AI summary by TwinMind
Lecture 21: Endogeneity and Instrument Variables
MIT OpenCourseWare
1h 9m 21s
Introduction to Instrumental Variables
- Transition from machine learning back to causality and coefficient estimation
- Focus on establishing causal relationships when regression-based methods insufficient
- Problem: occasions where controlling for X2 variables, difference-in-differences, or regression discontinuity won't work
- Missing control variables
- No available regression discontinuity
- Endogeneity issues where X1 affects Y and Y affects X1
Endogeneity Problem and Examples
- TV and Unhappiness Study: correlation study making both causal claims
- TV causes unhappiness (passive activity, less beneficial activities)
- Unhappiness causes TV watching (low energy leads to TV consumption)
- Both channels plausible - mutual relationship problem
- Other Endogeneity Examples:
- State failure and government collapse
- Socioeconomic status and education
- Practice and skill improvement
- Democracy and economic growth
- Health and exercise
- Crosswords and cognitive decline
- Prices and quantity (supply/demand)
Education Returns Research Problem
- Research Question: causal effect of education on cognitive scores and wages
- Omitted Variable Bias: family background, ability, "spunk and drive," network connections
- Challenge: can't randomly assign education (too linked to personality/choice)
- Solution: randomly assign program that increases education likelihood
Potential Education Instruments Discussion
- After school programs
- Free school/meals
- Information campaigns about education benefits
- Technology access (laptops)
- Rewards for attendance/performance
- Scholarships (chosen as main example)
Ghana Secondary School Scholarship Study
- Setting: Ghana - free primary/junior high, expensive secondary school
- Design: 3,000 qualified but non-enrolled students, randomly selected 1,000 for scholarships
- Timeline: 2008 intervention, followed through 2012 graduation, interviews since 2013
- Tracking Success: 97% follow-up rate despite mobile adolescent population
First Stage Results
- Treatment vs Control Enrollment:
- Girls: 78% vs 45% ever enrolled (33 percentage point difference)
- Boys: higher take-up rates, 38 percentage point difference for completion
- Key Insight: scholarship doesn't translate one-to-one to school attendance
- Some scholarship recipients didn't attend
- Some control group found other funding sources
Cognitive Test Score Results
- Girls:
- Control group: -0.17 standard deviations
- Treatment group: ~0 standard deviations
- Difference: 0.18 standard deviations
- Boys:
- Control group: 0.18 standard deviations
- Treatment group: 0.30 standard deviations
- Difference: 0.125 standard deviations
Wald Estimator Derivation
- Structural Model: Y = α + βA + ε, where A is education, Y is outcome
- Key Assumption: E[ε|Z=1] - E[ε|Z=0] = 0 (no direct effect of instrument)
- Wald Estimate: β̂ = [E(Y|Z=1) - E(Y|Z=0)] / [E(A|Z=1) - E(A|Z=0)]
- Example Calculation: For girls, 0.18/0.33 ≈ 0.6 standard deviation increase in test scores per unit education
Instrumental Variable Validity Conditions
- First Stage: instrument must affect treatment (E[A|Z=1] ≠ E[A|Z=0])
- Easy to check empirically
- If zero, estimator goes to infinity
- Exclusion Restriction: instrument affects outcome only through treatment
- E[ε|Z=1] = E[ε|Z=0]
- Cannot be proven, must be argued
- Even with randomized instruments, this assumption not automatically satisfied
Threats to Validity Discussion
- Overconfidence Effect: scholarship recipients feel "chosen," affecting performance
- Income Effect: families save money, spend on nutrition/books
- Stress Reduction: less financial worry, more bandwidth for studies
- Sibling Effects: other children's education affected by family budget changes
- Survey Approach: extensive questionnaires to test alternative explanations
Instrument Evaluation Examples
- Flu Shot Reminder (Hens Study):
- Doctors randomly reminded to advise flu shots
- Problem: doctors may give additional health advice (wash hands, nutrition)
- Direct effects beyond vaccination
- Charter School Lottery:
- Generally good instrument
- Concerns: psychological effects of winning/losing lottery
- Sample limitation: only applicants, not general population
- Deworming Pills in Schools:
- Problem: externalities and contagion effects
- Spillover effects on untreated students in treatment schools
- Violates exclusion restriction
Problematic Instruments
- Laptops: multiple direct effects beyond education
- Study time changes
- Entertainment vs educational use
- Direct cognitive benefits
- Free Meals: nutrition effects independent of education
- Vitamins, iron improving alertness
- Direct health benefits affecting test scores
Heterogeneous Treatment Effects and LATE
- Problem: β may vary across individuals (βᵢ)
- Solution: Local Average Treatment Effect (LATE) interpretation
- Four Groups Framework:
- Always-takers: go to school regardless of instrument
- Never-takers: don't go to school even with instrument
- Compliers: respond to instrument
- Defiers: do opposite of instrument (assumed not to exist)
- LATE Result: IV estimates effect for compliers only
- Policy Relevance: compliers often policy-relevant group (those moved by cost reductions)
External Validity Considerations
- Sample Limitation: results apply to qualified students only
- Policy Implications:
- Good for making education free/cheaper
- May not apply to relaxing admission requirements
- Heterogeneity Analysis: examined effects across ability distribution
- Finding: consistent effects across qualification levels
