Labor Market Returns to Vocational Secondary Education

21,591 individuals in the regression-discontinuity estimation sample; about 90% (19,932) had ranked the general track first

The universe of Finnish students graduating compulsory education 1996-2000 who applied to both vocational and general upper-secondary tracks, followed 17 years to age 33 in population registers.

Fuzzy RD at centralised admissions cutoffs, optimal bandwidth with fixed-bandwidth robustness from 0.1 to 2.0, first stage about 50 percentage points, clean balance on more than 30 covariates including parental income and education and subject-level GPA, passes McCrary, robust to donut estimates. This is the single most important source in the topic because it contains the decay debate inside one dataset. Limits: the compliers sit mid-distribution and only 30% complete higher education, so this is not the effect for an able student; the treatment bundles curriculum with peer group and within-school rank; the sample is dual applicants of whom 90% preferred general, so the headline estimate is dominated by people pushed INTO vocational against preference, which if anything biases toward zero; and the analysis stops at 33, so the claim that lifetime returns never turn negative is a discounted-value extrapolation rather than a measurement.

Admission to the vocational track raised annual income by about €1,800 at age 33, seventeen years after admission — roughly 6% on a €29,000 counterfactual — with no downward trend across the seventeen-year window and no effect on months employed. It did NOT reduce the likelihood of ever completing higher education for the marginal applicant, and did NOT raise exposure to automation- or offshoring-prone occupations. The decisive fact for the decay debate: in the SAME Finnish population's raw means the Hanushek pattern is plainly visible — vocational admits are overtaken by general admits 11-12 years after admission, at ages 27-28, and earn about €4,000 LESS per year by 33. The RD reverses the sign of the descriptive pattern. For applicants who ranked vocational first, admission raised employment by nearly two months a year (being pushed into general cut their employment about 20%) and raised income about 25% at the margin.

Low for the causal estimate — admissions-cutoff assignment among applicants, with clean balance on parental income and education. The paper is simultaneously the archive's cleanest illustration of how large the selection confound is in this literature: the descriptive and causal estimates have opposite signs on the same people.

The no-decay conclusion is independently corroborated by Bertrand, Mogstad & Mountjoy (Norway) and by the null-difference RDs of Malamud & Pop-Eleches (Romania) and Žilić (Croatia). It directly contradicts Hanushek et al. (2017) and Hampf & Woessmann (2017) — but those are non-causal designs, so the disagreement is not symmetric.