Targeting with Agents Paul Niehaus Antonia Attanassova Marianne Bertrand Sendhil Mullainathan January 29, 2012 1 / 29
An Example Measure 2 of households, half poor and half rich Net benefit b from giving slots to the poor Net cost c from giving slots to the rich 2 / 29
An Example Universal eligibility pays off b c 3 / 29
An Example Targeting via a Proxy Means Test (PMT) can potentially help Large literature uses statistical decision theory to design targeting rules Let s suppose perfect targeting is (statistically) feasible. Gain of +c 4 / 29
Problem Why might this conclusion be too simplistic? 5 / 29
Problem Why might this conclusion be too simplistic? Officials often break rules Divert transfers (Reinikka & Svensson 2004; Olken 2006) Inflate participation figures (Niehaus & Sukhtankar 2009) Charge eligible individuals for permits (Svensson 2003) Sell permits to ineligible individuals (Bertrand et al. 2007) 5 / 29
An Example Suppose that with imperfect enforcement, universal eligibility pays off q U (b c) 6 / 29
An Example How will targeting affect the rich? If enforcement is at all effective, the rich will get fewer slots under targeting This yields a gain relative to universal eligibility of +(q U q I )c 7 / 29
An Example How will targeting affect the poor? One possibility: the poor are still eligible and hence unaffected (q E = q U ) In this case targeting unambiguously outperforms universal eligibility as before 8 / 29
An Example What if give slots to the poor is harder to enforce than give slots to everyone? The poor receive fewer slots even though their eligibility is unchanged Targeting is best only if (q U q I )c (q U q E )b > 0 There is a tradeoff between statistical accuracy and enforceability 9 / 29
Overview Theoretical framework for thinking about targeting with agents Three key questions 1. Is enforcement strong? 2. Are preferences aligned? 3. Is eligibility a sufficient statistic for allocation? If all no s, tradeoff between statistical accuracy and enforceability 10 / 29
Overview Theoretical framework for thinking about targeting with agents Three key questions 1. Is enforcement strong? 2. Are preferences aligned? 3. Is eligibility a sufficient statistic for allocation? If all no s, tradeoff between statistical accuracy and enforceability Empirical assessment for Below Poverty Line cards in Karnataka, India Observe rule violations and (in most cases) prices charged 10 / 29
Overview Theoretical framework for thinking about targeting with agents Three key questions 1. Is enforcement strong? No frequent violations 2. Are preferences aligned? 3. Is eligibility a sufficient statistic for allocation? If all no s, tradeoff between statistical accuracy and enforceability Empirical assessment for Below Poverty Line cards in Karnataka, India Observe rule violations and (in most cases) prices charged 10 / 29
Overview Theoretical framework for thinking about targeting with agents Three key questions 1. Is enforcement strong? No frequent violations 2. Are preferences aligned? No violations yield a less progressive distribution 3. Is eligibility a sufficient statistic for allocation? If all no s, tradeoff between statistical accuracy and enforceability Empirical assessment for Below Poverty Line cards in Karnataka, India Observe rule violations and (in most cases) prices charged 10 / 29
Overview Theoretical framework for thinking about targeting with agents Three key questions 1. Is enforcement strong? No frequent violations 2. Are preferences aligned? No violations yield a less progressive distribution 3. Is eligibility a sufficient statistic for allocation? No degrees of ineligibility matter If all no s, tradeoff between statistical accuracy and enforceability Empirical assessment for Below Poverty Line cards in Karnataka, India Observe rule violations and (in most cases) prices charged 10 / 29
Empirical Context Below Poverty Line cards in India Determine eligibility for a wide range of social programs, notably the Public Distribution System 33% of Indian households hold BPL cards (Dreze & Kheera 2010) Allocation process Central government estimates poverty, determines transfers to states States then choose their own eligibility criteria (and typically estimate more BPL households) 11 / 29
Eligibility in Karnataka One must not have Annual income more than Rs. 17,000 in urban areas or Rs. 12,000 in rural areas; A telephone (land line or mobile); A two-, three- or four-wheeler (e.g. motorcycle, auto-rickshaw, or car); A gas connection; A color TV; More than 5 acres of dry land; A water pump set; A household member who is a salaried government employee. Note this is a linear scoring rule with threshold 1 12 / 29
Karnataka s Eligibility Determination 2007 eligibility survey by Gram Panchayat officials (village accountant, teacher, health worker) Transparency measures Gram Sabha Public posting of eligibility status Enforcement via follow-up audits 13 / 29
Our Independent Assessment 812 villages Simple random sample of villages in 13 districts Treatment and control villages from a (failed) experiment in Raichur, Kolar districts 14,074 households Survey Sampled 21 per village with replacements; yielded 17 hhds/village Demographics and assets first, BPL status and politics last Bribe questions: what was the official fee and what (if any) extra fees were required 14 / 29
Process Implemented Process appears broken 50% recall being surveyed to determine eligibility 13% recall a Gram Sabha to discuss eligibility (62% recall no GS at all) 16% of these 13% said households could raise objections at the GS 2% said an eligibility list was posted in the village Awareness is low 35% self-described as familiar with the rules Accuracy on eligibility criteria ranged from 19% to 77%, on average exactly 50% 17% self-reported knowing what to do if they disagreed 15 / 29
Plan of Attack 1. How strong is enforcement? 2. How aligned are officials preferences? 3. Do degrees of (in)eligibility matter? 16 / 29
Plan of Attack 1. How strong is enforcement? 2. How aligned are officials preferences? 3. Do degrees of (in)eligibility matter? How would you test for the strength of enforcement? 17 / 29
Regression Framework We will use a modified version of the pricing equation from our model p hv = fh(x hv )+(α α)logy hv +λ v +η hv x hv are the eligibility criteria, y hv is income, and the λ v are village fixed effects We will also estimate analogous linear probability models for allocation a hv We worry about unobserved η hv placebo test using fake exclusion criteria: electricity, black and white television, bicycle. 18 / 29
How Strong is Enforcement? Allocation: overall 48% are misclassified (42% ignoring the income threshold) Pricing: Ineligible Eligible Total No BPL Card 2560 652 3212 (30%) (13%) (24%) Has BPL Card 5862 4419 10281 (70%) (87%) (76%) Total 8422 5071 13493 73% of households and 93% of BPL hhds reported the price charged Bribes are common: 75% of reported prices exceed Rs. 5 (0.2% of reported prices are less than Rs. 5) Bribes are small: mean Rs. 9, max Rs. 305 19 / 29
Regressor 1 2 3 4 5 6 7 Panel A: Prices Ineligible 2.935 0.868 2.002 (0.509) (0.668) (0.477) # Violations 1.277 1.059 1.029 1.281 1.009 (0.253) (0.343) (0.298) (0.255) (0.296) # Placebo Violations -0.222-0.35 (0.343) (0.366) Log Annual Income 1.564 0.946 1.05 (0.507) (0.577) (0.611) N 9608 9608 9608 9608 9608 9608 9608 Panel B: Quantities Ineligible -0.013 0.003-0.006 (0.003) (0.005) (0.003) # Violations -0.008-0.009-0.006-0.008-0.006 (0.002) (0.002) (0.002) (0.002) (0.002) # Placebo Violations -0.005-0.004 (0.003) (0.003) Log Annual Income -0.012-0.006-0.005 (0.005) (0.005) (0.005) N 9608 9608 9608 9608 9608 9608 9608 Panel C: Quantities Ineligible -0.215 0.008-0.107 (0.01) (0.012) (0.01) # Violations -0.097-0.099-0.079-0.096-0.081 (0.003) (0.004) (0.004) (0.003) (0.004) # Placebo Violations -0.038-0.03 (0.005) (0.006) Log Annual Income -0.146-0.062-0.054 (0.009) (0.009) (0.009) N 13183 13183 13183 13183 13183 13183 13183
Plan of Attack 1. How strong is enforcement? 2. How aligned are officials preferences? 3. Do degrees of (in)eligibility matter? How would you test whether the officials preferences are aligned with the government s? 21 / 29
How Aligned are Officials Preferences? 0.0 0.2 0.4 0.6 0.8 1.0 Eligible BPL 8 9 10 11 Log income is correlated 0.55 with eligibility, 0.23 with BPL status Similar picture within-villages, omitting income, etc. Faithful targeting on any one of phone ( 0.37), water pump ( 0.32), land ( 0.31), or gas connection ( 0.30) would do better 22 / 29
Regressor 1 2 3 4 5 6 7 Panel A: Prices Ineligible 2.935 0.868 2.002 (0.509) (0.668) (0.477) # Violations 1.277 1.059 1.029 1.281 1.009 (0.253) (0.343) (0.298) (0.255) (0.296) # Placebo Violations -0.222-0.35 (0.343) (0.366) Log Annual Income 1.564 0.946 1.05 (0.507) (0.577) (0.611) N 9608 9608 9608 9608 9608 9608 9608 Panel B: Quantities Ineligible -0.013 0.003-0.006 (0.003) (0.005) (0.003) # Violations -0.008-0.009-0.006-0.008-0.006 (0.002) (0.002) (0.002) (0.002) (0.002) # Placebo Violations -0.005-0.004 (0.003) (0.003) Log Annual Income -0.012-0.006-0.005 (0.005) (0.005) (0.005) N 9608 9608 9608 9608 9608 9608 9608 Panel C: Quantities Ineligible -0.215 0.008-0.107 (0.01) (0.012) (0.01) # Violations -0.097-0.099-0.079-0.096-0.081 (0.003) (0.004) (0.004) (0.003) (0.004) # Placebo Violations -0.038-0.03 (0.005) (0.006) Log Annual Income -0.146-0.062-0.054 (0.009) (0.009) (0.009) N 13183 13183 13183 13183 13183 13183 13183
Plan of Attack 1. How strong is enforcement? 2. How aligned are officials preferences? 3. Do degrees of (in)eligibility matter? How would you test whether degrees of (in)eligibility matter? 24 / 29
Do Degrees of (In)eligibility Matter? Mean BPL Fee 0 10 20 30 40 0 2 4 6 8 Proportion with BPL Cards 0.0 0.4 0.8 0 2 4 6 8 Number of Violations Number of Violations Prices and allocations move monotonically with number of criteria violated 25 / 29
Regressor 1 2 3 4 5 6 7 Panel A: Prices Ineligible 2.935 0.868 2.002 (0.509) (0.668) (0.477) # Violations 1.277 1.059 1.029 1.281 1.009 (0.253) (0.343) (0.298) (0.255) (0.296) # Placebo Violations -0.222-0.35 (0.343) (0.366) Log Annual Income 1.564 0.946 1.05 (0.507) (0.577) (0.611) N 9608 9608 9608 9608 9608 9608 9608 Panel B: Quantities Ineligible -0.013 0.003-0.006 (0.003) (0.005) (0.003) # Violations -0.008-0.009-0.006-0.008-0.006 (0.002) (0.002) (0.002) (0.002) (0.002) # Placebo Violations -0.005-0.004 (0.003) (0.003) Log Annual Income -0.012-0.006-0.005 (0.005) (0.005) (0.005) N 9608 9608 9608 9608 9608 9608 9608 Panel C: Quantities Ineligible -0.215 0.008-0.107 (0.01) (0.012) (0.01) # Violations -0.097-0.099-0.079-0.096-0.081 (0.003) (0.004) (0.004) (0.003) (0.004) # Placebo Violations -0.038-0.03 (0.005) (0.006) Log Annual Income -0.146-0.062-0.054 (0.009) (0.009) (0.009) N 13183 13183 13183 13183 13183 13183 13183
Summary 1. Enforcement is weak 2. Weak evidence on soft targeting 3. Degrees of ineligibility matter Coarser but more enforceable rules are worth trying 27 / 29
An Emerging Trend? Jean Dreze & Reetika Khera (EPW, 2010)...we consider four simple ways of combining exclusion and inclusion criteria to construct a SAB list (analogous to the current BPL list ). A common feature of these different approaches is that every household can attribute its inclusion in, or exclusion from, the list to a single criterion. This would involve statements such as, I am on the SAB list because I am landless, or I am not on the SAB list because I own a car. This feature can be of great help in facilitating participatory verification of the BPL list, and in preventing fraud. In this respect, the primary method contrasts with the current scoring methods... 28 / 29
Discussion How would you test whether simpler rules really do work better? Do you think simplicity would still be important if bottom-up complaints and not top-down audits were the main enforcement mechanism? What would you want to know about Karnataka to understand whether the results might be externally valid for another Indian state? Does ease of determining who is eligible matter in the US? What are some other examples of things governments do that they might need to do differently in a setting with corruption? 29 / 29