Post on 04-Jan-2016
Nigeria Impact EvaluationCommunity of Practice
Abuja, Nigeria, April 2, 2014
Measuring Program Impacts ThroughRandomization
David Evans (World Bank)
2
Objective
Evaluate the causal impact of a program or an intervention on some outcome
Examples How much did free distribution of
bednets decrease malaria incidence?
Which of two supply chain models was most effective at eliminating drug shortages?
3
Counterfactual Criteria
Treated & comparison groups…Have identical average
characteristics (observed & unobserved)
The only difference is the treatment
Therefore the only reason for any difference in outcomes is the treatment
Key question: What would participant look like if she hadn’t received the program?
4
Perfect Experiment
1. Identify target beneficiaries
2. Clone them!
Identical on the outside (observable)
Identical on the inside (unobservable)
Chief Ahun
We’re both middle-aged
chiefs
We both love to take up new
health interventions!
Chief Batun
5
Perfect Experiment
Give the intervention to one set of clones
Ahun Batun
6
Perfect Experiment
Observe some time later
Because the groups are identical (inside & out), the difference is due to the bednets!
Ahun
Batun
7
Back to Reality
What would Batun look like if he didn’t receive the bednet?
Room For Improvement Control Groups
Before – After Participants – Non Participants
???
RFI: Before-After
BEFORE BEDNETS
6 malaria episodes in 6 months
AFTER BEDNETS
2 malaria episodes in 6 months
What else might be going on besides the bednets?
• Seasonal differences• Rising incomes: Households invest in other measures
Too many other factors!
Impact of bednets = ???
RFI: Before-After
Important to monitor before-after Insufficient to show impact of
programToo many factors changing over time
Example of cash transfers in Nicaragua!
Counterfactual: What would have happened in the absence of the project, with everything else the same
10
RFI: Participants vs Non-Participants
Compare recipients of a program to People who were not eligible for the
program People who chose not to enroll in the
program
Home births Clinic births
Example: Complications in childbirth
Impact of clinic births?
What else might explain the difference?
11
RFI: Participants vs Non-Participants
Observable differences Income Education
Unobservable differences Heard rumor about hospitals Neighbor available to care for other
children
Mercy Patience
12
RFI: Participants vs Non-Participants
How much of difference is because of clinic?
Impact of clinic births = ???
Home births Clinic births
Example: Complications in childbirth
Impact of clinic births
Other factors!
13
Selection bias
People who choose to join the program are different!
If we cannot account completely for those differences in our data…
We never can How do you capture attitudes toward
health systems? Initiative? …then our comparison will not show
the true impact of the program
What should we do?
Gold standard:Randomized experimental design
15
Randomized Experimental Design
Randomly assign potential beneficiaries to be in the treatment or comparison group
Treatment and comparison have the same characteristics (observed and unobserved), on average, so…
Any difference in outcomes is due to treatment
16
Why Randomization Works
Randomization with two doesn’t work!
But differences average out in a big sample
On average, same number of Ahuns and Patiences Observable AND unobservable
Result: Measure true impact of program
ComparisonTreatment
Comparison
Treatment
17
Random Sample or Assignment?
RANDOM SAMPLE
Select randomly who to gather data on
Gives unbiased average of the group NOT of impact
If take random sample of group: Half women, half men – Sample should be about ½ women, ½ men
RANDOM ASSIGNMENT
Randomly assign who gets the program
Gives unbiased estimate of program impact
Why? Treatment & comparison
are IDENTICAL (on average)
T
T
C
C
18
LET’S RANDOMIZE!
1. Identify the eligible participants
What is the impact of
receiving a new car on body-mass
index?
19
LET’S RANDOMIZE!
2. Generate a random number for each one
20
LET’S RANDOMIZE!
2. Generate a random number for each one
21
LET’S RANDOMIZE!
3. Re-order based on the random number
22
LET’S RANDOMIZE!
4. Assign the first ten to receive cars
I really wanted a car!
23
LET’S RANDOMIZE!
5. Check for balance across treatment & control
Treatment# Drs: 4
Control# Drs: 6
Total observations21
Not a very big sample!
How close? 2/3
24
LET’S RANDOMIZE!
What if we had 500 observations?
Treatment# Drs: 114
Control# Drs: 116
How close? 98/100
25
Is there more?
That’s a simple way to randomize Works with BIG samples
You can help randomization by stratifying Randomize within each sub-group Ensure that each group is equally
represented
26
Stratifying
1. Identify the characteristic(s) to stratify on
27
Stratifying
2. Sort on those characteristics
28
Stratifying
3. Generate those random numbers
29
Stratifying
4. Sort on those characteristics
30
Stratifying
5. Assign treatment within each sub-group
Result: Equal doctors in each group
31
Can we really randomize?
Randomization does not mean denying people the benefits of the project
Usually existing constraints in project roll-out allow randomization
Randomization often the fairest way to allocate treatment
32
Use Staggered Roll-out
Roll-out to 200 clinicsRoll-out to 400 more
clinics
Jan 2014
July 2014
Jan 2015
• Randomize the order in which clinics receive program
• Compare Jan 2014 group to Jan 2015 group at end of first year
33
Some groups must get the program
34
Vary treatment
INTENSITY OF TREATMENT
Malaria Information Campaign•100 villagesMalaria Information Campaign + SMS Reminders•100 villages
NATURE OF TREATMENT
Radio campaign•100 villages
Newspaper campaign•100 villages
35
Randomization is often the fairest
Watch the movie![link]
Randomization of an early child development program in Côte d’Ivoire
36
What if randomization is impossible?
Think again: It often is possible on some level, and it’s the best way to get a clear measure of impact Always begin the IE with imagining what the
ideal would look like
With a national policy Use randomization to test implementation
37
Key takeaway #1
The single best way to evaluate the unbiased average impact of an activity is by randomizing treatment.
38
Key takeaway #2
It is more ethical to test programs rigorously before universally implementing them than it is to use scarce public resources to implement a universal program with uncertain benefits.
39
Key takeaway #3
Randomization is more flexible than you think:
It does not require withholding of benefits.
It can take advantage of necessary staggered roll-out.
It can test different reforms or packages of services across groups at the same time (so all receive at least some package).
40
Let’s randomize
41
Thank you!
BONUS SLIDES
43
44
LET’S RANDOMIZE!
What if we had 100 observations?
Treatment# Drs: 26
Control# Drs: 22
Total observations21
Not a very big sample!