Policy Research Working Paper                 9482




         Does Patient Demand Contribute
       to the Overuse of Prescription Drugs?
                             Carolina Lopez
                             Anja Sautmann
                             Simone Schaner




Development Economics
Development Research Group
November 2020
Policy Research Working Paper 9482


  Abstract
 This study conducted an experiment in Mali to test whether                         letting doctors decide to share this information, increased
 patients pressure doctors to prescribe medical treatment                           discount use by 35 percent and overall malaria treatment
 they do not necessarily need. The experiment varied patients’                      by 10 percent. These marginal patients rarely had malaria,
 information about a discount for antimalarial tablets and                          worsening the illness-treatment match. Providers did not
 measured demand for both tablets and costlier antimalar-                           use the information advantage to sell injections—their use
 ial injections. The study finds evidence of patient-driven                         fell in both information conditions.
 demand: informing patients about the discount, instead of




 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
      Does Patient Demand Contribute to the
          Overuse of Prescription Drugs?
                   Carolina Lopez and Anja Sautmann
                           and Simone Schaner∗




JEL Codes: I12, I18, O12
Keywords: malaria treatment, demand for prescription drugs, healt care
overuse, doctor-patient interaction, Mali




    ∗
      We are grateful to Pascaline Dupas, Rema Hanna, seminar participants at the Univer-
sity of Toronto, Loyola Marymount University, MIT, USC, Rice University, UCSD Rady,
the Barcelona Summer Forum, the BREAD pre-conference, the Stanford Rosenkranz
Global Health Policy Symposium, McGill, Brown, Microsoft Research, FGV Sao Paulo,
PUC Rio, the University of Cape Town, the University of Gothenburg, and the Stock-
holm School of Economics for many helpful comments. We thank the staff of IPA Mali
and especially Alassane Koulibaly for expert field coordination. This project would not
have been possible without the advice and support of Dr. Seydou Doumbia and Dr. Is-
saka Sagara (Malaria Research and Training Center, University of Bamako) as well as
Dr. Seydou Fomba (Programme National de Lutte contre le Paludisme (PNLP)). We
thank the Ministry of Health and the participating community clinics for allowing us to
conduct this research. The study was funded by ESRC/DFID Development Frontiers
Award ES/N00583X/1. The research protocol was approved by the IRB of La Facult
de Mdecine de Pharmacie et d’Odontostomatologie de Bamako (FMPOS) (protocol No.
2016/47/CE/FMPOS) and registered in the AEA RCT registry (AEARCTR-0001707).
The findings, interpretations, 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. All errors are our own. Carolina Lopez: Brown University,
carolina lopez@brown.edu. Anja Sautmann: World Bank, asautmann@worldbank.org. Si-
mone Schaner: University of Southern California, BREAD, and NBER, schaner@usc.edu.
    Low-income countries must address two opposing challenges when craft-
ing health policy: many patients do not seek care when they are ill (White-
head, Dahlgren and Evans, 2001; Sautmann, Brown and Dean, 2017), but
those who do often consume a great deal of unnecessary treatment (Das and
Hammer, 2014; Li et al., 2012). One solution could be to subsidize critical
treatments in the public sector, while staffing facilities with trained health-
care providers to act as gatekeepers to subsidized care.
    Yet providers often fail to live up to their gatekeeping mandate. For ex-
ample, in an audit study Das et al. (2016) find that 74 percent of public
sector providers in India dispensed at least one unnecessary treatment; the
Centers for Disease Control and Prevention estimate that 30 percent of an-
tibiotic prescriptions in the United States are unnecessary (Fleming-Dutra
et al., 2016); and studies across Sub-Saharan Africa document large shares
(36-81 percent) of malaria-negative patients receiving antimalarials at public
health centers (Reyburn et al., 2004; Hamer et al., 2007; Bisoffi et al., 2009;
Ansah et al., 2010). Providers’ gatekeeping ability is key for healthcare poli-
cies that increase demand, such as subsidies: only if healthcare providers are
good gatekeepers will such policies expand access to those who need treat-
ment while limiting overuse among those who do not need treatment.
    Why are so many providers such poor gatekeepers? Most existing liter-
ature studies doctor-driven reasons for overtreatment. These include mis-
aligned financial incentives (e.g. Iizuka, 2012; Currie, Lin and Meng, 2014;
Das et al., 2016) and knowledge gaps, compounded by low motivation (e.g.
Banerjee, Duflo and Glennerster, 2008; Das, Hammer and Leonard, 2008).1
Doctors and other healthcare providers can overprescribe to patients because
health care is a credence good: since patients cannot verify the true cause of
their illness, they must rely on the provider’s recommendation. But by a sim-
ilar argument, there is scope for patient-driven demand for unnecessary treat-
ment: if the provider cannot persuade the patient that a medication is not
needed, she may write a prescription just to keep the patient satisfied. Anec-
dotal evidence suggests patient demand matters: for example, in 2016 the
U.S. pharmaceutical industry spent $6 billion on direct-to-consumer market-
ing of prescription drugs, 22 percent of total marketing spending (Schwartz
and Woloshin, 2019), and explicit or perceived patient demand is often cited
by providers as a reason for overtreatment (Kotwani et al., 2010; Linder et al.,
2014). Yet there has been much less rigorous research on this channel.

   1
       For a detailed review of this literature, see Dupas and Miguel (2017).


                                              2
    Our paper aims to fill this gap. We begin by developing a theoretical
model of doctor-patient interactions, where doctors and patients may have
different preferences about the best course of medical treatment for an illness,
and may in particular disagree about whether a certain treatment is truly
needed. Patient-driven demand occurs when the patient consumes more
treatment than the doctor would like. In our model, this can happen when
doctors change their prescription behavior to avoid “gatekeeping costs”. Pa-
tients inflict this utility cost (e.g., in time, hassle, or lost future business) on
doctors when they learn that the doctor’s prescription does not allow them
to obtain their preferred treatment option.2 We then conduct an experiment
focused on malaria treatment. Within the experiment, the model generates
predictions for when and how a provider will selectively share information
with patients to reduce gatekeeping costs and steer treatment towards her
own preferences. The theory guides our analysis, allowing us to (1) test for
the presence of gatekeeping costs, (2) assess to what extent patients drive
increased use of antimalarials in response to a price subsidy, and (3) assess
the extent to which providers drive demand for more expensive antimalarial
injections. Finally, we estimate how these changes in demand impact the
misallocation of treatment.
    The experiment was conducted with 2,055 patients visiting 60 public-
sector health clinics in Bamako, Mali. Mali’s policies for malaria control are
in line with international consensus and dictate that patients should only be
prescribed an antimalarial if they have a positive malaria test (Minist`     ere de
la Sant´e, 2013). The vast majority of malaria patients present with “sim-
ple” or “uncomplicated” malaria, which should be treated with artemisinin
combination therapy (ACT) tablets. A small share of patients present with
“severe” malaria, which requires more intensive (and expensive) medication,
usually delivered via injection or IV drip. We document widespread overuse
of antimalarials even though diagnostic tests are readily available: in our
control group, 58 percent of malaria-negative patients received an antimalar-
ial prescription, and 41 percent of these prescriptions were for severe malaria
treatment.
    The experiment introduced a discount for simple malaria treatment, by
giving a voucher for a free course of ACT tablets. This subsidy increased the

   2
     Our conceptualization of gatekeeping costs is related to Prendergast’s (2003) model
of bureaucracy, where the risk of complaints distorts the decisions of a bureaucrat tasked
with allocating a good.


                                            3
attractiveness of simple malaria treatment relative to both no treatment and
severe treatment. The experiment varied (within clinic, across different days)
whether this discount was a provider’s private information or known to both
provider and patient. On “Patient Voucher” days, all acutely ill patients
visiting the health clinic were given a voucher. On “Doctor Voucher” days,
the same vouchers were given to providers to dispense at their discretion.
This means the healthcare provider could choose to make simple treatment
more attractive to the patient by revealing the voucher.3
    Our model shows that treatment outcomes will not vary across the Doc-
tor and Patient Voucher conditions unless there are gatekeeping costs. Intu-
itively, the patient-facing costs and clinic revenues associated with different
treatments are the same in the Doctor and Patient Voucher conditions; con-
sequently a doctor (or other healthcare provider’s) preferred action based
on these factors will also be the same in both conditions. The doctor will
sometimes deviate from her preferred prescription to satisfy the patient and
avoid gatekeeping costs. However, relative to the Patient Voucher condition,
the doctor’s information advantage in the Doctor Voucher condition may
help her avoid making such compromises. Consider, for example, a doctor
who does not want to prescribe an antimalarial: she may give in to patient
pressure for free treatment in the Patient Voucher condition, but can conceal
the discount and avoid prescribing in the Doctor Voucher condition. This
provides us with a test of the case where patients prefer more simple malaria
treatment than doctors (“patient-driven demand”).
    By contrast, suppose the provider would like to induce as much treatment
as possible, for example because she wants to increase clinic revenue: in this
case she will reveal the voucher to patients who would otherwise not buy
malaria treatment, but conceal the voucher from patients who are prepared
to buy the more expensive severe malaria treatment. Thus, we can also test
for the case when providers have a stronger preference for severe malaria
treatment than patients (“doctor-driven demand”). Note that our primary
tests of doctor- and patient-driven demand rely on different margins (simple
versus no malaria treatment for patient-driven demand; simple versus severe
malaria treatment for doctor-driven demand). It is therefore ex ante possible
to find support for one, both, or neither type of misalignment.
   3
     Vouchers in both conditions had identical appearances, terms, and conditions. We
used clinics’ regular stocks of malaria medicine and reimbursed clinics for all free medica-
tion dispensed through the voucher program at the prevailing retail price. This kept the
financial incentives for doctors constant across the three treatment groups.


                                             4
    We find robust evidence that gatekeeping costs affect providers’ choices:
patients are nine percentage points (35 percent) more likely to redeem a
voucher on Patient as compared to Doctor Voucher days. This implies that
provider and patient preferences are frequently misaligned. We also find con-
sistent evidence that excess demand for vouchers on Patient Voucher days is
driven by patients’ preferences for simple treatment. First, providers are 4
percentage points more likely to write an antimalarial prescription (for sim-
ple or severe treatment) on Patient Voucher versus Doctor Voucher days,
and patients are 6 percentage points more likely to purchase malaria treat-
ment. Second, the excess demand for antimalarials on Patient Voucher days
is driven by patients with the fewest malaria symptoms – antimalarial pur-
chases by this group are 9 percentage points (23 percent) higher. This sug-
gests providers prefer to withhold treatment from these marginal low-risk
cases, but accede to demand when patients know that treatment is free.
    In contrast, we find no evidence of doctor-driven demand for severe
malaria treatment. There are no significant differences in rates of severe
malaria treatment across Patient and Doctor Voucher days, and we find
no evidence that providers use their information advantage in the Doctor
Voucher condition to increase clinic revenues.4
    Next, we assess the match between underlying illness and treatment with
data from follow-up home-based malaria tests with patients. In the Control
group, we estimate that 57 percent of patients received “appropriate” malaria
treatment, in that they either were malaria positive and purchased malaria
treatment, or were negative and did not purchase treatment. The Patient
Voucher condition reduced the match by 8 percentage points, compared to
only 4 percentage points in the Doctor Voucher condition.5 Over half of
the worsening of the match in the Patient Voucher condition is driven by a
change in providers’ prescribing behavior that is not observed in the Doctor
Voucher condition; in other words, patient-driven demand for ACTs leads to
overtreatment.6

    4
      Although doctors in our setting are salaried, they have an incentive to increase revenue
because this funds their salaries, and bonuses are often tied to revenues. See Section 1 for
more detail.
    5
      We reject equality of these two treatment effects at the one percent level.
    6
      The remainder is driven by an increase in antimalarial purchases among individuals
who would have received prescriptions even absent the price discount (some patients do
not purchase prescribed medication, which in our context works to improve the match,
since most patients do not actually have malaria).


                                              5
    Our experiment provides some of the first well-identified evidence of
patient-driven demand for excess medical treatment. Most directly related
to our work are Kravitz et al. (2005) and Currie, Lin and Meng (2014), who
conduct audit studies where “simulated patients” or SPs (trained actors who
are not actually sick) request a specific treatment. Kravitz et al. find that
doctors are more likely to prescribe antidepressants when SPs ask for them.
Currie, Lin and Meng find that direct SP requests for an antibiotic increased
prescription rates at hospitals in China, although this effect disappeared
when SPs indicated they would purchase their medications outside the hos-
pital pharmacy. Our contribution is to develop a method to identify whether
this provider compliance with patient requests for medication is willing (the
doctor prefers to prescribe) or reluctant (the doctor would rather not pre-
scribe, but goes along to avoid gatekeeping costs). Our experiment also has
the advantage that it uses real patients, allowing for natural doctor-patient
interactions where patients themselves choose whether to express demand for
treatment in response to the subsidy.
    Our work also contributes to the literature on the pricing of health goods
in developing countries. Most recent work in this area focuses on preventive
health products like mosquito bednets and water purification, where benefi-
ciary demand is low, use generates positive externalities, and large subsidies
typically improve outcomes.7 Use of curative products, by contrast, can gen-
erate negative externalities (e.g. by contributing to disease resistance (WHO,
2014)). Recent work on the pricing of curative health products has found
mixed results. Cohen, Dupas and Schaner (2015) find that subsidizing anti-
malarials at private-sector “drug shops” dramatically improves access to life-
saving medication, but making the subsidy too high generates large amounts
of overtreatment with only marginal additional gains in access. Importantly,
these shops were staffed by retail workers, who were not expected to perform
a gatekeeping function. In the public sector, Sautmann, Brown and Dean
(2017) find that subsidizing health care for children (combined with close
cost control) can improve access at the extensive margin significantly. Our
paper contributes by offering evidence on misallocation – and its drivers –
conditional on seeking care.
    Our tests are specific to marginal changes in patient demand in response
to the ACT subsidy and do not allow us to extrapolate how the different

   7
    For bednets see Cohen and Dupas (2010), (Dupas, 2014), and (Tarozzi et al., 2014).
For water purifications see (Ashraf, Berry and Shapiro, 2010) and (Kremer et al., 2011).


                                          6
mechanisms contribute to overall levels of overtreatment. Even so, public
clinics are a prime delivery mechanism for acute malaria care across Sub-
Saharan Africa, meaning this is precisely the margin of interest for poli-
cymakers deciding whether to subsidize malaria treatment. On this margin,
we show subsidies induce patient-driven demand that leads to overtreatment.
Thus, interventions that strengthen the provider’s authority and reduce gate-
keeping costs (like patient communication tools) could help sustain subsidies
while limiting overtreatment.
    The rest of the paper is structured as follows: in the next section we pro-
vide background on health care and malaria treatment in Mali and present
some stylized features of status quo malaria treatment. We develop a model
of doctor-patient interactions in section 2. Section 3 describes our experimen-
tal design, section 4 presents and discusses our empirical results, including
alternative explanations for our findings, and section 5 concludes.


1     Background
1.1    Health Care and Malaria in Mali
The public health system in Mali is organized around decentralized community-
based primary care funded by user fees. At the foundation of this system
are Centres de Sant Communautaire (CSComs) – community-based primary
care clinics managed by a local health association. A typical clinic in our
sample is staffed with one or two physicians and around five medical trainees
and five other staff who can prescribe malaria, along with additional techni-
cal and administrative staff. Clinics with a laboratory have a lab technician
who conducts blood tests. Table B1 in the Appendix gives an overview of the
staffing numbers in our sample. The local health association retains revenues
from sales of medications and other user fees, which are used to fund the op-
erations of the CSCom. Qualitative research suggests this system leads staff
to prioritize profit generation, since clinic revenues can be used to fund end-
of-year physician bonuses (Tour´  e, 2015; Escot, 2012). Moreover, pharmacy
sales often directly fund staff salaries – this is the case in 87 percent of the
clinics in our sample.
    Similar community-funded public healthcare is found all over West Africa,
and in other countries like China, where high rates of antibiotic use in the
public sector have been attributed to physicians’ incentives to generate bonus


                                      7
revenue (Currie, Lin and Meng, 2014). This structure – where medical staff
are salaried but still have an incentive to generate revenue to sustain the
clinic – exists in many other settings across the developing world, including
private, mission, and NGO-run medical facilities.
     CSComs are one of the most important sources of care in Mali: according
to the 2012-2013 Demographic and Health Survey, 47 percent of mothers
in Bamako who sought care for a child under five with fever or cough took
the child to a CSCom. The quality of care across CSComs varies and is
typically poorest in rural areas. Our study was conducted in the capital city
of Bamako and a nearby suburb, where clinics offer a higher standard of
care and are usually headed by a trained medical doctor (as compared to a
national average of roughly one third (PMI, 2016)). All clinics in our sample
have at least one salaried physician on staff, and in over 90 percent of clinics
that have a pharmacy, there is a physician present when the pharmacy is
open (see Appendix Table B1).
     One of the most commonly-treated illnesses in CSComs is malaria. De-
spite recent progress, malaria is the second largest cause of mortality in Mali,
accounting for roughly 14 percent of all deaths and 19 percent of deaths
among children under five (IHME, 2018). Although the parasite is endemic
in all parts of the country except the sparsely populated northern desert,
rates of transmission are substantially lower in urban areas. For example, in
2015 the estimated prevalence of the parasite in children under five in Ba-
mako was six percent, as compared to 36 percent nationwide (PNLP et al.,
2016).
     Malaria infections are classified as either “simple/uncomplicated” or “se-
vere”. Simple malaria is not life threatening if treated promptly, and is
characterized by non-specific, flu-like symptoms including fever, chills, and
headache. If left untreated, simple malaria can progress to severe malaria.
In this stage of the disease, patients often suffer convulsions and experience
life-threatening complications, including loss of consciousness/coma, respira-
tory failure, renal failure, and severe anemia (Trampuz et al., 2003). Patients
with severe malaria require prompt, aggressive treatment to avoid death and
should be hospitalized until their symptoms stabilize.
     Mali’s national malaria policy requires that suspected malaria cases be
confirmed via a positive microscopy or rapid diagnostic test (RDT) before
dispensing treatment (Minist`                 e, 2013).8 RDTs are meant to be
                                ere de la Sant´

   8
       This recommendation was first made in the five-year strategic plan for 2007-2011.


                                           8
free in public health facilities (including CSComs) to ensure that cost is not
a barrier to accurate diagnosis.9 Artemisinin combination therapies (ACTs)
are recommended for simple malaria, while severe malaria cases should be
treated with injectable artesunate followed by a dose of ACTs once the pa-
tient is stable.10 These policies are in line with WHO recommendations
(WHO, 2014) and are designed to limit the spread of drug resistance, which is
hastened by overtreatment. Limiting overuse of antimalarials is particularly
important, as drug resistance has rendered past generations of antimalarials
ineffective across much of Sub-Saharan Africa and Asia (Arrow, Panosian
and Gelband, 2004). This externality is one of the largest (potential) costs
to overtreatment, though overtreatment also strains government budgets for
drug subsidies, wastes resources of poor families, causes side effects, and can
slow diagnosis and treatment of non-malarial illnesses.
    In public facilities, ACTs are meant to be free for children under five
and subsidized for older individuals, but there are no subsidies for other
components of severe malaria treatment (PMI, 2016). In our study area, the
ACT subsidy policy is only partially enforced, as evidenced by the fact that
37 percent of children under five receiving ACTs in our control group paid a
positive price.

1.2     Status Quo Malaria Treatment in the Study Sam-
        ple
We use data from our Control group to quantify overuse of malaria treatment
at the 60 clinics that participated in our study. Here we give a brief overview
of the data; we provide additional detail on the sampling frame and data
collection protocols in section 3. We stationed surveyors at each clinic for

Previously, presumptive treatment of any fever cases was the main approach to malaria
control (Kon´ e et al., 2015).
    9
      In our data, RDTs were free of charge 70 percent of the time, while microscopy tests
were free less than 3 percent of the time.
   10
      Artesunate is a derivative of artemisinin. Artemisinin-based antimalarials are the
most effective treatments for malaria in Sub-Saharan Africa, where the emergence of drug-
resistant parasites has rendered earlier generations of antimalarials ineffective. Quinine
can also effectively treat both simple and severe malaria infections in this region, but the
drug is less effective than artemisinin and has more side effects (Achan et al., 2011). Malian
policy generally reserves quinine for pregnant mothers, though national policy allows for
an initial dose of quinine to treat severe malaria if an injectable artemisinin-based therapy
is not available.


                                             9
6 days over a two week period. Enumerators administered a short “intake”
interview to all consenting patients seeking care for an acute illness. After
the patient’s consultation with the provider was complete, the enumerator
recorded details of all blood tests performed, medications prescribed, medi-
cations purchased, and fees paid. A randomly chosen subset of these patients
were selected for a more detailed home-based follow up survey on the day
after their clinic visit. As part of this visit, a trained enumerator performed
an RDT on the patient to confirm his/her malaria status.11
    Table 1 uses the 628 patient observations in the Control group to give
an overview of how acute illnesses are treated in clinics in Bamako. Panel A
shows average indicators regarding malaria beliefs and treatment outcomes.
Panel B reports treatments that often accompany malaria medications: an-
tibiotics, which are often given for fevers but are not indicated for malaria,
as well as injections/IV, which are used in severe malaria treatment. Panel B
also shows how often patients were referred to a hospital, the recommended
course of action for severe malaria. Finally, Panel C shows average costs in
West African CFA (at the time of the experiment CFA 610 was approximately
USD 1), split by the type of treatment received.
    Looking at means in the whole sample of 628 (column 2), treatment for
malaria is very common, with 31 percent of patients prescribed treatment
for simple malaria and 30 percent for severe malaria. Twenty-one and 26
percent of patients purchase medication for simple and severe malaria, re-
spectively (Panel A).12 Malaria treatment is expensive: while the average
patient not treated for malaria reported spending CFA 2,793 ($4.58 at the
average exchange rate of CFA 610 per USD that prevailed during the study
period), simple malaria patients paid CFA 4,475 ($7.34) and severe malaria

  11
       RDTs detect antigens for malaria, which remain in the bloodstream after the infection
is cured. We used CareStart HRP2(Pf) tests, which detect an antigen that typically takes a
couple of weeks to become undetectable in blood samples (Humar et al., 1997; Kyabayinze
et al., 2008). This test detects the P. falciparum malaria parasite, which accounts for 92
percent of all malaria infections in Mali (PMI, 2016). In WHO quality assurance testing,
these RDTs correctly identified malaria infection in 91 percent of low parasite density
blood samples and 100 percent of high parasite density samples. False positives occurred
less than 1 percent of the time, and the invalid test rate was 0 percent (WHO, 2017).
    12
       We classify treatment as indicating severe malaria treatment if the patient received
artemisinin monotherapy or quinine, (1) as an injectable or (2) recorded in the survey
as tablets, and received an ACT. These cases are consistent with guidelines for severe
malaria treament which require administering monotherapy/quinine into the bloodstream
or rectally.


                                            10
patients paid CFA 10,353 ($16.97). To put this in perspective, the fee for
severe malaria patients amounts to 47 percent of average per capita monthly
income in our sample (Panel C).
    We quantify overtreatment among the subset of patients who received a
valid RDT result at the home follow up. Column 3 shows average outcomes
for all patients who took a test, while columns 4 and 5 show means for pa-
tients who tested malaria positive and negative, respectively. Overtreatment
is widespread, with 58 percent of malaria negative patients being prescribed
antimalarials, as compared to 76 percent of malaria positive patients. Severe
treatment accounts for 78 percent of prescriptions for those who tested pos-
itive for malaria and 41 percent for those who tested negative, even though
donors estimate that only 10 percent of malaria cases should be severe (PMI,
2015). Use of antibiotics is also widespread, with over 40 percent of malaria
positive and negative patients purchasing the drugs. Finally, only 56 per-
cent of Control group patients prescribed an antimalarial took a test, even
though all clinics in our study had malaria testing capability. These facts
suggest gatekeeping failures, especially given the very straightforward treat-
ment guidelines for malaria (prescribe only for patients with a positive test).
    There are some indications that overprescription might be driven by
providers. The patterns of severe malaria treatment we observe match anec-
dotal reports that doctors in our sample treat positive malaria tests as a
sign of severe malaria, while interpreting negative tests accompanied by
malaria-like symptoms as a sign of simple malaria. This, coupled with the
fact that microscopy tests for malaria are not very accurate in field settings
(Wongsrichanalai et al., 2007), could explain some of the misallocation we ob-
serve.13 The revenue especially from severe treatment provides an additional
motive if doctors are concerned with the clinic’s financial health.14
    Moreover, 25 percent of patients do not purchase their antimalarial pre-
   13
      We find indications for this in our sample as well. Appendix Figure B1 shows that
the match between malaria status in our at-home test and receipt of an antimalarial
prescription is better among patients who received an RDT at the clinic as compared
to microscopy. Specifically, we find higher rates of overtreatment after microscope-based
tests, which suggests clinics may see more false positives with microscopy.
   14
      While we do not observe profits, higher revenue from an individual patient will indi-
cate higher profits provided profit is positively correlated with revenue. This likely holds
in our case, especially since the higher cost of severe treatment includes service fees e.g. for
injections. Our estimates suggest service fees are 35-47 percent higher for severe (as com-
pared to simple) malaria. Since there was very limited patient queuing in our setting, the
short-run marginal cost of providing such services is close to zero.


                                              11
scription, with the prescription-purchase gap mostly driven by individuals
declining to purchase available treatment (77 percent of cases) rather than
stock outs.15 This could be a consequence of credit constraints – a reason to
subsidize treatment in the first place – but may also be a sign of misalignment
between doctor prescriptions and patient preferences.
    On the other hand, there are also signs of patient-driven overprescription.
First, patients over-estimate their malaria risk: 55 percent of respondents re-
port suspecting malaria before they see the provider; among those with a
valid home-based RDT, this proportion is 61 percent, even though only 25
percent actually test positive. Second, the share of patients prescribed treat-
ment exceeds providers’ average estimate of true malaria prevalence (48 per-
cent) by 13 percentage points (see Appendix Table B2). Finally, 57 percent of
health workers report feeling pressure from patients to prescribe unnecessary
medication, with over half mentioning antimalarials specifically.16


2        Theory
In order to investigate the role of patients pressuring health workers to pre-
scribe, we use a simple model of a doctor-patient interaction to predict dif-
ferences between two experimental pricing conditions. In both conditions,
we provide vouchers that make simple malaria treatment free of charge to
the patient, while keeping the clinic’s and doctor’s financial incentives con-
stant. In the first condition, which we refer to as “Patient Voucher” or “P V ”,
both the doctor (she) and the patient (he) know that simple malaria treat-
ment is free. In the second condition (“Doctor Voucher” or “DV ”), only the
doctor knows about the discount going into the consultation, and she can
decide whether to reveal this information to the patient. In section 3, we
explain how we operationalized these pricing treatments. The remainder of
this section introduces the model and outlines the key testable predictions
that identify patient- and doctor-driven demand. We leave the full model
and formal analysis for Appendix A.




    15
      Our home survey found that just 6 percent of individuals who declined to purchase
at the clinic obtained an antimalarial from another source after leaving the clinic.
   16
      Appendix B gives additional detail on the health worker survey.


                                          12
2.1     Model Setup
Malaria Risk and Patient Preferences. Patient and doctor observe the
patient’s symptoms and assign probability π that the patient has malaria.17
Patients receive different utilities from taking either simple malaria treatment
(t = L) or severe malaria treatment (t = H ). The expected utility of t ∈
{L, H } is the expected benefit minus the cost:
                             Ut (π, Pt ) = πBt − Ct (Pt ) .
where Bt denotes the benefit from treatment t to a malaria-positive patient,
and Pt is the price of treatment to the patient, with PL ∈ {0, P }, depending
on whether the patient has a voucher. The full cost of treatment, Ct (Pt ),
includes both monetary and non-monetary costs such as side effects. We
assume CL (0) < CL (P ) < CH (PH ). We normalize the expected utility of no
treatment (t = N ) to zero.
    Figure 1, Panel A illustrates patient preferences with and without the
                                            L (·)
simple malaria discount. Assuming that CB    L
                                                    CH
                                                  < BH
                                                       , there are three regions
of the malaria probability π from low to high: individuals in N prefer no
treatment; patients in L prefer simple treatment; and those in H prefer severe
treatment over simple or no treatment. This captures that the benefits of
both simple and severe treatment – and thus the patient’s willingness to buy
them – are increasing in the severity of symptoms, but that the benefits of
severe treatment only outweigh the higher costs when symptoms are most
pronounced. Panel A also shows that when the voucher lowers the price PL
from P to 0, L expands, so that more patients prefer simple treatment over
both no and severe treatment.
    Note that there is an area within L, denoted L2, where patients prefer
simple treatment but would be willing to purchase severe (whereas in L1 they
would rather get no treatment at all). This means doctors could “oversell”
severe treatment to at least some patients.

Doctor Utility. Doctors may disagree with patients’ preferences. This is
expressed by a wedge ∆t , making the doctor’s direct utility from the patient
taking treatment t

                             Vt (π, Pt ) = Ut (π, Pt ) + ∆t .
  17
    The full model in the appendix allows for private signals about malaria risk; see also
below.


                                           13
The wedge may be positive, for example if the doctor does not care about
financial costs paid by the patient, or if the doctor’s utility factors in clinic
profits. A positive ∆t also captures non-financial reasons for preferring to
treat, such as risk aversion or fear of missing a diagnosis. On the other hand,
∆t could be negative, meaning the doctor gets less utility from treatment t
than the patient. This may reflect internalized costs of public provision of
health care, anticipated side effects, concerns about creating disease resis-
tance, or discomfort with disregarding official treatment policy.
    From low to high π , preferences Vt generate regions N      ˆ , and V
                                                             ˆ, L       ˆ where
doctors prefer no, simple, and severe treatment, respectively (assuming BLC  L
                                                                            +∆L
                                                                                 <
   CH
BH +∆H
        ). Unless doctors are completely altruistic and ∆L = ∆H = 0, these
regions will be shifted relative to N , L, and V . When ∆t is positive, the
threshold for treatment t shifts to the left, and doctors prefer more of this
treatment than patients. When ∆t is negative, the threshold for treatment t
shifts to the right and doctors have a lower preference for t than patients.18
    The cases of disagreement are our object of interest. In what follows, we
say patients drive (marginal) demand for treatment when the doctor pre-
scribes (and patients purchase) more aggressive treatment than the doctor
herself thinks is optimal. We say the doctor drives demand when the doctor
leads patients to purchase more (powerful) treatment than patients find op-
timal. The extent to which this drives misallocation will depend on whether
doctor versus patient preferences are better aligned with the social planner.
We can also assess misallocation empirically.

Gatekeeping Costs. If the doctor only cares about her own utility V ,
patients have no way to influence her prescription. To allow for a deviation
between what the doctor prescribes and what she truly prefers, we assume
that the doctor incurs a gatekeeping cost if she fails to prescribe the patient’s
most preferred treatment. This utility cost encompasses factors like the risk
of losing the patient to another clinic, any direct cost of dealing with an upset
or angry person in the office, and the effort cost of explaining the prescription
and cajoling the patient to comply. We assume that the gatekeeping cost the
doctor experiences is a share g of the patient’s lost utility relative to his

   18
     It may be natural to assume that doctors prefer “more treatment” or “less treatment”
than patients across both treatment options, but it is in principle possible to have wedges
with different signs, e.g. ∆L > 0 and ∆H < 0, which would imply that doctors want to
start treating with simple at lower levels of π , but prefer to switch to severe at a higher π .


                                              14
first-best treatment choice.19
    Patient-driven demand for simple treatment can occur for example when
the doctor intrinsically prefers no treatment, but the patient strongly prefers
simple, so that the gatekeeping costs exceed the utility loss the doctor incurs
by writing a prescription: gUL (π, PL ) > |UL (π, PL ) + ∆L |.20 In a similar
manner, there may be patient-driven demand for severe treatment over simple
or no treatment. Doctor-driven demand for severe treatment occurs when
the patient prefers simple treatment, but the doctor prefers, and prescribes,
severe treatment (note that this can only lead to a purchase when patients
are prepared to buy severe treatment, i.e. π ∈ L2). Not only are preferences
reversed between patients and doctors, but the utility the doctor gains from
prescribing severe must outweigh the gatekeeping cost. This requires a higher
wedge in doctor preferences in favor of severe over simple: ∆H > ∆L .
    The range of beliefs π for which prescribing no or severe treatment incurs
a gatekeeping cost increases when patients know that simple treatment is
free (in the Patient Voucher condition). This reflects the fact that patient
demand for simple treatment is higher when the price is lower; i.e. L ex-
pands. However, in the Doctor Voucher condition the doctor has the option
to selectively conceal the voucher to reduce her gatekeeping costs from pre-
scribing no treatment or severe treatment. We use this below to identify
patient-driven and doctor-driven demand.
    In the full model, we assume that patients and doctors have private signals
about the patient’s probability of being malaria positive. Unobserved signals
allow for uncertainty on both sides: patients do not know in advance what
prescription will be given, and the doctor does not know whether the patient
will buy the prescription she makes, as we observe frequently in the data.
This also means the doctor cannot be completely sure what gatekeeping cost
she will incur when prescribing treatment t. In the face of this uncertainty, we
permit doctors to avoid gatekeeping costs completely by letting the patient
choose between simple and severe treatment. If the doctor takes this option,
treatment is completely determined by patient preferences.21


   19
      Note that when a patient prefers not to receive treatment, he can always choose not
to purchase the prescription. This form of overprescription therefore incurs no gatekeeping
cost.
   20
      Note that this requires ∆L < 0 since UL (π, PL ) > 0 > UL (π, PL ) + ∆L .
   21
      In an earlier version of this paper, we discuss the model without this option (available
on request) – the key insights are unchanged.


                                             15
2.2    Comparing Doctor Voucher and Patient Voucher
       Treatments
Our first testable prediction focuses on detecting gatekeeping costs when
doctors and patients have different prescription preferences:

Prediction (1) If there are gatekeeping costs and doctors and patients do
    not always agree on the optimal prescription, the use of vouchers will
    be higher in P V than DV .

    Obtaining this result is straightforward: absent gatekeeping costs, the
doctor will always choose the action that maximizes her utility, V . This
action is independent of voucher condition, which in turn implies that voucher
use will be the same in P V and DV . When gatekeeping costs are present,
note two things: First, if it is optimal to use the voucher under DV , it is also
optimal to use the voucher under P V . Second, under DV the doctor can
strategically conceal the voucher from patients to whom she does not want
to prescribe simple treatment and thereby avoid the increased gatekeeping
costs when simple treatment is free. In P V , by contrast, these additional
gatekeeping costs may lead her to switch her prescription to simple, despite
her own preferences.
    We motivate the remaining predictions with two stylized examples, illus-
trated in Panels B and C of Figure 1. For simplicity, the examples consider
a case where gatekeeping costs are so high that the doctor is never willing to
override the patient’s preferences. Our predictions hold more generally; we
formally establish these results in Appendix A.

Prediction (2) If the share of patients who are prescribed and purchase
    any kind of malaria treatment is lower in DV than P V , this indicates
    patient-driven demand for (simple) treatment.

    This prediction maps to Figure 1, Panel B, where we consider the optimal
strategy of a doctor who would like to sell as little medication as possible
(i.e. nothing simple severe). The uppermost bar at the bottom of the
figure shows patient demand when PL = P (no voucher), while the middle
bar shows patient demand when PL = 0. Since gatekeeping costs are high,
this bar also reflects prescription and purchase outcomes under P V . The
lowermost bar shows outcomes per the doctor’s optimal strategy under DV .

                                       16
To minimize demand, the doctor conceals the voucher from lower risk patients
with π ∈ NP (at price PL = P these patients prefer to buy nothing, so no
gatekeeping cost is incurred), but shares the voucher and prescribes simple
malaria treatment to patients with π ∈ L1P ∪ L20 . This strategy lets the
doctor sell less malaria treatment under DV than under P V , in line with
prediction (2). In the appendix, we formally show that any substitution
from no malaria treatment in DV to malaria treatment in P V will be driven
by doctors who prefer not to prescribe, but give in to patient pressure when
the discount is known. Note that we formulate this as a prediction about the
demand for any treatment vs. no treatment, because, as the next prediction
shows, the change in the rate of simple treatment is ambiguous.

Prediction (3) If the share of patients who are prescribed and purchase
    severe treatment is lower in P V than DV , this indicates doctor-driven
    demand (for severe treatment).

    Figure 1, Panel C, illustrates the optimal strategy of a doctor who would
like to sell as much medication as possible (i.e. severe simple nothing).
This is our example of doctor-driven demand. In this case, the doctor will
share the voucher with lower risk patients with π ∈ L10 , since this maximizes
their demand for antimalarials. Thus, in contrast to Panel B, the share of
patients receiving any antimalarial does not differ between P V and DV .
However, in DV the doctor will conceal the voucher from higher risk patients
on the margin between preferring simple and severe treatment (π ∈ L20 ∩
HP ), thereby increasing the share of patients purchasing severe treatment
relative to P V . The appendix establishes the general case: that any shift
from simple in P V to severe in DV will be driven by doctors who prefer
severe, selling to patients who prefer simple.
    Our final prediction summarizes observations made above about the “marginal”
patients receiving vouchers under P V but not DV :

Prediction (4) Higher voucher use and substitution into prescription and
    purchase of simple treatment in PV are driven by lower malaria prob-
    abilities if patients drive demand for treatment, but by higher malaria
    probabilities if doctors drive demand for treatment.

    This prediction highlights that our experiment can only detect patient-
driven demand at the margin between no treatment and simple treatment,

                                    17
and doctor-driven demand at the margin between severe treatment and sim-
ple treatment. Within our model, doctor-driven demand at the margin of
any versus no treatment is not possible because doctors cannot force pa-
tients to buy unwanted treatment (although doctor-driven demand induce-
ment through influencing the patient’s beliefs, which is independent of the
experimental interventions and therefore not part of our model, might still
occur).
    While our predictions work on different margins, finding evidence of de-
mand driven by one side suggests that this side has a stronger preference
for treatment than the other. For example, in the presence of patient-driven
demand as in prediction (2), doctors have a weaker preference for simple
treatment than patients, and so there would not be doctor-driven demand
at this margin. Similarly, if there is doctor-driven demand for severe over
simple treatment as in prediction (3), it implies that patients have a weaker
preference for severe and so there would not be patient-driven demand for
severe treatment.
    All proofs are in Appendix A, where we also discuss additional results on
prescribing simple treatment without using the voucher, and on doctors who
are purely revenue motivated. We now turn to the experiment we designed
to identify the role of patient and doctor preferences in driving demand for
malaria treatment.


3     Experimental Design and Sample
3.1    Experimental Design
Sampling Frame. At the outset of our study, we obtained an administrative
list of all CSComs in the city of Bamako and nearby Kati and Kalaban Coro
in Koulikoro. After conducting a census of these clinics, we dropped clinics
that had closed or were more than 15 kilometers away from Bamako, and
removed 21 clinics that were working with a local NGO to offer subsidized,
improved care to mothers and children. This yielded a final sample of 60
clinics. Four health workers at each clinic were invited to attend a refresher
training that covered Mali’s official malaria guidelines and a hands-on train-
ing on how to administer an RDT. The training materials were prepared
by the research team in cooperation with researchers at the University of
Bamako, as well as the Programme National de Lutte contre le Paludisme

                                     18
(PNLP), Mali’s department of malaria control, and the Malaria Research
and Training Center Bamako. The trainings were conducted by five trainers
from the PNLP and one trainer from the regional health directorate. Clinics
were trained in six groups in November 2016; all clinics sent at least two and
on average 3.9 providers to the training.

Doctor and Patient Voucher Treatments. The simple malaria treat-
ment discounts were delivered via vouchers for a free course of ACTs, and
randomized within each clinic across six days during a two-week observation
period. The objective of the vouchers was to reduce the cost of treatment
for simple malaria, while leaving both the revenue to the clinic and the cost
of other types of treatment to the patient fixed. The vouchers were delivered
by trained “intervention officers” who arrived at the clinic in the morning.
Intervention officers did not perform any survey activities and were stationed
at a separate part of the clinic from the survey team. Intervention eligibility
was not tied to survey participation.
    In the Patient Voucher condition, vouchers were distributed directly to
patients when they arrived at the clinic, before they went to consult the doc-
tor. Patients and/or caregivers were informed that the voucher would pay for
simple malaria treatment (ACT tablets), provided the provider determined
that this was the appropriate course of action. Study officers were carefully
trained to limit potential endorsement effects: patients were informed that
the decision of whether to use the voucher was up to the treating health
worker, and that the patient could not and should not use the voucher if
they were not diagnosed with malaria. In the Doctor Voucher condition, pa-
tients were not informed about voucher availability before the consultation.
The vouchers were directly given to the provider(s) on staff, who could assign
the vouchers to patients at their discretion.
    Intervention officers always brought more than enough vouchers to cover
a day’s demand, so voucher rationing was not an issue. Both conditions used
the same voucher design, which required the provider’s signature to verify
the simple malaria diagnosis. The voucher was not valid unless used on the
same day. Signed vouchers were processed at the clinic pharmacy after the
consultation was complete, and a copy had to be returned to the interven-
tion officer by the patient, who verified that the correct medication and full
discount was received. At the end of the day, unused vouchers were collected
(in DV ), and the clinic was reimbursed for all vouchers redeemed that day
at their standard sales price.

                                      19
Information Treatments. The Doctor and Patient Voucher treatments
were cross-cut with a third within-clinic treatment. This “Patient Informa-
tion” (P I ) was conveyed through a short video, which emphasized that all
suspected malaria cases should be confirmed with a malaria test, noted that
RDTs should be available for free at the clinic, and described recommended
treatment for simple and severe malaria. The video also demonstrated an
RDT test and how to recognize a positive test result. Finally, half the clin-
ics were randomly selected to receive a more intensive training designed to
increase providers’ trust in the diagnostic accuracy of RDTs. Both informa-
tion treatments are analyzed in detail in a separate paper. We find limited
evidence that the information treatments impacted the size of P V and DV
treatment effects (see Appendix Tables B3 and B4 ).
    The within-clinic randomization and associated data collection were con-
ducted after the health worker trainings in November and December 2016,
covering the end of the rainy season and therefore the period of highest
malaria risk. We divided the 60 clinics into three cohorts of 20 based on
geography. Each of the three cohorts rotated through two weeks of data
collection and experimental intervention. Within each cohort, we randomly
assigned each clinic to one of 20 intervention schedules depicted in Figure
2. Each clinic received two Control days, two Doctor Voucher days, and two
Patient Voucher days. Although all clinics were informed of the upcoming
study activities and interventions in advance, clinic staff did not receive prior
notice of the actual intervention schedule – rather, our field officers informed
clinic staff of the day’s intervention on the morning of an observation day.

3.2    Data Collection
In order to identify the source of over-demand for malaria treatment, we re-
quire detailed data on prescription and treatment outcomes from patients at
the clinic, as well as patients’ true malaria status. Our main analysis uses
data from three different sources.

In-Clinic Survey. On clinic observation days, enumerators approached
all acutely ill patients at clinic intake and requested consent to participate
in the study. We classified a patient as “acutely ill” if they were feeling
sick and exhibited any of the following symptoms: fever, chills, excessive
sweating, nausea, vomiting, diarrhea, poor appetite, headache, cough, weak-

                                      20
ness, fatigue, or reduced consciousness. Before the doctors’ consultation,
enumerators recorded patients’ basic demographic details, symptoms, and
information on any prior treatment/diagnosis.22 After the consultation was
complete, the enumerator recorded details of all prescribed and purchased
medications, blood tests taken, and costs paid to the clinic. All told, we
completed 2,055 interviews.

Home Follow-Up Survey. We randomly selected 1,735 patients for a more
detailed follow-up survey conducted in the patient’s home the day after the
clinic visit, of whom 1,495 (86 percent) were successfully interviewed.23 The
home survey collected information on changes in the illness, any treatments
or tests obtained after the clinic visit, and recorded whether patients were
taking the medications that were purchased at the clinic. During this sur-
vey, 1,126 patients (75 percent of interviewed patients/65 percent of patients
targeted for the home survey) gave separate consent to take an RDT and
obtained a valid test result.24 Appendix Table B5 shows that there are no
significant differences in the probabilities of being selected for home survey,
taking the home survey, or having a valid home-based RDT by treatment
arm.

Administrative Voucher Data. Finally, we use administrative data on
voucher redemption to quantify how informing patients about vouchers im-
pacted redemption rates.

3.3     Background Characteristics and Randomization Ver-
        ification
Table 2 presents summary statistics on the study population and verifies
randomization balance. The first column of the table displays means and
standard deviations of baseline characteristics in the Control group (patients

  22
      We interviewed the patient whenever s/he was aged 15 or older and well enough to
take the interview; otherwise a guardian answered the survey.
   23
      From November 14-17 2016 we randomly selected up to six patients per clinic per
study day to interview at home. Between November 18 and December 30 2016 (the end
of the study) we randomly selected up to six patients per clinic per study day, and then
oversampled at clinics where more than six patients were seen, until 60 patients per study
day were sampled.
   24
      1,158 patients gave consent; 32 tests were inconclusive.


                                           21
who visited the clinic on days where no vouchers were available). Columns 2
and 3 present coefficients from the following regression specification, which
we use throughout the paper:

                    yict = β0 + βP P Vct + βD DVct + γt + εict                    (1)

where yict is the outcome of interest for patient i visiting clinic c on day
t, P Vct and DVct are dummy variables indicating the patient voucher and
doctor voucher interventions for clinic c on day t, and γt are date fixed
effects.25 The treatment rotation schedules were assigned at the clinic level,
so we cluster standard errors at this level as well. Since the relative difference
between giving vouchers to health workers versus patients will be of special
focus in the analysis, column 4 presents the p-value from an F-test of the
null hypothesis that βP = βD . Column 5 presents the p-value from an F-test
of the null hypothesis that the vouchers treatment effects are jointly equal
to zero (βP = βD = 0). Each row shows results from a separate regression.
    The first row of Table 2 shows that 5-6 acutely ill patients with eligible
symptoms visited the clinic on the average observation day. The voucher
treatments could affect outcomes not just through a direct effect on prescrip-
tion and purchase behavior, but also through a selection effect, whereby the
pool of patients visiting the clinic changes in response to treatment. We ran-
domized the voucher treatments within-clinic and administered them only
for a day at a time and at unannounced dates in order to minimize this se-
lection effect. Nonetheless there is a 0.63-0.76 patient increase in traffic on
Doctor and Patient Voucher days (significant at the 5 percent level in both
cases). Critically, there is no evidence of a differential increase on PV versus
DV days, which is most important for our testable predictions.
    Panels B and C show that the average characteristics of patients (and
their households) are similar across treatment groups. Forty percent of pa-
tients are male. The average patient is 17 years old, reports 3.6 symptoms
and has been sick for 4.3 days. The most commonly reported symptoms
are fever (80 percent of patients) and headache (62 percent of patients).
Thirty percent of patients report chills or excessive sweating. These symp-
toms are all commonly associated with malaria (Taylor et al., 2010), but only
25 percent of patients taking a home RDT tested positive. Individuals in the
Patient Voucher group were 6 percentage points less likely than those in the
  25
    We include date fixed effects because the proportion of P V , DV , and C days in the
sample varies depending on the date.


                                         22
Doctor Voucher treatment to have a positive RDT. Malaria-positive patients
are more likely to get treated for malaria and especially severe malaria, so
this difference would bias results against patient-driven and towards doctor-
driven demand. This is the opposite of what we find.
    Panel C of Table 2 shows that patients come from relatively poor families:
just under one-third of survey respondents (either the patient or the care-
giver accompanying the patient to the clinic, in the case of minor children)
are literate; households are quite large, with 10.3 members on average, and
monthly per capita income is just under CFA 22,000 ($36). Forty-one per-
cent of household members are children under the age of 15, and households
own 0.49 mosquito nets per capita.
    Appendix Table B6 shows that the characteristics of individuals selected
for the home survey versus not, and interviewed versus not (conditional on
selection), are balanced. There is, however, significant selection into taking
the home RDT: patients who consented to take the home test are sicker and
more likely to have purchased an antimalarial than patients who did not give
consent.

3.4     Predicting Malaria Risk
We would like to study how prescription and treatment outcomes vary with
patients’ underlying malaria risk (π in the model). We follow Cohen, Dupas
and Schaner (2015) and estimate the relationship between RDT result and
observed patient characteristics in the clinic survey with the following probit
specification:
                        E [posict | xict ] = Φ xict λ                       (2)
where posict is a dummy variable equal to one if a patient tests RDT positive
and xict is a vector including dummy variables for symptoms, days since onset
of illness, patient age, a dummy equal to one if the patient is under age 5, the
interaction between age and the under 5 dummy, patient gender, and patient
pregnancy status.26 The results of this regression, shown in Appendix Table

  26
     We expect malaria risk to change discontinuously at age 5 because the Malian gov-
ernment had a policy of seasonal malaria chemoprevention for children under 5 at the time
of the study (PMI, 2017). We also control for demographic characteristics indicative of
socio-economic status (which correlates with malaria risk), including the survey respon-
dent’s ethnicity, ability to speak French, literacy in French, and education. Finally we
control for the patient and respondent being different people. Results are similar if we


                                           23
B7, are used to impute a malaria risk for each patient. Appendix Figure B2
shows that the distribution of predicted malaria risk by treatment group is
very similar, and Table 2 finds no significant difference in average predicted
risk across treatment groups.
    The probit is estimated off the selected sample of individuals who con-
sented to a home-based test. If this selection is partly based on unobservables
that are informative about malaria status, the malaria risk of tested patients
may be different, conditional on recorded symptoms, from that of the av-
erage patient. However, our most important tests do not require estimated
risk to be unbiased; rather we require a proxy that is strongly correlated with
true risk. The regression assigns positive weight to symptoms known in the
medical literature to indicate malaria, like fever, chills, nausea, and headache
(CDC, 2015), which suggests that imputed malaria risk is correlated with true
risk. Consistent with this, predicted risk is highly correlated with malaria
treatment outcomes. Figure 3 uses data from the Control group to plot local
linear regressions, where predicted malaria risk is the running variable. The
share of individuals prescribed an antimalarial steadily increases with risk;
lower risk patients are more likely to receive simple treatment, while higher
risk patients are more likely to receive severe. As a proportion of prescrip-
tions, the share of patients who decline the prescribed treatment is decreasing
in predicted malaria risk for both simple and severe treatment. Appendix
Figure B3 shows that these patterns hold for those with valid home-based
RDT tests and those without.


4     Main Results
4.1     Empirical Approach
We use equation 1 as our core specification for estimating treatment effects.
To improve precision and address any concerns of patient selection on PV
and DV days, we augment equation 1 to control for additional clinic and
patient-level characteristics. To avoid specification searching, we select these
additional covariates using double lasso (Belloni, Chernozhukov and Hansen,
2014).27 Appendix Tables B8 and B9 show that results with no additional

omit the basic demographic characteristics from the probit.
   27
      Candidate controls include a set of clinic fixed effects, as well as all individual char-
acteristics included in Table 2, except those only measured at the home survey: symptom


                                             24
controls are very similar. In addition, Appendix Tables B10 and B11 indicate
that results are analogous when using covariates without the lasso procedure.
   We estimate heterogeneous treatment effects with respect to predicted
malaria risk using the following specification:
  yict = δ0 + δP H P V × highict + δDH DV × highict + δP L P V × lowict +(3)
         δDL DV × lowict + θhighict + xict α + λt + νict
where highict and lowict are dummy variables identifying patients with above/below
median predicted malaria risk, xict is the vector of clinic and individual-level
controls (again selected using double lasso), and λt are date fixed effects.
Since highict and lowict are generated regressors, we bootstrap standard er-
rors, clustering at the clinic level.28
    We now turn to the data to assess our theoretical predictions empirically.

4.2     Impacts on Antimalarial Prescriptions and Pur-
        chases
Overall Impacts. Table 3 shows treatment effects on voucher use, anti-
malarial prescriptions, and antimalarial purchases. We report the coefficients
on the treatment dummies βP and βD and the p-value of a two-sided test of
H0 : βP = βD . We also indicate the theoretical mechanism being tested (exis-
tence of gatekeeping costs, doctor-driven, or patient-driven overprescription –
GC, DD, or PD) and whether we find significant evidence of that mechanism
or not.
    The first column looks at voucher redemptions. We find strong, significant
evidence in favor of gatekeeping costs: voucher redemption is 9 percentage
dummies, illness duration (top-coded at the 99th percentile), patient age, an under 5
dummy, patient gender, a dummy to identify pregnant women, a dummy to identify pa-
tients who were also survey respondents, survey respondent gender, respondent ethnicity,
education, literacy and knowledge of French, and a dummy variable identifying patients
interviewed in the home follow-up survey. Missing values are dummied out and recoded
to the sample mean. We also include squared terms for patient age and illness duration,
as well as pairwise interactions between the aforementioned individual-level controls. We
partial out date fixed effects, so they are included in all regressions.
   28
      First we use double lasso to select covariates. Then we create 1,000 bootstrapped
samples by sampling 60 clinics with replacement. Then we re-estimate predicted malaria
risk in each bootstrap sample, reconstruct highict and lowict based on the new risk esti-
mate, and run the regression. Standard errors are based on the distribution of coefficients
across bootstrapped samples.


                                           25
points (35 percent) higher in P V compared to DV . The next two columns
study the share of patients who were either prescribed (column 2) or pur-
chased (column 3) any malaria treatment. The results are consistent with
patient-driven demand: more patients are prescribed and purchase malaria
treatment in P V . This means the subsidy led to stronger substitution out
of no treatment when patients were aware of the ACT discount. Individuals
who visited the clinic on P V days (relative to Control days) were 5.2 percent-
age points more likely to leave with a malaria prescription and 14 percentage
points more likely to purchase an antimalarial treatment. By contrast, DV
had no impact on prescriptions but increased purchases by 8.5 percentage
points. Thus, providers used vouchers for patients to whom they were giving
prescriptions already, but not to make new prescriptions. We reject equality
of βP and βD at the 10 and 5 percent levels for prescriptions and purchases
respectively, indicating significant evidence of patient-driven demand.
    The last two columns turn to severe malaria treatment to test for evidence
of doctor-driven demand. Both voucher treatments are associated with a de-
cline in severe malaria treatment. While we cannot reject equality of βP and
βD , point estimates are, if anything, larger in magnitude under DV . This is
the opposite of our theoretical prediction for doctor-driven demand for severe
treatment.

Impacts by Predicted Malaria Risk. If patients pressure providers to
prescribe, higher rates of voucher use and antimalarial consumption under
P V (relative to DV ) should be driven by patients at the lower end of the
predicted risk distribution. If providers drive demand for severe malaria
treatment, we should find substitution from severe to simple treatment in
P V when malaria risk is higher, because providers in this condition cannot
conceal the vouchers to maintain demand for severe treatment. Figure 4 plots
the relationship between treatment outcomes and predicted malaria risk by
voucher condition using local linear regressions.29 For reference, vertical
lines demarcate the 25th, 50th, and 75th percentiles of the predicted risk
distribution.
    Again, our predictions for patient-driven demand for simple treatment are
borne out in practice. Panel A shows that voucher use is persistently higher
in P V up to the 75th percentile of predicted malaria risk. Panel B shows

  29
    To account for clustering and the fact that predicted risk is a generated regressor,
confidence intervals are bootstrapped using the procedure described in sub-section 4.1.


                                          26
the share of patients purchasing any malaria treatment is higher in P V up
until the 75th percentile. Panel C shows that the Doctor Voucher group has
lower purchase rates for severe malaria at higher levels of predicted malaria
risk. This cuts against the prediction for doctor-driven demand.
    To quantify these patterns, Table 4 looks at heterogeneity in treatment
effects by above versus below median predicted risk. While rates of voucher
use among “high risk” patients in DV and P V are very similar, voucher use
by “low risk” patients is 15 percentage points (71 percent) higher when pa-
tients know about the vouchers before consulting the doctor. We see similar
patterns for any antimalarial prescription and purchase, where P V increases
low-risk prescriptions by 8.1 percentage points (16 percent) and purchases
by 9 percentage points (23 percent) relative to DV . These differences are
significant at the 5 percent level. Finally, columns 4 and 5 explore effects for
severe malaria treatment. While we formally reject H0 : δP H = δDH at the 1
percent level in column 5, the point estimates run counter to the theoretical
predictions for doctor driven demand.

Adherence to Treatment Guidelines. Why would use of severe malaria
treatment – especially among high risk patients – be lower in DV relative
to P V ? Figure 5 offers an explanation: providers were more likely to bend
the vouchers’ rules when patients knew about discounts. Specifically, Panel
A graphs the share of patients who purchased severe malaria treatment by
experimental arm. In the Control group, none of the 26 percent of patients
purchasing severe malaria treatment used a free ACT voucher. None of
the severe malaria patients in the Doctor and Patient Voucher treatments
should have used a voucher either, since the vouchers were only valid for
simple malaria prescriptions. This condition, however, relied on provider
certification as well as patients reporting their prescription correctly to the
intervention officer, so could in practice be violated.
    The third bar of Panel A shows that providers and patients almost never
broke the rules in DV : just one percent of patients in this condition pur-
chased severe malaria treatment and used a voucher. In contrast, over four
percent of patients in P V purchased severe malaria treatment and used a
voucher. Columns 1 and 2 in Appendix Table B12 verify that the rate of
rule breaking is significantly higher in P V versus DV . This behavior is also
consistent with patient-driven demand: providers are sometimes willing to
violate the terms of the voucher when patients push for the discounted treat-
ment.

                                     27
Not Using the Voucher for Simple Treatment. Columns 3 and 4 ex-
plore the inverse of this phenomenon: patients purchasing simple malaria
treatment without a voucher in DV versus P V . Non-voucher purchases may
have occurred because antimalarials other than ACTs were given, because
patients visited the clinic when the intervention officer was absent, or be-
cause the provider or patient forgot to apply the voucher. Additionally, our
model predicts that providers will sometimes intentionally prescribe simple
treatment without a voucher, and that they do so more often in DV (see pre-
diction (5) in Appendix A). Intuitively, a provider who prefers to prescribe
something different from simple treatment, but suspects that her patient
has a strong preference for it, may offer simple malaria treatment without
the voucher. This “hedging” lets her avoid gatekeeping costs if the patient
turns out to have a strong preference for simple treatment, while ensuring
the patient will not purchase simple if his preferences are not strong. The
assumption that the doctor cannot observe patient beliefs, π , is critical for
obtaining this result – otherwise there would be no reason to hedge. Consis-
tent with prediction 5, Appendix Table B12 shows that patients in DV were
three percentage points more likely than patients in P V to purchase simple
malaria treatment without a voucher (significant at the 10 percent level).

Other Explanations. Alternatively, it could be that providers in DV sim-
ply forgot about the vouchers in some cases, whereas in P V they were re-
minded by patients. This could lower voucher use in DV relative to P V
and explain why patients in DV are relatively less likely to receive/purchase
malaria treatment (Table 3). This mechanism cannot, however, explain why
the reduction in voucher use in DV is concentrated among low risk patients,
while prescriptions for severe treatment are if anything lower in DV than
in P V . With limited attention we would expect to see less use of vouchers
across the risk distribution under DV versus P V , meaning more use of severe
malaria treatment.
   Another possibility is that knowledge of the vouchers led patients to ex-
aggerate their symptoms, either intentionally or subconsciously. If this led
providers to overestimate malaria risk and therefore prescribe more anti-
malarials, this would amount to a form of patient-driven demand that op-
erates through changing providers’ direct utility, rather than gatekeeping



                                     28
costs.30 However, the fact that providers almost never violated the rules of
the vouchers in DV , but did in P V , suggests that gatekeeping costs matter
in our context.
    A final question is whether differences between P V and DV reflect dif-
ferential rates of patients “stockpiling” ACTs for future use. Differences
between P V and DV due to stockpiling would be consistent with patient-
driven demand, but may not occur in a long-run equilibrium where subsidies
are permanent. We check for stockpiling using our home survey data, where
we cycled through the list of all medications purchased at the clinic and
asked if the patient was currently taking the medication.31 Appendix Table
B13, column 2, shows that 94 percent of individuals who purchased ACTs for
simple malaria at the clinic reported taking ACTs during the home survey,
with no significant differences by voucher condition. Barring the possibility
of poor self-reporting, this suggests that stockpiling is not a significant driver
of the higher purchase rates in P V relative to DV .

Hawthorne and Experimenter Demand Effects. It is also important to
ask how our results might be influenced by Hawthorne or experimenter de-
mand effects. This could be especially relevant for doctor behavior; although
all clinics were formally told that the study was not a performance review and
that all data would be kept strictly confidential, clinics were informed that the
study was being conducted in collaboration with the National Malaria Con-
trol Program and the National Directorate of Health. Clinics also knew that
the pre-intervention training was part of the study; this training emphasized
the importance of adhering to official Malian policy when treating malaria,
meaning that only patients with confirmed positive test results should receive
an antimalarial prescription. We therefore expect Hawthorne effects to have
increased the likelihood of providers adhering to official treatment guidelines.
    Since treatments rotated within clinic, this would impact behavior in all
treatment arms, which would bias us away from finding P V -DV differences
in outcomes. Consistent with this (and the hypothesis that Hawthorne ef-
fects decay across observation days, see e.g. Leonard and Masatu (2010)),
  30
      A related possibility would be if providers perceived more patients in P V to be
“likely compliers” with prescribed simple malaria treatment, if for example more patients
explicitly requested simple malaria treatment.
   31
      ACT treatment courses last for three days and individuals were interviewed the day
after the clinic visit, so the treatment course should be ongoing at the time of the home
survey.


                                           29
Appendix Figure B4 shows that P V -DV differences are larger after the first
couple of clinic observation days. At the same time, we find that only 56
percent of patients prescribed an antimalarial on Control days received a
malaria test, and there is no significant difference in this “test given pre-
scription” rate across treatment arms, even though testing before treatment
was a major focus of the training and a requirement for using the voucher.
Overall this suggests the implicit monitoring that took place during the ex-
periment had limited effects on behavior that, if anything, bias us away from
detecting effects.
    A related worry is that P V -DV differences are driven by experimenter
demand effects on the patient side (e.g. if patients interpreted receiving the
voucher as a signal that “buying an antimalarial is encouraged by an outside
NGO”). While this would not invalidate our tests of gatekeeping costs and re-
luctant compliance by doctors, it is important for interpretation. We believe
experimenter demand effects were minimal in our setting for two reasons.
First, study staff were carefully trained to explain that the vouchers were
not valid without a provider’s signature, and that the doctor was the one
who would make the malaria diagnosis. Second, data from the patient infor-
mation treatment (P I ) suggests patients were not influenced by experimenter
demand effects. The P I video communicated that suspected malaria cases
should only be treated if they received a positive test. However, Appendix
Table B14 shows that P I actually decreased the share of total patients and
antimalarial prescription-holders who took a malaria test, while leaving the
share of patients who were prescribed an antimalarial unchanged. Panel B
shows there is no significant interaction effect with the voucher treatment
arms, which suggests that effects are not simply driven by patients refusing
tests in the hope of redeeming a voucher. These results are inconsistent with
experimenter demand effects changing patient preferences, but are consistent
with patients having preferences of their own about whether to get tested.

4.3    Do Doctors Use Vouchers to Increase Revenue?
Financial motives are one of the most cited reasons for doctor-driven demand
for drugs. We argued above that doctor-driven demand for severe treatment
would arise if (relative to patient preferences) providers get substantially
more utility from severe over simple treatment (∆H > ∆L ). A health worker
who cares mostly about revenue would be a plausible explanation for this,
since severe treatment is much pricier and incurs more high-margin service

                                     30
fees. Having said that, providers could opt to use the vouchers to prescribe
simple treatment instead of severe because this strategy leads to an almost
guaranteed purchase. In this case, providers might be revenue driven and
motivated to oversell treatment, but they may not be “driving demand”
(prescribing more treatment than patients would like) per our definition.
    Fortunately we can test whether providers are revenue motivated inde-
pendent of whether this leads doctors to drive demand for severe malaria
treatment: prediction (6) in Appendix A shows that vouchers will lead to an
increase in clinic revenue when doctors are purely revenue motivated, and
prediction (7) says that revenue-maximizing doctors will use the informa-
tional advantage in the DV treatment to increase clinic revenues relative to
P V .32
    In practice, we find no significant differences in (per-patient) clinic rev-
enues across treatment arms (Appendix Table B15). Moreover, the average
patient in both P V and DV pays roughly CFA 500 less than in the Control
group (significant at the 5 percent level for DV only). Lastly, from the re-
sults we already saw, ACT prescriptions and purchases for low-risk patients,
voucher use, and overall treatment rates in DV do not increase at the same
rate as in PV. This is not consistent with providers who are simply trying
to sell as much treatment as possible. All told, our results do not paint a
picture of providers primarily motivated by revenue, at least on the margin.

4.4     Quantifying Contributions to Overtreatment
What do doctors’ gatekeeping failures mean for overuse of antimalarials? Ide-
ally, we would use the home-based RDTs to directly assess the match between
prescriptions/purchases and the patient’s underlying malaria status. Unfor-
tunately, patients who consented to the home-based test were significantly
sicker and more likely to have malaria, based on our predicted risk metric,
than patients who refused the test (Appendix Table B6). Since P V -DV dif-
ferences in malaria prescriptions and purchases are concentrated among those
with the lowest predicted malaria risk, relying on the home-tested subsample
could understate P V -DV differences in appropriate treatment.
    As an alternative, we focus on “expected match quality” based on pre-
dicted malaria risk (denoted π ˜ ), which we decompose into two parts. First,

  32
     The greater number of patients on voucher days versus control days complicates our
test of prediction (6), so we place more weight on prediction (7).


                                          31
the probability of a “correct positive”, i.e. that an antimalarial is given to
a truly sick patient, is given by mpict = π  ˜ict × antimalict , where antimalict
is a dummy variable equal to one if patient i at clinic c on day t was either
prescribed or purchased an antimalarial of any type. Second, the proba-
bility that an antimalarial is withheld from a patient without malaria – a
“correct negative” – is mnict = (1 − π  ˜ict ) × (1 − antimalict ). We consider
treatment effects on both components separately, as well as the “expected
match”, which for the sake of illustration puts equal weight on both desirable
outcomes: matchict = mpict + mnict . In practice, the social planner may put
very different weights on these two outcomes.
    Table 5 shows treatment effects on the expected match and its two compo-
nents, for both antimalarial prescriptions and purchases. While the voucher
treatments did not increase the number of correct positives (mpict ) for pre-
scriptions, they did for purchases (1.8-2.3 percentage points, significant at
the 1 percent level). This amounts to a 15-19 percent increase over the mean
in the Control group and underscores the rationale of subsidizing health care:
some truly sick patients may not purchase treatment due to credit constraints
or because they value treatment below its cost.
    The higher treatment rates for malaria-positive patients come at a price,
however. While mn does not get worse for prescriptions in DV , it does in
P V (column 2); and mn significantly declines for both voucher conditions
when considering purchases. Column 5 shows that an estimated 45 percent
of patients do not have malaria and do not purchase an antimalarial in the
Control group. In P V , this share is 11 percentage points (24 percent) lower,
in DV it is 6.3 percentage points (14 percent) lower. The P V -DV difference
is significant at the 1 percent level, underscoring the concern that patient-
driven demand at low drug prices leads to overuse.
    Turning to the expected match, we estimate that 48 percent of patients
are prescribed the proper (malaria-related) treatment in the control group,
and 57 percent purchase it. In P V , this share is significantly lower for both
prescriptions (3.2 percentage points) and purchases (8.3 percentage points),
while it is lower for purchases only in DV (3.9 percentage points). The
8.3 percentage point lower match rate amounts to a 19 percent increase in
misallocation, which is substantial, and driven entirely by patient behavior:
doctor preferences (captured by the effect of DV on the expected match
for prescriptions) have no impact on misallocation, while distortions in pre-
scription behavior from patient-driven demand and the decline in patients’
non-purchase rate each account for roughly half of the additional misalloca-

                                       32
tion we observe with the ACT subsidy.
    Changes in mismatch rates relative to the Control need to be interpreted
with caution, given that more patients visit clinics on P V and DV days.
More generally, we cannot extrapolate our findings in the voucher treatments
to mismatch rates we would observe in a long-term equilibrium with subsi-
dized ACTs, since the composition of patients visiting the clinic would likely
change. However, we have no evidence of differential selection in P V and
DV , and the analysis indicates that patient-driven demand in our sample
significantly worsens the treatment-illness match.
    One weakness of our expected match measure is that it assumes no private
information on the part of doctors. Doctors may have access to private signals
– in particular malaria tests – that improve upon π ˜ when making prescrip-
tion decisions.33 This would bias our expected match measures downwards,
because private signals would improve the match of treatment to illness con-
ditional on π˜ . Assessing the consequences for estimated treatment effects
is more difficult, since this depends on what drives “marginal” prescriptions
in the different experimental conditions. For example, suppose all marginal
prescriptions in P V were driven by malaria negative patients. In this case,
we would falsely estimate a positive effect on mp (since marginal patients
          ˜ > 0) and underestimate the negative effect on mn.
will have π
    Appendix Table B16 uses the home-tested RDT subsample to get an idea
of how much private information skews our expected match estimates. We
find strikingly similar treatment effects on the expected and actual match for
the home-tested subsample.Control group means are also similar, at 0.48 for
the expected match and 0.51 for the actual match. Analogous numbers for
purchases are 0.56 and 0.61. While these numbers are consistent with some
private information, the impacts on match estimates are very modest, likely
because providers often do not use, or do not comply with malaria tests.


5         Conclusion
A critical challenge for developing country health systems is to craft policy
that simultaneously avoids undertreatment and overtreatment. Undertreat-
ment remains a major problem; for example, treatable conditions like pneu-
monia, diarrhea, and malaria account for 48 percent of global postnatal child
mortality (WHO, 2018). Overtreatment is problematic for a range of reasons:
    33
         This is captured by the private signal η in our model in the Appendix.


                                              33
poor households with very high marginal utilities of income waste precious
resources, overuse of vital drugs like antimalarials and antibiotics promotes
disease resistance, and unnecessary consumption of government-subsidized
care strains already-tight public budgets.
    Doctors (and other health workers) are the first line of defense against
overtreatment when care is subsidized. Yet a growing body of evidence makes
it abundantly clear that providers in both the public and private sectors are
imperfect gatekeepers, prescribing unnecessary treatment at very high rates.
It is therefore essential to understand the factors that limit doctors’ ability
to target treatments to the right patients.
    The existing literature has largely focused on doctor-driven overtreat-
ment. The implicit assumption is that doctors prefer to overtreat, for ex-
ample because they do not bear the cost of treatment, because it is in their
financial interest to increase revenue, or because they do not care to exert
effort to arrive at a better diagnosis. Our paper broadens the scope of investi-
gation by asking whether patients’ preferences may also drive overtreatment:
while the doctor would prefer not to prescribe a medication, she is willing to
do so in order to avoid the “gatekeeping costs” of dealing with an unhappy
or distrustful patient.
    We begin by developing a model that can generate both doctor- and
patient-driven demand for treatment, and delivers testable implications for
both channels (albeit at different margins of substitution). We test the
model’s predictions with an RCT that we conducted with 60 public health
clinics in Bamako, Mali. Our context is well-suited to studying this issue,
as misallocation of malaria treatment in public clinics is both verifiable (via
follow-up RDTs), and rampant: 75 percent of patients with confirmed RDTs
tested negative, yet 58 percent of these malaria-negative patients were pre-
scribed some form of malaria treatment.
    We find that vouchers for free malaria treatment are much more likely
to be used when both patients and providers (as opposed to just providers)
know about their availability, indicating that gatekeeping costs are empiri-
cally important. Patients with the lowest levels of malaria risk receive more
prescriptions and buy more treatment, which is consistent with patient-driven
demand for simple malaria treatment. We also find that this additional de-
mand significantly worsens overtreatment. On the other hand, we find no
evidence that physicians in the Doctor Voucher treatment exploit their infor-
mational advantage to upsell patients more expensive severe malaria treat-
ment, or increase clinic revenues more generally.

                                      34
    Our approach is agnostic about the underlying factors that might drive
doctors’ or patients’ preferences for administering treatment even when malaria
risk is low; and our results do not rule out the possibility that doctor prefer-
ences are driving a share of the baseline misallocation we see in our sample.
However, our findings do show that marginal misallocation from price subsi-
dies is attributable to patient preferences and behavior. Doctors are no more
likely to write an antimalarial prescription when only they know about the
price discounts; this margin only moves when both providers and patients are
informed. This evidence is not just important for enriching our understand-
ing of doctor preferences, patient preferences, and how they interact during
the consultation process; our results also contribute to the ongoing debate
over how to price health products in the developing world. Our findings im-
ply that cost sharing in public sector settings can reduce overtreatment by
reducing patient demand, but at the cost of fewer malaria patients getting
appropriate treatment. Policies and tools that improve physicians’ gatekeep-
ing capabilities could help sustain higher subsidies while limiting overuse;
additional research on the design and efficacy of such interventions is an
important area for future work.
    Our model highlights how the effects of gatekeeping costs will be deter-
mined by the nature of mismatch between doctor and patient preferences –
in our setting patients prefer more simple malaria treatment than providers
would like to prescribe. Doctor-patient preference gaps may be different in
other contexts where providers face different incentives and patients face dif-
ferent prices. Both our model and empirical approach can be adapted to
other settings to generate customized evidence – and a more nuanced under-
standing of how gatekeeping failures mediate healthcare outcomes worldwide.




                                      35
References
Achan, Jane, Ambrose O Talisuna, Annette Erhart, Adoke Yeka,
 James K Tibenderana, Frederick N Baliraine, Philip J Rosenthal,
 and Umberto D’Alessandro. 2011. “Quinine, an Old Anti-Malarial
 Drug in a Modern World: Role in the Treatment of Malaria.” Malaria
 Journal, 10(1): 144.

Ansah, Evelyn K, Solomon Narh-Bana, Michael Epokor, Sam-
 son Akanpigbiam, Alberta Amu Quartey, John Gyapong, and
 Christopher JM Whitty. 2010. “Rapid testing for malaria in settings
 where microscopy is available and peripheral clinics where only presump-
 tive treatment is available: A randomised controlled trial in Ghana.” BMJ,
 340: c930.

Arrow, K., C. Panosian, and H. Gelband, ed. 2004. Saving Lives, Buy-
 ing Time: Economics of Malaria Drugs in an Age of Resistance. Washing-
 ton, DC:Institute of Medicine: National Academies Press.

Ashraf, Nava, James Berry, and Jesse M. Shapiro. 2010. “Can Higher
 Prices Stimulate Product Use? Evidence from a Field Experiment in Zam-
 bia.” American Economic Review, 100(5): 2383–2413.

Banerjee, Abhijit V, Esther Duflo, and Rachel Glennerster. 2008.
 “Putting a Band-Aid on a Corpse: Incentives for Nurses in the Indian Pub-
 lic Health Care System.” Journal of the European Economic Association,
 6(2-3): 487–500.

Belloni, Alexandre, Victor Chernozhukov, and Christian Hansen.
 2014. “High-Dimensional Methods and Inference on Structural and Treat-
 ment Effects.” Journal of Economic Perspectives, 28(2): 29–50.

Bisoffi, Zeno, Bienvenu Sodiomon Sirima, Andrea Angheben, Clau-
  dia Lodesani, Federico Gobbi, Halidou Tinto, and Jef Van den
  Ende. 2009. “Rapid malaria diagnostic tests vs. clinical management of
  malaria in rural Burkina Faso: safety and effect on clinical decisions. A
  randomized trial.” Tropical Medicine & International Health, 14(5): 491–
  498.

CDC. 2015. “Malaria.” http://www.cdc.gov/malaria/about/disease.html.

                                    36
Cohen, Jessica, and Pascaline Dupas. 2010. “Free Distribution or Cost-
 Sharing? Evidence from a Randomized Malaria Prevention Experiment.”
 The Quarterly Journal of Economics, 125(1): 1–45.

Cohen, Jessica, Pascaline Dupas, and Simone Schaner. 2015. “Price
 Subsidies, Diagnostic Tests, and Targeting of Malaria Treatment: Evi-
 dence from a Randomized Controlled Trial.” American Economic Review,
 105(2): 609–45.

Currie, Janet, Wanchuan Lin, and Juanjuan Meng. 2014. “Addressing
 Antibiotic Abuse in China: An Experimental Audit Study.” Journal of
 Development Economics, 110: 39–51.

Das, Jishnu, Alaka Holla, Aakash Mohpal, and Karthik Muralidha-
 ran. 2016. “Quality and Accountability in Health Care Delivery: Audit-
 Study Evidence from Primary Care in India.” American Economic Review,
 106(12): 3765–99.

Das, Jishnu, and Jeffrey Hammer. 2014. “Quality of Primary Care in
 Low-Income Countries: Facts and Economics.” 6: 525–553.

Das, Jishnu, Jeffrey Hammer, and Kenneth Leonard. 2008. “The
 Quality of Medical Advice in Low-Income Countries.” The Journal of Eco-
 nomic Perspectives, 22(2): pp. 93–114.

Dupas, Pascaline. 2014. “Short-Run Subsidies and Long-Run Adoption of
 New Health Products: Evidence from a Field Experiment.” Econometrica,
 82(1): 191–228.

Dupas, Pascaline, and Edward Miguel. 2017. “Impacts and Determi-
 nants of Health Levels in Low-Income Countries.” In Handbook of Eco-
 nomic Field Experiments. Vol. 2, 3–93. Elsevier.

Escot, Fabrice. 2012. “Bilan des diagnostics socio-sanitaires et des
                                e de la r´
  ressources humaines de la sant´        egion de Sikasso.”

Fleming-Dutra, Katherine E, Adam L Hersh, Daniel J Shapiro,
  Monina Bartoces, Eva A Enns, Thomas M File, Jonathan A
  Finkelstein, Jeffrey S Gerber, David Y Hyun, Jeffrey A Linder,
  et al. 2016. “Prevalence of Inappropriate Antibiotic Prescriptions among
  US Ambulatory Care Visits, 2010-2011.” JAMA, 315(17): 1864–1873.

                                   37
Hamer, Davidson H, Micky Ndhlovu, Dejan Zurovac, Matthew
 Fox, Kojo Yeboah-Antwi, Pascalina Chanda, Naawa Sip-
 ilinyambe, Jonathon L Simon, and Robert W Snow. 2007. “Im-
 proved diagnostic testing and malaria treatment practices in Zambia.”
 JAMA, 297(20): 2227–2231.
Humar, Atul, Colin Ohrt, Mary Anne Harrington, Dylan Pillai,
 and Kevin C Kain. 1997. “ParaSight R F Test Compared with the
 Polymerase Chain Reaction and Microscopy for the Diagnosis of Plasmod-
 ium Falciparum Malaria in Travelers.” The American Journal of Tropical
 Medicine and Hygiene, 56(1): 44–48.
IHME. 2018. “Global Burden of Disease Study 2017 (GBD 2017) Results.”
  Online, Accessed May 22, 2020.
Iizuka, Toshiaki. 2012. “Physician Agency and Adoption of Generic Phar-
  maceuticals.” American Economic Review, 102(6): 2826–58.
Kon´e, Diakalia, Drissa Coulibaly, Ogobara Doumbo, Ibrahima-Soc´     e
 Fall, Eliud Kibuchi, Bernard Mitto, Gilbert Sang, David Kyalo,
 Robert W. Snow, and Abdisalan M. Noor. 2015. An Epidemiolog-
 ical Profile of Malaria in Mali. Programme National de Lutte contre le
 Paludisme (PNLP).
Kotwani, Anita, Chand Wattal, Shashi Katewa, PC Joshi, and
 Kathleen Holloway. 2010. “Factors Influencing Primary Care Physicians
 to Prescribe Antibiotics in Delhi India.” Family Practice, 27(6): 684–690.
Kravitz, Richard L, Ronald M Epstein, Mitchell D Feldman,
 Carol E Franz, Rahman Azari, Michael S Wilkes, Ladson Hinton,
 and Peter Franks. 2005. “Influence of Patients’ Requests for Direct-to-
 Consumer Advertised Antidepressants: A Randomized Controlled Trial.”
 JAMA, 293(16): 1995–2002.
Kremer, Michael, Jessica Leino, Edward Miguel, and Alix Zwane.
 2011. “Spring Cleaning: Rural Water Impacts, Valuation, and Property
 Rights Institutions.” The Quarterly Journal of Economics, 126(1).
Kyabayinze, Daniel J, James K Tibenderana, George W Odong,
 John B Rwakimari, and Helen Counihan. 2008. “Operational Accu-
 racy and Comparative Persistent Antigenicity of HRP2 Rapid Diagnostic

                                    38
  Tests for Plasmodium Falciparum Malaria in a Hyperendemic Region of
  Uganda.” Malaria Journal, 7(1): 221.

Leonard, Kenneth L., and Melkiory C. Masatu. 2010. “Using the
  Hawthorne effect to examine the gap between a doctor’s best possible prac-
  tice and actual performance.” Journal of Development Economics, 93: 226–
  234.

Linder, Jeffrey A, Jason N Doctor, Mark W Friedberg,
  Harry Reyes Nieva, Caroline Birks, Daniella Meeker, and
  Craig R Fox. 2014. “Time of Day and the Decision to Prescribe An-
  tibiotics.” JAMA Internal Medicine, 174(12): 2029–2031.

Li, Yongbin, Jing Xu, Fang Wang, Bin Wang, Liqun Liu, Wanli
  Hou, Hong Fan, Yeqing Tong, Juan Zhang, and Zuxun Lu. 2012.
  “Overprescribing in China, Driven by Financial Incentives, Results in Very
  High Use of Antibiotics, injections, and corticosteroids.” Health affairs,
  31(5): 1075–1082.

      ere de la Sant´
Minist`               e. 2013. Politique Nationale De Lutte Contre Le
 Paludisme. Republique du Mali.

PMI. 2015. “President’s Malaria Initiative Mali: Malaria Operational Plan
 FY 2016.” USAID.

PMI. 2016. “President’s Malaria Initiative Mali: Malaria Operational Plan
 FY 2017.” USAID.

PMI. 2017. “President’s Malaria Initiative Mali: Malaria Operational Plan
 FY 2018.” USAID.

PNLP, INSTAT, INFO-STAT, INRSP, and ICF International. 2016.
 “Enquete sur les Indicateurs du Paludisme au Mali (EIPM) 2015.” IN-
 STAT, INFO-STAT and ICF International, Rockville, Maryland, USA.

Prendergast, Canice. 2003. “The Limits of Bureaucratic Efficiency.” Jour-
 nal of Political Economy, 111(5): 929–958.

Reyburn, H, R Mbatia, C Drakeley, I Carneiro, E Mwakasungula,
 O Mwerinde, K Saganda, J Shao, A Kitua, R Olomi, BM Green-
 wood, and CJ Whitty. 2004. “Overdiagnosis of Malaria in Patients with

                                    39
  Severe Febrile Illness in Tanzania: A Prospective Study.” British Medical
  Journal, 329(7476): 1212.

Sautmann, Anja, Samuel Brown, and Mark Dean. 2017. “Subsidies,
  Information, and the Timing of Children’s Health Care in Mali.”

Schwartz, Lisa M., and Steven Woloshin. 2019. “Medical Marketing in
  the United States, 1997-2016.” JAMA, 321(1): 80–96.

Tarozzi, Alessandro, Aprajit Mahajan, Brian Blackburn, Dan
  Kopf, Lakshmi Krishnan, and Joanne Yoong. 2014. “Micro-loans,
  Insecticide-Treated Bednets and Malaria: Evidence from a randomized
  controlled trial in Orissa (India).” The American Economic Review,
  104(7): 1909–41.

Taylor, Steve M, Malcolm E Molyneux, David L Simel, Steven R
  Meshnick, and Jonathan J Juliano. 2010. “Does This Patient Have
  Malaria?” JAMA: The Journal of the American Medical Association,
  304(18): 2048–2056.

Tour´e, Laurence. 2015. “User fee exemption policies in Mali: sustainability
  jeopardized by the malfunctioning of the health system.” BMC Health
  Services Research, 15(3): S8.

Trampuz, Andrej, Matjaz Jereb, Igor Muzlovic, and Rajesh M
  Prabhu. 2003. “Clinical Review: Severe Malaria.” Critical Care, 7(4): 315.

Whitehead, Margaret, G¨   oran Dahlgren, and Timothy Evans. 2001.
 “Equity and Health Sector Reforms: Can Low-Income Countries Escape
 the Medical Poverty Trap?” The Lancet, 358(9284): 833–836.

WHO. 2014. “Emergence and Spread of Artemisinin Resistance Calls for
 Intensified Efforts to Withdraw Oral Artemisinin-Based Monotherapy from
 the Market.” World Health Organization.

WHO. 2017. “Malaria rapid diagnostic test performance: results of WHO
 product testing of malaria RDTs: round 7 (2015-2016).” World Health
 Organization.

WHO. 2018. “MCEE Methods and Data Sources for Child Causes of Death
 2000-2016.”

                                    40
Wongsrichanalai, Chansuda, Mazie J Barcus, Sinuon Muth, Awal-
 ludin Sutamihardja, and Walther H Wernsdorfer. 2007. “A Re-
 view of Malaria Diagnostic Tools: Microscopy and Rapid Diagnostic
 Test (RDT).” The American Journal of Tropical Medicine and Hygiene,
 77(6 Suppl): 119–127.




                                41
                           Figure 1: Patient Preferences and Doctor Voucher-Sharing Strategies


          A. Effect of Simple Malaria Price Reduction on                      B. Voucher Sharing Strategy That Minimizes                             C. Voucher Sharing Strategy That Maximizes
                         Patient Preferences                                        Patient Demand for Treatment                                           Patient Demand for Treatment

                                                            πBH -CH                                                                  πBH -CH                                                                πBH -CH

                                                            πBL -C(0)                                                                πBL -C(0)                                                              πBL -C(0)

                                                            πBL -C(P)                                                                πBL -C(P)                                                              πBL -C(P)
42




      0                          πH                     1     π         0                             πH                         1     π         0                           πH                         1     π




               NP          L1P    L2P         HP                                   NP           L1P    L2P             HP                                 NP           L1P    L2P           HP

          N0         L10                L20        H0                         N0          L10                   L20         H0                       N0          L10                L20          H0

                                                                                   N                        L               H                        N                 L                    H
                                                                            conceal voucher                reveal voucher                                  reveal voucher             conceal voucher



     Notes: Diagonal lines graph patient utility net of costs. Shaded bars with P and 0 subscripts indicate patient
     preferences when the price of simple malaria treatment is P , and 0 respectively. N denotes “no antimalarial
     purchase”, L1 denotes “simple malaria treatment, unwilling to purchase severe malaria treatment”, L2
     denotes “simple malaria treatment, willing to purchase severe malaria treatment”, and H denotes “severe
     malaria treatment”. The bottom shaded bars in panels B and C show the outcome when the doctor adopts
     the strategy specified in the panel title.
                                  Figure 2: Within-CSCOM Randomization Design

                                            WEEK 1                                                           WEEK 2
      CSCOM
      Number     Mon       Tues     Weds      Thurs      Fri       Sat        Sun     Mon     Tues   Weds     Thurs    Fri     Sat    Sun
        1         C         --       PV         --       DV         --         --      PI      --    PI-PV      --    PI-DV     --     --
        2        DV         --        C         --       PV         --         --    PI-DV     --      PI       --    PI-PV     --     --
        3        PV         --       DV         --        C         --         --    PI-PV     --    PI-DV      --      PI      --     --
        4         C         --       DV         --       PV         --         --      PI      --    PI-DV      --    PI-PV     --     --
        5        DV         --       PV         --        C         --         --    PI-DV     --    PI-PV      --      PI      --     --
        6         PI        --      PI-PV       --      PI-DV       --         --      C       --     PV        --     DV       --     --
        7       PI-DV       --        PI        --      PI-PV       --         --     DV       --      C        --     PV       --     --
        8       PI-PV       --      PI-DV       --        PI        --         --     PV       --     DV        --      C       --     --
        9         PI        --      PI-DV       --      PI-PV       --         --      C       --     DV        --     PV       --     --
        10      PI-DV       --      PI-PV       --        PI        --         --     DV       --     PV        --      C       --     --
        11        --        C         --       PV         --       DV          --      --      PI      --     PI-PV     --    PI-DV    --
        12        --       DV         --        C         --       PV          --      --    PI-DV     --       PI      --    PI-PV    --
43




        13        --       PV         --       DV         --        C          --      --    PI-PV     --     PI-DV     --      PI     --
        14        --        C         --       DV         --       PV          --      --      PI      --     PI-DV     --    PI-PV    --
        15        --       DV         --       PV         --        C          --      --    PI-DV     --     PI-PV     --      PI     --
        16        --        PI        --      PI-PV       --      PI-DV        --      --      C       --      PV       --     DV      --
        17        --      PI-DV       --        PI        --      PI-PV        --      --     DV       --       C       --     PV      --
        18        --      PI-PV       --      PI-DV       --        PI         --      --     PV       --      DV       --      C      --
        19        --        PI        --      PI-DV       --      PI-PV        --      --      C       --      DV       --     PV      --
        20        --      PI-DV       --      PI-PV       --        PI         --      --     DV       --      PV       --      C      --


                                       LEGEND
         --    No data collection or interventions at CSCOM
         C     Data collection at CSCOM, no interventions
        DV     Doctor vouchers and data collection at CSCOM
        PV     Patient vouchers and data collection at CSCOM
         PI    Patient information and data collection at CSCOM
       PI-DV   Patient information, doctor vouchers, and data collection at CSCOM
       PI-PV   Patient information, patient vouchers, and data collection at CSCOM


     Notes: Before the interventions launched, clinic staff received trainings on RDTs on November 2, 3, and
     4 2016. The interventions listed above ran between November 14-December 30 2016 in three two-week
     blocks, with 20 CSCOMs active in each two-week block.
                            Figure 3: Treatment Outcomes by Predicted Malaria Risk in Control Group

                                      A. No Malaria Treatment                                                          B. Simple Malaria Treatment                                                     C. Severe Malaria Treatment
                   .8




                                                                                                        .8




                                                                                                                                                                                        .8
                   .6




                                                                                                        .6




                                                                                                                                                                                        .6
      Share Patients




                                                                                           Share Patients




                                                                                                                                                                           Share Patients
                   .4




                                                                                                        .4




                                                                                                                                                                                        .4
44




                   .2




                                                                                                        .2




                                                                                                                                                                                        .2
                   0




                                                                                                        0




                                                                                                                                                                                        0
                        0            .2                    .4                         .6                     0           .2                   .4                      .6                     0           .2                   .4                      .6
                                      Predicted Malaria Risk                                                             Predicted Malaria Risk                                                          Predicted Malaria Risk
                              No Treat. Prescribed              No Treat. Purchased                              Treatment Prescribed           Treatment Purchased                              Treatment Prescribed           Treatment Purchased




     Notes: Results from local linear regressions. Regressions are run on the full sample, but graphs omit
     results for top and bottom 2.5 percent of malaria risk distribution to avoid influence of outliers. Vertical
     dashed lines indicate 25th, 50th, and 75th percentiles of predicted malaria risk respectively.
                                 Figure 4: Treatment Outcomes by Predicted Malaria Risk and Voucher Group

                                        Used Voucher                                          Purchased Any Malaria Treatment                                      Purchased Severe Malaria Treatment




                                                                                                                                                          .8
                        .8




                                                                                     .8




                                                                                                                                                          .6
                        .6
       Share Patients




                                                                    Share Patients




                                                                                                                                         Share Patients
                                                                                     .6




                                                                                                                                                          .4
                        .4




                                                                                     .4
45




                                                                                                                                                          .2
                        .2




                                                                                     .2




                                                                                                                                                          0
                        0




                                                                                     0
                             0          .2             .4      .6                         0           .2             .4             .6                         0             .2             .4          .6
                                      Predicted Malaria Risk                                        Predicted Malaria Risk                                                 Predicted Malaria Risk


                                                                                          Doctor Voucher                  Patient Voucher


     Notes: Results from local linear regressions. Dotted lines give 90 percent confidence intervals. The standard
     errors used to calculate confidence intervals are bootstrapped by re-sampling clinics and recalculating
     predicted malaria risk on each of 1,000 bootstrap replications. Regressions are run on the full sample, but
     graphs omit results for top and bottom 2.5 percent of malaria risk distribution to avoid influence of outliers.
     Vertical dashed lines indicate 25th, 50th, and 75th percentiles of predicted malaria risk respectively.
                            Figure 5: Use of Vouchers for Simple and Severe Malaria by Treatment Arm

                                        A. Simple Malaria Treatment                                                           B. Severe Malaria Treatment
                       .4




                                                                                                          .4
                       .3




                                                                                                          .3
      Share Patients




                                                                                         Share Patients
46




                       .2




                                                                                                          .2
                       .1




                                                                                                          .1
                       0




                                                                                                          0
                              Control        Patient Voucher          Doctor Voucher                             Control           Patient Voucher          Doctor Voucher


                                                                       No Voucher Used                         Voucher Used


     Notes: Each panel graphs the share of patients receiving the specified treatment by voucher condition.
                                       Table 1: Descriptive Statistics from Control Group

                                                                         (1)      (2)         (3)      (4)       (5)
                                                                           Full Sample      Has Valid RDT (Home Survey)
                                                                                              All       RDT         RDT
                                                                         N       Mean       Patients   Positive    Negative
      Panel A. Malaria Beliefs and Treatment Outcomes
       Respondent Suspects Illness is Malaria (Pre-Consultation)         628     0.551       0.610       0.791       0.550
       Simple Malaria Treatment: Prescribed                              627     0.313       0.295       0.163       0.338
       Simple Malaria Treatment: Purchased                               627     0.206       0.197       0.140       0.215
       Severe Malaria Treatment: Prescribed                              627     0.303       0.327       0.593       0.238
       Severe Malaria Treatment: Purchased                               627     0.255       0.283       0.535       0.200
      Panel B. Other Treatment Outcomes
       Purchased Antibiotics                                             627     0.426       0.442       0.430       0.446
       Received Injection or IV                                          628     0.398       0.454       0.674       0.381
       Patient Referred to Hospital or Placed Under Observation          628     0.139       0.179       0.395       0.108
      Panel C. Patient Costs
47




       Total Cost No Malaria Treatment (CFA)                             301 2793.349           –          –           –
       Total Cost Simple Malaria (CFA)                                   126 4474.873           –          –           –
       Total Cost Severe Malaria (CFA)                                   147 10352.714          –          –           –
      Notes : Column 1 reports the number of non-missing observations of variables among all individuals in the Control group.
     Column 2 reports means for this sample. Columns 3-6 limit the sample to Control group individuals who were administered
     a valid RDT test at home, with column 3 reporting means for all patients with a valid test, column 4 reporting means for
     patients with a positive test, and column 5 reporting means for patients with a negative test. Sample sizes for columns 3, 4,
     and 5 are 346, 86, and 260 respectively. Variables measured in CFA top-coded at the 99th percentile. CFA610 ≈ USD1.
                      Table 2: Demographic Characteristics and Randomization Verification

                                              (1)          (2)          (3)               (4)              (5)       (6)
                                                         Regression Coefficients                  P-Values
                                        Control     Patient      Doctor     Joint Test              Joint Test
                                          Mean      Voucher     Voucher      PV=DV                  PV=DV=0           N
 A. Sample Frame (Clinic × Day-Level Observations)                                                                          .
                                                          ∗∗          ∗∗                                         ∗
  Number Eligible Logged Patients         5.233     0.757        0.634        0.725                   0.054          360
                                         [3.322]    (0.348)     (0.312)
 B. Patient Characteristics (Individual-Level Observations)                                                                 .
  Number of symptoms                      3.559       0.023      -0.090       0.114                    0.270         2055
                                         [1.578]    (0.081)     (0.086)
  Fever                                   0.801      0.045∗       0.012       0.096∗                   0.130         2055
                                         [0.400]    (0.024)     (0.022)
  Chills or Excessive Sweating            0.298      -0.045     -0.051∗∗      0.782                   0.092∗         2055
                                         [0.458]    (0.027)     (0.024)
  Nausea, Vomiting, or Diarrhea           0.497      -0.004      -0.024       0.455                    0.599         2055
                                         [0.500]    (0.028)     (0.026)
  Poor Appetite                           0.505      -0.023      -0.040       0.501                    0.292         2055
                                         [0.500]    (0.031)     (0.026)
  Headache                                0.621      -0.006      -0.007       0.963                    0.951         2055
                                         [0.486]    (0.023)     (0.023)
  Cough                                   0.363       0.027      -0.005       0.230                    0.403         2055
                                         [0.481]    (0.025)     (0.028)
  Weakness/Fatigue                        0.475       0.029       0.025       0.881                    0.538         2055
                                         [0.500]    (0.028)     (0.027)
  Duration of Illness in Days             4.299       0.180      -0.007       0.444                    0.648         2055
                                         [4.276]    (0.233)     (0.268)
                                                          ∗∗           ∗                                         ∗
  Age                                    16.603     1.545        1.282        0.760                   0.078          2055
                                        [15.096]    (0.760)     (0.740)
  Under 5 Years Old                       0.314      -0.041      -0.034       0.746                    0.179         2055
                                         [0.464]    (0.025)     (0.022)
                                                                      ∗∗
  Male                                    0.395       0.031      0.058        0.377                    0.102         2055
                                         [0.489]    (0.030)     (0.026)
  Patient is Pregnant                     0.105       0.003      -0.016       0.331                    0.567         1139
                                         [0.307]    (0.022)     (0.020)
  Positive RDT (home)                     0.249      -0.037       0.028      0.015∗∗                  0.051      ∗
                                                                                                                     1126
                                         [0.433]    (0.027)     (0.027)
  Predicted Malaria Risk                  0.212      0.008        0.008       0.981                    0.361         2055
                                         [0.160]    (0.007)     (0.007)
 C. Respondent and Household Characteristics (Individual-Level Observations)                                                .
  Patient Answered Clinic Survey          0.470       0.003       0.003       0.990                    0.990         2055
                                         [0.499]    (0.026)     (0.024)
  Male                                    0.264      0.017       0.039∗       0.350                    0.213         2055
                                         [0.441]    (0.024)     (0.022)
  Bambara                                 0.377      0.006        0.011       0.850                    0.926         2053
                                         [0.485]    (0.023)     (0.027)
  Speaks French                           0.518      -0.015       0.000       0.550                    0.822         2055
                                         [0.500]    (0.031)     (0.027)
  Literate (in French)                    0.264      -0.005      -0.012       0.772                    0.901         2055
                                         [0.441]    (0.030)     (0.028)
  Primary School or Less                  0.424       0.046       0.047       0.962                    0.248         2055
                                         [0.495]    (0.033)     (0.030)
                  +
  Household Size                         10.308      0.347        0.729       0.500                    0.483         1491
                                         [8.111]    (0.506)     (0.602)
                       +
  Share HH Under 15                       0.412       0.017       0.013       0.777                    0.397         1485
                                         [0.191]    (0.013)     (0.014)
                                +
  Share HH Members Working                0.258      -0.004      -0.002       0.886                    0.937         1485
                                         [0.188]    (0.010)     (0.011)
                              +                                          ∗∗
  Monthly income per capita            21987.613 -2143.183 -2928.525          0.522                    0.135         1432
                                      [25298.328] (1388.442) (1459.962)
                         +
  Rental Value of Home                 63823.469 -2452.218     -5053.802      0.607                    0.694         1469
                                      [90722.250] (5457.584) (5934.901)
                            +
  Mosquito Nets Per Capita                0.491      -0.016      -0.012       0.830                    0.766         1482
                                         [0.342]    (0.022)     (0.021)
 Notes : Robust standard errors clustered at the clinic level in parentheses. All regressions include clinic visit date fixed
effects. + indicates that variable was recorded in the home survey only. Variables measured in CFA and duration of
illness top-coded at the 99th percentile. CFA610 ≈ USD1. *, **, and *** denote statistical significance at the 10, 5, and
1 percent levels respectively.




                                                           48
                              Table 3: Impacts on Malaria Treatment Outcomes

                                                (1)        (2)        (3)                 (4)        (5)
                                                         Any Malaria Treatment         Severe Malaria Treatment

                                              Used       Prescribed Purchased          Prescribed Purchased
                                             Voucher
 βP : Patient Voucher                        0.35∗∗∗      0.052∗∗       0.14∗∗∗        -0.046∗∗       -0.022
                                             (0.030)      (0.025)      (0.027)         (0.018)       (0.020)
 βD : Doctor Voucher                         0.26∗∗∗       0.016       0.085∗∗∗        -0.054∗∗     -0.050∗∗∗
                                             (0.024)      (0.025)      (0.025)         (0.021)       (0.019)
 P-values and theory-driven tests
  βP = βD                                    0.011∗∗       0.079∗      0.018∗∗           0.735        0.193
  Test for mechanism:                          GC           PD           PD               DD           DD
  Significant evidence of mechanism:            Yes          Yes          Yes              No           No

 Mean (Control)                               0.000        0.616        0.461            0.303        0.255
 N                                            2055         2053         2053             2053         2053
 Notes : Robust standard errors clustered at the clinic level in parentheses. All regressions control for clinic visit
date fixed effects. We use double selection lasso to choose additional controls. Eligible controls include clinic
dummies, symptom dummies, duration of illness (topcoded at the 99th percentile) and its square, patient age
and its square, a dummy for patients under 5, patient gender, a dummy to identify pregnant patients, a dummy
to identify whether the patient (versus a caregiver) answered the survey, the gender of the survey respondent,
an ethnicity (Bambara) dummy, a dummy for French speaking respondents, a dummy for literate respondents, a
dummy for respondents with a primary education or less, a dummy to identify patients in the home-based follow
up survey, and pairwise interactions between all previously-listed patient and respondent controls. Missing values
are recoded to the sample mean and separately dummied out. These missing dummies are also used to construct
pairwise interactions. GC, PD, and DD indicate tests of gatekeeping costs, patient-driven, and doctor-driven
demand respectively. *, **, and *** denote statistical significance at the 10, 5, and 1 percent levels respectively.




                                                         49
     Table 4: Impacts on Malaria Treatment Outcomes - Heterogeneity by Predicted Malaria Risk

                                                (1)        (2)        (3)                 (4)        (5)
                                                         Any Malaria Treatment         Severe Malaria Treatment

                                              Used       Prescribed Purchased          Prescribed Purchased
                                             Voucher
 δP H : Patient Voucher × High Risk          0.34∗∗∗       0.035       0.12∗∗∗          -0.055∗       -0.020
                                             (0.039)      (0.035)      (0.037)          (0.031)      (0.033)
 δDH : Doctor Voucher × High Risk            0.32∗∗∗       0.021       0.078∗∗          -0.12∗∗∗     -0.12∗∗∗
                                             (0.038)      (0.034)      (0.036)          (0.036)      (0.033)
 δP L : Patient Voucher × Low Risk           0.36∗∗∗      0.092∗∗      0.18∗∗∗           -0.026       -0.016
                                             (0.041)      (0.036)      (0.041)          (0.027)      (0.029)
 δDL : Doctor Voucher × Low Risk             0.21∗∗∗       0.011       0.086∗∗          -0.0019       0.0085
                                             (0.033)      (0.035)      (0.034)          (0.030)      (0.026)
 θ: High Malaria Risk                        0.0029       0.13∗∗∗      0.13∗∗∗           0.090∗      0.093∗∗
                                             (0.041)      (0.046)      (0.048)          (0.048)      (0.043)
 P-values and theory-driven tests
  δP H = δDH                                  0.749        0.689        0.275            0.101      0.007∗∗∗
  Test for mechanism:                        GC/DD           –            –               DD          DD
  Significant evidence of mechanism:            No            –            –               No           No

  δP L = δDL                                 0.000∗∗∗     0.030∗∗      0.014∗∗           0.481        0.435
  Test for mechanism:                        GC/PD          PD           PD                –            –
  Significant evidence of mechanism:            Yes          Yes          Yes               –            –

 Mean (Control, Low Risk)                     0.000        0.486        0.329            0.154        0.116
 N                                            2055         2053         2053             2053         2053
 Notes : Robust standard errors clustered at the clinic level in parentheses. All regressions control for clinic visit
date fixed effects. We use double selection lasso to choose additional controls. See notes to Table 3 for a list
of potential controls. Standard errors are based on 1,000 bootstrap replications, with re-sampling at the clinic
level. Predicted malaria risk is re-calculated on each bootstrap replication. GC, PD, and DD indicate tests
of gatekeeping costs, patient-driven, and doctor-driven demand respectively. *, **, and *** denote statistical
significance at the 10, 5, and 1 percent levels respectively.




                                                         50
                            Table 5: Impacts on Match Between Treatment and Illness

                                   (1)        (2)          (3)                 (4)       (5)          (6)
                                   Expected Match: Prescribed                  Expected Match: Purchased

                                 Malaria       Malaria        Overall       Malaria       Malaria        Overall
                                 Positive      Negative       Match         Positive      Negative       Match
      βP : Patient Voucher       0.0096        -0.043∗∗       -0.032∗       0.023∗∗∗       -0.11∗∗∗      -0.083∗∗∗
                                (0.0070)       (0.021)        (0.017)       (0.0079)       (0.022)        (0.017)
      βD : Doctor Voucher        0.0073        -0.0074        0.0049        0.018∗∗∗      -0.063∗∗∗       -0.039∗∗
                                (0.0076)       (0.021)        (0.017)       (0.0071)       (0.020)        (0.016)
      P-values
       βP = βD                    0.718         0.034∗∗       0.006∗∗∗       0.537        0.008∗∗∗       0.002∗∗∗
51




       βP = βD = 0                0.384         0.050∗        0.015∗∗       0.007∗∗∗      0.000∗∗∗       0.000∗∗∗

      Mean (Control)              0.153          0.326         0.479         0.122          0.449         0.570
      N                           2053           2053          2053          2053           2053          2053
      Notes : Robust standard errors clustered at the clinic level in parentheses. All regressions control for clinic
     visit date fixed effects. We use double selection lasso to choose additional controls. See notes to Table 3
     for a list of potential controls. The expected match for malaria positive is equal to predicted malaria risk
     times the relevant malaria treatment/purchase dummy. The expected match for malaria negative is equal
     to one minus predicted malaria risk times one minus the malaria prescription/purchase dummy. The overall
     expected match is the sum of these two variables. *, **, and *** denote statistical significance at the 10, 5,
     and 1 percent levels respectively.