Journal of Economic Literature 2024, 62(4), 1349–1421

https://doi.org/10.1257/jel.20241646

Social Protection in the

Developing World†
Abhijit Banerjee, Rema Hanna,
Benjamin A. Olken, and Diana Sverdlin Lisker*

Social protection programs have become increasingly widespread in low- and middle-in-
come countries, with their own distinct characteristics to match the environments in
which they are operating. This paper reviews the growing literature on the design and
impact of these programs. We review how to identify potential beneficiaries given the
large informal sector, the design and implementation of redistribution and income sup-
port programs, and the challenges and potential of social insurance. We use our frame-
works as a guide for consolidating and organizing the existing literature and also to high-
light areas and questions for future research. (JEL E26, H23, I13, I32, I38, O12, O16)

1. Introduction their children, has more than doubled (World

Bank 2018). These programs have increas-

S ocial assistance programs have been rap-

idly evolving in the developing world.
Today, safety net programs cover about
ingly taken the place of b
­ road-based subsidies
of basic commodities, promising increased
efficiency and an enhanced ability to redis-
2.5 billion people. More than 120 low- and tribute income to the poor. The COVID-19
­middle-income countries run cash transfer pandemic further expanded these policies,
programs for poor families, and more than 70 suggesting their increased role in policy in
of them run social pension programs. In just the future (Gentilini et al. 2020).
the past decade, the number of developing Designing these programs for a developing
countries running conditional cash transfer country context, however, entails challenges
programs, which condition benefits on house- that differ from those faced in more devel-
holds making human capital investments in oped economies. For example, one of the
biggest questions in the public economics
literature is how to target social assistance
*
Banerjee: Massachusetts Institute of Technology.
Hanna: Harvard Kennedy School. Olken: Massachusetts
Institute of Technology. Sverdlin Lisker: World Bank. We do not necessarily represent the views of the individuals
thank Amy Finkelstein, Nathan Hendren, the JEL edi- or organizations acknowledged here. Sverdlin Lisker was
tors, and four anonymous referees for helpful comments. generously supported by the National Institute on Aging,
We thank Cristine von Dessauer, Emily Romano, Beatriz Grant Number T32-AG000186 and is based upon work
Velho and especially Kim Lan Mallon for outstanding supported by the National Science Foundation Graduate
research assistance. Financial support from the Wellspring Research Fellowship under Grant No. 1745302.
†
Philanthropic Fund is gratefully acknowledged. The Go to https://doi.org/10.1257/jel.20241646 to visit the
views expressed in this paper are the authors’ alone and article page and view author disclosure statement(s).

1349
Copyright 2024 American Economic Association.
1350 Journal of Economic Literature, Vol. LXII (December 2024)

programs. However, unlike in high-income programs, which seek to assist the poor, and
countries, information on income is hard to social insurance programs, which aim to
observe, partly because informal work and insure against shocks, while acknowledging
self-employment is common, resulting in that many programs could incorporate both
many households being excluded from the features and that the two goals are often
income tax system ( Jensen 2022). This means interrelated.1
that the standard frameworks used for think- The paper is organized as follows.
ing about targeting benefits—for instance, Sections 2 and 3 focus on programs with pri-
income eligibility thresholds and endoge- marily redistributive goals, that is, programs
nous labor responses to effective marginal that aim to assist the poor. In section 2, we
tax rates (e.g., Mirrlees 1971, Saez 2002)—do begin by outlining what we call the “ex post”
not readily apply in a developing country set- problem; that is, after types have been real-
ting. This has led to the rise of other targeting ized, how should the social planner think
approaches in developing countries, which about improving welfare through redistribu-
in turn has necessitated research designed tion? In particular, section 2 asks the ques-
­specifically to understand these contexts (e.g., tion of who should receive transfers and the
Alatas et al. 2012, Hanna and Karlan 2017). degree to which transfers should be depen-
Similarly, program design may also need to dent on estimated or actual incomes for pro-
be radically different in a developing country grams with redistributive aims. We briefly
context. For example, one successful devel- discuss the standard Mirrlees-type model as
opment intervention is to give households a a benchmark, but then mainly focus on what
productive livestock asset—chickens, a goat, happens in contexts where the usual prox-
a cow—and support them in taking care of it ies for type in developed countries (such as
(e.g., Banerjee et al. 2015, 2016; Banerjee, income) are not easily observable. We briefly
Duflo, and Sharma 2021; Bandiera et al. 2017; describe a theoretical framework for thinking
Bedoya et al. 2019). This kind of program is though targeting issues in this context, and
based on the theory of poverty traps—that is, then review the empirical evidence on a wide
the idea that with non-convexities in produc- range of approaches for determining who
tion and savings constraints, a large lumpy should benefit from these types of programs.
transfer may push households into a higher Once you decide who needs help, how
steady-state income level. While poverty should you provide such help? How do you
traps could occur in more developed settings know the help is working? In section 3, we dis-
as well, the particular nature of the interven- cuss both how to evaluate the impact of redis-
tions to relieve these shocks may differ: An tributive programs, and what we know about
intervention that provides productive live- the impacts of the various types of programs.
stock may make sense in low-income coun- Specifically, we discuss how both the design
tries, given that 63 ­percent of the population and evaluation of programs may be influenced
is engaged in agriculture (ILOSTAT 2019), by frictions between the social planner’s util-
but would make less sense in a developed ity function and individuals’ utility functions,
nation where the vast majority of the poor when there is a cost of take-up, and when
and near-poor are not in agriculture. non-convexities mean that the differential
This paper aims to both outline theoreti-
cal frameworks for understanding the social
1 While active labor market programs, such as minimum
protection challenges in developing coun-
wage policies or job retention programs, often fall under
tries as well as provide a review of the litera- social protection programs in the policy space, we do not
ture to date. We focus both on redistributive cover them in this review.
 Banerjee et al.: Social Protection in the Developing World 1351

timing of transfers may have different impli- conditional on the realizations of incomes,
cations. We then review the existing empir- how to transfer resources from those with
ical literature on some of the additional low marginal utility of consumption to those
choices that one can make in designing these with high marginal utility. In this section,
programs in light of this framework. We close we start with what is often the key question
out this section with a discussion of gover- in such models—how to identify who has
nance concerns, which affect how programs high marginal utility and would benefit from
are delivered in the end, particularly in the additional resources—and discuss how this
context of low state capacity. problem differs in low- and middle-income
In section 4, we then turn to an ex ante countries from the way it has been concep-
framework—that is, before types are real- tualized for high-income countries.
ized—that allows us to consider the value The benchmark public finance model
of social insurance. Again, we start from the assumes there is unobserved heterogene-
standard frameworks and discuss the reasons ity among taxpayers and the social planner
private insurance markets may be incom- would like to tax those with high ability to
plete, and then discuss how the problems give transfers to those with lower ability, con-
and solutions may differ in countries with sidering the incentives for high-ability earn-
large informal sectors. We then review the ers to claim lower incomes in order to pay
evidence on the set of common programs fewer taxes and get more transfers. This is
that insure against expenditure loss and complicated by the fact that the planner does
income loss. Section 5 concludes. not observe earning ability and therefore
In reviewing the literature we seek to draw bases decisions on realized earnings. This
attention to important, but yet unanswered introduces a reason for high earners to claim
questions and new frontiers. To take just a lower earnings potential and undersupply
few examples: the fact that the large infor- effort (Mirrlees 1971).2
mal sector in developing countries demands As highlighted by Saez (2002), this
a different and perhaps more sophisticated approach identifies the elasticity of labor
approach to workplace accident and unem- supply with respect to taxes and with respect
ployment insurance schemes; how govern- to income as the key parameters in the design
ments should modify their social assistance of welfare programs. There is a large liter-
programs as income grows; and how the rise ature on this point for developed countries
of digital technology and mobile payments (e.g. Saez, Slemrod, and Giertz 2012; Meghir
will change options for both targeting and and Phillips 2010; Currie and Gahvari 2008),
implementation. We elaborate on these and motivated partly by the fact that many US
other topics throughout the review. welfare programs (especially prior to the
1996 welfare reform bill) implicitly put a
high tax rate on earnings through a steep
2. Targeting for Redistribution Once
phasing-out of benefits.
Income Has Been Realized
However, the literature suggests that
labor supply distortions might not be as rel-
2.1 Targeting When Income Is Observed:
evant for developing countries. Two recent
The Classic Public Finance Approach
and Its Applicability to Developing
Countries 2 We focus here on income as the key metric, following
the public finance tradition of Mirrlees (1971). However,
We begin by thinking about social pro- the government may have broader metrics of utility,
tection from the ex post perspective; that is, including physical and mental health.
1352 Journal of Economic Literature, Vol. LXII (December 2024)

experimental studies of wage elasticity, both sector: data-based approaches using eas-
from Malawi, report low values of short- ily observed characteristics of households
run labor supply elasticity (Goldberg 2016, (e.g., location and asset ownership), self-se-
McKenzie 2017). Moreover, Banerjee et al. lection, and community information. In the
(2017) reanalyze data from seven different rest of this section, we discuss the theory
experimental trials of large-scale, condi- and practice of these methods.
tional cash transfer programs where earnings
2.2 A Model of Targeting When Income Is
were explicitly not part of the conditionality
Imperfectly Observed
and find no evidence of income effects on
work. Baird, McKenzie, and Özler’s (2018) What to do when governments cannot
systematic literature review came to a sim- observe incomes? In a classic paper, Akerlof
ilar conclusion. The fact that both income (1978) discusses how “tags”—that is, fixed
effects and the wage elasticity are small sug- individual characteristics—can be used
gests that it is less important to focus on the to improve the targeting performance of
endogeneity of earnings in designing trans- self-selection schemes. The idea of using tags
fer schemes in developing countries com- forms the theoretical basis of what is known
pared to in high-income countries. as proxy-means testing, in which assets and
Governments in the developing world face other observable characteristics are used to
a different challenge: They cannot observe predict people’s incomes for targeting. This
earnings for informal workers and for those approach forms the basis for targeting in
with relatively low earnings, which collec- many developing countries, as we discuss in
tively comprise a large share of the pop- more detail in section 2.3.
ulation (Glewwe 1992, Jensen 2022). For These proxy-means tests allow the gov-
workers in the formal sector—which encom- ernment to target the poor without income
passes most of the workforce in high-income data. However, one challenge is mistar-
countries—third-party reporting means that geting: Targeting in such a system is based
governments can obtain reliable information on the statistical model relating the tags to
on people’s incomes. But since most work- incomes. This model, like all statistical mod-
ers in low- and middle-income countries are els, is imperfect, and the data on the tags is
informal, or below income tax reporting lim- often noisy, leading to both inclusion errors
its, this approach does not work. Many gov- (including those who are not poor) and exclu-
ernments of these countries therefore worry sion errors (excluding those who are poor).
that if program eligibility was based solely on This is exacerbated by the fact that the data
income or employment status, the govern- for targeting is only updated every few years
ment would need to rely on self-reports, and due to concerns about collection costs (in
many nonpoor people would simply report a contrast, most rich countries use frequently
low income to claim benefits. The key issue updated administrative data).
in many low- and middle-income countries Governments therefore face a choice
is therefore not how income-based eligibility between using these noisy proxies, using
rules distort actual labor supply, but rather potentially biased self-reports, or doing uni-
the difficulty of observing income at all. versal transfers and not trying to target at all
For this reason, instead of targeting (or some combination of these approaches).
transfers based on income or employment, In this section, we present a simple model
as in high-income countries, many devel- to think through these trade-offs. Our model
oping countries use a combination of three accounts for the two features of developing
options, at least for those outside the formal countries: First, incomes and employment
 Banerjee et al.: Social Protection in the Developing World 1353

are hard to measure precisely, since much basis for the targeted transfers, that is, what
work is in the informal sector (Jensen 2022). it should use as the “predicted income” ​​y​​  p​​  .
Second, as discussed above, labor supply Define the predicted income to be a lin-
elasticities tend to be low. Unlike in tradi- ear combination of a person’s self-reported
tional models such as Saez (2002), where the income ​​ỹ ​​and the government’s estimate of
trade-offs are primarily driven by the labor his or her income, y​​ ​​ e​​, where the superscript​
supply response, we will assume that labor e​refers to the government’s income estimate
supply is fixed and unaffected by redistri- from an audit:
bution. Instead, the key driver in our model
is the government’s ability to learn people’s (1) ​​y​​  p​  = α​  ỹ ​  + ​(1 − α)​ ​y​​  e​​.
true incomes in a low-information environ-
ment, potentially with noise, when people The audit here encompasses whatever
can misreport their incomes.3 procedure the government uses to observe
an estimate of income, such as a proxy-means
2.2.1 The Baseline Model
test (described in detail in section 2.3 below)
The problem is one of how to best redis- where the government proxies for income
tribute to a population of mostly low-income are based on assets it can observe. The esti-
citizens who are outside the formal sector. mate from the audit data is noisy, so the
We assume that the government can iden- estimated income ​​y​​  e​  = y + ϵ,​where ​ϵ​ has
tify the population that is broadly the target expectation zero and a fixed finite support.
of this redistribution scheme in that it can We study the choice of α ​ ​, that is, how much
exclude those with incomes above the pov- weight to put on self-reported income versus
erty line ​​y​​  *​​.4 Let ​h(​ y)​​represent the density the noisy audit-based estimate of income.
of the true underlying income distribution of If people misreport their income, there
those who might be eligible for the program, is a cost ​F = ​(a/2)​ ​​(​y​​  e​  − ​ỹ ​)​​​  2​,​based on the
​y​​  *​
so ​​∫0​  ​​h(​ y)​dy = 1​. difference between audited income estimate​​
We presume B ​ ​is funded by a combina- y​​  e​​and self-reported income. The cost may
tion of direct taxes on the formal sector and be a probability of getting caught multiplied
indirect taxes potentially on the entire pop- by some kind of a penalty. Here we assume
ulation. If the indirect taxes were linear in that the penalty does not contribute to the
the amount spent, nothing would change, program budget, so it could be a psychic cost
so think of the incomes in our model as or a bribe to a government official.5
income net of indirect taxes. We also pre- We follow Saez (2002) and Saez
sume that the budget ​B​is not so large that and Stantcheva (2016) to introduce the gov-
it could totally eliminate poverty with per- ernment’s preference for redistributing to
fectly targeted transfers to everyone, that is, the neediest by assuming that the govern-
​y​​  *​
​B < ​∫0​  ​​h(​ y)​​(​y​​  *​− ​y​​  p​)​dy​. ment’s maximand puts weight ​g(​ y)​​ on the
A key question that we will examine is
what data the government should use as the
5 We find similar results in the opposite extreme case
where the penalty is financial, such as a fine, and the pen-
3 For related work exploring optimal targeting schemes, alties help relax the government’s budget constraint, as in
see also Kleven and Kopczuk (2011). Chetty (2009b). This would also be the case if, for example,
4 This helps clarify that we are not designing the entire the fine was a bribe, as in the example in the text, but the
tax and transfer system here; we focus on the part that mat- salaries of the bureaucrats are adjusted downward to take
ters for the low income population. We also could consider account of their income from bribes in the spirit of Becker
a threshold y ​​ˆ ​  > ​y​​  *​​, or even set y
​​ˆ ​  = ∞​and consider the and Stigler (1974), so in effect a bribe is just a payment to
entire population, but that is not the object of interest here. the government; see the online appendix.
1354 Journal of Economic Literature, Vol. LXII (December 2024)

after-transfer utility of someone with true The formal analysis of the model is in the
income ​ y​, with ​​g′ ​(​ y)​  < 0.​We adopt the online appendix. The first result applies to
​y​​  *​
normalization ​​∫0​  ​​h(​ y)​g(​ y)​dy = 1.​6 In addi- the baseline model where we assume that
tion, we allow for the government to care everyone has the same cost of misreport-
about equity among those who have the same ing their income and that the government
true income (that is, we allow for “horizontal cares about horizontal equity (​u​is strictly
equity” considerations). We introduce this concave).
by assuming that the government’s utility
function for those who have a post-transfer Result 1: In the baseline model with homo-
income of ​z​is ​u​(z)​,​where ​u​is a (potentially geneous preferences over the penalty for
strictly) concave function. This formulation misreporting, the optimal redistributive
of the government’s maximand nests the stan- scheme always has the maximum feasible
dard utilitarian social planner if the amounts slope with respect to earnings given the bud-
involved in the redistribution are relatively get constraint and has no universal compo-
small but is much more general, allowing for nent ​​(T = 0)​​. The optimal scheme relies
a range of political economy considerations in entirely on self-reports (​​ α = 1)​​. Moreover,
the government’s maximand. this scheme delivers the first-best level of
One important way that ours differ from targeting.
traditional analyses (such as Feldstein 1999)
is that in our baseline model we assume that This stark result has a simple intuition.
the government’s maximand does not take As long as all potential beneficiaries have
into account the beneficiary’s disutility from the same cost of misreporting incomes, the
misrepresenting their income (i.e., does not amount by which they distort is predict-
include ​F = ​(a/2)​ ​​(​y​​  e​  − ​ỹ ​)​​​  2​​).7 This simpli- able and the optimal targeting mechanism
fies our analysis considerably and perhaps fits can take that fully into account, and hence
the context better as well. For example, if the can undo any misreporting that exists. The
penalty takes the form of a bribe to the gov- government can thus avoid using the noisy
ernment official who discovered the discrep- audit estimates because it can perfectly infer
ancy, as it often may, it is not unreasonable to people’s real incomes from their (uniformly)
assume that from the point of view of the misreported self-reports. Given that the gov-
government it nets out to zero.8 ernment can perfectly back out incomes, it
is optimal to use targeted transfers rather
than make suboptimal transfers to richer
6 As described by Saez and Stantcheva (2016), for con- households.
sidering small tax changes, the approach of using social This stark result also implies that
welfare weights ​g​(y)​​nests the idea of using a utilitarian
social welfare function, since one can set g​ ​(y)​​ equal to ­heterogeneity in the cost of misreporting, that
the marginal utility of income at ​y​. However, they argue is, heterogeneity in a​ ​, is key to understand-
for this approach as it is more general. It also allows us to ing why governments may want to use audit-
differentiate between a preference for redistribution (i.e.,
people with lower g​ (​ y)​​) and a preference for horizontal based estimates, why there are inclusion and
equity, which turns out to be important, as we describe exclusion errors in the targeting process, and
below.
7 This is related to the discussion in Chetty (2009b) on
why there may be reasons to limit the extent
social costs of taxation versus transfers. to which transfers are income dependent,
8 Even if the cost is primarily psychic, it seems unrea- because heterogeneity in ​a​means that the
sonable to assume that the government is very concerned government can no longer perfectly back out
about its people being forced to lie about its income.
Alternatively, we explore what happens when misreporting true incomes from distorted self-reports. We
costs enter the planner’s welfare function in appendix A.5. explore this next.
 Banerjee et al.: Social Protection in the Developing World 1355

2.2.2 Introducing Heterogeneity in the Cost on ​​α​​  2​​and therefore, for ​α​small enough,
of Misreporting raising ​α​has a second-order negative wel-
Suppose now that a​​ takes two values,​​ fare effect through this channel but a
a​1​​​ and ​​a​2​​​, with ​​a​1​​  < ​a2​ ​​​. Let ​μ​denote the first-order positive welfare effect because
fraction of the population that are ​​a​1​​​ types. it reduces the impact of the noise (​ ϵ​).
The cost of misreporting a​​is unobserved Hence, in this model, the government will
by the government. We assume that both not completely ignore self-reports, though
types have the same income distribution it can clearly now be optimal to combine
h (y). Heterogeneity in the cost of prevari- self-reports with audit-based estimates.9
cation introduces an important additional Turning next to ​t​, as before, a higher ​t​
consideration: Transfer schemes that rely still means more transfers go to the poor.
on ­self-reports, in effect, redistribute from However, unlike in the previous model,
those with high values of ​a​(those who do not there are now two forces pushing in the
misreport incomes very much) to those with opposite direction. First, the social losses
low values of a​ ​(those who do). That is, the from redistributing from ​​a​2​​​ to ​​a1​ ​​​ types are
government can no longer fully unravel mis- greater as ​t​is higher. Second, since we no
reports, which helps those who are willing to longer have ​α = 1​, there are also losses
misreport more. due to targeting noise ϵ​ ​, and these are also
This model therefore features a trade-off. stronger when ​t​is higher. Hence, relative
On the one hand, as the government relies to the case where there is no heterogeneity,
more on the unbiased but noisy audit-based the optimal value of ​t​may be lower.10 That
estimates and less on self-reports (i.e., as is, there is less redistribution when there is
we lower ​ α)​
, the redistribution from ​​ a​2​​​ more heterogeneity in the ability to misre-
types to ​​a​1​​​types falls. On the other hand, port income.
as before, the more the government relies Summing up, we have the following result
on audit-based estimates, the greater the
social welfare loss induced by the noise Result 2: If there is heterogeneity in the cost
(i.e., the ​ϵ​’s) in the audit process. The net of misrepresenting incomes, it can be optimal
effect of lowering ​α​depends on the relative for the government to use noisy audit-based
sizes of these effects. In the limit holding estimates as part of the targeting process. The
other parameters fixed, as the audit process weight placed on audit-based estimates rela-
becomes better and better (​var​[ϵ]​ → 0)​, tive to self-reports is increasing in heteroge-
the government will rely entirely on audit- neity in misreporting costs and decreasing in
based estimates (i.e., ​α = 0​); conversely, noise in the audit process. Moreover, it can
as the two types become more similar the now be optimal for the benefits to have a uni-
government will rely entirely on self-re- versal component and not be entirely income
ports (i.e., α ​ = 1​). based.
When both forces are present—that is,
when ​var[ϵ] > 0​but the two types are suffi-
ciently different—it can be optimal for the 9 In practice, if there is a fixed cost of collecting the
government to use the audit-based estimates self-report data, and if the optimal α​ ​is very small absent
(i.e., it can be the case that α ​ < 1​  , unlike these fixed costs, the government may not bother p
the fixed costs and may choose to ignore self-reports
­ aying

before), but it will always be optimal to rely altogether.

at least a bit on self-reports, so ​0 < α < 1​. 10 Of course, we may still be at the maximum value of ​t​,
but it is now possible that the optimal solution will feature
This is because the welfare loss from redis- some lump-sum component ​T​and less than the maximum
tribution from a​ ​​ 2​​​ types to ​​a​1​​​ types depends level of ​t​.
1356 Journal of Economic Literature, Vol. LXII (December 2024)

A numerical example showing how the may opt for some universal component in
optimal choices of ​α​and ​t​change as we social transfers because of a combination of
change the extent of heterogeneity and the concern for equity (concave ​u​), heteroge-
​var​[ϵ]​​can be found in online appendix A.4. neity in misreporting costs, and noise in the
We have so far assumed that those who audit process (​var​[ϵ]​ > 0​).
have a lower cost of misreporting have the
2.3 Targeting on Observable Characteristics
exact same income distribution as those with
in Practice
a higher cost. In online appendix A.3, we also
examine the case where heterogeneity in
2.3.1 Geographic Targeting
misreporting costs is correlated with income.
Doing so yields the additional insight that, in The simplest version of characteristic-
this case, we can drop the assumption that based targeting is geographic targeting, that
the government cares about horizontal equity is, targeting programs to poorer regions.
(i.e., that u​ ​is strictly concave) and still get Its main attraction is its simplicity and the
the result that the government relies at least administrative ease of implementing it, since
partially on audit data to deal with the pos- it does not require differentiating among
sibility that those with higher income have individuals within a region. The “character-
a lower misreporting cost. Indeed, we show istic” used for targeting is just the region
that with linear u ​ ​, if there are two types, with where someone lives.
the cost of misreporting negatively correlated The main question involved in geographic
with income, it will be optimal to completely targeting is how to choose the geographies in
ignore self-reports and set α ​ = 0​. a data-driven way (Baker and Grosh 1994).
If one has representative survey data at a
2.2.3 Summary
particular administrative level, one can use
Our model, though stylized, helps high- that dataset to estimate average incomes at
light the trade-offs in targeting transfers in that administrative level and target based
the presence of noisy data. We highlight that on that. To use finer geographic resolution,
the key challenge is heterogeneity in mis- one can combine survey sample data with
reporting costs, which prevents the govern- census data to predict the poverty status of
ment from backing out people’s true incomes smaller geographical areas (Elbers, Lanjouw,
from their self-reports. Given this, even the and Lanjouw 2003; Elbers et al. 2007), not-
optimal transfer scheme will typically involve ing the statistical caveats raised by Tarozzi
both inclusion and exclusion errors. The fact and Deaton (2009) or by using newer remote-
that the government responds to the imper- sensed satellite or administration datasets
fections in self-reported data by using noisy that help predict smaller regions (e.g., Jean
audit-based estimates of income generates et al. 2016; Blumenstock, Cadamuro, and On
additional targeting errors. 2015). These geographic-based approaches
We also see why it may be optimal to rely at typically produce less accurate estimates than
least partially on self-reports even when they individual targeting (since they miss all the
are biased, but also why if the beneficiaries poor people not in chosen regions), but are
are heterogeneous in terms of the cost of mis- much less data intensive and are administra-
reporting, self-reported data may be down tively easier to implement.
weighted or even ignored in favor of the data
2.3.2 Proxy-Means Tests
the government collects, even if that data is
noisy. The model also suggests reasons why Characteristic-based targeting at the
even a redistribution-minded government ­individual/household level is typically referred
 Banerjee et al.: Social Protection in the Developing World 1357

to as a proxy-means test (PMT). The imple- Targeting) of in Peru, Listahanan in the

mentation of a PMT is based on two datasets. Philippines, the NSER in Pakistan, and
One starts with a training dataset that includes Takaful and Karama in Egypt.
measures of what the government is actually The way proxy-means tests are imple-
trying to target (e.g., per capita household mented connects to the challenges high-
consumption, income, etc.), as well as the lighted by the model in section 2.2.1. First,
characteristics—usually demographics and the predicted poverty scores, which corre-
assets—that will be used for targeting. The spond to the audit-based estimates in the
idea is to select a set of observable character- model, are imperfect predictors of poverty.
istics (​X′​​) that, unlike income, the government To illustrate this, Hanna and Olken (2018)
can actually collect. This training dataset is simulate this type of targeting for two coun-
usually a nationally representative household tries—Indonesia and Peru—using house-
survey that is collected for research or gen- hold survey data and a set of assets and
eral statistical purposes; as such, households household characteristics that are typically
typically have no reason to lie about their used in proxy-means tests. They show that
actual consumption or income. Using these the prediction equation has substantial pre-
data, one then estimates a model of the form​​ dictive power—the ​​R​​  2​​is between 0.53 and
y​i​​  = X′ β + ​ϵi​​​​. 0.66—but it nonetheless has errors of both
To identify beneficiaries, one then needs inclusion and exclusion. Other papers show
a much larger dataset of the ​​X​i​​​ on the entire similar findings (see, e.g., Brown, Ravallion,
relevant population, that is, a census of the and van de Walle 2018).
entire country or a “social register” that con- There is also no reason to think that the
tains information from anyone who may plau- relationship between underlying income and
sibly be eligible. This step may also include the types of assets used in the PMT algorithm
some self-selection. For example, in some is the same for everyone. The model used to
cases the government will collect the char- predict PMT scores makes no allowance for
acteristics ​​X​i​​​only for those who apply, and differences in preferences: Someone might
hence only applicants may be deemed eligi- end up in a bigger house than what would
ble. In other cases the government attempts normally correspond with their income
to collect the ​​ X​i​​​for almost everyone by because, for example, they feel obligated to
having census enumerators go door-to-door provide temporary housing to their extended
throughout the country. Using this much family. As in the model, such heterogeneity
larger dataset, one calculates a predicted in the link between true income and PMT-
poverty score, ​​​yˆ ​​i​​​, using the characteristics ​​X​i​​​ predicted income is a major challenge in tar-
and the estimates of β ​​ˆ ​​, for each individual i​​ geting algorithms.
that is then used for targeting. Recent advances have sought to reduce
This type of proxy-means testing was first noise in PMTs in two ways. First, one can
used in the 1980s in Chile for the targeting replace the ordinary least squares (OLS)
of its Ficha CAS program and is now quite prediction equation y​ ​​i​​  = X′ β + ​ϵi​​​​ with
common (Coady, Grosh, and Hoddinott more sophisticated machine-learning pre-
2004a). Well-known examples of proxy- diction algorithms. However, most papers
means tests include the System for the find little to no improvements using these
Identification of Potential Beneficiaries of techniques relative to the common methods
Social Programs (SISBEN) in Colombia, currently used (e.g., McBride and Nichols
the Sistema de Focalización de Hogares 2018; Baez, Kshirsagar, and Skoufias 2019;
(SISFOH, or the System for Household Areias and Wai-Poi 2022).
1358 Journal of Economic Literature, Vol. LXII (December 2024)

Second, one can use new types of admin- To investigate these issues with individual-
istrative data—rather than collecting a level targeting in a modern PMT, Banerjee,
­door-to-door census—that allow more fre- Hanna et al. (2020) conduct a nationwide
quent, and hence up-to-date, data. For randomized experiment, in cooperation
example, Abelson, Varshney, and Sun with the Indonesian Government’s Central
(2014) use remote sensing to differenti- Bureau of Statistics, which administers the
ate between thatched and metal roofs in census used for actual PMT targeting. To
Kenya’s GiveDirectly program. Using data test for endogenous distortions, the s­ tatistics
from Togo, Aiken et al. (2022) find that bureau randomly added questions on flat-
using mobile phone metadata for targeting screen televisions and/or the number of cell
has substantive predictive power, though it phone SIM cards owned to the targeting
is not as accurate as a more traditional PMT. census in some randomly selected provinces,
The fact that one can use administrative data but not in others. While ­self-reports of tele-
such as these, which are continually and vision ownership fell in treated provinces six
automatically updated, suggests that target- months later, the effect quickly died out, and
ing using these data may be more responsive more importantly, there were no changes
to shocks.11 In fact, some countries (e.g., in actual television purchases or active SIM
Pakistan and Togo) used this approach to cards in treated areas.
target assistance during the 2020 COVID-19 In the framework of the model in sec-
crisis (Gentilini et al. 2022). Understanding tion 2.2, the fact that people do not seem
whether this method was effective is an to change their consumption decisions in
important area of ongoing work. response to the PMT questionnaire may
correspond to the case where the cost of dis-
2.3.3 Endogenous Characteristics?
torting choices (what we call a​ ​in the model)
A further challenge is that many charac- is very large. But this is not the only inter-
teristics used in this type of targeting are, pretation. First, recall that in the model,​
themselves, choice variables. This intro- a​referred to the cost of changing one’s
duces an additional potential source of reported income, not one’s real income.
inefficiency, since households could distort Along these lines, one interpretation is that
their consumption to remain eligible. For reported consumption can change even if
example, after England imposed a “window actual consumption does not (e.g., people
tax” in 1696—a tax levied on the number can hide the television when the enumerator
of windows a house had, which proxied for comes around). The Banerjee, Hanna et al.
wealth—people built new houses with very (2020) result had hints of this: Reported
few windows, leading to many dark houses ownership of televisions declined in gov-
(Oates and Schwab 2015). Even geography is ernment data six months later, even though
not an immutable characteristic: Households actual television ownership was unchanged.
could choose where to live in response to Similarly, Martinelli and Parker (2009) exam-
geographic targeting incentives. In practice, ine ­self-reported versus enumerator-verified
are these distortions substantial in modern PMT data in Mexico and find more under-
contexts? reporting of assets in the self-reported data
compared to the enumerator-verified data
when the monetary stakes are higher.
11 Others (e.g., Baez, Kshirsagar, and Skoufias 2019;
Beyond this, there is a second important
Mills and del Ninno 2015) suggest augmenting PMT tar-
geting systems with remote sensed and/or administrative consideration. In the model, we assume
data to help capture shocks. that the household knows exactly what to
 Banerjee et al.: Social Protection in the Developing World 1359

do to claim the benefits. In practice, the line, filling out lengthy paperwork, or going
fact that there are many different variables through an interview process). However, this
interacting in complex ways in the PMT approach imposes costs on the poor—for
formula means that households have lim- instance, lower quality food than they would
ited control over the outcome and there- prefer or time wasted standing in pointless
fore low returns to manipulation. In fact, lines. A key question is whether the target-
governments typically keep the formulas ing benefits, which would allow governments
secret for this reason: Studying the PMT in to deliver more aid to the poor, outweigh
Colombia, Camacho and Conover (2011) these additional costs, so that the poor indi-
find that local politicians were only able viduals themselves can nevertheless be bet-
to manipulate the data after the formula ter off even once these costs are taken into
became public, in order to sign up more of consideration.
their constituents for federal government
transfers.12 Workfare.—One of the most common
Finally, in contrast to PMTs, there is some ordeal mechanisms is workfare, that is,
evidence that geographic targeting can lead requiring beneficiaries to work in exchange
to real distortions. In particular, place-based for payments. These programs have a long
policies can induce migration to areas with history, dating back at least to the nineteenth
generous transfer programs and/or prevent century in England, where transfers were
out-migration from these areas. For example, granted through residence in a workhouse
Imbert and Papp (2020a) show that India’s (Besley and Coate 1992), and to the United
public works program, the Mahatma Gandhi States’ Civilian Conservation Corps and
National Rural Employment Guarantee Act Works Progress Administration in the 1930s
(MGNREGA)—which was geographically (Aizer et al. 2020).
targeted to rural areas and is discussed more Workfare programs are a classic example
in depth in section 2.4—reduced seasonal of an ordeal mechanism because work takes
migration to urban areas. time. If the wages for the workfare job are
low or the tasks are unpleasant or difficult,
2.4 Self-Selection Mechanisms
this will generate self-selection: Those who
Another targeting approach is s­ elf-selection. can get a more attractive job will select out
The idea is to pair benefits with an action (Ravallion 1991, Besley and Coate 1992).
where the required action is relatively less However, if the wage is set too high, the
costly for the targeted population (the poor) program runs the risk of crowding out more
than for the rest (the nonpoor) (Nichols productive private sector work. The net
and Zeckhauser 1982). If so, then as in the efficiency consequences also depend on
model, the selection rule allows the govern- whether the work being done is productive.
ment to make an inference about people’s Using workfare labor to build roads may have
income (perhaps noisily) from their self-cho- little efficiency loss; using workfare labor to
sen actions. The required action can include dig ditches no one needs would also gener-
buying low-quality food, work requirements, ate self-selection, but at much larger social
and administrative ordeals (e.g., standing in efficiency costs.
Perhaps the largest workfare program in
the world is India’s MGNREGA, which offers
12 Recent theoretical advances suggest ways to improve
100 days of paid employment per year to any-
targeting by anticipating problems and taking the poten-
tial responses of agents into account (Björkegren, one in rural areas who is willing to do casual
Blumenstock, and Knight 2021; Ball 2023). manual labor. Perhaps unsurprisingly, given
1360 Journal of Economic Literature, Vol. LXII (December 2024)

the hard work the program entails, Dutta to apply. This saves the government the cost
et al. (2014) find substantial ­self-selection in of verifying these applicants. Second, bene-
program take-up. ficiaries will use their own estimate of their
Given that workfare can crowd out other likelihood of passing the screening when
work, estimating the efficiency costs are deciding whether to apply. This estimate
challenging. Murgai, Ravallion, and van de will be a mix of the information used in the
Walle (2016) investigate this in the Indian screening mechanisms plus their private
state of Bihar by asking workfare participants information about their true income level.
their best estimate of what their earnings Thus, if people have imperfect information
would have been in the absence of workfare. or beliefs about the screening mechanism,
They find that those who joined MGNREGA asking them to decide to apply may induce
gave up income equivalent to about them to reveal some of this private informa-
30–35 ­ percent of the workfare income tion to the government.
received. Similarly, Bertrand et al. (2021) Alatas, Purnamasari et al. (2016) develop a
experimentally examine a youth workfare model that captures this idea and then they
program in urban Cote d’Ivoire. They find experimentally test a self-targeting mecha-
that while the program led to earnings gains, nism (application with verification done by
earnings gains are only 53 percent of the PMT) against automatic enrollment based on
transfer due to the crowding out of private a PMT done by census enumerators within
employment. Both programs set the work- the context of the expansion of the govern-
fare wage at the formal minimum wage, ment’s real conditional cash transfer (CCT)
which—unlike in many developed country program. They found that self-targeting led
settings—may be above the prevailing wage, to dramatically poorer beneficiaries, driven
given the large informal sector. This sug- in large part by those who are wealthier on
gests that workfare programs may have more dimensions that would be missed by the
effective targeting properties in places with PMT being less likely to apply. Surprisingly,
a lower formal minimum wage and with less this approach also reduces exclusion error
informal employment. We further discuss compared to the PMT—while the govern-
workfare programs in section 3.2.4. ment makes its best efforts in the automatic
enrollment system, some of the very poor,
Self-Selection with Small Costs.—One who live on the margins of society, can be
challenge with ordeal mechanisms such as missed; in the self-targeting group, these
workfare is that they impose very large costs individuals apply (see figure 1). This suggests
on beneficiaries—for example, hard labor all that self-selection at the application stage
day under the hot sun. Is it possible to get can reveal important information, poten-
the gains from self-selection without impos- tially decreasing both inclusion and exclu-
ing such large costs? sion errors.
One possible selection mechanism is to ask
2.5 Community Information Based
people to apply for a program, rather than
Targeting
to be automatically enrolled. Upon appli-
cation, programs then often add additional An alternative approach seeks to leverage
screening mechanisms, such as a PMT. This local information about people’s poverty
can affect selection in two ways. First, to the status by allowing communities to partici-
extent that prospective beneficiaries under- pate in choosing beneficiaries (Alderman
stand the screening mechanism, those who 2002; Galasso and Ravallion 2005). But, if
know they will not be eligible will not bother governments try to use this information for
 Banerjee et al.: Social Protection in the Developing World 1361

0.2
0.8

Automatic screening

Probability to receive benefits

0.6 Self-targeting 0.15
Poverty line
log consumption distribution
Density

0.4 0.1

0.2 0.05

0 0
11 12 13 14 15 16
log per capita consumption

Figure 1. Probability of Receiving Benefits for Those in Automatic Enrollment versus Self-selection Villages,
Based on Alatas et al. (2016e)

Notes: The graph shows the probabilities of receiving benefits in both the automatic enrollment and self-tar-
geting treatments described in Alatas, Purnamasari et al. (2016) (right axis), along with a histogram of log per
capita consumption (x-axis). The red vertical line represents the poverty line.
Source: Data from Alatas, Banerjee, Hanna et al. (2016).

community-based targeting, is the outcome $2/day, although the households whose iden-
better than a proxy-means test? tification switched were sufficiently close
Alatas et al. (2012) experimentally compare to the poverty line that both approaches
PMT-based targeting to community-based would perform similarly in terms of most
targeting. In their setting, community meet- social welfare functions. On the other hand,
ings were held in each sub-village and house- community targeting did better in terms of
holds were ranked based on their relative identifying households who self-identify as
poverty levels. Those ranked as among the poor. As a result, citizens in villages random-
poorest households were given transfers. ized to community targeting reported higher
The results suggest that the optimal satisfaction with both the targeting process
approach may depend on the government’s and outcomes than in those with PMT-based
objective function, and in particular, on targeting.13
whether per capita consumption or a differ- These results suggest that communi-
ent, perhaps more locally based, notion of ty-based targeting can work. However, there
poverty is the “correct” measure of needi- are often concerns about elite capture in
ness. On the one hand, the authors find that
the PMT did better at identifying households
13 The gain in legitimacy from community target-
based on per capita consumption—the PMT
ing may be important. For example, Cameron and Shah
had about a 10 percent lower error rate in (2014) show that mistargeting in Indonesia’s cash transfer
terms of identifying households below PPP programs increased crime.
1362 Journal of Economic Literature, Vol. LXII (December 2024)

community-based targeting. That is, local PMTs tend to do at least a somewhat ­better
leaders might disproportionately choose job at identifying households with lower
themselves, family members, or others who objective per capita consumption, and there
are not necessarily the most in need. There is some heterogeneity in the extent to which
is therefore a potential theoretical trade-off community-based targeting does better on
between the superior local information and other metrics, such as self-assessed welfare,
greater capture of these programs by local which in itself may capture broader metrics
elites (e.g., Bardhan and Mookherjee 2000). of well-being relative to consumption.
Empirically, however, elite capture seems Premand and Schnitzer (2021), for exam-
to be small, at least in comparison to other ple, ran an experiment in Niger randomiz-
sources of targeting failures. In a field exper- ing villages about to receive a national cash
iment in Indonesia, Alatas et al. (2019) find transfer to be targeted using a PMT, a com-
that increasing elite decision-making over munity-based approach, or a formula that
targeting does not substantially change out- aims for food-insecure households. They
comes. More generally, while they do find find that households selected by the PMT
some evidence of elite preference in some are 8 percentage points more likely to be
programs, it is small in magnitude, so that poor than those selected by the community.
the gains one would achieve by reducing elite Like Alatas et al. (2012), they also find that
capture are entirely dwarfed by the gains community-based targeting does a better job
that one could make by improving the qual- matching self-assessed welfare status and the
ity of PMT data.14 Similarly, Basurto, Dupas, perceptions of others (though the difference
and Robinson (2020) study the targeting of is not large) and identifying households with
subsidies for both agricultural inputs and recent shocks.
food in Malawi, which is done by local chiefs, Several other studies (Stoeffler, Mills,
and compare it to a hypothetical targeting and del Ninno 2016; Beaman et al. 2021;
done through a PMT. While they find that Dupas, Fafchamps, and Houeix 2022) also
nepotism exists, the welfare consequences find that community-based targeting does
are small, again because those who receive worse than PMTs when the desired outcome
preferential treatment are also relatively is per capita consumption, though these
poor; moreover, there are potential gains as studies do not typically assess their compar-
chiefs use their local knowledge to target. ative effectiveness vis à vis self-assessed wel-
While these studies suggest that commu- fare. Other studies suggest that the efficacy
nity-based targeting has potential, especially of community-based targeting compared to
when assessed against local objective func- a PMT depends on the variables included
tions (such as self-assessed welfare) or to in the PMT (e.g., ­Sabates-Wheeler, Hurrell,
fill in the gaps in the PMT (e.g., detecting and Devereux 2015).
shocks, finding productive households), it These papers suggest directions for future
does not strictly dominate PMTs, and it may research. First, are there ways to make
not be appropriate in all settings. Indeed, community selection more effective? For
several studies find results that confirm that example, Alatas et al. (2012) find that house-
holds that were randomized to be ranked
14 The welfare losses from elite capture are essentially first are ranked more accurately than those
the product of (a) how much richer elites are (9 ­percent), ranked later, suggesting people get tired.
(b) how much more likely elites are to obtain bene- This suggests that the microstructure of the
fits (19 percent in the worst program considered), and
(c) elites’ proportion in the general population. Multiplying algorithm used to elicit the community rank-
yields a welfare cost less than 1 percent. ing can affect the ranking’s accuracy, and
 Banerjee et al.: Social Protection in the Developing World 1363

­ nderstanding this better can help improve

u costs vary across people. For example, find-
the information elicitation process. Second, ing appropriate childcare in order to spend
does the efficacy of community targeting an afternoon signing up for a program might
differ systematically? For example, Alatas, be more difficult for some. Moreover, not
Banerjee, Chandrasekhar et al. (2016) find everyone who we think is poor is actually
that the community-based targeting treat- poor—as described above, the data is hardly
ment in Alatas et al. (2012) was more effec- perfect.
tive at identifying households within areas However, the magnitude of incomplete
with more diffusive interpersonal network take-up in social protection programs sug-
structures. More generally, being able to gests that the problem goes beyond optimal
predict in which types of locations commu- self-targeting. For example, Bhattacharya
nity-based targeting is likely to be effective et al. (2015) report that two-thirds of women
is an important direction. A third question is eligible for a noncontributory social pension
whether community-based targeting can be for life are not enrolled in the program. The
compatible with cross-community targeting. generosity of the pension (the equivalent of
That is, one can allocate a fixed number of 70 percent of median per capita consump-
slots to a community (for example, based tion across urban areas) makes it unlikely
on geographic targeting, a past census, or that the low take-up could be rationalized by
another PMT approach) and ask them to a cost–benefit calculation.
choose which households should receive
2.6.2 Application Costs and Information
those slots. But if a community is asked to
Barriers
identify who is poor on an absolute level, it
has little incentive not to list many people as The fact that so many of the poor miss
poor in order to maximize government trans- out on benefits suggests that some non-
fers. Designing incentive-compatible proce- take-up might be inefficient. Finkelstein
dures for community-based targeting that and Notowidigdo (2019) argue that in the
reveal the level of poverty, not just who is the United States, misperception of the benefits
poorest, seems like an important mechanism and costs might explain part of the problem.
design problem for future research. They find that benefit information nearly
doubles take-up (from a low base); adding
2.6 Incomplete Take-Up and What We Can
sign-up assistance triples take-up.15
Learn from It
There are a number of related experi-
The flip side of self-targeting—through ments in the developing world that exam-
which take-up decisions can help improve ine how much of the take-up challenge can
targeting—is that many people do not take be attributed to the (perceived) hassles of
up the program even when they are entitled applying to programs by providing enhanced
to it. How should we think about this? Does take-up assistance to targeted households.
this reflect a rational cost–benefit calculus, These studies typically find that while
information, or other optimization frictions, take-up assistance can meaningfully increase
behavioral constraints (e.g., procrastination),
stigma, or something else? Is this optimal? 15 Finkelstein and Notowidigdo (2019) also note that
2.6.1 How Prevalent Are Take-Up improving take-up on the margin does not automatically
improve welfare if households are fully optimizing and
Challenges? rationally not signing up for programs given the costs of
doing so. However, this result depends on the fact that
The fact that not everyone takes up pro- households are fully optimizing, which is often not the case
grams is not, per se, surprising. Self-selection in many contexts.
1364 Journal of Economic Literature, Vol. LXII (December 2024)

program take-up compared to not having application assistance increased attempted

assistance, it does not close the majority of enrollments substantially—by 16 percentage
the take-up gap. points, compared to just 1.8 percent in the
For example, Carneiro, Galasso, and Ginja control group—the vast majority of those
(2019) study a program known as Chile who attempted to enroll failed. As a result,
Solidario that targeted extremely poor providing assistance only increased enroll-
households with extensive home visits by a ment by 2.4 percentage points.
social worker over a two-year period, includ- A common theme among these papers is
ing guidance on how to apply for government that the take-up problem is not easily solved
programs as well as a financial stipend. The by helping households apply. Instead, the
program itself had imperfect take-up; only fact that there is any paperwork may itself be
about 20 percent of those eligible for Chile an important barrier. This suggests that pro-
Solidario enrolled. For those who availed grams may want to radically simplify appli-
themselves of Chile Solidario assistance, cation requirements, such as by pre-filling
take-up of a separate government family forms using administrative data. However,
child allowance program (SUF) increased this only works if the underlying state sys-
by 17 percentage points, from a base of 53 tems are accurate. Many of the problems
percent, 4 years later. The intensive take-up Indonesian households encountered in
assistance of Chile Solidario mattered, but it Banerjee et al. (2021) came about because
only closed about one-third of the take-up the government’s underlying family records
gap. were inaccurate, a problem that could only
In a second example, Gupta (2017) looks be fixed by a costly trip to a government
at widows and divorcees in Delhi, India, and office. Improving the underlying adminis-
investigates why they fail to enroll in the trative capacity to allow for simpler enroll-
pension programs. Applying is a substantial ment procedures may be challenging, but
hassle: Widows must fill out forms, provide important.
supporting documentation (e.g., proof of
2.6.3 Stigma
marital status, unique identification card,
proof of five-year residency), get a local pol- A common refrain in the developed world
itician’s signature, and submit all this to the is that low take-up reflects low demand for
local district government. Intensive assis- government services due to the stigma asso-
tance also helped in this case, but it did not ciated with receiving benefits (Moffitt 1983).
substantially close the take-up gap: 22 per- While there is comparatively little research
cent of eligible women receiving intensive on stigma in developing countries, what
mediation successfully obtained the pension, exists suggests less concern about stigma.
compared to 15 percent of control house- For example, Osman and Speer (2024) con-
holds. While this is a 47 percent increase, it duct experiments in Egypt to recruit young
still left behind 80 percent of eligible women. people to labor market assistance programs,
Banerjee et al. (2021) examine related varying the recruitment messages to distin-
questions in Indonesia in the context of guish the effects of different types of stigma.
national health insurance. Universal health Several messages designed to reduce “pro-
insurance was launched in 2014 and by 2015 fessional” or “social” stigma actually seem to
the contributory portion of the program had have primed respondents to be more alert to
enrolled less than 20 percent of the intended these concerns. However, an explicit “wel-
population despite being officially manda- fare stigma” framing—that is, that the cost
tory. Banerjee et al. (2021) find that while of the job training program was subsidized
 Banerjee et al.: Social Protection in the Developing World 1365

“to help those in financial hardship”—had 2018). Several papers use these techniques
no effect whatsoever on program take-up. to show that they can identify heterogeneous
Nevertheless, understanding the role of treatment effects to improve the target-
stigma in more detail—and in particular ing of immunizations (Chernozhukov et al.
whether or not program participation is pub- 2018), workfare (Bertrand et al. 2021), and
lic information—is an important direction cash transfers (Haushofer et al. 2022). Caria
for future research. et al. (2024) combine this approach with a
dynamic experimental design to simultane-
2.7 Targeting on Treatment Effects, Not
ously learn which treatments work best opti-
Levels
mally and which work best for which types
The discussion thus far has focused pri- of respondents for job training programs for
marily on providing programs to the poor. Syrian refugees in Jordan.
Implicit in this view is that anti-poverty pro- A second option is to encourage
grams are more effective at increasing pri- self-selection on the basis of treatment
­
vate or social utility when the household is effects. Dupas et al. (2016), for example,
poorer. In our model, this was the result of investigate a program that provides free
the assumptions that ​g′​(y)​  < 0​ and/or that​ chlorine in Kenya. They compare the free
u​(y)​​ is concave. distribution of chlorine versus the distribu-
Targeting is more complicated, however, tion of vouchers redeemable at nearby shops
when there is heterogeneity in program for free chlorine. Redeeming a voucher is
treatment effects. Imagine that a program’s a screening mechanism as only those who
only effect is to generate income, but that intend to use the chlorine will redeem it.
the effect on income is heterogeneous by Indeed, they find similar rates of chlorine
individuals, that is, we write ​d​y​i​​​. Now, opti- in water in the two groups, yet the vouchers
mal targeting would want to find those indi- saved 60 percent of the cost by not distribut-
viduals for whom ​g(​ ​y​i​​)​ ​u′ ​(​ ​yi​​​)​d​yi​​​​ is highest, ing chlorine to those who would not use it.
which in this particular formulation is a These examples illustrate the potential to
combination of the individual’s specific treat- apply these approaches to social protection
ment effect and income. Even for programs more broadly. For example, the graduation
that do not target income, the point is that approach (discussed in section 3.2.6) may
one is no longer interested in targeting the be appropriate for some types of households
program just based on an observable charac- and not others. Can one identify them based
teristic (i.e., poverty score), but also on the on ex ante characteristics? Or find a way for
program’s treatment effect, d ​ ​y​i​​​. people to self-select in? On the flip side, for
This is a harder problem, because gener- social insurance products, one does not want
ally speaking d ​ ​y​i​​​ is unobserved. However, to encourage either adverse selection or
there are still several options. One option is so-called “selection on moral hazard” (Einav
to predict d ​ ​y​i​​​based on observable covari- et al. 2013). For example, making enrollment
ates, for example, from a randomized trial. to health insurance easier could potentially
Bhattacharya and Dupas (2012) consider this encourage the healthier to enroll (Banerjee
approach in the context of subsidies for anti- et al. 2021).
malarial bed nets, where they account for the There are two caveats worth noting. First,
probability of bed net use. Other recent work for this approach to work, it is important to
shows how to use machine learning to esti- be able to identify who could benefit the
mate heterogeneity in treatment effects (e.g., most, and that may not be feasible in all
Wager and Athey 2018, Chernozhukov et al. settings. For example, Carter, Tjernström,
1366 Journal of Economic Literature, Vol. LXII (December 2024)

and Toledo (2019) study the Rural Business potential to reduce administrative costs
Development Program in Nicaragua and (by obviating the need to collect household
show that observable characteristics do not data), though as a practical matter these costs
do a good job at predicting which households are usually very small compared to the ben-
would benefit the most. Second, we need to efits being given out. Finally, they may also
think carefully about the ethics of target- improve the political economy of redistribu-
ing on productivity. For example, there is a tion (Banerjee, Niehaus, and Suri 2019).
concern that targeting on work productivity However, a key challenge is that for a given
for a cash transfer program could screen out budget, universally eligible programs transfer
people who really need help but may not be substantially fewer resources to each benefi-
able to work, such as people who have severe ciary than targeted programs. That is, solving
disabilities. More work needs to be done to the problem of eliminating exclusion error
understand these issues. by giving a program to all can be expensive
(Hoynes and Rothstein 2019). While this is
2.8 Should We Even Target at All?
true everywhere, this trade-off is particularly
Universal versus Targeted Programs
severe in developing countries because the
All of the targeting methods discussed money is not easy to tax back given the infor-
thus far suffer from exclusion and inclusion mal economy. As a result, universal transfers
errors (Brown, Ravallion, and van de Walle end up giving the same net transfer quite
2018). Given these challenges, as described high up the income distribution in develop-
in section 2.2, another option is to make pro- ing countries (Hanna and Olken, 2018). This
grams universal, such as through a universal makes universal programs particularly expen-
basic income (UBI) program or a universal sive relative to targeted programs.
in-kind program. These programs involve Deciding which type of program is pref-
giving each individual a fixed amount of erable (targeted versus universal) ultimately
money or goods—regardless of income— depends on the social welfare function and
and are financed through proportional or the relative importance that the government
progressive taxation. Conceptually, one can gives to inclusion errors, exclusion errors,
think of them as shifting the intercept of the and per capita benefits, as described in the
tax schedule and modifying the rates (i.e., theory in section 2.2.
the slope of the tax schedule) in such a way To examine these trade-offs in the data,
that it satisfies the government’s budget con- Hanna and Olken (2018) calibrate this
straint (Saez 2002; Hanna and Olken 2018). using standard constant relative risk aver-
These programs could literally be available sion (CRRA) utility functions and find that
to everyone, as in a UBI, or they could be for a given budget, programs targeted using
available to everyone in a particular category standard PMTs substantially outperform
(e.g., the elderly, children), but the key dis- universal transfers because they transfer sub-
tinction is that they are not conditioned on stantially more resources to the poor. But,
income or proxies for income. they do entail more horizontal inequity, so
Universal transfers have several advan- the trade-off ultimately comes down to how
tages. In principle, they have no exclusion much one values increasing welfare of the
error, though in practice take-up may be far poor on average versus reducing exclusion
from universal. As such, they also avoid hor- errors and eliminating horizontal inequity.
izontal inequity problems (i.e., the c­ hallenge One important caveat is that while tar-
that similar people may not receive like geted programs may be more effective in
benefits). Universal transfers also have the improving welfare on average, even in the
 Banerjee et al.: Social Protection in the Developing World 1367

case of imperfect targeting (Coady, Grosh, administrative costs of going to relatively

and Hoddinott 2004b; Hanna and Olken rich regions with few of the poor) and then
2018), this assumes that the budget is fixed. use PMTs within these regions to find those
As the elasticity between the total budget most in need. Others augment the PMTs—
and the number of beneficiaries increases, which we know are imperfect—with com-
universal transfers may become more attrac- munity methods to fill in the gaps and reduce
tive politically (Klasen and Lange 2016). exclusion errors.
Many have argued that the durability of pro- While research has been growing in this
grams like Social Security and Medicare in area, substantial questions exist for future
the United States comes from their universal work, from how to best conduct dynamic tar-
eligibility, which makes them politically pop- geting, whether we can target on productivity
ular. Understanding the political response to or target different programs toward those who
targeted as opposed to universal programs, would gain the most from a particular type
or programs (like Social Security) that are of program, how to best use newer adminis-
both universal and progressive, is an import- trative and satellite data, how to reduce the
ant dimension for future work.16 take-up problem, how targeting affects politi-
cal and budgetary outcomes, and when stigma
2.9. Summary of Targeting
matters most for take-up.
Targeting in developing countries often is
quite different than in developed countries,
3. Transfers for Redistribution:
given information constraints and high levels
Understanding Program Types and Impacts
of informality. The growing literature pro-
vides indications of the trade-offs between The previous section discussed who should
methods under different contexts: That is, receive assistance. We now turn to under-
what “works” really depends on the extent of standing how to think about the impacts of
development and informalization (e.g., Chile transfer programs for redistribution as well
may be different than Malawi), what the as the impacts of various types of programs.
government is trying to target on (e.g., sys- In section 3.1, we begin by clarifying some of
tematic poverty, shocks, productivity), insti- the theoretical issues in evaluating the impact
tutional capacity, and the kinds of frictions of transfers. Section 3.2 then discusses the
that impact take-up. evidence on the many types of transfer pro-
It is also worth noting that given that grams in low- and middle-income coun-
each method has different strengths and tries. In section 3.3, we discuss governance
weaknesses, and many of these may be challenges in making sure the aid is actually
complementary, in practice, many devel- delivered, which is an overarching issue in
oping countries “mix and match” ­methods the impacts of social protection programs,
to improve ­targeting efficiency and reduce particularly in the low institutional capacity
administrative costs (Coady, Grosh, environments observed in many low- and
and Hoddinott 2004a). For example, many middle-income countries.
countries first geographically target pro-
3.1 Theory: Welfare Analysis of
grams to hard-hit regions (to save on the
Redistribution Programs

16 More broadly, understanding the political economy 3.1.1 A Simple Framework

of social protection programs is important for policy design
and future research. We refer the interested readers to We have so far assumed that a dollar of
online appendix B for an overview. benefits is worth a dollar to the person who
1368 Journal of Economic Literature, Vol. LXII (December 2024)

gets them. In the public discourse, however, household to make certain other choices in
there is a lot of discussion about both the order to receive the benefit.
form the transfer should take and how the We assume that a household is fully
household makes use of the benefits. We will sophisticated in its understanding of these
argue that under a set of quite standard con- constraints and how they are affected by b ​ ​  .18
ditions, the expression for the social value of Denote the indirect utility function gener-
a transfer is actually very simple and can clar- ated by the household’s constrained util-
ify a number of vexing issues. ity maximization by v​(​ y, b, t)​​. We assume
Assume that the utility function of a no spillovers so that no one else’s utility is
household is given by ​u​(​c1​ ​​, . . . ​cn​ ​​, b, l)​​, where affected by the benefits going to a particular
​​c​1​​, . . .​c​n​​​is a vector of consumption goods, ​b​ household.
is publicly provided benefits, and ​l​is labor The key assumption is that the social wel-
supply. The household maximizes this utility fare function is a weighted average of the
with respect to a budget constraint that is individual indirect utilities:
given by
W = ​∫0​  g​​ (​ y)​h(​ y)​v(​ y, b, t)​dy​.
​y​​  *​
n (3) ​
(2) ​​ ∑​ ​​​  c​i​​  ≤ y + wl + b − t​(wl + y − ​y​​  *​)​​,
i=1
Under these assumptions, the social ben-
where y​ ​is nonlabor income, ​w​is the wage
efit from a household with income y​​ get-
rate, b​ ​is the monetary value associated with
ting a small increase in b ​​is given by​
the government benefit, t​​is the tax/­transfer
rate, and ​​y​​  *​​is the poverty line.17 Writing util- g​(y)​h​(y)​​  __  ​  db = g​(y)​h​(y)​​  dv
dv __  ​ ​  dv/db
_____  ​ db.​
ity in this way captures a whole range of pos- db dy dv/dy
sibilities: The government benefit could be a dv/db
The ratio ____ ​​   ​​ represents the household’s
cash transfer that is conditional on an action, dv/dy
willingness to pay for ​db​, which we denote
in which case b ​ ​enters the utility function
by ​WTP​(y)​​. ​dv/dy​is the household’s mar-
negatively, or it could be a cash transfer
ginal utility of income, which we denote by
combined with a pyschosocial intervention
​MU​(y)​​  .19
in which ​b​would enter positively. The condi-
tionality is also allowed to impose some time Denote the incremental cost to the gov-
costs, which would then increase the disutil- ernment to providing a small additional
​∂​​  2​  u  ​ < 0 )​. benefit d ​b​by ​ ϕ′​(b)​db.​Then the cost of
ity of labor (i.e., ​​ ____
∂ l∂ b providing it to everyone with income y​ ​ is​
In addition, the household may face an h​(y)​ϕ′​(​ b)​.​If there is no labor supply
additional constraint on its choice, which response from a change in benefits d ​ b​, then
takes the form ​​ (​c1​ ​​, . . . ​cn​ ​​, l)​  ∈ Φ​(b)​.​ This the key statistic that determines the social
could represent a credit constraint or a con- desirability of a particular intervention under
straint that comes with the program—for these assumptions is given as follows.
example, a CCT program would require the

17 Note that we use y

​ ​to denote nonlabor income here, 18 While the government can vary ​b​by income level,
whereas in section 2.2.1 we used ​y​to denote total income. i.e., ​​b​i​​​, we suppress this to simplify notation.
We split out labor and nonlabor income here because we 19 Formally, this should be W ​ TP​(y, b, t)​​ and ​MU​(y, b, t)​​  ,
want to discuss how to think about changes in labor income but we suppress the dependence on b ​ ​and ​t​to simplify
in response to benefit changes explicitly. notation.
 Banerjee et al.: Social Protection in the Developing World 1369

Result 3: If the government welfare func- Implications for Evaluating Cash

tion can be written as a function of individ- Transfers.—The usual assumption with cash
ual utility functions with no spillovers, as in transfers is that the willingness to pay for ​$​1
equation (3), and if individuals are optimiz- is always ​$​1 (i.e., ​WTP​(y)​  = 1)​, though as
ing when they make their work and spending we will see, there may be exceptions to this
choices, then denoting willingness to pay for rule. The cost of paying an additional $​ ​1 is
a program from household with income ​y​by​ also a standard quantity in public economics.
WTP​(y)​​, a program is socially beneficial if What that leaves in the government’s deci-
and only if sion problem is the term that picks up the
household’s poverty, ​g​(y)​MU​(y)​​.
g​(y)​MU​(y)​WTP​(y)​
 ​  > 1​.
(4) ​​ __________________
      This has several key implications. First,
ϕ′​(​ b)​ assuming we know just how poor the house-
hold is, understanding the impacts of cash
As noted above, in developing country set- transfers on household consumption choices
tings, the labor supply reduction in response (e.g., whether a household buys food versus
to additional public benefits appears limited, a television) is not first order in understand-
so expression (4) is the key ratio from the ing a program’s welfare impacts. Second,
government’s perspective. Even if their labor from a welfare perspective, it does not really
supply does change, the utility consequences matter if cash transfers cause people to work
of this are second order by the envelope the- less unless doing so has implications for taxes
orem, so this does not change expression (4) (i.e., the transfers affect work, which in turn
as long as it does not affect government tax affects tax revenues). This is because, other
receipts. So, as long as either (a) labor sup- than the tax revenue impact on the govern-
ply does not change or (b) if it does change, ment, leisure is just another consumption
this does not affect net transfers or taxes choice.
collected by the government—a reasonable In other words, even though there is a
assumption for most of the poor in devel- large literature that focuses on the impacts
oping countries—expression (4) is the key of cash transfers on various household out-
expression of interest.20 comes—from food consumption, to health
outcomes, to whether one bought a TV—
3.1.2 Implications
from a welfare perspective, none of this
should matter if (a) the government wel-
This simple observation that the marginal
fare function can be written as a function
benefits from a government intervention are
g​(y)​MU​(y)​WTP​(y)​
of individual household utility functions
______________
given by   
​​   ​​ has a number of and (b) there are no spillovers to other
​ϕ′ ​​(b)​
households.
important implications for evaluation. However, understanding a program’s tar-
geting—which we discussed in section 2—is
20 If, however, labor supply changes and this affects first order since what matters is the degree
taxes or transfers, then we do need to modify expression to which cash transfers are redistributed to
(4) to account for this. Denote the change in labor sup- the poor. What also matters is the program’s
dl
ply as ​​ _ ​.​In this case, the cost of providing the benefit
db governance and administration (discussed
in the denominator becomes ( db )
​​ ϕ′​(b)​  − tw ​ _
dl
​ ​db.​ This is below), that is, whether the poor actually
then akin to the marginal value of public funds (MVPF) got their full cash transfer or whether some
discussed by Hendren and Sprung-Keyser (2020), but
augmented by g​(​ y)​MU​(y)​​to take distributional conse- got lost to corruption. But how beneficiaries
quences into account. choose to spend $​​1 of benefits—on nicer
1370 Journal of Economic Literature, Vol. LXII (December 2024)

food or a new television—is first-order not 3.1.3 Limitations to This Approach:

welfare relevant. Wedges, Spillovers, and Optimization
In short, the theory implies that while it Frictions
may be academically interesting to under-
stand how cash benefits affect consumption The willingness-to-pay approach is a
choices and work, and while these factors benchmark with three important assump-
may be important to the political sustainabil- tions (stated in result 3) for this logic to go
ity and messaging of these programs, from through: (a) that the government welfare
a pure welfare perspective this matters less. function can be written as a function of indi-
Unless, of course, the conditions outlined vidual utility functions, (b) that there are no
above fail—which they often do, as we dis- spillovers to other households, and (c) that
cuss more in detail. individuals are sophisticated about their own
preferences when they make their work and
Implications for Programs Other Than spending choices. We discuss each of these
Cash Transfers.—For programs other than in turn.
cash transfers, the basic framework is the
same, except that one needs to measure Wedges between Individual Utilities and
willingness to pay for the program rather Social Utility.—One key assumption in the
than assuming it is equal to 1. In situations framework is that the social planner max-
where one needs to estimate the willing- imizes the sum of each household’s utility,
ness to pay, it may be possible to do so using as given by equation (3). There are two
quasi-experimental program variation. For
­ important ways in which this assumption
example, in the United States, Finkelstein, may be an important oversimplification,
Hendren, and Shepard (2019) exploit nat- and where there may in fact be wedges
ural variation in the co-pay price of subsi- between the social planner’s utility and the
dized health insurance in Massachusetts to sum of individual utilities. First, the “house-
directly estimate an individual’s willingness hold” is not necessarily a decision-maker
to pay for the government-subsidized insur- with a well-defined utility function, and
ance, as do other studies of health insurance the household’s choices may not reflect the
demand (Thornton et al. 2010, Asuming utility of the individual household mem-
2013, Banerjee et al. 2021). Alternatively, bers. For example, husbands and wives may
one can experimentally measure the willing- disagree about how to spend money and
ness to pay for non-pure cash social protec- make other choices, and this may or may
tion programs. For example, one could offer not aggregate into a well-defined utility
some households the choice between the function (Browning and Chiappori 1998).
program and cash and then vary the rate at And even if a household where the husband
which households trade off cash versus the makes all decisions while ignoring his wife’s
program. If one cannot e­ stimate ­willingness preferences has a well-defined utility func-
to pay directly, one can try to impute it.21 tion, the social planner may not choose to
put that particular utility function in the
social ­welfare function. Instead, it may base
21 For example, Hendren and Sprung-Keyser (2020) its decisions about the choice of benefits on
give examples of how to impute willingness to pay for some other preferences that, for example,
133 programs in the United States, arguing that the will- give equal weight to both spouses. In that
ingness to pay for a program is equal to the change in
net present value of income (NPV) from receiving the case, it will need to look beyond the house-
program. hold’s willingness to pay.
 Banerjee et al.: Social Protection in the Developing World 1371

Second, the social planner may have in welfare due to the price changes. Note
explicit preferences over types of consump- that price changes affect everyone—benefi-
tion. That is, the social welfare function ciaries and non-beneficiaries alike—so both
may include the consumption choices c​​ need to be accounted for.
directly as an argument, above and beyond
how they affect ​u(​ c, b)​​. For example, voters Does Measured Willingness to Pay
often seem to prefer that welfare recipients Capture True Household Valuations?—The
do not spend cash assistance on “tempta- third potential limitation of this framework
tion goods” such as alcohol or cigarettes is the failure of the sophistication assump-
(Currie and Gahvari 2008). More generally, tion, in which case a household’s choices
one could imagine that voters would prefer may not reflect the true utility value of
a household receiving cash assistance to buy potential decisions, and hence, measured
food rather than a TV, even if a TV generates willingness to pay may not be an accu-
a lot of utility. Similarly, some voters seem rate measure of its true utility gain from a
to prefer that those who receive assistance choice. This failure can take many forms.
work. Therefore, while we adopt the view For example, one important constraint
that households should be free to choose is information: Households may not fully
how to spend their transfers as a benchmark, understand the benefit or the value that
we revisit this issue in which policymakers or they will receive from it, and so they may
voters may have particular preferences on make mistakes in forecasting their willing-
beneficiaries’ consumption decisions as we ness to pay (Banerjee, Duflo, and Hornbeck
discuss various program types and designs 2018). In fact, some households may not
below. understand a benefit’s value until they actu-
ally experience it, so their willingness to pay
Spillovers.—A second issue is spillovers. before receiving the benefit may not reflect
Households that receive benefits may their willingness to pay after.
spend them, creating Keynesian multipliers Behavioral frictions may also affect will-
that can have positive spillovers onto other ingness to pay. For example, if benefits have
households (Sadoulet, de Janvry, and Davis endowment effects, it would imply that one’s
2001; Angelucci and De Giorgi 2009; Egger willingness to pay for a benefit would be
et al. 2022). In principle, the theoretical higher once one actually has it than before
framework above can incorporate spillovers one has it (see, for example, Carney et al.
if one is careful to measure them. In the 2022). Time-inconsistent preferences are
social planners’ problem above, one can con- another type of behavioral friction (we return
sider the impact of a policy change b ​ ​on all to this in more detail later).
households, not just on those that receive the A third friction comes from credit con-
benefits. However, in the presence of such straints and lumpy investment opportuni-
spillovers, measuring willingness to pay by ties, which can also affect whether measured
beneficiaries alone will be insufficient. willingness to pay captures the true value to
A related issue is whether the transfers the household. Consider a lumpy invest-
are large enough to cause prices to rise ment with a co-payment of ​$1​ 0 that would
(Cunha, De Giorgi, and Jayachandran 2019; pay a return of ​$​100 tomorrow. The value to
Egger et al. 2022). If so, the right conceptual the household of this investment (above the
approach is to add the equivalent variation​ co-payment) should properly be $​​90. But
EV​to the direct benefits given b ​ ​, since the a household that only has $​ ​5 on hand and
equivalent variation captures the difference that cannot borrow will only have a stated
1372 Journal of Economic Literature, Vol. LXII (December 2024)

willingness to pay of ​$​5. This is what one in section 3.2.1. We then examine different
would find if one elicited willingness to pay program types and features. In so doing, we
directly.22 focus on whether and how different programs
A more subtle issue arises if the poor types and designs affect welfare in the senses
face obstacles navigating the bureaucratic discussed in the previous section: affecting
process of obtaining benefits. For example, future income above and beyond the transfer
households may be illiterate, or just intim- itself (e.g., by raising wages, or allowing the
idated by government bureaucracy (Gupta household to overcome credit constraints);
2017). They may highly value the benefits, affecting future taxpaying decisions; chang-
but simply not be able to take them up due ing intra-household allocations (e.g., between
to these frictions. spouses, or between parents and children);
In short, one needs to be careful to think changing patterns of consumption in ways
about household decision-making when ana- a social planner may care about, even if the
lyzing the impact of government programs. household does not; spillovers to other house-
For many programs such as cash transfers, holds, both pecuniary and otherwise; and
the value of the program is simply the value affecting welfare by allowing households to
of the transfer multiplied by the marginal overcome information constraints, credit con-
utility of income for those receiving it—so straints, or other frictions.
understanding targeting of benefits is often Note that we are not attempting a sys-
more important than carefully measuring tematic meta-analysis of all papers; we refer
which consumption choices respond. For interested readers to several recent reviews,
programs other than cash, measuring the such as Bastagli et al. (2016) and Ralston,
willingness to pay for the program, again Andrews, and Hsiao (2017). Instead, we aim
multiplied by the marginal utility of income, to capture key themes from the literature, to
provides a useful benchmark. That said, one relate these themes to the framework above,
also needs to pay close attention to impacts and to highlight where more evidence is
on within-household allocation, the poten- needed.
tial for spillovers, and the degree to which
3.2.1 Welfare Effects of Unconditional Cash
optimization frictions can artificially lower
Transfers
observed willingness to pay.
We begin by considering the simplest pro-
3.2 Empirics: Evidence on Design Choices
gram: unconditional cash transfers (UCTs).
for Transfer Programs
The benchmark program would be a transfer
In the sections that follow, we review that, once allocated, is completely uncondi-
the evidence on a wide range of social pro- tional—that is, a household’s ability to receive
tection programs with this framework in the transfer does not depend on any action.
mind. We start with the simplest program As described above, from a welfare perspec-
­conceptually—unconditional cash transfers— tive, the basic question of an unconditional
cash transfer program is who receives the
benefits. This is because a household’s will-
22 One could in principle design experiments to esti-
ingness to pay for ​$​1 in transfers should be​
mate willingness to pay where the credit constraint does
not bind—for example, by enrolling them in a lottery with $​1 (subject to the caveats above), with the
a small chance of winning enough to cover any reasonable social value of these transfers inflated by a
value of their willingness to pay and asking them to com- measure of ​g(​ y)​MU​(y)​​. To first order, the
mit to a willingness to pay conditional on winning the lot-
tery—but this is not the usual approach taken to measure key parameter for evaluating the impact of
willingness to pay. unconditional cash transfers is therefore the
 Banerjee et al.: Social Protection in the Developing World 1373

0.8
0.75
0.7 Year
0.65 1980
0.6 1990
Percentage of countries

0.55 2000
0.5 2010
0.45 2020
0.4
0.35
0.3
0.25
0.2
0.15
0.1
0.05
0
Low Lower−middle Upper−middle High
<$1,046 $1,046−$4,095 $4,096−$12,695 >$12,695
Income classification (based on GDP in constant US$2015)

Figure 2. Percentage of Countries That Enacted UCTs Covering at Least 1% of the Population by Income
Classification, over Time

Notes: This figure graphs the percentage of countries that enacted UCTs covering at least 1 percent of the
population by each year considered against income classification, following the World Bank’s 2021 income
classification thresholds as measured by GDP per capita (in constant US$2015). We define a UCT as a gov-
ernment-implemented, large-scale program that enforces no conditions on recipients and is means tested.
Source: Income classification, GDP per capita, and population data is from World Bank (2021a, b, c).

targeting of the transfers, which affects g​ (​ y)​ it in comparison to other programs, such as
MU​(y)​​. Baird, McIntosh, and Özler’s (2011) study in
Figure 2 graphs the percentage of coun- Malawi. There are also many nonrandom-
tries at different income levels with UCTs ized studies; Bastagli et al. (2016) provide a
covering at least 1 percent of the popula- systematic review of 201 cash transfer stud-
tion for countries at different levels of real ies (25 percent of which are UCTs).23
constant per capita GDP. The figure shows In general, these programs find substan-
that at the beginning of the period we study tial evidence that cash transfers improve
(1980), such programs were mostly found in different aspects of the welfare of recipi-
the wealthiest countries. By the end of the ent households. For example, Haushofer
period, UCTs had become much more com- and Shapiro (2018) document that recipients
mon and almost equally likely to be found in of cash transfers in Kenya report increases
rich or poor countries alike. in self-reported psychological well-being and
Several randomized studies directly
evaluate the impacts of UCTs, such as
23 An early and important UCT program was The South
the Haushofer and Shapiro (2016, 2018)
African Child Support grant (see, for example, Agüero,
and Egger et al. (2022) studies of the Carter, and Woolard 2006; Case, Hosegood, and Lund
GiveDirectly program in Kenya, or consider 2005).
1374 Journal of Economic Literature, Vol. LXII (December 2024)

increases in food security, assets, and nondu- of intra-household issues (e.g., an addict may
rable expenditures. Haushofer and Shapiro grab the transfer to fund his or her addic-
(2016) show that many of these effects persist tion), or simply because voters disapprove of
even several years after the transfers ended. these goods. Evans and Popova (2016) exam-
More generally, the systematic review from ine 19 cash transfer studies (a mix of UCTs
Bastagli et al. (2016) finds that most cash and CCTs), however, and find no evidence
transfer programs lead to increases in both that cash transfers increase expenditures on
overall expenditures and food expenditures. these goods.

Effects on Adult Labor Supply.—As dis- Effects on Investment and Future

cussed above, changes in labor supply them- Incomes.—The value of the cash transfer
selves are not welfare relevant under the may understate its true welfare effects in
assumptions in result 3: The labor/leisure the presence of credit constraints and lumpy
choice can be considered one of many con- investment opportunities (though if the
sumption choices, and a household choosing household understands them, they may still
to work less in response to a transfer can be be reflected in a properly measured willing-
viewed in a social welfare sense as no differ- ness to pay). In the presence of these credit
ent from choosing to buy new clothing or a constraints and lumpy investment opportu-
better roof. Nevertheless, there is substan- nities, one cannot use the envelope theorem
tial interest in this question from a policy to argue that future income increases are
perspective, particularly around the political second order. Instead, one should properly
support of these programs. However, the count any future income increases in addi-
evidence overwhelmingly suggests that cash tion. Given the ample evidence that the
transfers—at least how they are designed poor are credit constrained (e.g., Banerjee
in developing country settings—do not and Duflo 2010) and the plausibility of fixed
have measurable impacts on labor supply costs in investment (e.g., needing to buy a
(Banerjee et al. 2017, Handa et al. 2018; see machine), these effects may be substantial.24
also the discussion in section 2.1). Indeed, there are a number of studies that
suggest that households use cash transfers
Effects on Consumption of “Temptation to make future lumpy investments. Gertler,
Goods”.—A common concern about cash Martinez, and Rubio-Codina (2012) study
transfers is that people will buy “tempta- the Oportunidades cash transfer program
tion goods” such as alcohol and tobacco. in Mexico and find that households con-
However, the same arguments that apply sume 74 percent of the transfer received and
to labor supply apply here: To the extent invest the rest.25 They then compare house-
that governments’ utility function is a func- holds who were treated 18 months earlier,
tion of individual household utilities and
households are rationally choosing to con- 24 Another reason that cash transfers may increase risky
sume more of these goods, this should not investment making is if transfers make households less risk-
matter. Nonetheless, voters appear to be averse, which would be true if households have declining
less enthusiastic about funding transfer absolute risk aversion. See the related discussion in Janzen,
Carter, and Ikegami (2021). If there are no non-convexities
programs that would lead to increased con- in investment, however, then this effect is second order by
sumption of these goods, either because of the envelope theorem, and so we should not additionally
fears that these choices reflect short-term value investment that occurs for this reason above and
beyond the value of the cash transfer.
temptations rather than rational decisions 25 Oportunidades is a conditional cash transfer, as are sev-
(Banerjee and Mullainathan 2010), because eral other papers that we discuss in this section. Here, we
 Banerjee et al.: Social Protection in the Developing World 1375

rather than later, and find that those treated considering the net welfare effects of these
earlier have higher incomes, even 4 years programs. These happen through several dif-
later. They interpret this increase as a return ferent mechanisms.
on the extra 18 months’ worth of investments
these early-treated households had, com- Informal Insurance: One type of
pared to the later-treated households. spillover effect is through informal insur-
How should this sort of investment be val- ance arrangements in the spirit of Townsend
ued from a welfare perspective? Consider (1994) and Ligon, Thomas, and Worrall
the Gertler, Martinez, and Rubio-Codina (2002). If households are implicitly insur-
(2012) case. The envelope theorem approach ing one another, this could mean that if one
would be to value a transfer of $1 at $1. If household receives a transfer but another
so, the fact that the investments increase household does not, they may share the
future incomes is already included in the proceeds of the transfer to some extent.
$1, by the envelope theorem (i.e., house- Angelucci and De Giorgi (2009), for exam-
holds were already optimizing and the solu- ple, document these kinds of spillover effects
tions are continuous in the parameters), and in Mexico’s PROGRESA program. In con-
so one should not also include the investment trast, Evans and Kosec (2023) show that in
income, as this would be double-counting. In Tanzania’s CCT, while beneficiary house-
the presence of credit constraints and lumpy holds were more likely to have someone with
investments, however, one should instead a personal problem turn to them for assis-
value the initial consumption (e.g., $0.74 out tance, they were no more likely to actually
of $1), and add to it the discounted future provide assistance to others.
increases in consumption from those invest-
ments. In the Gertler, Martinez, and Rubio- Prices: A second channel through which
Codina (2012) case, they estimate that each cash transfer programs can affect the welfare
$1 transferred leads to a $0.016 higher con- of non-beneficiaries is through price changes.
sumption per month starting 12 months However, since those receiving benefits tend
after the transfer. The net welfare effect is to spend the benefits on a wide variety of
therefore given by ​0.74 + PDV​ (​0.016​ per items, the demand shock for any given item
month). With a 10 percent annual discount is likely to be small. This, combined with the
rate, for example, and assuming the effects fact that supply of those items may be at least
started 1 year after the transfer and lasted sometimes elastic suggests that, in practice,
until year 10, the net welfare effect of this price effects for pure cash transfers may
transfer would be $1.8 instead of $1. While be small. Indeed, several studies, such as
the precise numbers depend on the assump- Cunha, De Giorgi, and Jayachandran (2019)
tions, the point is that if one considers these and Egger et al. (2022) find either zero or
effects separately from a welfare perspective, economically very small impacts of general-
investment effects can make a big difference. ized cash transfers on prices.26
One important counterexample is Filmer
Spillovers.—Cash transfers have been et al. (2023), who study a cash transfer pro-
shown to have several important types of gram in the Philippines in which many, but
spillovers on other, non-recipient house-
holds, which are important to include when
26 By contrast, Cunha, De Giorgi, and Jayachandran
(2019) and Banerjee et al. (2023) do find price impacts of
discuss the cash component of these programs; we discuss in-kind transfers, which are a much larger supply shock to
conditionality in section 3.2.2. a small set of commodities; see section 3.2.2.
1376 Journal of Economic Literature, Vol. LXII (December 2024)

not all, households in treated areas received only part of the effect, as they do not include
benefits; on average 65 percent of households the reduction in demand in the areas where
were treated. As a result of this high satura- taxes would be increased to finance such a
tion, the program raised aggregate incomes scheme. Also, note that in most models,
in treated villages by about 15 ­percent. They long-run outcomes are determined by sup-
show that this led the prices of protein-rich ply and not demand, so it is unclear that the
perishable foods (e.g., eggs and fish) to experiment is providing a good estimate of
increase by 6–8 percent, with effects con- what the long-run will look like.
centrated in villages where the share treated Gerard, Naritomi, and Silva (2021) study
was high and in remote areas where supply related questions in their study of the labor
was less elastic. They then show that ineligi- market effects of a large-scale expansion
ble children in treated villages demonstrated of a CCT program, Bolsa Família. Unlike
increased rates of stunting, which they argue other programs, Bolsa Família is explicitly
is consistent with declines in protein con- means tested (i.e., those with formal sector
sumption for non-beneficiaries associated income above a certain level are ineligible),
with the higher prices. so one may be particularly concerned about
the labor market effects of such a program.
Demand Effects: A third source of spill- However, consistent with the presence of
overs is through Keynesian-type demand mul- Keynesian multipliers, they find that expan-
tipliers. That is, transfers may be spent locally, sions in Bolsa Família substantially increased
boosting incomes of those who provide goods labor market participation, even among
or services to beneficiaries, who then spend workers who were never part of the program.
some of their income locally, and so on. Since they do not observe consumption nor
Egger et al. (2022) use the fact that the everyone’s income, they do not formally esti-
transfer program they study was random- mate a Keynesian multiplier, but the results
ized at the village level to estimate a village are consistent with cash transfer programs
level “fiscal multiplier” of this sort. They having strong local multipliers.
find that each dollar transferred leads to a
$2.5–2.8 total increase in local Other Considerations.—In addition to
incomes in those villages due to this local these effects, UCT programs may have
multiplier effect. This suggests that these important welfare effects within households
types of demand-side spillovers may be that are not ­captured by this simple frame-
important in local economies, at least in the work. For example, these programs may
context of the large (US$1,000), one-time affect the education and health of children
transfers that they study. (which may or may not be fully counted in
Note, however, that since transfers are a household utility function, depending on
paid for externally, this estimate does not whether parents are completely altruis-
capture the net general equilibrium “mul- tic towards their children) and the relative
tiplier” if one were to impose such a policy bargaining weights and outcomes among
nationally. That is, if the transfers were paid spouses. We return to these issues in sections
for by increased taxes, the taxes also could 3.2.2 and 3.2.5.
have both direct and multiplier effects on
3.2.2 Conditional Cash Transfers and
economic activity. From the context of the
Intergenerational Investments
full general equilibrium effects of a domes-
tically financed program, the Egger et al. Perhaps the single largest innovation in
(2022) results therefore may be capturing social protection programs in the developing
 Banerjee et al.: Social Protection in the Developing World 1377

world in the past 30 years has been the intro- the human capital investments above the
duction of CCTs. These programs provide a degree to which they are valued by house-
regular stream of cash transfers as in many holds. To be precise, consider an action ​D​
UCTs, but condition the transfer on the that the planner would like the household
household fulfilling a set of criteria, usually to take. Suppose the household’s utility if it
related to human capital investments for kids. takes the requisite action is ​u(​ c)​  + aD​, with​
In appendix table 1, we present the condi- 0 ≤ a < 1​, but the planner’s utility from a
tions of 67 CCTs from 45 countries. The most household taking the action D ​ ​is ​u​(c)​  + D​  .
common conditions are school enrollment Suppose further that that the size of the
and attendance (52 CCTs), health check- benefit is B ​ ​, that taking the action has a con-
ups (38), complete vaccination schedule sumption cost to the household ​F​, and that
(22), and attendance of training sessions or the income effect of the benefit wouldn’t by
workshops (20). itself cause it to take the action. The benefit
CCTs were pioneered in the 1990s with to the household of getting the assistance (in
Brazil’s Bolsa Família, Mexico’s PROGRESA, the case where it chooses the action) is u ​​ ′ ​(​ y)​​
and Bangladesh’s Female Secondary School [B − F]​  + aD​. However, the benefit to the
Assistance Project being early examples. planner is ​​u′ ​(​ y)[​​ B − F]​  + D​. So, what we’d
Since then, over 60 countries have started want to count is the net benefit to the house-
their own CCTs (World Bank 2018). Figure 3 hold—​​u′ ​​(y)[​​ B − F]​  + aD​ —plus the unin-
shows the percentage of countries with CCT ternalized piece from the benefits that the
programs (i.e., covering at least 1 percent of planner values but the households don’t—
the population), by quartile of GDP per cap- that is, ​​(1 − a)​D​.27
ita (in constant US$(2015)). CCTs went from Mexico’s PROGRESA program was eval-
being essentially ­nonexistent in 1980 to exist- uated using a phase-in design. Surveys were
ing in between 10 and 35 percent of coun- conducted in 320 treated municipalities as
tries in the bottom three GDP per capita well as 186 control ones, which received the
quartiles by 2020. Unlike with UCTs shown program about 18 months later (Skoufias
­
in figure 2, these programs are less common 2005). This phase-in design allows research-
in the wealthiest countries. ers to understand if a CCT led to human
The theory behind these programs is that capital investments in health (Gertler 2004),
the cash transfers provide assistance to the education (Schultz 2004), and other out-
current generation, while a combination of comes (Skoufias 2005). When CCTs were
the cash payments and the requirements for instituted in other countries, a number of
human capital incentives aim to break the them were experimentally evaluated during
intergenerational cycle of poverty. Implicit early stages.28
in these requirements is a statement about
welfare: that the government values child
27 We thank David Romer for suggesting this
human capital investments more than the
formulation.
household decision-maker would, absent the 28 There are now randomized evaluations of CCT pro-
incentives (if not, then there would be no grams in Colombia (Barrera-Osorio, Linden, and Saavedra
need for incentives). 2019), Honduras (Benedetti, Ibarrarán, and Mcewan
2016; Galiani and McEwan 2013; Glewwe and Olinto
Thus, from the welfare perspective out- 2004; Morris et al. 2004), Indonesia (Alatas 2011, Cahyadi
lined above, evaluating these programs et al. 2020), Nicaragua (Barham and Maluccio 2009;
requires not only measuring the immediate Macours, Schady, and Vakis 2012; Macours and Vakis
2014), the Philippines (Filmer et al. 2023, Kandpal et al.
cash transfer value to the household, but 2016) and Tanzania (Evans et al. 2014; Evans, Holtemeyer,
also separately valuing the impacts from and Kosec 2019).
1378 Journal of Economic Literature, Vol. LXII (December 2024)

0.8
0.75
0.7 Year
0.65 1980
0.6 1990
Percentage of countries

0.55 2000
0.5 2010
0.45 2020
0.4
0.35
0.3
0.25
0.2
0.15
0.1
0.05
0
Low Lower−middle Upper−middle High
<$1,046 $1,046−$4,095 $4,096−$12,695 >$12,695
Income classification (based on GDP in constant US$2015)

Figure 3. Percentage of Countries That Enacted CCTs Covering at Least 1% of the Population by Income
Classification over Time

Notes: This figure graphs the percentage of countries that enacted CCTs covering at least 1 percent of the
population, by each year considered (1980, 1990, 2000, 2020) against income classification, following the
World Bank’s 2021 income classification thresholds, as measured by GDP per capita (in constant US$2015).
We define CCT as a government-implemented, large-scale program that enforces health or education-related
conditions on recipients.
Source: Income classification, GDP per capita, and population data is from World Bank (2021a, b, c).

These evaluations often find impacts of 2011; Schady and Araujo 2006; Benedetti,
CCTs on child health and schooling. In Ibarrarán, and Mcewan 2016).
the short run, there is evidence that these
programs lead to an increase in height Long-Run Impacts of Conditional Cash
and a reduction in stunting (Gertler 2004, Transfers.—Several studies suggest that
Attanasio et al. 2005, Kandpal et al. 2016); these human capital benefits may persist.
increases in elementary school enrollment Two studies, for example, examine exper-
(Schultz 2004, Todd and Winters 2011, imental variation. Cahyadi et al. (2020)
Baird et al. 2014); and increases in cognitive measure the impacts of a CCT program in
outcomes for children (Macours, Schady, Indonesia after six years, finding persistent
and Vakis 2012).29 They also show reduc- human capital gains: 6 years after the start
tions in child labor (Barrera-Osorio et al. of the program, truancy and child labor
fell by half, and stunting—which requires
29 Parker and Todd (2017) provide an in-depth discus- cumulative investments in health—fell by
sion of the many studies of the PROGRESA program. 23 ­percent. Similarly, Millán et al. (2020)
 Banerjee et al.: Social Protection in the Developing World 1379

examine the CCT in Honduras, which ran examine outcomes at 19–22 years old. They
for 5 years in treatment municipalities but find that girls with early exposure to the pro-
never in the control areas. Eight years after gram had higher educational attainment,
the program ended, they find large increases higher earnings, and lower fertility com-
in education—particularly secondary and pared to girls with late treatment. In Barham,
university enrollment—for non-Indige- Macours, and Maluccio (2018b), the authors
nous groups, but more muted effects for examine similar-aged boys and find similar
Indigenous groups. They also find that the education and labor market effects.
CCT increased international migration for More recently, Araujo and Macours (2021)
men who experienced the program as older use this same design to follow the experi-
cohorts. They find little robust evidence on mental PROGRESA cohorts about 20 years
wages.30 later and find that educational attainment
A second empirical strategy uses the increases. For the children who were school
fact that some programs started earlier in age when the program started, labor income
some places than others. Behrman, Parker, is higher, particularly for the top of the
and Todd (2011) use this strategy, compar- income distribution. Those in the early treat-
ing those who received an “extra” 18 months ment group were more likely to migrate, par-
of Mexico’s PROGRESA program, after 5.5 ticularly to the United States, which could in
years of exposure in treatment areas, as well part account for the higher incomes.
as comparing these areas to additional nev- A third empirical strategy is to use the
er-treated areas using a matching design. general phase-in of programs, coupled with
They find increases in schooling attainment the differential ages of children when the
for those who received the program earlier. CCT expands to their municipality. The
Barham, Macours, and Maluccio (2013, variation in when the CCT enters gener-
2018a, 2018b) use a related design in ates a difference in difference in total expo-
Nicaragua, using a randomized design and sure to the CCT while controlling for age.
comparing early treatment locations with late Parker and Vogl (2023) use this strategy to
treatment locations. In Barham, Macours, analyze PROGRESA. They find that expo-
and Maluccio (2013), the authors focus sure to Mexico’s PROGRESA in primary
on boys and compare the impact of being school leads to higher educational attain-
exposed to a CCT in utero through age two ment, greater labor force participation, more
with being exposed at ages two–five. They migration, and higher earnings for women.
find that boys exposed early in life had bet- They also find modest impacts on education
ter cognitive outcomes, with no impacts on and migration for men.
stunting. In Barham, Macours, and Maluccio
(2018a), they compare girls treated at ages Does Conditionality Matter? Can It Be
9–12 with those treated at ages 11–14, and a Two-Edged Sword?—These human cap-
ital improvements provide evidence that
30 The idea that nutritional investments can have long- CCTs can change household behavior. But
run effects can be seen in the long-run follow-up study perhaps poor households, when faced with
of protein supplementation done starting in 1969 in an infusion of cash, would have increased
Guatemala, which reduced stunting for 0–3 year olds by
more than 50 percent. The study found that, in 2002–04 investments in human capital due to income
(so more than 30 years later), wages were 46 percent higher effects even without the incentives. Do
for people who were ages 0–3 in the two villages random- the conditions themselves matter? Several
ized to receive protein supplements, compared to compa-
rable people from the two control villages (Hoddinott et al. studies have found evidence that house-
2008). holds respond to conditionality per se on
1380 Journal of Economic Literature, Vol. LXII (December 2024)

the targeted outcomes. Studying random denying cash transfers to noncompliers who
variation in the warnings for violating the could benefit from the program. Ultimately,
conditions on school attendance in Brazil’s how one balances these trade-offs in choos-
Bolsa Família program, Brollo, Kaufmann ing the program design depends on the pol-
and La Ferrara (2017) find that child atten- icy objective the social planner wishes to
dance increases in response to the threat of achieve.
penalties. Similarly in Nicaragua, Macours, Combined, these studies suggest that
Schady, and Vakis (2012) argue that a treat- conditions may matter for child human
ment arm that gave additional cash over capital outcomes, but that there may be
the basic CCT did not lead to additional a downside from cutting off assistance to
improvements in child outcomes, suggesting noncompliers. This presents the intrigu-
that the conditions drove the effects in the ing possibility that a labeled transfer pro-
basic treatment. gram—where households are told that there
The CCT conditions could also help those are conditions, but they are not rigorously
children who are more disadvantaged within enforced—could achieve a similar impact as
families: Akresh, de Walque, and Kazianga conditionality without its downsides. Such
(2013) conduct a randomized trial in Burkina an approach could also reduce monitoring
Faso, where villages were randomized into costs. Benhassine et al. (2015) suggest this
typical UCTs or CCTs. For boys and older may be the case: In Morocco, they com-
children, who tend to be prioritized by pare a labeled cash transfer with a CCT and
parents, both programs led to increases in find that the labeled transfer substantially
enrollments. However, they find that the reduced the drop-out rate and increased
CCT led to large increases in enrollments for school enrollment among those who had
girls and younger children, whom they iden- dropped out at baseline, whereas the tradi-
tify as more “marginal” for education, while tional conditional cash transfer had smaller
the UCT had no impacts on enrollment for positive effects. However, more research
these groups. is needed on labeling to better understand
Could there be downsides to condition- when it works, how long it persists, and how
ality? Baird, Mcintosh, and Özler (2011) to structure the labels.
­randomize communities in Malawi to a UCT,
3.2.3 In-Kind versus Cash Transfers versus
a CCT (based on school attendance), or a
Vouchers
control group. The results show that condi-
tionality matters, but also entails risks. On Conceptual Issues.—An alternative
the one hand, the CCT conditions improved approach to delivering assistance is to provide
targeted indicators: While both the UCT and households with goods or services directly,
CCT led to a reduction in drop-out rates for rather than cash. There are two ways to do
the adolescent girls, the reduction was more so. The first is to physically transfer a good
than twice as large in the CCT. Likewise, the or service (e.g., hand people loaves of bread,
CCT led to increases in reading comprehen- distribute bed nets, etc). A second way is to
sion, while the UCT did not. On the other give people a voucher to purchase a specific
hand, those in the UCT arm experienced good or service (e.g., give people a debit card
larger impacts in terms of delaying marriage to use to purchase rice, give people a cou-
and reducing pregnancy rates, particularly pon to pick up a bed net at a clinic). Both
among adolescent girls who dropped out types of in-kind programs create frictions
of school. This suggests that the success of that can influence recipients’ consumption
the conditionality could come at the cost of decisions, but they potentially differ in terms
 Banerjee et al.: Social Protection in the Developing World 1381

of how they affect the market for the goods in that area. This means that, if supply is
in question. inelastic, prices for the subsidized goods may
There are reasons why cash could be pre- be lower in some kinds of in-kind programs
ferred to in-kind transfers; after all, house- than in a cash program (Coate, Johnson,
holds can use cash to purchase whatever and Zeckhauser 1994; Basu 1996). Second,
will increase their utility most. Moreover, governments typically set in-kind benefits
for households that would purchase more in terms of quantities (i.e., 5 kg of rice per
of the in-kind good than is provided by month), whereas they typically give cash in
the transfer anyway, providing in-kind specific amounts each month or quarter (i.e.,
goods should be the same as cash, since 500 rupees per month). Setting benefits in
households can just reallocate their cash terms of quantities rather than prices can
elsewhere. provide implicit price insurance to beneficia-
However, theory provides some reasons ries (Gadenne et al. 2024). While the govern-
why policymakers may prefer in-kind. The ment could in principle adjust the quantity
first has to do with specific social prefer- of in-kind benefits based on prices to hold
ences over types of consumption discussed value constant, or conversely adjust the value
in section 3.1.2. If, for example, the social of vouchers or cash payments to keep their
decision-maker cares explicitly about certain real value constant, these adjustments may
types of “merit goods,” such as nutrition or not happen.
health care, rather than just household util- Third, in-kind transfers could potentially
ity, the social decision-maker may explicitly improve self-targeting if they increase partic-
want to influence beneficiaries’ consumption ipation costs more for the wealthy (Nichols
baskets. When households are constrained— and Zeckhauser 1982). For example, if the
so they are given more of the in-kind good government provides poor-quality, subsi-
than they would consume otherwise, and dized bread, richer households may not want
cannot easily resell it—in-kind transfers it.
can encourage consumption of these merit Finally, there are administrative differ-
goods. This strategy can also work if house- ences between these programs. In direct
holds are simply not optimizing or are using distribution programs, the infrastructure
mental accounting (Hastings and Shapiro required to deliver millions of tons of goods
2018). is different than the infrastructure required
A second reason policymakers may prefer to run a voucher system using electronic
in-kind transfers has to do with how these debit cards (Banerjee et al. 2023) or to give
programs interact with prices. There are two out cash payments, and these administrative
issues. First, in the case where the govern- differences may really matter in low state
ment provides the goods in question (i.e., capacity settings.
provides goods directly, rather than vouch-
ers), it can increase the overall supply of the Experimental Comparisons of In-Kind
goods in question. This depends, of course, and Cash Programs.—Several studies exam-
on how goods are sourced—for example, if ine these issues by experimentally allo-
the government buys food in local markets cating areas into cash, voucher, or in-kind
and then gives it away, supply in that local distribution programs. In Mexico, for
market is unaffected. However, if the gov- example, the government ran a random-
ernment buys food far away and transports ized controlled trial (RCT) that random-
it into a particular area, there may be a pos- ized villages into receiving in-kind food
itive supply shock of the good in question transfers, approximately equivalent cash
1382 Journal of Economic Literature, Vol. LXII (December 2024)

transfers, or a control group. Several stud- which had a much wider range of products—
ies have examined this experiment. Overall, led to higher dietary diversity. For example,
Skoufias, Unar, and ­González-Cossío (2008) voucher households consumed more vegeta-
find that both types of transfers have sim- bles, eggs, and dairy than those that received
ilar effects on food consumption, total the in-kind transfer. The in-kind program
consumption, male labor supply, and pov- was also more than twice as expensive (in
erty outcomes; Cunha (2014) finds similar terms of costs per outcome achieved) as
consumption effects of in-kind and cash either the cash or the voucher program.
programs. Third, Banerjee et al. (2023) examine an
These experiments, however, reveal two experiment that compared an in-kind rice
potentially important differences between program to an electronic voucher program
cash and in-kind programs. First, Leroy et al. that allowed recipient households to purchase
(2010) find (and Cunha 2014 confirms) that an approximately equivalent amount of rice
in-kind programs lead to increases in both and eggs (in value) from private providers.
total calories and micronutrients consumed Importantly, the Banerjee et al. (2023) study
through stickiness in consumption choices. is an at-scale experiment: 105 districts, with
Second, Cunha, Giorgi, and Jayachandran a combined 3.4 million beneficiary house-
(2019) find evidence that in-kind transfers holds, were included, with both programs
cause price declines of about 5 percent in implemented as usual by the government
remote areas. They argue that, in remote bureaucracy. They find a dramatic difference:
locations, in-kind rather than cash increases Voucher programs delivered concentrated
the real value of transfers provided by assistance to targeted households, whereas
14 percent (much larger than the 5 percent in-kind aid was spread much more widely. As
price decline) because it reduces prices on a result, targeted households received 45 per-
all purchases of affected goods by everyone cent more assistance in voucher districts than
in these locations, not just those financed by in in-kind areas. For households in the bot-
the transfer program. The authors find no tom 15 percent at baseline, poverty fell by 20
price changes in more developed locations.31 percent. These results suggest an important
A second experiment comparing cash, additional dimension, namely that voucher
in-kind food, and voucher food programs was programs may be more reliably administered
done by Hidrobo et al. (2014) in Ecuador. in settings with low state capacity.
They also use a randomized design to com- These studies together present a nuanced
pare the impact and cost-effectiveness of understanding of the relative costs and ben-
the three alternatives in urban areas with efits of in-kind, voucher, and food programs.
well-functioning markets. They find similar Cash remains an important benchmark, since
effects of the three programs on both food households can choose to spend it on what
and nonfood consumption, but they find they prefer. To the extent that policymakers
differences in food composition: Relative to believe households are not privately opti-
the in-kind program, the voucher program— mizing their nutrition choices or otherwise
prefer food programs, the evidence suggests
31 This is consistent with Michelson et al. (2012), who
that food vouchers can have impacts on the
argue that rural villages have difficulty responding to set of foods that households consume. In-kind
cash transfers stemming from less market access and less programs are more costly to administer and
competition among suppliers, and Jiménez-Hernández harder to administer faithfully, but may be
and Seira (2023), who show that direct provision by gov-
ernment introduces competition, and hence is effective useful particularly in very rural, isolated areas
when there is local market power in the provision of goods. where supply is inelastic or noncompetitive.
 Banerjee et al.: Social Protection in the Developing World 1383

3.2.4 Workfare Programs employment, consistent with monopsony in

In workfare programs, the government local labor markets.
provides public employment, typically at a Similarly, Franklin et al. (2024) study a
low wage, for those who want it. As discussed workfare program that was randomly phased
in section 2.4 above, the idea is that work in across neighborhoods in Addis Ababa,
requirements can be a screening mechanism: Ethiopia. The program provided an hourly
Those who have better work options than wage that was about 64 percent higher than
those offered by the government will opt out the private market, which led households to
of the government’s workfare program. shift to public employment. The program led
In these programs, the government is to an increase in private market wages that
directly intervening in the labor market, accounted for most of the increase in wel-
which has several implications for evalu- fare of program beneficiaries. Moreover, the
ating these programs. First, in addition to authors experimentally show that workfare
differences in workers’ preferences (i.e., increased amenities in treated neighbor-
how much they prefer a workfare job com- hoods, suggesting that the investments they
pared to their alternative private sector study were productive.
job), one also needs to consider the produc-
3.2.5 Programs Targeting Specific
tivity of the work. If the work is productive
Individuals within the Household
(i.e., building public infrastructure with
the same efficiency as the private market), If the household behaves like the unitary
this type of program can be efficient; if the model predicts (i.e., the household solves its
work is unproductive (e.g., slowly build- optimization problem as a single unit) then
ing “bridges to nowhere”), then the social the decision about which household mem-
efficiency loss also needs to be accounted ber should receive transfers would be irrel-
for in any cost–benefit analysis. Second, evant: No matter who receives the income,
there may be important spillovers through the money would be spent in the same
labor market effects on wages, or they may way. However, there is reason to think that
affect migration (see, for example, Imbert the recipient matters (see, e.g., Browning
and Papp 2020a, b; Berg et al. 2018; Azam and Chiappori 1998, Robinson 2012).
2012; Bertrand et al. 2021). Indeed, this is one of the frictions that we
In an at-scale study, Muralidharan, discuss in section 3.1.2 that moves us away
Niehaus, and Sukhtankar (2023) examine from the benchmark case. If money is spent
the general equilibrium market effects of differently in the household depending on
the MGNREGA program. They conduct a who receives it, then the recipient matters
randomized experiment at the subdistrict for policy. The empirical evidence on the
level that improved program administration degree to which this matters in a real-world
(by providing biometric smart cards, which policy context is, however, mixed.
reduced leakage, reduced payment delays, A number of studies explicitly randomize
and generally made the program function whether transfers are given to men or women
better). They find that private sector wages in developing countries, finding little effect
increased. As a result, while the reform overall.32 For example, Akresh, de Walque,
raised the income of low-income households
by 13 percent, 90 ​percent​of the gain came
32 These studies build on Lundberg, Pollak, and Wales’s
from equilibrium effects on the private mar-
(1997) quasi-experimental study in the United Kingdom.
ket. They find, perhaps surprisingly, that Other important nonrandomized studies include Duflo’s
this also led to an increase in private sector (2003) study of pensions, discussed in section 4.3.2.
1384 Journal of Economic Literature, Vol. LXII (December 2024)

and Kazianga (2016) randomize whether IPV, with only two studies showing mixed or
the transfer is given to mothers or fathers in adverse overall impacts.
Burkina Faso. They find no differences on
3.2.6 Poverty Traps and Lumpy Transfers
child health or education outcomes on aver-
age. That said, they do find that giving trans- An important decision when designing a
fers to mothers leads to lower levels of child transfer program is the transfer size and fre-
labor. They also find some evidence that giv- quency. Thinking about this question in the
ing money to fathers improves outcomes in developing world context is important given
poor rainfall years and leads to more house- that extreme poverty, together with saving
hold investment in livestock, cash crops, and constraints, may be more likely to create
housing. poverty traps (Dasgupta and Ray 1986). This
Likewise, Benhassine et al. (2015) and suggests a role for lumpy transfers for some
Haushofer and Shapiro (2016), in their types of people, rather than streams of pay-
studies of cash transfers, randomize the ments to spur investments.
recipient to be men or women, and both
studies find few differences on average. Conceptual Issues.—To fix ideas, con-
Haushofer and Shapiro (2016) do, however, sider a very simple Solow-style (1956) model
find that giving transfers to women leads to where individuals save a fixed fraction of
higher reported psychological well-being income and invest the rest. The capital stock
and greater female empowerment, though in period t​​ is ​​k​t​​​. Each period an individual
they find no differences on many other produces f​​(​kt​)​​ ​​ and consumes ​​(1 − s)​  f (​ ​kt​)​​ ​​  .
dimensions (e.g., food security, health, and Next period’s capital stock is given by ​​k​t+1​​
education). = ​(1 − δ)​ ​kt​​​  + s f  ​(​kt​)​​ ​​, where ​δ​is the depre-
However, Field et al. (2021) show how ciation rate of capital. The steady state level
female empowerment that stems from con- of capital is given by setting k​ ​​ t+1​​  = ​kt​​​​, which
trol over one’s money in the form of bank yields ​f (​ k)​  / k = s / δ​.
accounts could have real effects. Studying The key is to understand the shape of
MGNREGA, they experimentally show that the production function. If f​(​ k)​​ is globally
women who were set up to receive direct weakly concave, this yields a unique steady
deposits of their transfers into their own state—that is, there is no poverty trap. But if
accounts (rather than their husbands’), as ​f (​ k)​​has convex regions (i.e., has an S shape),
well as training on account use, worked more there is the possibility of multiple steady
in both the public sector program as well as states. In particular, there is a tipping point​​
in private sector jobs. Three years later, this k​​  *​​: If capital is just below ​​k​​  *​​, the individual
treatment even shifted community norms of will dis-accumulate capital and return to a
female employment. low equilibrium, but if capital is just above​​
Finally, another serious gender-related k​​  *​​ , the individual will accumulate capital and
concern is whether cash transfers themselves, move to a higher equilibrium.
by causing discord within households over The key point here is that if there are these
spending preferences (particularly if women threshold tipping points k​​ ​​ *​​, one-time shocks
receive the program), could lead to increases to capital can have long-run implications
in intimate partner violence (IPV). Buller for households: A one-time negative shock
et al. (2018) review fourteen quantitative can lead them to a long-run poverty trap or,
and eight qualitative studies and find little conversely, a one-time positive capital infu-
evidence of this. In fact, they find that most sion can lead them to escape poverty and
studies demonstrate evidence of decreased end up at a new, higher income level. From
 Banerjee et al.: Social Protection in the Developing World 1385

a social protection perspective, whether or in Bangladesh and are now present in over
not these tipping points exist—or whether, 40 countries (Banerjee, Duflo, and Sharma
alternatively, we are in a world with a unique 2021). The program usually consists of a lumpy
steady-state—matters substantially.33 productive asset—in the form of livestock—
The idea of threshold tipping points and bundled with skills training, savings, health
poverty traps is very general, but the details education, coaching, and cash payments for a
of the transition path can differ depending period of time. The goal is to alleviate a house-
on the particular model being analyzed. In hold’s capital and skill constraints to “push”
particular, note that we are simplifying here them out of poverty. Several studies have
by assuming that savings is a fixed fraction of found positive results, including Bandiera
income, as in Solow (1956). In a more gen- et al. (2017) in Bangladesh; Banerjee et al.
eral formulation with lumpy investments but (2015) in Ethiopia, Ghana, Honduras, India,
endogenous savings, there can still be pov- Pakistan, and Peru; and Bedoya et al. (2019)
erty traps, but the dynamics near the critical in Afghanistan. In the short run (three- to
point may differ (see, e.g., Skiba 1978, Buera fouryear follow-up one to two years after the
2009, Ikegami et al. 2018), with much more program ends), there is an improvement in
rapid transitions just above and below the consumption, food security, asset holding,
tipping point than in the very simple model and savings. In the medium run (seven-year
we have sketched here. follow-up and five years after the program
ends), both individual and household out-
Empirics: Which Model Are We In?—In comes remain higher (Banerjee et al. 2016,
an RCT in Bangladesh that provided assets Bandiera et al. 2017). In the long run (10 year
to a subset of poor households, Balboni et al. follow-up), there appears to be no additional
(2022) document the existence of poverty growth, but persistent effects for those who
traps. They sort people based on their ini- received the program (Banerjee, Duflo,
tial wealth, consider the shock of the capi- and Sharma 2021). While the p ­ rograms show
tal transfer, and compare assets at baseline long-run effects, they are also expensive, and
(​​kt​​​​) with assets 4 years later (​​k​t+1​​)​. They find so whether the returns are large enough to
that there is a unique steady state in control suggest moving households out of a poverty
villages, but that transferring assets creates trap, as opposed to households just receiving
multiple steady states—so not only does the regular returns from a large transfer (as in
transfer temporarily increase incomes in the the permanent income hypothesis), depends
short run, but for some, it can tip them into a on the discount rate assumed. In India and
substantially higher steady state.34 Bangladesh, the program costs are relatively
More generally, a number of studies have low and the impacts are large, so that the net
examined big-push “graduation” programs— effect is positive for most plausible discount
called “graduation” due the idea that they can rates, but this may not be true elsewhere,
help people “graduate” from poverty. These where the discount rate may matter more.
types of programs were pioneered by BRAC Two specific questions that arise from these
papers are: (i) Can the government imple-
33 This relates to the discussion of how to value a $
ment these programs? (ii) To make the pro-
​ ​1
transfer. On the margin, the ​$​1 is valued at ​$​1, but a lump- grams cheaper and more cost-effective, can
ier transfer could be valued at a different rate given how it you scale down some program components?
will be spent and its potential returns.
34 Lybbert et al. (2004), Carter and Lybbert (2012), and
Recent experimental evidence from Botea
Banerjee, Breza et al. (2019) also provide evidence of the et al. (2021) in Zambia answers “yes” to both
existence of poverty traps. questions. First, they show that governments
1386 Journal of Economic Literature, Vol. LXII (December 2024)

can run the programs: They show increases food security, while lump-sum transfers
in consumption, assets, and mental health as increased assets. After three years, there
a result of the full government-run package. were no differences between the lump sum
But, importantly, they show similar results and monthly transfer groups. It is import-
between the group that received a complete ant to note that the monthly payments were
graduation treatment and the group that only spread over nine months and thus may
received the asset (and a savings tool) but no understate differences when compared to a
training. steady-state transfer program.
In Ghana, however, Banerjee et al. (2022) More generally, the evidence discussed
find that neither giving people just assets here suggests the possibility that there are
nor just the savings tool has a positive long- some households that are constrained, and
run impact, while the whole package does. for whom a one-time intervention could have
The Zambia and Ghana papers’ results are long-lasting effects. We note two outstand-
consistent if both the asset and the savings ing questions. First, not all households are
tool are important for program success, but likely to benefit from one-time interventions.
the training is not. While the Zambia results Which households are “trapped,” and how
suggest that perhaps some of the training can can we identify them? Second, what, exactly,
be stripped off these programs, reducing the keeps these households in the poverty trap?
cost, the Ghana results suggest that further Is it a rational story given lumpy investments,
testing may be useful in unbundling these or are they unable to save optimally, as sug-
programs. gested by, for example, Augenblick et al.
Blattman, Fiala, and Martinez (2014, (2022)?
2020) study a different type of lump-sum
3.3 Governance in the Delivery of
transfer program. Upon submitting a busi-
Assistance
ness plan, young adults in Uganda form
groups and receive a one-time grant. The A crucial challenge that developing coun-
authors find substantial increases in invest- tries face when it comes to social protection
ment, work, and income four years later. is governance, that is, making sure that bene-
However, in the nine-year long-run follow fits get delivered to the targeted individuals.
up (Blattman, Fiala, and Martinez 2020) In many countries, this is a nontrivial prob-
they see the gains dissipate as the control lem. In India’s MGNREGA workfare pro-
group catches up with them. Although both gram, for example, Muralidharan, Niehaus
groups converge in employment, earnings, and Sukhtankar (2016) find a 30.7 percent
and consumption levels, those who get the leakage rate in Andhra Pradesh in 2012.
grants have a lasting impact on durable asset In Indonesia’s subsidized rice program,
stocks and skilled work. Banerjee et al. (2018) calculate that eligi-
These evaluations suggest that lumpy ble households only receive a third of the
transfers can have long-run effects. But is intended subsidy. Camacho and Conover
it the lumpiness per se that matters, or the (2011) find cheating on the PMT formu-
complementary investments that go along las, and Niehaus et al. (2013) estimate that
with it? To test this, Haushofer and Shapiro 70 percent of ineligible households in India
(2016, 2018) experimentally compare the have below-poverty line cards and many
effects of a lump-sum cash transfer (​$​404 households report paying small bribes. Of
in PPP terms) to an equivalent cash pay- course, these are particular examples and
ment in nine-month installments. At nine researchers often study leakage where it is
months, the monthly transfers increased thought to be high, so one should not take
 Banerjee et al.: Social Protection in the Developing World 1387

these as unbiased estimates of overall leak- leakage by 41 percent, and quite impor-
age rates. And some of this “leakage” goes tantly increased program take-up by poor
to other households near the poverty line. households. A second study, Banerjee
However, these facts do suggest that mean- et al. (2023) (described in more detail in
ingful improvements to programs can be section 3.2.3) studied the conversion of
made by improving service delivery. an Indonesian food subsidy program from
Corruption risk can also distort program an in-kind program administered by local
choices, as people may prefer programs with officials to digital vouchers (a debit card)
a lower risk of corruption. For example, in a redeemable at a network of bank agents.
study in Indonesia, Kyle (2018) shows that The digital vouchers led targeted house-
the poor tend to prefer targeted social assis- holds to receive substantially more benefits,
tance programs to broad-based energy sub- which in turn bumped many poor house-
sidies, for which the benefits accrue more to holds over the poverty line. These effects
the middle class. But, in districts where local occurred even without biometric identifica-
politicians are corrupt, poor citizens tend to tion, suggesting that the digital card, rather
support fuel subsidies, for which local cor- than the biometrics, may have been import-
ruption matters less. ant in this case.
A traditional approach to improving gov- A word of caution, though: Muralidharan,
ernance often focuses on monitoring and Niehaus, and Sukhtankar (2023) show that
audits (Olken 2007). More recent evidence a reform in Jharkhand, India, that began
suggests that reforms to payment systems, to require biometric authentication for the
private sector involvement, improvements subsidized food program had the unfor-
in bureaucracy, and information provision tunate side effect of dropping many poor
can also help. We next discuss the existing people from the system who had not linked
evidence on these approaches in the social their biometric information to their food
protection context. benefits account. Banerjee et al. (2021) sim-
ilarly find that many people could not enroll
Digital Payment Systems.—In many in Indonesia’s health insurance because of
developing countries, the authentication of errors in the underlying identification data
beneficiaries and payments systems is man- systems. These studies emphasized the need
ual: There is a list of beneficiaries, some for universal access and robust administra-
local official needs to verify who receives tive data to complement reforms.
benefits, and cash is given. This makes it Finally, it is worth noting that the degree
hard for the central government to monitor to which people prefer digital systems may
whether targeted beneficiaries receive their depend on the overall level of adoption of
full payments. those systems in the country more broadly.
Several studies have examined the Berkouwer et al. (2021) show that during the
implications of digitizing this process. COVID-19 crisis, people in Kenya—where
Muralidharan, Niehaus, and Sukhtankar there is high mobile money adoption—pre-
(2016) examine the randomized rollout of ferred mobile money over electricity subsi-
“smart cards”—biometrically authenticated dies. However, it was the opposite in Ghana,
payments—for receiving benefits from where mobile money is less widespread:
India’s workfare program (MGNREGA) Here, nearly half of recipients preferred elec-
and pension programs (SSP) in Andhra tricity transfers. In fact, many were willing to
Pradesh, India. The smartcards worked: forgo high levels of subsidy in order to receive
They reduced payment delays, reduced electricity instead of mobile money.
1388 Journal of Economic Literature, Vol. LXII (December 2024)

Back-End Payment Systems.—A related villages to outsource the last mile deliv-
type of reform is linking delivery accounts ery for Indonesia’s rice subsidy program.
to back-end payment systems, which are the Privatization of delivery leads to increased
background systems that ensure the flow efficiency with no drop in quality, though the
of payments from government programs to effects are small. It is worth noting, however,
people. Banerjee, Duflo et al. (2020) study that this study was about privatization to var-
these systems in the case of MGNREGA in ious small-scale local vendors, not large com-
Bihar, which simplified the payment process panies; larger-scale privatizations remain an
between central and local governments. The open question for further study.
changes reduced program expenditure by
24 percent while maintaining the same pay- Improvements in Bureaucratic Quality.—
ments to workers. However, it is worth noting He and Wang (2017) study what happens
that these programs must be implemented when villages receive college graduate vil-
with care. For example, the reform studied lage officials, who are more educated than
by Banerjee, Duflo et al. (2020) delayed pay- typical village officials and supposedly free
ments to beneficiaries. While these payment from local interest groups. Using the stag-
systems are important, such transitions must gered timing of assignment of these officials
therefore be managed with care. to local villages, they show that in villages
with college graduate officials, more house-
Information Provision.—If “leakages” holds are registered as poor or as including
arise from local officials not following the disabled residents and, most importantly,
national government’s rules, one fix is to more poor households benefit from subsi-
provide information directly to beneficiaries dies targeted to the poor.
so that they can demand their full benefits
3.4 Summing Up
from local officials. Banerjee et al. (2018)
experimentally test this hypothesis in a food The basic welfare framework discussed
transfer program in Indonesia and find in section 3.1 and result 3 is powerful in
that simply mailing out identification cards its simplicity: It says that for most pro-
dramatically increases the subsidy eligible
­ grams, the value of $​ ​1 in cash transfers to
households receive. They find additional a household is $​ ​1, and hence what matters
benefits from posting information publicly, most is who gets the support, rather than
creating “common knowledge” and thus the details of the form the ​$1​ in transfers
increasing accountability of local officials. takes. This suggests that understanding the
targeting of programs, as described in sec-
Outsourcing Delivery.—While social pro- tion 2, is fundamental to understanding the
grams are typically run by the government, welfare performance of a wide variety of
there is scope for the private sector to inter- social protection programs. Likewise, the
vene. Theoretically, there can be efficiency governance issues outlined in section 3.3
gains from involving the private sector, are first-order welfare relevant since they
since the government can provide stron- can affect whether the $1 actually reaches
ger incentives to contractors. Nonetheless, the intended household.
if there is little competition in tenders, However, there are some circumstances
then the private sector may provide lower where this result may not hold. For exam-
quality than the government. Empirically, ple, to the extent that parents are not perfect
Banerjee, Hanna et al. (2019) examine what agents for their children, or perhaps are credit
happens when they randomly allow some constrained, one may care about the impacts
 Banerjee et al.: Social Protection in the Developing World 1389

directly on children’s human capital acquisi- The studies reviewed suggest a “possibility
tion. The large literature on CCTs, for exam- result”—poverty traps may exist and, for at
ple, suggests that these programs, by explicitly least some people, a particular program can
conditioning transfers on children going to work. However, this is by no means a pan-
school and receiving regular health checkups, acea or the right solution for everyone, and
can improve children’s human capital, with not everyone needs all the program’s compo-
the potential to help break the cycle of pov- nents. Untangling this knot and figuring out
erty. These benefits may be valuable above approaches that work for others is an import-
and beyond the value of the cash to recipi- ant area for future work.
ent households. The presence of spillovers to A fourth area is interactions with labor
other households and poverty traps can also markets. Programs that provide low-wage
change the welfare calculus. work, such as MGNREGA in India, are one
The discussion here also suggests import- approach to linking work to social protection
ant directions for additional research. First, goals. However, one can think more broadly
different programs may be appropriate about how labor market interventions are
for different types of people. For exam- part of the social protection system, such as
ple, CCTs may work well if households are a minimum wage, job training, job matching
near the margin of take-up for the incentiv- programs, et cetera. While a full review of
ized behavior; for households far from the these types of labor market policies is out-
threshold, imposing conditions that they side our scope, understanding these issues in
cannot comply with could make them worse developing countries is an important area for
off by denying them program access (Baird, research.
McIntosh, and Özler 2011). Figuring out Finally, the programs discussed in this
how to assign the right programs to the right section have primarily been conceptualized
people is an important area for future work. as responses to long-run poverty. But to the
On a related theme, if households are not extent that eligibility for these programs is
fully optimizing—for example, due to behav- dynamic, they can also help provide insur-
ioral or cognitive frictions, perhaps made ance against shocks. We tackle the question
worse by poverty (Mullainathan and Shafir of how to provide insurance more systemati-
2013; Dean, Schilbach, and Schofield 2017), cally in the next section.
design elements of the program may be
first-order welfare-relevant even if they are
4. Mitigating Risks
not reflected in households’ willingness to
pay. Understanding how to leverage behav- Poverty is not a static concept: someone
ioral nudges to best improve program out- may belong to the middle class one day, but
comes is thus important. a loss of income due to a bad health shock,
A third area for future research concerns a loss of a job, or even a single bad har-
poverty traps. The literature suggests the vest may suddenly create real challenges.
possibility of sustained, long-run changes in Governments worldwide often step in to
outcomes from “graduation” programs. But provide social insurance programs to help
this is clearly not for everyone: The standard people manage these kinds of risks. They
“graduation approach” model may work in do so in part because private insurance
a rural area for someone with some ability markets are often incomplete, so even if
in animal husbandry, but this may not work individuals wanted to insure against these
for people in urban sectors, or for others risks on the private market, doing so would
who may want other types of employment. be difficult or expensive. There are often
1390 Journal of Economic Literature, Vol. LXII (December 2024)

substantial aggregate shocks, especially discuss aggregate shocks, in particular those

due to climate and natural disaster-re- related to climate and natural disasters.
lated issues—droughts, storms, floods, and These shocks affect many people at once—
earthquakes—that can pose challenges for changing the nature of the response, and
private insurance, especially in low- and limiting the ability of traditional risk pooling
middle-income countries, which are not to solve the problem—and can affect both
as well integrated into global reinsurance the income and expenditure sides simultane-
markets. And, for whatever reason, even if ously. Finally, in section 4.5, we discuss how
private insurance was supplied at actuarially risk interacts with broader issues of social
fair prices, insurance demand can be quite protection system design.
limited, so people may end up leaving risks Note that while the literature on social
uninsured that ex post can cause them seri- insurance in developed countries is exten-
ous challenges. sive, it is relatively less developed for low-in-
These problems, of course, are not lim- come countries. There is also substantial
ited to the developing world, but they end heterogeneity across countries in the scope
up in many ways more severe. Life in many of government-provided social insurance,
low- and middle-income countries is par- with social insurance programs more com-
ticularly risky. For example, 78 percent of mon in middle-income economies than
poor ­people reside in rural areas and mostly low-income ones.35 In light of this, we also
subsist on agriculture, so they face much discuss directions for future research.
more income risk than employees (World
4.1 The Challenges of Insurance Provision
Bank 2014). Since there is, on net, much less
and Rationales for Social Insurance
insurance against these shocks, households
end up taking much more severe actions
4.1.1 Basic Theoretical Issues: Moral
to smooth shocks (see, for example, Chetty
Hazard and Adverse Selection
and Looney 2007).
In this section, we broaden our frame- In section 2, we discussed a model where
work to consider social insurance policies. the government redistributes between
In section 4.1, we begin by summarizing households with different realized incomes.
the basic theoretical challenges of privately The premise of insurance markets, by con-
providing insurance (particularly in devel- trast, is that much of this redistribution can
oping country settings), discussing the basic be achieved by ex ante contracting between
rationale for these programs. Next, as many households.
low-income and emerging economies are We start with the simplest model, where
characterized by informal insurance, we dis- there are only two available types: with prob-
cuss their limitations and the places where ability ​​pi​​​​, a person will be a high-ability type
formal insurance could help fill the gaps. (earning higher income ​​y​hi​​​), and with prob-
We then review the current state of knowl- ability ​(1 − ​p​i​​)​, she will be a low-ability type
edge on social insurance schemes in devel- (earning lower income ​​y​li​​​). If the probabili-
oping countries. In section 4.2, we examine ties are public information, then the individ-
insurance against income shocks, for exam- ual can purchase insurance that pays ​​y​hi​​  − ​y​li​​​
ple, unemployment insurance and parental in the low state by paying an actuarially fair
benefits. In aection 4.3, we discuss insur-
ance programs that help cushion expendi- 35 For these reasons, social insurance programs are,
ture shocks, for instance, health insurance for example, much more developed in countries in Latin
and pensions/annuities. In section 4.4, we America. See Levy and Schady (2013) for a nice review.
 Banerjee et al.: Social Protection in the Developing World 1391

premium ex ante equal to ​​ m​i​​  = (1 − ​pi​​​) actuarially fair insurance contract, any risk-
(​  yh​ i​​  − ​y​li​​)​ . In a model with risk-averse indi- averse individual would buy it. Indeed,
viduals, symmetric information about types private insurance markets in developed
and actions, and actuarially fair insurance countries are large: the private insurance
markets, individuals will fully smooth their market accounts for about 9 percent of GDP
consumption across states by purchasing across the OECD (OECD 2020). By con-
insurance. In this world, the private mar- trast, this market is much smaller in many
ket will provide many types of insurance developing countries. Insurance premium
contracts: health insurance, unemployment volumes represent 1.63 percent of GDP in
insurance, crop insurance, et cetera. low- and middle-income countries, com-
As soon as we deviate from symmetric pared to 4.63 percent in high-income coun-
information, two market failures arise that tries (World Economic Forum 2019). This
undermine the private provision of insur- section explores several possible explanations
ance: adverse selection and moral hazard. for low insurance demand, that is, why many
Adverse selection arises because the individ- consumers in developing countries may not
ual has private information about their distri- want to buy even actuarially fair insurance.
bution of outcomes that the insurer cannot These explanations include information,
observe and thus cannot properly price; credit constraints, basis risk, and trust.
that is, either the probabilities p​ ​​ i​​​ or the risk
​​(​y​hi​​  − ​y​li​​)​​is at least partially private infor- Information, Trust, and Experience.—
mation. If the insurance company prices its One challenge is that people may not fully
contract to at least break even with high-risk understand insurance products, and there-
types, then those with lower risk may not buy fore may be unwilling to purchase them.
into the product offered. In an extreme case, Buying insurance requires a large degree
the market can completely unravel such of trust: individuals need to pay money now
that no insurance contracts exist (Stiglitz in the hope that in the future, if they have
and Weiss 1981). Government action can a problem, the insurer will cover their loss
help address this concern. (and not deny the claim). Sometimes a lack
The second market failure is moral haz- of trust is unwarranted, and comes from con-
ard, that is., the idea that insuring individuals fusion over how the products work and when
against adverse events can encourage risky claims should be received. In other cases,
behavior. For example, if individuals are however, profit-maximizing firms may have
insured against unemployment spells, they an incentive to limit claims, even if they are
might put less effort into finding a new job. legitimate under the rules. To create more
This can be included in the model by endog- trust, many developed countries have reg-
enizing ​​p​i​​​such that the probability of being ulation and legal processes to ensure that
in each state of the world is affected by an individuals know their rights and have mech-
individual’s behavior (Stiglitz 1974). If there anisms for recourse; even there, denial of
is heterogeneity in moral hazard, this can insurance claims is still a common concern.
also be a rationale for government interven- These challenges may be much more severe
tion (Einav et al. 2013). in the developing world, where financial reg-
ulation and the legal system are much less
4.1.2 Low Formal Insurance Demand in
developed. Some lack of trust may therefore
Developing Countries
be more rational on the part of citizens. For
The benchmark model above would pre- example, several papers show that lack of
dict that, when faced with the offer of an trust hinders insurance take-up by exploring
1392 Journal of Economic Literature, Vol. LXII (December 2024)

how experience effects (either from oneself “basis risk”—that is., the difference between
or one’s network) can improve insurance real losses incurred and losses insured based
take-up (Cole, Stein, and Tobacman 2014; on index values—can reduce demand for
Cai, de Janvry, and Sadoulet 2020). insurance products substantially.
It is worth noting that this problem is not
limited to private insurance markets. People Summing Up.—Experience, trust, credit
may not understand or trust insurance prod- constraints, and basis risk may all stifle
ucts from the government. Providing bene- demand for insurance products in develop-
fit information is often not sufficient if lack ing countries, even if actuarially fair. There
of trust is an issue (e.g., Dercon, Gunning, is some evidence that government-pro-
and Zeitlin 2019; Banerjee et al. 2021). In vided programs are also not immune to
this case, it is possible that experience with these issues, especially when they include
insurance matters more than just infor- co-pays. However, it is an open question
mation about government-provided insur- whether governments can help solve issues
ance programs that require co-payments. that private insurers may not be able to
Asuming, Kim, and Sim (2021) and Banerjee fully overcome, either by providing social
et al. (2021) find that temporary subsidies insurance directly or by working in con-
can lead to longer-run insurance demand junction with private providers to address
through experience. the challenges and concerns that limit
demand.
Credit Constraints and Timing Issues.—
4.1.3 Insurance Supply Issued in
Insurance premiums are typically paid up
Developing Countries
front. If households are credit constrained—
which many poor households in developing The literature on the supply challenges of
countries are—they may be less likely to insurance in low- and middle-income coun-
take up insurance against income shocks if tries is less developed, but the problem can
they need to pay premiums before income be characterized by two key factors. First, the
is realized. For example, in the context of limited information environments observed
insurance premiums for farmers in Kenya, in developing countries could make it chal-
Casaburi and Willis (2018) find that the lenging to administer insurance products in
take-up rate for pay-at-harvest insurance is practice. For example, in developed coun-
72 percent, compared to 5 percent for the tries, in addition to weather-based insur-
standard pay-up-front contract. ance, agricultural insurance contracts are
often written as a function of output or prof-
Basis Risk.—Basis risk is another rea- its from nearby farms. This type of contract
son why demand for insurance may be low, may have lower basis risk, but requires good
particularly for agricultural insurance con- information: the insurer needs to be able to
tracts (see Miranda and Farrin 2012, Carter observe the output or profits of all nearby
et al. 2017, Jensen and Barrett 2017). Given farmers (perhaps through tax or other data
private information concerns, instead of collection). By contrast, there is little data on
insuring against loss directly, agricultural the output of most subsistence farmers, so
insurance contracts are often written to pay these contracts are not practical to write in
out as a function of predicted losses. These most developing country contexts.
predictions typically only account for exog- A similar example comes from property: in
enous determinants of risk (e.g., weather), the United States, for example, 93 percent
and are therefore imperfect. The resulting of homeowners have property insurance
 Banerjee et al.: Social Protection in the Developing World 1393

(Insurance Information Institute 2020). Yet, are sustained in relatively closed networks
the fraction of people in developing coun- (Ligon, Thomas, and Worrall 2002). As vil-
tries who insure these risks is likely to be lages become more interconnected, or as
tiny. One reason why this may be different in the population increasingly lives in urban
lower-income countries may be a supply-side environments, these systems may become
constraint: Property insurers need to know less effective over time (Townsend 1995),
the market value of a house to insure it. The suggesting that the need for formal insur-
less-formalized property markets in devel- ance products may increase with develop-
oping countries may make such valuation ment. Moreover, informal insurance systems
challenging. can “trap” people in low-opportunity areas,
Second, challenges in contract enforce- and there may be inefficiencies if house-
ment, the legal system, and the regulatory holds need to remain in rural areas in order
environment may also impede insurance pro- to take advantage of these informal net-
vision. As described in section 4.1.2, insur- works (Banerjee and Newman 1998, Munshi
ance requires trust—i.e., that the insurer and Rosenzweig 2016).
will deliver on the contract and not renege Even when informal insurance exists, it
if a claim is made. In developed countries, does not necessarily provide the level of cov-
there is typically a regulatory role for the erage that households need, since income
government to make sure insurers have suf- shocks have a strong spatial correlation,
ficient capital to pay claims (and reinsurance especially in agricultural areas. Ideally, risk
to cover unexpected losses). There is also pooling would be done across villages, where
a functioning legal system through which shocks are independent. However, such
people can sue if they feel they are wrongly arrangements are difficult in practice, given
denied a claim. These systems are substan- that monitoring costs to prevent moral haz-
tially less developed in most developing ard are very high. Spatially correlated shocks
countries. may be particularly challenging in the con-
text of natural disasters or conflict, necessi-
4.1.4 Informal Insurance
tating large-scale disaster relief programs, as
A third complication for formal insurance we discuss in section 4.4.
provision in developing countries, compared All of these arguments suggest that some
to the developed world, is how it interacts forms of formal insurance are likely to be
with complex, preexisting informal insur- useful, even where informal mechanisms
ance arrangements. Many poor households are present. This then makes it necessary
in developing countries—while not for- to think through the potential interactions
mally insured—engage in various forms of between (formal) social insurance and infor-
risk sharing. Indeed, villagers are better at mal insurance, especially given the evidence
smoothing idiosyncratic consumption shocks that as economies grow, many households
than one may expect, given the absence of may no longer be able to rely as much on
formal insurance markets (Townsend 1994). informal networks. Ultimately, there are
However, from the perspective of a potential two key questions: first, whether offering
social insurance designer, it is important to formal social insurance crowds out informal
note that the existence of informal insurance insurance; and second, what welfare conse-
does not preclude the need for formal insur- quences arise from the interaction between
ance schemes. the two.
One key concern with relying on infor- Regarding the first question, the evi-
mal insurance mechanisms is that they dence on crowding out is mixed (Albarran
1394 Journal of Economic Literature, Vol. LXII (December 2024)

and Attanasio 2003; Strupat and Klohn 2018; literature on the degree to which this con-
Takahashi, Barrett, and Ikegami 2019). The ditioning of benefits on future employment
level of crowding out may vary based on discourages job search in high-income
characteristics of the insurance: for exam- countries. This literature focuses on (i) the
ple, while Huang and Zhang (2021) find no trade-offs between the welfare gains from
crowding out of private transfers in the con- insurance and their potential disincentive
text of China’s rural pension scheme, Jensen effects, and (ii) whether the benefits allow
(2004) finds large crowd-out effects of South for better job matching (e.g., Chetty 2009a;
Africa’s public pension. Huang and Zhang Nekoei and Weber 2017; Farooq, Kugler,
argue that the difference may be driven by and Muratori 2020). Regarding (i), a key
smaller pension benefits in China. question is whether policy design choices
As for the second question, Arnott made within UI programs (i.e., the level and
and Stiglitz (1991) theorize that whether duration of benefits) can improve welfare.
the existence of formal and informal insur- Examples of papers estimating these trade-
ance systems is welfare improving depends offs in low- and middle-income countries
on how well informal networks monitor oth- include van Ours and Vodopivec (2008);
ers to reduce moral hazard. They analyze Huneeus, Leiva, and Micco (2012); and
two extreme cases: one with perfect moni- González-Rozada and Ruffo (2023).36
toring and one with no monitoring at all. In While the same policy questions exist in
the first case, welfare improves since mon- many low- and middle-income countries,
itoring by informal networks helps mitigate they are complicated by the existence of a
moral hazard and thus improves risk shar- large informal sector, which prevents the
ing. At the other extreme, welfare decreases government from monitoring job entry
since formal insurers know that there is no and exit. This condition has two important
monitoring and adjust premiums and pay- implications. First, the government cannot
outs accordingly. Building on this analysis, provide benefits that begin conditional on
Attanasio and Rıos-Rull (2000) show how unemployment for those in the informal sec-
even a well-intentioned policy could reduce tor, so informal workers may not actually be
welfare. The intuition behind their result covered by UI. Second, the government may
is that the introduction of formal insurance be unable to condition the end of benefits on
ruptures the social fabric, leaving households reemployment (since it only observes formal
more vulnerable to risks not covered by for- employment). Workers receiving UI could
mal insurance. choose to seek employment in the informal
sector in order to continue receiving their
4.2 Insurance against Income Loss
benefits. In this way, UI programs may dis-
We next turn to understanding the chal- incentivize formalization, which would have
lenges of insuring against specific types of consequences for productivity, taxation, and
risk. In this section, we covers risks of income workplace safety.
loss. Section 4.3 covers expenditure risks. Gerard and Gonzaga (2021) examine the
relationship between informality and unem-
4.2.1 
Insurance against Involuntary Job Loss
ployment insurance in Brazil. They find,
Unemployment insurance (UI) is designed
to help people smooth consumption between
36 A number of recent papers also examine the spill-
jobs. However, a key challenge is that unem-
over effects of unemployment insurance on domestic
ployment benefits typically continue until the violence (Bhalotra et al., 2021) and crime (Britto, Pinotti,
worker resumes work. There is an e­ xtensive and Sampaio 2022).
 Banerjee et al.: Social Protection in the Developing World 1395

­ erhaps counterintuitively, that the presence

p may disincentivize firms from hiring, and a
of a large informal sector reduces the effi- single lump sum means that workers bear
ciency losses from moral hazard in UI. Indeed, the risk of being unemployed for longer than
they estimate that the efficiency costs of UI average. Using data from São Paulo, Brazil,
are 5 times lower in Brazil than in the United Gerard and Naritomi (2021) show that work-
States. This difference arises because the ers appear to overspend from lump-sum
presence of the informal sector allows work- severance payments relative to their optimal
ers to keep working (albeit informally) while consumption profile. They suggest that a
retaining UI benefits, so while workers do stream of payments (which could, of course,
lose formal protections from taking informal be unconditional) may lead to better con-
jobs, informal employment reduces the over- sumption smoothing than a lump sum.
all efficiency consequences of the UI tax on Another alternative is to create unem-
reemployment. Of course, informal employ- ployment savings accounts that workers
ment comes at a cost, because informal jobs could access in case of job loss, funded by
have lower earnings than formal ones. This mandatory contributions by workers and/or
finding is echoed in Liepmann and Pignatti’s firms (Feldstein and Altman 2007). The idea
(2024) study in Mauritius. is that individual accounts could help align
Given these challenges, one often observes incentives and reduce the scope for moral
alternative policies to UI to help to insure hazard. These accounts would not have the
workers against job loss. A common policy risk-pooling features of traditional insur-
is mandated severance pay: lump-sum pay- ance; they also have the challenge that work-
ments disbursed upon termination of a labor ers who only just starting paying in may not
relationship. These payments are not condi- have sufficient balances to cover an unem-
tioned on future employment, and thus do ployment shock (though governments could
not distort future employment decisions. in principle supplement accounts with too
Severance programs also have challenges. little money, guaranteeing a minimum bal-
First, someone needs to verify that workers ance). Countries such as Colombia, Chile,
were, indeed, working, and that they were Indonesia, and Mexico have implemented
terminated. If firms are paying directly, the this type of system. Kugler (2005) and Nagler
government needs to ensure that they do; (2013) examine how the shift from severance
if the government collects unemployment to these types of linked accounts in Colombia
taxes and pays the severance, it needs to col- and Chile, respectively, impacted wages and
lect the taxes to do so. Weak institutions may job duration.
exacerbate these enforcement challenges. In short, the challenges of a large infor-
For example, Sadka, Seira, and Woodruff mal sector mean that the design of insurance
(2024) find that in Mexico, despite a strong programs for unemployment in developing
de jure severance policy, many workers do countries may need to be fundamentally
not receive their full entitlement and delays different from those in high-income ones.
and misinformation abound in the court These challenges also suggest avenues for
system. The authors show that improving future research, particularly on policies that
institutional quality could help: Providing are not conditioned on reemployment (e.g.,
information about likely court outcomes individual unemployment-triggered savings
substantially shortens settlement times and accounts, severance, and lump-sum unem-
improves workers’ food security. ployment insurance paid by the government).
In addition, the need to make large sev- Moreover, given information constraints,
erance payments in the event of termination countries are also designing self-targeting
1396 Journal of Economic Literature, Vol. LXII (December 2024)

techniques that condition benefits on costly countries, suggesting that much of this risk
(but potentially productive) activities. For remains uninsured. Moreover, the relative
example, Indonesia requires job training to paucity of formal pensions—which often
receive benefits. Understanding the impacts come with survivors’ benefits to insure
of these policies on both insurance and ulti- spouses against the loss of income from their
mately labor market activities is important. primary earner in high-income countries—
leaves more of this risk uninsured. As with
4.2.2 Insurance against Disability and
disability, widow status is often considered in
Death
PMT formulas, but again, this only provides
Disability insurance is designed to help insurance to the extent that households are
insure workers against accidents or illnesses near the poverty threshold.
that may either temporarily or perma-
4.2.3 Insurance against Agricultural Loss
nently remove them from the labor market.
However, similar to UI, the large informal Agriculture plays a large role in low- and
sector hinders governments from provid- middle-income countries. But agriculture is
ing universal disability coverage through risky: crops can fail and prices are volatile, so
employers. Therefore, one often observes those engaged in agriculture as their primary
alternative policies to provide transfers to the occupation face much more income risk than
disabled. For example, many cash transfer those who work in manufacturing or other
programs, such as Argentina’s Programa de sectors. These risks are not covered by tra-
Ciudadanía Porteña, Chile’s Subsidio Único ditional unemployment insurance schemes.
Familiar, Indonesia’s Program Keluarga For these reasons, there has been a big
Harapan, and Tanzania’s Productive Social push among governments to develop pro-
Safety Net, give weight to having a disabled grams that help farmers manage agricul-
household member in the eligibility criteria. tural risk. Some programs involve providing
This has two challenges: (i) it only provides technologies that ensure more stable yields
insurance for those near the poverty thresh- and prevent crop loss, be it investments in
old (not the middle class) and (ii) if targeting irrigation or weather-resistant seeds. Other
is done infrequently, the sign-up period may policies aim to provide agricultural insurance
not align with the need for assistance. to help farmers smooth consumption during
A related issue is workplace accident periods of crop loss. There are a number of
insurance. This is often separate from dis- excellent reviews on these topics; see, for
ability insurance, in part due to attempts to example, Cole and Xiong (2017); Ali, Abdulai,
link risks back to employers. While some and Mishra (2020); and N ­ shakira-Rukundo,
middle-income countries are trying to pro- Kamau, and Baumüller (2021).
vide this type of insurance to workers in the
4.2.4 Parental Benefits
formal sector, this topic is comparatively
underexplored in the economics literature. Parental benefits are designed to help
Finally, we touch briefly on life insurance. insure families against income loss associ-
The death of a primary income earner often ated with pregnancy and the early period of
implies a substantial income loss. Indeed, a child’s life, when parents reduce labor sup-
in many developing countries widowhood is ply to care for the newborn child. These ben-
closely associated with poverty. Formal life efits also protect parents from employment
insurance markets tend to be much smaller termination during this period of life. One
(as a share of GDP) in low- and middle-in- common type of parental benefit provides a
come countries compared to high-income mechanism in which workers can take time
 Banerjee et al.: Social Protection in the Developing World 1397

off around pregnancy and a child’s birth by public preschool expansions in Indonesia
mandating paid time off and/or subsidizing and find positive effects on women’s unpaid
firms for the worker’s time off. A second family work. They argue that limited day care
benefit takes the form of subsidizing child- hours made it difficult for women to find jobs
care to allow parents to reenter or stay in the outside their homes.
workforce. Recent experimental studies (de Barros
While virtually all countries (all but seven) et al. 2011, Martínez and Perticará 2017,
have some type of maternity leave policy Clark et al., 2019) have also studied access to
(Iqbal 2018), benefit levels in developing childcare in Brazil, Chile, and Kenya, respec-
countries are often very low for mothers, and tively, with all three finding positive impacts
paternity leave is often nonexistent. Many of of childcare access on women’s work. In fact,
those in the large informal sector do not have Bjorvatn et al. (2022) find that income gains
access to any formal benefits. from childcare led to at least as large a gain
In contrast to the extensive literature on as an equivalent cash transfer, while also
these policies in high-income countries, there improving child development. Interestingly,
is relatively little micro-empirical research they find that the earnings increase comes
on paid maternity or paternity leave in lower- from higher productivity, largely in self-em-
and middle-income countries. There are, ployment, rather than longer hours.
however, a few notable exceptions. In Chile, There are still many open questions in this
Albagli and Rau (2019) show that a doubling area. For example, how should parental ben-
of maternity leave (from 12 to 24 weeks) efits be paid and who should pay for them?
led women to take substantially more time Should they be universal or dependent on
off after childbirth, increased breastfeeding labor status, especially in countries where
duration, and improved cognitive outcomes. the prevalence of informal firms may make
Interestingly, women with the enhanced it hard to mandate benefits? How should
leave policy were also more likely to be in countries think about provision of childcare
the labor force one year later. In Vietnam, as larger extended family networks—which
Vu and Glewwe (2022) show that a more often provided this type of childcare—break
generous maternity leave shifted potentially down? And more broadly, what are the fer-
eligible women from informal work to for- tility impacts of changing parental support,
mal work and, in particular, to public-sector particularly as many of these countries are
jobs where they could access benefits. More undergoing demographic transitions with
research is needed to understand how preg- rapidly falling fertility rates?
nancy affects both consumption smoothing
4.3 Insurance against Expenditure Loss
and women’s career trajectories, and how
effective leave policies can be designed given The other type of risk is expenditure
the presence of informal labor markets. risk—that is, the risk of unexpected large
In contrast, free or subsidized childcare outlays. In this section, we outline common
has become increasingly common, particu- social insurance programs that aim to allevi-
larly in middle-income countries. A number ate these risks.
of studies have tried to examine its impacts.
4.3.1 Health Insurance
For example, Calderón (2014) shows that
access to childcare in Mexico increased wom- Government-led health insurance systems
en’s likelihood of working and reduced their are becoming increasingly common in many
likelihood of earning zero income. Similarly, developing countries, as governments aim
Halim, Johnson, and Perova (2022) examine to increase health-care utilization, improve
1398 Journal of Economic Literature, Vol. LXII (December 2024)

health outcomes, and help households man- insurance and can help households manage
age income and consumption risks that may financial shocks from health events (King
arise from health shocks. Nearly 190 coun- et al. 2009; Levine, Polimeni, and Ramage
tries have some sort of public health insur- 2016; Gruber, Lin, and Yi 2023). In fact, del
ance system. Valle (2021) argues that health insurance
The literature on health insurance in low- provides another additional economic ben-
and middle-income countries tends to fall efits: the author shows that by reducing the
into two buckets. The first bucket examines severity of health shocks, Seguro Popular in
the impact of health insurance on insured Mexico reduced the degree to which other
households. This includes their ability to household members needed to drop out
smooth out the economic consequences of of the labor force to directly provide care.
health shocks, as well as their health-care Other work provides suggestive evidence that
utilization and health outcomes. The second health insurance may reduce financial stress
bucket explores the challenges of insurance above and beyond its impact on finances per
design and provision; in particular, how to se (Haushofer et al. 2020).
design and fund public insurance systems
given the constraints of informal employ- Health Care Utilization and Health.—
ment, adverse selection, moral hazard, Health insurance also has the potential to
information failures, and trust. See also Das change health-care utilization, since it can
and Do (2023) for a recent review of this reduce the price of health care, which could
literature. in turn generate health impacts. In some
cases, one may actually be concerned that
Impacts of Health Insurance.—It is worth this leads to overconsumption of health care,
noting that conceptually, the primary impact since consumers do not face the true mar-
one would expect from health insurance is ginal cost of the care and, indeed, in some
on consumption smoothing. That is, health developed countries alternative schemes
insurance is primarily a financial product with high deductibles have been developed
that provides payment in the event of an in an attempt to provide financial insur-
expenditure shock and should therefore ance while ensuring that households face
make the financial consequences of poor the correct price of care on the margin. On
health less severe (Finkelstein and Mcknight the flip side, if households are liquidity con-
2008). Health insurance can also affect strained (as we believe many households in
health-care consumption; since health insur- developing countries are), households may
ance typically reduces the marginal cost of under-consume health care without insur-
health-care utilization (as opposed to, say, ance. Understanding how to help households
lump-sum cash transfers, which would mit- achieve their optimal, distortion-free level of
igate the financial consequences of illness health consumption remains an important
without changing prices on the margin), one challenge.
would expect utilization to increase. This Does health insurance actually affect
increased health-care utilization could, in health consumption, and health, in develop-
turn, potentially affect health outcomes. ing countries? The experimental evidence
here is mixed, though one challenge is that
Health Insurance and Financial Shocks.—A some of the experimental studies may not
number of papers from low- and middle-in- be sufficiently powered to detect small but
come countries suggest that health insurance economically meaningful effects (King et al.
does indeed provide meaningful financial 2009, Haushofer et al. 2020). A recent study
 Banerjee et al.: Social Protection in the Developing World 1399

by Malani et al. (2021) highlights another to more generous government financing on

important policy issue: While the authors net. Using administrative mortality records,
find some increase in insurance usage in the authors show that prior to the reform,
their experiment, many beneficiaries had infant mortality rates were related to prov-
challenges using their insurance. For exam- inces’ wealth. After the reform, however,
ple, they often had problems with their insur- resources were equalized across provinces,
ance cards, did not know how to use them, leading this correlation to disappear.
or forgot them. Perhaps unsurprisingly, the Importantly, Gruber, Lin, and Yi (2023)
authors observe no health effects. In short, and Gruber, Hendren, and Townsend (2014)
most of the experimental studies to date find argue that part of why large health effects
little overall impact of health insurance on were seen is that insurance reforms, by
health outcomes. increasing government financing for health,
Given the need for large samples sizes to increased the supply of health services, as
measure impact, a number of other studies has been shown in developed countries
focus on quasi-experimental variation from (e.g., Finkelstein 2007). They argue that
large-scale public health insurance reforms, this increase in supply helps explain the
using large administrative datasets on mor- large health effects of insurance reforms. In
tality and other health outcomes. For exam- evaluating health insurance systems exper-
ple, Gruber, Lin, and Yi (2023) examine the imentally, it may therefore be important to
rollout of China’s public health insurance for randomize across hospital catchment areas
rural households, which covered as many or health markets, rather than individuals or
as 800 million people. They find a signifi- villages. Such designs would allow for estima-
cant decline in aggregate mortality, which tion of the full general equilibrium impacts
they argue explains 78 percent of the entire that could arise from insurance expansions
increase in life expectancy in China during or enhancements.
this period. Using survey data, they find large
effects on health-care utilization and a host Challenges with Health Insurance
of other health outcomes. Design.—The second bucket of work on
Gruber, Hendren, and Townsend (2014) health insurance centers around how to
study an alternative method of health insur- design and fund health insurance systems.
ance: the expansion of free or heavily subsi- For example, many countries choose not to
dized care at public facilities. They examine universally cover health insurance premiums
Thailand’s 2001 health-care reform, known through the government budget and instead
as the “30 Baht” program. Prior to this set up national health insurance programs
reform, the poor, young and old, were given where the poor are covered directly and
free public health care, but the system was everyone else has to pay a mandatory contri-
seen as chronically underfunded. The reform bution. Contributions are often collected for
changed the system in two ways. First, it pro- formal workers through payroll taxes remit-
vided universal access to public facilities at a ted by employers.
co-pay of 30 Baht (about US$0.75) per visit. A common challenge with these systems,
Since the co-pay was already waived for the however, is how to handle informal work-
poor, in practice this reform largely led to ers (for whom premia cannot be collected
reduced health access costs for the informal, from employers). There are, broadly speak-
nonpoor. Second, the reform provided hos- ing, three options. First, the government
pitals with a universal capitation payment could opt not to provide health insurance to
based on their province’s population, leading nonpoor, informal workers. This, of course,
1400 Journal of Economic Literature, Vol. LXII (December 2024)

would mean that these workers would contributions for formal workers, and
remain vulnerable to the economic impacts find evidence of both effects. The reform
of health emergencies. Moreover, if firms extended the SIA’s coverage for dependent
wanted to evade paying benefits to formal children, making the program more attrac-
workers and/or workers that do not fully tive. At the same time, the reform increased
value insurance, mandating insurance ben- the payroll tax contribution deducted from
efits only for formal workers could increase employees’ salaried earnings, which could
informality. Second, one could mandate that have caused an increase in underreporting
nonpoor, informal workers contribute to of wages for formal employees. The authors
insurance. This would be difficult to enforce, find both effects: the greater benefits drew
however, particularly in countries with lim- people into the system, but there was also
ited administrative capacity. Moreover, the an increase in wage misreporting. On net,
resulting adverse selection problems (i.e., the fiscal revenue gain from higher levels of
only signing up when sick) could financially formal employees was much larger than the
strain the insurance system. Third, one loss of revenue due to underreporting.
could decide that it is too hard to collect Camacho, Conover, and Hoyos (2014)
contributions from informal workers and examine the introduction of subsidized
instead extend free insurance to most, if not insurance for informal workers on labor mar-
all, informal workers. Once again, concerns ket choices in Colombia. A reform was insti-
about encouraging informality would arise as tuted that made workers who were below a
well as cost considerations. Poverty Index Score eligible for noncontrib-
Below, we first discuss the evidence on utory health insurance, but those who were
the interplay between health insurance and formally employed were ineligible regardless
informality. We then discuss the evidence on of score (and needed to contribute through
adverse selection, as well as policy tools that their employers), creating an incentive to
have been used to mitigate it. We conclude become informal. They find that the reform
with open research questions. increased informal employment by about 3-4
percentage points.
Health Insurance and Informality.— Finally, a number of papers have exam-
Ex ante, the effect of employer-provided (or ined the effect of Mexico’s Seguro Popular
subsidized) health insurance for formal-sec- program on formalization. Prior to the roll-
tor works on formality itself is ambiguous. If out of Seguro Popular, insurance was tied
(a) workers value the insurance and (b) there to payroll contributions, and many people
is a cost savings (from the workers’ perspec- were uncovered. Seguro Popular aimed to
tive) of gaining insurance by becoming a provide universal coverage. Analyzing the
formal employee, as opposed to paying an staggered rollout of the program and sur-
individual premium, this could increase for- vey data, Aterido, Hallward-Driemeier,
malization. But, if workers do not value the and Carmen (2011) and del Valle (2021),
insurance at cost or if alternative health-care among others, find small effects on formal-
arrangements are cheaper, employer-pro- ization, while Azuara and Marinescu (2013)
vided insurance could increase informality. find no effect on average, but small effects
A number of papers examine these con- for unskilled workers. More recently, exam-
cepts. For example, Bergolo and Cruces ining the rollout with social security data,
(2014) examine a large-scale policy reform Bosch and Campos-Vazquez (2014) show
in Uruguay’s social insurance adminis- that it slowed the social security registration
tration (SIA) that increased benefits and of employers and employees in small and
 Banerjee et al.: Social Protection in the Developing World 1401

medium firms (up to 50 employees), reduc- Bundling.—A second approach to reduce

ing revenues paid into the social security sys- adverse selection is bundling: by tying the
tem as well as VAT payments. purchase of health insurance to that of
another product, or bundling insurance
Adverse Selection, Information and Trust purchases for a household together, house-
with Imperfectly Enforced Mandates.— holds’ ability to buy insurance only for the
Many countries have tried to mandate that sick can potentially be limited. For exam-
nonformal, nonpoor workers purchase insur- ple, in their study with an NGO in Pakistan,
ance, but limited enforceability remains a Fischer, Frölich, and Landmann (2023)
challenge. Several studies have examined experimentally vary whether people can
what can be done to mitigate this issue. choose to enroll individuals or whether they
must enroll the entire household, under the
Subsidies.—Can time-limited subsidies idea that households will have to enroll all
substitute for unenforceable mandates? In members (not just the sick). They also exam-
Ghana’s health insurance scheme, Asuming ine a community bundled contract where at
(2013) and subsequently Asuming, Kim, least 50 percent of the community must sign
and Sim (2024) randomize partial and full up in order to activate the insurance. These
subsidies of premiums for one year and track bundled contracts reduce “expected costs”
the results for three years. Subsidies increase among those who sign up (where expected
take-up, an effect that persists even after the costs are claims predicted from baseline
subsidies end. covariates), suggesting that this type of group
Banerjee et al. (2021) also study time-lim- insurance may be effective. In many devel-
ited subsidies in the context of Indonesia’s oped countries, workplaces are the “group”
national health insurance, randomizing for health insurance purposes; this study
offers of partial and full premium subsidies suggests other groups as an alternative in
(and a control) for one year. Larger subsidies countries with high levels of informality.
bring healthier individuals into the health- The downside of bundling, however, is that
care system, consistent with undoing adverse if demand for insurance is low, then it can
selection. They also document a pattern of reduce demand for other bundled products.
dynamic selection where those who enroll For example, Banerjee, Duflo, and Hornbeck
in the no-subsidy condition are much more (2018) find that a substantial proportion
likely to immediately file large claims. Since of people were apparently willing to forgo
the subsidies attract healthier individuals renewing their microcredit just to avoid pur-
who then also pay premia in the post-subsidy chasing the required health insurance bundle.
period, they allow the government to cover
more people at the same total cost. Fischer, Information and Trust.—Lack of informa-
Frölich, and Landmann (2023) in Pakistan tion (or misinformation) about the benefits
also experimentally vary premia and find that of insurance and limited trust in the system
adverse selection is higher when premia are (some warranted and some unwarranted)
higher. In short, while subsidies are by no may be another reason for low insurance
means a panacea, they can help ameliorate demand, as discussed in section 4.1.2. For
adverse selection.37 information provision, the evidence is mixed,
with Giles et al. (2021) and Asuming (2013)
finding positive impacts, for example, and
37 Two other important papers on subsides are Thornton Banerjee et al. (2021), and Dercon, Gunning,
et al. (2010) and Wagstaff et al. (2016). and Zeitlin (2019) finding none.
1402 Journal of Economic Literature, Vol. LXII (December 2024)

Dercon, Gunning, and Zeitlin (2019) mea- 96 have both (see online appendix table 2).
sure trust using a trust game at baseline and These programs typically have three related
then offer composite health insurance to goals: (a) to help individuals save for their old
individuals with varying subsidies. They find age, either individually or through a tax-and-
that low generalized trust is negatively asso- transfer scheme; (b) to provide annuities that
ciated with insurance demand and that the insure against the risk of living too long as
purchase decisions of individuals with low compared to one’s savings; and (c) to provide
trust are significantly sensitive to price. some amount of redistribution so that even
In sum, despite the benefits of health those with low incomes are not too poor in
insurance for risk smoothing, health insur- old age.
ance provision in developing countries As seen in figure 4, governments often
remains a challenge. The combination develop these systems as incomes rise,
of lack of demand for insurance, adverse perhaps coincident with the rise of more
selection, a large informal sector, and the modern economies and increased life expec-
relative challenges of enforcing a man- tancies. That said, figure 4 reveals that even
date to purchase insurance means that it conditional on real GDP per capita, these
is challenging to ensure widespread insur- programs are becoming more prominent in
ance coverage. Many countries respond to low- and middle-income countries.
these issues through public-sector systems A substantial literature has studied the
with low prices, with those who want better effects of pension transfers on recipient
coverage opting out at their own expense. households, primarily from the perspective
However, this approach leads to substantial of labor supply and consumption choices.
uninsured risk. Working out how to move The literature on the design of these pro-
towards more comprehensive coverage grams is less developed. We review each of
remains an important direction for future these literatures in turn.
research.
Effects on Consumption and Well-
4.3.2 Pensions and Annuities
Being.—One strand of research is focused
Old age is often associated with poverty. on understanding the impact of pensions
Health typically falters as people age, so the on beneficaries’ well-being. For exam-
elderly are less able to work; this is particu- ple, examining the rollout of China’s New
larly true for those engaged in physical labor. Rural Pension Scheme, Huang and Zhang
Saving (or saving enough) may be difficult, (2021) find that the program increased
particularly for those outside the formal sec- income while reducing mortality. Similarly,
tor. Even for those who do save, aging entails Galiani, Gertler, and Bando (2016) eval-
risk: in particular, the risk that one will out- uate the impact of Mexico’s Adultos
live one’s savings. Traditionally, these chal- Mayores Program, a noncontributory cash
lenges have been borne by family members transfer provided to the elderly, and find
(e.g., children caring for parents), but these increases in consumption and reductions in
networks may be incomplete and may break depression.
down with urbanization. Consumption effects need not be limited
Accordingly, many governments around to pensioners themselves, as consumption
the world have public pension programs resources are shared within the household.
with a mix of systems. Indeed, 181 countries Duflo (2003), for example, shows that pen-
have some type of contributory pension sys- sions received by older women in South
tem, 102 have noncontributory systems, and Africa improve the anthropometrics of
 Banerjee et al.: Social Protection in the Developing World 1403

0.8
0.75
0.7 Year
0.65 1960

0.6 1980
Percentage of countries

0.55 2000

0.5 2020

0.45
0.4
0.35
0.3
0.25
0.2
0.15
0.1
0.05
0
Low Lower−middle Upper−middle High
<$1,046 $1,046−$4,095 $4,096−$12,695 >$12,695
Income classification (based on GDP in constant 2015 US$)

Figure 4. Percentage of Countries That Enacted Noncontributory Pension Systems by Income Classification
over Time

Notes: This figure graphs the percentage of countries that enacted noncontributory pension systems by each
year considered (1960, 1980, 2000, 2020) against quantile of GDP per capita (in constant US$2015). GDP per
capita quantiles are calculated as an absolute scale across the time periods considered.
Source: Income classification, GDP per capita, and population data is from World Bank (2021a, b, c).

female children in the household.38 Studying however, looks at ex ante child investments.
the same program, Edmonds (2006) finds She argues, using data from Indonesia and
increases in schooling and declines in child Ghana, that the introduction of public pen-
labor for boys, as well as declines in domestic sions make parents less likely to invest in
labor for girls when men in the household children’s education, since they no longer
become eligible. Similarly, de Carvalho Filho need to rely on their children for old-age
(2012) studies an unexpected pension reform support.
in Brazil and shows that pension eligibility
leads to an increase in school enrollment Labor Supply Effects.—A second set
for girls. Combined, these papers suggest of papers examines the impact of pension
important intergenerational spillovers from receipt on labor supply and/or retirement
pension receipt within the household. decisions. To the extent that these are
This set of papers focuses on what happens income effects (as opposed to price effects
when income starts flowing. Bau (2021), induced by features of the benefit formula),
labor supply reductions associated with pen-
38 Case and Deaton (1998) provide a rich description of sion receipt are likely to be welfare improv-
South Africa’s noncontributory pension program. ing, as people appear to have been working
1404 Journal of Economic Literature, Vol. LXII (December 2024)

longer than they otherwise would have in the A second question concerns investment
absence of a pension. choices. Some contributory systems allow
Several papers find that pensions reduce individuals to choose how their assets are
labor supply. De Carvalho Filho (2012) stud- invested. Hastings, Hortaçsu, and Syverson
ies the reform to the Brazilian Social Security (2017) document very little attention to
System discussed above and shows that it fees among plan participants in Mexico,
reduced the retirement age of rural men. so that fees charged were extremely high.
This is an example of a pure income effect, Combined, these fees meant that a ­100-peso
as the transfer was not means or retirement deposit that earned a 5 percent annual return
tested. Galiani, Gertler, and Bando (2016) would be worth only 95.4 pesos after 5 years.
study Mexico’s Adultos Mayores Program They document that a key constraint was
and also find that paid work declines, but that workers were not particularly price sen-
that those who stop doing paid work switch sitive, so firms primarily competed on non-
to family businesses. Huang and Zhang price attributes. This suggests that it can be
(2021) show that in China, pensions challenging to structure market incentives so
decreased labor supply for rural people older that competition leads to low prices, and that
than 60 (i.e., the eligible), particularly for care must be paid in structuring private man-
farm work.39 agement of individual account systems.
Third, there are questions about the
Program Design.—There is comparatively interactions of program design with savings
less work, however, on program design ques- incentives. An individual contributory sys-
tions in low- and middle-income countries, tem encourages private savings but is less
but we highlight several papers that suggest redistributive. A country may therefore want
some ways in which low- and middle-income to add a substantial component to cover the
contexts may be somewhat different from informal system, that is, a minimum pen-
higher-income contexts. sion floor. The challenge is to do this with-
One important question concerns contribu- out discouraging savings. Attanasio, Meghir,
tions: do firms report the correct social secu- and Otero (2014), for example, document
rity contributions, and pay taxes accordingly, these trade-offs in the context of Chile’s 2008
or do they underreport? Kumler, Verhoogen, reform, which introduced a minimum pen-
and Frías (2020) show that system design may sion floor.
matter: when the pension system became Fourth, there are questions of take-up.
largely contributory, payments for younger Many pension programs suffer from the
workers (who had more of an incentive to take-up challenges described above for cash
ensure their wages were reported accurately) or in-kind transfer programs. Gupta (2017)
increased. experimentally shows how administrative
challenges hinder the take-up of a widow’s
pension program in India, and that improv-
39 The income effects from a pension could also affect ing this process can increase the provision of
the labor supply decisions of other members. Studying benefits.
South Africa’s pension program Bertrand, Mullainathan, These challenges, however, are only the tip
and Miller (2003) find a sharp drop in the working hours
of prime-age men in the household, particularly when the of the iceberg and there is substantially more
pensioner is a woman. Posel, Fairburn, and Lund (2006) work to be done. A broad question concerns
argue that the pensions also drive migration among prime- the informal sector: as with health insurance,
age workers. Including migrants in the analysis, Ardington,
Case, and Hosegood (2009) find small positive increases in if pensions are introduced in the formal sec-
work among prime-age adults. tor and funded through payroll taxes, will this
 Banerjee et al.: Social Protection in the Developing World 1405

create a further wedge between formal and property insurance (Insurance Information
informal sectors? How does the presence of Institute 2020). For certain types of perils
the large informal sector interact with the where there are aggregate shocks, coverage
decision of whether pension systems should remains low. However, the fraction of peo-
be contributory or noncontributory? And as ple in developing countries who insure these
these systems grow over time, what will the risks is tiny despite their important effects
implications be for the many economic deci- (Anttila-Hughes and Hsiang 2013).
sions—marriage, investments in children, One potential reason is that adverse selec-
savings—that are related to how people plan tion and moral hazard may be a challenge. For
for their old age? example, in the United States, private prop-
erty insurers maintain the Comprehensive
4.3.3 Funeral Insurance
Loss Underwriting Exchange (CLUE) data-
Funerals are often a large, not always fore- base of claims so that insurers can price idio-
seen cost that can devastate households, par- syncratic risk into future premia. We are not
ticularly in bad times. For example, analyzing aware of similar systems in most developing
data from South Africa, Case et al. (2013) country contexts.40
find that households spend about a year’s Moreover, like other forms of insurance,
income for an adult funeral. Funeral insur- low demand often prevails due to lack of
ance can help manage these risks, whether trust in whether institutions will pay out
through community associations or burial (Reynaud, Nguyen, and Aubert 2018). More
clubs (Dercon et al. 2006, Case et al. 2013, generally, understanding why this market is
Berg 2018), religious institutions (Auriol limited seems important for future research.
et al. 2020), or private insurance companies
4.4 Natural Disasters and Climate-Related
(Berg 2018). In the Case et al. (2013) study
Shocks
of South Africa, about 28 percent of people
who died had some form of funeral insurance Natural disaster and climate-related
(either through burial clubs, funeral parlors, shocks, such as floods, earthquakes, droughts,
or private insurance). However, many more and hurricanes, pose their own set of chal-
needed to rely on savings, and/or borrow lenges. These shocks tend to be aggregate in
(about a quarter) to help cover the costs. nature, affecting a large number of people
As Dercon et al. (2006) discuss, many of in a particular geographic area at once. This
these informal insurance or burial clubs have means that the informal insurance networks
sets of rules and institutions to promote fair- discussed in section 4.1.4, which provide sub-
ness. Nonetheless, an important question is stantial risk-smoothing in low- and middle-in-
whether certain types of households can be come countries more generally, are not able
excluded and whether there is a role for gov- to deal with these risks well.41 In the absence
ernment-provided funeral insurance to com- of well-developed reinsurance markets, the
plement many of these informal structures.
4.3.4 Property Damage
40 An exception is TransUnion, South Africa’s similar
Risk from property damage poses a chal-
insurance claims and policy database.
lenge, whether it be from flooding, fire, 41 While local risk-pooling networks will be ineffective,
crime, or other perils. In wealthy countries, distant—and indeed even international­—risk-pooling net-
private insurance often covers these idiosyn- works can respond to these types of shocks, i.e., through
changes in remittances or in migration; see Yang (2008)
cratic risks: In the United States, for exam- and Mahajan and Yang (2020), respectively, for evidence
ple, over 90 percent of homeowners have on these channels.
1406 Journal of Economic Literature, Vol. LXII (December 2024)

private sector in low- and middle-income famine, despite the substantial amount of aid
countries may have challenges as well.42 delivered.
Many of these shocks are simultaneously
income and expenditure shocks, destroying Ex Ante versus Ex Post Assistance.—
livelihoods as well as property. They can there- These results underscore a key challenge
fore have longer-term effects (see, for exam- with ex post emergency aid: delays in the
ple, Baez and Santos 2007, A ­ nttila-Hughes time it takes for aid to arrive. To make this
and Hsiang 2013, Cas et al. 2014). process smoother, programs can be set up in
advance and triggered based on prespecified
Ex Post Emergency Relief.—After these thresholds. For example, del Valle, de Janvry,
shocks occur, governments tend to intervene and Sadoulet (2020) study Mexico’s Fonden
and provide some amount of emergency program, which provides transfers to munic-
relief to households, either in cash or in-kind, ipalities when rainfall exceeds a prespecified
though the degree to which they do so varies. threshold. Using a regression discontinuity
There is some evidence that this type of design based on the rainfall cutoff, they show
compensation can make an important differ- that insurance payments lead to substan-
ence. For example, Gignoux and Menéndez tially more economic activity as measured
(2016) study earthquakes in Indonesia and by night-lights. Subsequent work shows that
find that households are harmed by these this program has also reduced mortality by
shocks, but that the effects dissipate after restoring access to public health infrastruc-
two to five years. They find that government ture more quickly (del Valle 2024). Clarke
aid flows following the earthquakes may and Dercon (2016) also discuss Fonden and
have helped mitigate the effects. Likewise, other programs that similarly provide index-
Laurito, Frankenberg, and Thomas (2022) based drought insurance (e.g., a program for
find substantial positive effects of housing pastoralists in Kenya).
assistance in Indonesia following the 2004 More generally, there is increasing pol-
Indian Ocean tsunami. icy discussion around the idea of codifying,
However, there is evidence that if aid is not ex ante, that there would be expansions in
timely enough, children—who may be in a the types of social protection programs dis-
critical window when the shock occurs—can cussed in section 3 in the event of shocks (see
suffer long-term consequences. For exam- O’Brien et al. 2018, for a discussion). For
ple, Andrabi, Daniels, and Das (2023) find example, the government of Kenya explic-
that, while adults affected by the Kashmir itly relaxes eligibility rules for its food secu-
earthquake suffered no change four years rity program in times of drought to make it
later (perhaps because they received trans- more responsive (Gardner et al. 2017). All
fers equal to 150 percent of their annual con- households in the four counties covered by
sumption), there were persistent shortfalls in the Hunger Safety Net Program were regis-
the physical and cognitive development of tered during the expansion of this program,
children who were young at the time of the regardless of eligibility status, to facilitate
earthquake. Likewise, Dercon and Porter efforts to rapidly scale up the transfer during
(2014) find long-run negative consequences emergencies. This has allowed 50 percent to
for child survivors of the 1984 Ethiopian 75 percent of the population in these coun-
ties to receive cash transfers during droughts.
42 Note that there are also quite substantial interna-
But this program was unusual in that
tional efforts as well, such as the Caribbean Catastrophe all households were registered ex ante, in
Risk Insurance Facility and Africa Risk Capacity. order to facilitate expansion during times of
 Banerjee et al.: Social Protection in the Developing World 1407

shocks. More generally, programs must be if individuals were offered contracts priced
able to rapidly identify who needs assistance fairly for the entire population, take-up may
because of the shock, accounting for the fact be low.
that who needs assistance may have changed Moreover, the existence of these substan-
since the registration period due to the shock tial uninsured risks has broader implications
itself. for the design of social protection systems as
Given the chance that governments may a whole. Recall the poverty traps example
provide assistance ex post (as it is difficult to discussed in section 3.2.6. A further impli-
stand by and do nothing), households may be cation of the poverty trap model is that for
reluctant to purchase insurance themselves, households in the “good” equilibrium, an
a classic case of a Samaritan’s dilemma. If uninsured negative shock can push them
governments anticipate the need to bail out across the threshold into the bad equilib-
households ex post, they may be interested in rium, where they may spiral down into fur-
formalizing this commitment, allowing them ther poverty. This implies that solving the
to recoup some of their costs via mandatory insurance challenges discussed here may
premiums. One of the few examples we know have spillover effects, reducing the number
of in low- and middle-income countries is of long-run poor who need the redistribu-
the Turkish compulsory earthquake insur- tive programs discussed in section 3. Indeed,
ance scheme, established in 2000 (Natural Ikegami et al. (2018) and Janzen et al. (2021)
Disaster Insurance Institution 2022). argue that, for this reason, governments
A new frontier in this type of assistance is may be better off diverting some of their
to not just commit ex ante to deliver assis- standard targeted social protection funds to
tance, but to actually deliver assistance before providing additional social insurance. Even
shocks actually occur. A new wave of rapid to the extent that these programs do not
detection technologies mean that for some entirely mitigate financial shocks, condition-
predictable shocks (e.g., floods in the down- ality can direct households to smooth them
stream part of a river), aid could be delivered in ways that are less socially costly, such as by
before the shock is realized, allowing house- encouraging them to keep children in school
holds to take preventative action. Pople et al. (de Janvry et al. 2006). This approach can
(2021) examine a pilot study of this approach have protective intergenerational effects,
in Bangladesh and find that such ex ante preventing a shock in one generation from
assistance can make a difference. But doing creating a long-run cycle of poverty.
so could mean there is a risk of delivering aid Conversely, to the extent that targeting
to those who do not end up experiencing the can be made more dynamic, the same types
shock; that is, the realization of the shock is of programs discussed in section 3 can pro-
also a targeting device. vide some insurance. A challenge with many
of the targeting approaches discussed in
4.5 Summing Up, and Broader Implications
section 2 is that the targeting list is updated
for Social Protection System Design
infrequently. When targeting is done, it often
Life in developing countries is risky. While focuses on the permanent component of
insuring against risk is a challenge every- income (e.g., assets), making it unresponsive
where, the particular challenges in low- and to shocks.
middle-income countries highlighted This issue can be remedied in several
above—low insurance demand, low trust, ways. At the individual level, to the extent
liquidity constraints, challenges of state ver- that targeting can be made more responsive
ification, and informality—mean that even (for example, by using some combination of
1408 Journal of Economic Literature, Vol. LXII (December 2024)

high-frequency administrative data like elec- provide assistance to the poor, and social
tricity and mobile phone usage, ­on-demand insurance of a variety of types, than they
applications, and ­
community-driven were 50 or even 20 years ago.
approaches to verify shocks), the same pro- But the fundamental nature of these
grams that provide assistance to the poor can countries’ economies—e.g., the large infor-
also provide assistance to those who receive mal sector and the deep absolute poverty
shocks. This can help fill some of the gaps in level—means that the way these programs
insurance highlighted in this section. are designed fundamentally differs from how
At the aggregate level, as discussed in sec- they are designed in high-income countries.
tion 4.4, governments can use the types of One important difference is identifying
programs discussed in section 3 to respond beneficiaries. In high-income countries,
to aggregate shocks by changing eligibility governments typically choose beneficiaries
thresholds or by increasing transfer amounts. using income-based criteria. They can do
During the COVID-19 crisis, when this was this because income data is reported through
an extreme issue, many countries responded a third party, the tax system. With the
in creative ways to adapt their existing pro- large informal sector, this is not possible in
grams (many of which were of the type dis- ­lower-income countries. Instead, countries
cussed in section 3) to address these issues. face a trade-off between using noisy prox-
Indonesia, for example, both relaxed eligibil- ies for income and using potentially biased
ity rules and expanded transfer amounts in self-reports. We discuss how to think about
its food voucher program. It also created a this trade-off theoretically, and then review
new community-targeted cash transfer pro- the evidence on the trade-offs between these
gram in all rural villages. Likewise, Pakistan approaches in practice.
built on its flagship cash transfer program Once one figures out who to provide assis-
for women, expanding eligibility and using tance to, what should the form of assistance
cell phone metadata to help determine eli- look like? Again, the particular context of
gibility (Lone, Shakeel, and Bischler 2020). low- and middle-income countries has influ-
These examples were forged rapidly in a cri- enced how these programs look. For exam-
sis. Presumably, with advance planning, gov- ple, given the relatively low levels of human
ernments can do much more to make their capital investment, conditional cash trans-
existing redistribution programs function to fers—which tie cash payments to ensuring
smooth shocks. that family members of recipients meet a set
of health and educational conditions—are
far more common in low- and middle-in-
5. Conclusion
come countries than in high-income coun-
Social protection programs are becom- tries. Indeed, these programs have been
ing increasingly prominent in low- and shown to have substantial effects on the
middle-income countries worldwide. This incentivized behaviors. Given how poor the
is partly due to the fact that countries are poor can be, there is also a focus on finding
becoming richer, so some countries that program designs that can help break poverty
were recently quite poor are now middle traps, whether within a single generation or
income, and with substantially larger tax across generations.
bases that can fund more sophisticated sys- Finally, the substantial informal sector
tems. It also reflects the fact that, condi- has important implications for the design of
tional on a country’s income level, countries social insurance schemes. Unemployment
are more likely to have systems in place that insurance, for example, cannot easily be
 Banerjee et al.: Social Protection in the Developing World 1409

of Conditionality. World Bank Policy Research Work-

conditioned on remaining unemployed if ing Paper 6340.
one can simply find informal employment Akresh, Richard, Damien de Walque, and Harounan
instead. Likewise, health insurance can- Kazianga. 2016. “Evidence from a Randomized Eval-
not be provided through employer-based uation of the Household Welfare Impacts of Condi-
mandates and pensions cannot be funded tional and Unconditional Cash Transfers Given to
Mothers or Fathers.” World Bank Policy Research
through payroll tax systems for the vast num- Working Paper 7730.
ber of employees who are not in the formal Alatas, Vivi. 2011. “Program Keluarga Harapan: Main
sector. And the fact that informality remains Findings from the Impact Evaluation of Indonesia’s
an option for many businesses means that Pilot Household Conditional Cash Transfer Pro-
governments must tread carefully before gram.” World Bank Group Working Paper 72506.
Alatas, Vivi, Abhijit Banerjee, Arun G. Chandrasekhar,
imposing too many such mandates on firms, Rema Hanna, and Benjamin A. Olken. 2016. “Net-
lest they increase informality rates. work Structure and the Aggregation of Information:
Despite the challenges involved in deliv- Theory and Evidence from Indonesia.” American
ering these programs, there has been tre- Economic Review 106 (7): 1663–1704.
mendous growth in both social assistance Alatas, Vivi, Abhijit Banerjee, Rema Hanna, Benjamin
A. Olken, Ririn Purnamasari, and Matthew Wai-Poi.
programs for redistribution as well as social 2016. Data and Code for: “Self-Targeting: Evidence
insurance in low- and middle-income coun- from a Field Experiment in Indonesia.” Distributed
tries. This trend suggests that increasing by MIT Department of Economics, Cambridge,
attention to the unique challenges of devel- MA. https://economics.mit.edu/sites/default/files/
oping social protection in these contexts is publications/ST_final_data_final.zip (accessed Feb-
ruary 13, 2024).
likely to be an important area for ongoing Alatas, Vivi, Abhijit Banerjee, Rema Hanna, Benjamin
research in the years to come. A. Olken, Ririn Purnamasari, and Matthew Wai-Poi.
2019. “Does Elite Capture Matter? Local Elites
References and Targeted Welfare Programs in Indonesia.” AEA
Papers and Proceedings 109: 334–39.
Abelson, Brian, Kush R. Varshney, and Joy Sun. 2014.
Alatas, Vivi, Abhijit Banerjee, Rema Hanna, Benja-
“Targeting Direct Cash Transfers to the Extremely
min A. Olken, and Julia Tobias. 2012. “Targeting the
Poor.” In Proceedings of the 20th ACM SIGKDD
Poor: Evidence from a Field Experiment in Indone-
International Conference on Knowledge Discovery
sia.” American Economic Review 102 (4): 1206–40.
and Data Mining, 1563–72. New York: Association
Alatas, Vivi, Ririn Purnamasari, Matthew Wai-Poi,
for Computing Machinery.
Abhijit Banerjee, Benjamin A. Olken, and Rema
Agüero, Jorge M., Michael R. Carter, and Ingrid
Hanna. 2016. “Self-Targeting: Evidence from a Field
Woolard. 2006. “The Impact of Unconditional Cash
Experiment in Indonesia.” Journal of Political Econ-
Transfers on Nutrition: The South African Child
omy 124 (2): 371–427.
Support Grant.” Southern Africa Labour and Devel-
Albagli, Pinjas, and Tomás Rau. 2019. “The Effects of
opment Research Unit Working Paper 06/08.
a Maternity Leave Reform on Children’s Abilities
Aiken, Emily, Suzanne Bellue, Dean Karlan, Chris
and Maternal Outcomes in Chile.” Economic Journal
Udry, and Joshua Blumenstock. 2022. “Machine
129 (619): 1015–47.
Learning and Phone Data Can Improve Targeting of
Albarran, Pedro, and Orazio P. Attanasio. 2003. “Lim-
Humanitarian Aid.” Nature 603 (7903): 864–70.
ited Commitment and Crowding Out of Private
Aizer, Anna, Shari Eli, Adriana Lleras-Muney, and
Transfers: Evidence from a Randomised Experi-
Keyoung Lee. 2020. “Do Youth Employment Pro-
ment.” Economic Journal 113 (486): C77–85.
grams Work? Evidence from the New Deal.” NBER
Alderman, Harold. 2002. “Do Local Officials Know
Working Paper 27103.
Something We Don’t? Decentralization of Targeted
Akerlof, George A. 1978. “The Economics of ‘Tagging’
Transfers in Albania.” Journal of Public Economics
as Applied to the Optimal Income Tax, Welfare Pro-
83 (3): 375–404.
grams, and Manpower Planning.” American Eco-
Ali, Williams, Awudu Abdulai, and Ashok K. Mishra.
nomic Review 68 (1): 8–19.
2020. “Recent Advances in the Analyses of Demand
Akresh, Richard, Damien de Walque, and Harounan
for Agricultural Insurance in Developing and
Kazianga. 2013. Cash Transfers and Child Schooling:
Emerging Countries.” Annual Review of Resource
Evidence from a Randomized Evaluation of the Role
1410 Journal of Economic Literature, Vol. LXII (December 2024)

Economics 12: 411–30. 2014. “The Design of a Multi-Tier Contributory Pen-

Andrabi, Tahir, Benjamin Daniels, and Jishnu Das. sion System: The Distributional Impact of the 2008
2023. “Human Capital Accumulation and Disasters: Chilean Pension Reform.” In Social Insurance, Infor-
Evidence from the Pakistan Earthquake of 2005.” mality, and Labor Markets: How to Protect Workers
Journal of Human Resources 58 (4): 1057–96. While Creating Good Jobs, edited by Markus Frölich,
Angelucci, Manuela, and Giacomo De Giorgi. 2009. David Kaplan, Carmen Pagés, Jamele Rigolini, and
“Indirect Effects of an Aid Program: How Do Cash David Robalino, 291–346. Oxford, UK: Oxford Uni-
Transfers Affect Ineligibles’ Consumption?” Ameri- versity Press.
can Economic Review 99 (1): 486–508. Attanasio, Orazio, and José-Vıctor Rıos-Rull. 2000.
Anttila-Hughes, Jesse, and Solomon Hsiang. 2013. “Consumption Smoothing in Island Economies: Can
“Destruction, Disinvestment, and Death: Economic Public Insurance Reduce Welfare?” European Eco-
and Human Losses Following Environmental Disas- nomic Review 44 (7): 1225–58.
ter.” http://dx.doi.org/10.2139/ssrn.2220501. Augenblick, Ned, Kelsey Jack, Supreet Kaur, Felix
Araujo, María Caridad, and Karen Macours. 2021. Masiye, and Nicholas Swanson. 2022. “Budget
“Education, Income and Mobility: Experimental Neglect in Consumption Smoothing: A Field Experi-
Impacts of Childhood Exposure to Progresa after 20 ment on Seasonal Hunger.” Unpublished.
Years.” Paris School of Economics Working Paper Auriol, Emmanuelle, Julie Lassébie, Amma Panin,
­2021-57. Eva Raiber, and Paul Seabright. 2020. “God Insures
Ardington, Cally, Anne Case, and Victoria Hosegood. Those Who Pay? Formal Insurance and Religious
2009. “Labor Supply Responses to Large Social Offerings in Ghana.” Quarterly Journal of Economics
Transfers: Longitudinal Evidence from South 135 (4): 1799–1848.
Africa.” American Economic Journal: Applied Eco- Azam, Mehtabul. 2012. “The Impact of Indian Job
nomics 1 (1): 22–48. Guarantee Scheme on Labor Market Outcomes: Evi-
Areias, Ana, and Matthew Wai-Poi. 2022. “Machine dence from a Natural Experiment.” IZA Discussion
Learning and Prediction of Beneficiary Eligibility for Paper 6548.
Social Protection Programs.” In Revisiting Targeting Azuara, Oliver, and Ioana Marinescu. 2013. “Informal-
in Social Assistance: A New Look at Old Dilemmas, ity and the Expansion of Social Protection Programs:
edited by Margaret Grosh, Phillippe Leite, Matthew Evidence from Mexico.” Journal of Health Econom-
Wai-Poi, and Emil Tesliuc, 507–39. Washington, DC: ics 32 (5): 938–50.
World Bank. Baez, Javier E., and Indhira V. Santos. 2007. “Children’s
Arnott, Richard, and Joseph E. Stiglitz. 1991. “Moral Vulnerability to Weather Shocks: A Natural Disaster
Hazard and Nonmarket Institutions: Dysfunctional as a Natural Experiment.” https://conference.iza.org/
Crowding Out of Peer Monitoring?” American Eco- conference_files/chldc2007/baez_j3321.pdf.
nomic Review 81 (1): 179–90. Baez, Javier, Varun Kshirsagar, and Emmanuel Skou-
Asuming, Patrick O. 2013. “Getting the Poor to Enroll fias. 2019. “Adaptive Safety Nets for Rural Africa:
in Health Insurance, and Its Effects on Their Health: Drought-Sensitive Targeting with Sparse Data.”
Evidence from a Field Experiment in Ghana.” http:// World Bank Policy Research Working Paper 9071.
www.columbia.edu/~poa2102/jmp_asuming_2012. Baird, Sarah, Francisco H. G. Ferreira, Berk Özler,
pdf. and Michael Woolcock. 2014. “Conditional, Uncon-
Asuming, Patrick O., Hyuncheol B. Kim, and Armand ditional and Everything in Between: A Systematic
Sim. 2024. “Selection and Behavioral Responses of Review of the Effects of Cash Transfer Programmes
Health Insurance Subsidies in the Long Run: Evi- on Schooling Outcomes.” Journal of Development
dence from a Field Experiment in Ghana.” Health Effectiveness 6 (1): 1–43.
Economics 33 (5): 992–1032. Baird, Sarah, Craig McIntosh, and Berk Özler. 2011.
Aterido, Reyes, Mary Hallward-Driemeier, and Pagés “Cash or Condition? Evidence from a Cash Trans-
Carmen. 2011. “Big Constraints to Small Firms’ fer Experiment.” Quarterly Journal of Economics
Growth? Business Environment and Employment 126 (4): 1709–53.
Growth across Firms.” Economic Development and Baird, Sarah, David McKenzie, and Berk Özler. 2018.
Cultural Change 59 (3): 609–47. “The Effects of Cash Transfers on Adult Labor Mar-
Attanasio, Orazio, Luis Gómez, Patricia Heredia, and ket Outcomes.” IZA Journal of Development and
Marcos Vera-Hernández. 2005. “The Short-Term Migration 8: 22.
Impact of a Conditional Cash Subsidy on Child Baker, Judy L., and Margaret E. Grosh. 1994. “Poverty
Health and Nutrition in Colombia.” Institute for Fis- Reduction through Geographic Targeting: How Well
cal Studies Report Summary: Familias 03. Does It Work?” World Development 22 (7): 983–95.
Attanasio, Orazio, Costas Meghir, and Andres Otero. Balboni, Clare, Oriana Bandiera, Robin Burgess,
 Banerjee et al.: Social Protection in the Developing World 1411

Maitreesh Ghatak, and Anton Heil. 2022. “Why Do Banerjee, Abhijit, Rema Hanna, Jordan Kyle, Benjamin
People Stay Poor?” Quarterly Journal of Economics A. Olken, and Sudarno Sumarto. 2018. “Tangible
137 (2): 785–844. Information and Citizen Empowerment: Identifica-
Ball, Ian. 2023. “Scoring Strategic Agents.” https://doi. tion Cards and Food Subsidy Programs in Indone-
org/10.48550/arXiv.1909.01888. sia.” Journal of Political Economy 126 (2): 101–37.
Bandiera, Oriana, Robin Burgess, Narayan Das, Selim Banerjee, Abhijit, Rema Hanna, Jordan Kyle, Benja-
Gulesci, Imran Rasul, and Munshi Sulaiman. 2017. min A. Olken, and Sudarno Sumarto. 2019. “Private
“Labor Markets and Poverty in Village Economies.” Outsourcing and Competition: Subsidized Food Dis-
Quarterly Journal of Economics 132 (2): 811–70. tribution in Indonesia.” Journal of Political Economy
Banerjee, Abhijit, Emily Breza, Esther Duflo, and Cyn- 127 (1): 101–37.
thia Kinnan. 2019. “Can Microfinance Unlock a Pov- Banerjee, Abhijit, Rema Hanna, Benjamin A. Olken,
erty Trap for Some Entrepreneurs?” NBER Working Elan Satriawan, and Sudarno Sumarto. 2023. “Elec-
Paper 26346. tronic Food Vouchers: Evidence from an At-Scale
Banerjee, Abhijit, and Esther Duflo. 2010. “Giving Experiment in Indonesia.” American Economic
Credit Where It Is Due.” Journal of Economic Per- Review 113 (2): 514–47.
spectives 24 (3): 61–80. Banerjee, Abhijit, Rema Hanna, Benjamin A. Olken,
Banerjee, Abhijit, Esther Duflo, Raghabendra Chat- and Sudarno Sumarto. 2020. “The (Lack of) Distor-
topadhyay, and Jeremy Shapiro. 2016. “The Long tionary Effects of Proxy-Means Tests: Results from
Term Impacts of a ‘Graduation’ Program: Evidence a Nationwide Experiment in Indonesia.” Journal of
from West Bengal.” https://www.povertyactionlab. Public Economics Plus 1: 100001.
org/sites/default/files/research-paper/working- Banerjee, Abhijit, Dean Karlan, Robert Osei, Hannah
paper_92_Long-Term-Effects-of-Graduation- Trachtman, and Christopher Udry. 2022. “Unpack-
Approach_India_Sept2016.pdf. ing a Multi-Faceted Program to Build Sustainable
Banerjee, Abhijit, Esther Duflo, Nathanael Goldberg, Income for the Very Poor.” Journal of Development
Dean Karlan, Robert Osei, Parienté William, Jeremy Economics 155: 102781.
Shapiro, Bram Thuysbaert, and Christopher Udry. Banerjee, Abhijit, and Sendhil Mullainathan. 2010.
2015. “A Multifaceted Program Causes Lasting Prog- “The Shape of Temptation: Implications for the
ress for the Very Poor: Evidence from Six Countries.” Economic Lives of the Poor.” NBER Working Paper
Science 348 (6236): 1260799. 15973.
Banerjee, Abhijit, Esther Duflo, and Richard Horn- Banerjee, Abhijit, and Andrew Newman. 1998. “Infor-
beck. 2018. “How Much Do Existing Borrowers mation, the Dual Economy, and Development.”
Value Microfinance? Evidence from an Experiment Review of Economic Studies 65 (4): 631–53.
on Bundling Microcredit and Insurance.” Economica Banerjee, Abhijit, Paul Niehaus, and Tavneet Suri.
85 (340): 671–700. 2019. “Universal Basic Income in the Developing
Banerjee, Abhijit, Esther Duflo, Clément Imbert, San- World.” Annual Review of Economics 11: 959–83.
thosh Mathew, and Rohini Pande. 2020. “E-Gov- Bardhan, Pranab, and Dilip Mookherjee. 2000. “Cap-
ernance, Accountability, and Leakage in Public ture and Governance at Local and National Levels.”
Programs: Experimental Evidence from a Financial American Economic Review 90 (2): 135–39.
Management Reform in India.” American Economic Barham, Tania, Karen Macours, and John A. Maluccio.
Journal: Applied Economics 12 (4): 39–72. 2013. “Boys’ Cognitive Skill Formation and Physi-
Banerjee, Abhijit, Esther Duflo, and Garima Sharma. cal Growth: Long-Term Experimental Evidence on
2021. “Long-Term Effects of the Targeting the Critical Ages for Early Childhood Interventions.”
Ultra Poor Program.” American Economic Review: American Economic Review 103 (3): 467–71.
Insights 3 (4): 471–86. Barham, Tania, Karen Macours, and John A. Maluccio.
Banerjee, Abhijit, Amy Finkelstein, Rema Hanna, 2018a. “Are Conditional Cash Transfers Fulfilling
Benjamin A. Olken, Arianna Ornaghi, and Sudarno Their Promise? Schooling, Learning, and Earnings
Sumarto. 2021. “The Challenges of Universal Health after 10 Years.” CEPR Discussion Paper 11937.
Insurance in Developing Countries: Experimental Barham, Tania, Karen Macours, and John A. Maluccio.
Evidence from Indonesia’s National Health Insur- 2018b. “Experimental Evidence of Exposure to a
ance.” American Economic Review 111 (9): 3035–63. Conditional Cash Transfer during Early Teenage
Banerjee, Abhijit, Rema Hanna, Gabriel Kreindler, and Years: Young Women’s Fertility and Labor Market
Benjamin A. Olken. 2017. “Debunking the Stereo- Outcomes.” CEPR Discussion Paper 13165.
type of the Lazy Welfare Recipient: Evidence from Barham, Tania, and John A. Maluccio. 2009. “Eradicat-
Cash Transfer Programs.” World Bank Research ing Diseases: The Effect of Conditional Cash Trans-
Observer 32 (2): 155–84. fers on Vaccination Coverage in Rural Nicaragua.”
1412 Journal of Economic Literature, Vol. LXII (December 2024)

Journal of Health Economics 28 (3): 611–21. Policy 7 (3): 86–125.

Barrera-Osorio, Felipe, Marianne Bertrand, Leigh L. Berg, Erlend. 2018. “Funeral Insurance: An Inter-Gen-
Linden, and Francisco Perez-Calle. 2011. “Improv- erational Commitment Device?” Journal of African
ing the Design of Conditional Transfer Programs: Economies 27 (3): 321–46.
Evidence from a Randomized Education Experi- Berg, Erlend, Sambit Bhattacharyya, D. Rajasekhar,
ment in Colombia.” American Economic Journal: and R. Manjula. 2018. “Can Public Works Increase
Applied Economics 3 (2): 167–95. Equilibrium Wages? Evidence from India’s National
Barrera-Osorio, Felipe, Leigh L. Linden, and Juan E. Rural Employment Guarantee.” World Development
Saavedra. 2019. “Medium- and Long-Term Edu- 103: 239–54.
cational Consequences of Alternative Conditional Bergolo, Marcelo, and Guillermo Cruces. 2014. “Work
Cash Transfer Designs: Experimental Evidence from and Tax Evasion Incentive Effects of Social Insurance
Colombia.” American Economic Journal: Applied Programs: Evidence from an Employment-Based
Economics 11 (3): 54–91. Benefit Extension.” Journal of Public Economics 117:
Bastagli, Francesca, Jessica Hagen-Zanker, Luke ­211–28.
Harman, Valentina Barca, Georgina Sturge, Tanja Berkouwer, Susanna B., Pierre E. Biscaye, Eric Hsu,
Schmidt, and Luca Pellerano. 2016. Cash Transfers: Oliver W. Kim, Kenneth Lee, Edward Miguel, and
What Does the Evidence Say? London: Overseas Catherine Wolfram. 2021. “Money or Power? Finan-
Development Institute. cial Infrastructure and Optimal Policy.” NBER
Basu, Kaushik. 1996. “Relief Programs: When It May Working Paper 29086.
Be Better to Give Food Instead of Cash.” World Bertrand, Marianne, Bruno Crépon, Alicia Marguerie,
Development 24 (1): 91–96. and Patrick Premand. 2021. “Do Workfare Programs
Basurto, Maria Pia, Pascaline Dupas, and Jonathan Live Up to Their Promises? Experimental Evidence
Robinson. 2020. “Decentralization and Efficiency of from Côte d’Ivoire.” World Bank Policy Research
Subsidy Targeting: Evidence from Chiefs in Rural Working Paper 9611.
Malawi.” Journal of Public Economics 185: 104047. Bertrand, Marianne, Sendhil Mullainathan, and Doug-
Bau, Natalie. 2021. “Can Policy Change Culture? las Miller. 2003. “Public Policy and Extended Fami-
Government Pension Plans and Traditional Kinship lies: Evidence from Pensions in South Africa.” World
Practices.” American Economic Review 111 (6): Bank Economic Review 17 (1): 27–50.
1880–1917. Besley, Timothy, and Stephen Coate. 1992. “Work-
Beaman, Lori, Niall Keleher, Jeremy Magruder, and fare versus Welfare: Incentive Arguments for Work
Carly Trachtman. 2021. “Urban Networks and Tar- Requirements in Poverty-Alleviation Programs.”
geting: Evidence from Liberia.” AEA Papers and American Economic Review 82 (1): 249–61.
Proceedings 111: 572–76. Bhalotra, Sonia, Diogo G. C. Britto, Paolo Pinotti, and
Becker, Gary S., and George J. Stigler. 1974. “Law Breno Sampaio. 2021. “Job Displacement, Unem-
Enforcement, Malfeasance, and Compensation of ployment Benefits and Domestic Violence.” CEPR
Enforcers.” Journal of Legal Studies 3 (1): 1–18. Discussion Paper 16350.
Bedoya, Guadalupe, Aidan Coville, Johannes Haush- Bhattacharya, Debopam, and Pascaline Dupas. 2012.
ofer, Mohammad Isaqzadeh, and Jeremy P. Shap- “Inferring Welfare Maximizing Treatment Assign-
iro. 2019. “No Household Left Behind: Afghanistan ment under Budget Constraints.” Journal of Econo-
Targeting the Ultra Poor Impact Evaluation.” World metrics 167 (1): 168–96.
Bank Policy Research Working Paper 8877. Bhattacharya, Shrayana, Maria Mini Jos, Soumya
Behrman, Jere R., Susan W. Parker, and Petra E. Todd. Kapoor Mehta, and Rinku Murgai. 2015. “From
2011. “Do Conditional Cash Transfers for Schooling Policy to Practice: How Should Social Pensions Be
Generate Lasting Benefits? A Five-Year Followup Scaled Up?” Economic and Political Weekly 50 (14):
of PROGRESA/Oportunidades.” Journal of Human ­60–67.
Resources 46 (1): 93–122. Björkegren, Daniel, Joshua E. Blumenstock, and Sam-
Benedetti, Fiorella, Pablo Ibarrarán, and Patrick J. Mce- sun Knight. 2021. “Manipulation-Proof Machine
wan. 2016. “Do Education and Health Conditions Learning.” https://doi.org/10.48550/arXiv.2004.03865.
Matter in a Large Cash Transfer? Evidence from a Bjorvatn, Kjetil, Denise Ferris, Selim Gulesci, Arne
Honduran Experiment.” Economic Development and Nasgowitz, Vincent Somville, and Lore Vandewalle.
Cultural Change 64 (4): 759–93. 2022. “Childcare, Labor Supply, and Business Devel-
Benhassine, Najy, Florencia Devoto, Esther Duflo, Pas- opment: Experimental Evidence from Uganda.”
caline Dupas, and Victor Pouliquen. 2015. “Turning CEPR Discussion Paper 17243.
a Shove into a Nudge? A ‘Labeled Cash Transfer’ for Blattman, Christopher, Nathan Fiala, and Sebastian Mar-
Education.” American Economic Journal: Economic tinez. 2014. “Generating Skilled Self-Employment in
 Banerjee et al.: Social Protection in the Developing World 1413

Developing Countries: Experimental Evidence from Economic Review 110 (8): 2422–53.
Uganda.” Quarterly Journal of Economics 129 (2): Calderón, Gabriela. 2014. “The Effects of Child Care
697–752. Provision in Mexico.” Banco de México Working
Blattman, Christopher, Nathan Fiala, and Sebastian Paper 2014-07.
Martinez. 2020. “The Long-Term Impacts of Grants Camacho, Adriana, and Emily Conover. 2011. “Manip-
on Poverty: Nine-Year Evidence from Uganda’s ulation of Social Program Eligibility.” American Eco-
Youth Opportunities Program.” American Economic nomic Journal: Economic Policy 3 (2): 41–65.
Review: Insights 2 (3): 287–304. Camacho, Adriana, Emily Conover, and Alejandro
Blumenstock, Joshua, Gabriel Cadamuro, and Rob- Hoyos. 2014. “Effects of Colombia’s Social Protec-
ert On. 2015. “Predicting Poverty and Wealth from tion System on Workers’ Choice between Formal
Mobile Phone Metadata.” Science 350 (6264): and Informal Employment.” World Bank Economic
­1073–76. Review 28 (3): 446–66.
Bosch, Mariano, and Raymundo M. Campos-Vazquez. Cameron, Lisa, and Manisha Shah. 2014. “Can Mistar-
2014. “The Trade-Offs of Welfare Policies in Labor geting Destroy Social Capital and Stimulate Crime?
Markets with Informal Jobs: The Case of the ‘Seguro Evidence from a Cash Transfer Program in Indone-
Popular’ Program in Mexico.” American Economic sia.” Economic Development and Cultural Change
Journal: Economic Policy 6 (4): 71–99. 62 (2): 381–415.
Botea, Ioana, Markus Goldstein, Corinne Low, and Caria, A. Stefano, Grant Gordon, Maximilian Kasy,
Gareth Roberts. 2021. “Supporting Women’s Live- Simon Quinn, Soha Osman Shami, and Alexander
lihoods at Scale: RCT Evidence from a Nationwide Teytelboym. 2024. “An Adaptive Targeted Field
Graduation Program.” World Bank Working Paper Experiment: Job Search Assistance for Refugees in
175851. Jordan.” Journal of the European Economic Associa-
Britto, Diogo G. C., Paolo Pinotti, and Breno Sampaio. tion 22 (2): 781–836.
2022. “The Effect of Job Loss and Unemployment Carneiro, Pedro, Emanuela Galasso, and Rita Ginja.
Insurance on Crime in Brazil.” Econometrica 90 (4): 2019. “Tackling Social Exclusion: Evidence from
­1393–1423. Chile.” Economic Journal 129 (617): 172–208.
Brollo, Fernanda, Katja Maria Kaufmann, and Eliana Carney, Kevin, Michael Kremer, Xinyue Lin, and Gau-
La Ferrara. 2017. “Learning about the Enforcement tam Rao. 2022. “The Endowment Effect and Collat-
of Conditional Welfare Programs: Evidence from eralized Loans.” NBER Working Paper 30073.
Brazil.” IZA Discussion Paper 10654. Carter, Michael R., Alain de Janvry, Elisabeth Sadoulet,
Brown, Caitlin, Martin Ravallion, and Dominique van and Alexandros Sarris. 2017. “Index Insurance for
de Walle. 2018. “A Poor Means Test? Econometric Developing Country Agriculture: A Reassessment.”
Targeting in Africa.” Journal of Development Eco- Annual Review of Resource Economics 9: 421–38.
nomics 134: 109–24. Carter, Michael R., and Travis J. Lybbert. 2012. “Con-
Browning, Martin, and Pierre-André Chiappori. 1998. sumption versus Asset Smoothing: Testing the Impli-
“Efficient Intra-Household Allocations: A General cations of Poverty Trap Theory in Burkina Faso.”
Characterization and Empirical Tests.” Economet- Journal of Development Economics 99 (2): 255–64.
rica 66 (6): 1241–78. Carter, Michael R., Emilia Tjernström, and Patricia
Buera, Francisco J. 2009. “A Dynamic Model of Entre- Toledo. 2019. “Heterogeneous Impact Dynamics of
preneurship with Borrowing Constraints: Theory and a Rural Business Development Program in Nicara-
Evidence.” Annals of Finance 5 (3–4): 443–64. gua.” Journal of Development Economics 138: 77–98.
Buller, Ana Maria, Amber Peterman, Meghna Ran- Cas, Ava Gail, Elizabeth Frankenberg, Wayan Suri-
ganathan, Alexandra Bleile, Melissa Hidrobo, and astini, and Duncan Thomas. 2014. “The Impact of
Lori Heise. 2018. “A Mixed-Method Review of Cash Parental Death on Child Well-Being: Evidence from
Transfers and Intimate Partner Violence in Low- and the Indian Ocean Tsunami.” Demography 51 (2):
Middle-Income Countries.” World Bank Research ­437–57.
Observer 33 (2): 218–58. Casaburi, Lorenzo, and Jack Willis. 2018. “Time ver-
Cahyadi, Nur, Rema Hanna, Benjamin A. Olken, Rizal sus State in Insurance: Experimental Evidence from
Adi Prima, Elan Satriawan, and Ekki Syamsulhakim. Contract Farming in Kenya.” American Economic
2020. “Cumulative Impacts of Conditional Cash Review 108 (12): 3778–3813.
Transfer Programs: Experimental Evidence from Case, Anne, and Angus Deaton. 1998. “Large Cash
Indonesia.” American Economic Journal: Economic Transfers to the Elderly in South Africa.” Economic
Policy 12 (4): 88–110. Journal 108 (450): 1330–61.
Cai, Jing, Alain de Janvry, and Elisabeth Sadoulet. 2020. Case, Anne, Anu Garrib, Alicia Menendez, and Analia
“Subsidy Policies and Insurance Demand.” American Olgiati. 2013. “Paying the Piper: The High Cost of
1414 Journal of Economic Literature, Vol. LXII (December 2024)

Funerals in South Africa.” Economic Development Jayachandran. 2019. “The Price Effects of Cash Ver-
and Cultural Change 62 (1): 1–20. sus In-Kind Transfers.” Review of Economic Studies
Case, Anne, Victoria Hosegood, and Frances Lund. 86 (1): 240–81.
2005. “The Reach and Impact of Child Support Currie, Janet, and Firouz Gahvari. 2008. “Transfers in
Grants: Evidence from KwaZulu-Natal.” Develop- Cash and In-Kind: Theory Meets the Data.” Journal
ment Southern Africa 22 (4): 467–82. of Economic Literature 46 (2): 333–83.
Chernozhukov, Victor, Mert Demirer, Esther Duflo, Das, Jishnu, and Quy-Toan Do. 2023. “The Prices in
and Iván Fernández-Val. 2018. “Generic Machine the Crises: What We Are Learning from 20 Years
Learning Inference on Heterogeneous Treatment of Health Insurance in Low- and Middle-Income
Effects in Randomized Experiments, with an Appli- Countries.” Journal of Economic Perspectives 37 (2):
cation to Immunization in India.” NBER Working ­123–52.
Paper 24678. Dasgupta, Partha, and Debraj Ray. 1986. “Inequality as
Chetty, Raj. 2009a. “Is the Taxable Income Elasticity a Determinant of Malnutrition and Unemployment:
Sufficient to Calculate Deadweight Loss? The Impli- Theory.” Economic Journal 96 (384): 1011–34.
cations of Evasion and Avoidance.” American Eco- de Barros, Ricardo Paes, Pedro Olinto, Trine Lunde,
nomic Journal: Economic Policy 1 (2): 31–52. and Mirela Carvalho. 2011. “The Impact of Access
Chetty, Raj. 2009b. “Sufficient Statistics for Welfare to Free Childcare on Women’s Labor Market Out-
Analysis: A Bridge Between Structural and Reduced- comes: Evidence from a Randomized Trial in
Form Methods.” Annual Review of Economics 1: Low-Income Neighborhoods of Rio de Janeiro.”
­451–88. https://www.projectenportfolio.nl/images/archive/
Chetty, Raj, and Adam Looney. 2007. “Income Risk 1/1b/20231013183812%21Barros-Olinto_creche_
and the Benefits of Social Insurance: Evidence from rio_2011.pdf.
Indonesia and the United States.” In Fiscal Policy de Carvalho Filho, Irineu Evangelista. 2012. “House-
and Management in East Asia, Vol. 16, edited by hold Income as a Determinant of Child Labor and
Takatoshi Ito and Andrew K. Rose, 99–121. Chicago: School Enrollment in Brazil: Evidence from a Social
University of Chicago Press. Security Reform.” Economic Development and Cul-
Clark, Shelley, Caroline W. Kabiru, Sonia Laszlo, and tural Change 60 (2): 399–435.
Stella Muthuri. 2019. “The Impact of Childcare on de Janvry, Alain, Frederico Finan, Elisabeth Sadoulet,
Poor Urban Women’s Economic Empowerment in and Renos Vakis. 2006. “Can Conditional Cash Trans-
Africa.” Demography 56 (4): 1247–72. fer Programs Serve as Safety Nets in Keeping Chil-
Clarke, Daniel J., and Stefan Dercon. 2016. Dull Disas- dren at School and From Working When Exposed to
ters? How Planning Ahead Will Make a Difference. Shocks?” Journal of Development Economics 79 (2):
Oxford, UK: Oxford University Press. ­349–73.
Coady, David, Margaret Grosh, and John Hoddinott. Dean, Emma Boswell, Frank Schilbach, and Heather
2004a. Targeting of Transfers in Developing Coun- Schofield. 2017. “Poverty and Cognitive Function.”
tries: Review of Lessons and Experience. Washing- In The Economics of Poverty Traps, edited by Chris-
ton, DC: World Bank. topher B. Barrett, Michael Carter, Jean-Paul Chavas,
Coady, David, Margaret Grosh, and John Hoddinott. and Michael R. Carter, 57–118. Chicago: University
2004b. “Targeting Outcomes Redux.” World Bank of Chicago Press.
Research Observer 19 (1): 61–85. del Valle, Alejandro. 2021. “The Effects of Public
Coate, Stephen, Stephen Johnson, and Richard Zeck- Health Insurance in Labor Markets with Informal
hauser. 1994. “Pecuniary Redistribution through Jobs: Evidence from Mexico.” Journal of Health Eco-
In-Kind Programs.” Journal of Public Economics nomics 77: 102454.
55 (1): 19–40. del Valle, Alejandro. 2024. “Saving Lives with Indexed
Cole, Shawn, Daniel Stein, and Jeremy Tobacman. Disaster Funds: Evidence from Mexico.” American
2014. “Dynamics of Demand for Index Insurance: Economic Journal: Economic Policy 16 (2): 442–79.
Evidence from a Long-Run Field Experiment.” del Valle, Alejandro, Alain de Janvry, and Elisabeth
American Economic Review 104 (5): 284–90. Sadoulet. 2020. “Rules for Recovery: Impact of
Cole, Shawn, and Wentao Xiong. 2017. “Agricultural Indexed Disaster Funds on Shock Coping in Mex-
Insurance and Economic Development.” Annual ico.” American Economic Journal: Applied Econom-
Review of Economics 9: 235–62. ics 12 (4): 164–95.
Cunha, Jesse M. 2014. “Testing Paternalism: Cash ver- Dercon, Stefan, and Catherine Porter. 2014. “Live Aid
sus In-Kind Transfers.” American Economic Journal: Revisited: Long-Term Impacts of the 1984 Ethiopian
Applied Economics 6 (2): 195–230. Famine on Children.” Journal of the European Eco-
Cunha, Jesse M., Giacomo De Giorgi, and Seema nomic Association 12 (4): 927–48.
 Banerjee et al.: Social Protection in the Developing World 1415

Dercon, Stefan, Jan Willem Gunning, and Andrew Experimental Evidence from Tanzania.” World Bank
Zeitlin. 2019. “The Demand for Insurance under Economic Review 37 (2): 221–34.
Limited Trust: Evidence from a Field Experiment in Evans, David K., and Anna Popova. 2016. “What Really
Kenya.” Centre for the Study of African Economies Works to Improve Learning in Developing Coun-
Working Paper 2019-06. tries?: An Analysis of Divergent Findings in System-
Dercon, Stefan, Joachim De Weerdt, Tessa Bold, and atic Reviews.” World Bank Research Observer 31 (2):
Alula Pankhurst. 2006. “Group-Based Funeral Insur- ­242–70.
ance in Ethiopia and Tanzania.” World Development Farooq, Ammar, Adriana D. Kugler, and Umberto
34 (4): 685–703. Muratori. 2020. “Do Unemployment Insurance
Duflo, Esther. 2003. “Grandmothers and Granddaugh- Benefits Improve Match and Employer Quality?
ters: Old-Age Pensions and Intrahousehold Alloca- Evidence from Recent U.S. Recessions.” NBER
tion in South Africa.” World Bank Economic Review Working Paper 27574.
17 (1): 1–25. Feldstein, Martin. 1999. “Tax Avoidance and the Dead-
Dupas, Pascaline, Marcel Fafchamps, and Deivy weight Loss of the Income Tax.” Review of Econom-
Houeix. 2022. “Measuring Relative Poverty through ics and Statistics 81 (4): 674–80.
Peer Rankings: Evidence from Côte D’Ivoire.” Feldstein, Martin, and Daniel Altman. 2007. “Unem-
NBER Working Paper 29911. ployment Insurance Savings Accounts.” Tax Policy
Dupas, Pascaline, Vivian Hoffmann, Michael Kremer, and the Economy 21: 35–63.
and Alix Peterson Zwane. 2016. “Targeting Health Field, Erica, Rohini Pande, Natalia Rigol, Simone Scha-
Subsidies through a Nonprice Mechanism: A ner, and Charity Troyer Moore. 2021. “On Her Own
Randomized Controlled Trial in Kenya.” Science Account: How Strengthening Women’s Financial
353 (6302): 889–95. Control Impacts Labor Supply and Gender Norms.”
Dutta, Puja, Rinku Murgai, Martin Ravallion, and American Economic Review 111 (7): 2342–75.
Dominique van de Walle. 2014. Right to Work? Filmer, Deon, Jed Friedman, Eeshani Kandpal, and
Assessing India’s Employment Guarantee Scheme in Junko Onishi. 2023. “Cash Transfers, Food Prices,
Bihar. Washington, DC: World Bank. and Nutrition Impacts on Ineligible Children.”
Edmonds, Eric V. 2006. “Child Labor and Schooling Review of Economics and Statistics 105 (2): 327–43.
Responses to Anticipated Income in South Africa.” Finkelstein, Amy. 2007. “The Aggregate Effects of
Journal of Development Economics 81 (2): 386–414. Health Insurance: Evidence from the Introduc-
Egger, Dennis, Johannes Haushofer, Edward Miguel, tion of Medicare.” Quarterly Journal of Economics
Paul Niehaus, and Michael Walker. 2022. “General 122 (1): 1–37.
Equilibrium Effects of Cash Transfers: Experimental Finkelstein, Amy, Nathaniel Hendren, and Mark
Evidence from Kenya.” Econometrica 90 (6): 2­ 603–43. Shepard. 2019. “Subsidizing Health Insurance for
Einav, Liran, Amy Finkelstein, Stephen P. Ryan, Paul Low-Income Adults: Evidence from Massachusetts.”
Schrimpf, and Mark Cullen. 2013. “Selection on American Economic Review 109 (4): 1530–67.
Moral Hazard in Health Insurance.” American Eco- Finkelstein, Amy, and Robin Mcknight. 2008. “What
nomic Review 103 (1): 178–219. Did Medicare Do? The Initial Impact of Medicare
Elbers, Chris, Tomoki Fujii, Peter Lanjouw, Berk on Mortality and out of Pocket Medical Spending.”
Özler, and Wesley Yin. 2007. “Poverty Alleviation Journal of Public Economics 92 (7): 1644–68.
through Geographic Targeting: How Much Does Finkelstein, Amy, and Matthew J. Notowidigdo. 2019.
Disaggregation Help?” Journal of Development Eco- “Take-Up and Targeting: Experimental Evidence
nomics 83 (1): 198–213. from SNAP.” Quarterly Journal of Economics
Elbers, Chris, Jean O. Lanjouw, and Peter Lanjouw. 134 (3): 1505–56.
2003. “Micro-Level Estimation of Poverty and Fischer, Torben, Markus Frölich, and Andreas Land-
Inequality.” Econometrica 71 (1): 355–64. mann. 2023. “Adverse Selection in Low-Income
Evans, David K., Stephanie Hausladen, Katrina Kosec, Health Insurance Markets: Evidence from an RCT
and Natasha Reese. 2014. Community-Based Con- in Pakistan.” American Economic Journal: Applied
ditional Cash Transfers in Tanzania: Results from a Economics 15 (3): 313–40.
Randomized Trial. Washington, DC: World Bank. Franklin, Simon, Clément Imbert, Girum Abebe, and
Evans, David K., Brian Holtemeyer, and Katrina Kosec. Carolina Mejia-Mantilla. 2024. “Urban Public Works
2019. “Cash Transfers and Health: Evidence from in Spatial Equilibrium: Experimental Evidence from
Tanzania.” World Bank Economic Review 33 (2): Ethiopia.” American Economic Review 114 (5):
­394–412. ­1382–1414.
Evans, David K., and Katrina Kosec. 2023. “Cash Gadenne, Lucie, Samuel Norris, Monica Singhal,
Transfers, Trust, and Inter-Household Transfers: and Sandip Sukhtankar. 2024. “In-Kind Transfers
1416 Journal of Economic Literature, Vol. LXII (December 2024)

as Insurance.” American Economic Review 114 (9): Economics 38 (2): 297–321.

­2861–97. Glewwe, Paul, and Pedro Olinto. 2004. “Evaluating the
Galasso, Emanuela, and Martin Ravallion. 2005. Impact of Conditional Cash Transfers on Schooling:
“Decentralized Targeting of an Antipoverty Pro- An Experimental Analysis of Honduras’ PRAF Pro-
gram.” Journal of Public Economics 89 (4): 705–27. gram.” https://pdf.usaid.gov/pdf_docs/pnadt588.pdf.
Galiani, Sebastian, Paul J. Gertler, and Rosangela Goldberg, Jessica. 2016. “Kwacha Gonna Do? Exper-
Bando. 2016. “Non-Contributory Pensions.” Labour imental Evidence about Labor Supply in Rural
Economics 38: 47–58. Malawi.” American Economic Journal: Applied Eco-
Galiani, Sebastian, and Patrick J. McEwan. 2013. “The nomics 8 (1): 129–49.
Heterogeneous Impact of Conditional Cash Trans- González-Rozada, Martín, and Hernán Ruffo. 2023.
fers.” Journal of Public Economics 103: 85–96. “The Welfare Effects of Unemployment Insurance
Gardner, Clare, Caroline Riungu, Clare O’Brien, and in Argentina: New Estimates Using Changes in the
Fred Merttens. 2017. Evaluation of the Hunger Schedule of Transfers.” Journal of Human Resources.
Safety Net Programme Phase 2: The Legacy of HSNP https://doi.org/10.3368/jhr.0622-12372R1.
Phase 2: Systems, Practices and Lessons Learned. Gruber, Jonathan, Nathaniel Hendren, and Robert M.
Oxford, UK: Oxford Policy Management. Townsend. 2014. “The Great Equalizer: Health Care
Gentilini, Ugo, Mohamed Almenfi, Hrishikesh Iyengar, Access and Infant Mortality in Thailand.” American
Yuko Okamura, John Austin Downes, Pamela Dale, Economic Journal: Applied Economics 6 (1): 91–107.
Michael Weber, et al. 2022. “Social Protection and Gruber, Jonathan, Mengyun Lin, and Junjian Yi. 2023.
Jobs Responses to COVID-19: A Real-Time Review “The Largest Insurance Program in History: Saving
of Country Measures.” World Bank Working Paper One Million Lives Per Year in China.” Journal of
168386. Public Economics 226: 104999.
Gerard, François, and Gustavo Gonzaga. 2021. “Infor- Gupta, Sarika. 2017. “Perils of the Paperwork: The
mal Labor and the Efficiency Cost of Social Pro- Impact of Information and Application Assistance
grams: Evidence from Unemployment Insurance in on Welfare Program Take-Up in India.” https://
Brazil.” American Economic Journal: Economic Pol- www.povertyactionlab.org/sites/default/files/
icy 13 (3): 167–206. research-paper/Perils-of-the-Paperwork_Gupta_
Gerard, François, and Joana Naritomi. 2021. “Job Jobmarketpaper_November2017.pdf.
Displacement Insurance and (the Lack of) Con- Halim, Daniel, Hillary C. Johnson, and Elizaveta
sumption-Smoothing.” American Economic Review Perova. 2022. “Preschool Availability and Women’s
111 (3): 899–942. Employment: Evidence from Indonesia.” Economic
Gerard, François, Joana Naritomi, and Joana Silva. Development and Cultural Change 71 (1): 39–61.
2021. “Cash Transfers and Formal Labor Markets: Handa, Sudhanshu, Silvio Daidone, Amber Peterman,
Evidence from Brazil.” CEPR Discussion Paper Benjamin Davis, Audrey Pereira, Tia Palermo, and
16286. Jennifer Yablonski. 2018. “Myth-Busting? Confront-
Gertler, Paul J. 2004. “Do Conditional Cash Transfers ing Six Common Perceptions about Unconditional
Improve Child Health? Evidence from Progresa’s Cash Transfers as a Poverty Reduction Strategy
Control Randomized Experiment.” American Eco- in Africa.” World Bank Research Observer 33 (2):
nomic Review 94 (2): 336–41. ­259–98.
Gertler, Paul J., Sebastian W. Martinez, and Marta Hanna, Rema, and Dean Karlan. 2017. “Designing
Rubio-Codina. 2012. “Investing Cash Transfers to Social Protection Programs: Using Theory and
Raise Long-Term Living Standards.” American Eco- Experimentation to Understand How to Help Com-
nomic Journal: Applied Economics 4 (1): 164–92. bat Poverty.” In Handbook of Economic Field Exper-
Gignoux, Jérémie, and Marta Menéndez. 2016. “Ben- iments, Vol. 2, edited by Abhiji Vinayak Banerjee and
efit in the Wake of Disaster: Long-Run Effects of Esther Duflo, 515–53. Amsterdam: North-Holland.
Earthquakes on Welfare in Rural Indonesia.” Journal Hanna, Rema, and Benjamin A. Olken. 2018. “Uni-
of Development Economics 118: 26–44. versal Basic Incomes versus Targeted Transfers:
Giles, John, Xin Meng, Sen Xue, and Guochang Anti-Poverty Programs in Developing Countries.”
Zhao. 2021. “Can Information Influence the Social Journal of Economic Perspectives 32 (4): 201–26.
Insurance Participation Decision of China’s Rural Hastings, Justine, Ali Hortaçsu, and Chad Syverson.
Migrants?” Journal of Development Economics 150: 2017. “Sales Force and Competition in Financial
102645. Product Markets: The Case of Mexico’s Social Secu-
Glewwe, Paul. 1992. “Targeting Assistance to the Poor: rity Privatization.” Econometrica 85 (6): 1723–61.
Efficient Allocation of Transfers When Household Hastings, Justine, and Jesse M. Shapiro. 2018. “How
Income Is Not Observed.” Journal of Development Are SNAP Benefits Spent? Evidence from a Retail
 Banerjee et al.: Social Protection in the Developing World 1417

Panel.” American Economic Review 108 (12): ­223–56. Chicago: University of Chicago Press.
­3493–3540. ILOSTAT. 2019. “Employment in Agriculture (% of
Haushofer, Johannes, Matthieu Chemin, Chaning Total Employment)—Low Income.” International
Jang, and Justin Abraham. 2020. “Economic and Labour Organization. https://ilostat.ilo.org (accessed
Psychological Effects of Health Insurance and Cash January 2, 2020).
Transfers: Evidence from a Randomized Experiment Imbert, Clément, and John Papp. 2020a. “Costs and
in Kenya.” Journal of Development Economics 144: Benefits of Rural-Urban Migration: Evidence from
102416. India.” Journal of Development Economics 146:
Haushofer, Johannes, Paul Niehaus, Carlos Paramo, 102473.
Edward Miguel, and Michael W. Walker. 2022. “Tar- Imbert, Clément, and John Papp. 2020b. “Short-Term
geting Impact versus Deprivation.” NBER Working Migration, Rural Public Works, and Urban Labor
Paper 30138. Markets: Evidence from India.” Journal of the Euro-
Haushofer, Johannes, and Jeremy Shapiro. 2016. “The pean Economic Association 18 (2): 927–63.
Short-Term Impact of Unconditional Cash Transfers Insurance Information Institute. 2020. “Insured Homes
to the Poor: Experimental Evidence from Kenya.” by State.” https://www.iii.org/table-archive/21285
Quarterly Journal of Economics 131 (4): 1973–2042. (accessed August 22, 2022).
Haushofer, Johannes, and Jeremy Shapiro. 2018. “The Iqbal, Sarah. 2018. Women, Business and the Law
Long-Term Impact of Unconditional Cash Trans- 2018. Washington, DC: World Bank.
fers: Experimental Evidence from Kenya.” https:// Janzen, Sarah A., Michael R. Carter, and Munen-
www.povertyactionlab.org/sites/default/files/research- obu Ikegami. 2021. “Can Insurance Alter Poverty
paper/The-long-term-impact-of-conditional-cash- Dynamics and Reduce the Cost of Social Protec-
tranfer_Kenya_Haushofer_Shapiro_January2018.pdf. tion in Developing Countries?” Journal of Risk and
He, Guojun, and Shaoda Wang. 2017. “Do College Insurance 88 (2): 293–324.
Graduates Serving as Village Officials Help Rural Janzen, Sarah A., Nicholas Magnan, Conner Mul-
China?” American Economic Journal: Applied Eco- lally, Soye Shin, I. Bailey Palmer, Judith Oduol, and
nomics 9 (4): 186–215. Karl Hughes. 2021. “Can Experiential Games and
Hendren, Nathaniel, and Ben Sprung-Keyser. 2020. “A Improved Risk Coverage Raise Demand for Index
Unified Welfare Analysis of Government Policies.” Insurance? Evidence from Kenya.” American Jour-
Quarterly Journal of Economics 135 (3): 1209–1318. nal of Agricultural Economics 103 (1): 338–61.
Hidrobo, Melissa, John Hoddinott, Amber Peterman, Jean, Neal, Marshall Burke, Michael Xie, W. Matthew
Amy Margolies, and Vanessa Moreira. 2014. “Cash, Davis, David B. Lobell, and Stefano Ermon. 2016.
Food, or Vouchers? Evidence from a Randomized “Combining Satellite Imagery and Machine Learn-
Experiment in Northern Ecuador.” Journal of Devel- ing to Predict Poverty.” Science 353 (6301): 790–94.
opment Economics 107: 144–56. Jensen, Anders. 2022. “Employment Structure and
Hoddinott, John, John A. Maluccio, Jere R. Behrman, the Rise of the Modern Tax System.” American Eco-
Rafael Flores, and Reynaldo Martorell. 2008. “Effect nomic Review 112 (1): 213–34.
of a Nutrition Intervention during Early Childhood Jensen, Nathaniel, and Christopher Barrett. 2017.
on Economic Productivity in Guatemalan Adults.” “Agricultural Index Insurance for Development.”
The Lancet 371 (9610): P411–16. Applied Economic Perspectives and Policy 39 (2):
Hoynes, Hilary, and Jesse Rothstein. 2019. “Universal ­199–219.
Basic Income in the United States and Advanced Jensen, Robert T. 2004. “Do Private Transfers ‘Displace’
Countries.” Annual Review of Economics 11: 929–58. the Benefits of Public Transfers? Evidence from South
Huang, Wei, and Chuanchuan Zhang. 2021. “The Africa.” Journal of Public Economics 88 (1–2): 89–112.
Power of Social Pensions: Evidence from China’s Jiménez-Hernández, Diego, and Enrique Seira. 2023.
New Rural Pension Scheme.” American Economic “Should the Government Sell You Goods? Evidence
Journal: Applied Economics 13 (2): 179–205. from the Milk Market in Mexico.” FRB of Chicago
Huneeus, Cristóbal, Silvia Leiva, and Alejandro Micco. Working Paper 2013-19.
2012. “Unemployment Insurance and Search Effort Kandpal, Eeshani, Harold Alderman, Jed Friedman,
in Chile.” Inter-American Development Bank Work- Deon Filmer, Junko Onishi, and Jorge Avalos. 2016.
ing Paper 313. “A Conditional Cash Transfer Program in the Philip-
Ikegami, Munenobu, Michael R. Carter, Christopher pines Reduces Severe Stunting.” Journal of Nutrition
B. Barrett, and Sarah Janzen. 2018. “Poverty Traps 146 (9): 1793–1800.
and the Social Protection Paradox.” In The Eco- King, Gary, Emmanuela Gakidou, Kosuke Imai, Jason
nomics of Poverty Traps, edited by Christopher B. Lakin, Ryan T. Moore, Clayton Nall, Nirmala Rav-
Barrett, Michael R. Carter, and Jean-Paul Chavas, ishankar, et al. 2009. “Public Policy for the Poor?
1418 Journal of Economic Literature, Vol. LXII (December 2024)

A Randomised Assessment of the Mexican Uni- Wales. 1997. “Do Husbands and Wives Pool Their
versal Health Insurance Programme.” The Lancet Resources? Evidence from the United Kingdom
373 (9673): 1447–54. Child Benefit.” Journal of Human Resources 32 (3):
Klasen, Stephan, and Simon Lange. 2016. “How Nar- ­463–80.
rowly Should Anti-Poverty Programs Be Targeted? Lybbert, Travis J., Christopher B. Barrett, Solomon
Simulation Evidence from Bolivia and Indonesia.” Desta, and D. Layne Coppock. 2004. “Stochastic
Courant Research Centre Discussion Paper 213. Wealth Dynamics and Risk Management among
Kleven, Henrik Jacobsen, and Wojciech Kopczuk. 2011. a Poor Population.” Economic Journal 114 (498):
“Transfer Program Complexity and the Take-Up of ­750–77.
Social Benefits.” American Economic Journal: Eco- Macours, Karen, Norbert Schady, and Renos Vakis.
nomic Policy 3 (1): 54–90. 2012. “Cash Transfers, Behavioral Changes, and
Kugler, Adriana D. 2005. “Wage-Shifting Effects of Cognitive Development in Early Childhood: Evi-
Severance Payments Savings Accounts in Colombia.” dence from a Randomized Experiment.” American
Journal of Public Economics 89 (2–3): 487–500. Economic Journal: Applied Economics 4 (2): 247–73.
Kumler, Todd, Eric Verhoogen, and Judith Frías. 2020. Macours, Karen, and Renos Vakis. 2014. “Changing
“Enlisting Employees in Improving Payroll Tax Households’ Investment Behaviour through Social
Compliance: Evidence from Mexico.” Review of Eco- Interactions with Local Leaders: Evidence from a
nomics and Statistics 102 (5): 881–96. Randomised Transfer Programme.” Economic Jour-
Kyle, Jordan. 2018. “Local Corruption and Popular nal 124 (576): 607–33.
Support for Fuel Subsidy Reform in Indonesia.” Mahajan, Parag, and Dean Yang. 2020. “Taken by
Comparative Political Studies 51 (11): 1472–1503. Storm: Hurricanes, Migrant Networks, and US
Laurito, Maria M., Elizabeth Frankenberg, and Dun- Immigration.” American Economic Journal: Applied
can Thomas. 2022. “Effects of Housing Aid on Psy- Economics 12 (2): 250–77.
chosocial Health after a Disaster.” International Malani, Anup, Phoebe Holtzman, Kosuke Imai, Cyn-
Journal of Environmental Research and Public thia Kinnan, Morgen Miller, Shailender Swamina-
Health 19 (12): 7302. than, Alessandra Voena, Bartosz Woda, and Gabriella
Leroy, Jef L., Paola Gadsden, Sonia Rodríguez-Ramírez, Conti. 2021. “Effect of Health Insurance in India:
and Teresa González de Cossío. 2010. “Cash and A Randomized Controlled Trial.” NBER Working
In-Kind Transfers in Poor Rural Communities in Paper 29576.
Mexico Increase Household Fruit, Vegetable, and Martinelli, César, and Susan Wendy Parker. 2009.
Micronutrient Consumption but Also Lead to Excess “Deception and Misreporting in a Social Program.”
Energy Consumption.” Journal of Nutrition 140 (3): Journal of the European Economic Association 7 (4):
­612–17. ­886–908.
Levine, David, Rachel Polimeni, and Ian Ramage. 2016. Martínez, Claudia, and Marcela Perticará. 2017.
“Insuring Health or Insuring Wealth? An Experimen- “Childcare Effects on Maternal Employment: Evi-
tal Evaluation of Health Insurance in Rural Cambo- dence from Chile.” Journal of Development Econom-
dia.” Journal of Development Economics 119: 1–15. ics 126: 127–37.
Levy, Santiago, and Norbert Schady. 2013. “Latin McBride, Linden, and Austin Nichols. 2018. “Retool-
America’s Social Policy Challenge: Education, Social ing Poverty Targeting Using Out-of-Sample Valida-
Insurance, Redistribution.” Journal of Economic Per- tion and Machine Learning.” World Bank Economic
spectives 27 (2): 193–218. Review 32 (3): 531–50.
Liepmann, Hannah, and Clemente Pignatti. 2024. McKenzie, David. 2017. “Identifying and Spurring
“Welfare Effects of Unemployment Benefits When High-Growth Entrepreneurship: Experimental Evi-
Informality Is High.” Journal of Public Economics dence from a Business Plan Competition.” American
229: 105032. Economic Review 107 (8): 2278–2307.
Ligon, Ethan, Jonathan P. Thomas, and Tim Worrall. Meghir, Costas, and David Phillips. 2010. “Labour
2002. “Informal Insurance Arrangements with Lim- Supply and Taxes.” In Dimensions of Tax Design:
ited Commitment: Theory and Evidence from Village The Mirrlees Review, edited by James A. Mirrlees
Economies.” Review of Economic Studies 69 (1): and Institute for Fiscal Studies, 202–74. New York:
­209–44. Oxford University Press.
Lone, Tanya, Maheen Shakeel, and Jana Bischler. Michelson, Hope, Erin C. Lentz, Richard Mulwa,
2020. Towards Shock-Responsive Social Protection: Mitchell Morey, Laura Cramer, Megan McGlinchy,
Lessons from the COVID-19 Response in Pakistan. and Christopher B. Barrett. 2012. “Cash, Food, or
Oxford, UK: Oxford Policy Management. Vouchers? An Application of the Market Information
Lundberg, Shelly J., Robert A. Pollak, and Terence J. and Food Insecurity Response Analysis Framework
 Banerjee et al.: Social Protection in the Developing World 1419

in Urban and Rural Kenya.” Food Security 4 (3): Nichols, Albert L., and Richard J. Zeckhauser. 1982.
­455–69. “Targeting Transfers through Restrictions on Recip-
Millán, Teresa Molina, Karen Macours, John A. Malu- ients.” American Economic Review 72 (2): 372–77.
ccio, and Luis Tejerina. 2020. “Experimental Long- Niehaus, Paul, Antonia Atanassova, Marianne Ber-
Term Effects of Early-Childhood and School-Age trand, and Sendhil Mullainathan. 2013. “Targeting
Exposure to a Conditional Cash Transfer Program.” with Agents.” American Economic Journal: Eco-
Journal of Development Economics 143: 102385. nomic Policy 5 (1): 206–38.
Mills, Bradford and Carlo del Ninno. 2015. Safety Nets Nshakira-Rukundo, Emmanuel, Juliet Wanjiku Kamau,
in Africa: Effective Mechanisms to Reach the Poor and Heike Baumüller. 2021. “Determinants of
and Most Vulnerable. Washington, DC: World Bank Uptake and Strategies to Improve Agricultural Insur-
and Agence Française de Développement. ance in Africa: A Review.” Environment and Devel-
Miranda, Mario J., and Katie Farrin. 2012. “Index opment Economics 26 (5-6): 605–31.
Insurance for Developing Countries.” Applied Eco- Oates, Wallace E., and Robert M. Schwab. 2015. “The
nomic Perspectives and Policy 34 (3): 391–427. Window Tax: A Case Study in Excess Burden.” Jour-
Mirrlees, James A. 1971. “An Exploration in the Theory nal of Economic Perspectives 29 (1): 163–80.
of Optimum Income Taxation.” Review of Economic O’Brien, Clare, Zoë Scott, Gabrielle Smith, Valentina
Studies 38 (2): 175–208. Barca, Andrew Kardan, Rebecca Holmes, Carol
Moffitt, Robert. 1983. “An Economic Model of Welfare Watson, and Jenny Congrave. 2018. Shock-Respon-
Stigma.” American Economic Review 73 (5): ­1023–35. sive Social Protection Systems Research: Synthesis
Morris, Saul S., Rafael Flores, Pedro Olinto, and Juan Report. Oxford, UK: Oxford Policy Management.
Manuel Medina. 2004. “Monetary Incentives in Pri- OECD. 2020. “Insurance Spending.” https://www.
mary Health Care and Effects on Use and Coverage oecd.org/en/data/indicators/insurance-spending.html
of Preventive Health Care Interventions in Rural (accessed August 22, 2022).
Honduras: Cluster Randomised Trial.” The Lancet Olken, Benjamin A. 2007. “Monitoring Corruption:
364 (9450): 2030–37. Evidence from a Field Experiment in Indonesia.”
Mullainathan, Sendhil, and Eldar Shafir. 2013. Scar- Journal of Political Economy 115 (2): 200–249.
city: Why Having Too Little Means so Much. New Osman, Adam, and Jamin D. Speer. 2024. “Stigma and
York: Henry Holt. Take-Up of Labour Market Assistance: Evidence from
Munshi, Kaivan, and Mark Rosenzweig. 2016. “Net- Two Field Experiments.” Economica 91 (361): ­123–41.
works and Misallocation: Insurance, Migration, and Parker, Susan W., and Petra E. Todd. 2017. “Conditional
the Rural-Urban Wage Gap.” American Economic Cash Transfers: The Case of Progresa/Oportunidades.”
Review 106 (1): 46–98. Journal of Economic Literature 55 (3): ­866–915.
Muralidharan, Karthik, Paul Niehaus, and Sandip Parker, Susan W., and Tom Vogl. 2023. “Do Conditional
Sukhtankar. 2016. “Building State Capacity: Evi- Cash Transfers Improve Economic Outcomes in the
dence from Biometric Smartcards in India.” Ameri- Next Generation? Evidence from Mexico.” Eco-
can Economic Review 106 (10): 2895–2929. nomic Journal 133 (655): 2775–2806.
Muralidharan, Karthik, Paul Niehaus, and Sandip Pople, Ashley, Ruth Hill, Stefan Dercon, and Ben
Sukhtankar. 2023. “Identity Verification Standards Brunckhorst. 2021. “Anticipatory Cash Transfers in
in Welfare Programs: Experimental Evidence from Climate Disaster Response.” Centre for the Study of
India.” Review of Economics and Statistics. https:// African Economies Working Paper 2021-07.
doi.org/10.1162/rest_a_01296. Posel, Dorrit, James A. Fairburn, and Frances Lund.
Murgai, Rinku, Martin Ravallion, and Dominique van 2006. “Labour Migration and Households: A Recon-
de Walle. 2016. “Is Workfare Cost-Effective against sideration of the Effects of the Social Pension on
Poverty in a Poor Labor-Surplus Economy?” World Labour Supply in South Africa.” Economic Modelling
Bank Economic Review 30 (3): 413–45. 23 (5): 836–53.
Nagler, Paula. 2013. “How Unemployment Insurance Premand, Patrick, and Pascale Schnitzer. 2021. “Effi-
Savings Accounts Affect Employment Duration: ciency, Legitimacy, and Impacts of Targeting Meth-
Evidence from Chile.” IZA Journal of Labor and ods: Evidence from an Experiment in Niger.” World
Development 2: 1–25. Bank Economic Review 35 (4): 892–920.
Natural Disaster Insurance Institution. 2022. “Policy Ralston, Laura, Colin Andrews, and Allan Jer-Yu Hsiao.
Production.” https://dask.gov.tr/en/policy-production 2017. “The Impacts of Safety Nets in Africa: What
(accessed August 22, 2022). Are We Learning?” World Bank Policy Research
Nekoei, Arash, and Andrea Weber. 2017. “Does Extend- Working Paper 8255.
ing Unemployment Benefits Improve Job Quality?” Ravallion, Martin. 1991. “Reaching the Rural Poor
American Economic Review 107 (2): 527–61. through Public Employment: Arguments, Evidence,
1420 Journal of Economic Literature, Vol. LXII (December 2024)

and Lessons from South Asia.” World Bank Research Experimental Evidence from Rural Mexico.” World
Observer 6 (2): 205–31. Bank Policy Research Working Paper 4778.
Reynaud, Arnaud, Manh-Hung Nguyen, and Cécile Solow, Robert M. 1956. “A Contribution to the Theory
Aubert. 2018. “Is There a Demand for Flood Insur- of Economic Growth.” Quarterly Journal of Econom-
ance in Vietnam? Results from a Choice Experi- ics 70 (1): 65–94.
ment.” Environmental Economics and Policy Studies Stiglitz, Joseph E. 1974. “Incentives and Risk Shar-
20 (3): 593–617. ing in Sharecropping.” Review of Economic Studies
Robinson, Jonathan. 2012. “Limited Insurance within 41 (2): 219–55.
the Household: Evidence from a Field Experiment Stiglitz, Joseph E., and Andrew Weiss. 1981. “Credit
in Kenya.” American Economic Journal: Applied Rationing in Markets with Imperfect Information.”
Economics 4 (4): 140–64. American Economic Review 71 (3): 393–410.
Sabates-Wheeler, Rachel, Alex Hurrell, and Stephen Stoeffler, Quentin, Bradford Mills, and Carlo del
Devereux. 2015. “Targeting Social Transfer Pro- Ninno. 2016. “Reaching the Poor: Cash Transfer Pro-
grammes: Comparing Design and Implementation gram Targeting in Cameroon.” World Development
Errors Across Alternative Mechanisms.” Journal of 83: 244–63.
International Development 27 (8): 1521–45. Strupat, Christoph, and Florian Klohn. 2018. “Crowd-
Sadka, Joyce, Enrique Seira, and Christopher Wood- ing Out of Solidarity? Public Health Insurance versus
ruff. 2024. “Information and Bargaining through Informal Transfer Networks in Ghana.” World Devel-
Agents: Experimental Evidence from Mexico’s opment 104: 212–21.
Labour Courts.” Review of Economic Studies. https:// Takahashi, Kazushi, Christopher B. Barrett, and
doi.org/10.1093/restud/rdae003. Munenobu Ikegami. 2019. “Does Index Insurance
Sadoulet, Elisabeth, Alain de Janvry, and Benjamin Crowd in or Crowd Out Informal Risk Sharing? Evi-
Davis. 2001. “Cash Transfer Programs with Income dence from Rural Ethiopia.” American Journal of
Multipliers: PROCAMPO in Mexico.” World Devel- Agricultural Economics 101 (3): 672–91.
opment 29 (6): 1043–56. Tarozzi, Alessandro, and Angus Deaton. 2009. “Using
Saez, Emmanuel. 2002. “Optimal Income Transfer Census and Survey Data to Estimate Poverty and
Programs: Intensive versus Extensive Labor Supply Inequality for Small Areas.” Review of Economics
Responses.” Quarterly Journal of Economics 117 (3): and Statistics 91 (4): 773–92.
­1039–73. Thornton, Rebecca L., Laurel E. Hatt, Erica M. Field,
Saez, Emmanuel, Joel Slemrod, and Seth H. Giertz. Mursaleena Islam, Freddy Solís Diaz, and Martha
2012. “The Elasticity of Taxable Income with Respect Azucena González. 2010. “Social Security Health
to Marginal Tax Rates: A Critical Review.” Journal of Insurance for the Informal Sector in Nicaragua: A
Economic Literature 50 (1): 3–50. Randomized Evaluation.” Health Economics 19 (S1):
Saez, Emmanuel, and Stefanie Stantcheva. 2016. ­181–206.
“Generalized Social Marginal Welfare Weights For Todd, Jessica E., and Paul Winters. 2011. “The Effect
Optimal Tax Theory.” American Economic Review of Early Interventions in Health and Nutrition on
106 (1): 24–45. On-Time School Enrollment: Evidence from the
Schady, Norbert, and Maria Caridad Araujo. 2006. Oportunidades Program in Rural Mexico.” Economic
“Cash Transfers, Conditions, School Enrollment, and Development and Cultural Change 59 (3): 549–81.
Child Work: Evidence from a Randomized Experi- Townsend, Robert M. 1994. “Risk and Insurance in Vil-
ment in Ecuador.” World Bank Policy Research lage India.” Econometrica 62 (3): 539–91.
Working Paper 3930. Townsend, Robert M. 1995. “Financial Systems in
Schultz, T. Paul. 2004. “School Subsidies for the Poor: Northern Thai Villages.” Quarterly Journal of Eco-
Evaluating the Mexican Progresa Poverty Program.” nomics 110 (4): 1011–46.
Journal of Development Economics 74 (1): 199–250. van Ours, Jan C., and Milan Vodopivec. 2008. “Does
Skiba, Aleksandr K. 1978. “Optimal Growth with a Reducing Unemployment Insurance Generosity
Convex-Concave Production Function.” Economet- Reduce Job Match Quality?” Journal of Public Eco-
rica 46 (3): 527–39. nomics 92 (3–4): 684–95.
Skoufias, Emmanuel. 2005. PROGRESA and Its Vu, Khoa, and Paul Glewwe. 2022. “Maternity Benefits
Impacts on the Welfare of Rural Households in Mex- Mandate and Women’s Choice of Work in Vietnam.”
ico. Washington, DC: International Food Policy World Development 158: 105964.
Research Institute. Wager, Stefan, and Susan Athey. 2018. “Estimation
Skoufias, Emmanuel, Mishel Unar, and Teresa and Inference of Heterogeneous Treatment Effects
González-Cossío. 2008. “The Impacts of Cash and Using Random Forests.” Journal of the American
In-Kind Transfers on Consumption and Labor Supply: Statistical Association 113 (523): 1228–42.
 Banerjee et al.: Social Protection in the Developing World 1421

Wagstaff, Adam, Ha Thi Hong Nguyen, Huyen Dao, World Bank. 2021b. GDP per Capita (Constant 2015
and Sarah Bales. 2016. “Encouraging Health Insur- US$). Washington, DC: World Bank DataBank.
ance for the Informal Sector: A Cluster Randomized https://web.archive.org/web/20220706011113/
Experiment in Vietnam.” Health Economics 25 (6): https://api.worldbank.org/v2/en/indicator/NY.GDP.
­663–74. PCAP.KD?downloadformat=csv (accessed July 18,
World Bank. 2014. “For up to 800 Million Rural Poor, 2022).
a Strong World Bank Commitment to Agriculture.” World Bank. 2021c. Population, Total. Washington, DC:
World Bank News, November 12. https://www. World Bank. https://data.worldbank.org/indicator/
worldbank.org/en/news/feature/2014/11/12/for-up- SP.POP.TOTL (accessed September 14, 2022).
to-800-million-rural-poor-a-strong-world-bank- World Economic Forum. 2019. “GCI 4.0: Insur-
commitment-to-agriculture. ance Premium.” https://www3.weforum.org/docs/
World Bank. 2018. The State of Social Safety Nets 2018. WEF_TheGlobalCompetitivenessReport2019.pdf
Washington, DC: World Bank. (accessed July 13, 2022).
World Bank. 2021a. Country Classification by Income Yang, Dean. 2008. “Coping with Disaster: The Impact
Level. Washington, DC: World Bank DataBank. of Hurricanes on International Financial Flows,
https://web.archive.org/web/20220621100831/ 1970–2002.” BE Journal of Economic Analysis and
https://databank.worldbank.org/data/download/site- Policy 8(1).
content/CLASS.xls (accessed June 21, 2022).