The effect of sample selection and initial conditions in duration models: Evidence from experimental data on training

We investigate the separate effects of a training program on the duration of participants' subsequent employment and unemployment spells. This program randomly assigned volunteers to treatment and control groups. However, the treatments and controls experiencing subsequent employment and unemployment spells are not generally random (or comparable) subsets of the initial groups because the sorting process into subsequent spells is very different for the two groups. Standard practice in duration models ignores this sorting process, leading to a sample selection problem and misleading estimates of the training effects. We propose an estimator that addresses this problem and find that the program studied, the National Supported Work Demonstration, raised trainees' employment rates solely by lengthening their employment durations.