Summary
- Causal/treatment effect is defined in terms of potential outcomes
- Same unit, same moment in time post-treatment
- Fundamental problem: Only one potential outcome is observable!
- Problem of causal inference is a missing data problem
- Conceptualize potential outcomes as having trajectories over time
- Estimation requires filling in missing outcome(s)
- Scientific solutions:
- Estimate individual causal (treatment) effects
- Compare with different unit or earlier/later outcome of same unit
- Assumptions: unit homogeneity, temporal stability, causal transience
- Statistical solution:
- e.g., randomization
- Estimate average causal (treatment) effect
- Compare averages between treatment and control groups
- Assumptions: IA, SUTVA
- Framework extendable to whatever units (countries, towns, physial objects) and treatments with several levels
- Conceptualize treatment variable as an intervention/action
- …makes alternative values and timing clearer! (e.g. “going to college”)