
14 minute read
Reflections on Patterns of Use of Targeting Methods
Choosing among Targeting Methods | 265
upper-middle-income programs and 23 percent of high-income ones). Perhaps mirroring income levels, Sub-Saharan African countries are most likely to use CBT and least likely to use means testing, which is particularly predominant in Eastern Europe and Central Asia. PMT is generally used in around 6–13 percent of the programs, less so in Sub-Saharan Africa, and infrequently in Eastern Europe and Central Asia.
The rest of the chapter looks at the different considerations in choosing which methods to use. The wide range of factors to be thought through in different contexts helps explain the diversity of use of methods.
Reflections on Patterns of Use of Targeting Methods
The tenor of the conversation around self-targeting seems to have changed over the years with respect to price subsidies. In the era of state-led development with major interventions in the agricultural markets in the form of import/export controls, state marketing boards, strategic grain reserves, and the like, general food price subsidies were fairly common and part of interventions with a mix of agricultural, social protection, and statebuilding objectives. The choice of food commodities to subsidize has become a less common thread of the social protection practice due to reduced interventions in markets and eventually due to the development of cash transfer programs. In 2008, during the general food and fuel price crisis, food subsidies or tax exemptions were a common response, especially in lowincome countries that did not have well-established alternative social assistance–28 countries increased their subsidies and 84 reduced taxes on selected foodstuffs (IMF 2008). That is the case in fewer countries now, and although the responses to COVID-19 included food commodity distribution programs with various degrees of targeting, resorting to general food price subsidies and tax reductions has not been so common. In addition, there has been a secular move to cash rather than in-kind benefits over the years. For example, Indonesia’s subsidized but low-quality rice program (variously named Implementation of Special Market Operation, Raskin, and then Rastra), which has been in place since the Asian Financial Crisis, was phased out in 2017 and replaced with a digital food voucher Food Assistance Program, Bantuan Pangan Non-Tunai (BPNT), then called Sembako (Holmemo et al. 2020; Banerjee et al. 2021).
Of course, every program is self-targeting in the sense that people have to deem the benefits worthy of the costs of participating. This is the flip side of the issues of transaction costs and stigma discussed in chapter 2. The growing prevalence of human rights viewpoints and their concern with inclusion and dignity rules out the always quite rare suggestion that delivery systems should be purposefully inconvenient. The accumulated work
266 | Revisiting Targeting in Social Assistance
on delivery systems should be improving inclusion, although there is still a journey to go. Thus, the main way in which people are purposefully induced to self-exclude due to transaction costs is in setting low wages for temporary employment on public works projects, which are discussed later in this chapter. Occasionally, self-exclusion from the top is encouraged for nearly universal programs, for example, India’s Give It Up campaign, which was implemented as part of the liquid petroleum gas–related cash program, reaching 177 million people, and successfully promoted the self-exclusion of about 10 million wealthy individuals (Gelb and Mukherjee 2019).
In its traditional mode, geographic targeting was about rationing, locating programs exclusively, or concentrating caseloads in the neediest areas. Because this could be done with off-the-shelf data, it was much simpler than household-specific targeting, although of course, rationing carries explicit errors of exclusion.
As adaptive social protection gains prominence, especially in the contexts of natural disasters and climate change, geographic targeting has become a central method for preparing for and responding to such covariate shocks. Early warning systems, disaster risk management, and risk profiling analyses are helping countries to identify ex ante the needs of the population exposed to a shock and the areas that are more likely to be affected. This facilitates policies to help improve the resilience of communities and people. Better interoperability of an early warning system and a disaster risk management system can help improve preparedness and response to shocks.
As big data (for example, from satellite imagery, mobile phones, or digital content) has become more pervasive, the sophistication of information that can be brought to bear on geographic targeting has increased enormously,4 but big data do not completely solve the problem of householdspecific targeted income support programs. Big data make poverty mapping possible in places without recent censuses or surveys, with faster updates and more granularity than was previously possible. For example, poverty maps based on night lighting (sensing) may use a kilometer grid (see, for example, Skoufias, Strobl, and Tveit 2017). Even for poverty maps that can observe some dwelling-specific characteristics, such as the material of the roof, this is only a single characteristic and not often used alone in determining eligibility or setting benefit levels. The sole, although quite prominent, example for cash transfers5 known to the authors is the Give Directly program, which initially used thatch roofs as its targeting criterion (Abelson, Varshney, and Sun 2014; Haushofer and Shapiro 2016) but eventually added additional criteria (Ohlenburg 2020). The World Development Report 2021 (World Bank 2021b) envisages the use of mobile phone and social media data to target households directly, but several steps will be required to ensure access to these usually private sector data (see chapter 6 for
Choosing among Targeting Methods | 267
further discussion). More promising is the use of sensing data to trigger vertical or horizontal program expansions in response to droughts or floods.
Nonetheless, machine learning and big data approaches require rigorous benchmarking and assessment. The accuracy of big data poverty maps needs to be assessed; as chapter 6 discusses, much of their evaluation has been in-sample. Most importantly, machine learning does not replace the need to invest in people and human process, as Joshua Blumenstock (CEGA6 Faculty Co-Director) highlighted in an interview7: “The algorithms are sort of the shiny object, and they receive a lot of attention. But when it comes to actually implementing social protections, going the last mile to put money in the hands of people who need it, the algorithms are just one small link in a much larger chain of humanitarian assistance. Most of the other links are human. Algorithms can help surface relevant information, but humans must decide what to do with it.”
An important number of programs rely on demographic (also known as categorical age-based) targeting. Age (for children or the elderly) is used as the sole standard or combined with other criteria for eligibility for many programs, and benefit levels may be customized by age as well. There seem to be three variants in the reasoning to support demographic targeting, which are sometimes not clearly distinguished or acknowledged: • First, some programs simply consider that all members of a group with a (usually) simple-to-observe characteristic deserve public support no matter their individual money-metric welfare or that of their families.
Examples include veterans who provided service to their country and merit support/recognition for that. The blanket argument accords well with rights-based arguments as well, along the lines that children are inherently vulnerable and the precious future, society must nurture them, providing health, education, water, social protection, and so forth.
Similarly, the elderly are vulnerable and deserving of support and respect for their service and wisdom. • A second rationale acknowledges that not every member of a group requires public assistance within a money-metric notion of welfare but sees that, on the whole, members of the group and the families with which they live are poorer than average. In this line of argument, demographic targeting may have significant errors of inclusion, but the use of a single, easy-to-observe proxy is simpler to implement and more transparent than many other methods. Moreover, the groups selected resonate with societal views of deservingness in most places. Thus, it may be relatively easy to build consensus in support of such programs and any errors of inclusion tend not to offend. • A third variant of demographic targeting is when programs or benefits that use a money-metric gradient for narrowly targeting only admit and provide for families that have members of the defined category. In the
268 | Revisiting Targeting in Social Assistance
United States, for example, the welfare state is both heavily means tested and markedly focused on children or the elderly. Many programs provide support to families with poor children or elderly members, but only to these families. Social assistance units composed of only poor, primeage adults were long excluded altogether; even now, prime-age adults are required to meet much more stringent thresholds or work requirements, recertify more often, and so forth. There is enough consensus over societal responsibility for children and the elderly to support social protection programs in their favor, but there is much less consensus on supporting prime-age adults, especially if they are not working, except possibly for those living with serious disabilities.
Means testing is a common method used to differentiate eligibility and benefits, especially in highly formalized economies. Means testing is widely used in Western Europe and the long-standing Organisation for Economic Co-operation and Development members, such as Australia, Canada, Japan, New Zealand, and the United States, where verification of means is possible. Unverified means testing is used in some countries with substantial informality, such as Brazil and South Africa. Means testing is deemed something of a gold standard among household-specific methods because, unlike other methods, it contains no inherent measurement error, although of course, various errors creep in during implementation. Means testing has been shown to be very accurate in its assessments.
An allied strand of targeting practice grew from the European and Central Asian transition economies with the HMT. These countries had high but decreasing levels of formality and an orientation toward Western Europe with its high formality and use of verified means testing. The countries invented what Tesliuc et al. (2014) call an HMT that uses declared income that can be verified with an imputation or proxy for other sources of income that are not easily verifiable. The method has not spread as widely as PMT, but as the data revolution has increased the scope and decreased the cost of databases everywhere, it may be pertinent in places not using it now. HMT may be especially useful in moderately formal economies or for programs that use affluence tests, which try to screen out the top of the income distribution, which may have formal incomes or assets, more than trying to focus on the very poorest.
One of the strands of modern targeting practice in the developing world today is PMT, which originated in Latin America. Countries in that region had high income inequality, high levels of informality, relatively strong government and information, and largely adequate physical access to health and education services but big gaps in human capital outcomes. They also had years of fiscal, economic, and societal scarring from the debt crisis of the 1980s and all that ensued. Given the inequalities and limited fiscal space, household-level targeting was desired, but with high levels of informality, means testing as traditionally practiced in Western Europe and
Choosing among Targeting Methods | 269
North America seemed unreachable. Thus, countries such as Chile, Colombia, and Mexico started using PMT for a variety of programs, their conditional cash transfers most prominently but also subsidized health insurance, sometimes social pensions, and many more. PMTs are based on data analytics from household surveys but not on verification of household-specific information from existing government databases. The PMT method spread not just through a great deal of Latin America, but far beyond, sometimes to relatively similar settings (for example, the Philippines) and sometimes to far different ones, with many PMTs built in the lower income, lower inequality African countries.
PMT is something of a Rorschach test for those who think about targeting. Many, especially in ministries of finance or planning, see it as modern, scientific, data driven, replicable, and thus good for preventing patronage politics in social programs and safeguarding their reputation. Some communities find it a black box, a mystery. Analysts and observers have mixed opinions. Some find it a realistic, if imperfect, solution to a problem without perfect solutions; some find it anathema for its inbuilt statistical errors or lack of transparency. Some associate it with static survey sweeps and dislike those (although several countries using PMT have on-demand applications and dynamic registries). Many observers (the authors of this volume included) find that conversations about social policy proposals often jump far too quickly to issues around PMT, with insufficient discussion of the policy problem to be solved, the range of programming options, the range of possible eligibility determination methods that might be used, or how improvements in delivery systems could improve outcomes.
CBT is perhaps the oldest of the household-specific assessment techniques, but today, a much smaller share of programs apply CBT as a standalone approach. Conning and Kevane (2002) cite the use of CBT in historical events such as to support the 1834 English system of poor relief, in which local parishes performed some functions of local civil government, including the administration of poor relief, and the use of “native authorities” by the French and British as local leaders. In recent times, CBT is still among the most commonly used methods, especially in low-income countries, but it is rarely used alone and there are different ways of implementing it. In McCord’s (2013) review of CBT experience (still the most recent comprehensive review), over half of the cases she identifies are in Africa, a plurality in Asia, a few in Latin America, and only one in Eastern Europe and Central Asia. She found information on 57 programs using CBT for which there were sufficient data for further analysis. In these programs, CBT was used alone in only three programs and paired with geographic targeting in half of the remainder. Some programs use CBT as a filter prior to using other methods, to narrow the pool of households still further. Others use community validation only at the end, taking steps to avoid reintroducing elite capture at this stage, such as allowing households
270 | Revisiting Targeting in Social Assistance
that were excluded during the original CBT process to be included or only allowing subtractions but not additions.
Despite the data revolution and increasing availability of other options, CBT remains a favored choice for its ties to the country political context. The community has been part of eligibility determination processes for years in many countries; thus, there is a strong sense that the community must still be part of the eligibility determination process. Beegle, Coudouel, and Monsalve Montiel (2018) highlight the importance of considering the possible trade-off between political and technical imperatives in designing targeting methods. In the limited comparative treatment arm experiments that have been done, CBT may be preferred by communities or lend cohesion.
CBT and PMT are sometimes used in the same program; a third of the CBT-based programs in ASPIRE also use PMT. There are multiple views on the logic of this, depending in part on the functions carried out by the community. Sometimes the community helps with functions such as outreach, prelisting households to be surveyed, and even data collection on a standard PMT form. The logic of using community members to complement and support administrative staff in conducting a PMT is clear, although in such cases, perhaps the nomenclature exaggerates the role of the community and it might be more accurate to label these instances as PMT methods. Where communities play a decision role jointly with a PMT, the logic is often stated as the community process guarding against errors of exclusion and the PMT guarding against errors of inclusion. This has a ring to it, but the logic is not unassailable. Running both processes may increase the costs and risk contradictions. In the Ghana study on this issue presented by Pop (2015), for example, CBT worked first to create a prelist of potential beneficiaries; then, PMT was brought to bear. By ruling out households put forth by the community, PMT lowered errors of inclusion somewhat, but it also introduced significant errors of exclusion with respect to the CBT-based lists. Moreover, in overruling the community, PMT undermines its power in decision making, and it may not abet the acceptability of decisions, hence creating some social tensions. Other evidence supports using CBT as a filter in cases where budgets or capacity are severely constrained or information sources limit the ability of more quantitative methods, such as PMT, to predict welfare status with enough accuracy. Adding PMT to a CBT process can reduce crosscommunity variation by bringing the following: (1) a common definition of poverty with more weight on money-metric poverty8 to the process, (2) more training for community agents on the objective of the program; and (3) strengthened foundations of programs with preparation of clear operational manuals, information campaigns, and accountability mechanisms.
Another of the roots of current social assistance practice is humanitarian assistance, which is usually provided in response to some sort of emergency, financed by donors, and has a temporary vision and sometimes improvised methods. Because the programs are usually (initially) viewed as temporary,