Do life-skills courses — money, cooking, driving, health — change what students actually do?

Verdict

mixed, and the mixture has a shape. Four different life-skills subjects have been tested with real designs, and they line up on a single axis: how far in the future is the behaviour you are trying to change?

• Knowledge always moves. Financial education raises financial knowledge by 0.204 SD across 76 RCTs and 0.25 SD in schools — comparable to what a decent maths or reading intervention buys. Abstinence programmes improved students' ability to identify STDs. Schools can teach this content.
• Behaviour close to the instruction sometimes moves. State personal-finance graduation requirements shift college financing from private to federal loans, cut credit-card balances, and are associated with higher credit scores and sharply lower delinquency in the first years of adult credit life.
• Behaviour far from the instruction does not move. The same kind of mandate, studied on cohorts who graduated between 1957 and 1982 and observed decades later, produces nothing on any financial outcome. Four abstinence programmes of 50+ contact hours, randomised, followed four to six years, produced identical sexual behaviour in both arms on every measure.
• One subject is worse than nothing. Driver education gets teenagers licensed sooner and does not make them safer.

What the evidence shows

Driver education is the strongest and most uncomfortable finding in this file. Three randomised trials, 17,965 young people, three continents, official licensing and police crash records rather than self-report — this is a better evidence base than most things schools do, and it says the course works only on the outcome nobody wanted. The DeKalb County trial alone randomised 16,338 students with concealed central allocation. Crash relative risks were 1.01, 1.03 and 1.10: every point estimate above one. The reviewers' arithmetic is worth quoting in a budget meeting — if a schools programme raised the proportion of licensed teenagers by just 2%, an additional 27 British teenagers might be killed or seriously injured each year because of the road-safety programme.

Note what the Cochrane authors refused to do, because it is a methodological lesson that generalises across this whole topic: they excluded driving skill as an outcome, on the grounds that treating improved skill as a proxy for fewer crashes assumes the very relationship under test. That is the same move as separating financial knowledge from financial behaviour, and it is the move that dissolves most of the optimism here.

The financial-literacy dispute is real and it is not settled by picking a side. Fernandes et al. found that interventions explain 0.1% of the variance in the behaviours they target, and that number became a stylised fact. Kaiser et al. re-ran the question on 76 randomised experiments — five times as many as Fernandes had — and got 0.100 SD on behaviour, some 3.6 to 5.5 times the original estimate, robust to publication-bias correction, to restriction to top journals, and to restriction to adequately powered studies. They also document four coding errors in the original RCT classification.

Three things stop that from being a clean win for financial education, and a school builder needs all three:

1. It is not a school literature. Only 7.5% of the estimates concern children under 14 and 72.4% concern adults over 25; 60.8% come from developing economies. The school-age cell is small.
2. The school-age behavioural effect is the smallest one in the paper — 0.064 SD for under-14s, against 0.1203 for 14-25 year-olds. Whatever financial education does, it does it least to the children schools most want to reach early.
3. The dose is tiny. Mean 11.71 hours, median 7. So 0.10 SD is what seven hours buys — which is genuinely good value at $22.90 median cost per participant, and is also not what a year-long graduation requirement is.

The two halves reconcile through timing, and that is the topic's organising fact. Fernandes's decay meta-regression finds effects falling with delay and no significant effect of even 24 hours of instruction at 18.5 months or more. Kaiser et al. dispute the steepness (0.0574 SD at two-plus years, not statistically distinguishable from the short-run estimate) but their long-horizon cell is seven studies. And the two natural-experiment literatures split exactly along this line: mandates observed on 18-22 year-olds show effects; the same mandates observed on people in their thirties and forties show none.

The Cole, Paulson & Shastry result should change what a school actually schedules. They killed the founding result of the field — Bernheim, Garrett & Maki's saving effect does not survive the inclusion of state fixed effects — and then found that the curriculum requirement which did move adult financial outcomes was an extra mathematics course: $1,500-3,000 more home equity on a $15,500 base, 0.4pp less foreclosure on a 9% base, surviving controls for income, education and occupation. If the school day is zero-sum, that is the comparison that matters, and it does not favour the life-skills course.

Hereditarian-lens assessment

Risk: low for the verdict, which rests on randomised trials (driver education, abstinence education, cooking) and on cohort-level policy mandates that no family chose.

But the lens explains why this field looked promising for twenty years before the trials arrived. Fernandes's central contrast is a demonstration of exactly the confound this archive's premise predicts: measured financial literacy correlates with good financial behaviour at r ≈ 0.13 (1.8% of variance), while manipulated financial literacy moves it by r ≈ 0.03 (0.1%). The gap is not a puzzle. Financial literacy scores proxy numeracy, propensity to plan, confidence and risk tolerance — all substantially heritable — and the paper shows the partial effects of literacy collapsing when those traits are controlled or when literacy is instrumented. Within the measured-literacy studies, instrumental-variable estimates run smaller than OLS on the same data (0.059 versus 0.091 where the instruments are non-weak). The correlation between financially literate adults and financially successful adults is largely a fact about people, not a lever schools can pull.

The same pattern is visible in the cooking literature — reviews of school cooking classes report small positive effects on self-efficacy and vegetable intake from a study set that is roughly three-quarters uncontrolled, and the one cluster-randomised test with validated instruments and diet recall returns four negative point estimates.

Boundaries & what critics say

• The steelman for financial education is Kaiser et al. 2022, and it is a good one: RCTs only, robust to five sensitivity analyses, and cheap per head. The reason it is not decisive here is the population and the dose, not the design.
• Independence is genuinely murky on both sides. Fernandes et al. was funded by the National Endowment for Financial Education — an advocacy funder that paid for a null, which strengthens it. Kaiser et al. are the field's principal proponents meta-analysing a literature that includes their own trials; they check robustness to excluding their own papers, which is the right check, but no independent team has re-coded either side of this dispute. Both halves of the most consequential disagreement in the topic have been adjudicated by one of the two disputants.
• The mandate estimates deserve more scepticism than they usually get. Three treated states, synthetic controls, and outcome years 2008-2012 — which is to say, the mandates took effect in 2007 and the credit outcomes are measured through the financial crisis. Quarter-by-year fixed effects absorb the national shock but not Georgia's housing bust. The effect grows monotonically with each cohort in every state, which the authors read as teachers learning the curriculum and which a differential post-treatment trend would reproduce exactly. Twenty-eight credit-score points is a large number for seven hours of teaching.
• Driver education has changed since 1984. The three trials are from 1982-84 and tested intensive courses with behind-the-wheel training. Contemporary school programmes are often a single 50-minute presentation. The Cochrane reviewers argue this is worse, not better: it could still pull licensing forward while touching driving skill even less. That argument is untested, and it is the main live objection to reading this evidence onto today's programmes.
• Graduated driver licensing is a different intervention and is not graded here. It is a licensing policy, not a course, and the driver-education evidence is an argument for it rather than against road safety.
• Abstinence education is not all of sex education. Trenholm tested four programmes bound by a statutory eight-point abstinence-only content definition; it is not a test of comprehensive sex education, and two of the four control groups had access to rich alternative school health services, making it a test of the marginal effect over what schools already provided.
• Nothing here says the skills are worthless. Knowing how to cook, how to read a loan agreement and how to drive are worth having. The claim under test is that a school course is how you get them, and that the course changes what students do.

Practical guidance

• Put personal finance in the last year or two of school, keep it short, and tie it to a decision the student is about to make — student loans, a first credit card, a first pay cheque. Seven to twelve hours is the evidenced dose; $25-60 per student is the evidenced cost. That is a defensible purchase.
• Do not run driver education as a school programme. It is the one item in this archive where the best available randomised evidence points at net harm, and the mechanism — earlier licensing without better driving — is well identified.
• Judge every life-skills course on a behavioural outcome, never on a knowledge post-test. The knowledge post-test will pass. It passed for abstinence education, whose students got better at naming STDs and behaved identically to controls for six years.
• Ask when the behaviour happens before you buy the course. If the answer is "in fifteen years", the evidence says buy something else.
• If it competes with a mathematics course, teach the mathematics. That is not a rhetorical flourish; it is the finding of the one study that estimated both from the same policy variation.

Open questions

• Nobody has randomised a full-year, high-school personal-finance graduation requirement. The RCT literature averages seven hours; the mandate literature has no randomisation. The intervention schools actually run has never been tested with the design that would settle it.
• The whole mandate result rests on cohorts observed at 18-22. Whether the credit-score advantage persists into the thirties is exactly what Cole et al.'s null suggests it does not, and no study follows the post-2007 mandate cohorts far enough to check.
• No modern randomised trial of driver education exists, and the reviewers' argument that today's lighter programmes would be worse rather than better has never been tested.
• Home economics is essentially unevaluated as an academic subject. One cluster RCT of a food-literacy programme in 26 primary schools is the entire causal record found here, and it was null on its own knowledge outcome.
• First aid and CPR instruction were not reached in this pass and are a live gap — they are the one life skill whose target behaviour is both rare and immediately consequential, which is a different problem shape from everything graded above.