
4 minute read
Conclusion
How to Harness the Power of Data and Inference | 445
For example, a teacher may be involved in selecting poor children for scholarships, but also must interact with the parents of all students on their learning, including those of children not selected.
Clear protocols for oversight and control procedures are needed even when communities are given significant authority in beneficiary selection. A regular monitoring and evaluation system, which includes spot checks, process evaluations, and independent audits, will need to be defined, as well as defining a rotation of members to (1) share the burden of selection, and (2) ensure that different community members can participate in the process, increasing citizen engagement and local governance. A grievance system to respond to beneficiary complaints and ensure a high level of social accountability is desirable. Together these can detect systemic problems to be addressed. These processes require thorough training of and guidance for community committees and an adequate and easily accessible information system for enrollment, transactions, and proper grievance and appeal mechanisms to compensate for the inherent errors and limitations occurring in implementation.
Conclusion
It is hard to summarize both fully and succinctly such a long chapter, spanning diverse targeting methods and with many technical details for each, but this conclusion points out some common threads.
Policy makers must make considered judgments. Chapter 5 discusses the choice of targeting method(s). Even after that choice is made, further judgments are needed in designing the implementation of any given method— about things like the choices over the unit of assistance, the weight put on errors of exclusion versus errors of inclusion, and the emphasis on targeting accuracy versus administrative costs versus incentive effects versus transparency or ownership.
There is no reason to be purist about targeting methods. Many, many countries and programs use multiple methods. Moreover, the line between means testing and HMT is blurry as is that between HMT and PMT. Further, CBT and PMT are increasingly combined.
Data matter. Traditional data, in the sense of government-held administrative data, data from applicant interviews, or community members’ knowledge of their neighbors, still dominate in targeting practice. The revolution in people holding and using foundational or functional IDs, especially eIDs, and increased computing power are making it far easier to create integrated or interoperable data systems that lower costs and increase the dynamism of social registries. As data coverage and quality improves, more countries will meet the minimum conditions to move
446 | Revisiting Targeting in Social Assistance
toward means testing or HMT. Meanwhile, although further work is needed to develop fully and assess its accuracy, big data, such as remote sensing, CDR, and social media data, hold the potential for more frequent and cheaper geographic targeting for various purposes—allocation of program benefits, allocation of program administrative resources, and blending of spatial analysis into formulae for eligibility assessment. Further data advances are on the horizon; how soon and to what extent they can be brought to bear in social assistance programs depends on a series of factors, many of them more about culture and regulation than data science.
Inference matters. There is not a single recipe to guide the modeling in HMT, PMT, or PMTplus, but there is a well-developed body of statistics with applications to targeting that helps guide the modeler, and this chapter has reprised some of the basics. The bottom line for the moment is that there may be a sort of “Goldilocks” range in terms of sophistication—in means testing, capturing most income and verifying some may be enough, and pushing to the extreme may be counterproductive; in PMT modeling, while the added sophistication of quantile regressions or logistic regressions with Lasso selection over principal component analysis or OLS is usually preferable, so far, the complexity of machine learning does not seem to pay off well, at least on the traditional static data used in PMTs. However, machine learning can be used in data preprocessing to create “deep features” within traditional survey data, which may then improve PMT performance regardless of whether traditional regressions or machine learning algorithms are used.
Customization matters. There are a lot of principles that apply and tricks of the trade. But in the end, what makes sense to do must account for specific features of the setting—the goals, design, and budget of the social protection programs; the shape of the welfare distribution; the availability, quality, and details of the data available; national capacities; and political economy.
Good data and inference are important but not sufficient; all targeting mechanisms build on the rest of the delivery system and all require a bridge between the central administration and the potential population. In the more formal systems, building these bridges to bring the local authorities as close as possible to the program implementation activities seems to be largely a matter of delivery systems focused on establishing rules, data systems, and training professional staff. In all, the participation of government staff (locally posted staff of federal agencies or staff at the municipal level) with proper resources and incentives and proper citizen engagement legitimizes the process and helps to improve program outcomes.
The development of good targeting systems takes a diverse skill set. As this chapter has made clear, data/stats nerds clearly have their place on a team. However, as prior chapters have made clear, there is also a need for