Dynamic Selection Bias in Duration Models
Expert TheoryReview
In randomized experiments tracking time-to-event data over time, initial randomization balances unobserved worker characteristics at baseline. However, as individuals exit unemployment at different rates, the remaining treatment and control samples become compositional unequivalent over time.
Picture this
Imagine two identical lines of runners where one line receives an incentive that causes the fastest runners to leave early. As time passes, the remaining runners in the incentivized line consist disproportionately of slower runners, falsely making the treatment group appear less capable than the control group.
What the evidence says
Uncorrected raw Kaplan-Meier hazard rates provide only a lower bound on treatment efficacy due to dynamic selection, whereas MPH duration models correctly isolate true 20–40% exit rate increases.
- Who was studied
- N = 4,513 unemployed individuals across two Danish counties tracked over weekly intervals.
- How
- Mixed Proportional Hazard (MPH) duration modeling specifying conditional unobserved heterogeneity distributions.
What to do
Apply mixed proportional hazard duration models that explicitly incorporate unobserved heterogeneity parameters when evaluating dynamic social experiment panel data.
From the source
"once the experiment starts affecting the exit rates from unemployment, the distribution of unobservables among the survivors in unemployment will start to differ between the treatment and control groups."
bc5b372e-5470-4162-9864-568cdcec4a00-Experimental Evidence on the Nature of the Danish Employment Miracle.pdf