The Problem of Selection Bias
To measure the true impact of an action, we need to compare groups that are identical in everything except for that action. But in real life, the group that chooses (or is chosen) to receive a treatment is different from the start.
The Solution: Propensity Score Matching (PSM)
PSM allows us to build an artificial control group that resembles the treated group as closely as possible, enabling a fair comparison.
The Matching Process
1. Estimate the Score
Compute the probability of receiving the treatment.
2. Common Support
Discard individuals without comparable counterparts.
3. Match
Find the most similar “twin” for each treated unit.
4. Check Balance
Verify that the groups are now similar.
5. Estimate the Effect
Compute the difference in the final outcome.
Simulation
Balance and Results Analysis
Covariate Balance (SMD)
Final Results
--
Average Treatment Effect on the Treated (ATT)
-- / --
Final / Initial Sample