In Simple Words
Imagine tracking runners in two different high schools across two different semesters. To figure out if a mandatory training program in School A hurt the non-trained runners on the same track, comparing trained versus untrained runners in School A is insufficient. Untrained runners in School A must be compared to runners in School B across both semesters, controlling for overall weather changes across semesters without assuming how fast runners normally tire out.
Nonparticipants observed across Danish counties
- Counties without the trial
- Trend comparison group
- Counties with the activation trial
- Exit rate about 5% lower
Pre-trial-adjusted county comparison. Counties were not randomly selected; overall estimate was marginally significant, not a guaranteed 5% loss.
See source →Research question
Standard duration models evaluate job search assistance by comparing participants and non-participants within the same region, failing to identify treatment spillovers because non-participants' baseline hazard rates are contaminated by program externalities.
Participants
N = 89,466 unemployed benefit spells (40,403 experiment period, 49,063 pre-experiment period) across 14 Danish counties.
Study Design
Stratified partial likelihood proportional hazard estimation comparing non-participants in trial counties (Storstrøm and South Jutland) to comparison counties, stratifying duration dependence by calendar cohort.
Results
Participation in the activation program reduced non-participants' unemployment exit hazard rate by approximately 5% ($\gamma = -0.048$, p < 0.10 in full covariate model; $\gamma = -0.079$, p < 0.05 in South Jutland sub-sample).
Takeaway
Apply stratified partial likelihood proportional hazard models using non-experiment comparison regions to isolate non-treated spillover effects without imposing parametric hazard assumptions.