Public Finance II: Targeting

14.740x: Foundations of Development Policy

Professor Ben Olken

Introduction

Redistribution and safety nets are common in developed countries,

and becoming increasingly common in developing countries
Examples from the US?
Examples from developing countries?

Olken () PF Lecture 2 2 / 34
Targeting

Basic challenge of implementing these programs: lack of information

about who is really poor.
This is a problem everywhere.
In the US literature, the problem is typically framed that we observe
income, not true earning ability.
Optimal taxes are set taking into account this asymmetric information
(Mirrlees 1971, Saez 2001).
If we know more characteristics about individuals that predict poverty
(e.g., widowhood), we can “tag” these individuals and assign them
di¤erent tax schedules (Ackerlof 1978)
But the problem is particularly severe in developing countries: we
don’t even observe income!
So what can you do? Ideas?

Olken () PF Lecture 2 3 / 34
Targeting

Targeting options if income is not observable:

Proxy-means tests (more generalized version of “tagging”)
Community-based targeting
Self-targeting

Olken () PF Lecture 2 4 / 34
Proxy-Means Tests

This is the main way individual targeting is done in most developing

countries. (E.g, Progresa).
Concept: can’t target based on consumption directly, since you could
easily lie on a survey
Instead: do a survey where you collect data on assets (land, house,
motorcycle, etc)
Assets capture permanent component of income
And they are hard to falsify on a survey
Use survey data to estimate relationship between consumption and
assets, and used predicted consumption for targeting
Problems
R2 much less than 1, so you don’t get poverty exactly right (horizontal
equity)
Corruption among surveyors
Costly: need to do a census
Olken () PF Lecture 2 5 / 34
Community-Based

Allow local community to identify poor households

Idea: local community has much more information than central
government
This is the premise behind informal insurance, micro…nance, etc.
Problem:
If you are using this information to target bene…ciaries, this information
may not get revealed. Instead, elites may capture the project.
Potential tradeo¤: better local information vs. more elite capture

Olken () PF Lecture 2 6 / 34
Comparing PMT and Community Approaches
Alatas, Banerjee, Hanna, Olken, and Tobias (2009): “How to Target The Poor: Evidence
from a Field Experiment in Indonesia”

Randomized experiment compares three targeting methods:

Proxy-means test
Community ranking
Hybrid: community ranking, followed by proxy-means test on bottom
50% (to prevent elite capture)
Villages randomized to one of above treatments, used to give out real
one-time $3 transfer
Sub-treatments to tease out why community and PMT may di¤er
Elite capture: let elites run meetings or invite full community
E¤ort: randomize order of ranking and see if going …rst matters, start
with identifying 10 poorest …rst
Preferences: vary time of meeting to encourage more women in some
meetings
How do you randomize these?
Baseline survey measures true consumption, endline for satisfaction
Olken () PF Lecture 2 7 / 34
Community treatment

Olken () PF Lecture 2 8 / 34
Community treatment

Olken () PF Lecture 2 9 / 34
Experimental
VOL. 102 NO. 4 design Alatas et al.: Targeting the Poor 1215

Table 1—Randomization Design

Community/hybrid subtreatments Main treatments

Community Hybrid PMT
Elite 10 poorest first Day 24 23
Night 26 32
No 10 poorest first Day 29 20
Night 29 34

Whole community 10 poorest first Day 29 28

Night 29 23
No 10 poorest first Day 28 33
Night 20 24

Total 214 217 209

Notes: This table shows the results of the randomization. Each cell reports the number of sub-
villages randomized to each combination of treatments. Note that the randomization of subvil-
How do you analyze this design?
lages into main treatments was stratified to be balanced in each of 51 strata. The randomization
of community and hybrid subvillages into each subtreatment (elite or full community, 10 poor-
est prompting or no 10 poorest prompting, and day or night) was conducted independently for
each subtreatment, and each randomization was stratified by main treatment and geographic
stratum.

survey was completed (December 2008 and January 2009). Fund distribution, the
collection of the complaint form boxes, and interviews with the subvillage heads
occurred during February 2009. Finally, the survey company conducted the endline
Olken () PF Lecture 2 10 / 34
Metrics

First evalute targeting based on headcount:

MISTARGET = 0 if poor and didn’t receive transfer or rich and did
receive it, 0 otherwise
Evaluate targeting results based on four metrics:
Consumption (ug )
How households ranked each other on baseline survey (uc )
How village head ranked households at baseline (ue )
Self-assessment (us )
Also evaluate impact on satisfaction and legitimacy (many di¤erent
measures)

Olken () PF Lecture 2 11 / 34
Speci…cation

For mistargeting:

MISTARGETvhk = α + β1 COMMUNITYvhk + β2 HYBRIDvhk + γk + ε

Rank-correlations:
Convert each metric to a rank-ordering within village
Each targeting treatment de…nes a rank-ordering within village
So for each village v , compute RANKCORRvkw as the correlation
between the targeting outcome in village v and welfare metric w
Then regress

RANKCORRvkw = α + β1 COMMUNITYvk + β2 HYBRIDvk + γk + ε

Olken () PF Lecture 2 12 / 34
Results
VOL.on mistargeting
102 NO. 4 (headcount)
Alatas et al.: Targeting the Poor 1219

Table 3—Results of Different Targeting Methods on Error Rate Based on Consumption

By income status By detailed income status Per capita

Full Inclusion Exclusion Middle Near Very consumption
Sample: population error error Rich income poor poor of beneficiaries
(1) (2) (3) (4) (5) (6) (7) (8)
Community treatment 0.031* 0.046** 0.022 0.028 0.067** 0.49 −0.013 9.933
(0.017) (0.018) (0.028) (0.021) (0.027) (0.038) (0.039) (18.742)
Hybrid treatment 0.029* 0.037** 0.009 0.020 0.052** 0.031 −0.008 −1.155
(0.016) (0.017) (0.027) (0.020) (0.025) (0.037) (0.037) (19.302)

Observations 5,753 3,725 2,028 1,843 1,882 1,074 954 1,719

Mean in PMT treatment 0.30 0.18 0.52 0.13 0.23 0.55 0.48 366

Notes: All regressions include stratum fixed effects. Robust standard errors in parentheses, clustered at the village
level. All coefficients are interpretable relative to the PMT treatment, which is the omitted category. The mean of the
dependent variable in the PMT treatment is shown in the bottom row. All specifications include stratum fixed effects.
*** Significant at the 1 percent level.
 ** Significant at the 5 percent level.
  * Significant at the 10 percent level.

erty lines based on the PPP$2 per-day consumption threshold, and then classify a
­household as incorrectly targeted if its per capita consumption level is below the
poverty line and it was not chosen as a beneficiary, or if it was above the poverty
line and it was identified as a recipient (Errorivk). We then examine which method
minimized the error rate by estimating the following equation using OLS:

(1)
OlkenERROR
() ivk  =  α + β1 COMMUNITY ivk + β
PF Lecture 2 2 HYBRIDivk + γk  + εivk , 13 / 34
Results on alternative welfare metrics

Communities target worse

1232 basedECONOMIC
THE AMERICAN on consumption,
REVIEW but targetjune
better
2012
based on local welfare metrics
Table 9—Assessing Targeting Treatments Using Alternative Welfare Metrics

Consumption Community Subvillage head Self-assessment

(​rg​   ​) survey ranks (​rc​  ​ ) survey ranks(​re​  ​ ) (​rs​   ​)
(1) (2) (3) (4)
Community −0.065** 0.246*** 0.248*** 0.102***
  treatment (0.033) (0.029) (0.038) (0.033)
Hybrid −0.067** 0.143*** 0.128*** 0.075**
  treatment (0.033) (0.029) (0.038) (0.033)

Observations 640 640 640 637

Mean in PMT 0.451 0.506 0.456 0.343
 treatment

Notes: The dependent variable is the rank correlation between the treatment outcome (i.e.,
the rank ordering of households generated by the PMT, community, or hybrid treatment) and
the welfare metric shown in the column, where each observation is a village. Robust standard
errors are shown in parentheses.
*** Significant at the 1 percent level.
 ** Significant at the 5 percent level.
  * Significant at the 10 percent level.

Olken () PF Lecture 2 14 / 34
Results on satisfaction and legitimacy
1226 THE AMERICAN ECONOMIC REVIEW june 2012
All metrics of satisfaction are higher with community treatment
Table 6—Satisfaction

Panel A. Household endline survey

Is the method Are you
applied to satisfied with the Are there
determine targeting any poor
the targeted activities in HH that Number of Number of
households this subvillage should be added HH that HH that
appropriate? in general? to the list? should should be p-value
(1 = worst, (1 = worst, (0 = no, be added subtracted from
4 = best) 4 = best) 1 = yes) to list from list joint test
(1) (2) (3) (4) (5) (6)

Community 0.161*** 0.245*** −0.189*** −0.578*** −0.554*** < 0.001

 treatment (0.056) (0.049) (0.040) (0.158) (0.112)
Hybrid treatment 0.018 0.063 0.020 0.078 −0.171 0.762
(0.055) (0.049) (0.042) (0.188) (0.129)

Observations 1,089 1,214 1,435 1,435 1,435

Mean in PMT 3.243 3.042 0.568 1.458 0.968
 treatment

Panel B. Subvillage head endline survey

Is the method In your opinion,
applied to are villagers Are there any Are there any
determine satisfied with the poor HH nonpoor HH
the targeted targeting activities that should that should
households in this subvillage be added be subtracted
appropriate? in general? to the list? from the list?
(0 = no, (1 = worst, (0 = no, (0 = no,
1 = yes) 4 = best) 1 = yes) 1 = yes)
Community 0.378*** 0.943*** −0.169*** −0.010 < 0.001
  Olken
treatment() (0.038) (0.072) PF (0.045)
Lecture 2 (0.020) 15 / 34
Summary

Interpretation: community has di¤erent concept of welfare, and

community targeting allows them to achieve it. Outcome matches
local welfare function, hence higher satisfaction.
Other results:
Elite capture: no elite capture
Elite connected households no more likely to get transfer
In fact, if anything reverse discrimination in community treatment
But might be di¤erent if more money were at stake
Information:
Communities have some information about that PMT does not
Conclusions:
Suggests that tradeo¤ for community targeting is more about what
welfare function you want to maximize
If your goal is to minimize poverty headcount, want to use PMT
If your goal is to maximize utility (ie.., W = W (u1 , u2 , ..., un )), then
community approach may be better
Olken () PF Lecture 2 16 / 34
Self-Targeting

Nichols and Zeckhauser (1982): “Ordeals” can be used to target the

poor
Suppose you need to wait in long line to get unemployment bene…ts
Unemployed have low opportunity cost of time, so they are more likely
to wait in line
Waiting in line therefore serves as a screening device
Other examples?

Olken () PF Lecture 2 17 / 34
Self-Targeting Theory

Self-targeting context: you need to apply (wait in line, etc) in order to

get screened
This creates an additional complication: you don’t know for sure if
you will get the bene…ts
Imagine your income is y = y o + y u , where y o is the part the
government can observe through PMT and y u is the residual
De…ne µ (y o ) to be the probability you pass the PMT as a function of
observable income and λ(y ) to be the probability you pass the PMT
as a function of your total income
What do these look like in practice?

Olken () PF Lecture 2 18 / 34
 PMTs in practice
Figure 1. Probability of Obtaining Benefits vs. Log Per Capita Consumption and PMT score
.4

1 .8
Probability to Receive Benefits

Probability to Receive Benefits

.3

.6
.2

.4
.1

.2
0

0
11 12 13 14 15 16 11 12 13 14 15
Log Consumption Predicted Log Consumption (PMT score)

(a) Probability of Obtaining Benefits vs. (b) Probability of Obtaining Benefits vs.
Log Per Capita Consumption PMT score
Panel (A) shows the predicted probability of receiving the benefit, conditional on applying, from a probit model of receiving the benefit as a fu
capita consumption. Panel (B) repeats the same exercise replacing log per capita consumption by the predicted values from the PMT using ba
sset data. The predicted values from Panel (B) are the µ(yio ) that we use in the model. We include urban/rural interacted with district fixed
bit equations in both panels, since the PMT cutoff for inclusion varies slightly for each urban/rural times district cell.

Olken () PF Lecture 2 19 / 34
Self-targeting theory

When will someone apply? Suppose that person only knows y (they
don’t really understand what government can and can’t observe).
De…ne c as cost of applying and b as bene…t
They apply if
c + λ(y )(b ) > 0
Suppose that only cost is time. Then cost is proportional to your
wage So then c = ty .
So apply if
ty + λ(y )(b ) > 0
Two key drivers of targeting:
As I get richer, costs go up relative to bene…ts
As I get richer, expected chance I pass the test falls

Olken () PF Lecture 2 20 / 34
Simple example with constant lambda
Figure 1. 1Illustration of utility gain with no errors

applied
gain

1
0

close far close

far

y* y** consumption y* y** income

(a) Gain vs. consumption for close and (b) Targeting improves as length of ordeal
far subtreatments increases

Figure 2. Illustration of utility gain with log-logistic errors

2.5
1
.8

ication rates
2
ion rate
.6

Olken () PF Lecture 2 21 / 34
Problems

Why might this not work in practice?

Olken () PF Lecture 2 22 / 34
 l' l'' intensity of ordeal
Di¤erential utilityNotes: Increasing ordeal within l’ to l”, marginal cost for rich is lower than marginal cost for the poor.

Figure 4. Illustration of utility gain with concave utility

applies
gain
close far

apply

don’t apply

close
far

consumption consumption

(a) Gain vs. consumption for close and (b) Targeting can worsen as length of or-
far subtreatments deal increases

Olken () PF Lecture 2 23 / 34
 Travel costs

Figure 3. Non-Linearities in Travel Costs

cost
rich, walking

rich, bus

poor, walking

poor, bus

l' l'' intensity of ordeal

tes: Increasing ordeal within l’ to l”, marginal cost for rich is lower than marginal cost for the p

Olken () PF Lecture 2 24 / 34
What happens in practice?

Randomized experiment in Indonesia

400 villages newly eligible for Indonesian conditoinal cash transfer
program. Targeted to bottom 10% of HH based on PMT
Randomized into PMT (with some pre-screening done by villages) vs.
self-targeting, where you had to go to central meeting place to apply
for program
Measure true poverty based on baseline survey conducted before we
start

Olken () PF Lecture 2 25 / 34
Explaining the program

Olken () PF Lecture 2 26 / 34
Application process

Olken () PF Lecture 2 27 / 34
Who shows up
Figure 5. Showup Rates Versus Log Per Capita Consumption

.8
.6
Showed up
.4
.2
0

11 12 13 14 15
Log per capita consumption

Aside: this is a Fan local regression. What is that?

Notes: Figure provides a non-parametric fan regression of the probability of applying for PKH against baseline log per capita
consumption in the 200 self-targeting villages. Bootstrapped standard error bounds, clustered at the village level, are shown
Olken () PF in dashes.2
Lecture 28 / 34
On observables...
Figure 6. Showup Rates Versus Observable and Unobservable Components of Log
Per Capita Consumption

.8
.6
showup
.4
.2
0

12 13 14 15
PMTSCORE

0
(a) Showup as a function of observable consumption (Xi β)
.8

Olken () PF Lecture 2 29 / 34
 PMTSCORE
And unobservables...
(a) Showup as a function of observable consumption (X β)
0
i

.8
.6
showup
.4
.2
0

−2 −1 0 1 2
Epsilon

(b) Showup as a function of 45

unobservable consumption (εi )
Notes: Figures provide non-parametric fan regressions of the probability of applying for PKH against components of baseline
log per capita consumption in the 200 self-targeting villages. Bootstrapped standard error bounds, clustered at the village
level, are shown in dashes.
Olken () PF Lecture 2 30 / 34
Comparison to actual (pre-selected) PMT...
CDF

Figure 7. Experimental Comparison of Self Targeting and Automatic Enrollment

Treatments

1
.8
.6
CDF
.4
.2
0

11 12 13 14 15
Log per capita Consumption

Automatic Enrollment Self−Targeting

(a) CDF of log per capita consumption of beneficiaries

.4

CDF plots Pr(y < x ). How do we read this?

.3

Olken () PF Lecture 2 31 / 34
fit
 11 12 13 14 15
Log per capita Consumption
Comparison to actual (pre-selected) PMT... Automatic Enrollment Self−Targeting

Probability of getting bene…ts

(a) CDF of log per capita consumption of beneficiaries

.4
.3
Get benefit
.2
.1
0

11 12 13 14 15
Logconsumption

Automatic Enrollment Self−Targeting

46 of log per capita consumption

(b) Receiving benefit as a function
Notes: Panel A shows a CDF of log per capita consumption of beneficiaries. Kolmogorov-Smirnov test of equality yields a
p-value of 0.10. Panel B present a non-parametric fan regression of benefit receipt on log consumption per capita.
Key result: the rich don’
Bootstrapped standard t bother
errors, clustered atto apply
the village level, are shown in dashes.

Olken () PF Lecture 2 32 / 34
Concluding thoughts

Common theme for taxation and redistribution: lack of information

True everywhere, but particularly true in developing countries
As a result, tax and redistribution policies look very di¤erent in
developing countries

Olken () PF Lecture 2 33 / 34
References

Alatas, Banerjee, Hanna, Olken, Purnamasari, and Wai-Poi (2014).

"Self Targeting: Evidence from a Field Experiment in Indonesia"
Alatas, Banerjee, Hanna, Olken, and Tobias (2012). "Targeting the
Poor: Evidence from a Field Experiment in Indonesia"
Alatas, Banerjee, Hanna, Olken, Purnamasari, and Wai-Poi (2013).
"Ordeal Mechanisms In Targeting: Theory and Evidence from a Field
Experiment in Indonesia"

Olken () PF Lecture 2 34 / 34