Performance Pay and
Malnutrition: Evidence from an
Experiment targeting Child
Malnutrition in West Bengal
RAND Health Seminar, 13th April 2015
Prakarsh Singh
Amherst College/University of Pennsylvania
Sandip Mitra
Indian Statistical Institute
Motivation-I
Child undernutrition in India
• An estimated 1.27 million children die in India
every year due to malnutrition
• Malnutrition levels are higher than in SubSaharan Africa and have remained relatively
stagnant despite rapid economic growth
• Undernutrition related strongly to higher child
mortality and lower labor productivity
Motivation-II
Public economics and incentives
• Performance pay appears to have a positive effect on productivity
of workers in the private sector (Lazear, 2001)
• Little evidence on impact of performance pay in the public sector in
health in developing countries
• Rare to find an exogenous change in compensation schemes and a
contemporaneous control group
• Mullen et al. (2010) find no adverse impact on quality in a P4P
program in California but higher bonuses lead to better outcomes
• Review by Hasnain, Manning and Pierskalla (2012) for developing
countries: The two studies they cite in this sub-field are Basinga et
al. (2011) in Rwanda and Singh (2015) in India
Motivation – III
Health, incentives and development
• In developing countries, the role of
performance pay to health providers is still in
its infancy and there is scant evidence on how
to optimally structure contracts on the supplyside (Miller and Babiarz, 2013).
• Performance incentives based on outcomes can help local
providers use their knowledge productively even though
the outcome is only partly under the control of the
provider.
• Health outcomes can often be costly to measure as
compared to inputs and providers may lack
knowledge and skills to work towards obtaining the
performance bonus.
• In our context, it is difficult to measure inputs objectively as
the demand-side could collude with the providers when
reporting their effort.
• Performance pay can be based on piece rates
where a payout is linked to only individual
productivity or output
• Marginal effects of piece rates are understudied
• Aims of the project
– Test for gradient of performance with respect to
absolute performance pay in the public sector
– Test for the difference between a basic level of
absolute and relative performance pay
Incentive arms and control group
1. Basic Absolute Treatment
2. High Absolute Treatment
3. Basic Relative Treatment
4. Control group
Background
• There are 1.31 million child day care centers in India
under the Integrated Child Development Services
(ICDS), a government-sponsored program.
• Each center is staffed by a childcare worker who is
responsible for supervising children between ages 36 years during day-time, providing them
mid-day meals, teaching mothers about child
nutrition, imparting non-formal pre-school education
and facilitating health check-ups by doctors.
• Workers can affect health of the child through
two channels:
– Providing mid-day meals to children
– Advising mothers on a nutritious diet
• Yet, leakage of food to non-beneficiaries and
almost no guidance to parents (Gragnolati et
al. 2005, Hungama Report, 2011)
• "Although communication for behavior
change through the childcare worker is a
crucial weapon in the fight against
malnutrition, it appears that the information
the worker is conveying to the mothers is not
being communicated effectively enough to
positively affect mothers' behavior."
• Quality of service delivery has often come under
scrutiny with experts describing expenditure on
the program as if going into a bottomless pit
(Chaudhury et al. (2006) and Das and Hammer
(2005)).
• 96 percent of the villages are served by these
centers, although only 50 percent provide food
on the day of survey and 19 percent of the
mothers report that the center worker provides
nutrition counselling (HUNGaMA report, 2011).
Setting
• In our context, workers employed by the West
Bengal government in Maheshtala
Municipality
• In Maheshtala, total population is close to
400,000 and 11% of the population is under 6
years of age (Census, 2001)
Unique features of the project
•
•
•
•
•
Based entirely within one geographic block
Same institutional setting
Similar tasks, quality of infrastructure and wages
No endogenous selection in to the treatment
Exogenous treatments assigned at the same time
with a pure control group
• Treatments randomly assigned at the cluster level
to minimize any spillover effects between centers
[34 clusters, 209 centers]
Main results
• High absolute treatment reduces severe
malnutrition by 6.3% points and significantly
increases average weight
• Low absolute and relative treatments have no
effects on weight
• High absolute treatment effect driven by boys
In the Public Sector
• Relative may dominate absolute performance
pay because:
– Shocks to output may be salient
– Workers may be more risk averse (Buurman, Dur
and Bossche, 2009)
Theory
• Under risk neutrality, there is no difference
between absolute and relative incentives
(Lazear and Rosen, 1981)
• Under risk aversion and common shocks,
relative incentives can dominate absolute
incentives (Green and Stokey, 1983)
• The reason is that relative pay filters the
common shock making agents face lower risk
as compared to absolute pay
Framework
• Agents choose effort levels 0, 1 or 2
• Two possible individual outputs: 0 and 1
• One risk neutral principal, n risk averse agents
Production function
• If agent chooses 0 effort, output is 0.
• If agent chooses effort = 1, he gets output = 1
with probability 0.5 and output 0 with
probability 0.5.
• If agent chooses effort = 2, he gets output = 1.
• Common shock with probability γ, in which
case output = 0 regardless of the effort level
Public sector assumption
• We assume that workers get paid a minimum
utility even when they input effort = 0
• This is relevant in our context as these are
public sector workers who are on fixed wages
Results and Intuition
• The cost to principal of implementing high
effort levels = 2 for all agents under a relative
treatment does not depend on γ (immune to
common shock)
• The cost to principal of implementing high
effort levels = 2 for all agents under a absolute
treatment increases with γ (no longer immune
to common shock)
Cost
𝑈 = ln(𝑥)
𝑈 = 𝑥 1/2
𝑈 = 𝑥 1/5
γ
Results and Intuition
• The cost of implementing relative and an
absolute scheme is the same when there is no
common shock
• However, when there is a common shock,
relative treatment always dominates the
absolute treatment with risk averse agents
• This is a surprising result that comes from
combining the public sector assumption with
existing theory on common shocks
How is malnutrition measured?
• Moderate (more than 2 SD away from the WHO sexspecific weight-for-age mean)
• Severe (more than 3 SD away from the WHO sexspecific weight-for-age mean)
• All Workers given goal cards listing target weights for
each child in their AWC.
• Goal was the threshold for moderate if child is severely
malnourished and normal if child is moderately
malnourished
• If child was normal, a maximum threshold was
provided below which the child would become
malnourished
Absolute Incentives
• N for each worker = # children who jump at
least one grade - # children who drop at least
one grade in an Anganwadi
• What if more children get worse? [Solution:
ignore if N<0]
• Payout in Basic Absolute = Rs. 100 * N
• Payout in High Absolute = Rs. 200 * N
• AWW Salary is Rs. 4350 ($79.10) in Kolkata
and increases to Rs. 4413 ($80.24) if she has
10 years of service.
• The average payout in Chandigarh for the
basic absolute incentive with recipe book had
been Rs. 291 per worker ($5.30) but worker
salary was Rs. 2000 per month ($36.36) in
2010.
Relative Incentive
• N for each worker = # children who jump at
least one grade - # children who drop at least
one grade
• Let T = sum of all positive N in that treatment
group
• Payout = 55 workers * 291 * (N/T)
= 16000 (N/T)
• Manual randomization of clusters through
slips by a lottery into 4 boxes.
• This process was to be carried out till each box
had at least 50 centers allocated to it.
• When the total number of centers would cross
50, the assistant was asked to shut the box.
Groups
• Basic Absolute Incentive + Recipe Book
– [50 AWCs/8 clusters]
• High Absolute Incentive + Recipe Book
– [53 AWCs/8 clusters]
• Basic Relative Incentive + Recipe Book
– [55 AWCs/8 clusters]
• Control group (no incentive or recipe book)
– [51 AWCs/10 clusters]
Timeline
• Baseline: March – May, 2012
• Endline: August – September, 2012
• At Baseline and Endline – a team of
enumerators measured weights of children,
interviewed their mothers and Anganwadi
workers
Summary statistics
Variables
Mother's age
Mother can read
Monthly income
Number of rooms
Mobile phone
Water tap at home
Weight of child
Age of child
Fraction female
Total Siblings
Toilet in AWC
Drinking water in AWC
BA
28.09
[5.11]
0.73
[0.44]
4556.27
[2608.34]
1.43
[1.08]
0.84
[0.37]
0.35
[0.48]
13.65
[2.05]
4.23
[0.83]
0.51
[0.50]
1.29
[1.35]
0.39
[0.48]
0.68
[0.47]
HA
27.64
[5.43]
0.67
[0.47]
4059.73
[2567.92]
1.39
[0.88]
0.66
[0.47]
0.14
[0.35]
13.45
[2.09]
4.21
[0.84]
0.52
[0.50]
1.15
[1.29]
0.30
[0.46]
0.67
[0.47]
BR
27.38
[5.23]
0.75
[0.43]
4400.02
[2823.25]
1.43
[0.90]
0.83
[0.37]
0.18
[0.39]
13.62
[2.28]
4.23
[0.81]
0.50
[0.50]
1.21
[1.26]
0.49
[0.50]
0.80
[0.40]
C
27.47
[5.67]
0.73
[0.44]
3867.94
[2513.68]
1.40
[0.87]
0.66
[0.47]
0.14
[0.35]
13.49
[2.10]
4.18
[0.85]
0.50
[0.50]
1.43
[1.30]
0.39
[0.49]
0.74
[0.44]
Normalized Differences
BA-C
HA-C
BR-C
0.08
0.02
-0.01
0.00
-0.09
0.03
0.19
0.05
0.19
0.02
-0.01
0.02
0.30
0.00
0.28
0.35
0.00
0.08
0.05
-0.01
0.04
0.04
0.03
0.04
0.01
0.03
0.00
-0.07
-0.15
-0.12
0.00
-0.13
0.14
-0.09
-0.11
0.10
• Mother’s age, monthly income, occupation,
child’s age, child’s weight: similar to
Chandigarh
• Mother’s literacy is 30% points higher in
Maheshtala
• Ownership of mobile phones (high) and water
filters (low): similar to Chandigarh
Attrition
BA
HA
BR
C
Baseline children
1333
1555
1369
1264
Endline children
971
1127
1031
933
Attrition (%)
0.27
0.28
0.25
0.26
Specification
weightict = αpostt + Σβitreatmentc + Σγipostt*treatmentc + θict + ε ict
ict = {individual, cluster, time}
γi = difference-in-difference estimates of treatments 1, 2 or 3
θict = individual-level and anganwadi-level controls
Table 2: Main Results without Control Variables
(1)
(2)
Weight
Malnourished
Post
(3)
Grade
(4)
Severe Malnourished
0.283***
(0.0611)
-0.00291
(0.0206)
0.00517
(0.0249)
0.00807
(0.00835)
Basic Absolute (BA)
0.165
(0.234)
-0.0123
(0.0355)
-0.0413
(0.0521)
-0.0289
(0.0186)
High Absolute (HA)
-0.0411
(0.168)
0.0195
(0.0247)
0.0290
(0.0409)
0.00946
(0.0191)
Basic Relative (BR)
0.126
(0.173)
-0.0231
(0.0318)
-0.0404
(0.0491)
-0.0172
(0.0202)
Post*BA
-0.154
(0.153)
-0.00809
(0.0258)
-0.00205
(0.0346)
0.00604
(0.0149)
Post*HA
0.191**
(0.0899)
-0.0465
(0.0301)
-0.0953**
(0.0444)
-0.0488**
(0.0189)
Post*BR
0.107
(0.0842)
0.00682
(0.0239)
-0.00385
(0.0300)
-0.0107
(0.0136)
Constant
13.49***
(0.127)
0.331***
(0.0172)
0.416***
(0.0301)
0.0848***
(0.0142)
9376
9376
9376
9376
Observations
Table 3: Main Results with Control Variables
(1)
Weight
(2)
Malnourished
(3)
Grade
(4)
Severe Malnourished
Post
0.271**
(0.114)
0.0234
(0.0363)
0.0317
(0.0416)
0.00830
(0.0121)
Basic Absolute (BA)
0.0954
(0.155)
-0.0117
(0.0304)
-0.0401
(0.0453)
-0.0284
(0.0178)
High Absolute (HA)
0.176
(0.126)
-0.0174
(0.0227)
-0.0177
(0.0373)
-0.000215
(0.0182)
Basic Relative (BR)
0.0193
(0.145)
0.00793
(0.0372)
-0.0160
(0.0512)
-0.0239
(0.0187)
Post*BA
0.0398
(0.176)
0.00229
(0.0430)
0.00103
(0.0551)
-0.00126
(0.0194)
Post*HA
0.253*
(0.130)
-0.0557
(0.0405)
-0.119**
(0.0502)
-0.0630***
(0.0175)
Post*BR
0.225
(0.141)
-0.0439
(0.0437)
-0.0530
(0.0518)
-0.00906
(0.0156)
Constant
8.688***
(0.341)
0.0624
(0.0695)
-0.0232
(0.110)
-0.0856
(0.0694)
x
x
x
x
5342
5342
5342
5342
Control Variables
Observations
Placebo Check
VARIABLES
Weight
Basic Absolute (BA)
0.351
(0.269)
High Absolute (HA)
0.016
(0.192)
Basic Relative (BR)
0.223
(0.206)
Placebo Post
0.291**
(0.111)
Placebo Post*BA
-0.187
(0.163)
Placebo Post*HA
-0.061
(0.163)
Placebo Post*BR
-0.094
(0.183)
Constant
13.202***
(0.110)
Observations
10021
Robustness checks
• Results robust to controlling for reversion-tothe-mean, pre-trends, Moulton correction,
Lee bounds and Propensity Score Matching
• Do not find selective targeting (perhaps due to
disincentive for lowering weights)
• Iron supplementation and deworming drugs have been
amongst the most effective in reducing malnutrition.
• Bobonis et al. (2006) find that an increase in weight of 0.5
kg in five months for 2-6 year old slum children in Delhi due
to iron supplementation and deworming.
• In the HA treatment, the average weight gain per month is
84 gms as opposed to 100 gms per month in the iron and
deworming treatment.
• The cost for our treatment is much lower than the World
Bank (2006) figures for a nutrition education program of
$2.50 per person per year at $1.80 per child per year.
• The estimated net benefit-cost ratio for the HA treatment
turns out to be 12.8.
Diminishing returns of repeated
treatments
• Mechanical diminishing returns due to fewer
remaining malnourished children in the
Anganwadi.
• Participation rates (and thereby wages) are
unlikely to increase as a child becomes normal
weight.
• Biological diminishing returns as the body adapts
to increased calorie intake.
• Workers may need to be motivated with higher
incentives if their intrinsic motivation keeps
getting crowded out.
Table 8: Main Results for Girls after including Control Variables
(1)
(2)
Weight
Malnourished
(3)
Grade
(4)
Severe Malnourished
Post
0.359**
(0.137)
0.00229
(0.0479)
0.00207
(0.0593)
-0.000212
(0.0159)
Basic Absolute (BA)
0.167
(0.198)
-0.0394
(0.0424)
-0.0983
(0.0601)
-0.0589**
(0.0223)
High Absolute (HA)
0.255*
(0.147)
-0.0278
(0.0334)
-0.0699
(0.0513)
-0.0421*
(0.0221)
Basic Relative (BR)
0.157
(0.176)
-0.0111
(0.0476)
-0.0693
(0.0647)
-0.0582**
(0.0228)
Post*BA
-0.0711
(0.171)
0.0245
(0.0542)
0.0373
(0.0705)
0.0128
(0.0272)
Post*HA
0.0984
(0.173)
-0.0204
(0.0531)
-0.0566
(0.0661)
-0.0361
(0.0215)
Post*BR
0.239
(0.175)
-0.0358
(0.0589)
-0.0424
(0.0731)
-0.00659
(0.0220)
Constant
7.399***
(0.475)
0.191**
(0.0922)
0.191
(0.136)
-0.000154
(0.0818)
x
x
x
x
2748
2748
2748
2748
Control Variables
Observations
Table 9: Main Resultsfor Boys after including Control Variables
(1)
(2)
Weight
Malnourished
(3)
Grade
(4)
Severe Malnourished
Post
0.173
(0.130)
0.0488
(0.0321)
0.0661
(0.0508)
0.0174
(0.0300)
Basic Absolute (BA)
0.0289
(0.201)
0.0131
(0.0371)
0.0111
(0.0527)
-0.00196
(0.0208)
High Absolute (HA)
0.0901
(0.204)
-0.00723
(0.0331)
0.0315
(0.0485)
0.0387
(0.0247)
Basic Relative (BR)
-0.150
(0.185)
0.0328
(0.0396)
0.0450
(0.0561)
0.0123
(0.0211)
Post*BA
0.150
(0.220)
-0.0220
(0.0439)
-0.0361
(0.0727)
-0.0141
(0.0396)
Post*HA
0.428***
(0.148)
-0.0966**
(0.0406)
-0.189***
(0.0627)
-0.0920**
(0.0350)
Post*BR
0.205
(0.156)
-0.0548
(0.0380)
-0.0645
(0.0557)
-0.00973
(0.0312)
Constant
9.707***
(0.513)
-0.0945
(0.116)
-0.302*
(0.158)
-0.207**
(0.0784)
x
x
x
x
2594
2594
2594
2594
Control Variables
Observations
Table 12: Gender Difference
(1)
Weight
(2)
Malnourished
(3)
Grade
(4)
Severe Malnourished
Post*BA
0.142
(0.216)
-0.0226
(0.0428)
-0.0339
(0.0712)
-0.0113
(0.0391)
Post*HA
0.432***
(0.145)
-0.0989**
(0.0397)
-0.191***
(0.0620)
-0.0924**
(0.0350)
Post*BR
0.219
(0.164)
-0.0548
(0.0371)
-0.0682
(0.0548)
-0.0133
(0.0307)
Gender*BA*Post
-0.197
(0.163)
0.0475
(0.0469)
0.0678
(0.0887)
0.0203
(0.0548)
Gender*HA*Post
-0.345*
(0.193)
0.0830*
(0.0483)
0.140*
(0.0780)
0.0567
(0.0435)
Gender*BR*Post
0.0188
(0.176)
0.0202
(0.0426)
0.0283
(0.0726)
0.00811
(0.0424)
8.784***
(0.349)
0.0442
(0.0699)
-0.0635
(0.109)
-0.108
(0.0708)
x
x
x
x
5342
5342
5342
5342
Constant
Control Variables
Observations
(1)
Pulses
(2)
Fish
(3)
Meat
(4)
Green
(5)
Sweet
Post
-0.0523
(0.0435)
0.0271
(0.0470)
-0.0308
(0.0499)
-0.0466
(0.0362)
-0.123**
(0.0516)
Basic Absolute (BA)
0.0617
(0.0742)
0.122
(0.0882)
0.0330
(0.0767)
0.0900
(0.0700)
-0.0252
(0.0608)
High Absolute (HA)
-0.117
(0.0931)
0.0373
(0.0872)
-0.0446
(0.0815)
-0.0309
(0.0906)
-0.0874
(0.0648)
Basic Relative (BR)
0.0752
(0.0735)
0.133
(0.105)
0.0183
(0.0701)
0.0291
(0.0730)
0.00818
(0.0793)
Post*BA
0.0797
(0.0964)
-0.00203
(0.100)
0.101
(0.0989)
0.0189
(0.0914)
0.0607
(0.100)
Post*HA
0.217**
(0.0803)
0.0169
(0.106)
0.127*
(0.0716)
0.000472
(0.0876)
-0.00706
(0.0706)
Post*BR
0.109
(0.0689)
-0.102
(0.0790)
0.0221
(0.0695)
0.0342
(0.0583)
-0.0183
(0.111)
Constant
0.235
(0.179)
0.177
(0.166)
-0.156
(0.164)
0.345**
(0.152)
0.130
(0.0972)
x
x
x
x
x
5421
5430
5446
5441
5328
Control Variables
Observations
Table 7: Number of Visits with Control Variables
(1)
(2)
Total Visits
Worker Visits
(3)
Home Visits
Post
0.548
(1.272)
0.128
(0.393)
0.412
(0.889)
Basic Absolute (BA)
-0.573
(0.648)
-0.113
(0.237)
-0.400
(0.481)
High Absolute (HA)
1.032
(0.820)
0.716**
(0.264)
0.347
(0.587)
Basic Relative (BR)
1.022
(0.821)
0.448
(0.346)
0.567
(0.537)
Post*BA
0.651
(1.695)
0.176
(0.757)
0.429
(1.035)
Post*HA
-0.906
(1.290)
-0.380
(0.406)
-0.541
(0.906)
Post*BR
-0.726
(2.080)
-0.0807
(0.905)
-0.631
(1.228)
x
x
x
5.666***
(0.978)
2.754***
(0.431)
2.947***
(0.655)
5365
5420
5397
Control Variables
Constant
Observations
Channels
• Food at home goes up for both boys and girls
• Why does it go up? Recipe books + Worker
tells mothers to increase food at home
[though she does not increase visits to home]
• Weight goes up only for boys
• Food must have gone up for boys in the AWC
where workers have high absolute incentives
Why?
• Boys have greater appetite or ask for more (so
are able to get more food from the incentivized
worker)
• Expectation that boys will show more weight gain
• Expectation that mothers will be less likely to cut
down on food for boys leading to
complementarity
• A gender preference for food provision that gets
accentuated with the incentives
Conclusion
• High Absolute incentive works to reduce
malnutrition by about 6.5% points along with
information to mothers and this effect is
driven by the severely malnourished moving
to moderately malnourished
• First order effects important but gender
inequality may be exacerbated
Related work
• Informational Spillovers (Singh, 2014)
• Training of workers and its interaction with the recipe
treatment in rural villages of Punjab (Singh, 2015)
• Ongoing Gates Foundation project testing different
performance incentive schemes and informational
spillovers in Chandigarh
– Staggered phase-in of different incentive schemes to
capture long-run effects
– Better measures of Intrahousehold allocation, food
distribution at centers and worker effort
– Randomize the intensity of informational treatment by
center (No/Low/Medium/High)
Change in weight
0
.2
Density
.4
.6
Kernel density estimate
-5
0
changeinweight
Control
Basic Absolute
High Absolute
Basic Relative
kernel = epanechnikov, bandwidth = 0.1870
5
10