Using the Longitudinal Structure of Earnings to Estimate the Effect of Training Programs
Causal Question / Estimand
The effect of CETA (1976) job-training participation on trainees’ earnings — the average effect of treatment on the treated (Causal-Estimand). The paper’s deeper target is how sensitive such an estimate is to the model of how people select into training.
Identification Strategy
Exploits the longitudinal structure of earnings: fit a components-of-variance model to a comparison group’s earnings, then use a participation model to predict what trainees would have earned absent training, yielding both an effect estimate and an overidentification test. This is the fixed-effects / pre-post-with-comparison logic that underpins difference-in-differences (Parallel-Trends).
Key Assumptions
Parallel-Trends (absent training, trainee earnings would track the comparison group’s), a correctly specified participation/selection model, and SUTVA. The paper’s contribution is to show how badly results move when the selection model is varied.
Threats to Validity
The famous Ashenfelter-Dip: trainees’ earnings dip in the year(s) just before enrollment, so a naive pre-period baseline overstates the counterfactual and biases DiD estimates. Because the dip is driven by transitory shocks that also drive selection, Parallel-Trends fails. Estimates range from $200 to $2000 depending on the participation model.
Setting / Data
Longitudinal Social Security earnings records for participants in 1976 CETA training programs and a comparison group (CPS-based), spanning multiple pre- and post-program years.
Key Claims
- Pre-program earnings of trainees systematically dip (the Ashenfelter dip), contaminating before/after and DiD comparisons.
- Program effect estimates are highly sensitive to the assumed participation model — an order-of-magnitude range.
- Because no observational selection model is clearly right, randomized trials are needed to pin down training effects reliably.
Connections
- See also: Meyer1995-NaturalAndQuasiExperiments (mean reversion as a DiD threat), Abadie2005-SemiparametricDiD (conditioning on covariates to handle non-parallel dynamics), DiD
- Cited by the credibility-revolution surveys: AngristKrueger1999-EmpiricalStrategiesLaborEconomics
Citation
Ashenfelter, O., & Card, D. (1985). Using the Longitudinal Structure of Earnings to Estimate the Effect of Training Programs. The Review of Economics and Statistics, 67(4), 648–660. https://doi.org/10.2307/1924810