This preview shows page 1. Sign up to view the full content.
Unformatted text preview: Randomized Experiment: Deworming in Kenya
Estimation
Next Time Lect 2: Health and Introduction to Randomization
Nancy Qian January 12, 2011 Nancy Qian Econ 404/776: Lecture 2 Randomized Experiment: Deworming in Kenya
Estimation
Next Time Agenda
1 Randomized Experiment: Deworming in Kenya
The Basic Problem
Baseline Survey
Sample Attrition
Bundled Treatments
Comparison Group 2 Estimation
Compliance
Derivation
Results 3 Next Time
Nancy Qian Econ 404/776: Lecture 2 Randomized Experiment: Deworming in Kenya
Estimation
Next Time Basic Problem
Baseline Survey
Sample Attrition
Bundled Treatments
Comparison Group Outline
1 Randomized Experiment: Deworming in Kenya
The Basic Problem
Baseline Survey
Sample Attrition
Bundled Treatments
Comparison Group 2 Estimation
Compliance
Derivation
Results 3 Next Time
Nancy Qian Econ 404/776: Lecture 2 Randomized Experiment: Deworming in Kenya
Estimation
Next Time Basic Problem
Baseline Survey
Sample Attrition
Bundled Treatments
Comparison Group Program Evaluation Question: What are the eﬀects of deworming on school
attendance?
To address the question of whether health has a causal impact
on schooling, Miguel and Kremer (2004) randomly assigned
deworming drugs (across schools) to children in Kenya.
The PSDP project introduced deworming drugs into 3 groups
of schools. Groups 1& 2 will receive the drug and Group 3 is
the control group
The unit of randomization is at the school level Nancy Qian Econ 404/776: Lecture 2 Randomized Experiment: Deworming in Kenya
Estimation
Next Time Basic Problem
Baseline Survey
Sample Attrition
Bundled Treatments
Comparison Group Why Randomize? Why not just control for factors?
Are there pitfalls to randomization?
Why randomize across schools? Nancy Qian Econ 404/776: Lecture 2 Randomized Experiment: Deworming in Kenya
Estimation
Next Time Basic Problem
Baseline Survey
Sample Attrition
Bundled Treatments
Comparison Group Outline
1 Randomized Experiment: Deworming in Kenya
The Basic Problem
Baseline Survey
Sample Attrition
Bundled Treatments
Comparison Group 2 Estimation
Compliance
Derivation
Results 3 Next Time
Nancy Qian Econ 404/776: Lecture 2 Randomized Experiment: Deworming in Kenya
Estimation
Next Time Basic Problem
Baseline Survey
Sample Attrition
Bundled Treatments
Comparison Group Baseline Survey MK check the baseline characteristics by comparing the
characteristics of the students in the control and treatment groups
before the program begins
Why is this important?
What do you expect to see?
See Table I Baseline Characteristics
Baseline characteristics should be similar between samples. Nancy Qian Econ 404/776: Lecture 2 Randomized Experiment: Deworming in Kenya
Estimation
Next Time Basic Problem
Baseline Survey
Sample Attrition
Bundled Treatments
Comparison Group Descriptive Statistics One of the biggest mistakes empirical economists make is to not
closely examine the descriptive statistics
Many beneﬁts
Check randomization
Check data quality
Check that trends and patterns are realistic
See means – why is this important?
Check that there is variation – why is this important? Nancy Qian Econ 404/776: Lecture 2 Randomized Experiment: Deworming in Kenya
Estimation
Next Time Basic Problem
Baseline Survey
Sample Attrition
Bundled Treatments
Comparison Group Outline
1 Randomized Experiment: Deworming in Kenya
The Basic Problem
Baseline Survey
Sample Attrition
Bundled Treatments
Comparison Group 2 Estimation
Compliance
Derivation
Results 3 Next Time
Nancy Qian Econ 404/776: Lecture 2 Randomized Experiment: Deworming in Kenya
Estimation
Next Time Basic Problem
Baseline Survey
Sample Attrition
Bundled Treatments
Comparison Group Sample Attrition One problem in administering the program was that they could only
give students medication if students came to school that day
Poorer students who attend school less will be less likely to
receive treatment
Look at Table III shows that in treatment schools.
How many students in the treatment group receive treatment?
In the untreated group?
70% of the students receive treatment Nancy Qian Econ 404/776: Lecture 2 Randomized Experiment: Deworming in Kenya
Estimation
Next Time Basic Problem
Baseline Survey
Sample Attrition
Bundled Treatments
Comparison Group Does this pose a problem for the interpretation? Nancy Qian Econ 404/776: Lecture 2 Randomized Experiment: Deworming in Kenya
Estimation
Next Time Basic Problem
Baseline Survey
Sample Attrition
Bundled Treatments
Comparison Group Yes & No: The comparison will not tell us the eﬀect of the
program on the population at large
It will tell us the eﬀect of the program on students who receive
the drugs
As long as the same population of absent students exist in the
control group schools Nancy Qian Econ 404/776: Lecture 2 Randomized Experiment: Deworming in Kenya
Estimation
Next Time Basic Problem
Baseline Survey
Sample Attrition
Bundled Treatments
Comparison Group Outline
1 Randomized Experiment: Deworming in Kenya
The Basic Problem
Baseline Survey
Sample Attrition
Bundled Treatments
Comparison Group 2 Estimation
Compliance
Derivation
Results 3 Next Time
Nancy Qian Econ 404/776: Lecture 2 Randomized Experiment: Deworming in Kenya
Estimation
Next Time Basic Problem
Baseline Survey
Sample Attrition
Bundled Treatments
Comparison Group Bundled Treatments In addition to deworming medication, students at treatment
schools received health classes on cleanliness and sanitation
What problem does this pose for the interpretation of the
results?
How do you propose to resolve it? Nancy Qian Econ 404/776: Lecture 2 Randomized Experiment: Deworming in Kenya
Estimation
Next Time Basic Problem
Baseline Survey
Sample Attrition
Bundled Treatments
Comparison Group Outline
1 Randomized Experiment: Deworming in Kenya
The Basic Problem
Baseline Survey
Sample Attrition
Bundled Treatments
Comparison Group 2 Estimation
Compliance
Derivation
Results 3 Next Time
Nancy Qian Econ 404/776: Lecture 2 Randomized Experiment: Deworming in Kenya
Estimation
Next Time Basic Problem
Baseline Survey
Sample Attrition
Bundled Treatments
Comparison Group Comparison Group
What is the right comparison group? why?
Those who took the pill versus those who did not?
Those who took the pills versus the comparison group?
All of those initially in the treatment group (all of those who
were supplosed to take the pills) versus all of those assigned to
the comparison groups?
This comparison is called the Intention to Treat (ITT)
estimate How do we obtain the average eﬀect? Attrition
What would happen to the sample if the treatment students
were much healthier because of the experiment and the
comparion group saw no improvement?
Nancy Qian Econ 404/776: Lecture 2 Randomized Experiment: Deworming in Kenya
Estimation
Next Time Compliance
Derivation
Results Estimation
Denote
school attendance Y
receiving the deworming pill P
assignment into treatment T
assignment into control C If compliance was 100%, then the eﬀect of deworming on
school attendance is simply the ITT
E [Y T ] − E [Y C ]
But compliance is not perfect. Table III shows that compliance
is around 80%
How do we take this into account?
Nancy Qian Econ 404/776: Lecture 2 Randomized Experiment: Deworming in Kenya
Estimation
Next Time Compliance
Derivation
Results Outline
1 Randomized Experiment: Deworming in Kenya
The Basic Problem
Baseline Survey
Sample Attrition
Bundled Treatments
Comparison Group 2 Estimation
Compliance
Derivation
Results 3 Next Time
Nancy Qian Econ 404/776: Lecture 2 Randomized Experiment: Deworming in Kenya
Estimation
Next Time Compliance
Derivation
Results Compliance We can scale the ITT by the compliance to treatment
E [Y T ] − E [Y C ]
E [pill T ] − E [pill C ]
It is easy to see that in the perfect experiment,
E [pill T ] − E [pill C ] = 1 Nancy Qian Econ 404/776: Lecture 2 Randomized Experiment: Deworming in Kenya
Estimation
Next Time Compliance
Derivation
Results Outline
1 Randomized Experiment: Deworming in Kenya
The Basic Problem
Baseline Survey
Sample Attrition
Bundled Treatments
Comparison Group 2 Estimation
Compliance
Derivation
Results 3 Next Time
Nancy Qian Econ 404/776: Lecture 2 Randomized Experiment: Deworming in Kenya
Estimation
Next Time Compliance
Derivation
Results Derivation
To see how we address imperfect compliance, assume the
following linear relationship
Y = C + αP + ε
And that the probability of taking a pill is the probability of
taking the pill absent the treatment plus the probability of
taking the pill conditional on being treated.
K + γT + η
Then
E [Y T ] = E [C + α P + ε T ]
= E [C + α P T ]
= E [C + α (K + γ T + η )T ]
= C + αγ
Nancy Qian Econ 404/776: Lecture 2 Randomized Experiment: Deworming in Kenya
Estimation
Next Time Compliance
Derivation
Results Why did ε T and η T disappear?
Why did α K T disappear? Nancy Qian Econ 404/776: Lecture 2 Randomized Experiment: Deworming in Kenya
Estimation
Next Time Compliance
Derivation
Results Why did ε T and η T disappear?
Treatment is uncorrelated with error term due to
randomization Why did α K T disappear?
MK show that no one received treatment before the
intervention Nancy Qian Econ 404/776: Lecture 2 Randomized Experiment: Deworming in Kenya
Estimation
Next Time Compliance
Derivation
Results Wald Estimator
Similarly, using Table III shows that there are no pill recipients
in the control group
E [Y C ] = E [C + α P + ε C ]
= C + E [ε C ] = C
Hence the ITT
E [Y T ] − E [Y C ] = αγ
ˆ
α= ˆ
ITT
ˆ
γ Wald Estimator
ˆ
α is called the Wald Estimator , a binary 2SLS or instrumental
variables estimator.
Nancy Qian Econ 404/776: Lecture 2 Randomized Experiment: Deworming in Kenya
Estimation
Next Time Compliance
Derivation
Results Outline
1 Randomized Experiment: Deworming in Kenya
The Basic Problem
Baseline Survey
Sample Attrition
Bundled Treatments
Comparison Group 2 Estimation
Compliance
Derivation
Results 3 Next Time
Nancy Qian Econ 404/776: Lecture 2 Randomized Experiment: Deworming in Kenya
Estimation
Next Time Compliance
Derivation
Results Results Table V shows the results Nancy Qian Econ 404/776: Lecture 2 Randomized Experiment: Deworming in Kenya
Estimation
Next Time Compliance
Derivation
Results Externalities Outside Schools
Reducing the worm load of a student can reduce the probability
of infection of those near him, in school and outside of school.
How does Miguel and Kremer estimate the externality eﬀect?
How would you test this?
What if you were able to design an experiment, what would
you do?
What do they ﬁnd?
Costbeneﬁt:
Including the externality beneﬁts, the cost per additional year
of school participation is only $3.50 Nancy Qian Econ 404/776: Lecture 2 Randomized Experiment: Deworming in Kenya
Estimation
Next Time Next Time More on reduced form estimation and identiﬁcation.
Acemoglu and Johnson (2007) and Bleakeley (2007) Nancy Qian Econ 404/776: Lecture 2 ...
View
Full
Document
This note was uploaded on 08/12/2011 for the course ECON 404 taught by Professor Nancyqian during the Spring '11 term at Yale.
 Spring '11
 NANCYQIAN
 Economics

Click to edit the document details