The Long Run Effects of Early Childhood Deworming On Literacy and Numeracy: Evidence From Uganda
The Long Run Effects of Early Childhood Deworming On Literacy and Numeracy: Evidence From Uganda
The Long Run Effects of Early Childhood Deworming On Literacy and Numeracy: Evidence From Uganda
The
long
run
effects
of
early
childhood
deworming
on
literacy
and
numeracy:
Evidence
from
Uganda
Kevin
Croke
Department of Global Health and Population, Harvard School of Public Health
Acknowledgements:
Thanks
to
Owen
Ozier,
for
suggesting
the
Uganda
deworming
project
as
a
candidate
for
long
run
follow
up
and
for
providing
the
list
of
treatment
and
control
parishes;
to
Harold
Alderman
and
Günther
Fink
for
valuable
feedback;
and
to
seminar
p articipants
at
the
Harvard
School
of
Public
Health.
1
Abstract:
This
paper
analyzes
the
long
run
impact
of
a
cluster-‐randomized
trial
in
eastern
Uganda
that
provided
mass
deworming
treatment
to
a
sample
of
preschool
aged
children
from
2000
to
2003.
An
early
impact
evaluation
of
this
intervention
found
that
the
treatment
group,
comprised
of
children
aged
1-‐7,
showed
increased
weight
gain
compared
to
controls
(Alderman
et
al.
2007).
Since
there
is
now
a
large
literature
linking
early
life
health,
often
proxied
by
weight,
to
long
run
outcomes
(including
cognitive,
educational,
health,
and
labor
market
outcomes),
I
use
data
collected
in
these
communities
7-‐8
years
after
the
end
of
the
deworming
trial
to
see
whether
children
in
treatment
communities
have
higher
scores
than
children
in
control
communities
on
simple
numeracy
and
literacy
tests.
I
find
that
children
who
lived
in
treatment
communities
during
the
period
in
question
have
test
scores
0.2-‐0.4
standard
deviations
higher
than
those
in
control
parishes.
Effects
are
larger
for
math
than
for
English
literacy
scores.
The
effect
is
robust
to
a
wide
range
of
alternate
specifications
and
inclusion
of
socioeconomic
control
variables,
and
to
a
placebo
treatment
test.
Girls
and
children
from
the
poorest
income
quintiles
experience
relatively
larger
gains.
1 Introduction
While
mass
provision
of
deworming
treatment
to
school
aged
children
living
in
highly
endemic
regions
has
long
been
a
WHO-‐recommended
policy,
debate
about
the
effectiveness
of
this
intervention
has
recently
emerged,
tied
to
the
2012
publication
of
an
updated
Cochrane
Systematic
Review
(Taylor-‐Robinson
et
al
2012)
which
finds
essentially
no
evidence
of
benefit
to
mass
deworming
programs.
This
paper
adds
new
evidence
to
the
debate
about
the
relationship
between
early
childhood
deworming
and
educational
and
cognitive
outcomes.
I
analyze
the
long
run
impact
of
a
cluster-‐randomized
trial
in
eastern
Uganda
that
provided
mass
deworming
treatment
to
a
sample
of
preschool
aged
children
from
2000
to
2003.
An
early
impact
evaluation
of
this
project
found
that
the
treatment
group,
comprised
of
children
aged
1-‐
7,
showed
increased
weight
gain
compared
to
controls
(Alderman
et
al.
2007).
Since
there
is
now
a
large
literature
linking
early
life
health
to
long
run
outcomes
(including
cognitive,
educational,
health,
and
labor
market
outcomes),
I
use
data
collected
in
these
communities
7-‐8
years
after
the
end
of
the
deworming
trial
to
see
whether
children
in
treatment
communities
have
higher
scores
than
children
in
control
communities
on
simple
numeracy
and
literacy
tests.
I
find
that
children
who
lived
in
treatment
communities
during
the
period
in
question
have
test
scores
0.2-‐0.4
standard
deviations
higher
than
those
in
control
parishes.
Effects
are
larger
for
math
than
for
English
literacy
scores.
The
effect
is
robust
to
a
wide
range
of
alternate
specifications
and
inclusion
of
socioeconomic
control
variables,
and
to
a
placebo
treatment
test.
When
the
clustering
of
errors
is
adjusted
for
the
relatively
small
(22
total)
number
of
clusters,
the
treatment
effect
on
math
results
is
unchanged,
while
the
effect
on
literacy
and
combined
scores
are
significant
at
the
10%
level.
As
the
following
discussion
will
show,
I
cannot
precisely
identify
the
mechanism
through
which
this
gain
is
transmitted.
However,
I
can
rule
out
2
any
health
differences
that
operate
through
a
long
run
school
enrollment
or
contemporaneous
school
attendance
channels,
since
these
do
not
differ
between
treatment
and
control
communities.
As
many
as
two
billion
people
are
estimated
to
suffer
from
intestinal
worms
(also
known
as
soil-‐transmitted
helminths)
such
as
roundworm,
hookworm,
or
whipworm
(Bundy
et
al
2009).
The
largest
burden
of
disease
is
found
in
sub-‐Saharan
Africa
and
in
South
Asia.
Worm
infection,
although
rarely
fatal,
is
associated
with
a
wide
range
of
health
problems
including
stunted
growth,
reduced
food
absorption,
loss
of
appetite,
listlessness,
and
anemia.
Among
pre-‐school
age
children,
worm
infection
is
associated
with
slowed
growth,
and
among
school
age
children
it
has
been
linked
to
poor
attendance
at
school
and
reduced
performance
on
cognitive
tasks.
Soil-‐transmitted
helminths
are
largely
transmitted
through
fecal-‐oral
contact,
with
the
result
that
the
disease
burden
is
highest
among
populations
with
poor
sanitation
and
inadequate
access
to
health
care,
and
in
conditions
of
poverty.
Inexpensive
and
effective
treatment
is
available
for
worms,
in
the
form
of
a
single
dose
of
drugs,
such
as
albendazole
or
mebendazole.
These
drugs
are
safe,
have
minimal
side
effects,
and
are
well
tolerated
by
both
infected
and
non-‐infected
populations.
The
availability
of
inexpensive
treatment
that
has
no
significant
side
effects
for
uninfected
children
–
and
the
fact
that
screening
children
for
infection
is
significantly
more
expensive
than
treating
them
–
has
pointed
global
health
policymakers
towards
a
policy
of
mass
treatment
in
high
prevalence
populations,
such
as
school
age
children.
In
fact,
the
World
Health
Organization
advocates
for
annual
deworming
of
all
children
in
regions
where
STH
prevalence
is
over
20%,
and
twice
annually
where
prevalence
is
above
50%.
Several
recent
papers
have
found
large
benefits
to
interventions
following
this
policy
in
Western
Kenya.
Miguel
and
Kremer
(2004),
for
example,
find
large
impacts
of
school-‐based
mass
deworming
on
school
attendance
in
western
Kenya,
reducing
absenteeism
by
7.5
percentage
points.
Baird
et
al
(2011)
track
these
cohorts
over
time
through
the
Kenya
Life
Panel
Survey
(KLPS),
and
find
positive
educational
and
labor
market
outcomes
for
the
cohorts
that
were
dewormed
more
frequently
(the
original
study
was
a
phase-‐in
randomization
so
there
is
no
pure
control
in
the
long
run).
Ozier
(2012)
returns
to
the
original
study
schools
approximately
10
years
later
and
finds
large
spillover
effects
on
children
who
were
under
one
year
of
age
when
their
community
was
dewormed,
with
increased
scores
on
cognitive
tests
equivalent
to
a
full
year
of
schooling.
Moving
across
the
border
from
Western
Kenya
to
the
program
in
eastern
Uganda
that
this
paper
focuses
on,
the
short
run
impact
evaluation
mentioned
above
(Alderman
et
al
2007)
found
that
beneficiaries
of
the
deworming
program
(e.g
those
who
attended
at
least
two
Child
Health
Days
over
the
3
program
duration
in
the
treatment
group)
gained
5-‐10%
more
in
weight
than
Child
Health
Day
attendees
in
the
control
communities,
where
deworming
was
not
offered.1
The
evidence
from
these
studies,
and
from
earlier
biomedical
trials,
has
led
to
various
international
expert
groups
to
rank
mass
deworming
very
highly
on
a
number
of
policy
prioritization
and
cost
effectiveness
exercises.2
For
example,
the
2012
Copenhagen
Consensus
ranked
school-‐based
deworming
as
its
fourth
most
important
priority
for
development
funders.3
Similarly,
researchers
affiliated
with
MIT’s
Poverty
Action
Lab
found
that
it
was
the
second
most
cost
effective
way
to
increase
school
attendance,
producing
13.9
additional
years
of
schooling
for
every
$100
spent.4
These
recommendations,
however,
are
in
tension
with
the
findings
from
a
series
of
Cochrane
Collaboration
systematic
reviews
(Dickson
et
al.
2000;
Taylor-‐Robinson
et
al.
2007;
and
Taylor-‐
Robinson
et
al,
2012).
The
most
recently
updated
(2012)
Cochrane
review
states
clearly
that,
contrary
to
the
policy
recommendations
of
WHO
and
others,
they
find
no
convincing
evidence
of
benefits
of
mass
administration
of
deworming
drugs,
concluding
that
“it
is
probably
misleading
to
judge
contemporary
deworming
programs
based
on
evidence
of
consistent
benefit
on
nutrition,
hemoglobin,
school
attendance,
or
school
performance
as
there
is
simply
insufficient
reliable
information
to
know
whether
this
is
so
(2).”
One
of
the
authors
of
the
review,
in
a
subsequent
interview,
stated
that
he
would
“love
[for
deworming]
to
work.
But
to
claim
that
it
does
on
the
basis
of
the
evidence
available
is
simply
misleading”
(Hawkes
2013).
Another
skeptical
perspective
comes
from
a
comprehensive
review
by
the
charity
rating
organization
Givewell,
which
concludes
that
“overall,
evidence
for
the
impact
of
deworming
on
short-‐term
general
health
is
thin.”
With
respect
to
longer-‐term
developmental
effects,
they
conclude
that
“empirical
evidence…is
very
limited,”
although
they
note
that
it
is
based
on
“two
relatively
well-‐known
and
well-‐executed
studies”
(referring
to
Bleakley
2007
and
Baird
et
al
2011).5
1
The
2012
Cochrane
review
downgrades
Alderman
et
al.
(2007)
because
in
in
table
1
of
the
paper,
the
authors
did
not
cluster
standard
errors
for
their
comparison
of
treatment
and
control
means,
although
they
do
so
for
their
regression-‐based
estimates.
2
The
Disease
Control
Priorities
Project
chapter
on
deworming
(Hotez
et
al
2006)
for
example,
presented
a
stylized
case
in
which
disability
adjusted
life
years
associated
with
STH
could
be
prevented
for
just
$3.41,
which
would
make
deworming
by
far
one
of
the
most
cost
effective
interventions
available.
However,
researchers
at
Givewell
demonstrated
that
these
calculations
were
erroneous;
see
http://blog.givewell.org/2011/09/29/errors-‐in-‐dcp2-‐
cost-‐effectiveness-‐estimate-‐for-‐deworming/.
3
See
http://www.copenhagenconsensus.com/copenhagen-‐consensus-‐iii/outcome
4
See
http://www.povertyactionlab.org/policy-‐lessons/education/student-‐participation
5
See
http://www.givewell.org/international/technical/programs/deworming
4
A
long
back
and
forth
has
ensued
between
advocates
of
mass
deworming
and
the
authors
of
the
Cochrane
review.
In
response
to
the
2007
version
of
the
Cochrane
systematic
review,
Bundy
et
al
(2009)
criticized
the
Cochrane
findings
for
not
taking
sufficient
account
of
clustering
(by
overweighting
studies
that
had
individual-‐level
randomization
and
by
not
including
recent
cluster
randomized
trials
such
as
Miguel
and
Kremer
2004);6
for
insufficient
attention
to
data
quality
with
respect
to
school
attendance
and
cognitive
outcomes,
and
sample
attrition;
and
for
not
taking
into
account
observational
econometric
studies
with
strong
strategies
for
causal
identification,
such
as
Bleakley
(2007).
The
updated
Cochrane
review
(2012)
responded
to
these
criticisms
with
several
changes,
including
by
incorporating
the
Miguel
and
Kremer
study
(but
ranking
it
as
relatively
weak
evidence
based
on
the
Cochrane
evidence
ranking
criteria)
,7
and
by
adding
several
additional
trials
to
the
meta-‐analysis.8
In
response
to
the
negative
findings
about
mass
deworming
in
this
updated
Cochrane
review,
a
group
of
economists,
including
Miguel
and
Kremer
and
other
researchers
associated
with
the
Abdul
Latif
Jameel
Poverty
Action
Lab
(JPAL),
and
the
NGOs
Innovations
for
Poverty
Action
(IPA)
and
Evidence
Action,
responded
in
turn,
disagreeing
strongly
with
Taylor-‐Robinson
et
al.
over
the
interpretation
of
Miguel
and
Kremer
(2004)
and
over
the
non-‐inclusion
of
studies
that
use
experimental
or
quasi-‐experimental
variation
generated
by
the
original
Miguel
and
Kremer
study
(such
as
Baird
et
al
2011
and
Ozier
2012)
to
identify
impact.9
This
debate
has
led
some
to
criticize
the
NGO
Evidence
Action
(which
recently
incorporated
the
deworming-‐focused
NGO
Deworm
the
World)
for
scaling
up
school
based
deworming
in
the
face
of
this
contrary
evidence
provided
in
the
Cochrane
review.
For
example,
Waddingham
and
Leach
(2014)
wrote
on
the
International
Initiative
for
Impact
Evaluation
(3ie)
website
that
“On
balance,
the
evidence
does
not
favor
the
scaling
up
of
[deworming].”10
The
British
medical
Journal
published
an
article
in
2013
with
the
headline
“Deworming
Debunked.”
(Hawkes,
2013).
This
paper
does
not
attempt
to
adjudicate
this
debate,
but
rather
to
add
to
the
evidence
base.
One
way
to
increase
the
evidence
base
would
be
to
conduct
new
large
scale
randomized
trials.
In
the
meantime,
however,
policymakers
have
to
make
decisions
in
the
face
of
competing
6
Because
treatment
has
large
potential
spillovers
to
control
groups
by
reducing
community-‐level
worm
loads,
they
argue
that
individually
randomized
studies
likely
systematically
underestimate
the
benefits
of
deworming.
Therefore,
they
argue,
“the
primary
focus
of
a
review
should
be
studies
that
use
a
cluster
design.”
7
Cochrane
reviews
incorporate
six
potential
sources
of
bias:
random
sequence
generation
bias,
allocation
concealment/selection
bias,
blinding
(performance
bias
and
detection
bias);
incomplete
outcome/attrition
bias;
selective
reporting
bias,
and
“other”
bias.
8
This
study
was
apparently
excluded
from
the
2007
review
because
of
misunderstanding
about
the
extent
to
which
the
results
were
robust
when
the
sub-‐sample
of
communities
also
treated
for
schistosomiasis
was
excluded.
9
See
http://blogs.berkeley.edu/2012/07/20/cochranes-‐incomplete-‐and-‐misleading-‐summary-‐of-‐the-‐evidence-‐on-‐
deworming/
10
See
http://blogs.3ieimpact.org/how-‐much-‐evidence-‐is-‐enough-‐for-‐action/
5
evidence.
This
points
towards
the
value
of
using
already-‐collected
data,
when
possible,
to
answer
questions
about
the
impact
of
deworming
programs.
The
Uwezo
data
sets
(described
further
below)
are
particularly
helpful
in
this
regard.
They
are
large
scale,
high-‐resolution
education
surveys
conducted
in
areas
in
East
Africa
where
important
deworming
programs
have
been
implemented.
They
also
represent,
in
effect,
a
blinded
data
collection
exercise
with
respect
to
deworming,
since
(in
the
Uganda
case,
for
example)
the
Uwezo
team
had
no
idea
that
they
were
sampling
and
surveying
communities
that
had
been
part
of
a
deworming
project
7-‐10
years
prior.
The
deworming
debate
is
complex
and
there
is
a
wide
range
of
potential
outcomes
to
examine,
which
also
vary
in
relevance
according
to
the
age
of
treated
children.
This
paper
addresses
one
component
of
the
broader
question
about
the
comprehensive
effects
of
deworming:
it
examines
the
effect
of
deworming
pre-‐school
age
children
on
their
test
scores
later
in
life.
The
channel
through
which
deworming
at
young
ages
could
affect
young
adult
cognitive
function
is
thought
to
be
through
improved
nutrition
and
growth
during
the
critical
period
of
early
growth
and
brain
development.
As
mentioned
above,
the
Cochrane
Collaborative
meta-‐analysis
finds
no
convincing
evidence
that
cognitive
gains
or
school
performance
improvements
result
from
deworming.
However,
relatively
few
of
the
studies
examined
by
the
systematic
review
are
both
cluster
randomized
and
have
long
run
cognitive
outcomes.
This
suggests
that
the
evidence
is
not
yet
definitive
with
regard
to
this
outcome.
Furthermore,
one
recent
study
(Ozier
2011)
provides
strong
evidence
for
relatively
large
cognitive
improvements
among
a
cohort
of
Kenyan
children
whose
older
siblings
were
dewormed
as
part
of
the
project
described
in
Miguel
and
Kremer
(2004).
However,
Ozier
identifies
this
impact
as
operating
through
epidemiological
spillovers
rather
than
direct
treatment
of
preschool
age
children.
This
paper
seeks
to
add
to
the
evidence
base
by
examining
the
long
run
effects
of
a
program
that
treated
young
children
directly
with
deworming
medication,
rather
than
by
reducing
their
exposure
to
worms
through
an
epidemiological
spillover
mechanism.
The
deworming
project
that
is
the
focus
of
this
paper
took
place
in
48
parishes
in
5
districts
in
Eastern
Uganda
from
November
2000
to
June
2003.11
The
districts
were
chosen
because
they
were
identified
as
having
heavy
worm
loads,
with
at
least
60%
of
children
ages
5-‐10
infected,
primarily
with
hookworm
(Kabatereine
et
al
2001).
It
was
implemented
through
the
“Child
Health
Day”
(CHD)
delivery
platform.
Child
Health
Days
are
a
pre-‐defined
day
in
Uganda
and
other
developing
countries,
usually
every
6
months,
in
which
all
parents
in
a
given
catchment
area
are
requested
to
bring
all
pre-‐school
age
children
to
a
treatment
site
to
receive
a
set
of
11
Parishes
are
the
second
lowest
administrative
level
in
Uganda,
just
above
the
village
level
and
below
the
sub-‐
county
level.
6
basic
health
services
such
as
Vitamin
A
supplementation,
growth
monitoring,
and
often
any
standard
vaccines
that
they
have
not
yet
received.
Five
Child
Health
Days
were
held
over
the
course
of
the
project.
The
experimental
variation
was
introduced
in
the
following
way:
At
Child
Health
Days
in
the
control
group,
attendees
were
offered
the
standard
intervention
of
Vitamin
A
supplementation,
vaccines,
growth
monitoring,
and
complementary
feeding
demonstrations,
while
in
the
treatment
group
they
were
also
offered
deworming
treatment,
in
the
form
of
400
mg
of
albendazole.12
All
children
(except
those
who
were
ill
at
the
time
of
the
Child
Health
Day)
between
age
1
and
7
were
offered
albendazole
in
the
treatment
group,
while
none
were
offered
it
in
the
control
group.
The
intervention
was
delivered
by
community-‐based
organizations
(CBOs).
Randomization
was
at
the
parish
level
because
that
is
the
level
at
which
CHDs
are
typically
implemented
in
Uganda.13
Alderman
et
al.
(2007)
use
data
collected
at
the
Child
Health
Day
by
program
staff
to
measure
anthropomorphic
outcomes
such
as
height
and
weight,
and
use
baseline
and
end
line
surveys
to
measure
population
level
participation.14
In
regression
models,
they
find
that
weight
increases
as
a
function
of
the
number
of
CHDs
attended,
but
by
a
larger
increment
in
the
treatment
parishes
than
the
control.
For
children
with
long
gaps
between
treatment
(corresponding
roughly
to
annual
treatment)
weight
gain
was
5%,
compared
to
10%
when
treatment
frequency
was
roughly
once
per
year.
The
study
population
was
comprised
of
children
who
had
at
least
two
anthropomorphic
measurements,
which
means
that
the
sample
size
was
over
27,000
children
in
48
clusters.
4 Data
Uwezo
is
a
project
led
by
the
Tanzanian
NGO
Twaweza,
modeled
on
India’s
Annual
Status
of
Education
Report
(ASER).
Uwezo
does
large-‐scale
annual
surveys
to
test
basic
literacy
and
numeracy
in
Kenya,
Tanzania,
and
Uganda.
The
first
test
was
done
in
Kenya
in
2009
and
in
Tanzania
and
Uganda
in
2010;
for
each
country
two
years
of
data
are
in
the
public
domain,
and
available
at
http://www.uwezo.net.
Uwezo
aims
to
collect
data
that
is
representative
at
district
level,
which
means
that
in
each
year,
30
villages
per
district
are
sampled
in
each
country,
and
20
households
per
village
are
tested
on
basic
literacy
and
numeracy.
Over
the
course
of
its
2010
12
The
medicine
(Zentel
from
GlaxoSmithKline)
was
provided
in
chewable
form
13
According
to
the
Uganda
Service
Provision
Assessment
Survey
(2008),
the
catchment
area
of
a
parish-‐level
health
facility
is
up
to
5,000
people.
14
Since
only
children
who
attended
Child
Health
Days
(CHDs)
two
times
were
measured,
there
is
potential
selection
into
treatment.
However
since
both
treatment
and
control
were
offered
the
standard
CHD,
differential
selection
would
have
to
be
based
on
the
inclusion
of
deworming
in
the
Child
Health
Day.
This
is
possible,
if
we
assume
that
parents
know
their
childrens’
worm
infection
status
and
are
more
likely
to
bring
them
to
CHDs
when
they
learn
that
deworming
treatment
is
available.
7
and
2011
surveys
in
Uganda,
Uwezo
sampled
22
out
of
the
48
parishes
that
had
participated
in
the
deworming
study
in
2000-‐2003,
surveying
1,097
children
between
the
ages
of
6
and
16.
763
out
of
these
1,097
children
surveyed
by
Uwezo
in
2010
or
2011
were
between
the
ages
of
1
and
7
during
the
deworming
study
period
(2000
to
2003),
and
were
therefore
eligible
to
be
dewormed.15
10
of
the
parishes
sampled
by
Uwezo
were
treatment
parishes,
and
12
were
control.
Table
1
(appendix)
shows
the
relationship
between
age
at
survey
and
the
age
that
the
respondent
had
attained
during
the
period
of
program
implementation,
from
2000-‐2003.
Table
2
shows
the
means
across
a
range
of
socioeconomic
variables
between
treatment
and
control
parishes,
demonstrating
that
on
observed
socioeconomic
variables
in
the
Uwezo
data
there
are
no
significant
differences
between
respondents
in
the
treatment
and
control
parishes
in
2010-‐
2011.
Table 1: comparison of means, treatment and control parishes, combined 2010 and 2011 data
5 Analysis
Given
the
random
allocation
of
treatment,
and
the
balance
on
observables
between
treatment
and
control
parishes,
I
use
a
simple
econometric
framework
to
estimate
the
impact
of
deworming
treatment.
As
shown
in
table
2,
I
estimate,
in
a
regression
framework,
the
effect
of
being
a
child
in
a
deworming
(“treatment”)
parish
during
the
2000-‐2003
period.
This
means
that
I
include
all
children
in
Uwezo’s
sample
who
were
aged
1-‐7
during
the
study
period,
and
who
are
tested
in
2010
or
2011
in
one
of
the
treatment
or
control
parishes
(710
children
in
22
parishes).
Tests
scores
are
standardized,
with
mean
zero
and
standard
deviation
of
1.
The
main
result
is
that
I
observe
sizable
treatment
effect
from
being
dewormed,
with
treatment
coefficients
varying
from
0.16
to
0.36
standard
deviations.
The
coefficient
is
positive
in
all
specifications,
and
it
significant
at
the
5%
level
for
math
(with
controls),
at
10%
level
for
English
15
Note
that
no
control
parishes
were
sampled
by
Uwezo
in
2010.
8
(with
controls)
and
math
(without
controls)
and
at
5%
for
the
combined
math/English
score
(with
controls).
In
table
2,
odd
number
specifications
are
the
simple
difference
of
means
(using
only
the
treatment
indicator)
while
even
numbered
columns
include
controls
for
age,
gender,
survey
round,
and
all
interactions
of
age,
gender,
and
survey
year.16
Table 2: effect of deworming on test scores, all exposed cohorts
I
also
test
treatment
intensity,
or
whether
the
effect
size
increases
with
the
number
of
years
exposed
to
treatment.
This
can
be
tested
separately
from
age
at
exposure
since
different
age
cohorts
got
either
zero,
one,
two,
three,
or
four
years
of
deworming
depending
on
their
age
at
program
initiation
and
ending.
I
group
respondents
into
having
either
received
0-‐1
year
of
treatment,
or
receiving
2-‐4
years
of
treatment.
Effects
are,
in
most
specifications,
two
to
three
times
larger
in
magnitude
for
the
more
intensely
treated
group;
and
I
can
reject
that
the
effects
are
of
equal
magnitude
for
math
scores,
although
not
for
the
other
two
categories.
Standard errors in second row, robust standard errors clustered at parish level. Even numbered columns contain
(unreported) controls for age, gender, survey year, and interactions of all of these variables.
*
p < 0.1, ** p < 0.05, *** p < 0.01
16
For
comparability
I
follow
the
specification
used
by
the
other
paper
in
this
literature
that
attempts
to
measure
long
run
cognitive
impacts
of
early
childhood
deworming,
i.e.
Ozier
(2011).
9
6
Treatment
heterogeneity
In
this
section
I
test
for
two
potential
sources
of
treatment
heterogeneity:
gender,
and
wealth
quintile.
The
first
treatment
interaction
is
by
gender.
Coefficients
on
the
interaction
term
are
positive
in
all
specifications.
The
point
estimates
suggest
an
added
benefit
of
deworming
for
female
children;
and
the
difference
between
the
coefficient
for
male
and
the
female
interaction
term
is
significant
at
5%
level
for
math
(with
controls)
and
at
10%
for
total
score
(with
controls).
Second,
I
test
whether
there
are
interactions
with
household
wealth,
as
proxied
by
a
household
asset
wealth
index.
Respondents
are
coded
as
belonging
to
a
“poor”
household
if
they
fall
into
the
poorest
or
second-‐poorest
wealth
quintile.
Coefficients
on
the
interaction
of
treatment
and
the
poverty
indicator
are
uniformly
positive,
suggesting
an
added
positive
effect
for
the
poorest
households.
Equality
of
coefficient
and
interaction
can
be
rejected
at
5%
significance
for
math
(with
controls),
and
at
10%
significance
for
English
and
total
score
(both
with
controls).
10
7 Robustness checks
In
this
section,
I
present
a
series
of
robustness
checks.
First,
I
examine
the
data
for
potential
covariate
imbalance
(even
though
table
1
shows
no
statistically
significant
differences)
and
re-‐
estimate
our
preferred
specifications
using
only
observations
that
have
equal
values
of
the
baseline
value.
For
example,
although
none
of
the
components
of
the
asset
index
(ownership
of
television,
mobile
phone,
and
radio,
access
to
electricity,
access
to
water)
show
statistically
significant
differences
between
treatment
and
control,
the
one
variable
from
this
index
that
shows
any
substantive
(although
non-‐significant)
difference
between
treatment
and
control
is
access
to
water,
which
is
almost
10
percentage
points
higher
in
the
treatment
group.
Therefore
to
test
if
this
imbalance
drives
the
result,
I
restrict
the
sample
to
households
without
access
to
water.
The
results
are
robust
to
this
adjustment.
Another
variable
with
some
(though
not
statistically
significant)
baseline
imbalance
is
the
percentage
of
mothers
with
no
education,
which
is
higher
in
the
control
group.
I
again
re-‐estimate
the
main
specifications
among
mothers
with
at
least
some
education.
The
results
do
not
change.
Second,
I
include
district
fixed
effects.
In
this
specification,
we
can
reject
the
null
at
p=0.05
for
all
five
out
of
the
six
regressions.
Third,
I
generate
a
new,
“placebo”
treatment
group
from
among
the
list
of
other
parishes
in
the
study
area
which
were
sampled
by
Uwezo
but
which
were
not
chosen
as
participants
in
the
deworming
study.
In
this
model
the
“treatment”
variable
is
positive
but
the
effect
does
not
reach
statistical
significance,
either
in
the
main
specifications
or
when
socioeconomic
variables
are
used
as
controls.
I
also
estimate
regressions
using
data
from
only
the
2011
Uwezo
survey
round,
since
no
control
parishes
were
selected
in
2010,
and
I
(separately)
exclude
the
two
parishes
that
were
initially
selected
for
the
study
but
did
not
participate.
Finally,
I
re-‐estimate
the
main
specifications
with
controls
for
wealth
quintiles,
generated
using
the
first
principal
component
of
an
asset
index.
The
main
results
come
through
in
each
of
these
specifications,
and
in
most
instances
significance
is
strengthened.
Each
of
these
regression
tables
in
presented
in
the
appendix
(tables
10-‐13).
Another
potential
concern
is
that,
since
Uwezo
only
randomly
sampled
22
out
of
the
48
study
clusters
from
the
original
experiment,
normal
regression
estimates
with
clustered
standard
errors
might
be
overly
likely
to
reject
the
null
hypothesis,
unless
an
adjustment
for
small
number
of
clusters
is
applied.
Therefore
table
7
re-‐estimates
the
original
specifications
using
Wild
cluster
bootstrapped
standard
errors.
In
these
specifications
the
results
for
math
are
still
highly
significant
while
the
English
and
the
total
score
coefficients
are
significant
at
the
10%
level
when
district
fixed
effects
are
included.
11
Table
6:
main
model
with
Wild
cluster
bootstrapped
(Cameron
et
al.
2008)
standard
errors
A
final,
intuitive
robustness
check
is
to
simply
examine
the
unadjusted
pattern
of
scores
by
age.
The
youngest
cohorts
surveyed
in
the
treatment
group
were
too
young
to
benefit
from
multiple
years
of
the
program
(for
example,
child
aged
7
in
2010
were
born
in
2003
and
could
only
have
been
dewormed
twice
at
most),
and
the
oldest
children
were
too
old
(a
16
year
old
in
2003
would
have
aged
out
of
the
1-‐7
age
group
after
1
year
of
the
program),
it
seems
reasonable
to
expect
small
differences
between
treatment
and
control
among
the
oldest
groups
and
the
youngest
groups,
and
larger
differences
in
the
middle
age
groups.
That
is
exactly
the
pattern
that
can
be
seen
in
the
unadjusted
test
scores,
as
figure
1
demonstrates.
12
Figure
1:
total
scores
by
age
14.00
12.00
10.00
8.00
treat
6.00
control
4.00
2.00
0.00
6
7
8
9
10
11
12
13
14
15
16
age
8
Discussion
The
finding
that
early
childhood
deworming
appears
to
have
long
lasting
positive
impact
on
test
scores,
and
(implicitly)
cognitive
ability
is,
on
one
hand,
very
consistent
with
Ozier’s
(2011)
finding
that
children
in
Western
Kenya
who
were
dewormed
under
1
show
large
cognitive
benefits
compared
to
children
dewormed
at
older
ages,
and
the
long
run
labor
market
gains
shown
by
Baird
et
al
(2011).
And
they
are
consistent
with
the
significant
weight
gain
that
Alderman
et
al’s
(2007)
regression
models
suggest
was
the
effect
of
the
deworming
program
in
the
short
run.
On
the
other
hand,
they
are
in
tension
with
the
findings
of
the
Cochrane
Collaborative
systematic
review,
which
finds
no
evidence
of
weight
gain
or
cognitive
benefit.
How
can
these
findings
be
reconciled?
While
the
section
7
examined
the
statistical
robustness
of
the
estimates
presented
in
this
paper,
in
this
section
I
take
a
step
back
to
consider
the
plausibility
of
the
effect
sizes
identified,
potential
mechanisms,
and
consistency
with
the
broader
literature.
One
circumstance
under
which
these
estimates
might
not
be
plausible
would
be
if
the
original
deworming
had
only
reached
a
small
fraction
of
the
treatment
population.
Since
I
measure
what
are
in
effect
“intent
to
treat”
estimates,
and
yet
still
see
large
effects,
this
implies
that
the
effects
are
actually
proportionally
larger
for
those
children
actually
dewormed
(we
do
not
observe
actual
deworming
status
in
the
Uwezo
survey;
all
we
know
is
whether
children
live
in
parishes
which
were
provided
with
deworming).
For
example,
if
on
average
only
25%
of
children
were
ever
dewormed
by
the
program
in
treatment
parishes,
then
Uwezo’s
random
13
sample
would,
on
average,
be
comprised
of
75%
untreated
children
and
25%
“ever
treated”
children.
Assuming
no
spillovers
(such
that
only
treated
respondents
benefited
from
deworming),
large
observed
effects
at
population
level
would
imply
very
large,
perhaps
implausible,
“treatment
on
treated”
(TOT)
effects
–
on
the
order
of
0.8-‐1.2
standard
deviations.
But
program
coverage,
as
measured
by
Alderman
et
al
(2007)
was
actually
quite
high.
66%
of
children
between
ages
1-‐7
in
treatment
districts
report
being
dewormed
in
the
past
two
years,
and
74%
reported
attending
at
least
one
child
health
day
in
the
past
two
years.
The
average
child
attended
1.74
child
health
days
over
the
three
year
program
period.17
Thus,
depending
on
whether
one
uses
the
66%
figure
or
the
74%
figure
as
closer
approximation
of
population
coverage,
the
proportion
of
children
in
study
parishes
(and
in
the
sample)
that
were
dewormed
at
least
once
is
therefore
likely
between
two-‐thirds
and
three
quarters.
These
implied
TOT
effects
(with
the
conservative
assumption
of
no
spillovers),
are
therefore
large
but
not
so
large
as
to
be
ex
ante
implausible.
Another
factor
working
against
the
plausibility
of
the
estimates
presented
here
it
that
there
was
also
apparently
some
crossover
or
contamination
between
treatment
and
control.
While
coverage
of
deworming
tripled
between
2000
and
2003
(from
22%
to
66%)
in
treatment
parishes,
it
also
reportedly
increases
by
approximately
50%
in
the
control
parishes
(from
24%
to
35%).18
When
we
scale
our
implied
TOT
effects
to
account
for
this
crossover,
we
are
again
at
an
effect
size
close
to
1
standard
deviation.
But
again,
this
is
under
the
conservative
assumption
of
no
spillover
effects.
Yet
in
a
spatially
and
temporally
nearby
setting
(the
other
side
of
the
Kenyan
border,
3-‐5
years
earlier),
both
Miguel
and
Kremer
(2004)
and
Ozier
(2011)
document
sizable
spillovers
from
deworming
treatment.
Without
knowing
more
about
the
magnitude
of
the
spillover
in
the
Ugandan
case,
it
is
difficult
to
precisely
judge
the
plausibility
of
the
implied
effect
size.
8b Potential mechanisms
Although
the
fact
that
identification
in
this
study
is
based
on
the
original
randomization
should
generate
confidence
in
the
internal
validity
of
these
results,
we
unfortunately
lack
corroborating
information
that
could
suggest
the
precise
mechanism
might
be
driving
this
result.
For
example,
if
the
early
life
weight
gain
induced
by
randomization
translated
into
greater
observed
adolescent
height
in
the
treatment
group,
we
could
be
confident
that
we
are
observing
a
biomedical
pathway
linked
to
early
life
nutrition
(via
deworming),
and
not
a
statistical
artifact.
Similarly,
we
would
ideally
have
a
broad
range
of
cognitive
tests
to
better
identify
the
channel
through
which
improved
numeracy
and
literacy
is
manifested.
Unfortunately,
the
flip
side
of
Uwezo’s
extremely
large
sample
is
that
the
survey
instrument
itself
is
quite
limited.
This
means
that
we
do
not
have
any
evidence
on
health
outcomes
from
17
See
Alderman
et
al.
Table
4.
18
In
both
groups,
the
average
child
attended
1.74
child
health
days
over
the
three
year
period.
14
Uwezo,
such
as
anthropometry,
or
detailed
cognitive
measurements
beyond
basic
numeracy
and
literacy.
As
a
result,
we
cannot
test
whether
specific
health
channels
are
operational,
or
which
components
of
cognition
are
driving
improved
math
and
English
scores.
Nor,
since
we
do
not
have
information
on
program
attendance
at
the
original
deworming
program,
can
we
determine
whether
these
effects
are
driven
by
the
respondents
who
were
actually
dewormed,
or
whether
spillover
effects
play
a
large
role.
Another
limitation
is
that
this
study
did
not
compare
a
deworming
program
versus
a
pure
control,
but
rather
deworming
plus
a
standard
Child
Health
Day
(Vitamin
A,
growth
monitoring)
versus
a
standard
Child
Health
Day
without
deworming.
Therefore
if
there
are
any
interactions
between
deworming
and
other
components
of
the
Child
Health
Day
package,
the
results
from
this
context
should
be
interpreted
in
light
of
this.
Finally,
it
is
also
the
case
that
shortly
after
the
trial,
the
Ugandan
Ministry
of
Health
made
deworming
a
standard
component
of
Child
Health
Days.
Thus,
after
the
study
period,
both
groups
had
access
to
routine,
free
deworming
treatment.
Retuning
to
the
question
of
causal
mechanisms,
the
one
channel
that
can
be
tested
relates
to
school
enrollment.
If
there
were
persistent
health
effects
of
early
childhood
deworming
(as
distinct
from
cognitive
or
growth
effects)
we
might
see
that
children
dewormed
in
early
life
attend
school
more
(in
2009/2010)
or
are
less
likely
to
never
have
enrolled
in
school
between
when
they
were
dewormed
in
2000-‐2003
and
when
they
were
surveyed
in
2009/2010.
However
there
is
no
relationship
between
these
measures
and
treatment
status.
Soil
transmitted
helminths
are
an
infectious
disease,
so
an
individual’s
likelihood
of
infection
decreases
by
some
unknown
magnitude
as
his
neighbors’
levels
of
infection
decrease,
making
spillover
effects
a
potentially
important
factor.
However,
the
vast
majority
of
the
evidence
used
in
systematic
reviews
(and
driving
the
systematic
review
conclusion
that
mass
deworming
is
ineffective)
comes
from
studies
randomized
at
the
individual
level
(and
therefore
with
results
15
potentially
attenuated
by
spillovers).
For
example
in
the
second
Cochrane
review
(Taylor-‐
Robinson
et
al
2007),
only
3
out
of
34
included
trials
were
cluster
randomized
and
analyzed
with
design
effects19,
while
in
the
2012
updated
review,
8
out
of
42
were
cluster
randomized,
and
only
5
of
these
were
originally
analyzed
using
design
effects
(in
2012
the
review
team
re-‐
analyzed
using
imputed
design
effects
if
results
were
not
originally
reported
in
this
way.)
The
three
cluster
randomized
trials
in
the
2007
Cochrane
review
are
Stoltzfus
et
al
(1997);
and
two
trials
led
by
Awasthi
et
al.
(2000;
20001).
These
studies,
unlike
many
of
the
individually
randomized
ones,
find
positive
treatment
effects.
Stoltzfus
et
al.
found
positive
effects
on
growth
among
a
sub-‐group
(children
under
10)
school
aged
children
on
the
Zanzibari
island
of
Pemba
in
a
school-‐based
deworming
program
after
1
year,
using
a
single
dose
of
mebendazole.
Awasthi
et
al
(2001)
studied
twice-‐annual
doses
of
albendazole
in
60
urban
clusters
of
Lucknow,
in
Uttar
Pradesh
state
in
India.
After
1.5
years
treatment
group
showed
significant
gains
in
weight.20
Awasthi
et
al.
(2000)
examined
albendazole
treatment
in
the
same
slums
(in
a
separate
trial)
and
found
reduced
risk
of
stunting
in
the
treatment
group.
The
2012
Cochrane
review
added
in
Miguel
and
Kremer
(2004)
and
Alderman
et
al.
(2007),
both
of
which
were
cluster
randomize
trials
with
positive
treatment
effects.
It
also
added
in
Rousham
(1994)
and
Hall
(2006),
which
were
cluster
randomized
trials
without
positive
and
significant
effects.
Nonetheless
it
is
suggestive
that
while
the
overall
Cochrane
meta-‐analysis
shows
no
effect,
the
majority
of
clustered
trials
show
positive
effects.
This
is
consistent
with
the
critique
offered
by
Bundy
et
al
(2009).
However,
since
the
publication
of
the
2012
updated
systematic
review,
new
research
has
come
to
light
which
further
complicates
any
simple
narrative
about
clustered
versus
unclustered
trials.
In
2013
the
results
of
the
“DEVTA”
trial
from
northern
India,
the
largest
deworming
trial
to
date,
were
published.
The
DEVTA
trial
(which
was
a
factorial
design
trial
of
Vitamin
A
supplementation
and
deworming)
showed
no
significant
impact
of
deworming
on
children’s
weight,
in
a
large,
cluster-‐randomized
sample
in
Uttar
Pradesh.
(The
difference
in
weight
between
treatment
and
control
was
significant
at
the
10%
level
but
small
in
magnitude,
at
0.04
kg).
While
the
DEVTA
trial
took
place
in
an
environment
of
relatively
low
worm
prevalence,
it
is
difficult
to
say
whether
this
explains
the
differences
between
this
outcome
and
that
of
other
cluster-‐randomized
trials.
Further
trials
in
high
prevalence
settings,
and
with
specific
focus
on
pre-‐school
age
children,
seem
justified
at
this
stage.
9 Conclusion
19
Six
of
theses
trials
were
cluster
randomized,
but
three
of
these
did
not
use
design
effects
in
analysis
and
so
are
not
included
in
the
meta-‐analysis.
20
The
2012
review
includes
these
three
studies
plus
Miguel
and
Kremer
(2004);
Alderman
et
al
(2007);
Hall
2006;
and
Rousham
(1994).
16
Mass
deworming
of
school
age
children
is
a
highly
touted
policy
by
a
range
of
authoritative
sources
–
and
mass
deworming
of
preschool
age
children
has
seemed
to
be
a
promising
extension
of
this
policy.
However,
the
evidence
base
has
been
clouded
by
the
current
controversy
over
the
effect
of
deworming,
as
shown
by
the
disagreements
between
the
researchers
associated
with
the
Cochrane
Collaborative
review,
and
the
group
of
deworming
proponents,
associated
with
IPA,
JPAL,
Evidence
Action,
the
World
Bank,
and
other
institutions.
This
paper
exploits
a
new
data
source
to
identify
large
educational
benefits
to
a
group
of
school
children
dewormed
in
early
childhood
-‐
effects
which
are
present
despite
the
fact
that
educational
outcomes
are
measured
7-‐10
years
after
the
end
of
the
deworming
experiment.
It
avoids
weaknesses
in
previous
studies
by
exploiting
a
cluster-‐randomized
approach
(thus
avoiding
attenuation
of
effect
via
spillovers).
As
such,
it
strengthens
the
case
that
there
are
important
and
persistent
cognitive
benefits
to
mass
deworming
in
settings
of
high
worm
prevalence.
17
References
Alderman,
H,
et
al.
(2007).
Alderman,
H.,
J.
Konde-‐Lule,
I.
Sebuliba,
D.
Bundy,
and
A.
Hall
(2006):
“Effect
on
weight
gain
of
routinely
giving
albendazole
to
preschool
children
during
child
health
days
in
Uganda:
cluster
randomised
controlled
trial,”
British
Medical
Journal,
333,
122–127.
Awasthi
S,
Pande
VK.
Six-‐monthly
de-‐worming
in
infants
to
study
effects
on
growth.
Indian
Journal
of
Pediatrics
2001;68(9):823–7.
Awasthi
S,
Pande
VK,
Fletcher
RS.
Effectiveness
and
cost-‐
effectiveness
of
albendazole
in
improving
nutritional
status
of
pre-‐school
children
in
urban
slums.
Indian
Pediatrics
2000;37(1):19–29.
Baird
S,
Hicks
JH,
Kremer
M,
Miguel
E.
Worms
at
work:
long-‐run
impacts
of
child
health
gains.
2012.
Accessed
at
http://scholar.harvard.edu/kremer/publications/worms-‐work-‐long-‐run-‐
impacts-‐child-‐health-‐gains.
Bleakley,
Hoyt.
(2007).
“Disease
and
Development:
Evidence
from
Hookworm
Eradication
in
the
American
South.”
Quarterly
Journal
of
Economics,
122(1):73-‐117.
Bundy,
D.
A.
P.,
M.
Kremer,
H.
Bleakley,
M.
C.
H.
Jukes,
and
E.
Miguel
(2009):
“Deworming
and
Development:
Asking
the
Right
Questions,
Asking
the
Questions
Right,”
Public
Library
of
Science:
Neglected
Tropical
Diseases,
3(1),
e362.
Gerber
A,
Green
D.
2012.
Field
Experiments:
Design,
Analysis,
and
Interpretation.
New
York:
WW
Norton.
Hall
A,
Nguyen
Bao
Khanh
L,
Bundy
D,
Quan
Dung
N,
Hong
Son
T,
Lansdown
R.
A
randomized
trial
of
six
monthly
deworming
on
the
growth
and
educational
achievements
of
Vietnamese
school
children.
Unpublished
manuscript.
Hawkes,
N.
2013.
“Deworming
Debunked.”
British
Medical
Journal.
Accessed
at
http://www.bmj.com/content/bmj/346/bmj.e8558.full.pdf.
doi:
10.1136/bmj.e8558
Hotez
P
et
al.
“Helminth
Infections:
Soil
Transmitted
Helminth
Infections
and
Schistosomiasis.”
In
Disease
Control
Priorities
in
Developing
Countries.
2006.
Jamison,
DT
et
al
(eds).
Washington
DC:
World
Bank,
and
New
York:
Oxford
University
Press.
Kabatereine,
N.
B.,
Tukahebwa,
E.,
Brooker,
S.,
Alderman,
H.,
&
Hall,
A.
(2001).
Epidemiology
of
intestinal
helminth
infections
among
schoolchildren
in
Southern
Uganda.
East
African
Medical
Journal.
78(6),
283–286.
Miguel,
E.,
and
M.
Kremer
(2004):
“Worms:
Identifying
Impacts
on
Education
and
Health
in
the
Presence
of
Treatment
Externalities,”
Econometrica,
72(1),
159–217.
18
Ministry
of
Health
(Uganda)
and
Macro
International
(2008).
Uganda
Service
Provision
Assessment
2007.
Kampala,
Uganda:
Ministry
of
Health,
and
Calverton,
MD:
Macro
International.
Ozier,
O.
(2010).
“Exploiting
Externalities
to
Estimate
the
Long-‐Term
Benefits
of
Early
Childhood
Deworming”,
unpublished
working
paper,
U.C.
Berkeley.
Northrop-‐Clewes
CA,
Rousham
EK,
Mascie-‐Taylor
CN,
Lunn
PG.
Anthelmintic
treatment
of
rural
Bangladeshi
children:
effect
on
host
physiology,
growth,
and
biochemical
status.
American
Journal
of
Clinical
Nutrition
2001;73(1):
53–60.
Stoltzfus
RJ,
Albonico
M,
Tielsch
JM,
Chwaya
HM,
Savioli
L.
School-‐based
deworming
program
yields
small
improvement
in
growth
of
Zanzibari
school
children
after
one
year.
Journal
of
Nutrition
1997;127(11):2187–93.
Taylor-‐Robinson
D,
Jones
A,
Garner
P.
(2007)
“Deworming
drugs
for
treating
soil-‐transmitted
intestinal
worms
in
children:
effects
on
growth
and
school
performance.”
Cochrane
Database
of
Systematic
Reviews
2007,
4,
CD000371.doi:10.1002/14651858.CD000371.pub31.
Taylor-‐Robinson
DC,
Maayan
N,
Soares-‐Weiser
K,
Donegan
S,
Garner
P.
Deworming
drugs
for
soil-‐transmitted
intestinal
worms
in
children:
effects
on
nutritional
indicators,
haemoglobin
and
school
performance.
Cochrane
Database
Syst
Rev
2012;11:CD000371.
Waddington,
H
and
Leach,
B.
March
4,
2014.
“How
Much
Evidence
is
Enough
for
Action?”
Blog
post
accessed
at
http://blogs.3ieimpact.org/how-‐much-‐evidence-‐is-‐enough-‐for-‐action/.
19
Table
9:
age
at
survey
/
age
at
treatment
for
2010
and
2011
survey
rounds
survey
year
2011
2010
2011
2010
2011
2010
2011
2010
age
in
year:
2000
2001
2002
2003
6
-‐5
-‐4
-‐4
-‐3
-‐3
-‐2
-‐2
-‐1
7
-‐4
-‐3
-‐3
-‐2
-‐2
-‐1
-‐1
0
8
-‐3
-‐2
-‐2
-‐1
-‐1
0
0
1
9
-‐2
-‐1
-‐1
0
0
1
1
2
10
-‐1
0
0
1
1
2
2
3
11
0
1
1
2
2
3
3
4
12
1
2
2
3
3
4
4
5
13
2
3
3
4
4
5
5
6
14
3
4
4
5
5
6
6
7
15
4
5
5
6
6
7
7
8
16
5
6
6
7
7
8
8
9
20
Table
11a:
analysis
restricted
to
households
without
water
supply
at
home
(odd
columns)
or
mothers
without
education
(even
columns)
Table 11b: Full sample, controlling for lack of water supply/mother’s education
21
Table 12: Additional socioeconomic controls (wealth quintiles, private school attendance)
22
Appendix: Randomization inference figures
Tables generated using R code adapted from Gerber and Green (2012)
23
24
25