Do career and technical education tracks help students — and which students?

Verdict

moderate-support, and the support is narrow, male, and economic.

This is the one topic in the archive where test scores are the wrong outcome and the field knows it. The nine-site randomised trial moved earnings 17% for young men and moved educational attainment by exactly zero. Read that sentence twice, because it dissolves the usual mediation story: CTE's earnings effect does not run through credentials, and it does not run through tested achievement, both of which were measured and both of which were flat.

Four outcome classes, four different answers:

What the evidence shows

The single most important thing in this file is inside Silliman & Virtanen. In the same Finnish population, the same cohorts, the same outcome and the same years: the descriptive means show vocational admits overtaken by general admits at ages 27-28 and €4,000 a year behind by 33 — the Hanushek pattern, plainly visible. The regression discontinuity on the same people shows +6% at 33 with no downward trend over seventeen years. Same data, opposite sign. The only thing that changed is identification.

The lifecycle-decay claim, adjudicated. Hanushek et al. is the load-bearing paper and it has no exogenous variation in education type: its "difference-in-differences" differences across age within a cross-section, where age and cohort are collinear by construction. Hampf & Woessmann is a same-author extension to newer data, not a replication. Forster et al. is the only independent test of the mechanism and it fails it — no systematic variation by how occupationally specific a country's system is, which is Hanushek's central claim — while placing the crossover about a decade later. Meanwhile every design that generates exogenous variation kills or reverses the decay: Finland, Norway, Romania, Croatia. The honest verdict is that the decay finding is, on present evidence, a selection artefact — and that nobody has actually tested it, because the causal literature stops at 33 and the claimed crossover is at 43-50.

The gender null is the most replicated result in the field and it is almost certainly not about girls. Five independent designs find it. Connecticut supplies the mechanism: 73% of male technical-school students are in trades, 52% of female students are in cosmetology, hospitality and childcare. Norway confirms it at national scale, with men concentrated in higher-paying skilled trades and women almost exclusively in lower-paying service fields. This is a programme-mix problem, not a treatment-effect heterogeneity, and filing it as the latter has stopped anyone fixing it.

Both meta-analyses undercut the headline earnings story, and it matters how. The WWC-standards review could compute an earnings effect for only two studies (g = 0.164, not significant) — and it says plainly that several influential earnings studies were excluded because their reports lacked the information needed to compute an effect size. That is not a refutation of Connecticut or Career Academies; it is a demonstration that the computable American evidence base is nearly empty. The 39-study international meta is more damaging: against a general-education counterfactual, vocational education raises employment 4.8 points and moves wages not at all. Against no specific education it looks terrific (+11.7pp, +37.6%). The archive's baseline field is exactly this distinction, and most enthusiasm for CTE is quoting the second number for the first question.

Hereditarian-lens assessment

Risk: low for the verdict, high for half the literature, and the split is the most instructive thing here.

The verdict rests on lotteries, admissions cutoffs and birth-date discontinuities. A student one point above a cutoff and a student one point below have the same expected distribution of alleles; a Norwegian born on 31 December 1977 and one born on 1 January 1978 differ in policy eligibility and nothing else. Genes cannot explain those contrasts.

The entire lifecycle-decay cluster — Hanushek, Hampf & Woessmann, Forster et al., Golsteyn & Stenberg, and Kreisman & Stange — is high risk, and it is a textbook case for this archive's premise. These compare people who chose different tracks. Vocational interests are about 50% heritable in a twin design that includes pairs separated in infancy (Lykken 1993), and heritability rises with the level of aggregation exactly as it does for cognitive measures — so "who went to the vocational school" is substantially an expression of a heritable disposition, guided (as Lykken's own model has it) by heritable precursor traits of aptitude and personality through gene-environment correlation. Track choice correlates with literacy and maternal education in Hanushek's own appendix. And three separate regression discontinuities independently conclude that the observed vocational-general differential is selection.

Kreisman & Stange supply the tidiest demonstration. Their headline is that students positively sort into vocational coursework — the funnelling premise that motivates most detracking argument is not true in the NLSY97 — and their 2% wage premium per year of upper-level coursework should therefore be read as an upper bound on a confounded estimate, not as a return.

Reconciling with the archive's finding that early tracking harms low achievers

The archive grades early, rigid, between-school tracking as harmful to low achievers (grouping-tracking). A US CTE high school sorts children between schools at fourteen. Are these the same treatment? No, on four dimensions.

1. Age is most of the treatment. The harm evidence identifies the value of two extra comprehensive years against a German/Austrian sort at ten, made on teacher recommendations before achievement differences have stabilised. US CTE admission happens at fourteen on eighth-grade grades and test scores that actually exist. A four-year delay is much of what the harm literature is measuring.
2. Assignment versus application. European tracking assigns; US CTE high schools are oversubscribed and students apply. The estimand is a student who wanted in — the opposite selection direction from "shunted" — and Kreisman & Stange find positive selection into vocational coursework in national data.
3. Permeability is the operative variable, and the archive already says so. Its reading of the German evidence is that where tracks are permeable, marginal assignment has no long-run effect. That is what the CTE record shows: Connecticut's college cost fades to zero by 23, Finland finds no reduction in ever completing higher education, and Norway is the direct experiment — bolt a college-eligibility bridge onto the vocational track and you get more vocational enrolment, no loss of college-going, and +5% earnings for disadvantaged men. The European harm mechanism is foreclosure; US CTE high schools mostly do not foreclose.
4. They measure different outcome classes and agree where they overlap. The tracking topic's harm evidence is on test scores for low achievers. The CTE evidence is on attainment and earnings — and on test scores it agrees, finding roughly nothing.

Where there is genuine tension, record it. The archive treats three national detracking reforms (Sweden, Finland, France) as favourable evidence. But Žilić's Croatian reform — less tracking, more general content inside the vocational track — increased dropout and reduced male high-school and university completion. These are not contradictions; they are a different lever. The Nordic and French reforms removed between-school selection at 10-11. Žilić added academic content to a 16-19 vocational track. The second lever has a null-to-negative record, and it is the one a school founder is more likely to reach for.

One thing that should update the tracking topic when it is next revised. Norway's vocational expansion narrowed the class earnings gap by about 20% (30% among men) while widening the gender gap by 8%. The archive's equity framing for tracking is entirely about ability and class. Under current enrolment patterns, CTE expansion is class-progressive and gender-regressive simultaneously.

Boundaries & what critics say

• The estimand is not "CTE". Every regression-discontinuity estimate is the effect on a marginal applicant to an oversubscribed, selective, better-resourced technical high school — the best CTE in America, delivered through its least common model. Connecticut's own heterogeneity result shows the effect is 12% larger per SD fewer CTE offerings at the fallback school, so a substantial share of "the CTE effect" is a school effect. MDRC says the same about the randomised trial in plainer words: the model has proven challenging to implement at scale with high fidelity.
• Independence is not clean. Dougherty co-authors the Massachusetts RD, the Connecticut RD, and the systematic review that grades them. Not developer-led — he did not build the schools — but the field's evidence synthesis is not independent of the field's evidence.
• The American earnings case rests on two studies, both male-only, both single-state, both stopping in the mid-twenties. Everything else is graduation, employment or enrolment.
• The one US design whose academic signal is negative cannot currently be quoted. The Baltimore RD finds significantly lower GPA among CTE-centre admits, and it carries a published corrigendum that could not be read, so it is recorded unresolved and its numbers are not entered.
• Career Academies' applicant pool was positively selected — 50% of both arms earned a postsecondary credential against about 28% nationally — so the trial is efficacy on motivated applicants, not effectiveness at scale.
• Two of these results are on reformed vocational tracks, not vocational tracks. Norway's Reform 94 bundles more general content, a college bridge and more apprenticeships; P-TECH bundles a free associate's degree. Both are evidence about what has to be added before degrees and social mobility move.

Practical guidance

• Build depth, not electives. Two per cent per year of upper-level coursework and zero for introductory courses is the shape of the observational return, and every credible causal estimate is on a multi-year sequenced pathway inside a whole-school model.
• Fix the programme mix before you open. If the female pathways are cosmetology, hospitality and childcare, the evidence predicts precisely the zero that five studies have already found.
• Bolt a post-secondary bridge on. Ordinary CTE substitutes for four-year college and complements two-year college; it raises degrees only where the degree is architecturally built in (P-TECH: +5pp associate's, 13% versus 3% for men). Norway shows the bridge costs nothing in enrolment.
• Judge it on graduation, employment and earnings, and say so in advance. If you promise test scores you will lose, and the loss is avoidable — the randomised trial reported a flat zero on standardized maths and reading a quarter of a century ago.
• Do not quote the "vocational earnings decay after 45" finding as established. It comes from a design with no exogenous variation, its mechanism failed an independent test, and every discontinuity design contradicts it. Also do not quote its refutation as established: no causal study has reached the ages where the claim lives.

Open questions

• Nobody has measured CTE earnings past age 33 with a credible design, anywhere. The decay claim is located at 43-50. The causal literature stops a decade short of it. This is the single most valuable study the field could run and it requires only linking existing cutoff designs to older administrative earnings.
• Would the gender null close if girls enrolled in the trade pathways? No study has tested it, and the question is answerable — the enrolment mix varies across systems.
• What is the effect of ordinary CTE course-taking in a comprehensive high school? That is what most American CTE is, and there is no credible causal design on it. The best available source is explicit that students positively select into it.
• How much of the Connecticut effect is the school and how much is the curriculum? The authors' heterogeneity result implies a substantial school component, and nobody has decomposed it.
• The Baltimore RD needs re-reading against its corrigendum. It is the only US cutoff design whose academic signal points the wrong way, and its numbers are currently unusable.