190 likes | 208 Views
Causal Inference. David Evans Economist Africa Impact Evaluation Initiative. Goal. Evaluate the causal impact of a program or an intervention on some outcome How much did X move Y? How much did the new teacher contracts affect students performance? Not the same as correlation !
E N D
Causal Inference David Evans Economist Africa Impact Evaluation Initiative
Goal • Evaluate the causal impact of a program or an intervention on some outcome • How much did X move Y? • How much did the new teacher contracts affect students performance? • Not the same as correlation! • X and Y move together • Holding an umbrella & having wet shoes • Does the umbrella cause wet shoes? • Age and income
Evaluation problem • Compare same individual with & without a program at the same point in time • BUT Never observe same individual with and without program at same point in time
Solving the evaluation problem • Counterfactual: what would have happened without the program
A little more rigorously • Estimate counterfactual • find a control or comparison group • Counterfactual Criteria • Treated & counterfactual groups have identical initial average characteristics • Only reason for the difference in outcomes is due to the intervention Mick Pick B- A+
“Counterfeit” Counterfactual #1 • Before and after: • Same individual before the treatment • But what else is happening over time? Before the books
Before and After: Food aid • Compare income before and after • Income after food aid is lower than income before food aid • Did the program fail? • Received aid because of income fall • Cannot separate (identify) effect of food aid from the shock that led to needing it
Before & After:Cash transfers Income • Compare income before & after intervention Before & after counterfactual = B Estimated impact = B-A • Control for time varying factors True counterfactual = C True impact = A-C B-A is gives a false negative! Before After B Negative impact!! B A C Positive impact!! t-1 t Time Treatment (Transfers)
“Counterfeit” Counterfactual #2 • Compare participants to non-participants: • Those who choose not to enroll in program, or • Those who were not offered the program • Problem: We can not determine why Mick participates in the program and Don does not • Why didn’t Don get books?
Why participants and non-participants might differ? • People who claim unemployment benefits and people who do not? • People who go to the doctor and people who do not? The unemployed The sick
Health program example • Treatment offered • Who signs up? • Those who are sick • Areas with epidemics • Have lower health status than those who do not sign up • Healthy people/communities are a poor estimate of counterfactual
What's wrong? • Selection bias: People choose to participate for specific reasons • Many times reasons are directly related to the outcome of interest Mick gets better marks with books B- A+ But what if Mick’s received books because he was already getting better marks A+ B-
HIV: Effect of VCT on risk behaviors • What what if Mick went for VCT because he knows he engages in high-risk behaviors? (Even after, still higher than Don!) • Cannot separately identify impact of the program from these other factors/reasons No VCT VCT High risk behaviors Low risk behaviors
Possible Solutions… • Guarantee comparability of treatment and control groups • ONLY remaining difference is intervention • How? • Experimental design/randomization • Non-experimental (difference in difference, matching, regression discontinuity, instrumental variables, etc)
These solutions all involve… • EITHER Randomization • Give everyone an equal chance of being in control or treatment groups • Ensures that participants and non-participants will be similar on most characteristics* • Only difference is the intervention • OR Transparent & observable criteria for assignment into the program
Finding controls: opportunities • Budget constraints: • Eligible who get it = potential treatments • Eligible who do not = potential controls • Roll-out capacity: • Those who go first = potential treatments • Those who go later = potential controls
Finding controls: ethical considerations • No need to delay benefits if rollout based on budget/capacity constraints • Equity: equally deserving populations deserve an equal chance of going first • Transparent & accountable: If based on criteria, then criteria should be measurable and public • Ethics binds after you know it works
Conclusions • Seek to measure the causal effect of a program on some outcome • Need a counterfactual • Shun the counterfeits! • Before and after • Participants vs non-participants • Find a good control group • Would be the same in the absence of the intervention • Options • Randomize (coming soon!) • Use other methods with more assumptions (coming tomorrow!) THANK YOU