Abstract
This paper analyzes household data from nine programs in the Sahel region using a harmonized approach to compare Proxy-Means Testing (PMT) and Community-Based Targeting (CBT) as conducted in practice, once geographical targeting has been applied. Results show that the targeting performance measured depends critically on the definition of the targeting objectives, share of beneficiaries selected, and indices used to evaluate targeting. While PMT performs better in reaching the poorest households based on per capita consumption, it differs little from CBT, random or universal selection when distribution-sensitive measures are employed, or when food security is used as the welfare metric. Administrative costs associated with targeting represent only a small share of budgets. Results emphasize the importance of studying programs as implemented in practice instead of relying on simulations of targeting performance. They also suggest that PMT and CBT contribute little to poverty or food insecurity reduction efforts in poor and homogeneous settings.
Acknowledgements
We are very grateful to Diana Cheung, Mame Fatou Diagne, Yves Kameli, Papa Sosthène Konate, Patrick Premand, Solene Rougeaux and Kalilou Sylla for providing us with the data analyzed in this paper. We are also grateful to Arthur Alik-Lagrange, Christian Bodewig, Caitlin Brown, Juliette Cappicot, Margaret Grosh, and Patrick Premand for their comments. The datasets and code used for the analysis are available upon request to bona fide researchers.
Disclosure statement
No potential conflict of interest was reported by the author(s).
Notes
1 For opposite views on the topic, see for instance Del Ninno and Mills (Citation2015), Kidd, Gelders, and Bailey-Athias (Citation2017) or Grosh, Leite, and Wai-Poi (Citation2022). Some scholars question the very idea of targeting beneficiaries rather than implementing “universal” programs (Banerjee, Niehaus, & Suri, Citation2019; Brown, Ravallion, & van de Walle, Citation2018; Devereux, Citation2021; Mkandawire, Citation2005). Various organizations also have different paradigms and practices in terms of targeting, as can be observed in the Sahel region for example (Premand & Schnitzer, Citation2020).
2 Due to data limitations, we are unable to study these important dimensions of targeting performance. In addition, our study focuses on countries characterized by high poverty rates and low administrative capacities These are caveats to our analysis and calls for future research.
3 Ralston, Andrews, and Hsiao (Citation2017) includes only one study from Burkina Faso, while Devereux et al. (Citation2017) and Coady, Grosh, and Hoddinott (Citation2004) do not include any of the countries we study.
4 There are two datasets for Burkina Faso, Niger and Senegal, which we differentiate by a number (1 or 2).
5 See Table S1 in Supplementary Materials.
6 FCS was not available in two datasets where only the Household Dietary Diversity Score (HDDS) is available for food security. In Chad, consumption per capita is not available.
7 The Senegal 2 study was conducted to assess the targeting performance of the Unique National Registry (RNU), which was conducted nationally. The sample comes from 4 regions. Likewise, the health insurance project studied in Burkina Faso 2 was rolled out in several regions, and the sample comes from 2 distinct regions.
8 In low-income settings, welfare is often measured noisily and imperfectly (Deaton & Zaidi, Citation2002), and the binary distinction between “poor” and “non-poor” is not always meaningful (Ellis, Citation2012). Besides, the binary distinction between the correct and incorrect inclusion can bias the policy discussion by suggesting high rates of “non-deserving” individuals that receive transfers.
9 For instance, when comparing CBT and PMT, if 35% of the population is selected by CBT, then 35% of the population with the lowest per capita consumption is defined as “poor” (or as the target), and the PMT threshold is adjusted so that 35% of the households with the lowest PMT scores are selected.
10 Such adjustments are possible because per capita consumption and the FCS are continuous (or largely continuous) variables. However, HDDS is a lumpy measure of food security. In Mali, the CBT threshold does not correspond to a food security threshold. Consequently, some households with HDDS between 6 and 7 were randomly allocated to the “food insecure” group.
11 We use 35% because HDDS measures, which are lumpy, also marked a discontinuity at this threshold, for example jumping from 5 to 6 in Burkina Faso 2.
12 Specifically, CGH = where is the share of the transfers received by the poor population, and is the share of the population that is poor.
13 In addition, we compute the Distribution Characteristic Index (DCI) as a robustness check. The DCI also considers the full distribution of the well-being metric, but with some important limitations (see Appendix A). The results using DCI are broadly consistent with those obtained from our three main measures and available upon request.
14 This amount is close to the amount provided in most of the programs that we study and other cash transfer programs in Sub-Saharan Africa. Increasing or decreasing the amount received by each household does not affect the results qualitatively.
15 We compute bootstrap 90% confidence intervals. PMT and CBT confidence intervals do not overlap in most cases.
16 In Supplementary Materials Figure S2, we also compare PMT selection with geographic selection (Elbers, Fujii, Lanjouw, Özler, & Yin, Citation2007). Geographic selection consists in selecting all households in villages with the lowest, average PMT scores. Geographic selection performs relatively well in some datasets and for some selection rates (Supplementary Materials Figure S2). It achieves error rates that range generally between PMT and random selection error rates. However, the potential of geographical targeting tested here is relatively low given that the programs studied are implemented after some geographical targeting has already occurred.
17 DCI results are available upon request. Mechanically, DCI values depend critically on selection rates and on the initial distribution in each dataset: it is higher in countries with more heterogeneity such as Burkina Faso 2 and Mali. For other countries, DCI values are closer, and relatively flat within a dataset.
18 On implementation issues, see for instance Sabates-Wheeler, Hurrell, Devereux (2015), Devereux et al. (Citation2017), Olivier de Sardan and Piccoli (Citation2018) or Stoeffler, Fontshi, and Lungela (Citation2020).
19 Pooling all datasets together lead to similar results (Table S7 in Supplementary Materials).
20 Financial cost per individual was $5.73, while the economic cost (taking time into account) was $11.83.
21 The estimates include variable costs associated with collecting the information necessary to identify beneficiaries and exclude fixed costs that are linked to multiple program aspects other than targeting, such as government administrative costs. Under self-targeting-PMT, households had to physically come to an office to apply for the program and provide the information required to apply the PMT.
22 We simulated scenarios under a method that would perfectly identify the poorest households based on consumption per capita and compared these to PMT and CBT while holding all other factors the same (e.g., level of benefit, share of population selected). Results are available upon request.
23 When thresholds are adjusted so that the welfare threshold and the selection threshold are identical (e.g. 30% household selected by a method, 30% deprived households) then inclusion and exclusion errors at the household level are identical. In medical sciences, the terminology “sensitivity” and “specificity” are used instead of “inclusion” and “exclusion” errors (e.g. Ouédraogo et al., Citation2017).