AI summary by TwinMind
Lecture 14: Causality
MIT OpenCourseWare
1h 15m 57s
Introduction to Causality
- Three types of causal statements examined:
- Headache got better because she took pill (simple counterfactual)
- Got good job because went to MIT (fuzzy counterfactual - alternative unclear)
- Cannot get job interview because African-American (most complex - manipulation of race unclear)
- Causal statements involve counterfactuals - comparing what happens vs what would have happened
- Third example has multiple interpretations:
- Educational/background effects from growing up in certain circumstances
- Direct discrimination at interview level despite qualifications
Importance of Causal vs Non-Causal Questions
- Many economics/social science questions are causal:
- Does immigration lower native worker wages?
- Would trade increase inequality?
- Would wall between Mexico/US stop immigration?
- Non-causal questions also important:
- Early warning signs for at-risk children
- Age-grade mismatch predicting ADHD diagnoses (may be correlation, not causation)
- Business applications: Google ad targeting based on search patterns
- Distinction between research vs business applications:
- Research seeks generalizable insights
- Business may only need practical predictions
Directed Acyclic Graphs (DAGs)
- Proposed by Judea Pearl in "Causality" (2000)
- Graphical representation of causal relationships
- Example: Season → Sprinkler/Rain → Wet Pavement → Slippery
- Components:
- Nodes (variables)
- Directed arrows (causal relationships)
- Acyclic (no loops back to original variable)
- Terminology:
- Parents/children relationships between nodes
- Chains, forks, colliders
- Markovian parents
- "do" notation represents interventions vs conditional probabilities
- Warning: DAGs helpful for thinking but not substitute for careful causal reasoning
- Risk of "miracle happens" - writing DAGs doesn't automatically solve identification
Regression Framework for Causality
- Basic model: Y = βX + ε
- Often interpreted causally as ideal experiment
- With controls: Y = βT + γX + ε
- T = treatment variable of interest
- X = control variables (covariates)
- Example: Sprinkler effect on pavement wetness controlling for season
- Without controls, could get backwards result due to confounding
- Regression coefficients represent causal effects under assumptions
Rubin Causal Model (Potential Outcomes Framework)
- Due to Donald Rubin
- Each unit has potential outcomes for each treatment state
- Headache example: 4 possible types of people
- Aspirin helps (Y₁=0, Y₀=1)
- No effect either way (Y₁=Y₀)
- Aspirin hurts (Y₁=1, Y₀=0)
- Causal effect = Y₁ - Y₀ for individual
- Fundamental problem: only observe one potential outcome per person
- Need multiple units and assumptions about assignment mechanism
SUTVA (Stable Unit Treatment Value Assumption)
- Two parts:
- Potential outcomes don't vary with treatment of other units
- No different versions of same treatment
- Examples where SUTVA fails:
- Vaccination (herd immunity effects)
- Deworming medicine (transmission between children)
- Heating in apartment building
- Job training (limited positions available)
- Solution: Redefine units (e.g., communities instead of individuals)
Selection Bias and Randomization
- Observed difference = Treatment effect + Selection bias
- Selection bias = E[Y₀|W=1] - E[Y₀|W=0]
- Headache example: People with worse headaches take pills
- Randomization eliminates selection bias
- E[Y₀|W=1] = E[Y₀|W=0] under randomization
Types of Randomized Experiments
- Completely randomized: Simple random assignment
- Stratified randomization: Randomize within blocks/strata
- Paired randomization: Match similar units, randomize within pairs
- Cluster randomization: Randomize groups rather than individuals
- All types eliminate selection bias
- Stratification improves precision by ensuring balance on observables
Empirical Example: Traffic Policies Using Google Maps API
- Gabriel's research on congestion policies in Delhi and Jakarta
- Google Maps API provides travel time predictions
- Delhi odd-even policy (Jan 1-15, 2016):
- Cars with odd license plates drive odd days only
- Even plates drive even days only
- Method: Query 93 routes every 20 minutes
- Results: Clear reduction in travel time during policy period
- Jakarta "three-in-one" policy: Require 3+ passengers during peak hours
- Shows immediate effects when policy starts/stops
