Successfully reported this slideshow.
We use your LinkedIn profile and activity data to personalize ads and to show you more relevant ads. You can change your ad preferences anytime.

Informational Constraints on Antipoverty Programs: Evidence for Africa

834 views

Published on

Martin Ravallion - Georgetown University
ERF Eminent Lecturer Series on Global Poverty and Anti-Poverty Policies

Egypt, Cairo, ERF Premises December 10, 2017
www.erf.org.eg

Published in: Government & Nonprofit
  • Excellent. Thank you Martin & Colleagues for some academic rigour. I arrived at the basicincome conclusion through more of a thumb suck after my cash transfer experience in Malawi 1990s where targeting proved a major headache. You suggest India for BI which is on the cardds given GOI interest. We suggest Sierra Leone and kick it off with Development Aid. Already in discussions with Shadow Development sec in UK over this! We've done a fair amount of number crunching etc.
       Reply 
    Are you sure you want to  Yes  No
    Your message goes here
  • Be the first to like this

Informational Constraints on Antipoverty Programs: Evidence for Africa

  1. 1. Informational Constraints on Antipoverty Programs: Evidence for Africa Martin Ravallion Georgetown University Presentation at the Economic Research Forum, Cairo, December 2017
  2. 2. New hope for redistributive interventions using transfers in developing countries • World’s aggregate poverty gap “ …is roughly what Americans spend on lottery tickets every year, and it is about half of what the world spends on foreign aid.” (Annie Lowrie, NYT, Feb 2017). (Based on an estimate of $80 billion for global PG at MERs; at PPP it is $160 billion.) • The implication drawn is that it should be relatively easy to eliminate global poverty using transfers. Is that right? 9/23/16 Brown, Ravallion & van de Walle 2
  3. 3. Huge expansion in “social safety nets” (SSN) in the developing world • SSN: Direct non-contributory income transfers targeted to poor families • In last 15 years many developing countries have introduced new SSN programs. • Today almost every developing country has at least one SSN program. • Roughly one billion people currently receive assistance. • Using the World Bank’s ASPIRE database I estimate that population coverage of SSN programs (% receiving any help) is growing at 9% per annum (3.5% points). 3
  4. 4. Cruel irony: Poorer countries are less effective in reaching their poor 0 20 40 60 80 100 0 2000 4000 6000 8000 10000 12000 14000 16000 18000 20000 22000 GDP per capita at PPP for year of survey Safety net coverage for poorest quintile (%) Safety net coverage for whole population (%) Poorest quintile Population
  5. 5. One billion poor; one billion SSN recipients Living in poverty Receiving help from SSN 5 But mostly not the same people in poor countries!
  6. 6. How much does imperfect information constrain antipoverty policies in poor countries? 9/23/16 • The “poverty gap” calculation assumes that we can accurately identify poor people and tell how poor they are. This is a strong assumption. • Limitations of even the best household surveys • Policies in practice use a smaller set of poverty proxies • Reaching poor households does not mean we will reach all poor individuals • Many questions left begging: • How well can we do against poverty with such data? • What methods of aggregating these data into a score do best? • How much do implementation lags restrict performance against poverty? Brown, Ravallion & van de Walle 6
  7. 7. Questions for this presentation 1: How well can we identify poor households with the type of data routinely used by policy makers in poor countries? 2: How well can we identify poor individuals using such data? In both, focus on Sub-Saharan Africa; the region with highest poverty measures today + large expansion in social protection programs. 9/23/16 Brown, Ravallion & van de Walle 7
  8. 8. How well can we identify poor households with the data routinely used in poor countries? 9/23/16 Brown, Ravallion & van de Walle 8 Based on Caitlin Brown, Martin Ravallion and Dominique van de Walle, “Poor Means Test? Econometric Targeting in Africa,” NBER Working Paper
  9. 9. Targeting with imperfect information • Standard measures of economic welfare are not fully observed in the population • Instead, governments use proxies for targeting (location, family size, housing conditions,…) • Households receive a “score” = Proxy Means Testing (PMT) • Main challenge is setting the weights of that score 9/23/16 Brown, Ravallion & van de Walle 9
  10. 10. Econometric targeting • Regression for (log) consumption; h’hold 𝑖 in country 𝑗, date 𝑡, covariates 𝑥𝑖𝑗𝑡: 𝑦𝑖𝑗𝑡 = 𝛼𝑗𝑡 + 𝛽𝑗𝑡 𝑥𝑖𝑗𝑡 + 𝜀𝑖𝑗𝑡 (𝑖 = 1, … , 𝑁𝑗𝑡) • The PMT score: 𝑦𝑖𝑗𝑡 = 𝛼𝑗𝑡 + 𝛽𝑗𝑡 𝑥𝑖𝑗𝑡 • H’holds in population with scores below some cutoff point are eligible. • This is used by World Bank, other donors, and many governments. Supporters: “it produces the best targeting outcome, esp., in excluding the non-poor” (Grosh & Baker, 1995; Grosh, 1994; del Ninno & Mills, 2015). 9/23/16 Brown, Ravallion & van de Walle 10
  11. 11. Critics of econometric targeting • Econometric targeting has been criticized for its “considerable inaccuracy at low levels of coverage.” Kidd & Wylde (2011) • Transparency concerns (Adato & Rooppnaraine, 2004). “...the targeting process as a whole is poorly understood… One informant from a geographically-targeted community noted: ‘Well, some people wonder why they weren’t targeted even though they live in this same area. So we tell them that the Bible says that many are called but few are chosen.’” Adato & Rooppnaraine (2004) • Local social unrest generated by PMT (Cameron and Shah, 2004). • Erosion of local social capital; distrust of local administrators. • Can’t address all these concerns; but at the core: how well does it work? 9/23/16 Brown, Ravallion & van de Walle 11
  12. 12. Econometrics 101: Understanding regressions • By design, a standard regression line passes through the means of the data. • The residuals will be positively correlated with the dependent variable, and more so the higher the variance of the residuals: • Econometric targeting will tend to work less well near the extremes; tends to overestimate living standards for the poorest. • This is intrinsic to the use of regression. 9/23/16 Brown, Ravallion & van de Walle 12 If   xy with 0),( xCov then 0)ˆ()ˆ,(   VaryCov
  13. 13. In this paper we… … provide a systematic assessment of econometric targeting as a tool for social policies aimed at reducing poverty. We: • Study the most common form of “Basic PMT” • Consider alternative models using additional covariates: “Extended PMT” • Consider alternative estimation methods more appropriate to poverty reduction • Characterize optimally targeted transfers with imperfect information • Introduce plausible lags in implementation • Compare econometric targeting to other targeting methods 9/23/16 Brown, Ravallion & van de Walle 13
  14. 14. How do we assess performance? 9/23/16 Brown, Ravallion & van de Walle 14
  15. 15. The literature on targeting • “Targeting” tries to attain greater poverty impacts from transfers, or reduce the cost of a given impact. • One strand of literature: Choose transfers across types of households to minimize a measure of poverty subject to a budget constraint (Kanbur, 1987: Ravallion and Chao, 1989; Glewwe, 1992). • Others emphasize “targeting efficiency,” defined in terms of reducing certain “targeting errors” (Grosh and Baker, 1995; Coady et al., 2004). • Here we shall study both types of measures. 9/23/16 Brown, Ravallion & van de Walle 15
  16. 16. What we don’t do • We don’t question that household consumption is an adequate welfare indicator. PMT makes that assumption, and we accept it for our purpose. • However, we do allow for measurement errors (using panel data) • We do not consider alternatives such as self-targeting using work requirements (“workfare”) or community-based targeting in which local communities are engaged directly in deciding who is poor and who is not. • We do not consider the (economic, social and political) costs of targeting, which have received some attention in the literature. 9/23/16 Brown, Ravallion & van de Walle 16
  17. 17. Measures of targeting 9/23/16 Brown, Ravallion & van de Walle 17 2. Exclusion Error Rate (EER): Proportion of poor who are identified as non- poor: 𝐸𝐸𝑅 𝑗𝑡 = 𝑖=1 𝑁 𝑗𝑡 𝑤 𝑖𝑗𝑡1( 𝑦 𝑖𝑗𝑡>𝑧 𝑗𝑡|𝑦 𝑖𝑗𝑡≤𝑧 𝑗𝑡) 𝑖=1 𝑁 𝑗𝑡 𝑤 𝑖𝑗𝑡1(𝑦 𝑖𝑗𝑡≤𝑧 𝑗𝑡)
  18. 18. Measures cont. 3. Normalized Targeting Differential (NTD) Difference between proportion of poor predicted as poor and proportion of the non-poor predicted as poor 𝑁𝑇𝐷𝑗𝑡 = 1 − 𝐸𝐸𝑅 𝑗𝑡 − 𝑖=1 𝑁 𝑗𝑡 𝑤𝑖𝑗𝑡1( 𝑦𝑖𝑗𝑡 ≤ 𝑧𝑗𝑡|𝑦𝑖𝑗𝑡 > 𝑧𝑗𝑡) 𝑖=1 𝑁 𝑗𝑡 𝑤𝑖𝑗𝑡1(𝑦𝑖𝑗𝑡 > 𝑧𝑗𝑡) • When all and only the poor receive the transfer, 𝑁𝑇𝐷𝑗𝑡 = 1 • Under a uniform transfer scheme, 𝑁𝑇𝐷 𝑗𝑡 = 0 • When all and only the non-poor receive the transfer, 𝑁𝑇𝐷𝑗𝑡 = -1 9/23/16 Brown, Ravallion & van de Walle 18
  19. 19. Poverty reduction is the objective, not targeting per se • EER and NTD better predict impact on poverty than IER (Ravallion, 2009) • However, with the same data we can measure impacts on poverty directly, bypassing the targeting measures. • Three poverty measures => 1. Headcount index: 𝐻𝑗𝑡 = 𝑖=1 𝑁 𝑗𝑡 𝑤𝑖𝑗𝑡1(𝑦𝑖𝑗𝑡 ≤ 𝑧𝑗𝑡) 2. Poverty Gap Index: 𝑃𝐺𝑗𝑡 = 𝑦 𝑖𝑗𝑡≤𝑧 𝑗𝑡 𝑤𝑖𝑗𝑡(1 − 𝑦𝑖𝑗𝑡/𝑧𝑗𝑡) 3. Watts Index: 𝑊𝑗𝑡 = 𝑦 𝑖𝑗𝑡≤𝑧 𝑗𝑡 𝑤𝑖𝑗𝑡 ln(𝑧𝑗𝑡/𝑦𝑖𝑗𝑡) 9/23/16 Brown, Ravallion & van de Walle 19
  20. 20. Can we do better? 9/23/16 Brown, Ravallion & van de Walle 20
  21. 21. Can we do better by adding extra covariates? • “Basic PMT”: Basic consumer durables and assets, demographic variables, attributes of the head • “Extended PMT”: HH water source; housing materials; separate room for cooking; electricity; more details on the characteristics of the household • But maybe the more important margin for improving PMT is the method of aggregating information => 9/23/16 Brown, Ravallion & van de Walle 21
  22. 22. Can we do better by tailoring the estimator to the goal of poverty reduction? • Regression gives equal weight to all error variances, whether for poor or rich: two options considered: • Quantile regression setting the quantile at the overall poverty rate • Poverty-weighted least-squares (PLS) • Calibrate the PMT using OLS with poor households only • Also a model adding some h’holds above the poverty line (“poor + 20”) • Many other options (such as lasso regression). Instead we characterize the optimally differentiated transfers for the same information set. 9/23/16 Brown, Ravallion & van de Walle 22
  23. 23. How much better with optimal transfers? 9/23/16 Brown, Ravallion & van de Walle 23  Think of the information-feasible transfer as a function of the observed x’s.  The problem is to choose the parameters of a score for assigning the real- world transfers based on the x’s:    m k k ijt k jtijt x 0 0)(    The choice of the score parameters k jt is made to minimize the Watts index:   jtN i ijtijtjtijtjt yzwW 1 ))]/(ln(,0max[   The choice is constrained by the budget:    jtN i ijtijtjt wB 1   We solve this problem numerically (using Matlab’s “fmincon” program).
  24. 24. Data and measurement assumptions 9/23/16 Brown, Ravallion & van de Walle 24
  25. 25. Data from the LSMS Data from the World Bank’s Living Standards and Measurement Surveys • Burkina Faso, Ethiopia, Ghana, Malawi, Mali, Niger, Nigeria, Tanzania, Uganda • Survey years range from 2009 to 2014 • Between 2,600 and 10,000 observations per country 9/23/16 Brown, Ravallion & van de Walle 25
  26. 26. Measurement assumptions • Log per capita household consumption as dep. var. (spatially deflated) • Three poverty rates: 𝐻𝑗𝑡 = 0.2 and 0.4 and national poverty rate • 40% approximates average poverty rate in SSA • 20% allows for focus on the poorest households • Cut-off points for PMT scores calculated in two ways: 1. Fix poverty line: 𝑦𝑗𝑡 ≤ 𝑧𝑗𝑡 , 𝑧𝑗𝑡 ≡ 𝐹𝑗𝑡 −1 0.2 or 𝐹𝑗𝑡 −1 0.4 2. Fix poverty rate: assign lowest 20 or 40% of 𝑦𝑗𝑡 as poor • Under method 2, 𝐼𝐸𝑅 𝑗𝑡 = 𝐸𝐸𝑅 𝑗𝑡 • Each time a poor household is incorrectly identified as non-poor, a non-poor household is identified as poor (since the total rate of poor is kept constant) • We call this the targeting error rate, 𝑇𝐸𝑅 𝑗𝑡 9/23/16 Brown, Ravallion & van de Walle 26
  27. 27. Results 9/23/16 Brown, Ravallion & van de Walle 27
  28. 28. Basic PMT Variables in 𝑥𝑖𝑗𝑡 include: • Type of toilet a household has • Floor, wall and roofing material • Type of fuel most commonly used for cooking • Characteristics of the household head, including gender, education and occupation • Household’s religion and demographic composition • Dummy variables for categories of household size, age of head, month of survey and region of residence Average of 𝑅2 is 0.54, with a range from 0.32 (for Ethiopia) to 0.65 (Malawi) • Similar to what is found in existing literature 9/23/16 Brown, Ravallion & van de Walle 28
  29. 29. Covariates (x) 9/23/16 Brown, Ravallion & van de Walle 29 Basic Extras in Extended Toilet: flush Water: piped Toilet: pit Water: well or outside pipe Floor: finished Persons per room Wall: finished Separate room for kitchen Roof: finished Has radio Fuel: electricity gas or kerosene Has television Fuel: charcoal Has fridge or freezer Urban area Has bicycle Female head Has motorbike Head has at least primary education Has car or truck Head has at least secondary education Has telephone (owns) Divorced or separated head Has mobile phone (owns) Widowed head Has computer Head worked in a paid job (last week) Has video player Head worked in own non-farm enterprise Has any type of stove Muslim Has sewing machine Christian Has air conditioner or fan HH share: girls 5 and under Has iron HH share: boys 5 and under Has satellite dish HH share: girls 6 to 14 Has electric generator HH share: boys 6 to 14 Household owns dwelling HH share: women 65 and over HH share: widows HH share: disabled men 15+ HH share: disabled women 15+ HH share: orphan boys less than 15 HH share: orphan girls less than 15
  30. 30. Summary stats on regressions 9/23/16 Brown, Ravallion & van de Walle 30 Burkina Faso Ethiopia Ghana Malawi Mali Niger Nigeria Tanzania Uganda Basic PMT R2 0.644 0.319 0.561 0.573 0.418 0.634 0.581 0.585 0.498 N 10265 5017 4224 3931 3212 3833 3720 4753 2650 Extended PMT R2 0.724 0.381 0.587 0.674 0.520 0.718 0.666 0.647 0.596 N 10265 5017 4224 3931 3212 3833 3720 4753 2650
  31. 31. Actual and predicted values for Basic PMT 9/23/16 Brown, Ravallion & van de Walle 31 Exclusion errors Inclu- sion errors
  32. 32. Results for Basic PMT: Targeting errors 32 Inclusion error rate Exclusion error rate Inclusion error rate Exclusion error rate Targeting Error Targeting Error (IER) (EER) (IER) (EER) (TER) (TER) Fixed poverty line Fixed poverty rate z = F -1 (0.2) z = F-1 (0.4) H = 0.2 H = 0.4 Burkina Faso 0.401 0.751 0.304 0.375 0.522 0.329 Ethiopia 0.515 0.945 0.396 0.562 0.621 0.413 Ghana 0.354 0.628 0.257 0.350 0.428 0.288 Malawi 0.431 0.880 0.333 0.451 0.553 0.373 Mali 1.000 1.000 0.348 0.485 0.553 0.375 Niger 0.539 0.875 0.384 0.340 0.584 0.362 Nigeria 0.332 0.548 0.247 0.243 0.392 0.244 Tanzania 0.396 0.822 0.323 0.291 0.513 0.314 Uganda 0.357 0.663 0.350 0.294 0.455 0.335 Mean 0.481 0.807 0.309 0.359 0.505 0.319 About half those identified as poor are not About 80% of the poor are not identified as poor
  33. 33. Basic PMT reduces inclusion errors, at a cost to exclusion errors • For a poverty rate of 20% and a fixed line, the PMT method has nearly halved the IER that would be obtained with a uniform transfer payment. • However, this has come at the expense of exclusion errors. The average EER is sizeable, with 81% of those who are in the poorest 20% in terms of survey- based consumption being incorrectly identified as non-poor by PMT. • There is considerable variation across countries, with IER ranging from 33% to 100%, and EER from 55% to 100%. 9/23/16 Brown, Ravallion & van de Walle 33
  34. 34. Residuals for Basic PMT against log consumption 9/23/16 Brown, Ravallion & van de Walle 34 Figure 2: Residuals for Basic PMT and log consumption per capita
  35. 35. Econometric targeting using OLS substantially overestimates consumption by the poorest • As expected, the residuals tend to be lower (more negative) for poor people. • What is notable is how much the PMT regression is over-estimating the living standards of the poorest. • For the poorest 20% in terms of actual consumption, the mean residual ranges from -0.73 to -0.37, implying that the PMT regressions yield predicted consumptions for the poor between 50% and 100% above their actual consumption! 9/23/16 Brown, Ravallion & van de Walle 35
  36. 36. Results for poverty-focused methods • Two “poverty-focused” options to OLS: • Quantile regression using the poverty rate as the quantile • “Poverty-weighted least squares” (PLS) • Both => reduction in exclusion error rate using a fixed poverty line • This comes at the cost of higher inclusion errors, especially for the lower poverty line • The PLS method is better at covering the poor but has very high inclusion errors • The “poor + 20” PLS regression reduces the inclusion error relative to the PLS, but also increases the exclusion error 9/23/16 Brown, Ravallion & van de Walle 36
  37. 37. Results for Extended PMT • Adding variables increases 𝑅2 and improves targeting. But by how much? • Additional variables in 𝑥𝑖𝑗𝑡 include: • Household’s water source • More details on housing conditions, e.g. number of household members per room; separate room for cooking • Whether the household has electricity • Household assets • As expected, values of 𝑅2 are higher • Simple average is 0.61 with a range from 0.72 (Burkina) to 0.38 (Ethiopia) • In most cases the gains are relatively small, though the number of explanatory variables has almost doubled 9/23/16 Brown, Ravallion & van de Walle 37
  38. 38. Targeting errors for Extended PMT 9/23/16 Brown, Ravallion & van de Walle 38 Inclusion error rate Exclusion error rate Inclusion error rate Exclusion error rate Targeting Error Targeting Error (IER) (EER) (IER) (EER) (TER) (TER) Fixed poverty line Fixed poverty rate z = F -1 (0.2) z = F-1 (0.4) H = 0.2 H = 0.4 Burkina Faso 0.334 0.626 0.257 0.336 0.449 0.282 Ethiopia 0.419 0.857 0.373 0.455 0.541 0.405 Ghana 0.349 0.591 0.256 0.331 0.421 0.267 Malawi 0.439 0.698 0.295 0.374 0.470 0.315 Mali 0.444 0.951 0.344 0.444 0.572 0.341 Niger 0.458 0.751 0.328 0.323 0.539 0.327 Nigeria 0.330 0.496 0.228 0.217 0.384 0.223 Tanzania 0.403 0.670 0.283 0.279 0.481 0.281 Uganda 0.379 0.552 0.313 0.279 0.467 0.307 Mean 0.362 0.639 0.283 0.308 0.456 0292 Basic PMT Mean 0.481 0.807 0.309 0.359 0.505 0.319
  39. 39. Further results (see paper for details) • Using panel data for a subset of countries to measure time-mean consumption. • We also estimated the model separately for urban and rural households • Performs slightly worse than when model is estimated with full sample • Consider additional models with more variables than the Extended PMT • Food security variables • Community-level variables • Provide only modest improvements on targeting performance • Compare performance with “pruned” models • Stepwise regression and Extended PMT variables • Similar results to the Basic PMT 9/23/16 Brown, Ravallion & van de Walle 39
  40. 40. Targeting differentials • The mean NTD for Basic PMT is 0.21, meaning that if program participation was based on the PMT scores the participation rate for the poor would be 21% points higher than that for the non- poor. • The poverty-quantile method yields the highest NTD, at around 0.49 on average for Basic PMT, rising to 0.53 for the Extended PMT. • For all nine countries, the poverty- quantile regression method comes out best by this measure. 9/23/16 Brown, Ravallion & van de Walle 40 Mean Basic PMT covariates Basic PMT 0.214 Using means from panel data 0.366 Poverty quantile regression 0.485 Poverty weighted: Poor only 0.107 Poverty weighted: Poor + 20 0.421 PMT with Urban/Rural 0242 Extended PMT covariates Extended PMT 0.309 Using means from panel data 0.334 Poverty quantile regression 0.531 Poverty weighted: Poor only 0.196 Poverty weighted: Poor + 20 0.484 Stepwise (p=0.01) 0.249 HH Shocks + Food Security 0.333 Shocks, Food Security + Community Variables 0.331
  41. 41. Impacts on poverty for stylized transfer programs Setting the budget at each country’s aggregate poverty gap, we compare how well the different PMT specifications do at reducing the three poverty measures considered • We assume a poverty line corresponding to H = 0.2 • A uniform per capita transfer is given to all households who are predicted to be below the poverty line according to the PMT • Aggregate poverty gap divided by total number of people predicted to be poor • This transfer amount is distributed to households according to their size 9/23/16 Brown, Ravallion & van de Walle 41
  42. 42. Comparisons with other targeting methods We also compare the poverty-reducing performance of PMT with other targeting methods: • Uniform transfer (“basic income”) • Every household receives a transfer adjusted for household size • Categorical targeting • Use several different categories, for example elderly, widows, households with children • If a household has two members who fit the category, or one person fits two categories, the transfer amount is doubled (and so on for three or more categories) 9/23/16 Brown, Ravallion & van de Walle 42
  43. 43. Impacts on poverty rate (H=20%) • Most methods bring H down to 15-16%; more than 3/4 of the poor remain poor. • Basic PMT does only slightly better than UBI with the same budget, • Basic PMT does not do as well as the universal transfer in one third of countries. • Quantile regression does better, bringing the poverty rate down to about one % point below that for Basic PMT. • Categorical targeting does well. Targeting to h’holds with elderly, widows, disabled and children does as well as Basic PMT. • The poverty quantile regression method typically has the greatest impact on poverty. 9/23/16 Brown, Ravallion & van de Walle 43 Mean Universal (basic income) 0.171 Basic PMT covariates Basic PMT 0.163 Using means from panel data 0.159 Poverty quantile regression 0.155 Poverty weighted: Poor only 0.170 Poverty weighted: Poor + 20 0.162 PMT with Urban/Rural 0.159 Extended PMT covariates Extended PMT 0.154 Using means from panel data 0.155 Poverty quantile regression 0.154 Poverty weighted: Poor only 0.168 Poverty weighted: Poor + 20 0.157
  44. 44. Allowing for lags in implementation • Use panel data; between 1 and 2 years apart • Calibrate PMT using Round 1; two different lags 1. Use Round 2 covariates to determine eligibility – lag in estimation only 2. Use Round 1 covariates – lag in implementation • In general lags increase targeting errors, and make impact of program more uncertain • On average, poverty reduction is hampered by one percentage point on average due to lags 9/23/16 Brown, Ravallion & van de Walle 44
  45. 45. Allowing optimally differentiated transfers • So far: standard practice of giving the same transfer payment to all those predicted to be poor using PMT. • Differentiating the transfers could be expected to work better if the predicted poverty gaps are quite accurate. • However, we have already seen that this is not the case—that PMT works poorly in predicting the levels of living of the poorest. • So it is unclear on a priori grounds whether differentiated transfers will have larger impacts on poverty. • How much does the constraint of relying on uniform transfers to the “predicted poor” limit the effectiveness of PMT? 9/23/16 Brown, Ravallion & van de Walle 45
  46. 46. 9/23/16 Brown, Ravallion & van de Walle 46 Optimal transfers PMT Linear Non-linear Burkina Faso 0.044 0.038 0.036 Ethiopia 0.074 0.056 0.053 Ghana 0.066 0.048 0.044 Malawi 0.060 0.051 0.044 Mali 0.064 0.057 0.053 Niger 0.048 0.042 0.039 Nigeria 0.062 0.049 0.043 Tanzania 0.069 0.060 0.051 Uganda 0.083 0.061 0.053 Mean 0.066 0.054 0.049 Watts index post transfer using differentiated transfers
  47. 47. Modest gain only using differentiated transfers • Recall: optimal allocation based on the set of PMT covariates. Watts index is the objective function. • Objective function tends to be quite flat in the sub-set of the parameter space corresponding to the various solutions found. • Overall, filling the predicted gaps brings only a small further reduction in the poverty measures. 9/23/16 Brown, Ravallion & van de Walle 47
  48. 48. How well can we identify poor individuals using standard household level data? 9/23/16 Brown, Ravallion & van de Walle 48 Based on Caitlin Brown, Martin Ravallion and Dominique van de Walle, “Are Poor Individuals Mainly Found in Poor Households? Evidence Using Nutrition Data for Africa.” NBER Working Paper
  49. 49. Poverty is an individual deprivation, but individual consumption is rarely observed • Even if we can reach poor households, that does not guarantee that we reach poor individuals. And poverty is ultimately an individual state. • Unitary model of h’hold (single utility fn) finds little support empirically; various alternatives (Chiappori and Mazzocco; Baland and Ziparo). • These alternative models permit new sources of inequality within households, such as in reservation utility levels. • An extensive literature details intra-household inequalities in resource allocations and outcomes (reviewed in World Bank 2012). • Targeting poor h’holds may miss may disadvantaged individuals and targeted h’holds may not allocate the benefits to the neediest. 9/23/16 Brown, Ravallion & van de Walle 49
  50. 50. One clue to individual poverty: nutritional status • Anthropometric data at individual level • Nutritionally-vulnerable women and children are a high priority for social policies • Malnutrition has both immediate and long-term social and economic costs • However, individual level interventions are often intrusive and expensive • Difficult to observe undernourishment in large populations • Key question: is household-level targeting effective at reaching poor individuals as indicated by nutritional status? 9/23/16 Brown, Ravallion & van de Walle 50
  51. 51. Evidence on wealth and nutrition • ”Wealth effect” suggests poorer households are more likely to include undernourished individuals • Behrman and Deolalikar (1987); Ravallion (1990, 1992); Filmer and Pritchett (1999, 2000); Filmer and Scott (2012) • However, even strong correlations between wealth and nutrition may not mean wealth is a satisfactory predictor of individual nutrition • Other factors may mean non-poor households also include the undernourished • Intra-household inequality; local health environment; heterogeneous responses to shocks
  52. 52. Intra-household inequality constrains performance, but how much? • Our theoretical model formalizes the intuition that heterogeneity due to intra-household inequality or the local health environment diminishes the scope for reaching poor individuals (nutritionally vulnerable women and children) by targeting poor households. • But how much does this matter empirically? Is the wealth effect on individual nutritional status strong enough to allow satisfactory targeting of vulnerable women and children? 9/23/16 Brown, Ravallion & van de Walle 52
  53. 53. Data sources Demographic and Health Surveys (DHS) for 30 countries in SSA • 350,000 women and children • Women 15 to 49; children under 5 • Nutritional outcomes: BMI, height-for-age (stunting) and weight-for-height (wasting) • DHS wealth index • Additional household- and individual-level variables Wealth index constructed using variables related to a household’s welfare • Household assets, materials used in housing construction, access to water and sanitation etc. • Country specific, mean zero and standard deviation of one
  54. 54. Data cont. • Household consumption may be a better indicator of individual nutritional status than wealth index • DHS does not collect consumption data • Also draw from Living Standards and Measurement (LSMS) surveys • 7 countries with nutritional data • Focus on bottom 20% and 40% of households based on wealth index • 40% roughly coincides with overall poverty rate in SSA • Include other indicators relevant to nutritional outcomes • GDP per capita, poverty rate, female literacy rate, access to improved water and sanitation.
  55. 55. Nutritional outcomes and household wealth
  56. 56. Regression coefficients of individual nutritional outcomes on the DHS household wealth index BMI Height-for-age Weight-for-height Benin 0.194*** 0.189*** 0.080*** Burkina Faso 0.279*** 0.281*** 0.097*** Burundi 0.242*** 0.505*** 0.115*** Cameroon 0.285*** 0.451*** 0.257*** Congo 0.265*** 0.292*** 0.051** Cote D'Ivoire 0.203*** 0.279*** 0.043 DRC 0.276*** 0.378*** 0.073*** Ethiopia 0.352*** 0.374*** 0.229*** Gabon 0.182*** 0.397*** 0.056**
  57. 57. Proportion of undernourished individuals in the poorest 20% and 40% of households Poorest 20% of households Poorest 40% of households Underweight women Stunted children Wasted children Underweight women Stunted children Wasted children Benin 0.248 0.233 0.223 0.444 0.446 0.464 Burkina Faso 0.307 0.242 0.224 0.551 0.458 0.433 Burundi 0.276 0.249 0.281 0.464 0.451 0.506 Cameroon 0.396 0.326 0.364 0.637 0.594 0.630 Congo 0.221 0.310 0.232 0.460 0.534 0.465 Cote D'Ivoire 0.226 0.289 0.240 0.414 0.516 0.447 DRC 0.252 0.247 0.209 0.521 0.482 0.442 Ethiopia 0.235 0.218 0.259 0.461 0.445 0.534 Gabon 0.246 0.434 0.206 0.422 0.634 0.388
  58. 58. Joint probabilities of being undernourished and wealth poor Poorest 20% of Households Poorest 40% of Households Underweight women Stunted children Wasted children Underweight women Stunted children Wasted children Benin 0.016 0.095 0.032 0.028 0.181 0.067 Burkina Faso 0.047 0.072 0.031 0.085 0.136 0.060 Burundi 0.044 0.129 0.014 0.074 0.233 0.026 Cameroon 0.027 0.091 0.018 0.043 0.166 0.031 Congo 0.032 0.058 0.012 0.066 0.100 0.023 Cote D’Ivoire 0.018 0.069 0.017 0.032 0.123 0.032 DRC 0.036 0.090 0.015 0.075 0.177 0.032 Ethiopia 0.063 0.085 0.022 0.123 0.173 0.046 Gabon 0.018 0.057 0.007 0.031 0.083 0.012
  59. 59. Most of Africa’s nutritionally vulnerable women and children are not found in poor households • On average, 75% of undernourished individuals are not found in the poorest 20% of households using household wealth. • Similar results using consumption (slightly better for women) • About half of undernourished women and children are not found in the bottom 40%. • Joint probabilities for underweight women and wasted children close to zero. • Results are robust to: • Adjusting for demographic imbalance of wealth fractiles • Reweighting the wealth index • Adding additional household-level variables • Adding individual-level variables
  60. 60. How do these results vary with the rate of undernutrition? • Are a higher proportion of undernourished found in poor households in countries with a high rate of undernourishment? • Look at relationship between Pr 𝑤 < 1|𝑛 < 1 and Pr 𝑛 < 1 => 9/23/16 Brown, Ravallion & van de Walle 60
  61. 61. Countries with fewer underweight women have a higher proportion of those women in wealth-poor households .0 .1 .2 .3 .4 .5 .6 .7 .00 .05 .10 .15 .20 .25 .30 Poorest 20% (r=-0.31) Poorest 40% (r=-0.22) Share of women who are underweight Proportionofunderweightwomenfound inpoorhouseholds
  62. 62. Countries with fewer stunted children have a higher proportion of those children in wealth-poor households .0 .1 .2 .3 .4 .5 .6 .7 .0 .1 .2 .3 .4 .5 Incidence of stunted children Proportionofstuntedchildrenfound inwealth-poorhouseholds Poorest 20% (r=-0.47) Poorest 40% (r=-0.56)
  63. 63. Conclusions 9/23/16 Brown, Ravallion & van de Walle 63
  64. 64. 64 Exclusion errors Inclusion errors Targeting trade off • Strengths and weaknesses of econometric targeting • It substantially reduces inclusion errors in an antipoverty program • But this comes at the cost of a large rise in exclusion errors Uniform-untargeted (basic income) Basic PMT New PMT methods • The choice of method ultimately rests on the weight policy makers give to these two sources of errors • High weight on poverty => higher weight on reducing exclusion errors
  65. 65. Our findings on econometric targeting • This filters out the nonpoor but excludes many poor people, diminishing impact on poverty (relative to untargeted transfers). • More data and better methods (geared towards poverty reduction) do better, although the gains to the poor are typically modest. • In some cases, much simpler targeting methods dominate. • However, no method we consider performs especially well when poverty reduction is the goal. • Even with sufficient budget to eliminate poverty with full information, none of the methods bring the poverty rate below 75% of its initial value. • Prevailing methods do not reliably reach the poorest. 9/23/16 Brown, Ravallion & van de Walle 65
  66. 66. Household data are weak in identifying poor individuals • The bulk of individually poor women and children—as indicated by nutritional status—are not found in observationally poor households • On average, about 75% of undernourished women and children are not found in the poorest 20% using the wealth index • When the rate of undernourishment is high, these individuals are more spread out across the wealth distribution • Adding variables e.g. household-level, individual-level improves targeting performance, but still misses many. • Results are robust to: • Reweighting the wealth index • Adding additional household-level variables • Adding individual-level variables 9/23/16 Brown, Ravallion & van de Walle 66
  67. 67. Overall conclusion: Eliminating poverty using transfers is much harder than the total poverty gap suggests • With budget = aggregate poverty gap we will not be able to eliminate household poverty with the data routinely used by antipoverty policies. • And even when we have correctly identified poor households, the majority of poor individuals (as indicated by nutritional status) will be missed. • Much better data and/or much better targeting methods will be needed. 9/23/16 Brown, Ravallion & van de Walle 67
  68. 68. Time to reconsider basic income • Push for fine targeting has dominated social policy discussions in developing countries. 9/23/16 Brown, Ravallion & van de Walle 68 0 20 40 60 80 100 0 2000 4000 6000 8000 10000 12000 14000 16000 18000 20000 22000 GDP per capita at PPP for year of survey Safety net coverage for poorest quintile (%) Safety net coverage for whole population (%) Poorest quintile Population • However, the information constraints are severe. Weak administrative capabilities. • (Other constraints: incentive effects and political economy.) • Basic (full) income should be on the menu of options in developing countries. • India example: Basic income with same budget as NREGA has more impact on poverty! • Promotes social cohesion and sustainability of social policies. • So too should the development of administrative capabilities over longer term.
  69. 69. Further reading: Martin Ravallion, The Economics of Poverty: History, Measurement and Policy Oxford University Press, 2016 economicsandpoverty.com Thank you for your attention!

×