Professional Documents
Culture Documents
Program Evaluation
ProgEval 03 Randomization 2
Randomized Controlled Trial (RCT)
Random assignment of people into a treated group
and a non-treated group
Think of a medical experiment
Wi ǁ (Y(0)i, Y(1)i)
Where sign ǁ here means independence
The distribution of (Y(0)i, Y(1)i)
their joint distribution
ProgEval 03 Randomization 4
Randomized Assignment
Assignment independent from both Y(0) and Y(1)
Participation is unrelated to whether someone would
have higher or lower outcome if treated
Participation is unrelated to whether someone would
have higher or lower outcome if untreated
ProgEval 03 Randomization 5
Randomized Assignment
Whatever “randomness” means is not really
important: what we need here is independence
The idea of “true” randomness is controversial
Is the result of a coin flip random or is it predictable if we
know the forces, the properties of the coin, the wind, etc.?
In practice randomization is usually carried out using
computer-generated “pseudo-random sequences”
But that is not really important for us
ProgEval 03 Randomization 6
Independence of What from What?
Independence of treatment status from potential
outcomes
Of course, that does not mean that treatment is
independent of observed outcomes
If the program has positive effect, observed outcomes for
the treated may be better than the observed outcomes of
the untreated
Observed outcomes for the treated are Y(1) and observed
outcomes for the untreated are Y(0)
ProgEval 03 Randomization 7
Randomization In Practice
Create a group from the potential participants
List the potential participants
Split them into two groups
using randomization device (coin, computer-generated pseudo-
random number etc.)
ProgEval 03 Randomization 8
Consequences of Randomization
Potential outcomes are similar in the treatment group
and the control group
The average potential outcomes are the same
E[Y(1)i|Wi=1] = E[Y(1) i|Wi=0]
and E[Y(0) i|Wi=1] = E[Y(0) i|Wi=0]
So that
E [Y(1) i] = P(Wi=1)E[Y(1)i|Wi=1] + P(Wi=0)E[Y(1) i|Wi=0]
= E[Y(1) i|Wi=1] = E[Yi|Wi=1]
E [Y(0) i] = P(Wi=1)E[Y(0) i|Wi=1] + P(Wi=0)E[Y(0) i|Wi=0]
= E[Y(0) i|Wi=0] = E[Yi|Wi=0]
ProgEval 03 Randomization 9
Consequences of Randomization
So that the average treatment effect is identified
by the difference of observed outcomes in the
treated group vs. the untreated group
ProgEval 03 Randomization 12
The RCT Movement
The gold standard of evaluations
High internal validity
Easy to communicate
ProgEval 03 Randomization 13
The RCT Movement
There is a movement out there to propagate
randomized evaluations
The JPAL group
http://www.povertyactionlab.org/
ProgEval 03 Randomization 14
RCTs Beyond Program Evaluation
“Field experiments” vs. “lab experiments”
See the field experiment site by John List
www.fieldexperiments.com
Most experiments in econ are done in labs
ProgEval 03 Randomization 15
Case Study: NSW
The National Supported Work demonstration
U.S., 1970’s
Run and evaluated by non-profit organization Manpower
Demonstration Research Corporation (MDRC)
Source:
LaLonde (1986)
ProgEval 03 Randomization 19
Effect on Employment (Women)
Source:
Ham &
LaLonde
(1996)
ProgEval 03 Randomization 20
Things to Note On Graph
Treatment & control very similar before 0
Employment rates at months 26 are different
38% vs. 28%, cca. 10 pctage points difference
ProgEval 03 Randomization 24
Take-away
Randomized assignment of treatment
Treatment status independent of potential outcomes
Wi ǁ (Y(0)i, Y(1)i)