Estimating Treatment Effects from Matched Samples
What you'll learn: How to calculate the causal effect of a treatment after matching units to balance confounders.
Why Matching Isn't Enough
You've worked hard to create matched samples—pairing treated and control units that look similar on observed covariates. But matching is just the preparation. Now comes the payoff: estimating the actual treatment effect.
The Average Treatment Effect on the Treated (ATT)
The most common estimand from matched data is the ATT: the average effect of treatment for those who actually received treatment.
Think of it like this: You matched each treated patient with a similar untreated patient. Now subtract the control outcome from the treated outcome for each pair, then average those differences. That's your ATT.
Formula:
ATT = (1/N_treated) × Σ(Y_treated - Y_matched_control)
Where you sum over all treated units and their matched controls.
Interpretation
If your ATT = 5 (e.g., 5 points higher test scores), you conclude: "Among those who received the treatment, the treatment caused an average increase of 5 points compared to what would have happened without treatment."
Important nuances:
- ATT applies only to the treated group—it doesn't tell you what would happen if you treated the currently untreated
- Standard errors must account for matching (you can't treat matched pairs as independent)
- Check balance diagnostics first—if matching failed to balance covariates, your ATT estimate is biased
Practical Steps
- Verify balance after matching (previous lesson)
- Calculate outcome differences within matched sets
- Average those differences
- Compute appropriate standard errors (bootstrap or matched-pair methods)
Key Takeaway: The ATT is the difference in average outcomes between treated units and their matched controls, estimating the causal effect of treatment on those who were treated.