Professional Documents
Culture Documents
J. Lecture 6 - Quasi-Experimental Methods
J. Lecture 6 - Quasi-Experimental Methods
Rema Hanna
Associate Professor of Public Policy
Evidence for Policy Design Harvard Kennedy School
August 29, 2012
Lecture Overview
1. Quasi-Experimental Methods
a) b) c) Regression Discontinuity Matching Difference-in-Differences
Strengths Threats Validity Tests
2.
Differences-in-Differences
a) b) c)
3.
Quasi-Experimental Methods
Quasi-experimental methods are alternative evaluation methods which:
Aim to find experiments naturally in data Are not as rigorous as randomization
Quasi-Experimental Methods
1. Regression Discontinuity 2. Matching 3. Difference-in-Differences
1. Regression Discontinuity
Program is assigned using a cutoff score and the participants and non-participants are compared controlling for the cutoff criteria.
Compare people who just met the eligibility criteria and those who were just excluded
Assumption 1: Discontinuity in program participation at some eligibility threshold Assumption 2: No discontinuities in any other factor that impacts the outcome of interest at that point
Regression Discontinuity
Obtains microcredit
Land size
Regression Discontinuity
Obtains microcredit
Land size
Can answer: Should the program cut-off be expanded at the margin? Cannot answer: Should the program exist?
2. Matching
Construct a comparison group that is similar on observable characteristics, especially those factors that could influence likelihood of participating in the intervention. Use available baseline data on participants and non-participants, for example census or household survey data.
Matching
Vote 2002 Campaign
Have background characteristics on intervention group Have data on 2,000,000 unaffected individuals
Matching
Matching
Estimation Method: Compare outcome for those who are similar
Strengths: 1. Equivalent on observable characteristics Challenges: 1. Unobservable characteristics 2. Requires lots of data (expensive)
3. Difference-in-Differences
Compare change in outcomes for intervention group and change in outcomes for comparison group
Start by reviewing:
Pre-post Simple difference
Review: Pre-Post
Compares what happened before and after intervention group received program
Assumption: Outcome variable would have remained constant in the absence of program. Source of Bias: We dont know what the change in outcome would have been without program.
Intervention Group
Time
Intervention Group
Time
Time
Intervention Group
Comparison Group Time
Difference-in-Differences
Combines Pre-Post and Simple Difference by comparing change in outcome over time between intervention and comparison groups.
Simple Difference
Biased because groups may be systematically different.
After Intervention Comparison Difference P2 C2 P2-C2 Before
Difference-in-Differences
Subtract out systematic differences using pre-intervention data: (P2 - C2) - (P1 - C1)
After Intervention Comparison Difference P2 C2 P2-C2 Before P1 C1 P1-C1
Pre-Post
Pre-post biased because do not know trend in outcome.
After Intervention Comparison P2 Before P1 Difference P2-P1
Difference-in-Differences
Subtract out trend from comparison group: (P2 - P1) - (C2 - C1)
After Intervention Comparison P2 C2 Before P1 C1 Difference P2-P1 C2-C1
Graphically:
Primary Outcome Before After Intervention Intervention P2
P1
C2
C1
Time
Strengths
Eliminates Common Shocks: Eliminates bias that can affect both program and comparison groups the same way, i.e. change in national financial regulation Removes Systematic Differences: Subtracts out fixed differences in characteristics between the program and comparison groups over time, i.e. if the program group always has a larger proportion of small business owners than comparison
Requires that, in the absence of the program, the outcome variable would look similar in both intervention and comparison groups over time.
Before Intervention
After Intervention
C1
P1
P2
Before Intervention
After Intervention
C1
P1
P2
Before Intervention
After Intervention
Mean Reversion
Primary Outcome C1 P1
P2 C2
Mean Reversion
Primary Outcome C1 P1
P2 C2
Validity Tests
Impossible to prove validity of the parallel trends assumption. However, several tests can be used to assess the validity of the assumption. 1. Check pre-intervention trends 2. Placebo intervention group 3. Placebo outcome
1. Pre-Program Trends
If outcomes of intervention and comparison groups moved in parallel before the program started it is more likely that they would have continued to move in parallel in the absence of the program.
1. Pre-Program Trends
Primary Outcome
Intervention period
2. Placebo Intervention
Choose a fake intervention group that was not affected by the program.
Compare two regions with no microfinance program
Difference-in-Difference estimate for this group should be close to zero or there is a problem.
3. Placebo Outcome
Choose an outcome that cannot logically be affected by the program.
Example: Did microfinance program affect rainfall?
What is the effect of a large-scale school construction program on years of schooling obtained?
Between 1973 and 1978, the Indonesian government built more than 61,000 primary schools.
Source: Duflo, E. (2001). Schooling & Labor Market Consequences of School Construction in Indonesia: Evidence from an Unusual Policy Experiment.
Program Theory
Could improve educational outcomes by increasing the supply of education. Could also reduce quality of education because new teachers are not as well trained, canceling out supply effect.
Evaluation Method #1
Compare average education levels of 18year-olds in 1973 and 1990. What is the problem with this evaluation strategy?
Evaluation Method #2
Compare average education levels of 18year-olds in 1990. across areas where the program was more intensive and less intensive. What is the problem with this evaluation strategy?
Difference-in-Differences
More schools were built in regions with the fewest schools per capita initially. Children over age of 12 when program began were not affected. Compare the change in educational outcomes between younger and older children in low program intensity regions with the change in outcomes in high program intensity regions.
Results:
Estimated impact: D-in-D estimate of gain of 0.12 years of education.
Years of education: age 2-6 in 1974 High program intensity Low program intensity Difference 8.50 9.76 -1.26 Years of education: age 12-17 in 1974 8.02 9.40 -1.38
Difference
Difference
Key Takeaways
Difference-in-Differences can remove some bias found in Pre-Post and Simple Differences methods. D-in-D does not fully solve the problem of selection bias, so one must still be cautious D-in-D often requires more data than randomization (may be more expensive unless administrative data is available)