Policy Research Working Paper                9486




     Subsidies, Information, and the Timing
       of Children’s Health Care in Mali
                             Anja Sautmann
                             Samuel Brown
                              Mark Dean




Development Economics
Development Research Group
November 2020
Policy Research Working Paper 9486


  Abstract
 Sustained progress in reducing child mortality requires                            of nine weeks of daily symptom and health care use data
 better care for children who are acutely ill. This paper stud-                     to study the impact of each policy on demand conditional
 ies how health care subsidies and health workers providing                         on need for care, as defined by World Health Organization
 information on symptoms affect the overuse and underuse                            standards. Subsidies substantially increase care when it is
 of primary care, which depend not just on absolute levels                          medically indicated, while overuse remains rare. Health
 of demand, but also on whether care is received when the                           worker visits have no aggregate effect on demand, but
 child is actually sick. In a randomized controlled trial of                        they may help the youngest children take advantage of the
 1,768 children in Mali, the study collected a unique panel                         subsidy.




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




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


                                                       Produced by the Research Support Team
               Subsidies, Information, and the
       Timing of Children’s Health Care in Mali

                         , Samuel Brown† and Mark Dean‡
            Anja Sautmann∗

                         This version: November 2020


 JEL codes: I15, O12
 Keywords: demand for acute care, children’s health care, health care under-
use and overuse, health care subsidies, community health workers, Mali




   ∗
     The World Bank, corresponding author, email: asautmann@worldbank.org.
   †
     Wayfair Inc., email: sambrown@wayfair.com.
   ‡
     Columbia University, email: mark.dean@columbia.edu.
We received invaluable advice and support from Andrew Foster, Duncan Thomas, and Ira
Wilson. We would also like to thank Anna Aizer, Dan Björkegren, Joseph Camardo, Caitlin
Cohen, Jishnu Das, Seydou Doumbia, Chris Gill, Rachel Glennerster, Jessica Goldberg,
Bentley McLeod, Yusuf Neggers, Emily Oster, Simone Schaner, Dan Silverman, and Dean
Yang, and participants of the “Natural Experiments and Controlled Field Studies”workshop
2015. Special thanks go to Hamadoun Bocoum, Zorina Curry, Judith Kom, and Fred Wens-
ing, Mali Health and its staff and Board, and Innovations for Poverty Action. We are grateful
for funding from Brown University through the Brown Seed Fund Award, the Population
Studies and Training Center, and the Rhodes Center, from the Aga Khan Foundation, and
from the ESRC/DFID Development Frontiers Award (ES/K01207X/1). The findings, inter-
pretations, and conclusions expressed in this paper are entirely those of the authors. They
do not necessarily represent the views of the World Bank and its affiliated organizations, or
those of the Executive Directors of the World Bank or the governments they represent.
1   Introduction

In the last two decades, great strides have been made in reducing child mor-
tality through preventive care programs (e.g. Bhatt et al., 2015). Nonetheless,
over five million children under five years died in 2018 alone (Hug et al., 2019).
Current consensus is that further progress requires better care for children who
are acutely ill. Two demand-side policies have played a key role in this effort:
health care subsidies and community healthworker (CHW) home visits, which
serve to educate caretakers and monitor the health of the child.
 The policy debate on how to fund primary care for children is longstanding,
and important developments are currently underway. The Bamako consensus
from 1987 mandated charging user fees to maintain sustainability and quality
of care. This view has been reversed in policy circles in favor of universal
subsidization of basic primary care for small children and mothers, see e.g. Akin
et al. (1987); Litvack and Bodart (1993); Jimenez (1995); Save the Children
(2008); UK Secretary of State for International Development (2009). In 2019,
Mali’s Ministry of Health announced its intention to provide free children’s and
mothers’ health care by 2022, at an estimated cost of $120 million, although it
acknowledged funding gaps (Adepoju, 2019). While a rise in security incidents,
political turbulence, and the Covid-19 crisis have stalled the reforms, this
proposal will likely serve as a benchmark for future health policy.
 It is known that subsidy and information policies for preventive care can
greatly improve demand (Ashraf et al. (2010); Dupas and Cohen (2010); Ashraf
et al. (2013); Dupas (2014); Miguel and Kremer (2004), see also Kremer and
Glennerster (2012)). But unlike for prevention, whether and when acute health


                                       1
care is needed depends on the child’s health status, and it is up to the child’s
caretakers at home to monitor symptoms and make decisions about access to
care. This makes an assessment of the demand for acute care more difficult. As
African health care systems move towards greater subsidization (Yates (2007),
Ridde and Morestin (2012)), effective policy-making requires understanding
to what degree subsidies encourage health care use when the child is in need
of medical attention vs. not, and whether complementary health information
policies can improve utilization. Are unfilled health care needs met? What
share of expenses may end up paying for unnecessary treatment?
 A priori, subsidies for children’s primary care are likely to increase use. This
could improve the allocation of care if there are access barriers, for example
if caretakers face credit constraints, misjudge the child’s health care needs,
prioritize the costs to them over the benefits to the child, or disregard the
negative externalities of untreated infection. At the same time, subsidies may
cause the waste of scarce resources if they lead caretakers to overuse care,
in particular if they have little health education and geographical distance
is not a primary barrier, as is the case in urban areas. This is a particular
concern given that increased demand will in practice challenge existing supply
capacities. Health education and information provided by CHWs might in this
context help direct those children to formal care who need it most.1
 Central to interpreting any demand response is the ability to estimate the
   1
     While CHWs programs were initially conceived as a way to reach patients in remote
rural areas, many CHW work in practice in urban or peri-urban areas. Comprehensive data
is sparse, but even in national CHW programs that target rural populations, a significant
share of healthworkers operate in urban areas (e.g. 20% of CHW in Pakistan, Brook (2009)).
Besides directing children’s acute care, CHW typically also promote prevention and prenatal
care; aspects we do not focus on here. We document few differences in acute illness incidence
with and without CHW (see below).


                                             2
impact of subsidies and CHW visits on care seeking conditional on the health
status of the child. As an example, randomized trials of health insurance
subsidies in Nicaragua and Mexico have resulted in few health or utilization
effects, with low uptake and retention (King et al. (2009); Thornton et al.
(2010)). Information on conditional demand could help determine whether
this is because there is no significant unfilled need for care, or because the
intervention did not succeed in curbing underuse, perhaps due to unobserved
supply-side effects. In this study, we address this gap with the help of novel,
detailed daily health data collected at home.
 In partnership with the NGO Mali Health, we conducted a randomized con-
trol trial of a health care subsidy, which reduced primary care cost of chil-
dren under five years by 72% on average, and a community healthworker
(CHW) program, for which CHWs visit families biweekly and teach them basic
health knowledge. The interventions were cross-randomized at the compound
(dwelling) level, with a final sample of 1768 children from 642 compounds.
Comprehensive quality controls at Mali Health’s partner clinics and the small
sample mitigate general equilibrium effects and allow us to focus on the de-
mand response to the interventions.
 Nine months after intervention start, we collected nine weeks of daily data on
children’s health and health care. Building on the Integrated Management of
Childhood Illness (IMCI) guidelines for community healthworkers,2 we record
14 symptoms that are key in the detection of the main causes of child mortality
in resource-poor settings. We construct illness spells and observe demand for
  2
   The version adapted for the use by community healthworkers is called C-IMCI, see
WHO (2014); WHO Department of Child and Adolescent Health and Development (2005);
UNICEF Health Section (1999); Rosales and Weinhauer (2003) and Arifeen et al. (2009).


                                         3
formal health care during the spell. We can infer when the IMCI would instruct
a CHW to seek a formal medical evaluation by a doctor, based on the duration
and type of symptoms. The guidelines are quite intuitive – for example, they
prescribe immediate care if symptoms indicate a risk of malaria, but treat
coughing or diarrhea as harmless unless they last unusually long or cause
complications that point to pneumonia or dysentery. We observe children
whether or not they seek care and therefore can measure utilization relative
to the IMCI benchmark.
 The IMCI was developed by Unicef and the World Health Organization and
is used by public health systems and NGOs in over 80 countries, including
our partner NGO Mali Health, who based their CHW training on it. Clas-
sifying when an illness spell requires evaluation by a doctor according to the
IMCI allows an assessment of severity from a medical perspective and there-
fore relative value of care: a policymaker who wants to improve population
health should aim to increase care seeking on “care-required” days over other
days, all else equal. It also measures whether providing information about care
seeking guidelines actually increases parents’ adherence to these guidelines. In
this manner, the IMCI provides a benchmark for assessing the targeting of
care, given that the optimal solution to the social planner’s problem cannot
be directly observed (see also Section 3 for a discussion).
 We organize our empirical results with the help of a framework that treats
care seeking as a stopping problem, where parents decide to seek care during
a given illness spell conditional on spell length and symptoms. This captures
the intuition that a set of initial symptoms will become more concerning as
the illness persists or new symptoms arise, and eventually lead to a clinic visit.

                                        4
The value of waiting instead of seeking care immediately lies in the possibility
of avoiding a visit altogether. The policy maker solves a similar optimization
problem, but may disagree at what point care should be sought, e.g. because
of a different cost-benefit trade-off or a better understanding of the illness.
Parents may therefore overuse care relative to the policymaker’s assessment
by going to the doctor too early, before care is required, or underuse it, by
going too late. Subsidies will lower the cost threshold and lead to earlier (more)
care. Information can improve the allocation of care, as long as the parents’
cost-benefit assessment is aligned with that of the policymaker.
 We observe 3747 illness spells – more than two per child – with an average
length of 6.4 days, of which 3.0 are “care-required” days. We find that overuse
according to IMCI recommendations is rare in the control group, even though
there is surprisingly large scope: about half of all illness spells end without
ever requiring care.3 By contrast, underuse is rampant, with a probability
of care-seeking on a care-required day of six percent or lower. As a result,
only 11% of spells that reach “care required” status actually receive formal
care. Nonetheless, the probability of care is on average more than five times
higher on days when care is required than when it is not: parents do seem to
differentiate illness severity as embodied in the IMCI care recommendations,
even though they delay the doctor visit longer than recommended.
 The subsidy increased visits to the doctor by at least 175% and the total
   3
     Note that the idea of seeing the doctor too early hinges on the possibility that an illness
may pass harmlessly without entering “care required” status. Nonetheless, a spell that enters
“care required” status may end without treatment. “Care required” should be interpreted
to mean that the risk of harm or death is sufficiently high to warrant the cost of care, even
if the illness may ex post pass without apparent harm. Indeed, more than 85% of illness
spells in our sample end despite not receiving any formal care, including those that would
have required care according to the IMCI.


                                               5
value of received care by over 80%. Importantly, this increase does not crowd
out private spending on acute care, and the bulk of the increase occurs on
care-required days. Again, parent behavior emulates the IMCI assessment of
relative need for care, and it means that more than 70% of the doctor visits
that the subsidy generates are IMCI recommended.
 In our sample, providing health information to parents through CHW visits
has no aggregate effects on demand, consistent with parents already being able
to distinguish need for care. However, exploratory analysis uncovers substan-
tial heterogeneity. For children up to one year old who receive the subsidy,
CHW visits lead to a significant additional reduction in underuse of care (IMCI
based). This is important because it suggests that the subsidy and informa-
tion interventions complement each other in improving care for the age group
with the highest mortality risk.
 The CHW may simply alert parents that the subsidy is available, but another
possibility suggested by our model is an interaction of improved information
with the cost threshold for care that the caretakers apply. The difference
in care seeking when care is vs. is not required is smaller for the youngest
children (under one year of age), in line with qualitative evidence on the greater
difficulty parents have in recognizing illness in young children (Charlet et al.
(2017)). The CHW effect is also lower without the subsidy and exhibits a
significant negative age gradient, suggesting greater underuse for older, lower-
risk children for whom prices remain high. Taken together, it is possible that
the CHW are helping parents more accurately assess need for care according
to IMCI guidelines, but depending on how closely parents’ preferences are
aligned with these guidelines, this may reduce or increase underuse.

                                        6
 Overall, our results are consistent with cost constituting a primary barrier
to appropriate care in our population. The results are encouraging for the
planned roll-out of subsidized care in Mali, although they predict a surge
in health care utilization that could create bottlenecks in supply. Our work
complements a study in Bamako by the NGO Muso and the Malian Ministry
of Health in 2008-2015, combining CHW, user fee removal, and health system
strengthening measures (Johnson et al. (2013, 2018)). Cross-sectional surveys
find improvements in incidence of, and care seeking for, febrile illness, as well as
reductions in mortality after the intervention was introduced. Our experiment
can directly attribute health behavior changes to the interventions, pointing
to formal care demand as a potential pathway for the health outcome effects,
and highlighting the differential contributions of the CHW and the subsidy.
The ambiguous CHW effects call for further study, given the important role
of CHW interventions in health policy in general and in Mali in particular.
 Throughout the paper, we apply hazard models to spell data collected at
home in order to estimate the effects of our interventions on care seeking
conditional on health status. In section 5.4, we discuss the advantages of this
approach for making out-of-sample predictions and for understanding selection
effects that would arise in demand analysis based on health system contacts
only (e.g. records of health consultations or health insurance events).4 For
example, we show that children with the subsidy who visit the clinic have
been actually sick weakly longer, because the treatment succeeds at bringing
   4
    Since sickness spells can and do frequently end without treatment, selection bias in such
data is pervasive. In most other spell data economists use, e.g. unemployment spells, there
are no equivalent sources of censoring.




                                             7
children to the clinic whose parents tend to wait long before seeking care.5
 Our paper contributes to the literature on children’s acute health care in
low-income countries, aided by novel daily panel data. Research on acute care
is still fairly rare (but see Powell-Jackson et al. (2014)).6 The problem of
underuse is thought to be particularly severe for preventive goods, where the
benefits are nearly universal, but also often abstract or far off. Our results
suggest that underuse is similarly pervasive at the extensive margin of acute
care – the decision to visit a doctor – and that subsidies can substantially
improve utilization. This complements Cohen et al. (2015) and Lopez et al.
(2020), who examine the effect of subsidies for malaria drugs at the point
of purchase. There are significant rates of overtreatment that increase with
subsidization. This points to a challenging twin problem, where underuse at
the extensive margin, in the form of too few contacts with the health care
system, can occur simultaneously with overuse at the intensive margin, in the
form of overmedication conditional on such a contact, especially when there
are no cost controls such as those implemented by Mali Health.7
   5
     One of the few data sets we are aware of that has panel spell data, the 1987 National
Medical Expenditure Survey for the United States, includes only spells that lead to medical
care use or work absenteeism. Gilleskie (1998) recognizes the selection problem but can
only address it by estimating the probabilities of unobserved health events parametrically.
With administrative data e.g. from insurance claims (Manning et al. (1987); Finkelstein
et al. (2012)), illness spells can only be studied indirectly, by grouping claims into treat-
ment episodes, see Stoddart and Barer (1981); Hornbeck et al. (1985); Keeler et al. (1988);
Santos Silva and Windmeijer (2001). A public health literature studies the covariates of the
delay in seeking care for specific diseases (see Storla et al. (2008) and Nguyen et al. (2010)
for reviews).
   6
     Das and Sánchez-Páramo (2003) have weekly data on illness and document the vari-
ation in demand that comes from differential illness incidence, but do not study demand
interventions.
   7
     Das et al. (2016); Currie et al. (2011) conduct patient audit studies and show similarly
low-quality diagnostic and treatment choices, which often result in mis-targeted treatment.
Also related is work by Fischer et al. (2014), who test for learning and reference dependence
as they vary prices in door-to-door sales of over-the-counter drugs (but do not focus on over-


                                              8
 Section 2 describes the policy environment and the RCT intervention and
survey design. Section 3 discusses our use of the IMCI measures. Section 4
introduces a framework for demand for acute care and the effects of subsidy and
information policies. Section 5 describes the empirical results and 6 concludes.

2    Study Background and Data Collection

The Malian health care system builds on a network of community health clin-
ics or centres de santé communautaires (CSComs). A CSCom has typically
one or two trained doctors on duty and sells prescribed medications through
an attached pharmacy. At the time of the study, CSCom care was partly
subsidized, but primarily funded by user fees, in accord with the Bamako Ini-
tiative from 1987, which advocates self-sustaining, decentralized public health
care across West Africa. The public health care system is flanked by a private
formal sector with higher prices, and informal sources such as market stalls
(see below).
 This study was conducted in Sikoro, a peri-urban area of the capital of Mali,
Bamako. Most roads are unpaved, and many dwellings are not connected
to the water supply or the sewage system. In and after the wet season the
incidence of waterborne diseases (diarrhea) and malaria is highest. Mali has
high rates of maternal and child mortality, especially in rural areas. Poor urban
areas have better health care access and lower rates of mortality, but still often
lack basic health services. Mali’s urban areas resemble those elsewhere in West
and Sub-Saharan Africa, although literacy rates in Mali tend to be lower and
fertility and child mortality rates higher (see table C.1 in the online appendix).
or underuse).



                                        9
This makes child health interventions particularly important here.

2.1    Study Design

Mali Health started the Action for Health (AfH) program in 2010 in collabo-
ration with the two local clinics (CSComs) in Sikoro. The full AfH program
combines subsidized health care and community healthworker visits (CHW).
 The subsidy is administered via a personalized card that entitles the child to
unlimited free consultations at a partner clinic, and free treatment and medica-
tion for any illness due to diarrhea/malnutrition, malaria, vaccine-preventable
diseases, and respiratory infection (together causing the vast majority of child
deaths outside of neonatal conditions). Families have to cover remaining ex-
penses for example for services that are not part of the standard treatment
course for a given diagnosis.
 Families in the CHW program received biweekly CHW visits. The CHW
tracks a set of simple indicators including symptoms and malnutrition mea-
sures, teaches basic health practices, and advises families when to visit a doc-
tor or alternatively what care to provide at home, such as rehydration in mild
cases of diarrhea. CHWs are recruited locally and their training follows the
C-IMCI (Rosales and Weinhauer (2003)), a set of guidelines for community
healthworkers that incorporates the WHO’s and UNICEF’s “Integrated Man-
agement of Childhood Illness” recommendations for when to refer children to
formal health care (WHO (2014); WHO Department of Child and Adolescent
Health and Development (2005); see also below).8
   8
    Outside the focus of this paper, CHWs also monitor nutrition and advise households
on preventive care in general. They also dispense water chlorination tablets for households
with no access to clean water.




                                            10
 The research design took advantage of the second planned roll-out wave of
AfH. Mali Health conducted a census in their expansion area in mid-2012 to
enumerate all families with children under five years of age (or a pregnant
mother) who satisfied a proxy-means test designed to select approximately
the poorest third of households. Data were collected in two survey rounds in
2012 and 2013 during the rainy season (September-November). All households
identified in the census that were found at baseline were included in the random
assignment to the different treatment groups and revisited in 2013.
 In this study we use demographic, location, and household asset data col-
lected in the baseline survey, and daily health diary data on children collected
at follow-up. The unit of randomization is the compound. A compound may
house more than one household, and typically consists of a few rooms around
a common courtyard with shared latrines and other facilities. After stratify-
ing compounds by average household assets, number of eligible children, and
location, each was assigned to the subsidy-only group, the CHW-only group,
the full treatment group, or the control. The healthworkers in the two CHW
treatment groups were trained and managed separately, to avoid spillovers.9
 The focus of this study is to understand the demand for health care of the
typical quality a local medical professional (trained and certified in-country)
can provide, holding the supply side constant. There are two supply-side issues
   9
    It was not possible to assign healthworkers across the entire intervention area while
keeping their travel distances manageable. Mali Health therefore paired healthworkers of
similar experience and quality, and one of each pair was assigned to the healthworker only
or the healthworker and subsidy group. The pairs were then assigned in an overlapping
pattern by stratum. For example, pair 1 and pair 2 were each randomly assigned half of the
families in the same stratum, and half of the families in a different stratum, each shared with
another pair of healthworkers. In this manner, each stratum was assigned four healthworkers
(two in each treatment group), the quality of the healthworkers in the treatment groups was
matched, and no two strata had identical sets of healthworkers.


                                             11
we do not address here. The first is the problem of sometimes low quality of
care in public health care systems in developing countries, which may mean
in practice that patients receive care from an untrained individual or not at
all (e.g. if the doctor is actually absent). Such quality problems would of
course affect how desirable the use of the health care system is both from
the parents’ and the policymaker’s point of view (see e.g. Das (2011); Das
and Hammer (2005, 2007, 2014); Das et al. (2016); Das and Sohnesen (2006);
Leonard and Masatu (2010); Leonard et al. (2002)). The clinics in this study
were financially supported by Mali Health during the study and subject to
regular quality control. We assume that the standard of care provided at these
clinics reflects the level of care that the WHO expects when they formulate
treatment recommendations for low-income countries.
 The second issue is demand inducement and moral hazard by doctors, com-
mon concerns in particular when care is subsidized. Mali Health controlled
costs in the subsidy group using treatment guidelines for common diagnoses,
spot checks on diagnosis and prescriptions using bills submitted to Mali Health,
and the clinic’s treatment records and accounting. The subsidy did not change
the average value of services received per visit (see section 2.3).

2.2   Sample Population

Mali Health identified 1804 eligible children in the census. At baseline, 1732
children were included in the study. The roll-out of AfH started three months
after the baseline in early 2013. By the second survey round in Fall 2013, there
were three refusals and an additional 162 children could not be surveyed, in-




                                       12
cluding five who passed away.10 The remaining sample in 2013 contains 1567
children from 990 households in 642 compounds. Attrition between baseline
survey (and treatment assignment) and follow-up is 8-11%; the group differ-
ences are not significant. At follow-up, 201 children were newly reported by
the caretakers of existing study children. Newborn children in the treatment
households were enrolled in AfH either by the healthworkers, or by a program
officer who visited study households every three months. Since the caretakers’
treatment choices are our outcome of interest, we consider these children part
of the sample. We only included newly reported children 3-12 months old.
Table C.2 in the online appendix shows the sample over the study period.
 Table 1 shows sample characteristics and balance. We find few significant
differences between the treatment groups. There are on average more than
six people in a household, due in part to polygamous marriages. About half
of the household heads are literate and 73% speak the lingua franca Bam-
bara. Households report on average about US$6600 in assets, with an owner
occupancy rate of 41%. Only 13% of household heads earn a regular salary.
The closest source of formal health care is on average 546 meters away (less
than 0.4 mile). Average child age is about three years. The study children are
well below age-typical weight for height (normalized using W.H.O. reference
distributions). We control for these covariates in the analyses below, except
weight for height, which may be an outcome of parents’ health behaviors.
  10
     Our sample population moves a fair amount, but households that relocated inside the
program area were followed and remained eligible for Action for Health; only children who
left the area entirely or could not be tracked are not included.




                                           13
2.3    Health Calendar

We collected daily information on symptoms, health consultations, and med-
ications taken in weekly visits over the course of nine weeks from the child’s
primary caretaker (usually the mother). Caretakers could use pictorial diaries
(figure B.1 in the online appendix) to mark off symptoms and health-related
events. The surveyor then reconstructed the child’s health history during the
visit. There was a concern that parents would respond to the observation by
the surveyors, for example by watching their children more closely. Ultimately
this drawback was outweighed by the benefits of detailed coverage without
the problems associated with long recall periods (Das et al., 2011). The sur-
veyors were carefully instructed not to comment on the child’s health, and to
emphasize that they had no medical qualification. While we cannot rule out
all Hawthorne effects, we believe our precautions mitigated them, combined
with the fact that the data collection had been piloted over six weeks after the
baseline in 2012, thus creating some habituation.11

Symptom Calendar. The list of symptoms was designed in collaboration
with Mali Health staff and based on the C-IMCI. The C-IMCI is designed
for use by CHW who have no prior medical experience, and is used in Mali
Health’s own healthworker training. It consists of simple rule charts that use
important and easy-to-spot symptoms to broadly classify a child’s illness and
  11
    Since the survey took place towards the end of the rainy season, illness rates were
declining over the course of the panel, so we cannot assess if a time trend is due to a fading
Hawthorne effect. However, in the baseline pilot of the health diaries, initially elevated
symptom reporting rates stabilized at a lower level after the first week of surveying and
then declined at a slower and more continuous rate (data available on request). This is
consistent with data on doctors, who rapidly return to unobserved length and quality of
doctor-patient interactions even when observed by other doctors (Das et al., 2008).



                                             14
establish need for care. These properties mean we can collect symptom reports
from mothers and surveyors who are not medically trained, and map them
into an IMCI care seeking recommendation.The survey explicitly asks about
convulsions, fits, or spasms; lethargy or unconsciousness; inability to drink;
vomiting; coughing; difficulty breathing; more than three loose stools; blood
in the stool; sunken eyes; and unusually hot (cold) skin. We constructed four
more symptoms from the “other” category: skin conditions (e.g. spots, rash,
itch), cold symptoms, ear pain, and injuries. We observe on average 60 days
per child, and children exhibited symptoms on average on 18 days. Cold symp-
toms, coughing, and unusually hot skin were most frequent. Section 3 below
gives more detail on how symptoms translate into illness spells. Appendix B
provides a description of how the health calendar information was collected
and shows recorded symptom days by treatment group in table B.1.

Consultations. At each surveyor visit, mothers were asked if they discussed
their child’s health with anyone in the preceding week, and the surveyors
recorded role or occupation of the person seen, facility, waiting time, treat-
ments received and costs incurred.12
 Table 2 provides an overview of these health care “consultations”. The 735 for-
mal consultations in our data are provided by CSComs and CSREFs (reference
hospitals), and private doctors, hospitals, clinics, and pharmacies. Parents oc-
casionally see several providers in the same day, for example when they get
a prescription at a clinic and then purchase medication elsewhere. We group
  12
    In order to facilitate the recording of drugs prescribed, purchased, and taken, mothers
were asked to keep the packaging of any medications their child took. The surveyors could
search and fill the drug from a database of brand and generic names of about 300 medications
commonly sold in Mali.



                                            15
visit records that are linked in this manner and classify them by the “most
formal” consultation in the group. The value of care is approximated using
the full price of care.13 Note that the sizable difference in costs and value of
care even without a subsidy is due to families seeking outside help to pay for
the child’s care, e.g. from relatives or friends.
 The table shows that the subsidy program reduced the average cost of a visit
to the household by 71%, from CFA 2850 to 933 (476 CFA equaled 1 USD
in 2013). Among the formal-care visits in the control and CHW-only groups,
14% and 12% were reported as free, vs. 70% and 77% in the subsidy groups,
respectively. Parents paid if they visited other providers than the two partner
clinics or did not claim the AfH benefit, or the child received services not
covered under AfH.
 CSComs provide similar value per visit with and without the subsidy, sug-
gesting that the type and quality of health care received, conditional on vis-
iting a clinic, is similar. A priori, one might expect that doctors respond to
the subsidy by raising treatment costs or prescribing unnecessary treatments.
This is not the case here, likely due to Mali Health’s quality and cost control,
implemented after initial increases in costs during earlier program roll-outs.
 Wait times at the CSCom are somewhat longer with the subsidy. This could
be because subsidized patients have a less acute need for care, or because they
receive medical tests such as malaria tests, which require waiting. Another
  13
    Value of care is imputed using prices recorded for consultations and medications where
the respondent reported having paid the full price themselves. We use these prices combined
with provider information, brand name, and point of purchase to construct the median price
by source. If the respondent reported another person paying for care or was in the subsidy
group, we compare the sum of median prices for the services received with the private
expenses the respondent reported and use the higher value. This reflects the cost of care to
the family, the NGO, or a third party.


                                            16
possibility is that they see different personnel, or even are made to wait longer
than other patients. For this the clinic staff would need to know in advance of
the consultation which patients have the subsidy. Given the good relationship
of the cooperating clinics with Mali Health, the income stream AfH patients
provide, and Mali Health’s efforts at quality control, we deem this possibility
unlikely, although we cannot rule it out empirically.
 Private formal providers are pricier than CSComs and have longer wait times.
Informal visits and purchases are common and come at low monetary and
almost no time cost. The low prices (under 50 US cents) suggest that they
also provide little value. Typical informal sources are peddlers, stalls, or shops,
healers, or Islamic marabouts; individuals who are not trained to diagnose
illness. For our analysis, we treat these as home care. We also assign pharmacy
purchases not associated with a doctor visit to the “informal” category because
pharmacists do not carry out a diagnosis. Pharmacy visits alone are also not
subsidized by AfH or promoted by the CHW.
 Lastly, we recorded CHW visits if they were mentioned among the consul-
tations (we did not explicitly ask about them to prevent reporting bias from
associating the survey with Mali Health, for example if parents believe we
are monitoring the healthworkers). Parents report 0.46 visits per child in the
CHW-only group, but 1.64 in the full program group (24% and 43% of chil-
dren report any CHW visit). Similarly, at the start of the survey, parents in
the CHW-only and the subsidy-only groups report much lower enrollment in
a health program than the full program group. However, we find no differen-
tial effects of the full program on health care use (see below). Together, these
numbers suggest underreporting that is higher in the partial treatment groups.

                                        17
3        IMCI Classification and Illness Spells

Throughout the paper, we identify “overuse” and “underuse” of care relative
to the IMCI guidelines proposed by the WHO. The C-IMCI is part of the
community and family arm of the IMCI itself, the WHO’s and UNICEF’s pri-
mary children’s health program, which has been adopted by over 80 countries
(Bryce et al. (2004)). The C-IMCI goes through a set of simple checks in
order to spot danger signs and classify symptoms into gastrointestinal illness,
respiratory illness, malaria, and so on. It targets the causes of about 70% of
child deaths.14 All diagnostic tests can be operationalized in the field. The
effectiveness of these tests in identifying the underlying disease was rigorously
tested (see Gove (1997) and the studies therein).
 The C-IMCI algorithms make intuitive recommendations for when care should
be sought based on the duration and severity of symptoms during an illness
spell. For example, cough and cold symptoms typically point to a viral infec-
tion and do not require medical care, unless they persist for 14 days or longer,
on suspicion of a more serious illness like tuberculosis. Diarrhea alone can be
treated with home remedies, but in the presence of sunken eyes (dehydration)
or blood in the stool (dysentery) requires immediate attention. Applying these
rules to symptom records, every day in an illness spell can be classified as ei-
ther an “early” day (i.e. before care should be sought) or a “care required” day
(on or after the day at which care should be sought). Appendix B describes
the approach in detail. We focus on pre-care spell days up to the first con-
    14
    The five main causes are diarrhea, pneumonia, measles, malaria, and malnutrition.
According to Unicef estimates, around 50,000 children under five died in Mali in 2017, 17%
of whom died of pneumonia, 10% of diarrhea, and 14% of malaria.



                                           18
tact with the health system, since symptoms might be affected by treatment.
The C-IMCI is explicitly tailored to a developing-country context, and we as-
sume that its care-seeking recommendations take into account factors like the
expected quality of care and resource constraints.15
 What justifies the use of the IMCI to assess need for care? The object of
interest is the allocation of acute care. Demand for children’s health care is
likely distorted by market imperfections such as credit constraints. Parents
may not fully internalize their children’s welfare or externalities due to con-
tagion. Moreover, households may not be able to interpret symptoms and
assess the value of visiting the doctor in different health states. All of these
imply that parents’ health care demand is not socially (or even privately) op-
timal, and there is a role for policy intervention. However, solving the social
planner’s problem of optimal resource allocation requires too many parame-
ters for which there are no good estimates, from the production-side trade-off
between supplying health care and other goods, to the true preferences of the
household.
 Given the correlation of the IMCI diagnostic algorithms with serious illness,
they provide some guidance for assessing welfare. First, they constitute an
approximate solution to an optimization problem: conditional on the limited
information an outside observer can use to make a recommendation, for an
organization that is prepared to devote resources to children’s health, which
mapping from health states to care utilization has the greatest effects on ex-
  15
    Mayo Clinic recommendations for children in the United States, for example, have lower
thresholds for seeking medical advice. They advise to call a doctor after 24 hours of diarrhea,
and state on fever: “When in doubt, go ahead and call your child’s doctor, whether you think
your baby’s temperature is abnormally high or abnormally low.” (Mayo Clinic (2014, 2016)).



                                              19
pected mortality and morbidity? At a minimum, this provides a ranking ; from
a medical standpoint, care in the “care required” state is more valuable than
otherwise, even if the total amount of care is not globally optimal.
 Many policymakers may in addition agree that all children should receive care
when the IMCI recommends it. For example, a child with malaria symptoms
should be seen by a doctor regardless of parental preferences or wealth. More-
over, while parents may observe symptoms that require medical treatment
but are not covered by the IMCI, such as injuries, it is less likely that parents
have private information that indicates that no care is needed when the IMCI
recommends care. The “care required” designation may then be considered a
lower bound for underuse.
 Last but not least, comparing parents’ behavior to the IMCI recommenda-
tions indicates to what degree they follow the CHW’s care seeking advice.
Thus, even if they do not characterize the social optimum, the WHO guide-
lines capture important information, giving them a role in evaluating the effects
of health care policies on the targeting and level of health care. Studying de-
mand in detail complements measuring mortality or longterm health outcomes.
While ultimately the outcome of interest, these depend on many factors be-
sides acute care and do not give us direct information about actual behavior
and use of health care in response to policy.

Classifying illness spell days. As described, we group contiguous symp-
tom days into illness spells and classify each spell day as “early” or “care re-
quired” according to the IMCI, focusing on “pre-care” spells up to the first
formal consultation or recovery. For example, consider a child who is cough-



                                       20
ing for three days and takes cough drops from a peddler, then develops a
fever on day three, and is prescribed paracetamol at the CSCom on day five;
symptoms subside after day eight. This is a five day pre-care spell with three
care-required days (days 3 to 5). If the child had not received care, it would
be an eight-day spell.
 Table 3 shows untreated days per child (possibly in several illness spells),
and the percentage of days on which each symptom is present, in total and
split into early and care-required days. On average, 16.3 of 59.8 observed days
per child (27%) were (pre-care) illness spell days, and on six of those there
was a need for care. Convulsions, lethargy, inability to drink, and vomiting
should trigger immediate care, so they always occur on care-required days.
The biggest contributor to need for care is fever (unusually hot skin). While
cold symptoms and cough rarely in themselves indicate a need for care, they
co-occur on many “care required” days. The data show that the children in
our sample are often ill with symptoms that point to dangerous conditions like
malaria, but also often experience harmless symptoms, so they are in need of
a doctor evaluation only 37% of the time.
 The symptoms we collected do not cover all possible illnesses, but focus on
conditions that mothers and surveyors can easily recognize. As an example,
breaths per minute and “chest in-drawing” are IMCI indicators of respiratory
illness, but require experience and training (and a stopwatch) to measure.
We also interpret the IMCI guidelines conservatively, for example for diarrhea
(where the recommendation is for a CHW to return to the family after five
days to check on the child’s status, but we assume a need for care only when
complications arise). Moreover, the IMCI itself is not exhaustive. As a result,

                                      21
our classification as “care required” contains noise and generally represents a
lower bound for need for care.
 Appendix Table B.2 summarizes the observed pre-care spells. Due to spells
recorded in the first week of the survey or after survey interruptions, 464 illness
spells are potentially left-censored, that is, they may have started before the
first recorded day. We exclude these spells from our main analysis.16 This
leaves us with 3747 spells, with an average length of 6.43 days (6.92 st. dev.).
Within these spells, there are on average 3.4 early days and 3.0 care-required
days. Only 503 spells, or 13.4%, end with a consultation; in all other spells,
the child eventually recovers from the illness unaided. Spontaneous recovery
is an important feature of the data and of acute illness in general: section 4
describes that “waiting out” an illness can be optimal if unaided recovery is
possible, and in this case it may be overuse to visit a doctor early. By contrast,
if recovery could only occur with medical care, “too early” use of acute care
is not an issue; care should always be provided immediately, and any delay
would constitute underuse.17
 Spell incidence is fairly similar in the four treatment groups; between 907
and 967 spells (see table 5 below), despite the demand effects we document
in section 5. A probable reason is that the main causes of acute illness –
malaria, respiratory, and gastrointestinal – are infectious, and the children
in AfH constitute only a small percentage of the population (about 20% of
  16
     Left-censoring means that need for care cannot be determined, because the day of the
spell is unknown. Left-censored spells are on average 10.2 days long. Our main results are
robust to treating these spells as not left-censored (see also below).
  17
     We discuss below that recovery is a form of right-censoring, because we do not observe
when parents would have sought care if the child had not recovered. Right-censoring is here
an integral part of the data generating process, and important for understanding health care
demand, especially because a large number of spells do not end in formal care.


                                            22
people in the study households, which in turn make up a fraction of the local
population), so the program is unlikely to reduce infection rates.18
 Below, we analyze care seeking conditional on symptoms, so even differential
incidence would not be a concern, as long as conditional health care choices
remain unaffected. If we believe that illness incidence does affect conditional
demand, it is an advantage for comparing treatment groups that spell numbers
do not significantly differ between them, but external validity would be lower.
In Dean et al. (2017), we document a decrease in malnutrition rates and acute
diarrhea incidence among children who received the full program by the second
year, suggesting that incidence (preventive) effects of AfH may become more
important in the long term.

4        Demand for Acute Health Care

We sketch a simple model for the decision to seek health care in response to
acute illness, in order to derive notions of overuse and underuse and to make
predictions for the impact of AfH policies on health care demand (formal state-
ments in Appendix A). Our framework is based on two central observations.
First, demand for acute care arises in response to a negative health shock that
causes discomfort – and risk of harm – to the child until it passes. Second,
there is learning about the severity of an illness over time. For many illness
spells, it is likely that the child will recover without the aid of a clinic visit,
and it is worth tolerating some disutility from sickness while waiting. As time
goes by, the probability of recovery without treatment decreases, and further
    18
     This is consistent with the balanced symptom incidence by treatment group, reported
in table B.1 in the Appendix. Treatment rates remain overall low and treatment occurs
fairly late in the spell, so the program has only a small effect on pre-care spell length and
none on overall spell length (see also below for a discussion in the context of censoring).


                                            23
suffering (or worse) can only be avoided by seeking care. Thus, the demand
decision becomes when during an illness spell to visit a formal care provider.
 Formally, denote the parent’s belief that the illness will continue another day
by πt . Each day t of the spell, the parent decides to either visit a provider who
can evaluate and treat the child, or wait another day, and with probability
(1 − πt ) the illness passes on its own. With uncertainty about the underlying
illness, πt is increasing in illness days t: continued symptomaticity reveals an
illness with lower probability of recovery (Lemma 1). This implies a decreasing
value of postponing treatment, since it becomes less likely that the child will
recover on their own. As a result, parents optimally follow a simple cut-off
strategy: seek care when belief πt exceeds some threshold K (Proposition 1).
K depends on the benefits as well as the monetary and non-monetary costs of
care C . C is increasing in the (expected) price of care.
 We assume that the policymaker faces a similar stopping problem, although
with different parameters π ∗ and K ∗ . This agrees with expert care seeking
recommendations such as the IMCI (see previous section), which take into
account duration of symptoms as an indicator for health care needs. Such
recommendations cannot be explained without learning and the possibility of
                                  ∗
recovering without treatment: if πt (or πt ) was constant over time, the choice
to seek care would be a simple yes/no decision on the first day of illness,
because getting earlier care for the same illness is always better.

Effect of a subsidy: Consistent with current policy consensus, we assume
that the parents’ treatment threshold K is weakly greater than that of the
policymaker K ∗ . The policymaker may evaluate the financial costs differently



                                       24
from a parent with credit constraints, or she may incorporate the long-run
human capital benefits or reduced infection risk for others. This case is illus-
trated in Panel A of figure 1 (numerical details in the appendix). Since there
is learning over time, Bayesian updating implies that πt < πt+1 (solid black
line, parent and policymaker assumed to share beliefs π ). Conditional on not
having recovered, the policymaker would like the child to receive care in period
t + 1 when πt+1 > K ∗ , but since πt+1 < K , the parent will not take the child
to the doctor in t + 1. This is an incidence of underuse, or visiting the doctor
too late relative to policymaker preferences.
 A subsidy lowers the care-seeking threshold of parents by reducing the treat-
ment cost C . Panel A illustrates how this (weakly) increases utilization in all
periods. However, if K is reduced by too much, for example to K , overuse,
or early care, may now occur (since K > πt > K ). In a heterogeneous
population, a subsidy policy may eliminate some, but not all, underuse, yet
simultaneously create some overuse.

Effect of information: Disagreement may additionally occur due to dif-
                                                                  ∗
ferences in beliefs about the recovery probability, so that πt = πt . If these
differences are due to the parents’ poor information, a situation might arise as
in Panel B of figure 1: the policymaker can distinguish a more severe and a less
severe illness (π ∗S and π ∗L ) while the parent holds average beliefs π . Exam-
ples are diarrhea spells with or without signs of dehydration, or a generalized
fever vs. a fever combined with symptoms that point to malaria (see also ap-
pendix). Lack of information can lead to both overuse and underuse, because
parents take the same action regardless of illness severity. For example, at K



                                      25
the parent will always seek care in period t + 1 – but for many thresholds K ∗ ,
including K , the policymaker would prefer care for the severe illness in t, but
no care for the less severe illness in t + 1.
 Subsidies cannot address simultaneous overuse and underuse from poor infor-
mation, because they will only monotonically increase utilization. By contrast,
a successful information policy may improve allocative efficiency. Take behav-
ior at threshold K ∗ : an uninformed parent with belief π will always seek care
in t and t + 1, but an informed parent will exactly share policymaker prefer-
ences and delay care in period t in the case of the less severe illness. Thus, if
the CHW intervention aligns the beliefs of the parents with those of the policy-
maker it may complement the subsidy by directing care (only) to the children
who need it.
 To summarize, if the main barrier to a good allocation of care is high cost,
parents in the control group will seek care primarily on care-required days, but
we will see significant underuse. If poor information is an important allocative
barrier, then care seeking will be only weakly correlated with “care required”
status and we may see both overuse and underuse (observation 1). A sub-
sidy will increase utilization, thus decreasing underuse, but also potentially
increasing overuse (observation 2). If the subsidy aligns the cost thresholds K
and K ∗ and the CHW intervention aligns beliefs π and π ∗ , then the two inter-
ventions may complement each other in decreasing both underuse and overuse
(observation 3). Observation 3 is the motivation for combining subsidies with
information policies, delivered e.g. through CHW visits, as in AfH: subsidized
care gives families access to health services, while better information ensures
that they make efficient use of this benefit.

                                        26
 A few notes are in order. First, the degree to which the effects above are
realized depends on the severity of the various constraints families experience,
and is therefore an empirical question. Second, as the reader may notice from
the graph, whether information unambiguously improves the allocation of care
from the policymaker perspective depends on whether the cost cutoffs K and
K ∗ are aligned. This is because upon receiving information that alters π ,
parents will align their behavior with their own cutoff K , not with K ∗ . This
may be a positive or negative change from the policymaker perspective. We
will return to this point when we discuss age-specific CHW effects in section
5.3. Last, in Action for Health (as in most CHW programs), the CHW do
more than just providing information about when to visit a clinic. However,
in the context of health care demand conditional on symptoms, we interpret
this as the main potential channel through which they may affect care seeking.

5     Results

We begin the empirical analysis by documenting the impact of subsidies and
CHW on unconditional utilization of acute care. We then study the impact
of our three treatments on health care use conditional on health care status.
Finally, we present exploratory analysis which examines the impact of each
treatment by age and discuss potential explanations.

5.1   Unconditional Health Care Utilization

Table 4 shows regressions of a set of per-child outcomes on dummies for the
three treatment groups along with a statistical test that the effects in the
subsidy-only group are the same as in the full treatment group. All standard
errors are clustered at the compound level. Columns (1) to (3) show the

                                      27
number of formal consultations per child, in total and split into CSCom and
other formal care. (4) to (6) report the effect of the program on private
expenditure per child, again first in total and then for CSCom and other
formal care separately. Column (7) reports wait times, and (8) and (9) report
total and CSCom value of care (see above) per child.
 Subsidies increase the number of formal care visits by 175-202 percent (0.38-
0.44 visits), and the total value of care consumed by 83-84 percent (CFA1386-
1409) relative to the control. The demand elasticity is clearly high: the subsidy
increases the number of CSCom visits per child by 263-312%, relative to the
decrease in the CSCom visit cost by 71% (table 2). There may be some
substitution out of other formal care, but overall demand for private care is
low, and none of the effects are significant. Remarkably, the subsidy has almost
no effect on families’ total health expenditure on their children. This implies
that there is no crowding out; the subsidies fully translate into additional
health care consumption. The subsidy groups also pay higher non-pecuniary
costs for their child’s health care, as evidenced by the increase in time spent
waiting for treatment by about 17 minutes per child.
 In stark contrast to the effects of the subsidy, the CHW intervention has no
average effects on unconditional health care use.

5.2   Health Care Use Conditional on Need for Care

While the unconditional effect of the subsidy is encouraging, it could be a real
increase in health care access for sick children but also unnecessary doctor
visits. The CHW may have no impact on care seeking at all, or improve
the allocation of care without changing overall use. To differentiate these



                                       28
possibilities we examine conditional demand.

Analysis. In order to understand the impact of each treatment on condi-
tional demand, we present three complementary pieces of analysis. First, table
5 shows disaggregated pre-care spells by treatment group and “care required”
status, and the number and proportion that receive formal care.19
 Table 5 is the joint outcome of illness incidence and demand, and does not
allow us to isolate the effect of the policy interventions on parents’ behavior.
Based on our model of care seeking, we therefore estimate the daily hazard of
formal care, that is, the daily probability that the parents seek care on pre-care
spell day t conditional on not having seen a doctor yet.
 Figure 2 shows the simple average hazard of seeking care on each spell day, by
treatment group and “care required” classification, estimated by regressing care
seeking on a full set of interacted indicators for day of the pre-care spell, care-
required status, and treatment group. Since this non-parametric approach
does not show treatment effects in a concise way, we additionally estimate
a Cox proportional hazard model. Formally, the hazard at t conditional on
covariates xit is
                                    h(t|xit ) = λ0 (t)exit β .
  19
     Table C.3 in the appendix shows a complete breakdown of all formal consultations during
and outside a spell. Only 5% of visits occur outside a spell, reported by mothers as follow-
ups, visits for prevention (e.g. vaccinations), or “other”. Of the remaining consultations, 88%
(or 494) constitute the first formal visit during a spell of illness. These “acute care” visits are
the target of the WHO care-seeking recommendations. Visits after a first acute visit cannot
directly be judged using the C-IMCI standards. However, almost all of them occur after a
spell entered “care required” status and when the child is still exhibiting symptoms. This
might indicate a more hands-on treatment approach by doctors when care is subsidized.
Anecdotally, CSCom doctors lament that it is very hard to convince patients to come back
after a first visit, so that they often prescribe treatments for a range of conditions at once.
We also know that patients often do not buy everything prescribed at a doctor visit. With
subsidized care, doctors may be able to ask the parents to return to try a different treatment
approach if the child’s condition has not improved.


                                               29
 This assumes an unrestricted baseline hazard λ0 (t), but the covariates shift
the probability of seeking care proportionally on every day t of the illness
spell.20 Treatment effects are captured by CHW and subsidy indicators and
their interaction. Table 6 reports the estimated hazard ratios (the proportional
increase in the probability of seeking care). Models (2), (4), and (6) control for
basic household characteristics. All regressions include stratum fixed effects.21
 Specifications (3) to (6) separately interact the CHW and the subsidy×CHW
indicators with an indicator for “care required” and “early” in order to under-
stand to what degree the CHW may lead parents to behave differently on early
and care-required days. If the CHW reduce both overuse and underuse, the
effect of the CHW should be negative on early, but positive on care-required
days. Models (5) and (6) additionally interact subsidy and care-required indi-
cators. In appendix table C.7 we replicate the same estimates but including
left-censored spells as if they were not censored as a robustness check; the
results are very similar.

Results The illness profiles in our sample provide significant scope for both
underuse and overuse (IMCI based). Table 5 shows that out of 955 spells in
the control (2.1 per child), half (476) eventually require care. On the other
hand, 479 spells (about one per child) never enter care-required status before
recovery. The parents’ decision is a dynamic problem: 210 spells enter care-
required status only later on. Put another way, out of 689 spells that initially
do not indicate a need for care, 30.5% eventually require care.
  20
     As we show in appendix figure C.1, the proportionality assumption delivers a good
description of our data.
  21
     Of the controls, only assets affect health care demand positively and significantly, while
the distance to the closest provider as a measure of non-monetary costs has the expected
(negative) sign, but is not significant.


                                             30
 We organize our results using the observations from Section 4. First off,
parents behave very differently on early and care-required days. From table
5, spells that enter care-required status are three times more likely to get care
than other spells. Table 6 shows that a parent in the control group is 5.3 to 5.7
times more likely to seek care when it is required according to the C-IMCI. In
figure 2, the probability of care-seeking on any “early” day is below one percent,
but much higher when care is required, although never higher than six percent.
As a result, there is significant underuse but relatively little overuse. Table 5
shows that 89% of spells that require care do not get it, whereas 4% of “early
spells” receive treatment.22 These results suggest that information is not a
primary barrier to care. The stark difference in care seeking by care-required
status implies that parents are able to gauge the relative seriousness of their
child’s illness, much more so than perhaps expected ex ante. High levels of
underuse coupled with low levels of overuse instead suggest a high cost barrier
(Observation 1).
 Next, we examine the impact of the subsidy (Observation 2). The subsidy
should decrease underuse, but may also increase overuse. In models (1) to (4)
in table 6, the subsidy increases the conditional probability of care-seeking sig-
nificantly and by at least 249 percent. We cannot reject that the proportional
impact of the subsidy is roughly the same on early and care-required days (see
tests Se = Sr and Se + CHW e + SCHW e = Sr + CHW r + SCHW r). How-
ever, given the higher baseline hazard, this implies that the absolute increase
in care seeking is much larger on care-required days. This can be seen clearly
  22
    The difference in the care-seeking probabilities per day and per spell is driven by the
ratio of early to care-required days within a spell. Note that we are considering pre-care
spells (not illness spells), so that the consultation always occurs on the last day of the spell.


                                               31
in Figure 2 and it suggests again that an important barrier to needed health
care is simply the cost of care.
 As a result, 9% and 11% percent of “early” spells end in a consultation in
the subsidy treatment groups, vs. 29 percent of “care-required” spells. By this
metric, the subsidy increases overuse by 5-7 percentage points and decreases
underuse by 18-19 percentage points. The last column of Table 5 also shows
overuse vs. use when care is required as a proportion of consultations. 25
percent of consultations on average occur on an early day in the control, and
thus constitute overuse. This number is essentially unchanged in the sub-
sidy groups. However, our analysis reveals that there is little room to reduce
overuse, given that the hazard of care on any given early day is so low. By
contrast, with 71% of “care-required spells” without a consultation, there is
substantial room to decrease underuse further.
 Again, unlike the encouraging findings about subsidizing care, we find little
impact of the CHW treatment on conditional demand, either in overall terms
or separately on early and care-required days. The hazard rates plotted in
Figure 2 look similar between the groups with and without the CHW, a result
confirmed by the estimates in table 6; on average the CHW do not appear to
improve the allocation of care. Given the extremely low probability of early
care seeking to begin with, it is perhaps not surprising that the CHW do
not significantly decrease overuse; however, they also does not significantly
increase care-seeking on care-required days. There is also little evidence of an
interaction between CHW and subsidies.




                                      32
5.3   Health Worker Effects by Child’s Age

Very young children are much more vulnerable to illness than older children.
The World Bank reports in 2013 that 63% of all under-five mortality in Mali
occurred in infants. Worldwide, this proportion ranged from 57-93%. This
suggests that the benefits from subsidizing acute health care may be greatest
very early in life. The benefits of providing information may also be greater,
both because parents may be less experienced, and because it is harder to
determine illness severity in a non-verbal infant. We therefore conduct an
exploratory analysis in which we examine treatment effects by age. Figure 3
shows nonparametric care seeking probabilities by treatment group and care-
required status for ages 0 to 5. In the appendix, we also estimate Cox hazard
models for the youngest and oldest children separately (table C.5) and with a
linear age variable interaction (table C.6).
 A few observations about age-specific effects emerge. First, care-seeking prob-
abilities tend to decrease in age, suggesting that parents see care as more
valuable for younger children. The low probability of care seeking for age-0
children in the “subsidy only” group, especially when care is required, is a stark
break from this pattern, compared to other ages and compared to the full pro-
gram group. In addition, the point estimates suggest that the difference in
care-seeking for age-0 children between early and care-required days is some-
what smaller than for other children. Lastly, for older children, care seeking
appears lower in the groups with than without CHW. The Cox hazard point
estimates support this (detailed discussion in the appendix). In particular, for
the youngest children, the increase in care-seeking in the group that receives



                                       33
both interventions is large and significant.
 The sample is small and power low, so these results are tentative. However,
they support the notion that parents are less certain about when care is re-
quired when their child is very young, that is, information barriers are higher.
They also support that the CHW complement the subsidy for the youngest
children, in the sense that the full program group seeks care significantly more
often than the other groups. We find that this increases overuse somewhat,
but primarily reduces underuse. This suggests that the healthworkers may
have an important role for the success of the program among the youngest
children, even if there are no average effects on all children.
 There are several possible explanations for these effects. One is prosaic: the
high care-seeking probability for age-0 children in the full program group could
simply be the result of better access to the subsidy. Sample households were
visited in three-months intervals by Mali Health employees, and during those
visits, any newborn children or children of program mothers who moved (back)
into the household were enrolled into Action for Health. It is possible that
parents in the subsidy-only group more often missed the enrollment visits, or
did not understand the child’s eligibility change. Some features of the design
guarded against this issue;23 nonetheless, it remains possible that the most
valuable information the CHW provide is not about symptoms, but about
subsidy eligibility.
 Alternatively, the results could suggest that the CHW do provide informa-
tion that affects care seeking, but the effect is ambiguous, consistent with
  23
   For example, all parents already had at least one child in the program and thus knew
how the program worked. Our analysis is also restricted to children four months and older
who should have received at least one enrollment visit.


                                           34
the framework we proposed earlier. When policymaker and parent share the
cost threshold K ∗ , more accurate beliefs π have the potential to reduce both
overuse and underuse. But the earlier graph also shows a potential negative
effect when K and K ∗ are not aligned. Specifically, a parent with threshold
K > K ∗ who learns to distinguish severe and less severe illness will actually
stop seeking care for the less severe case in t + 1, and thus increase underuse.
They use new information to align better with their own preferences based on
K , not K ∗ , which is here counterproductive for the goals of the policymaker.
 IMCI care-seeking rules (and thus K ∗ ) are the same for all children under
five. For the youngest children with the subsidy, parents’ cost threshold K may
be close to that of the policymaker, and they may also find it more difficult
to assess their child’s need for care. Thus, the healthworkers improve the
allocation and may in particular reduce underuse (observation 3). However, for
older children without the subsidy, parents may apply a much higher K than
the policymaker, and so the healthworkers may actually increase underuse, as
observed for the oldest children in the CHW-only group. Essentially, parents
use the additional information to decide against health care visits that the
policymaker deems necessary.
 In summary, our results suggest that the CHW might create complemen-
tarities with subsidies for the youngest children, who are most vulnerable to
adverse consequences of serious illness, and contribute most to continued high
child mortality rates. Given the widespread use of and advocacy for CHW
interventions, our findings point to a need for further research about both
the overall CHW effects and the exact mechanisms behind them. Our model
suggests some important caveats to health information policies that aim to

                                      35
change caretaker choices about acute care.

5.4   Advantages of estimating demand conditional on health status

Censoring and Selection Bias. Our data make a strong case for collecting
illness data in the home and analyzing spells with a hazard model, due to the
issue of censoring with spontaneous recovery: we do not observe how long
parents would have waited to see a doctor when the child gets better without
care. Spontaneous recovery is integral to the decision to delay care in the first
place – there would be otherwise no benefit to waiting. Even care-required
spells often end with unaided recovery, illustrating the challenge of improving
child mortality through acute care: we need to treat many spells in order to
prevent the small portion that may become truly dangerous to the child.
 One effect of censoring is that the observed pre-care spell length system-
atically underestimates true time-to-care, and effects on spell length do not
fully reflect changes in daily care seeking probability. For example, there will
be spells that are censored (ending without care) at high K , but uncensored
(ending in care) at low K (leading to increases in utilization). Table C.4 in
the appendix correspondingly shows that the subsidy reduces (pre-care) spell
length by only 14 percent, even though we found that it increases the care
seeking probability five-fold.
 This problem is exacerbated in data that contains only uncensored spells—
unfortunately often the only data we have. Table C.4 shows in fact that
the average uncensored spell in our sample is weakly longer in all treatment
groups than in the control. Our hazard estimates confirm that this is entirely
selection bias: the large number of additional visits in the subsidy groups tend



                                      36
to be from families who wait on average longer for care, and so they raise
the overall time to care in the group.24 Even recognizing that selection bias
may occur, data of uncensored spells only would not permit the researcher to
distinguish it e.g. from moral hazard (parents waiting longer before getting
care with the subsidy).25 In other words, where underuse is rampant but
responds strongly to an intervention, we show that data collected at the point
of use (e.g. during consultations with the health care provider, or from health
insurance events as in Gilleskie (1998)) are subject to selection effects so strong
that they reverse the estimated impact of the intervention on some aspects of
care-seeking behavior.

Application: Predicting Care-Seeking for Alternative Illness Envi-
ronments. Our study estimates demand for care conditional on detailed
health information. A benefit of this approach are more accurate out-of-sample
predictions, e.g. in other disease environments. In Appendix C.6, we demon-
strate this by predicting health care demand for a set of hemorrhagic fever
spells (Ebola and Marburg disease). This illustrates how records of tcondi-
tional care-seeking behavior might help improve estimates of undetected in-
fections and time-to-care for those who contract a specific illness, for instance
  24
     Similar effects arise when asking retrospective information on health for a specific period
of time. An example are the Demographic and Health Surveys (DHS). The DHS question-
naire asks about diarrhea and fever spells and treatment sought within the last two weeks.
Shorter spells are observed in full, but longer spells are likely to be censored by the visit
date. Moreover, detailed questions about treatment are asked only about spells that are
ongoing at the time of the survey. See Heger Boyle et al. (2018), especially questions 518
and 527 and following.
  25
     Note that the Cox model assumes proportionality and so does not permit for demand
changes that exhibit both selection into care and an increase in time-to-care under the
subsidy (which might occur with a decrease in care seeking on early spell days but an
increase later in the spell). However, figure 2 and the proportionality test in Appendix C.5
show that the proportionality assumption reflects the data well.



                                             37
in the initial phases of new outbreaks. In Mali, policymakers might be for
example interested in predicted behavior in the dry season, when respiratory
illness is more prevalent than malaria and diarrhea.

6        Conclusion

The prevailing view in international health policy today is that subsidies are
needed to close remaining gaps in access to care. The potential benefits are
large: by some estimates, improved access to primary care could prevent
29%–40% of post-neonatal deaths in developing countries. Less than 26%
of malaria cases in children are estimated to receive adequate treatment, and
411,000 children died of malaria in 2013 in Africa alone.26 Many countries have
begun to remove health care user fees for children, and Mali just announced
that it would adopt this policy by 2022. However, we have relatively sparse
evidence how this affects patterns of utilization.
 This study aims to open the “black box” of health care demand. We estimate
health care use conditional on health status to remove confounding effects
of illness incidence, and benchmark care-seeking against the need for care
according to WHO medical standards. By surveying all (potential) health
care users we can account for selection bias due to spontaneous recovery.
 Our results have very encouraging implications for the subsidy debate. First,
a primary barrier to effective care seeking in the urban population we study is
indeed the cost of care. Underuse is rampant: care is either sought late in the
spell or, in the majority of cases, not at all. The welfare cost of subsidies could
be prohibitive if they lead to substantial mis-targeting and overuse of health
    26
   United Nations Inter-Agency and Expert Group on MDG Indicators (2015); WHO
Global Malaria Programme (2015); Bhutta et al. (2008).


                                        38
care, risking overburdening the severely limited capacity of health care systems
in poor countries and thus diverting resources away from those most in need.
However, while subsidies nearly triple health care usage in our sample, they do
not result in substantial overuse, because the additional demand comes largely
from children for whom medical care is required according to WHO guidelines.
One reason is that parents have some ability to recognize differences in illness
severity and are nearly six times more likely to seek care on ‘care required’
days compared to ‘early’ days.
 The degree to which parents can spot illness suggests that information about
the health status of the child a priori can only have limited additional ef-
fects, and in aggregate this is what we find; the healthworkers do not affect
care-seeking probabilities significantly. However, the healthworkers do reduce
underuse among the most vulnerable (youngest) children who receive the sub-
sidy. This points to an role for policies that guide parents’ use of existing
subsidies for young children. At the same time, the age-specific effects suggest
that parents may make differential use of information they are given, depend-
ing on the cost threshold they apply in deciding about seeking care, and this
may not always be in the interest of the policymaker. This could be a poten-
tial caveat for health information policies in general. For Action for Health, it
may mean that the CHW should focus their advice on the youngest children
for whom the effects are likely highest. Further research could confirm these
findings and help understand the underlying mechanisms.




                                       39
References

Paul Adepoju. Mali announces far-reaching health reform. The Lancet, 393
(10177):1192, 2019.
John Akin, Nancy Birdsall, and David de Ferranti. Financing Health Services
in Developing Countries: An Agenda for Reform, volume 34. World Bank,
1987.
Shams E Arifeen, DM Emdadul Hoque, Tasnima Akter, Muntasirur Rahman,
Mohammad Enamul Hoque, Khadija Begum, et al. Effect of the integrated
management of childhood illness strategy on childhood mortality and nutri-
tion in a rural area in Bangladesh: a cluster randomised trial. The Lancet,
374(9687):393 – 403, 2009.
Nava Ashraf, James Berry, and Jesse M. Shapiro. Can higher prices stimu-
late product use? Evidence from a field experiment in Zambia. American
Economic Review, 100(5):2383–2413, 2010.
Nava Ashraf, B. Kelsey Jack, and Emir Kamenica. Information and subsidies:
Complements or substitutes? Journal of Economic Behavior and Organiza-
tion, 88:133–139, 2013.
S. Bhatt, D. J. Weiss, E. Cameron, D. Bisanzio, B. Mappin, U. Dalrymple,
et al. The effect of malaria control on plasmodium falciparum in Africa be-
tween 2000 and 2015. Nature, 526(7572):207–211, 10 2015.
Zulfiqar A Bhutta, Samana Ali, Simon Cousens, Talaha M Ali, Batool Azra
Haider, Arjumand Rizvi, et al. Interventions to address maternal, newborn,
and child survival: what difference can integrated primary health care strate-
gies make? The Lancet, 372(9642):972–989, 2008.


                                    40
Simon Brook. Lady health worker programme evaluation 2007-09. Oxford
Policy Management Policy Expertise, 2009.
Jennifer Bryce, Cesar G Victora, Jean-Pierre Habicht, J Patrick Vaughan,
and Robert E Black. The multi-country evaluation of the integrated manage-
ment of childhood illness strategy: Lessons for the evaluation of public health
interventions. American Journal of Public Health, 94(3):406–415, 2004.
Danielle Charlet, Allisyn C. Moran, and Supriya Madhavan. Summary findings
from a mixed methods study on identifying and responding to maternal and
newborn illness in seven countries: implications for programs. Journal of
Health, Population and Nutrition, 36(1):48, 2017.
Jessica Cohen, Pascaline Dupas, and Simone Schaner. Price subsidies, diag-
nostic tests, and targeting of malaria treatment: evidence from a randomized
controlled trial. American Economic Review, 105(2):609–645, 2015.
Janet Currie, Wanchuan Lin, and Wei Zhang. Patient knowledge and antibiotic
abuse: evidence from an audit study in China. Journal of Health Economics,
30:933–949, 2011.
Jishnu Das.   The quality of medical care in low-income countries: From
providers to markets. PLoS Medicine, 8(4):1–2, 2011.
Jishnu Das and Jeffrey Hammer. Which doctor? combining vignettes and item
response to measure clinical competence. Journal of Development Economics,
78(2):348–383, 2005.
Jishnu Das and Jeffrey Hammer. Location, location, location: residence,
wealth, and the quality of medical care in Delhi, India. Health Affairs, 26
(3):w338–w351, 2007.
Jishnu Das and Jeffrey Hammer. Quality of primary care in low-income coun-

                                     41
tries: Facts and economics. Annual Review of Economics, 6:525–553, 2014.
Jishnu Das and Carolina Sánchez-Páramo. Short but not sweet: New evi-
dence on short duration morbidities from india. World Bank Policy Research
Working Paper, February 2003.
Jishnu Das and Thomas Pave Sohnesen. Patient Satisfaction, Doctor Effort
and Interview Location: Evidence from Paraguay. World Bank, 2006.
Jishnu Das, Jeffrey Hammer, and Kenneth L. Leonard. The quality of medical
advice in low-income countries. Journal of Economic Perspectives, 22(2):
93–114, 2008.
Jishnu Das, Jeffrey Hammer, and Carolina Sánchez-Peramo. The impact of
recall periods on reported morbidity and health seeking behavior. World Bank
Policy Research Working Paper, 5778, 2011.
Jishnu Das, Alaka Holla, Aakash Mohpal, and Karthik Muralidharan. Quality
and accountability in health care delivery: Audit-study evidence from primary
care in India. American Economic Review, 106(12):3765–3799, 2016.
Mark Dean, Anja Sautmann, Xinyi Zhang, and Pierre Pratley. Impact of
Health Worker Visits and Subsidized Primary Care on Health Outcomes and
Behavior - A Randomized, Controlled Trial in Bamako, Mali. mimeo, 2017.
Pascaline Dupas. Short-run subsidies and long-run adoption of new health
products: Evidence from a field experiment. Econometrica, 82(1):197–228,
2014.
Pascaline Dupas and Jessica Cohen. Free distribution or cost-sharing? Evi-
dence form a randomized malaria prevention experiment. Quarterly Journal
of Economics, 125(1):1–45, February 2010.
Amy Finkelstein, Sarah Taubman, Bill Wright, Mira Bernstein, Jonathan Gru-

                                    42
ber, Joseph P. Newhouse, Heidi Allen, Katherine Baicker, and Oregon Health
Study Group. The Oregon health insurance experiment: Evidence from the
first year. Quarterly Journal of Economics, 127(3):1057–1106, 2012.
Greg Fischer, Dean Karlan, Margaret McConnell, and Pia Raffler. To charge
or not to charge: evidence from a health products experiment in Uganda.
NBER Working Paper, 20170, 2014.
Donna B. Gilleskie. A dynamic stochastic model of medical care use and work
absence. Econometrica, 66(1):1–45, 1998.
Sandy Gove. Integrated management of childhood illness by outpatient health
workers: technical basis and overview. the who working group on guidelines
for integrated management of the sick child. Bulletin of the World Health
Organization, 75(Suppl 1):7, 1997.
Elizabeth Heger Boyle, Miriam King, and Matthew Sobek.                IPUMS-
Demographic and Health Surveys: Version 5. Model Questionnaire for Fe-
males, 10 2018.
Mark C. Hornbeck, Arnold V. Hurtado, and Richard E. Johnson. Health
care episodes: definition, measurement and use. Medical Care Research and
Review, 42(2):163–218, 1985.
Lucia Hug, David Sharrow, and Danzhen You. Levels & trends in child mor-
tality report 2019: Estimates developed by the UN Inter-agency Group for
Child Mortality Estimation. UNICEF, 2019.
Emmanuel Jimenez. Human and physical infrastructure: public investment
and pricing policies in developing countries, volume 3, chapter 43, pages 2773–
2843. Elsevier, 1995.
Ari D Johnson, Dana R. Thomson, Sidney Atwood, Ian Alley, Jessica L. Beck-

                                     43
erman, Ichiaka Koné, et al. Assessing early access to care and child survival
during a health system strengthening intervention in Mali: A repeated cross
sectional survey. PLoS One, 8(12):e81304, 2013.
Ari D Johnson, Oumar Thiero, Caroline Whidden, Belco Poudiougou, Djoumé
Diakité, Fousséni Traoré, et al. Proactive community case management and
child survival in periurban Mali. BMJ Global Health, 3(2), 2018.
Emmett B. Keeler, Joan L. Buchanan, John E. Rolph, Janet M. Hanley, and
David M. Reboussin. The demand for episodes of medical treatment in the
health insurance experiment. RAND Working Paper, 1988.
Gary King, Emmanuela Gakidou, Kosuke Imai, Jason Lakin, Ryan T. Moore,
Clayton Nall, et al. Public policy for the poor? a randomised assessment of
the Mexican universal health insurance programme. The Lancet, 373(9673):
1447–1454, 2009.
Michael Kremer and Rachel Glennerster. Improving health in developing coun-
tries: evidence from randomized evaluations. In Handbook of Health Eco-
nomics, volume 2, chapter 4, pages 201–315. Elsevier, 2012.
Kenneth L. Leonard and Melkiory C. Masatu. Using the Hawthorne effect to
examine the gap between a doctor’s best possible practice and actual perfor-
mance. Journal of Development Economics, 93:226–234, 2010.
K.L. Leonard, G.R. Mliga, and D. Haile Mariam. Bypassing health centres in
tanzania: revealed preferences for quality. Journal of African Economies, 11
(4):441, 2002.
Jenny I. Litvack and Claude Bodart. User fees plus quality equals improved
access to health care: Results of a field experiment in cameroon. Social Science
and Medicine, 37(3):369 – 383, 1993.

                                       44
Carolina Lopez, Anja Sautmann, and Simone Schaner. Does patient demand
contribute to the overuse of prescription drugs?     unpublished manuscript,
2020.
Willard G. Manning, Joseph P. Newhouse, Naihua Duan, Emmett B. Keeler,
and Arleen Leibowitz. Health insurance and the demand for medical care:
Evidence from a randomized experiment. American Economic Review, 77(3):
251–277, 1987.
Mayo Clinic.      Diseases and conditions:       Fever, May 2014.         URL
http://www.mayoclinic.org/diseases-conditions/fever/basics/symptoms/con-20019229.

Mayo Clinic.     Diarrhea:    Symptoms and causes, October 2016.          URL
http://www.mayoclinic.org/diseases-conditions/diarrhea/symptoms-causes/dxc-20232937.

E. Miguel and M. Kremer. Worms: identifying impacts on education and
health in the presence of treatment externalities. Econometrica, 72(1):159–
217, 2004.
Hoa L. Nguyen, Jane S. Saczynski, Joel M. Gore, and Robert J. Goldberg. Age
and sex differences in duration of prehospital delay in patients with acute my-
ocardial infarction: A systematic review. Circulation: Cardiovascular Quality
and Outcomes, 3:82–92, 2010.
Timothy Powell-Jackson, Kara Hanson, Christopher J.M. Whitty, and Eve-
lyn K. Ansah. Who benefits from free healthcare? evidence from a random-
ized experiment in Ghana. Journal of Development Economics, 107:305–319,
2014.
Valéry Ridde and Florence Morestin. A scoping review of the literature on
the abolition of user fees in health care services in Africa. Health Policy and
Planning, 27(3):1–11, 2012.

                                     45
Alfonso Rosales and Kristin Weinhauer. C-IMCI: Community Integrated Man-
agement of Childhood Illness. Catholic Relief Services, 2003.
João Santos Silva and Frank Windmeijer. Two-part multiple spell models for
health care demand. Journal of Econometrics, 104:67–89, 2001.
Save the Children. Freeing up Healthcare: A guide to removing user fees. Mc-
Pake, Barbara and Schmidt, Alice and Araujo, Edson and Kirunga-Tashobya,
Christine, London, 2008.
G. L. Stoddart and M. L. Barer. Analysis of demand and utilization through
episodes of medical services. McMaster University working paper, 81-06, 1981.
Dag Gundersen Storla, Solomon Yimer, and Gunnar Aksel Bjune. A system-
atic review of delay in the diagnosis and treatment of tuberculosis. BMC
Public Health, 8(15), 2008.
Rebecca L. Thornton, Laurel E. Hatt, Erica M. Field, Mursaleena Islam,
Freddy Solís Diaz, and Martha Azucena González. Social security health in-
surance for the informal sector in Nicaragua: a randomized evaluation. Health
Economics, 19(S1):181–206, 2010.
UK Secretary of State for International Development. Eliminating World
Poverty: Building our Common Future. Department for International De-
velopment White Paper, London, 2009.
UNICEF Health Section. Integrated Management of Childhood Illness: House-
hold and Community Component. Kenya-Mugisha, Nathan and Pangu, Kasa,
1999.
United Nations Inter-Agency and Expert Group on MDG Indicators. The
Millennium Development Goals Report 2015. United Nations, 2015.
WHO. Integrated Management of Childhood Illness: Chart Booklet. World

                                     46
Health Organization, 2014.
WHO Department of Child and Adolescent Health and Development. Hand-
book IMCI: Integrated Management of Childhood Illness. World Health Orga-
nization and UNICEF, 2005.
WHO Global Malaria Programme. World Malaria Report 2014. World Health
Organization, 2015.
Rob Yates. The impact of abolishing user fees in Africa – recent developments
in six African countries. SSRN Working Paper, June 2007.




                                     47
Tables and Figures


                                          Table 1: Demographics and balance.
                                   Child characteristics                         Characteritics of household head                  Household characteristics         Compound
                                                                                                                                                                         Log
                                                                                                                               Number                      Total      distance
                                                       Weight                     Speaks                   Has      Age >                     Owner
                             Gender     Age (yrs)                  Is literate               Gender                              of                      assets in   to closest
                                                     for height                  Bambara                  salary    median                   occupiers
                                                                                                                               members                   log USD       formal
                                                                                                                                                                        care
                               (1)          (2)         (3)           (4)           (5)         (6)         (7)       (8)         (9)          (10)        (11)          (12)
Subsidy only group (S)        0.016       -0.018       0.112         0.033        -0.062      -0.018      -0.000    -0.002       0.175         0.073      0.424        -0.066
                             (0.04)       (0.09)       (0.12)        (0.05)       (0.05)      (0.03)      (0.04)    (0.05)      (0.32)        (0.05)      (0.28)       (0.06)

CHW only group               -0.015       0.036        0.042         -0.027       -0.044      -0.026      0.001     -0.011      -0.073         0.074      0.266       -0.004
                             (0.04)       (0.10)       (0.14)        (0.05)       (0.05)      (0.03)      (0.03)    (0.05)      (0.32)        (0.05)      (0.27)      (0.06)

Subsidy x CHW group (SCHW)    0.025       -0.018       0.032        0.117∗∗       -0.079      -0.003      -0.037     0.043       0.321         0.042      0.149       -0.079
                             (0.03)       (0.10)       (0.13)        (0.05)       (0.05)      (0.03)      (0.03)    (0.05)      (0.31)        (0.06)      (0.27)      (0.07)
                                  ∗∗∗         ∗∗∗            ∗∗∗          ∗∗∗         ∗∗∗         ∗∗∗         ∗∗∗        ∗∗∗         ∗∗∗           ∗∗∗         ∗∗∗
Control group mean           0.515       2.706       -0.655         0.467        0.729       0.903       0.131      0.418       6.234         0.405      6.458       6.238∗∗∗
                              (0.02)      (0.06)       (0.09)        (0.04)       (0.04)      (0.02)      (0.03)     (0.03)      (0.21)        (0.04)     (0.18)      (0.04)
Observations                   1763        1768         1306          1242         1242        1241        1242       1256        1238          1208       1208        642

Notes: Coefficients from a regression of the outcome variable on treatment indicators.
Standard errors clustered at the compound level in parentheses. * p < 0.10, **
p < 0.05, *** p < 0.01.



Table 2: Consultation numbers, average costs to the patient’s family, value of
services received, and wait times per visit in hours.
                                   Number of                                                                                              Waiting time per visit in hours
                                                           Cost per visit (to household) Value per visit (total cost)
                                     visits                                                                                                         (caretaker)
CSCom (all)                          646                         1344                (106)              3767            (114)                 0.74             (0.04)
CSCom (with subsidy)                 508                        933***               (108)              3794            (129)               0.78**             (0.04)
CSCom (without subsidy)              138                         2850                (254)              3666            (252)                 0.59             (0.07)
Private provider                      89                         5470                (543)                -               -                   1.24             (0.19)
Informal/no consultation             2691                        218                  (11)                -               -                   0.04             (0.01)

Notes: CSCom refers to local clinic or associated reference hospital (CSRef ). Pri-
vate provider is a private doctor, hospital, or health center provided e.g. by the Red
Cross. Informal includes peddlers, market, traditional or religious healers, mid-
wives, and pharmacy visits without a formal consultation. Costs, prices, and wait
times are added for all connected consultations, e.g. doctor visit and medication pur-
chase. Costs per visit are for consultation and treatment as reported by the caretaker.
Value/total cost is imputed from private cost and median prices, by consultation type
and treatment received. Standard errors in parentheses. Significance levels for a
t-test between free care and no-free care consultations. *** p<0.01, ** p<0.05, *
p<0.1.




                                                                                     48
Table 3: Days of (untreated) illness per child, by need for care (IMCI); and
proportion of days on which each symptom is observed.
                         Pre-care days per child:                        All               Early                Care required
                                                                 Mean         SD       Mean      SD             Mean       SD
                         Total number                            16.31      (14.32)    10.34 (9.38)             5.97     (8.68)
                         Percentage of total number on which each symptom is present:
                         Convulsions, fits, or spasms        0.32%                0%                             0.66%
                         Lethargic or unconscious            3.95%                0%                            10.27%
                         Unable to drink or breastfeed       1.17%                0%                             3.29%
                         Vomiting everything                 5.39%                0%                            15.58%
                         Coughing                           33.42%            62.56%                            35.38%
                         Difficulty breathing                4.46%             8.58%                             5.85%
                         > 3 loose stools                    8.54%            11.84%                            12.75%
                         Blood in the stool                  0.61%             0.06%                             1.39%
                         Sunken eyes                         2.18%             0.31%                             5.27%
                         Unusually hot skin                 32.04%            17.68%                            74.00%
                         Other: rash, spots, or itch         2.85%             5.74%                             1.28%
                         Other: cold symptoms               50.76%            92.01%                            37.31%
                         Other: ear ache                     1.01%             1.13%                             1.90%
                         Other: wound, injury, or burn       3.62%             6.82%                             1.60%
                         Other symptoms                      4.90%             9.04%                             2.54%

Notes: The first four symptoms are danger signs and always occur on “care required”
days. Symptoms that do not themselves indicate need for care can co-occur on care-
required days



Table 4: Utilization, private expenditure, and value of treatment by treatment
group.
                                        Number of visits per child            Private cost per child (CFA)         Wait time   Value per child (CFA)
                                                               Other                                   Other       per child
                                  All formal   CSCom                           Total    CSCom                                      Total      CSCom
                                                               formal                                  formal         (hrs)
                                      (1)         (2)            (3)         (4)           (5)           (6)           (7)         (8)          (9)
Subsidy only group (S)             0.379∗∗∗    0.402∗∗∗        -0.023      -77.523      215.880       -129.134      0.275∗∗∗   1386.066∗∗∗   1665.570∗∗∗
                                    (0.06)       (0.05)       (0.03)       (275.29)     (134.08)     (175.97)        (0.10)       (374.75)    (263.11)

CHW only group                      0.022        0.017         0.005       459.660      207.295       -33.815        0.039        387.381     126.962
                                    (0.04)       (0.03)       (0.03)       (295.11)     (134.88)     (179.80)        (0.08)       (313.59)    (154.89)

Subsidy x CHW group (SCHW)         0.439∗∗∗     0.478∗∗∗      -0.039       -220.833     186.633      -231.538       0.307∗∗∗   1409.314∗∗∗   1773.584∗∗∗
                                    (0.06)       (0.05)       (0.02)       (245.47)     (128.40)     (166.57)        (0.09)       (315.97)    (221.90)

Control group mean                 0.217∗∗∗     0.153∗∗∗     0.065∗∗∗    1461.672∗∗∗   353.302∗∗∗    369.797∗∗      0.264∗∗∗   1671.343∗∗∗   523.071∗∗∗
                                    (0.03)       (0.02)       (0.02)      (192.35)      (60.64)      (153.38)        (0.06)     (204.40)      (81.12)
p-value S = SCHW                    0.421        0.286        0.309         0.565        0.859         0.343         0.771        0.953        0.739

Notes: Coefficients from a regression of the outcome variable on treatment indicators.
1768 observations. Standard errors clustered at the compound level in parentheses.
* p < 0.10, ** p < 0.05, *** p < 0.01.




                                                                         49
Table 5: All spells vs. spells with a consultation by treatment group; condi-
tional care-required status.
                                                           consultations                      consultations as
                                                            (spells with   consultations as      % of all
                                                spells     consultation)    % of all spells    consultations
                  Control
                     care never required        479            18                4%                25%
                     care required              476            53               11%                75%
                         entered during spell   210            22               10%
                     total                      955            71               7%                 100%
                  CHW group
                     care never required        402            13                3%                20%
                     care required              516            52               10%                80%
                         entered during spell   308            30               10%
                     total                      918            65               7%                 100%
                  Subsidy group
                     care never required        449            40                9%                23%
                     care required              458            133              29%                77%
                         entered during spell   195            57               29%
                     total                      907            173              19%                100%
                  Full program group
                     care never required        463            50               11%                26%
                     care required              504            144              29%                74%
                         entered during spell   275            80               29%
                     total                      967            194              20%                100%


Notes: includes only uncensored and right-censored spells. “Care never required”
indicates spells with early days only. “Care required” includes spells that either consist
entirely of care-required days, or start with an early day and eventually enter care-
required status (“entered during spell”). A spell cannot revert from care-required status
except through recovery.




                                                      50
            Table 6: Treatment effects in a Cox proportional hazard model.
                                           (1)         (2)        (3)        (4)        (5)       (6)
Care required                           5.284∗∗∗    5.366∗∗∗   5.592∗∗∗   5.714∗∗∗   5.367∗∗∗   5.509∗∗∗
                                         (0.63)      (0.64)     (0.88)     (0.92)     (1.40)     (1.45)

Subsidy                                 2.563∗∗∗    2.525∗∗∗   2.561∗∗∗   2.523∗∗∗
                                         (0.45)      (0.45)     (0.45)     (0.45)
Subsidy x early (Se)                                                                 2.450∗∗∗   2.426∗∗∗
                                                                                      (0.74)     (0.74)

Subsidy x care req. (Sr)                                                             2.599∗∗∗   2.557∗∗∗
                                                                                      (0.50)     (0.50)

CHW                                      0.915       0.886
                                         (0.20)      (0.19)
CHW x early (CHWe)                                              0.910      0.877      0.882      0.853
                                                                (0.34)     (0.33)     (0.36)     (0.35)

CHW x care req. (CHWr)                                          0.914      0.884      0.923      0.892
                                                                (0.21)     (0.20)     (0.21)     (0.21)

Subsidy x CHW                            1.218       1.240
                                         (0.31)      (0.31)
Subsidy x CHW x early (SCHWe)                                   1.366      1.414      1.428      1.471
                                                                (0.56)     (0.58)     (0.68)     (0.70)

Subsidy x CHW x care req. (SCHWr)                               1.176      1.192      1.159      1.176
                                                                (0.31)     (0.31)     (0.32)     (0.32)
Covariates                                              Y                    Y                     Y
p-value Se = Sr                                                                       0.855      0.872
p-value CHWe = CHWr                                             0.991      0.983      0.916      0.917
p-value CHWe+SCHWe=CHWr+SCHWr                                   0.568      0.532
p-value Se+CHWe+SCHWe = Sr+CHWr+SCHWr                                                 0.752      0.704
N                                        23836       23836      23836      23836      23836      23836

Notes: table reports hazard ratios, uncensored and right-censored spells only. Co-
variates include household assets; distance to closest formal care provider; gender,
age, literacy of household head; and child gender. Stratum FE included. Standard
errors clustered at the compound level in parentheses. * p < 0.10, ** p < 0.05, *
p < 0.01.




                                                   51
Figure 1: The effect of providing a subsidy (A) and of aligning parents’ beliefs
with those of the policymaker (B).




Notes: the probability of not recovering π is increasing conditional on the spell
continuing, due to learning about the illness. We assume parents have a higher cost
threshold than the policymaker (K ≥ K ∗ ). Panel A: beliefs are aligned so π = π ∗ ,
Panel B: parental beliefs π are the average of the policymaker’s beliefs π ∗S and π ∗L .


Figure 2: Average daily probability of care seeking by day of the spell and
treatment group, early vs. care-required classification according to the C-IMCI.
                                   Control                                                 Subsidy only
         .2




                                                                       .2
         .15




                                                                       .15
         .1




                                                                       .1
         .05




                                                                       .05




               1   2   3           4     5          6        7   >7          1   2   3           4   5          6        7   >7

                           early             care required                               early           care required



                               CHW only                                                  Subsidy & CHW
         .2




                                                                       .2
                                                                       .15
         .15




                                                                       .1
         .1




                                                                       .05
         .05




               1   2   3           4     5          6        7   >7          1   2   3           4   5          6        7   >7

                           early             care required                               early           care required




Notes: x-axis shows day of the spell; spell days after day seven are grouped. Graphs
show point estimates of a fully dummied-out linear regression model along with 95%
confidence intervals. Standard errors clustered at the compound level.




                                                                      52
Figure 3: Average daily probability of care seeking by age of the child in years
and treatment group, “early” vs. “care-required” classification according to the
C-IMCI.
                               Control                                        Subsidy only




                                                              .1
         .04




                                                              .08
                                                              .06
         .02




                                                              .04
                                                              .02
         0




                                                              0
               0   1      2              3           4   5          0   1      2        3           4   5

                       early             care required                      early       care required



                           CHW only                                         Subsidy & CHW




                                                              .1
         .04




                                                              .08
                                                              .06
         .02




                                                              .04
                                                              .02
                                                              0
         0




               0   1      2              3           4   5          0   1      2        3           4   5

                       early             care required                      early       care required




Notes: x-axis shows age of the child in years; children over 5 years are dropped.
Graphs show point estimates of a fully dummied-out linear regression model along
with 95% confidence intervals. Standard errors clustered at the compound level. Note
difference in scale on the y-axis in Control and CHW only group (left) vs. subsidy
only and full program group (right).




                                                             53