0
90)
(0.1
12)
NG
O
0.94
7***
0.33
2***
0.20
3**
0.26
7***
(0.0
80)
(0.0
43)
(0.0
90)
(0.0
86)
Gov
× N
GO
-0
.417
**-0
.471
***
-0.3
66**
*-0
.546
***
(0.1
67)
(0.0
85)
(0.1
21)
(0.1
41)
Obs
erva
tions
127
2,74
740
740
7R-
squa
red
0.85
10.
298
0.16
30.
313
Uni
t of O
bser
vatio
nV
illag
eH
ouse
hold
Hou
seho
ld
Hou
seho
ld
Stan
dard
Err
orN
ewey
-Wes
tC
lust
er V
illag
eC
lust
er V
illag
eC
lust
er V
illag
e
[1] G
ov ×
NG
O +
NG
O c
oef
0.53
0-0
.139
-0.1
63-0
.279
[2] G
ov ×
NG
O +
NG
O p
-val
ue<0
.001
0.01
30.
045
0.00
1
Dep
ende
nt V
aria
bles
Not
es: T
his t
able
pre
sent
s coe
fficie
nts e
valu
ated
at t
he m
ean
valu
e of
all
of th
e co
ntro
l var
iabl
es. A
ll re
gres
sions
cont
rol f
or a
rea
fixed
effe
cts a
nd th
e fo
llow
ing
addi
tiona
l con
trol
s, w
hich
are
all
mea
sure
d in
201
0 an
d de
mea
ned,
and
thei
r int
erac
tions
with
the
pres
ence
of a
gov
hea
lth w
orke
r in
2010
: vi
llage
size
and
com
posit
ion
(# o
f HH
, # o
f inf
ants
per
HH
). H
H ch
arac
teris
tics:
%H
Hs i
nvol
ved
in fa
rmin
g, %
HH
hea
ds w
ho co
mpl
eted
prim
ary
educ
atio
n), p
roxi
es fo
r wea
lth (a
vera
ge fo
od se
curit
y, a
vera
ge n
umbe
r of a
sset
s ow
ned
by a
HH
), ac
cess
to h
ealth
pro
vide
rs (p
rese
nce
of g
ov h
ospi
tal
with
in 1
0km
, pre
senc
e of
gov
hea
lth ce
nter
with
in 1
0km
, pre
senc
e of
priv
ate
clini
c with
in 1
0km
, pre
senc
e of
a d
rug
stor
e, p
rese
nce
of a
trad
ition
al
heal
er),
mor
talit
y (#
of c
hild
ren
<1 p
er H
H w
ho d
ied
in p
ast y
ear)
. ***
p<0
.01,
** p
<0.0
5, *
p<0.
1.
48
Tab
le8:
Results
onInfant
Mortality
(1)
(2)
(3)
(4)
(5)
At l
east
one
di
ed
#die
d as
shar
e of
birt
hs+
# di
ed p
er 1
,000
bi
rths
+
#die
d as
shar
e of
birt
hs++
#
died
per
1,0
00
birt
hs++
Mea
n D
ep.V
ar.
0.04
00.
053
69.0
690.
088
117.
215
Gov
-0
.026
**-0
.063
***
-76.
710*
**-0
.073
**-1
16.0
57**
*(0
.012
)(0
.021
)(2
7.31
4)(0
.033
)(4
3.36
2)
NG
O
-0.0
23**
-0.0
57**
*-5
8.25
1**
-0.0
72**
-74.
677*
(0.0
11)
(0.0
20)
(28.
351)
(0.0
32)
(39.
469)
Gov
× N
GO
0.
029*
0.06
2**
68.0
36*
0.08
8**
116.
083*
*(0
.016
)(0
.025
)(3
4.85
5)(0
.041
)(5
8.53
9)
Obs
erva
tions
2,74
71,
408
127
936
127
R-sq
uare
d0.
021
0.03
70.
248
0.05
80.
256
[1] G
ov ×
NG
O +
NG
O c
oef
0.00
60.
005
9.78
50.
016
41.4
06[2
] Gov
× N
GO
+ N
GO
p-v
alue
0.57
10.
693
0.60
00.
529
0.32
8
Uni
t of o
bser
vatio
nH
ouse
hold
Hou
seho
ldVi
llage
Hou
seho
ldVi
llage
Stan
dard
Err
orC
lust
er V
illag
eC
lust
er V
illag
eN
ewey
-Wes
tC
lust
er V
illag
eN
ewey
-Wes
t
Dep
ende
nt V
aria
ble:
Infa
nt (U
nder
-1)
Mor
talit
y
Not
es: A
ll re
gres
sion
s con
trol
for a
rea
fixed
effe
cts.
*** p
<0.0
1, **
p<0
.05,
* p<
0.1
+ In C
ol. 2
, we
calc
ulat
e th
e ho
useh
old-
leve
l inf
ant m
orta
lity
ratio
as (
# ch
ildre
n bo
rn 2
010-
2012
and
who
die
d be
low
age
one
201
0-20
12)/
(# ch
ildre
n bo
rn
2010
-201
2). T
he la
tter i
s # ch
ildre
n b
orn
2010
-201
2 an
d d
ied
belo
w a
ge o
ne 2
010-
2012
+ #
child
ren
born
201
0-20
12 a
nd
did
not d
ie d
urin
g 20
10-2
012
(i.e.
, chi
ldre
n le
ss th
an 3
yea
rs in
201
2). C
ol. 3
follo
ws a
sim
ilar a
ppro
ach
as C
ol. 2
but
ag
greg
ates
all
num
bers
at t
he v
illag
e le
vel r
athe
r tha
n at
the
hous
ehol
d le
vel a
nd fu
rthe
r mul
tiplie
s the
den
omin
ator
by
1,0
00. ++
Col
. 4-5
repl
icat
e th
e ap
proa
ch o
f Col
. 2-3
but
exc
lude
s fro
m th
e ra
tio a
ny ch
ild w
ho, i
n 20
12, i
s les
s tha
n on
e ye
ar o
ld o
r who
wou
ld h
ave
been
less
than
one
had
he/
she
not d
ied.
49
Tab
le9:
Decom
position
oftheInfant
MortalityResults
(1)
(2)
(3)
(4)
(5)
(6)
NG
OG
ovG
ov o
r NG
OA
t lea
st o
ne d
ied
#die
d as
shar
e of
bi
rths
+ #
died
per
1,0
00
birt
hs+
Mea
n D
ep.V
ar.
0.23
50.
313
0.45
70.
040
0.05
369
.069
Gov
-0.0
79**
0.50
1***
0.39
7***
-0.0
33**
*-0
.073
***
-87.
941*
**(0
.038
)(0
.052
)(0
.060
)(0
.012
)(0
.022
)(2
8.05
0)
NG
O
0.31
6***
-0.0
150.
290*
**-0
.023
**-0
.057
***
-58.
983*
*(0
.043
)(0
.017
)(0
.046
)(0
.011
)(0
.021
)(2
8.58
5)
Gov
× N
GO
× 1
hea
lth w
orke
r in
2012
(NG
O)
-0.0
68-0
.506
***
-0.5
45**
*0.
053*
*0.
096*
**11
0.54
9**
(0.0
55)
(0.0
47)
(0.0
71)
(0.0
21)
(0.0
31)
(43.
006)
Gov×
NG
O ×
2 h
ealth
wor
kers
in 2
012
(NG
O +
Gov
)0.
040
0.04
9-0
.275
***
0.00
30.
035
22.9
83(0
.068
)(0
.060
)(0
.070
)(0
.016
)(0
.025
)(3
3.61
3)
Obs
erva
tions
2,74
72,
747
2,74
72,
747
1,40
812
7R-
squa
red
0.19
60.
488
0.28
50.
024
0.04
20.
302
Uni
t of o
bser
vatio
nH
ouse
hold
Hou
seho
ldH
ouse
hold
Hou
seho
ldH
ouse
hold
Villa
geSt
anda
rd E
rror
Clu
ster
Vill
age
Clu
ster
Vill
age
Clu
ster
Vill
age
Clu
ster
Vill
age
Clu
ster
Vill
age
New
ey-W
est
[1] G
ov ×
NG
O x
1 w
orke
r + N
GO
coe
f0.
248
-0.5
21-0
.255
0.03
00.
039
51.5
67[2
] Gov
× N
GO
x 1
wor
ker +
NG
O p
-val
ue<0
.001
<0.0
01<0
.001
0.07
20.
090
0.09
4[3
] Gov
× N
GO
x 2
wor
kers
+ N
GO
coef
0.35
60.
034
0.01
5-0
.020
-0.0
22-3
5.99
9[4
] Gov
× N
GO
x 2
wor
kers
+ N
GO
p-v
alue
<0.0
010.
553
0.77
00.
070
0.08
50.
025
Notes: T
he u
nit o
f obs
erva
tion
is st
ated
at t
he b
otto
m o
f the
tabl
e. A
ll re
gres
sion
s con
trol
for a
rea
fixed
effe
cts.
*** p
<0.0
1, **
p<0
.05,
* p<
0.1
+ In C
ol. 5
, we
calc
ulat
e th
e ho
useh
old-
leve
l inf
ant m
orta
lity
ratio
as (
# ch
ildre
n bo
rn 2
010-
2012
and
who
die
d be
low
age
one
201
0-20
12)/
(# ch
ildre
n bo
rn 2
010-
2012
). Th
e la
tter i
s # ch
ildre
n b
orn
2010
-201
2 an
d di
ed b
elow
age
one
201
0-20
12 +
# ch
ildre
n bo
rn 2
010-
2012
and
did
not
die
dur
ing
2010
-201
2 (i.
e., c
hild
ren
less
than
3 y
ears
in 2
012)
. Col
. 6 fo
llow
s a
sim
ilar a
ppro
ach
as C
ol. 5
but
agg
rega
tes a
ll nu
mbe
rs a
t the
vill
age
leve
l rat
her t
han
at th
e ho
useh
old
leve
l and
furt
her m
ultip
lies t
he d
enom
inat
or b
y 1,
000.
Dep
ende
nt V
aria
ble
HH
soug
ht m
edic
al a
dvic
e fr
om th
e fo
llow
ing
= {0
, 1}
Infa
nt M
orta
lity
50
Figure
1:UnitPrice
&Marginpe
rProdu
ctSo
ldby
theNGO
Not
es: T
his f
igur
e pr
esen
ts th
e lis
t of p
rodu
cts s
old
by th
e N
GO
wor
kers
, ran
ked
from
low
est t
o hi
ghes
t uni
t m
argi
n. T
he u
nit m
argi
n eq
uals
the
pric
e at
whi
ch th
e he
alth
wor
ker s
ells
the
prod
ucts
in h
er co
mm
unity
("
heal
th w
orke
r pric
e") m
inus
the
pric
e at
whi
ch sh
e bu
ys th
e pr
oduc
ts fr
om th
e N
GO
("N
GO
pric
e").
Pric
es
are
expr
esse
d in
Uga
ndan
Shi
lling
s (1$
=3,7
13 U
GX)
.
0
500
1000
1500
2000
2500
3000
3500
4000
4500
5000
PRICE AND MARGIN (IN UGX)
NG
O P
RIC
Ehe
alth
wor
ker P
RIC
EU
NIT
MA
RGIN
51
Tab
leA.1:Lo
gitEstim
ation
Pois
son
(1)
(2)
(3)
(4)
(5)
(6)
(7)
NG
OG
ovN
GO
or G
OV
Del
iver
y w
as
assi
sted
by
a he
alth
pr
ofes
sion
al
Rece
ived
pos
t-nat
al
visi
t with
in tw
o da
ys o
f birt
h
At l
east
one
infa
nt
died
# To
tal h
ealth
w
orke
rs (G
ov o
r N
GO
) = {0
, 1, 2
}
Mea
n D
ep.V
ar.
0.23
50.
313
0.45
70.
533
0.26
50.
040
0.94
5
Gov
-0
.118
2.69
3***
2.09
6***
1.55
1***
2.11
6***
-0.7
25**
16.7
74**
*(0
.320
)(0
.291
)(0
.300
)(0
.482
)(0
.799
)(0
.298
)(0
.205
)
NG
O
2.55
6***
-0.3
891.
844*
**0.
850*
1.54
4*-0
.633
**16
.934
***
(0.2
99)
(0.3
09)
(0.2
49)
(0.4
66)
(0.7
96)
(0.2
76)
(0.1
92)
Gov
* N
GO
-0
.692
*-0
.906
*-2
.464
***
-1.1
10**
-1.9
82**
0.78
0*-1
6.49
7***
(0.3
67)
(0.4
68)
(0.3
54)
(0.5
09)
(0.9
22)
(0.4
08)
(0.2
10)
[1] G
ov ×
NG
O +
NG
O c
oef
1.86
4-1
.295
-0.6
20-0
.261
-0.4
380.
147
0.43
7[2
] Gov
× N
GO
+ N
GO
p-v
alue
0.00
00.
000
0.00
90.
270
0.34
20.
602
0.00
0
Obs
erva
tions
2,74
72,
747
2,74
740
735
72,
747
127
Uni
t of o
bser
vatio
nH
ouse
hold
Hou
seho
ldH
ouse
hold
Hou
seho
ldH
ouse
hold
Hou
seho
ldVi
llage
Stan
dard
Err
orC
lust
er V
illag
eC
lust
er V
illag
eC
lust
er V
illag
eC
lust
er V
illag
eC
lust
er V
illag
eC
lust
er V
illag
eN
ewey
-Wes
t
Logi
t
Notes:
All
regr
essi
ons c
ontr
ol fo
r are
a fix
ed e
ffect
s. **
* p<0
.01,
** p
<0.0
5, *
p<0.
1
HH
soug
ht m
edic
al a
dvic
e fr
om th
e fo
llow
ing
= {0
, 1}
Del
iver
y, P
ost-n
atal
Car
e, M
orta
lity
= {0
, 1}
1
Tab
leA.2:Multino
mialL
ogit
Estim
ation
(1)
(2)
(3)
(4)
12
34
Mea
n D
ep.V
ar.
0.30
60.
306
0.30
60.
306
Gov
-0
.298
0.42
50.
969
1.72
0(0
.327
)(0
.357
)(1
.042
)(1
.072
)
NG
O
-0.8
47**
0.57
21.
279
2.64
6**
(0.4
23)
(0.4
34)
(1.1
50)
(1.0
32)
Gov
× N
GO
0.
432
-0.7
11-1
.508
-2.4
03**
(0.4
56)
(0.4
46)
(1.1
08)
(1.1
31)
Obs
erva
tions
2,74
72,
747
2,74
72,
747
Dep
ende
nt V
aria
ble
Notes: O
bser
vatio
ns a
re a
t the
hou
seho
ld le
vel.
All
estim
ates
incl
ude
area
fixe
d ef
fect
s. St
anda
rd
erro
rs a
re cl
uste
red
at v
illag
e le
vel.
*** p
<0.0
1, **
p<0
.05,
* p<
0.1
Mul
tinom
ial l
ogit
(0 to
4)
0 =
No
wom
an d
eliv
ered
1
= W
oman
del
iver
ed, n
ot a
ssis
ted
by a
hea
lth p
rofe
ssio
nal,
with
out p
ost n
atal
care
2
= W
oman
del
iver
ed, a
ssis
ted
by a
hea
lth p
rofe
ssio
nal,
with
out p
ost n
atal
care
3
= W
oman
del
iver
ed, n
ot a
ssis
ted
by a
hea
lth p
rofe
ssio
nal,
with
pos
t nat
al ca
re
4 =
Wom
an d
eliv
ered
, ass
iste
d by
a h
ealth
pro
fess
iona
l, w
ith p
ost n
atal
care
2
Tab
leA.3:Heterog
eneous
Effe
ctsby
DrugStore
(1)
(2)
# To
tal h
ealth
wor
kers
(G
ov o
r NG
O) =
{0, 1
, 2}
HH
soug
ht m
edic
al a
dvic
e fr
om a
ny h
ealth
wor
ker i
n th
e pa
st y
ear =
{0, 1
}
Mea
n D
ep.V
ar.
0.94
50.
457
Gov
× N
GO
× D
rug
Stor
e-0
.042
0.22
7(0
.194
)(0
.142
)
Obs
erva
tions
127
2,74
7R-
squa
red
0.80
00.
274
Uni
t of O
bser
vatio
nVi
llage
Hou
seho
ldSt
anda
rd E
rror
New
ey-W
est
Clu
ster
Vill
age
Notes
: All
regr
essi
ons c
ontr
ol fo
r are
a fix
ed e
ffect
s and
the
full
set o
f int
erac
tion
term
s an
d un
inte
ract
ed te
rms:
gov,
NG
O, d
rug
stor
e, g
ov ×
dru
g st
ore,
ngo
× d
rug
stor
e, g
ov
× N
GO
. ***
p<0
.01,
** p
<0.0
5, *
p<0.
1
Dep
ende
nt V
aria
ble
3
Tab
leA.4:Association
betw
eenHealthServices
andInfant
mortality
(1)
(2)
(3)
(4)
(5)
(6)
Gov
-0.0
13-0
.100
**(0
.009
)(0
.045
) N
GO
-0.0
30**
*-0
.087
**(0
.010
)(0
.043
) G
ov o
r NG
O-0
.025
***
-0.1
52**
*(0
.009
)(0
.051
)
Obs
erva
tions
2,74
740
72,
747
407
2,74
740
7R-
squa
red
0.02
00.
069
0.02
30.
067
0.02
30.
097
Gov
-0.0
16-0
.048
**(0
.014
)(0
.022
) N
GO
-0.0
55**
*-0
.053
**(0
.012
)(0
.021
) G
ov o
r NG
O-0
.042
***
-0.0
80**
*(0
.014
)(0
.025
)
Obs
erva
tions
1,40
835
71,
408
357
1,40
835
7R-
squa
red
0.02
70.
064
0.04
10.
070
0.03
50.
092
Expl
anat
ory
Var
: HH
so
ught
med
ical a
dvice
fr
om th
e fo
llow
ing
in
the
past
yea
rPa
nel A
. Dep
ende
nt V
aria
ble
-- A
t lea
st o
ne in
fant
die
d
Pane
l B. D
epen
dent
Var
iabl
e --
# of
infa
nts w
ho d
ied
as a
shar
e of
birt
hs
Not
es:
Odd
num
bere
d co
lum
ns co
mpr
ise o
f all
hous
ehol
ds. E
ven
num
bere
d co
lum
ns co
mpr
ise o
f H
Hs w
ith w
oman
who
del
iver
ed in
pas
t yea
r. +I
n 0.
2% o
f the
hou
seho
lds,
mor
e th
an o
ne w
oman
de
liver
ed in
the
past
yea
r in
whi
ch ca
se w
e co
llaps
e w
oman
-leve
l dat
a at
the
hous
ehol
d le
vel.
All
regr
essio
ns in
clude
are
a fix
ed e
ffect
s. St
anda
rd e
rror
s are
clus
tere
d at
the
villa
ge le
vel a
nd a
re
pres
ente
d in
par
enth
eses
. ***
p<0
.01,
** p
<0.0
5, *
p<0.
1
4
Tab
leA.5:Results
onInfant
Mortality–Rob
ustnessto
Con
trols
(1)
(2)
(3)
(4)
(5)
At l
east
one
die
d #d
ied
as sh
are
of
birt
hs+
# di
ed p
er 1
,000
bi
rths
+
#die
d as
shar
e of
bi
rths
++
# di
ed p
er 1
,000
bi
rths
++
Mea
n D
ep.V
ar.
0.04
00.
053
69.0
690.
088
117.
215
Gov
-0.0
56**
*-0
.123
***
-119
.947
***
-0.1
43**
*-1
68.0
99**
(0.0
21)
(0.0
38)
(43.
991)
(0.0
54)
(72.
064)
NG
O
-0.0
43**
-0.0
98**
*-8
9.51
7**
-0.1
19**
*-1
28.3
96**
(0.0
18)
(0.0
32)
(38.
904)
(0.0
44)
(60.
598)
Gov
× N
GO
0.
062*
*0.
121*
**11
3.71
0**
0.15
8**
190.
433*
(0.0
25)
(0.0
43)
(53.
903)
(0.0
62)
(95.
937)
Obs
erva
tions
2,74
71,
408
127
936
127
R-sq
uare
d0.
033
0.06
70.
423
0.09
50.
397
Uni
t of O
bser
vatio
nH
ouse
hold
Hou
seho
ldVi
llage
Hou
seho
ldVi
llage
Stan
dard
Err
orC
lust
er V
illag
eC
lust
er V
illag
eN
ewey
-Wes
tC
lust
er V
illag
eN
ewey
-Wes
t
[1] G
ov ×
NG
O +
NG
O c
oef
0.01
90.
023
24.1
930.
039
62.0
38[2
] Gov
× N
GO
+ N
GO
p-v
alue
0.14
20.
238
0.31
70.
266
0.24
1
Dep
ende
nt V
aria
ble:
Infa
nt (U
nder
-1) M
orta
lity
Not
es: A
ll re
gres
sion
s con
trol
for a
rea
fixed
effe
cts a
nd th
e fo
llow
ing
addi
tiona
l con
trol
s, w
hich
are
all
mea
sure
d in
201
0 an
d de
mea
ned,
an
d th
eir i
nter
actio
ns w
ith th
e pr
esen
ce o
f a g
ov h
ealth
wor
ker i
n 20
10: v
illag
e si
ze a
nd co
mpo
sitio
n (#
of H
H, #
of i
nfan
ts p
er H
H).
HH
ch
arac
teris
tics:
%H
Hs i
nvol
ved
in fa
rmin
g, %
HH
hea
ds w
ho co
mpl
eted
prim
ary
educ
atio
n), p
roxi
es fo
r wea
lth (a
vera
ge fo
od se
curit
y,
aver
age
num
ber o
f ass
ets o
wne
d by
a H
H),
acce
ss to
hea
lth p
rovi
ders
(pre
senc
e of
gov
hos
pita
l with
in 1
0km
, pre
senc
e of
gov
hea
lth
cent
er w
ithin
10k
m, p
rese
nce
of p
rivat
e cl
inic
with
in 1
0km
, pre
senc
e of
a d
rug
stor
e, p
rese
nce
of a
trad
ition
al h
eale
r), m
orta
lity
(# o
f ch
ildre
n <1
per
HH
who
die
d in
pas
t yea
r). *
** p
<0.0
1, **
p<0
.05,
* p<
0.1
+ In C
ol. 2
, we
calc
ulat
e th
e ho
useh
old-
leve
l inf
ant m
orta
lity
ratio
as (
# ch
ildre
n bo
rn 2
010-
2012
and
who
die
d be
low
age
one
201
0-20
12)/
(# ch
ildre
n bo
rn 2
010-
2012
). Th
e la
tter i
s # ch
ildre
n b
orn
2010
-201
2 an
d d
ied
belo
w a
ge o
ne 2
010-
2012
+ #
child
ren
born
201
0-20
12 a
nd d
id n
ot d
ie d
urin
g 20
10-2
012
(i.e.
, chi
ldre
n le
ss th
an 3
ye
ars i
n 20
12).
Col
. 3 fo
llow
s a si
mila
r app
roac
h as
Col
. 2 b
ut a
ggre
gate
s all
num
bers
at t
he v
illag
e le
vel r
athe
r tha
n at
the
hous
ehol
d le
vel a
nd fu
rthe
r mul
tiplie
s the
den
omin
ator
by
1,00
0. ++
Col
. 4-5
repl
icat
e th
e ap
proa
ch o
f Col
. 2-3
but
exc
lude
s fro
m th
e ra
tio a
ny ch
ild
who
, in
2012
, is l
ess t
han
one
year
old
or w
ho w
ould
hav
e be
en le
ss th
an o
ne h
ad h
e/sh
e no
t die
d.
5
Tab
leA.6:Results
onDisease
Treatm
ent
(1)
(2)
(3)
Mal
aria
Dhi
arre
aPn
eum
onia
Mea
n D
ep.V
ar. (
1,00
0 U
gand
an S
hilli
ngs)
5.64
33.
418
3.35
7
Gov
-1.2
19-2
.019
*-1
.434
*(1
.115
)(1
.074
)(0
.839
)
NG
O
0.31
0-1
.276
-0.4
89(1
.279
)(1
.434
)(1
.105
)
Gov
× N
GO
0.
006
0.98
30.
596
(1.5
78)
(1.5
71)
(1.2
06)
Obs
erva
tions
2,23
61,
977
1,81
9R-
squa
red
0.10
50.
078
0.10
3
[1] G
ov ×
NG
O +
NG
O c
oef
0.31
7-0
.294
0.10
7[2
] Gov
× N
GO
+ N
GO
p-v
alue
0.72
50.
579
0.81
5
Dep
ende
nt V
aria
ble:
Pric
e of
trea
tmen
t (m
edic
atio
n an
d te
sts)
Notes: O
bser
vatio
ns a
re a
t the
hou
seho
ld a
nd y
ear l
evel
. All
regr
essi
ons i
nclu
de a
rea
fixed
effe
cts.
Stan
dard
err
ors a
re cl
uste
red
at
the
villa
ge le
vel.
*** p
<0.0
1, **
p<0
.05,
* p<
0.1
6
Tab
leA.7:Results
onOther
HealthOutcomes
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
(9)
Dep
ende
nt V
aria
bles
Mea
n D
ep.V
ar.
Coe
f.St
d. E
rr.
Coe
f.St
d. E
rr.
Coe
f.St
d. E
rr.
Obs
.R2
A. H
ealth
Beh
avio
r -- A
ll ho
useh
olds
Chi
ldre
n sl
eep
unde
r bed
net
0.72
50.
064*
*(0
.030
)-0
.003
(0.0
33)
-0.0
64(0
.049
)2,
747
0.08
4C
hild
ren
drin
k tr
eate
d w
ater
(e.g
., bo
iled,
0.
722
0.03
1(0
.032
)0.
008
(0.0
26)
-0.0
15(0
.047
)2,
747
0.21
0C
hild
ren
was
h ha
nds b
efor
e fo
od a
nd a
fter
0.87
80.
024
(0.0
25)
-0.0
02(0
.022
)-0
.040
(0.0
41)
2,74
70.
084
Chi
ldre
n ar
e fu
lly im
mun
ized
0.90
50.
012
(0.0
29)
0.01
9(0
.022
)-0
.027
(0.0
28)
2,04
50.
236
Cou
ple
uses
cont
race
ptiv
es (e
.g.,
cond
om, p
ill,
ster
iliza
tion)
0.46
9-0
.070
(0.0
46)
-0.0
25(0
.034
)0.
047
(0.0
57)
2,51
00.
036
Aver
age
stan
dard
ized
effe
ct o
f out
com
es a
bove
0.01
4(0
.046
)0.
011
(0.0
37)
-0.0
40(0
.063
)1,
915
0.12
1
B. D
iseas
e Inc
iden
ce --
Hou
seho
lds w
ith a
t lea
st o
ne ch
ild u
nder
5 y
ears
of a
geC
ough
0.59
3-0
.019
(0.0
39)
0.01
3(0
.028
)0.
013
(0.0
43)
1,78
30.
204
Dhi
arre
a0.
284
-0.0
44(0
.039
)-0
.011
(0.0
35)
0.01
0(0
.047
)1,
783
0.18
1W
orm
s0.
348
-0.0
08(0
.036
)-0
.007
(0.0
33)
-0.0
18(0
.046
)1,
783
0.27
5TB
0.03
80.
003
(0.0
26)
0.02
2(0
.014
)-0
.030
(0.0
26)
1,78
30.
111
Mal
aria
0.47
10.
001
(0.0
33)
0.00
3(0
.032
)-0
.013
(0.0
44)
1,78
30.
119
Aver
age
stan
dard
ized
effe
ct o
f out
com
es a
bove
-0.0
28(0
.064
)0.
023
(0.0
44)
-0.0
36(0
.069
)1,
783
0.15
8
Gov
NG
O
Gov
× N
GO
Not
es: O
bser
vatio
ns a
re a
t the
hou
seho
ld a
nd y
ear l
evel
. Sam
ple
rest
rictio
ns a
re st
ated
in th
e pa
nel h
eadi
ngs.
The
aver
age
stan
dard
ized
effe
ct is
an
equa
lly
wei
ghte
d av
erag
e of
z-s
core
s of i
ts co
mpo
nent
s. In
Pan
el B
, mor
e th
an o
ne u
nder
-5 ch
ild li
ves i
n th
e ho
useh
old
in 3
4% o
f the
hou
seho
lds,
in w
hich
case
we
colla
psed
child
-leve
l dat
a at
the
hous
ehol
d le
vel.
Each
row
is o
ne re
gres
sion
. All
regr
essi
ons i
nclu
de a
rea
fixed
effe
cts.
Stan
dard
err
ors a
re cl
uste
red
at th
e vi
llage
leve
l. **
* p<0
.01,
** p
<0.0
5, *
p<0.
1
7