Академический Документы
Профессиональный Документы
Культура Документы
Sharon Barnhardt
Assistant Professor, IFMR Research Director, Centre for Microfinance barnhardt@post.harvard.edu August 27, 2012
Lecture Overview
Evaluate impact through comparing outcomes to a counterfactual Methods for constructing the counterfactual:
Pre-post Simple Difference Difference-in-Difference Regression Analysis Randomization
Lecture Overview
Evaluate impact through comparing outcomes to a counterfactual Methods for constructing the counterfactual:
Pre-post Simple Difference Difference-in-Difference Regression Analysis Randomization
Counterfactual
The counterfactual represents the state of the world that program participants would have experienced in the absence of the program (i.e. had they not participated in the program)
Time
Impact
Counterfactual
Time
Primary outcome
Intervention
Time
Primary Outcome
T1
T2
Primary Outcome
Impact
T1
T2
We cant accurately estimate impact without finding a way to accurately represent the counterfactual
13
Lecture Overview
Evaluate impact through comparing outcomes to a counterfactual Methods for constructing the counterfactual:
Pre-post Simple Difference Difference-in-Difference Regression Analysis Randomization
Method 1: Pre-Post
Intervention
Primary Outcome
Measured Impact
Counterfactual
Time
Method 1: Pre-Post
Intervention
Primary Outcome
Actual Impact
Measured Impact
Counterfactual
Bias
Time
Method 1: Pre-Post
Intervention
Primary Outcome
Counterfactual
Bias
Time
Pre-post works when you can assume that there were no other factors that contributed to the change in outcomes over time
20
Lecture Overview
Evaluate impact through comparing outcomes to a counterfactual Methods for constructing the counterfactual:
Pre-post Simple Difference Difference-in-Difference Regression Analysis Randomization
22
Time
Actual Impact
Bias
Measured Impact
Time
Simple Difference
What are the potential problems with this? The comparison group might be inherently different (selection bias or otherwise)
Works when you can assume that non-participants: 1) Are identical to participants except for program participation 2) Were equally likely to enter program before it started
25
Lecture Overview
Evaluate impact through comparing outcomes to a counterfactual Methods for constructing the counterfactual:
Pre-post Simple Difference Difference-in-Difference Regression Analysis Randomization
Method 3: Difference-in-Difference
Primary Outcome
Intervention
Time
Method 3: Difference-in-Difference
Measured Impact
Intervention
Primary Outcome
Actual Impact
Time
Difference-in-Difference
Are there any potential problems with this?
Groups may have behaved differently over time Shock to one group and not the other
Works when you can assume that if the program didnt exist, the two groups would have had identical trajectories over this period.
30
Lecture Overview
Evaluate impact through comparing outcomes to a counterfactual Methods for constructing the counterfactual:
Pre-post Simple Difference Difference-in-Difference Regression Analysis Randoimzation
Regression Analysis
Special case of simple difference or difference in difference Run regression to predict outcome Have an explanatory variable indicating if the individual is treated or not treated AND other variables that could explain a difference in the outcome or account for selection into treatment.
Works when you have data on all of the other potential explanations.
Lecture Overview
Evaluate impact through comparing outcomes to a counterfactual Methods for constructing the counterfactual:
Pre-post Simple Difference Difference-in-Difference Regression Analysis Randomization
Nonrandomized
Randomized
Use random assignment of the program to create a control group which mimics the counterfactual.
Lecture Overview
Evaluate impact through comparing outcomes to a counterfactual Methods for constructing the counterfactual:
Pre-post Simple Difference Difference-in-Difference Regression Analysis Randomization
Proposed solution
Balsakhi Program: 1) Hire local teachers (balsakhis, female)
Given training to teach remedial classes in Hindi, Maths, English
Start of program
End of program
vs.
Comparison group should be as similar as possible Compare test score of these two groups at the end of the program.
56.27
-5.05
50 40
30 20 10 0
51.22
Enrolled in program
QUESTION: Under what conditions can the difference of -5.05 be interpreted as the impact of the Balsakhi program?
The Program
RECALL: Balsakhi program targeted towards lowest performing children Hire local teachers (balsakhis, female)
Given them training to teach remedial classes in Hindi, Maths, English
Method 3: Difference-in-difference
Find a comparison group to evaluate impact: Group 1 Enrolled in Balsakhi program Group 2
vs.
Compare test score of these two groups at the start and at the end of the program. Takes into account pre-program differences in the two groups
Method 3: Difference-in-difference
Average test scores
60 50 40 30 20 10 0 51.22 36.67
56.27
24.8
Start of program
End of program
Method 3: Difference-in-difference
Average test scores
60
50 40 30 20
26.42
19.60
10
0
Start of program
End of program
Method 3: Difference-in-difference
Average test scores
60
50 40 30 20
26.42
6.82 19.60
10
0
Start of program
End of program
vs.
Compare test score of these two groups at the start and at the end of the program. Control for additional variables like gender, class-size
-5
-10 (1) Pre-post *
* Significant at 5% level
-5.05
(3) Difference-inDifference *
Impact Estimate
26.42* -5.05*
(3) Difference-in-Difference
(4) Regression
*: Statistically significant at the 5% level
6.82*
1.92
Which of these methods do you think is closest to the best estimate of the program impact?
84%
a)
c)
Impact Estimate
26.42* -5.05*
(3) Difference-in-Difference
(4) Regression (5) Randomized Experiment
*: Statistically significant at the 5% level
6.82*
1.92 5.87*
Conclusion
Measuring impact is all about finding a convincing representation of the counterfactual
Identify or create a group that mimics the counterfactual Always test your assumptions: What else might be causing the measured impact? Randomization is often the easiest way to create a reliable counterfactual