Banerjee Et Al 2021 The Challenges of Universal Health Insurance in Developing Countries Experimental Evidence From

American Economic Review 2021, 111(9): 3035–3063
https://doi.org/10.1257/aer.20200523
The Challenges of Universal Health Insurance in

Developing Countries: Experimental Evidence from
Indonesia’s National Health Insurance†
By Abhijit Banerjee, Amy Finkelstein, Rema Hanna, Benjamin A. Olken,
Arianna Ornaghi, and Sudarno Sumarto*
To investigate barriers to universal health insurance in developing

countries, we designed a randomized experiment involving about
6,000 households in Indonesia who are subject to a government health
insurance program with a weakly enforced mandate. Time-limited
subsidies increased enrollment and attracted lower-cost enrollees,
in part by reducing the strategic timing of enrollment to correspond
with health needs. Registration assistance also increased enrollment,
but increased attempted enrollment much more, as over one-half of
households who attempted to enroll did not successfully do so. These
findings underscore how weak administrative capacity can create
important challenges in developing countries for achieving wide-
spread coverage. (JEL D82, G22, H51, I13, I18, O15)
As developing countries emerge from extreme poverty and enter m iddle-income

status, many aim to expand government-run social safety nets (Chetty and Looney
2006). An important part of this process is the creation of universal health insurance
programs, which have expanded to many lower- and m iddle-income countries over
recent decades (Lagomarsino et al. 2012). In expanding health insurance, however,
emerging countries may face particularly vexing versions of the challenges faced by
developed countries, because of the large informal sector operating o utside the tax
* Banerjee: MIT (email: [email protected]); Finkelstein: MIT (email: [email protected]); Hanna: Harvard
University (email: [email protected]); Olken: MIT (email: [email protected]); Ornaghi: University
of Warwick (email: [email protected]); Sumarto: TNP2K and SMERU (email: [email protected]).
Henrik Kleven was the coeditor for this article. We thank our partners at BPJS Kesehatan, Bappenas, TNP2K,
and KSP for their support and assistance. In particular, we wish to thank from BPJS Kesehatan Fachmi Idris,
Mundiharno, Tono Rustiano, Andi Afdal, Dwi Martiningsih, Citra Jaya, Togar Siallagan, Tati Haryati Denawati,
Atmiroseva, Muh. Syahrul, Golda Kurniawati, Jaffarus Sodiq, Norrista Ulil, and the many staff at regional BPJS
offices who provided assistance; Maliki and Vivi Yulaswati from Bappenas, Jurist Tan from KSP, and Bambang
Widianto and Prastuti (Becky) Soewondo from TNP2K. We thank the outstanding JPAL SEA team members
for their work on this study, in particular Ignasius Hasim, Masyhur Hilmy, Amri Ilmma, Ivan Mahardika, Lina
Marliani, Patrya Pratama, Hector Salazar Salame, Reksa Samudra, Nurul Wakhidah, and Poppy Widyasari. Yuanita
Christayanie provided excellent research assistance. We thank SurveyMeter for outstanding data collection and
fieldwork, especially Bondan Sikoki and Nasirudin. Funding from the Australian Department of Foreign Affairs and
Trade and KOICA is gratefully acknowledged. This project was approved by the MIT IRB (Protocol 1406006432).
This randomized control trial was registered in the American Economic Association Registry for randomized con-
trol trials under trial AEARCTR-0000815. The views expressed here are those of the authors, and do not necessarily
reflect those of the many individuals or organizations acknowledged here.
†
Go to https://doi.org/10.1257/aer.20200523 to visit the article page for additional materials and author
disclosure statements.
3035
3036 THE AMERICAN ECONOMIC REVIEW SEPTEMBER 2021
net (Jensen 2019). Some countries, such as Thailand, have created a single-payer-like
system funded entirely out of tax revenues and supplemented by small copays; this
has been shown to improve health-care utilization and outcomes, but faces funding
challenges (Gruber, Hendren, and Townsend 2014). Other countries, such as Ghana,
Kenya, the Philippines, and Vietnam—as well as Indonesia, which is the focus of
our study—have sought to create a contributory system with an individual mandate
to reduce the financial burden on the government. In these systems, the very poor
are subsidized by tax revenues, but everyone else is required to pay a premium, col-
lected through a payroll tax for formal sector workers and collected directly from
individuals for everyone else.
The challenge with contributory systems is that enforcing the insurance mandate
for those who directly pay premium is difficult. While the political and administra-
tive challenges of enforcing mandates are not unique to developing countries—for
example, the 2010 Obamacare mandate did not achieve universal coverage in the
United States (Berchick, Hood, and Barnett 2018) and the US tax penalties for fail-
ing to comply with the mandate were eliminated starting in 2019—they are partic-
ularly hard for developing countries, again because the majority of their citizens
are outside the tax net. This means that the types of penalties for noncompliance
used initially under Obamacare—fines collected through the personal income tax
system—are not even an option. Since developing countries have, perhaps rightly,
shown little appetite for enforcing the few possible remaining sanctions on this pop-
ulation (e.g., denying delinquent households the ability to enroll their children in
school), what they are left with is a toothless mandate.
A toothless mandate can create two related challenges for governments that are
trying to increase coverage: low program enrollment and adverse selection, where
the least healthy are most likely to enroll, raising program costs per enrollee above
the population average (Akerlof 1970, Einav and Finkelstein 2011). Indonesia, like
other countries, has experienced both problems: although mandatory, universal
health insurance was launched in 2014, one year after its introduction, the con-
tributory portion of the program, known as JKN \ Mandiri (Jaminan Kesehatan
Nasional Mandiri), had enrolled less than 20 percent of the targeted population;
moreover, because enrollees were much less healthy than the typical targeted popu-
lation, claims exceeded premiums by a ratio of 6.45 to 1.1
These facts motivate the question of whether and how developing country gov-
ernments can design supplemental policies to mitigate these challenges—to boost
national health insurance enrollment, while also reining in the financial costs to the
tax-funded government budget—in contexts with mandatory, but weakly enforced,
contributory health insurance programs. The aim is not necessarily for the govern-
ment to break even—some subsidies may be needed in order to make sure that there
is enough social protection against health shocks—but to limit government spend-
ing while insuring as many people as possible. With this perspective in mind, in
2015 and in cooperation with the Indonesian government, we designed a l arge-scale,
multiarm experiment—involving almost 6,000 households.
1
Enrollment rates are from authors’ calculation from the nationally representative 2015 Indonesian National
Household Survey, SUSENAS (BPS 2015). Claims-to-premium ratios are from LPEM-UI (2015).
VOL. 111 NO. 9 BANERJEE ET AL.: CHALLENGES OF UNIVERSAL HEALTH INSURANCE 3037
We designed three interventions that simple economic theory suggested could

increase enrollment and reduce adverse selection in this nationally mandated insur-
ance program. First, we examined the role of large, temporary subsidies. We ran-
domized households to receive subsidies of either 50 percent (“half subsidy”) or
100 percent (“full subsidy”) for the first year of enrollment. To be eligible for the
subsidy, households had to enroll within two weeks after they were offered it, akin to
governments offering a large, t ime-limited registration incentive. Second, we exam-
ined the role of transaction costs by randomly offering some households at-home
assistance with the online registration system, so that they did not have to travel to
a far-off insurance office to enroll. Third, we examined information constraints by
randomly advertising three different types of basic insurance information: the finan-
cial costs of a health episode and how they relate to insurance prices, the presence
of a two-week waiting period from enrollment to coverage (so that one could not
wait to get sick and immediately sign up), and the fact that insurance coverage is
legally mandatory.
To assess the impacts of these interventions, we primarily analyze government
administrative data. These data contain detailed information on all program enroll-
ees from the study group for 20 months after the intervention, including monthly
data on registration, premiums paid, and the amount and nature of any insurance
claims made. These extensive, high-frequency administrative data allow us to exam-
ine the dynamic responses to the interventions both during and after the subsidy
period. We supplement these data with a short baseline assessment survey in which
we collected data on demographics and s elf-reported health; this allows us to mea-
sure preintervention “health status” for all study participants, regardless of whether
they subsequently enrolled in the insurance program.
We use these data to study several key outcomes. First, we examine both
attempted enrollment, which we define as the household starting the initial regis-
tration process, and actual enrollment (hereafter “enrollment”), which we define
as successfully completing the initial registration process. Second, since the deci-
sion to stay enrolled is a dynamic one in which households need to pay a monthly
premium, we also examine the impact of the intervention on the time path of
insurance coverage, which we define as having paid the premium for a given
month. Third, we study the patterns of insurance claims among those who have
insurance coverage.
We have three main findings. First, there are both monetary and nonmonetary bar-
riers to enrollment. Both the time-limited, temporary subsidies and the registration
assistance increased enrollment. The full subsidy increased enrollment by 18.6 per-
centage points off of a control group mean of 7.8 percent. Moreover, even after the
subsidy ended, insurance coverage in the f ull-subsidy group remained about twice
as high as in the no-subsidy group, consistent with the idea of health insurance as
an “experience good” (Cai, de Janvry, and Sadoulet 2020; Dupas 2014; Delavallade
2017). Reducing hassles by assisting with i nternet-based registration also increased
enrollment, by a statistically significant 3.5 percentage points (45 percent), but none
of the information treatments affected enrollment. This suggests that lack of infor-
mation may not be a key barrier and, relatedly, while information and nudge cam-
paigns are often an attractive policy option given their low cost (Thaler and Sunstein
2009), they may be of limited effectiveness in this context.
Second, weak administrative state capacity is a central impediment to achieving

universal coverage, as capacity constraints generate additional hassles to enroll at a
government office. These hassles appear to be a substantial barrier: nearly as many
people attempted to enroll in the assisted internet-based registration with no subsidy
as did in the full-subsidy group with status quo office-based registration. This sug-
gests that the hassle costs involved with in-person enrollment are in some sense as
costly as a full year’s worth of premiums.
But an even more important finding is that many more people attempted to enroll
than were actually able to do so. For example, when offered both a full subsidy
and assisted registration, about 55 percent of households tried to enroll, but less
than one-half of these households were successfully able to do so. The large wedge
between attempted and actual enrollment was due to technical and administrative
challenges with the government’s online enrollment system. While also reminiscent
of the issues with HealthCare.gov in the United States, this particular challenge
stemmed from a problem common in many developing countries: Indonesia’s under-
lying state civil registry, i.e., the data on who is in each family, is often inaccurate
(Sumner and Kusumaningrum 2014). Because all family members must be enrolled
at once (to help mitigate adverse selection), inaccuracies in family definitions in
the civil registry meant that people would have to visit an office to fix these errors
before they could return to the website and sign up correctly. Since imperfect civil
registries are common throughout the developing world (see, e.g., Mikkelsen et al.
2015; Muralidharan, Niehaus, and Sukhtankar 2020), these types of challenges are
likely to be encountered in other contexts as well. Households could overcome these
issues by showing up in person to a district office and having an official override the
system in some way, but this was a significant additional hassle cost that appears
to have discouraged many households. In short, administrative capacity constraints
generate hassle costs that suppress demand.
Third, we find evidence of strategic timing in health insurance coverage that
also in part reflects the consequences of Indonesia’s administrative structure. In
Indonesia, individuals can enroll in health insurance at any point during the year.
This creates incentives to delay coverage until one gets sick and—despite penal-
ties if someone drops coverage and then tries to reactivate it—to drop coverage
once one recovers. Such strategic “wait till you need it” enrollment timing became
evident when we compared outcomes for the no-subsidy group to outcomes for
the group to whom we offered time-limited subsidies. Relative to enrollees in the
no-subsidy group, those enrolling in insurance in the full-subsidy treatment reported
better health at baseline and had fewer claims (and notably, fewer claims for chronic
conditions) during their first year of enrollment than typical new enrollees in the
no-subsidy group. These cost differences in part reflect strategic timing decisions by
the no-subsidy group, rather than just fixed health differences alone. The no-subsidy
enrollees submitted more claims than did full-subsidy enrollees in the first three
months after enrollment, and many enrollees in the n o-subsidy group subsequently
dropped coverage—i.e., stopped paying premiums—after a few months.
The fact that the subsidies brought in healthier people who were less likely to drop
coverage soon after an initial claim has potential implications for the longer-term
ability of the program to increase coverage. On net, once the subsidies had ended,
the government was able to cover substantially more people at no higher total cost, at
least up to the first 20 months since offer date, because the time-limited subsidies
brought in healthier enrollees with fewer claims.
A natural policy tool to limit such strategic timing is to allow individuals a limited
window each year to enroll. Such “open enrollment” periods are the norm in the
United States, but are absent in Indonesia and in several other developing countries
(e.g., Ghana, Kenya, the Philippines, and Vietnam) that also have a public health
insurance system in which the n onpoor informal sector may pay premiums to enroll.
One potential reason developing countries may not use limited enrollment windows
is that lumpy incomes and credit constraints may hinder households from making
the first payment if the timing of that payment is not aligned with the timing of their
income. In fact, take-up of various products in developing countries (e.g., fertiliz-
ers, scholarships) has been shown to be dramatically higher if households are given
the opportunity to make the first payment right after they receive income (e.g., see,
for example Duflo, Kremer, and Robinson 2011, for an example from agriculture).
Since the precise timing of when households receive incomes differs across the pop-
ulation, a fixed, limited enrollment window might substantially discourage enroll-
ment for many households. In fact, in Indonesia, the social agency’s understanding
of Indonesian law was that they were legally mandated to allow enrollment through-
out the year in order to allow for more flexible access. The inability to offer limited
enrollment windows—perhaps for this reason, or because of the practical challenges
associated with doing so—reveals another important challenge for developing coun-
tries in trying to administer a universal contributory health insurance program.
Our study builds on the large literature on participation in public health insurance
systems—and in social insurance programs.2 Existing evidence from both devel-
oped and developing countries points to increased participation in social insurance
programs, including health insurance, from monetary subsidies (e.g., Thornton et al.
2010; Asuming 2013; Fischer, Frölich, and Landmann 2018; Finkelstein, Hendren,
and Shepard 2019), reductions in transaction costs (e.g., Alatas et al. 2016, Bettinger
et al. 2012, Dupas et al. 2016), and information (e.g., Gupta 2017, Bhargava and
Manoli 2015). We examine the impact of all three of these commonly conjectured
participation barriers in the context of a l arge-scale government insurance program.
Our rich administrative data allow us to probe more deeply than is typically fea-
sible in developing countries to identify key challenges to achieving universal cov-
erage in the context of weak state capacity. For example, our ability to measure both
enrollment attempts and actual enrollment successes highlights the key obstacles
that imperfect civil registries can impose. In the same vein, our high-frequency data
on the dynamics of premium payments and claims allow us to identify the strate-
gic timing of insurance coverage that can occur when the state is unable to limit
enrollment to a short period of time each year;3 Diamond et al. (2019) explores
similar phenomena in the individual health insurance market in California subject to
2
It is particularly closely related to Thornton et al. (2010), which examines the impact on enrollment and health-
care utilization of randomly offered subsidies for contributory health insurance for informal workers in Nicaragua.
It finds impacts of subsidies on enrollment and health-care utilization.
3
Our use of administrative data allows us to go further than Asuming, Kim, and Sim (2019), which uses survey
data to assess the impact of o ne-year subsidies on enrollment and health behaviors during and after the subsidy in
Ghana. Our h igh-frequency administrative data allow us to unpack the dynamics of selection and show how differ-
ential retention affects our understanding of these health insurance markets.
the rules of the US Affordable Care Act. Cabral (2017) demonstrates how adverse
selection can be generated not only through the strategic timing of coverage (as we
and Diamond et al. 2019 focus on), but also through the strategic timing of treatment
decisions.
The remainder of the paper is organized as follows. Section I describes the setting.
Section II presents the experimental design and data. Section III presents the impacts
of the intervention on enrollment, coverage patterns over time, and attempted enroll-
ment. Section IV illustrates the strategic coverage timing under the status quo and
how time-limited subsidies reduce such behavior. The last section concludes.
I. Setting
A. The Indonesian Health-Care System
The Indonesian health-care system consists of a mix of public and private provid-
ers. The public sector provides all levels of care, including large tertiary care hos-
pitals in major cities, smaller secondary care hospitals in virtually every district of
Indonesia, and a vast network of clinics at the subdistrict and village levels. Private
hospitals and private clinics operate alongside these public providers, particularly
in urban areas. Publicly employed physicians and midwives are also allowed to
operate private practices after hours, and many do so, even outside of urban areas.
Nationwide, about 55 percent of inpatient visits are in public hospitals and clinics,
while 45 percent are in private hospitals and clinics; about 40 percent of outpa-
tient visits take place in public clinics, and 60 percent in private (online Appendix
Table 1).4
At the time of our intervention, the main provider of health insurance was the
government, through several different programs (JKN Mandiri, which we study
here, the branches of JKN that provide insurance for formal sector workers and
government employees, and the g overnment-run schemes [JKN and otherwise] that
provide free coverage for the poor from national or district governments); only
about 1 percent of the population had private insurance. In 2015, about one-half of
the Indonesian population lacked formal health insurance.5 Those without insur-
ance faced substantial risk of o ut-of-pocket health-care expenditures. For uninsured,
nonpoor informal sector households—the sample we will focus the study on—we
estimated average annual out-of-pocket health expenditures of Rp 894,018, or about
2.3 percent of total nonhealth household consumption. But this average masks con-
siderable variance; for example, while the median household spent only 0.4 percent
of total nonhealth consumption on health expenditures, the ninety-fifth percen-
tile household spent 9.1 percent, and the ninety-ninth percentile household spent
36.1 percent. By way of comparison, King et al.’s (2009) study of the uninsured in
Mexico defined “catastrophic” health expenditures as those that exceed 30 percent
4
All statistics in this section are authors’ calculation from the nationally representative 2015 Indonesian
National Household Survey, SUSENAS (BPS 2015) unless otherwise noted.
5
Since the time period that we started our study, there has been an increase in coverage, and as of 2020,
approximately 20 percent are still uncovered (https://bpjs-kesehatan.go.id/bpjs/). For the population that we focus
on—the informal, n onpoor who do not qualify for free insurance paid for by the government—over 60 percent
remain uninsured.
of annual household spending, after subtracting a subsistence food allocation. This

suggests that there is substantial risk to be insured.
B. Mandatory, Contributory Public Health Insurance: The JKN Mandiri Program
In January 2014, the Government of Indonesia launched Jaminan Kesehatan

Nasional (JKN), a national, mandatory, public health insurance program designed
to provide universal coverage by 2019. It consists of three different s ubprograms
based on income and employment status. The program we study—JKN Mandiri—
covers nonpoor informal workers and their households, which represent 30 percent
of the population, and is supposed to be funded by their premium contributions.6
Public insurance covers essentially all health-care costs incurred at all public
and affiliated private clinics and hospitals with no copays, although certain spe-
cific procedures (e.g., cosmetic surgery, infertility treatments, orthodontics, etc.) are
excluded. Primary care clinics are reimbursed under capitation based on the total
number of practitioners, the ratio of practitioners to beneficiaries, and operating
hours. Hospitals are reimbursed by case following a tariff system called I NA-CBG
(Indonesia Case Base Groups), in which amounts are determined jointly by primary
diagnosis and severity of the case. Having JKN insurance changes who pays for
care, but does not change access to care, other than requiring referrals from primary
care before accessing secondary or tertiary care services.
To receive coverage under JKN Mandiri, households must complete an initial
registration process and then subsequently pay their monthly premiums. To enroll,
households must register either in person at the Badan Penyelenggara Jaminan
Sosial Kesehatan (Social Security Administration for Health, or BPJS) office or
through the social security administration website; the latter option, however, is
often unavailable to households since they may lack internet access or find the
online process confusing. They must also register the entire family, as listed on the
household’s Kartu Keluarga (family card), which is maintained in the civil regis-
try by another ministry (Ministry of Home Affairs); registering a single individual
within the family is not allowed. Registration requires each individual’s national
ID number (Nomor Induk Kependudukan, or NIK) and family card number.
Those registering in person in a BPJS office must bring a photocopy of their
national identity card and government-issued family card, as well as a 3 centimeter
(cm) by 4 cm photo of each individual being registered. Meanwhile, those register-
ing online should input their family card number for the system to automatically
retrieve their national identity numbers and addresses from the national civil reg-
istry system. Households registering online must also provide an active telephone
number and must have a photo that can be uploaded or scanned to the system for
each household member. Registrants select a primary care location at the time of
registration.
To maintain coverage, the household must then pay premiums each month. The
premium can be paid at any social security administration office, ATM, or select
6
About 40 percent of households that are classified as poor and near poor receive insurance funded completely
out of general government revenues. The program for formal workers is funded jointly by employee and employer
contributions that are withheld by the tax system.
convenience stores, or mobile banking. Paying the premium by the tenth of a given
month ensures coverage for that calendar month. If no payment is made, coverage is
deactivated after a o ne-month grace period.
The per-person monthly premium for basic coverage (known as Class III) is
Rp 25,500 (~$2).7 This means that premiums for a family of four (Rp 100,000) are
roughly similar to the average household out-of-pocket health expenditures which,
we noted above, are Rp 894,018 per year, or about Rp 75,000 per month. However,
as we will see below, due to some combination of adverse selection and moral haz-
ard, the claims for those actually enrolled are substantially higher than premiums
paid. This high level of claims relative to out-of-pocket expenditures among the
uninsured is consistent with the insurance product being potentially valuable and
useful.
Institutional Features to Reduce Selection.—While insurance enrollment is

legally mandatory, the mandate is hard to enforce in practice. There are no penalties
assessed on households that do not enroll. However, as one method to combat poten-
tial adverse selection, the government requires households who enroll to register all
nuclear family members (e.g., father, mother, and children), as listed on their official
Kartu Keluarga (family card). In addition, if a household stops paying premiums
and drops coverage, there are financial penalties if they subsequently try to re-enroll.
The amount of these penalties are increasing with the amount of initial claims they
incur after re-enrolling.8
Despite these features, a key opening for selection is that households may register
for JKN Mandiri at any time of the year. After the program’s introduction in 2014,
the government became concerned that this might lead individuals to wait to enroll
in JKN until they had a health emergency. However, existing constitutional rulings
precluded it from limiting enrollment to a fixed window of time within the year, as
all social benefit programs must be open to enrollment by all citizens at all times.
Lacking the ability to limit enrollment to a short period of time within the year, in
September 2015 the government instead introduced a two-week waiting period after
enrollment before households can submit an insurance claim.
II. Experimental Design and Data
A. Study Setting and Sample
We carried out this project in two large Indonesian cities: Kota Medan in North
Sumatra and Kota Bandung in West Java. These are the fourth and third largest
7
There are three different classes of coverage that cover the same medical procedures, but offer different types
of accommodations should an inpatient procedure be required. The per-person monthly premium during the period
of the study was Rp 42,500 (~$3) for class II (3–5 beds per room) and Rp 59,500 (~$4.5) for class I (2–3 beds per
room). Class III (more than 5 beds) is the most common insurance among our population of interest, with 85 per-
cent of households in the control group enrolling in Class III insurance.
8
Specifically, for coverage to reactivate at a later date, the household must pay their premiums in arrears, up
to a maximum of six months. If no inpatient claims are submitted within 45 days from reactivation, there are no
additional fees beyond these p remiums in arrears. Otherwise, the household has to pay an additional penalty equal
to 2.5 percent of the inpatient claims times the number of inactive months, up to a maximum of 12 months or
Rp 30 million.
urban areas in Indonesia, respectively, each with a population of approximately

2.2–2.5 million people. We focused on an urban setting to abstract from s upply-side
issues that are likely to depress demand in rural areas. We chose Medan and Bandung
because a significant fraction of their population was uninsured. Moreover, selecting
cities both on and off Java helps ensure representativeness of Indonesia’s heteroge-
neity in culture and institutions (Ravallion and Dearden 1988).
Working with the government, we implemented the interventions in two
subdistricts in Medan and eight s ubdistricts in Bandung. Using the 2010 census, we
selected s ubdistricts from among those with the highest concentration of nonpoor
informal workers; within those subdistricts, we randomly selected neighborhoods
for the study.9 To identify JKN-eligible households within the sampled areas, we
targeted uninsured, nonpoor informal workers by administering a rapid eligibility
survey to all listed households. We excluded households that already had at least
one member covered by health insurance and those that were officially below the
poverty line (and thus qualified for free insurance). Of the 52,519 listed households,
7,629 (14.5 percent) satisfied the target population criteria.
When we matched our baseline survey data with the government’s administrative
data, we discovered that some households were already covered by health insurance,
even if they reported that they were not. This was mostly an issue for the city of
Medan, where the local government had recently expanded the set of poor house-
holds who qualified for free insurance, but had not yet communicated this to the
newly insured. Since households with at least one insured member were not eligible
for the study (and this was predetermined), we excluded those already enrolled. Our
final sample was 5,996 households, about t hree-quarters of whom were in Bandung.
We implemented the intervention in Medan in February 2015 and in Bandung
in November and December 2015. Because of the introduction of the two-week
waiting period from enrollment to coverage in September 2015, the households in
the Bandung sample were subject to the waiting period, whereas households in the
Medan sample were not. Otherwise, the health insurance program was identical in
the two cities.
B. Experimental Design
Upon identifying an eligible household, we administered a short baseline survey,

at the end of which the household was randomly assigned to three fully crossed
treatment arms affecting the insurance price, the hassle cost of registration, and the
information available. Figure 1 summarizes the experimental design for the cities of
Medan and Bandung separately.10
9
We excluded s ubdistricts with universities, large factories, or malls to avoid areas with a high concentration
of temporary residents. We then randomly selected 12 kelurahan (urban municipal units) in the 2 s ubdistricts in
Medan (out of 16 possible kelurahan) and 4 kelurahan in each subdistrict in Bandung (out of 41 possible kelu-
rahan). Within each kelurahan, we randomly selected the neighborhoods (rukun warga, also known as RW) to
enumerate.
10
The number of households differs in each treatment for two reasons. First, while in Medan we maximized
power to detect differences in enrollment, in Bandung we maximized power to detect differences in claims condi-
tional on take-up. Since we expected greater take-up with a larger subsidy, we randomized more households into
groups with smaller subsidy amounts. Second, a coding error meant that while the overall treatment probabilities
were as assigned, some combinations of treatments were more likely to be randomly assigned to households than
Status quo Assisted registration

Standard Extra Standard Extra Subsidy
information information information information treatment totals
Panel A. Medan
No subsidy 37 63 134 237 471
Half subsidy 171 215 26 66 478
Full subsidy 176 54 170 97 497
Registration treatment totals 716 730 1,446
Status quo Assisted registration

Standard Mandate Standard Mandate Subsidy
information information information information treatment totals
Panel B. Bandung
No subsidy
Standard information 236 307 241 274
Waiting period information 232 300 297 349 2,236
Half subsidy
Waiting period information 141 114 100 91 918
Full subsidy
Bonus subsidy
Registration treatment totals 2,269 2,281 4,550
Figure 1. Experimental Design
Notes: This figure shows the randomization into each treatment arm, by city. Each cell reports the number of house-
holds allocated to the specific treatment cell. For more information on the bonus subsidy treatment, see online
Appendix A.
Time-Limited, Temporary Subsidies.—Households were randomly selected to be

in one of three groups: a control group, a full-subsidy group covering the premiums
for all family members for one year, and a half-subsidy group covering one-half
of a family’s premiums for a year. Importantly, the subsidy offer was explicitly
time-limited: it was only available for up to two weeks after the offer was made.
Thus while the state does not have the capacity to enforce a limited annual window
for enrollment, our time limited subsidy was designed to approximate the idea of
there being a limited time period under which enrollment conditions are more favor-
able.11 For participants whom we enrolled within two weeks of the offer, subsidies
were administered for 12 months.
For logistical reasons, we could not pay half of each person’s premium. Instead,
we implemented the half subsidy through a “buy-one-get-one-free” scheme in which
others (this coding error was corrected partway through the Bandung experiment). We include in the analysis a
dummy for whether the old or new randomization was used, and reweight observations to obtain the intended
cross-randomization weights so that each main treatment group has the same mix of each crossed additional
treatment.
11
In each city, we additionally randomized a separate subtreatment within the subsidy design. The households
in these s ubtreatments are included in all our main analysis, but they are described in more detail and results pre-
sented in online Appendix A and online Appendix Table 3.
we paid the full premiums for one-half the family members for one year, and the
household was then required to pay for the other half.12 Households chose which
family members were subsidized. In theory, the government regulated that all imme-
diate household members be registered, so subsidizing half of the household mem-
bers was roughly equivalent to providing a 50 percent discount. The subsidy receipt
for the subsidized members was conditional on payment for the n onsubsidized
members for the first month, but unconditional thereafter in practice. Households in
the f ull-subsidy group were not required to make any payments during the subsidy
period.
Assisted Registration.—Registering for JKN Mandiri usually requires travel-

ing to the social security administration office in the district capital. To reduce the
one-time hassle costs of registration, we randomly offered one-half of the study
households the possibility of completing the registration process online at home with
the assistance of the study enumerator. The enumerators had internet-enabled lap-
tops that they used to access the official social security website. They then assisted
the household with gathering the correct documentation, taking pictures, and filling
in the forms on the website. Upon successful registration, the enumerators provided
information to the household on payment procedures. If the household wanted to
think more about their options, needed time to assemble the documentation, or had
technical registration problems, the enumerators returned within a few days to try to
assist with the enrollment process again.
Information.—All study households received basic information, such as what

the insurance covered, the premiums, and the procedure for registration. For ran-
domly selected households in each city, we provided additional types of general,
one-time information to test whether various forms of lack of knowledge con-
strained enrollment.
In Medan, we randomly assigned a group of households to receive additional
information on the financial costs of a health episode (“extra information treat-
ment”). Using a script and an accompanying booklet, we detailed the average
out-of-pocket expenditures for Indonesia’s most common chronic health conditions,
as well as the cost of having a heart attack.
In Bandung, all households received basic insurance information, as well as a dis-
cussion of the out-of-pocket expenditures associated with accessing care. However,
based on discussions with the government, we then randomly assigned households
to the following two treatments: (i) a “waiting period” treatment, in which we
informed households about the new two-week waiting period between enrollment
and the start of coverage; and (ii) a “mandate information” treatment, in which we
reminded households that enrollment is mandatory, and that there was a possibility
that the government would soon introduce regulations requiring proof of insurance
to be able to renew government documents, such as a passport or driver’s license.
12
If a family had an odd number of members, we randomly assigned the household to receive a subsidy for
(y + 1)/ 2or (y − 1)/ 2members with equal probability. If there was only one member, the member received a full
subsidy.
2015 2016 2017 2018

2 3 4 5 6 7 8 9 10 11 12 1 2 3 4 5 6 7 8 9 10 11 12 1 2 3 4 5 6 7 8 9 10 11 12 1 2 3 4 5 6 7 8
Medan
Baseline and treatment
Subsidy period
Administrative data
Bandung
Baseline and treatment
Subsidy period
Administrative data
Figure 2. Timeline
Notes: This figure provides a timeline by city. The baseline and treatment offer occurred in the same household
visit, in February 2015 in Medan and November/December 2015 in Bandung. Subsidies were disbursed over the
course of 12 months following the offer. We accessed the administrative data in April 2015, April 2016, May 2017,
and August 2018.
C. Data
We compiled two new datasets for this project: a baseline survey and government
administrative data. Figure 2 shows the time period for our data relative to the exper-
imental interventions, separately for each city.
We conducted a short baseline survey (Banerjee et al. 2021) in conjunction with
an independent and established survey firm (SurveyMeter). We administered the
baseline survey immediately following the listing questionnaire to determine eligi-
bility for the study (e.g., informal worker without health insurance). The baseline
survey collected information on the demographic characteristics of family members,
self-reported health and previous health-care utilization, and existing knowledge of
the program.13 Self-reported health was measured on a four-point scale from one
(unhealthy) to four (very healthy); we analyze average s elf-reported health across
household members. The survey was identical in Bandung and Medan, with the one
exception being that we added questions on income and employment in Bandung.
All of our outcomes are measured using detailed, high-frequency, high-quality
government administrative data from February 2015 to August 2018 (BPJS 2021).
We track all of our participants for either 20 months since the date of offer or date
of enrollment, depending on the analysis. We matched the study participants to the
administrative data using individuals’ unique national identification number (Nomor
Induk Kependudukan or NIK).14
We define enrollment to be the household’s successful completion of the registra-
tion process for the national insurance program. Since a household may enroll but
not actually pay any premiums, we then define coverage in a given month to mean
13
To minimize priming, the questions related to knowledge of the program were asked after the information
on health status. The consent form only mentioned SurveyMeter and Indonesia’s National Development Planning
Agency (Bappenas), the other partner in the study, but not the social security administration or JKN.
14
To ensure that we identify the correct individuals, we exclude matches when the year of birth reported in
the baseline and that reported in the administrative database differ by more one year. When the same NIK links to
two different membership numbers, we consider both observations as a match. When two different NIKs link to
the same membership number, we exclude the observation. When enrollment date or membership type changes in
subsequent extracts, we retain the information as reported in the first extract in which the individual appears. About
23 percent of the individuals surveyed did not have a NIK at baseline and cannot be matched to the administrative
data. We show in column 1 of online Appendix Table 2 that the probability that a household reports the NIK of at
least one of its members is not differential across treatment. Given that a NIK is a requirement of enrollment, those
without a NIK are likely not enrolled in JKN, and we treat them as such.
that the enrolled household’s premiums were paid that month. We use the admin-
istrative data on registration date to measure enrollment. We use the administrative
premium payment data, which report the date and value of each payment, to mea-
sure coverage in each month.
Since a household may also attempt to enroll but not succeed, we also measure
attempted enrollment. By construction, this is measured differently depending on
the treatment arm. For households in the assisted registration arm, we define them
as attempting to enroll if the enumerator records that the household accepted the
enumerator’s offer of assistance with enrollment and that the enumerator began
the internet-based process of helping them. For households assigned to the status
quo registration procedures, if they are in the subsidy treatments, we define them
as attempting to enroll if they showed up to the social security office to enroll.
Households in the subsidy treatments had to contact the study assistant at the social
security office in order to redeem their voucher, enabling us to record their attempt.
For households in the n o-subsidy group, we set attempted enrollment equal to actual
enrollment, a choice justified by the fact that the failure rate of enrollment attempts
for households assigned to the status quo registration in the subsidy treatments was
negligible (as social security officials within the office could manually fix family
card issues within the system).
Finally, to measure insured health-care utilization, we analyze data on all claims
that are covered by the JKN insurance in both hospitals and clinics. The hospital
claim data report start and end date, diagnosis, reimbursement value, and facility
where the claim was made.15 The clinic claims data report similar information to the
hospital claims data, except that—due to capitation—claim values are not available.
We also use diagnoses to code whether each claim is for a chronic versus emergent
condition.16
D. Empirical Specification
We estimate the following equation for a variety of outcome variables yi:
yi = β0  + β1 HALF SUBSIDYi  + β2 FULL SUBSIDYi

(1)
+ β3 ASSISTANCEi  + INFO  ′i   β4  + X  ′i   δ + εi ,
ALF SUBSIDYi, FULL SUBSIDYi, and ASSISTANCEiare dummy variables

where H
equal to one if household iwas randomly assigned to the respective treatment, and
15
A claim corresponds to an outpatient or inpatient event. Each event is associated with a series of diagnoses.
The hospital is reimbursed for the amount that corresponds to the primary diagnosis according to the INA-CBG
tariff. All exams and treatment needed for an event get reimbursed under the same claim.
16
We build our chronic condition classification from the Chronic Condition Indicator for the International
Classification of Diseases from the Healthcare Cost and Utilization Project (Agency for Healthcare Research and
Quality 2018). This database provides information on whether diagnoses included in the ICD-10-CM: 2018 can
be classified as chronic conditions. We link conditions in the ICD-10-CM (International Classification of Diseases,
Tenth Revision, Clinical Modification): 2018 to conditions in the ICD-10: 2008 (WHO 2008)—the classification
system followed by BPJS—using the first three digits of the diagnosis code. This is the lowest classification that
straightforwardly corresponds across the two systems. We consider a diagnosis as chronic if it belongs to a 3 -digit
code group with more than 75 percent chronic diagnoses.
INFOiis a vector of dummies equal to one if household i was randomly assigned to a

particular information intervention. The term X iis a matrix of h ousehold-level con-
trols that includes dummy variables for the assignment to the other subtreatments
(see footnote 11), a dummy variable for the randomization procedure (see foot-
note 10), and a dummy variable for city of residence. Regressions are weighted to
reflect the desired c ross-treatment randomization design (see footnote 10). Given
the household-level randomization, we report robust standard errors. Our exhibits
often report results for s ubsets of treatments, but the full set of indicator variables
is always included.
To assess balance across treatment arms, we estimate equation (1) using vari-
ous household characteristics measured in the baseline survey. Online Appendix
Table 2 shows the results. Only 6 out of the 54 coefficients are significantly differ-
ent from 0 at the 10 percent level, in line with what we would expect by chance.
III. Impacts on Enrollment, Coverage, and Enrollment Attempts
A. Enrollment
Table 1 presents the impact of the interventions on enrollment—i.e., successfully

completing the registration process. Subsidies substantially increased the probabil-
ity of enrollment during the 12 months after the offer date (while the subsidies
were still active), while assisted registration had a positive but much smaller impact
(panel A column 1). Only about 7.8 percent of the n o-subsidy, s tatus-quo registra-
tion group enrolled within the 12-month period. Relative to this, offering the full
subsidy increased enrollment by 18.6 percentage points. Offering the half subsidy
increased enrollment by 10 percentage points. By contrast, the assisted registration
treatment only increased enrollment by 3.5 percentage points.
Because the subsidy offer and the offer of registration assistance were time lim-
ited, it is possible that they shifted forward in time an enrollment decision that would
have occurred anyway (so-called “harvesting”). This dynamic response seems par-
ticularly plausible given that both the offer of registration assistance and the sub-
sidy offers were t ime limited. To examine this, we separately analyze the impact on
enrollment within the first eight weeks of the offer, and after eight weeks but within
the first year (the subsidy offers were only valid for up to two weeks, but we look
before and after eight weeks to allow for some margin of error in terms of data lags).
The results indicate a small, but statistically significant harvesting effect (column 3),
which accounts for slightly more than 10 percent of the total additional enrollment
effect from the first eight weeks (column 2).
Panels B and C report the same analyses for the information treatments. We report
the results separately by city because we tested different information treatments
in different cities, providing detailed information on heart attack costs in Medan
(panel B) and about the nature of insurance (i.e., that enrollment is mandatory and
that households must enroll at least two weeks in advance of a health claim) in
Bandung (panel C). We find no statistically significant effect of any of these informa-
tion treatments. We can rule out effect sizes respectively bigger than 8.5 percentage
points (information on heart attack costs), 2.5 percentage points (information on
mandates), and 3.2 percentage points (information on waiting period).
Table 1—Effect of Temporary Subsidies and Assisted Registration on Year 1 Enrollment
Decomposition
Enrolled Enrolled within Enrolled after 8
within 1 year 8 weeks of offer weeks, but within
date 1 year of offer date
(1) (2) (3)
Panel A. Subsidy and assisted registration treatments
Full subsidy 0.186 0.209 −0.023
(0.020) (0.018) (0.010)
Half subsidy 0.100 0.114 −0.014
(0.014) (0.013) (0.008)
Assisted registration 0.035 0.043 −0.008
(0.011) (0.009) (0.006)
Observations 5,996 5,996 5,996

No subsidy, status quo registration mean 0.078 0.018 0.060
p-value of test of hypothesis

Half subsidy = full subsidy 0.000 0.000 0.441
Assisted registration = full subsidy 0.000 0.000 0.265
Panel B. Information treatments, Medan

Information on cost of treatment for heart attack 0.029 0.034 −0.005
(0.029) (0.025) (0.016)

No information mean 0.190 0.131 0.059
Panel C. Information treatments, Bandung

Information on possible mandate penalties 0.004 −0.001 0.004
(0.011) (0.009) (0.007)
Information on 2-week waiting period 0.011 0.009 0.002
(0.011) (0.009) (0.007)

No information mean 0.123 0.078 0.045
Notes: This table shows the effect of subsidies, assisted registration, and the information treatments on enrollment
in year 1 since offer. We regress each outcome on indicator variables for assignment to all treatment arms, an indi-
cator variable for the randomization procedure used, and an indicator variable for the study location (equation (1)).
The omitted category is no subsidy for the subsidy treatments, status quo registration for the assisted registration
treatment, and no information for all information treatments. The p-values reported in panel A are from a test of the
difference between the half-subsidy and full-subsidy treatments (β1  = β2) and assisted registration and full sub-
sidy treatments (β1 = β3 ). All regressions are estimated by ordinary least squares (OLS) and weighted to reflect the
intended cross randomization. Robust standard errors are reported in parentheses.
B. Subsequent Coverage
After the initial decision to enroll examined in Table 1, households must decide
whether to continue to pay their monthly premiums to remain covered at any given
point in time. Figure 3 shows household coverage patterns over time for differ-
ent subsidy groups, shown by month since offer (panel A) and, for those who ever
enroll, by month since enrollment (panel B).17 Coverage is defined as the premium
having been paid in full for all its members that month. Payment may be made
17
Online Appendix Table 4 reports the magnitudes and statistical significance of the patterns observed, while
online Appendix Table 5 provides the underlying regression estimates for the p-values reported in the table.
Panel A. By month since offer Panel B. By month since enrollment

Share of HHs with coverage

0.3 1
0.75
0.2
0.5
0.1
0.25
0 0
1 3 5 7 9 11 13 15 17 19 1 3 5 7 9 11 13 15 17 19
2 4 6 8 10 12 14 16 18 20 2 4 6 8 10 12 14 16 18 20
Months since offer Months since enrollment
No subsidy Half subsidy Full subsidy
Figure 3. Insurance Coverage by Subsidy Treatment
Notes: This figure shows mean insurance coverage for households (HHs) assigned to different subsidy treatments,
by month since offer (panel A) and by month since enrollment (panel B), with 95 percent confidence intervals for
the mean. Means are weighted to reflect the intended cross randomization. Coverage for a household is defined as
the premium having been paid in full for all its members that month. The sample size in panel A is 5,996 house-
holds. In panel B, the sample is restricted to households who enrolled within a year since offer and had coverage for
at least one month over the same time period. The sample size is 749 households.
either independently by the household or by the study. Thus all households in the
full-subsidy group who successfully enroll are covered for 12 months but not there-
after; households in the no-subsidy and h alf-subsidy groups need to remit payments
each month to remain covered.
In the no-subsidy group, coverage slowly increased over time from 0.61 per-
cent in the first month of the experiment to 6.66 percent almost two years later
(panel A). However, among those who enrolled, many in the n o-subsidy group
quickly dropped coverage; one-quarter of enrolled households in the no-subsidy
group had stopped paying their premiums three months p ost-enrollment, and nearly
one-half had stopped paying their premiums a year post-enrollment (panel B). The
steady increase in coverage for the no-subsidy group in panel A implies that the rate
of new enrollment was large enough so that net coverage rates continued to increase
despite the dropout effect.18
These patterns look very different for the full-subsidy group. About 25 percent
of those offered the full subsidy enrolled in the first two months after the offer,
statistically significantly higher than the no-subsidy group (panel A). Their cover-
age mechanically remained constant during the first year, when the subsidies were
active,19 and fell substantially once the subsidy ended; by month 20, coverage in
the full-subsidy group was about 40 percent of the first year level, with most of this
decline occurring in the first few months after the subsidy ended (panel A).
18
The steady increase in enrollment of the no-subsidy group throughout the study period is in line with the
number of enrollees going from approximately 10 million in January 2015 to more than 15 million in January 2016.
19
The slight increase in coverage shown in Figure 3 panel A for the full-subsidy group during months 4–12
comes from the fact that a small number of households in this group enrolled after the subsidy period was over.
Panel A. By month since offer Panel B. By month since enrollment

0.125 1
0.1
0.75
0.075
0.5
0.05
Status quo registration 0.25 Status quo registration

0.025
Assisted registration Assisted registration
0 0
1 2 3 4 5 6 7 8 9 10 11 121314 1516 17 18 1920 1 2 3 4 5 6 7 8 9 10 11 121314 1516 17 18 1920
Months since offer Months since enrollment
Figure 4. Insurance Coverage by Assisted Registration
Notes: This figure shows mean insurance coverage for households assigned to different registration treatments by
month since offer (panel A) and by month since enrollment (panel B), with 95 percent confidence intervals for the
mean. Means are weighted to reflect the intended cross randomization. Coverage for a household is defined as the
premium having been paid in full for all its members that month. The sample size in panel A is 5,996 households. In
panel B, the sample is restricted to households who enrolled within a year since offer and had coverage for at least
one month over the same time period. The sample size is 749 households.
Despite the large declines after the subsidy ended, coverage in the full-subsidy
group remained higher than in the no-subsidy group. For example, the full-subsidy
group was 4.6 percentage points (87 percent; p-value < 0.001) more likely than the
no-subsidy group to have coverage at month 15, and 3.9 percentage points (58 per-
cent; p-value = 0.001) more likely than the no-subsidy group to have coverage at
month 20 (see online Appendix Table 4). This persistence in elevated coverage even
after the subsidies expired cannot be explained by inertia or switching costs—which
have been w ell-documented in other health insurance contexts (e.g., Handel 2013,
Polyakova 2016); in our context, staying enrolled required an active choice to pay
premiums each month at an office, ATM, or convenience store. There is a financial
penalty for dropping and re-enrolling (see footnote 8 above), which may explain
part of this persistence. The persistence is, however, consistent with health insurance
as an “experience good,” for which initial exposure increases household demand.
For example, individuals may not have understood the benefits of insurance until
they experienced it; indeed, we show in Section IV below that retention is higher
among those households who had a claim.
As one might expect, results for the h alf-subsidy group are somewhere between
the no-subsidy and full-subsidy results. Their coverage rate in the first year was
higher than the no-subsidy group, but far below the full-subsidy group. They also
experienced a drop in coverage when the subsidy ended, and while their coverage
level was roughly flat in the second year, the n o-subsidy group slowly caught up to
them. By the 2 0-month mark, their coverage rates appear similar.
Figure 4 shows coverage patterns over time separately for the assisted-registration
group compared to the status quo group. Relative to the status quo group, the
assisted-registration group saw a slight increase in coverage initially, but coverage
rates soon converged. After subsidies ended, there is some evidence of a larger cov-
erage decline for those in the assisted-registration group, which may indicate that
Table 2—Effect of Temporary Subsidies and Assisted Registration on Attempted Enrollment
Attempted to enroll within Enrolled within 8 weeks

8 weeks of offer date of offer date
(1) (2)
Panel A. Main effect
Assisted registration 0.237 0.043
(0.011) (0.009)
Status quo registration mean 0.079 0.077
Panel B. Interacted specification

Full subsidy and assisted registration 0.538 0.241
(0.026) (0.024)
Full subsidy and status quo registration 0.171 0.201
(0.025) (0.025)
Half subsidy and assisted registration 0.400 0.164
(0.028) (0.023)
Half subsidy and status quo registration 0.082 0.088
(0.014) (0.013)
No subsidy and assisted registration 0.161 0.024
(0.011) (0.007)
No subsidy, status quo registration mean 0.018 0.018
Note: This table shows the effect of subsidies and assisted registration on attempted and actual enrollment within
eight weeks since offer. The sample size is 5,996 households. In panel A, we regress each outcome on indicator vari-
ables for assignment to all treatment arms, an indicator variable for the randomization procedure used, and an indi-
cator variable for the study location (equation (1)). The omitted category is status quo registration. Panel B reports
estimates from a variant of equation (1) that additionally includes the fully interacted effects of subsidies and reg-
istration assistance. The omitted category is no subsidy and status quo registration. All regressions are estimated by
OLS and weighted to reflect the intended cross randomization. Robust standard errors are reported in parentheses.
some of the households brought into the insurance system by reducing hassles were
particularly sensitive to the hassles of paying each month.
C. Attempted Enrollment
The enrollment measures in Table 1 and the coverage measures in Figures 3 and 4
mask the fact that many more households, particularly those in the assisted registra-
tion treatment, attempted to enroll than were actually successful. Table 2 sheds light
on this by examining the impact of the interventions on both attempted enrollment
in the first eight weeks after the intervention and actual enrollment. Panel A shows
that, averaged across all subsidy treatments, assisted registration led to a 23.7 per-
centage point increase in attempted enrollment during the first 8 weeks (column 1),
but only a 4.3 percentage point increase in successful enrollment during that period
(column 2). In other words, fewer than one-fifth of the households induced by the
registration assistance to attempt enrollment were actually successful in doing so.
Panel B explores the fully interacted effects of subsidies and registration assis-
tance. It estimates an enhanced version of equation (1) that also includes a full set
of interactions between the (cross-randomized) subsidy treatments and the assisted
registration treatment. Even with a full subsidy and assisted registration, enrollment
during the first 8 weeks only reached 25.9 percent, compared to 1.8 percent in the
Table 3—Reasons for Failing to Enroll
Observations Percent
(1) (2)
Panel A. All issues
Family card issues 468 83.6
Other issues 71 12.7
Technical reasons (internet, website) 21 3.8
Panel B. Breakdown of family card issues

Supporting document issue 286 61.1
Family card does not match the family members listed in the online system 91 19.4
Family already has insurance according to the online system 80 17.1
Family card not registered in the online system 11 2.4
Note: The sample includes households assigned to the assisted registration treatment that attempted to enroll within
six weeks from offer but failed to complete the registration process. Data are from enumerators’ recording at the
end of the baseline survey.
n o-intervention status quo (column 2). However, 55.6 percent of households tried to
enroll when offered both free insurance for the year and assistance with registration
(column 1). This points to important challenges to successful enrollment.
To explore the sources of these challenges in more depth, Table 3 reports the rea-
sons for failing to enroll among households that attempted to enroll in the registra-
tion assistance intervention. These data are from the enumerators’ recording of the
reason for failed enrollment.20 Over 80 percent of failed enrollment attempts were
due to an issue with the family card, the official identification document. As men-
tioned in Section II, the family card was required for registration because house-
holds were required to enroll all nuclear family members (defined as those listed on
the family card) as a way to combat adverse selection. The family card information
was supposed to be pulled automatically from the digital records held by the Home
Affairs Ministry, but this turned out to be problematic if the family composition
had changed, but the card had not been updated. In practice, updating the card is
challenging—it cannot be updated online, and instead requires at least one trip to
a Home Affairs-linked administrative office, and can often incur delays and other
additional costs. During in-person enrollment, social security administration offi-
cials could use discretion to overrule the system for certain causes (e.g., if house-
holds had documentation that the Home Affairs record was inaccurate), but the lack
of flexibility in the online system made web enrollment nearly impossible for many.
IV. Strategic Coverage Timing and the Impact of Time-Limited Subsidies
As discussed, the Indonesian government instituted certain features to try to com-

bat adverse selection—such as requiring all nuclear family members to enroll and
20
Specifically, enumerators were given a structured choice of seven possible answers, or other. We grouped two
possibilities (internet connection issue and BPJS registration website issue) into “technical issues,” and four possi-
bilities related to family card issues (supporting documents issue, family card is not registered in the online system,
family already has insurance according to the online system, and family card does not match the family members
listed in the online system). We do not have data on the reasons for enrollment failure in the status quo arms, but as
can be seen in Table 2 panel A, less than 1 percentage point of people who tried to enroll in the subsidy interventions
with status quo registration failed to do so, compared to 20 percentage points in the registration assistance arms.
imposing financial penalties for those who re-enroll after dropping coverage—but
was not allowed to impose a limited annual enrollment period. This opened the door
for strategic timing of enrollment when a health emergency occurs. Since the pro-
hibition on a limited annual enrollment period precluded our studying its potential
impact directly, we approximated it in spirit by making our subsidy offer explicitly
time limited, only available for up to two weeks after the offer was made. Our results
suggest that this time-limited subsidy was able to reduce strategic coverage timing.
Indeed, because the t ime-limited subsidies brought in healthier enrollees with fewer
claims, they allowed the government to cover almost double the number of house-
holds at no higher total cost, at least with the 20-month period we examined.
A. Impacts on Selection
Table 4 shows that the subsidies brought in healthier, lower cost enrollees, as
standard models would predict (e.g., Akerlof 1970). The analysis is limited to
households who enrolled and had coverage for at least one month during the first
year (as measured in column 1 of online Appendix Table 4), and shows the means
for each group.
Column 1 shows that the marginal household who received coverage in response
to the subsidies had a higher level of self-reported health at baseline than enrollees
in the no-subsidy group. The self-reported health score is a Likert score ranging
from one to four, with four as the highest option. The average s elf-reported health
of those enrolling with either the full or half subsidy is about 4.5 percent higher
than those enrolling with no subsidy (these differences are statistically significant
at the 5 percent level). The effects of assisted registration were smaller, but in the
same direction of bringing in healthier enrollees, and statistically significant at the
10 percent level.
The remaining columns examine enrollee claims for the 12 months after the
enrollment date. We focus on four main indicators: whether the household had any
claim, the number of claims (overall and for chronic visits), the total value of claims,
and the number of days to first claim. The latter is a way to proxy for the value of
claims with greater precision (Aron-Dine et al. 2015).
Those who enrolled under the subsidies were lower cost.21 For example, in
the n o-subsidy group, 62 percent had any claim compared to 48 percent in the
full-subsidy group (column 2; p-value = 0.040). Column 3 shows that those enrolled
in the f ull-subsidy group were also less likely to have a claim for a chronic, ongo-
ing condition (17 percentage points) than those enrolled in the no-subsidy group
(27 percentage points; p-value of difference = 0.082). Compared to the n o-subsidy
group, the full-subsidy group also had 40 percent lower average claims (column 6
of Table 4; p-value of difference = 0.095) and waited 32 percent longer before
submitting their first claim (column 7; p-value of difference = 0.006).22 Results for
the half-subsidy group and the assisted-registration group are mostly qualitatively
21
While these two measures capture different objects—namely, health and health-care usage—perhaps not sur-
prisingly, enrollees with better self-reported health indeed tend to have fewer claims (see online Appendix Table 6).
22
Online Appendix Table 7 reports the underlying regression estimates for the p-values reported in Table 4.
Table 4—Self-Reported Health and Claims in 12 Months since Enrollment, by Temporary Subsidies
and Assisted Registration
Had a claim Total number of claims Claims

Self-reported Of any Of any Value of Days to
health type Chronic type Chronic claims first claim
(1) (2) (3) (4) (5) (6) (7)
Full subsidy 3.237 0.480 0.171 3.589 0.184 0.986 232.425
[0.452] [0.501] [0.377] [6.805] [0.429] [2.620] [141.273]
Half subsidy 3.244 0.511 0.231 4.698 0.292 1.879 213.850
[0.541] [0.501] [0.422] [10.326] [0.624] [4.712] [150.106]
No subsidy 3.099 0.622 0.272 6.167 0.339 1.637 176.272
[0.538] [0.486] [0.446] [9.712] [0.600] [4.064] [154.913]
Assisted registration 3.217 0.525 0.226 4.254 0.259 1.366 214.823

[0.524] [0.500] [0.419] [7.419] [0.512] [3.676] [147.576]
Status quo registration 3.149 0.555 0.214 5.176 0.267 1.519 200.836
[0.505] [0.498] [0.411] [10.136] [0.583] [3.853] [150.794]

Full subsidy = no subsidy 0.016 0.040 0.082 0.025 0.036 0.095 0.006
Half subsidy = no subsidy 0.014 0.155 0.639 0.362 0.676 0.625 0.107
Full subsidy = half subsidy 0.888 0.540 0.138 0.164 0.053 0.046 0.278
Assisted registration = status quo 0.083 0.451 0.786 0.117 0.815 0.576 0.269
Note: This table shows mean self-reported health and mean claims submitted in months 1 to 12 since enrollment
by temporary subsidies and assisted registration. Means are weighted to reflect the intended cross randomization.
Standard deviations are in brackets. The sample is restricted to households who enrolled within a year since offer
and had coverage for at least one month over the same time period. The sample size is 749 households. In column 1,
the outcome is the average self-reported health of all household members, where the self-reported health score is
a Likert score ranging from one to four, with four as the highest option (better self-reported health). The value of
claims in column 6, in thousand Rp, is winsorized at the 99 percent level and only refers to hospital claims. The
p-values at the bottom of the table are from regressions of each outcome on indicator variables for assignment to all
treatment arms, an indicator variable for the randomization procedure used, and an indicator variable for study loca-
tion (equation (1)). Standard errors are robust. The p-values are from a test of the difference between the no-subsidy
and full-subsidy treatments (β2   = 0), between the no-subsidy and half-subsidy treatments (β1   = 0), between the
half-subsidy and the full-subsidy treatments (β1   = β2), and between the status quo and assisted registration treat-
ments (β3  = 0). All regressions are estimated by OLS and weighted to reflect the intended cross randomization.
Online Appendix Table 7 provides the regression estimates behind the numbers reported in this table.
similar to the full-subsidy group, but smaller in magnitude and never statistically
significantly different from the n o-subsidy group.
Figure 5 shows that these cost differences in part reflect strategic timing decisions
by the no-subsidy group, rather than only fixed health differences across enrollees in
different intervention groups. It plots the number of claims by month since enroll-
ment, separately by subsidy treatment groups; as with Table 4, the analysis is limited
to households who enrolled and had coverage for at least one month during the first
year. Those who enrolled without the subsidy submitted more claims in the first
few months upon enrollment than those who enrolled in the f ull-subsidy group, but
over time this difference became less stark and, by the end of the period, the groups
displayed similar patterns in number of claims. Enrollees in the h alf-subsidy group
also submitted more claims than households in the full-subsidy group, and even sub-
mitted claims for a higher value than the no-subsidy group in a handful of months.
In addition to the full-subsidy group having fewer claims than the no-subsidy
group, Figure 6 shows that, in the first 12 months since enrollment, those who enroll
and have claims in the full-subsidy group are much less likely to have “large claims”
(which are suggestive of a substantial health emergency) than those who enroll and
1
No subsidy
Half subsidy
0.8 Full subsidy
Number of claims
0.6
0.4
0.2
0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Months since enrollment
Figure 5. Number of Claims, by Month since Enrollment and by Subsidy Treatment
Notes: The figure shows the mean number of claims in each month since enrollment with 95 percent confidence
intervals for the mean. Means are weighted to reflect the intended cross randomization. The sample is restricted to
households who enrolled within one year since offer and had coverage for at least one month over the same time
period. The sample size is 749 households.
have claims in the no-subsidy group. Indeed, the probability distribution function
of the value of inpatient claims submitted within 12 months since enrollment is
markedly left-shifted for the full-subsidy group relative to the n o-subsidy group.
Again, the same is true—although less pronounced—in comparing the half-subsidy
and no-subsidy groups. The differences across groups are statistically significant
according to a Kolmogorov-Smirnov test for equality of distribution functions
( p-value = 0.012 for the test of equality between the distribution of the h alf-subsidy
and no-subsidy groups and p-value = 0.001 for the test of equality between the
distribution of the full-subsidy and no-subsidy groups). In short, when they use the
health-care system, those whose coverage was heavily subsidized have less expen-
sive health incidents.
Combined with the payment patterns from Figure 3 panel B, these results suggest
that n o-subsidy households may have had large claims once they enrolled, but then
stopped paying premiums (i.e., dropped coverage). In contrast, the subsidy groups
brought in healthier people, who kept paying premiums longer in the first year while
the subsidies were active (see Figure 3), and had smaller claims throughout the year
(Figure 5).
We also explored how the subsidies affected selection in terms of who retained cov-
erage in the period after all subsidies are over, which is important for understanding
the l ong-run cost implications of temporary subsidies. For each treatment, we divide
those who enrolled in the first year into “dropouts”—those who did not still have
coverage in month 15—and “stayers”—those who did. The results in Table 5 indi-
cate that in the full-subsidy group, those who retained coverage had higher base-
line s elf-reported health than those who did not (column 1; p-value = 0.072), but
also were more likely to have had claims (column 2; p-value = 0.005) and to have
0.0004
No subsidy
Half subsidy
Full subsidy
0.0003
Density
0.0002
0.0001
0
0 5,000 10,000 15,000
Value of claims (thousand Rp.)
Figure 6. Distribution of Inpatient Claims in Year 1 since Enrollment, by Subsidy Treatment
Notes: This figure shows the probability distribution function of the value of inpatient claims submitted within
the first 12 months since enrollment by subsidy treatment. The unit of observation is a single claim. The sample is
restricted to 749 households who enrolled within one year since offer and had coverage for at least one month over
the same time period, which is the same sample we use in Figure 5 and Table 4. The sample size is 3,827 inpatient
claims.
had more visits (column 4; p-value = 0.002), particularly for chronic conditions
(column 5; p-value = 0.14). The half-subsidy group showed a similar pattern of
claims,23 while the patterns are more ambiguous for the n o-subsidy group. The fact
that stayers are more likely to have had claims is consistent with an “experience
effect,” whereby having had experience utilizing the insurance can increase future
demand. These patterns also highlight the potential—emphasized theoretically and
empirically by Diamond et al. (2019)—for strategic dropouts to generate ex post
adverse selection, even among consumers who may be ex ante identical in their
expected health-care utilization. However, on net, the overall selection effect from
the full subsidy compared to the no subsidy still dominates its differential retention
effect, so that on net, the stayers in the full-subsidy group are healthier than the
stayers from the no-subsidy group, both in terms of self-reported baseline health and
in terms of claims during the first year.
B. Implications for the Government’s Budget
Table 6 shows the effects of the subsidies on the net revenue for the government.
Government net revenue is the difference between premiums received (net of sub-
sidies) and claim expenditures. It reflects the amount that would need to be covered
from the general government budget. The outcomes we have previously studied—
enrollment, coverage retention, and claims patterns—all affect government net
23
Online Appendix Table 8 presents the equivalent results broken down by the assisted registration treatment,
and finds a similar pattern: stayers were more likely to have had claims.
Table 5—Year 1 Claims by Retention in Year 2, by Temporary Subsidies
Had a claim Total number of claims Claims

Self-reported Of any Of any Value of Days to
health type Chronic type Chronic claims first claim
(1) (2) (3) (4) (5) (6) (7)
Panel A. Full subsidy
Dropouts 3.198 0.414 0.154 2.552 0.154 0.755 251.238
[0.441] [0.494] [0.362] [4.905] [0.362] [2.333] [134.730]
Stayers 3.313 0.610 0.205 5.653 0.246 1.446 194.983
[0.467] [0.490] [0.406] [9.213] [0.535] [3.077] [147.208]

Dropouts = stayers 0.072 0.005 0.340 0.002 0.140 0.073 0.005
Panel B. Half subsidy

Dropouts 3.248 0.413 0.155 2.773 0.192 1.600 240.240
[0.548] [0.494] [0.364] [5.618] [0.481] [4.225] [147.498]
Stayers 3.238 0.674 0.357 7.922 0.459 2.345 169.670
[0.531] [0.471] [0.482] [14.737] [0.784] [5.423] [144.754]

Panel C. No subsidy
Dropouts 3.169 0.530 0.234 3.632 0.327 1.840 197.490
[0.507] [0.503] [0.426] [5.470] [0.643] [4.231] [157.035]
Stayers 3.055 0.680 0.296 7.788 0.346 1.507 162.699
[0.555] [0.468] [0.458] [11.376] [0.574] [3.968] [152.741]

Panel D. Stayers across subsidy groups

Full subsidy = no subsidy 0.001 0.346 0.158 0.138 0.207 0.905 0.161
Half subsidy = no subsidy 0.026 0.937 0.418 0.941 0.269 0.239 0.778
Note: This table shows mean self-reported health and claims in the first year since enrollment, separately by tem-
porary subsidies and by whether households kept or dropped coverage at month 15 since offer. Means are weighted
to reflect the intended cross- randomization. Standard deviations are in brackets. The sample is restricted to house-
holds who enrolled within a year since offer and paid for at least one month over the same time period. The sam-
ple size is 749 households. In column 1, the outcome is the average self-reported health of all household members,
where the self-reported health score is a Likert score ranging from one to four, with four as the highest option (better
self-reported health). The value of claims in column 6, in thousand Rp, is winsorized at the 99 percent level and only
refers to hospital claims. The p-values in panels A, B, and C are from a specification where the outcome is regressed
on an indicator variable for whether the household has coverage in month 15 and the sample is restricted to house-
holds assigned to the subsidy treatment specified. The p-values in panel D are from a specification where the out-
come is regressed on indicator variables for subsidy treatment assignment and the sample is restricted to households
with coverage in month 15. All regressions are estimated by OLS and weighted to reflect the intended cross random-
ization. Standard errors are robust. The coverage rates of these two groups are shown in online Appendix Table 4.
r evenue. We show the total cost to the government per capita (i.e., per household in
the population).
Panel A shows the results during the time period the subsidy was in effect. As we
have seen, full-subsidy households have higher coverage during the subsidy period.
In fact, column 1 shows that the mean number of h ousehold-months within insur-
ance coverage in the full-subsidy group is almost nine times that of the no-subsidy
group during the subsidy period. For the full-subsidy group, revenue (i.e., premi-
ums received net of subsidies) per-household month is of course lower than for the
Table 6—Expenditures and Revenues, by Temporary Subsidies
Per household-month
Mean (i.e., total cost to the government in thousand Rp)
number of
household Net revenues
coverage Claims Net including
months Revenues expenditures revenues capitation
(1) (2) (3) (4) (5)
Panel A. Months 1 to 12 since offer date
Full subsidy 2.887 1.056 7.459 −6.403 −11.938
[4.899] [10.123] [157.686] [157.884] [159.105]
Half subsidy 1.096 4.210 17.687 −13.477 −15.185
[3.013] [18.013] [477.729] [475.367] [475.804]
No subsidy 0.323 2.214 4.927 −2.712 −3.167
[1.535] [14.270] [214.599] [212.905] [213.125]
Observations 5,996 71,952 71,952 71,952 71,952

Full subsidy = no subsidy 0.000 0.003 0.906 0.890 0.254
Half subsidy = no subsidy 0.000 0.000 0.033 0.073 0.042
Half subsidy = full subsidy 0.000 0.000 0.114 0.252 0.543
Panel B. Months 13 to 20 since offer date

Full subsidy 0.951 8.548 11.104 −2.556 −5.057
[2.317] [29.350] [239.891] [238.591] [238.819]
Half subsidy 0.593 6.280 13.823 −7.543 −8.935
[1.911] [25.460] [321.711] [319.543] [320.074]
No subsidy 0.451 3.966 8.594 −4.628 −5.565
[1.680] [19.484] [341.563] [340.097] [340.430]
Observations 5,996 47,968 47,968 47,968 47,968

Full subsidy = no subsidy 0.000 0.000 0.141 0.657 0.411
Half subsidy = no subsidy 0.034 0.003 0.181 0.332 0.292
Half subsidy = full subsidy 0.000 0.042 0.828 0.504 0.653
Notes: This table shows mean revenues and expenditures by temporary subsidies, for months 1 to 12 from offer
(panel A) and for months 13 to 20 from offer (panel B). Means are weighted to reflect the intended cross random-
ization. Standard deviations are in brackets. Column 1 reports mean number of months with insurance coverage.
Observations are at the household level. Columns 2 to 5 show mean revenues (premiums paid by enrollees), expen-
ditures (total value of claims), net revenues, and net revenues including capitation payments, in thousand Rp for all
household-months. Observations are at the household-month level. The value of claims in column 3 is winsorized at
the 99 percent level and only refers to hospital claims. The p-values are from regressions of each outcome on indi-
cator variables for assignment to all treatment arms, an indicator variable for the randomization procedure used, and
an indicator variable for study location (equation (1)). In column 1 standard errors are robust, while in columns 2
to 5 standard errors are clustered at the household level. The p-values reported are from a test of the difference
between the no-subsidy and full-subsidy treatments (β2  = 0), between the no-subsidy and half-subsidy treatments
(β1   = 0), and between the half-subsidy and full-subsidy treatments (β1  = β2). All regressions are estimated by OLS
and weighted to reflect the intended randomization.
no-subsidy group (column 2),24 but the value of claims per household month is not
statistically different, even despite the larger number of h ousehold-months cov-
ered (column 3). The last two columns report net revenue per h ousehold-month
24
Revenues should be mechanically zero for the full-subsidy group while the subsidy is in effect, but are not
literally zero since a few households in this group enrolled after the time period the subsidy offer was in effect and
therefore had to pay premiums.
with (column 4) and without accounting for capitation payments (column 5).25
Accounting for capitation payments, net revenue per h ousehold-month are about
negative Rp 3,000 in the no-subsidy group, compared to about negative Rp 12,000
per household in the full-subsidy group (although these differences are not statisti-
cally indistinguishable; p -value of 0.25). This implies that while the subsidies are
in effect, the government covers nine times more household-months at four times
the cost.
The financial implications for the government improve once the subsidies are
withdrawn (panel B). We observe about twice as many h ousehold-months cov-
ered in the full-subsidy group as compared to no subsidy (column 1; p -value <
0.001), but now they are also paying premiums. Putting this together, net revenues
to the government are almost identical between the full- and no-subsidy groups
(column 5)—it costs the government about Rp 5,000 per household in both the
no-subsidy and the full-subsidy arms. In short, the full subsidy was able to substan-
tially expand coverage in both the subsidy year and the year after, at no higher cost
to the government.
V. Conclusion
As incomes have risen in emerging economies, there has been a growing move to
increase coverage of health insurance programs through mandated, national enroll-
ment. However, many countries are running into key challenges in expanding cover-
age, particularly among informal workers, who often comprise a large share of their
populations.
In this paper, we show that there are both monetary and nonmonetary barriers to
enrollment. Time-limited, temporary subsidies increased enrollment. In fact, house-
holds that received the subsidies were even more likely to stay enrolled after the sub-
sidies expired, consistent with an experience effect. Reducing hassles by providing
registration assistance to households also increases enrollment.
These enrollment gains, however, were muted by a weak administrative state
capacity. In particular, the hassles of status quo (i.e., without assistance) enrollment
at a government office appear to be a substantial barrier: nearly as many people
attempted to enroll in the assisted internet-based registration with no subsidy as
did in the full-subsidy group with status quo office-based registration. Even more
important, many more people attempted to enroll when assistance was provided
than were actually able to do so. For example, when offered both a full subsidy and
assisted internet registration, about 55 percent of households tried to enroll, but
fewer than one-half of these households were successful. The primary reasons were
challenges with the government’s online enrollment system arising from inaccura-
cies in the state’s underlying civil registries of families, a problem that is unfortu-
nately common throughout the developing world. Even the 55 p ercent attempted
25
Capitation payments depend on the number of enrollees that declare the facility as their primary provider,
the total number of practitioners, the ratio of practitioners to beneficiaries, and operating hours, and range between
Rp 3,000–6,000 per enrollee for puskesmas and Rp 8,000–10,000 for clinics. Given that approximately 80 percent
of JKN Mandiri enrollees declare puskesmas and 20 percent declare clinics as their primary health facility, for these
calculations we assume capitation payments to be Rp 6,800 per enrollee per month. Capitation payments are only
paid to health-care facilities in months in which the household paid the premium.
enrollment represents a lower bound on the barriers to enrollment imposed by

weak administrative capacity; some individuals, anticipating failure, may not even
attempt to enroll.
The state’s inability to limit enrollment to a fixed period of time also contrib-
uted to low enrollment by encouraging households to strategically time their cover-
age. Households who enrolled in the control group tended to have a large insurance
claim in the first three months and also had high dropout rates, relative to those who
enrolled when we offered time-limited subsidies. The state’s limited capacity here
is directly due to current law, which requires that individuals be allowed to enroll in
social insurance programs throughout the year. However, even absent such legal con-
straints, the state might well face challenges in trying to offer a limited enrollment
period. Challenges could come from the demand side, since households in devel-
oping countries often have lumpy incomes that may not correspond to prescribed
enrollment periods. There could also be substantial administrative challenges in
implementing these limited enrollment periods. For example, when there is a limited
time period for enrollment in other settings, such as the annual enrollment period for
the health insurance exchanges under the Affordable Care Act in the United States,
exceptions are always allowed for certain “qualifying events”—such as a change in
marital status, the birth of a child, or a relevant employment change; such l ife-events
may be harder for the state to verify in developing countries, or would impose an
additional administrative burden on citizens since getting these types of documents
can be challenging. Of course, even when states can limit enrollment to once a year,
individuals may still engage in some strategic coverage dropping after time-limited
health-care needs are met (Diamond et al. 2019).
The results suggest several key dimensions through which governments in emerg-
ing economies can improve health insurance coverage. First, time-limited subsidies
can be used to entice the healthy to try the insurance; our results indicate that some
of those who try it will remain. Second, governments can make complementary
investments in reducing the hassles associated with registration and in the underly-
ing state registry systems needed to make them work efficiently.
Neither of these steps, however, is likely sufficient for achieving universal cover-
age. Our most intensive efforts only yielded 55 percent attempted enrollment. While
collectively, such steps may allow governments to cover substantially more house-
holds and at lower costs per covered household, our results leave open the question
of what is needed to achieve universal coverage in this setting. This is an important
area for further work. Other barriers certainly exist, including the possibility of opti-
mization frictions if households, for example, misperceive the value of insurance or
the quality of health care, as well as heterogeneity across individuals in the value of
health insurance.
REFERENCES
Agency for Healthcare Research and Quality. 2018. “Healthcare Cost and Utilization Project (HCUP)
Chronic Condition Indicator.” www.hcup-us.ahrq.gov/toolssoftware/chronic/chronic.jsp.
Akerlof, George A. 1970. “The Market for ‘Lemons’: Quality Uncertainty and the Market Mecha-
nism.” Quarterly Journal of Economics 84 (3): 488–500.
Alatas, Vivi, Abhijit Banerjee, Rema Hanna, Benjamin A. Olken, Ririn Purnamasari, and Matthew
Wai-Poi. 2016. “Self-Targeting: Evidence from a Field Experiment in Indonesia.” Journal of
Political Economy 124 (2): 371–427.
Aron-Dine, Aviva, Liran Einav, Amy Finkelstein, and Mark Cullen. 2015. “Moral Hazard in Health
Insurance: Do Dynamic Incentives Matter?” Review of Economics and Statistics 97 (4): 725–41.
Asuming, Patrick O. 2013. “Getting the Poor to Enroll in Health Insurance: Evidence from a Field
Experiment in Ghana.” Unpublished.
Asuming, Patrick O., Hyuncheol Bryant Kim, and Armand Sim. 2019. “Long-Run Consequences of
Health Insurance Promotion when Mandates are not Enforceable: Evidence from a Field Experi-
ment in Ghana.” Unpublished.
Badan Penyelenggara Jaminan Sosial Kesehatan (BPJS). 2021. “Administrative Data on Membership,
Payments, and Claims [database].” Jakarta, Indonesia. Accessed August 16, 2018.
Badan Pusat Statistik (BPS). 2015. “Survei Sosial Ekonomi Nasional (SUSENAS).”
Banerjee, Abhijit, Amy Finkelstein, Rema Hanna, Benjamin A. Olken, Arianna Ornaghi, and Sudarno
Sumarto. 2017. “Achieving Universal Health Coverage with an Unenforceable Mandate: Evidence
from the Government of Indonesia’s JKN Mandiri Program.” AEA RCT Registry. https://doi.
org/10.1257/rct.815-3.0.
Banerjee, Abhijit, Amy Finkelstein, Rema Hanna, Benjamin A. Olken, Arianna Ornaghi, and Sudarno
Sumarto. 2021. “Replication Data for: The Challenges of Universal Health Insurance in Devel-
oping Countries: E xperimental Evidence from Indonesia’s National Health Insurance: Dataset.”
American Economic Association [publisher], Inter-university Consortium for Political and Social
Research [distributor]. http://doi.org/10.3886/E138701V1.
Berchick, Edward R., Emily Hood, and Jessica C. Barnett. 2018. Health Insurance Coverage in
the United States: 2017. Current Population Report P60-264. Washington, DC: US Government
Printing Office.
Bettinger, Eric P., Bridget Terry Long, Philip Oreopoulos, and Lisa Sanbonmatsu. 2012. “The Role
of Application Assistance and Information in College Decisions: Results from the H&R Block
FAFSA Experiment.” Quarterly Journal of Economics 127 (3): 1205–42.
Bhargava, Saurabh, and Dayanand Manoli. 2015. “Psychological Frictions and the Incomplete
Take-Up of Social Benefits: Evidence from an IRS Field Experiment.” American Economic Review
105 (11): 3489–3529.
Cabral, Marika. 2017. “Claim Timing and Ex Post Adverse Selection.” Review of Economic Studies
84 (1): 1–44.
Cai, Jing, Alain de Janvry, and Elisabeth Sadoulet. 2020. “Subsidy Policies and Insurance Demand.”
American Economic Review 110 (8): 2422–53.
Chetty, Raj, and Adam Looney. 2006. “Consumption Smoothing and the Welfare Consequences of
Social Insurance in Developing Economies.” Journal of Public Economics 90 (12): 2351–56.
Delavallade, Clara. 2017. “Quality Health Care and Willingness to Pay for Health Insurance Retention:
A Randomized Experiment in Kolkata Slums.” Health Economics 26 (5): 619–38.
Diamond, Rebecca, Michael J. Dickstein, Timothy McQuade, and Petra Persson. 2019. “Take-Up,
Drop-Out, and Spending in ACA Marketplaces.” NBER Working Paper 24668.
Duflo, Esther, Michael Kremer, and Jonathan Robinson. 2011. “Nudging Farmers to Use Fertilizer:
Theory and Experimental Evidence from Kenya.” American Economic Review 101 (6): 2350–90.
Dupas, Pascaline. 2014. “Short-Run Subsidies and Long-Run Adoption of New Health Products:
Evidence from a Field Experiment.” Econometrica 82 (1): 197–228.
Dupas, Pascaline, Vivian Hoffmann, Michael Kremer, and Alix Peterson Zwane. 2016. “Targeting
Health Subsidies through a Nonprice Mechanism: A Randomized Controlled Trial in Kenya.” Sci-
ence 353 (6302): 889–95.
Einav, Liran, and Amy Finkelstein. 2011. “Selection in Insurance Markets: Theory and Empirics in
Pictures.” Journal of Economic Perspectives 25 (1): 115–38.
Finkelstein, Amy, Nathaniel Hendren, and Mark Shepard. 2019. “Subsidizing Health Insurance for
Low-Income Adults: Evidence from Massachusetts.” American Economic Review 109 (4): 1530–67.
Fischer, Torben, Markus Frölich, and Andreas Landmann. 2018. “Adverse Selection in Low-Income
Health Insurance Markets: Evidence from a RCT in Pakistan.” IZA Discussion Paper 11751.
Gruber, Jonathan, Nathaniel Hendren, and Robert M. Townsend. 2014. “The Great Equalizer: Health
Care Access and Infant Mortality in Thailand.” American Economic Journal: Applied Economics
6 (1): 91–107.
Gupta, Sarika. 2017. “Perils of the Paperwork: The Impact of Information and Application Assistance
on Welfare Program Take-Up in India.” Unpublished.
Handel, Benjamin R. 2013. “Adverse Selection and Inertia in Health Insurance Markets: When
Nudging Hurts.” American Economic Review 103 (7): 2643–82.
Institute for Economic and Social Research, Faculty of Economics and Business, Universitas Indone-
sia (LPEM-UI). 2015. “The Study on Examining the Sustainability of Premium Payments of JKN
Self Enrolled Member.”
Jensen, Anders. 2019. “Employment Structure and the Rise of the Modern Tax System.” NBER
Working Paper 25502.
King, Gary, Emmanuela Gakidou, Kosuke Imai, Jason Lakin, Ryan T. Moore, Clayton Nall, Manett
Vargas, Martha María Téllez-Rojo, Juan Eugenio Hernández Ávila, Mauricio Hernández Ávila,
and Héctor Hernández Llamas. 2009. “Public Policy for the Poor? A Randomised Assessment of
the Mexican Universal Health Insurance Programme.” The Lancet 373 (9673): 1447–54.
Lagomarsino, Gina, Alice Garabrant, Atikah Adyas, Richard Muga, and Nathaniel Otoo. 2012.
“Moving Towards Universal Health Coverage: Health Insurance Reforms in Nine Developing
Countries in Africa and Asia.” Lancet 380 (9845): 933–43.
Mikkelsen, Lene, David E. Phillips, Carla AbouZahr, Philip W. Setel, Don De Savigny, Rafael Lozano,
and Alan D. Lopez. 2015. “A Global Assessment of Civil Registration and Vital Statistics Systems:
Monitoring Data Quality and Progress.” Lancet 386 (10001): 1395–1406.
Muralidharan, Karthik, Paul Niehaus, and Sandip Sukhtankar. 2020. “Identity Verification Standards
in Welfare Programs: Experimental Evidence from India.” NBER Working Paper 26744.
Polyakova, Maria. 2016. “Regulation of Insurance with Adverse Selection and Switching Costs:
Evidence from Medicare Part D.” American Economic Journal: Applied Economics 8 (3): 165–95.
Ravallion, Martin, and Lorraine Dearden. 1988. “Social Security in a ‘Moral Economy’: An Empiri-
cal Analysis for Java.” Review of Economics and Statistics 70 (1): 36–44.
Sumner, Cate and Santi Kusumaningrum. 2014. AIPJ Baseline Study on Legal Identity: Indonesia’s
Missing Millions. Australia Indonesia Partnership for Justice.
Thaler, Richard H. and Cass R. Sunstein. 2009. Nudge: Improving Decisions about Health, Wealth,
and Happiness. London: Penguin.
Thornton, Rebecca L., Laurel E. Hatt, Erica M. Field, Mursaleena Islam, Freddy Solis Diaz, and
Martha Azucena Gonzalez. 2010. “Social Security Health Insurance for the Informal Sector in
Nicaragua: A Randomized Evaluation.” Health Economics 19: 181–206.
World Health Organization (WHO). 2008. “International Statistical Classification of Diseases and
Related Health Problems 10th Revision Version: 2008.” https://icd.who.int/browse10/2008/en#/.
This article has been cited by:
1. Pascaline Dupas, Radhika Jain. 2023. Can beneficiary information improve hospital accountability?
Experimental evidence from a public health insurance scheme in India. Journal of Public Economics
220, 104841. [Crossref]
2. Mona Balesh Abadi, Kevin Devereux, Farah Omran. 2023. Correcting for transitory effects in
RCTs: Evidence from the RAND Health Insurance Experiment. Canadian Journal of Economics/Revue
canadienne d'économique 56:1, 288-305. [Crossref]
3. Haotian Zhang, Wei Zheng, Wei Zhu. 2023. Procrastinated Participation in Social Insurance
Expansion: Evidence from China’s New Rural Pension Scheme. SSRN Electronic Journal 113. .
[Crossref]
4. Aurélien Baillon, Owen O'Donnell, Stella Quimbo, Kim van Wilgenburg. 2022. Do time preferences
explain low health insurance take‐up?. Journal of Risk and Insurance 89:4, 951-983. [Crossref]
5. Stephen Kwasi Opoku Duku, Edward Nketiah‐Amponsah, Christine J. Fenenga, Wendy Janssens,
Menno Pradhan. 2022. The effect of community engagement on healthcare utilization and health
insurance enrollment in Ghana: Results from a randomized experiment. Health Economics 31:10,
2120-2141. [Crossref]
6. Kathrin Durizzo, Kenneth Harttgen, Fabrizio Tediosi, Maitreyi Sahu, August Kuwawenaruwa, Paola
Salari, Isabel Günther. 2022. Toward mandatory health insurance in low‐income countries? An analysis
of claims data in Tanzania. Health Economics 31:10, 2187-2207. [Crossref]
7. Thomas Rouyard, Yukichi Mano, Bocar Mamadou Daff, Serigne Diouf, Khadidiatou Fall Dia, Laetitia
Duval, Josselin Thuilliez, Ryota Nakamura. 2022. Operational and structural factors influencing
enrolment in community-based health insurance schemes: an observational study using 12 waves of
nationwide panel data from Senegal. Health Policy and Planning 37:7, 858-871. [Crossref]
8. Johannes G. Jaspersen. 2022. When full insurance may not be optimal: The case of restricted
substitution. Health Economics 31:6, 1249-1257. [Crossref]
9. Aurélien Baillon, Joseph Capuno, Owen O'Donnell, Carlos Antonio Tan, Kim van Wilgenburg. 2022.
Persistent effects of temporary incentives: Evidence from a nationwide health insurance experiment.
Journal of Health Economics 81, 102580. [Crossref]

Banerjee Et Al 2021 The Challenges of Universal Health Insurance in Developing Countries Experimental Evidence From

Uploaded by

Copyright:

Available Formats

Banerjee Et Al 2021 The Challenges of Universal Health Insurance in Developing Countries Experimental Evidence From

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Banerjee Et Al 2021 The Challenges of Universal Health Insurance in Developing Countries Experimental Evidence From

Uploaded by

Copyright:

Available Formats

American Economic Review 2021, 111(9): 3035–3063

The Challenges of Universal Health Insurance in

To investigate barriers to universal health insurance in developing

As developing countries emerge from extreme poverty and enter m ­ iddle-income

We designed three interventions that simple economic theory suggested could

Second, weak administrative state capacity is a central impediment to achieving

A. The Indonesian Health-Care System

of annual household spending, after subtracting a subsistence food allocation. This

B. Mandatory, Contributory Public Health Insurance: The JKN Mandiri Program

In January 2014, the Government of Indonesia launched Jaminan Kesehatan

Institutional Features to Reduce Selection.—While insurance enrollment is

II. Experimental Design and Data

A. Study Setting and Sample

urban areas in Indonesia, respectively, each with a population of approximately

Upon identifying an eligible household, we administered a short baseline survey,

Status quo Assisted registration

Status quo Assisted registration

Figure 1. Experimental Design

­Time-Limited, Temporary Subsidies.—Households were randomly selected to be

Assisted Registration.—Registering for JKN Mandiri usually requires travel-

Information.—All study households received basic information, such as what

2015 2016 2017 2018

We estimate the following equation for a variety of outcome variables ​​y​i​​​:

​yi​​​ = ​β​0​​ + ​β​1​​ ​HALF SUBSIDY​i​​ + ​β​2​​ ​FULL SUBSIDY​i​​

+ ​β​3​​ ​ASSISTANCE​i​​ + ​INFO​ ′i​ ​​ β​4​​ + ​X​ ′i​ ​ δ + ​ε​i​​ ​,

​ ALF SUBSIDY​i​​​, ​FULL SUBSID​Y​i​​​, and ​​ASSISTANCE​i​​​are dummy variables

INFO​i​​​is a vector of dummies equal to one if household i was randomly assigned to a

III. Impacts on Enrollment, Coverage, and Enrollment Attempts

Table 1 presents the impact of the interventions on enrollment—i.e., successfully

Table 1—Effect of Temporary Subsidies and Assisted Registration on Year 1 Enrollment

Observations 5,996 5,996 5,996

p-value of test of hypothesis

Panel B. Information treatments, Medan

Observations 1,446 1,446 1,446

Panel C. Information treatments, Bandung

Observations 4,550 4,550 4,550

Panel A. By month since offer Panel B. By month since enrollment

Share of HHs with coverage

Figure 3. Insurance Coverage by Subsidy Treatment

Panel A. By month since offer Panel B. By month since enrollment

Share of HHs with coverage

Status quo registration 0.25 Status quo registration

Figure 4. Insurance Coverage by Assisted Registration

Table 2—Effect of Temporary Subsidies and Assisted Registration on Attempted Enrollment

Attempted to enroll within Enrolled within 8 weeks

Status quo registration mean 0.079 0.077

Panel B. Interacted specification

No subsidy, status quo registration mean 0.018 0.018

Table 3—Reasons for Failing to Enroll

Panel B. Breakdown of family card issues

IV. Strategic Coverage Timing and the Impact of ­Time-Limited Subsidies

As discussed, the Indonesian government instituted certain features to try to com-

Had a claim Total number of claims Claims

Assisted registration 3.217 0.525 0.226 4.254 0.259 1.366 214.823

p-value of test of hypothesis

Figure 5. Number of Claims, by Month since Enrollment and by Subsidy Treatment

Figure 6. Distribution of Inpatient Claims in Year 1 since Enrollment, by Subsidy Treatment

B. Implications for the Government’s Budget

Table 5—Year 1 Claims by Retention in Year 2, by Temporary Subsidies

As developing countries emerge from extreme poverty and enter m iddle-income

Time-Limited, Temporary Subsidies.—Households were randomly selected to be

We estimate the following equation for a variety of outcome variables yi:

yi = β0  + β1 HALF SUBSIDYi  + β2 FULL SUBSIDYi

+ β3 ASSISTANCEi  + INFO  ′i   β4  + X  ′i   δ + εi ,

ALF SUBSIDYi, FULL SUBSIDYi, and ASSISTANCEiare dummy variables

INFOiis a vector of dummies equal to one if household i was randomly assigned to a

IV. Strategic Coverage Timing and the Impact of Time-Limited Subsidies