Does civic education produce civic knowledge, and does civic knowledge produce civic behaviour?

Verdict

Civics teaches civics. It does not make citizens. That sentence would be an unfair caricature if the dissociation were merely unproven — but it has been tested directly, under randomisation, and it failed.

Green et al. randomised 59 classrooms to a 21-lesson constitutional-principles curriculum. Knowledge of civil liberties moved d = 0.49. Support for civil liberties moved −0.071 (SE 0.051) — null in every wave, on the index and item by item. General political knowledge moved 0.023 (SE 0.076): no transfer. And two years later the knowledge gain itself had decayed to +0.221 (SE 0.269), not significant.

Narrow knowledge, temporarily. That is the product.

Meanwhile the outcomes people actually want — registering, voting, volunteering, participating — do move for some children, and the things that move them are not civics curricula. Two charter admission lotteries raised voter-file-validated turnout by 6 and 12 percentage points, and a six-year childhood psychosocial-skills programme with no political content at all raised ever-voted by 7.3 points in the late twenties.

Hence mixed, with the two halves kept strictly apart, because collapsing them is the error this topic exists to prevent.

What the evidence shows

outcome_class discipline is doing real work in this table. Knowledge is domain-skill. Voting, volunteering and participating are behaviour. Tolerance, efficacy and interest are non-cognitive. These do not move together.

The knowledge-to-behaviour link failed the conservative test, not the demanding one. Green et al. deliberately report the reduced form — assignment to attitudes — because it credits to the knowledge channel any effect the discussion-heavy curriculum might have had directly. It still returned null, in all three waves, with ample power.

Open classroom climate and deliberation — the field's most-cited positive finding — collapse under randomisation. Persson held content, texts and time constant, confirmed with IEA CivEd items that treated students genuinely did perceive a more open climate, and then moved none of knowledge, interest, values or discussion. The correlational literature that made "how you teach it matters more than how much" a truism has not survived the one experiment that tested it.

What moves validated turnout is school quality and noncognitive skill. Three independent designs converge. Cohodes tested five mechanisms and found the noncognitive index tracking the voting effect while the cognitive index did not — and, decisively for this topic, charters whose missions emphasise civics were not significantly better on voting (0.075 vs 0.034, p = 0.391) even though they raised AP US History and Government taking by 0.159 SD. More civics coursework, same voting. Holbein's Fast Track had no political content whatsoever and moved turnout by 7.3 points a decade and a half later.

"More schooling causes voting" is an American, self-reported result. Milligan's own UK arm is a precise null. Norway is null. Sweden's 417,000 register-validated voters are null on average. The US draft-lottery instrument on 247,876 men returns a non-significant estimate that flips sign on wider cohorts. And in both US positives, the IV estimate is statistically indistinguishable from OLS — which is what you see when the instrument is not doing any work.

The one pattern that recurs across every design is compensation. Lindgren's bottom SES quartile gains 11.62 points against a null average. Campbell & Niemi's knowledge gains appear for Latinos and immigrants. Holbein's and Cohodes' effects are largest where baseline participation is lowest. Civic education, where it does anything, does it for the children who arrive with least.

Hereditarian-lens assessment

Risk: low for the verdict, which rests on two admission lotteries and three cluster RCTs. But this is a topic where the premise has been measured rather than assumed, and the result is worth stating plainly.

Dinesen et al. compare monozygotic twins discordant in education. Genes and shared family environment turn out to substantially confound the raw education–participation gradient — the correlation that the entire observational civic-education literature is built on. A within-pair effect survives in the US and Denmark and is null in Sweden. That is exactly what this archive's premise predicts: an input correlated with a heritable trait, producing a gradient that shrinks and sometimes vanishes when the genotype is held constant.

Two further notes:

• The self-report problem compounds the genetic one. Milligan and Dee both measure self-reported voting. People who are more educated over-report voting more. The studies with validated outcomes — voter files, digitised election rolls — are the ones that return nulls (Sweden, Norway) or that credit something other than civics (the charter lotteries).
• Barr's +0.12 prosocial composite is a researcher-designed self-report in a developer-involved trial with no behavioural outcome measured. It is the best evidence any named civics programme has, and under this archive's rules it does not support a behavioural claim.
• Nothing in this topic bears on g, and nothing claims to.

Boundaries & what critics say

• Mandated service may backfire. Maryland's requirement raised 8th-grade service and later 12th-grade cohorts did less voluntary activity; sibling-fixed-effects analysis finds service-learning coursework null on adult turnout. The one supportive study measured intentions, in one school.
• Civics-test mandates are the cheapest possible policy and they do nothing measurable to turnout — at most 1.5 points, not significant, with a possible −9.7 points among Black youth that deserves attention rather than dismissal.
• Trickle-up to parents did not survive randomisation. Gill's 5,792-parent analysis is null despite a large observational Kids Voting literature claiming otherwise.
• The turnout effects are narrower than they sound. Cohodes' effect appears only in presidential elections, only for women (+11 pp vs ~0 for men), does not move registration (88% of controls were already registered), and has converged by the third election. It is an effect on when people start voting more than on whether.
• The behavioural evidence is all adult outcomes of childhood treatment, measured at 18-29. Nothing here tells you what a civics course does to a 16-year-old this year.
• A defender of civic education would say knowledge is the point — that an informed citizenry is a goal in itself and demanding turnout effects is a category error. That is a legitimate position, and it has to contend with the fact that the knowledge decayed to the control mean within two years and never transferred beyond the taught content.
• The strongest study is a working paper in transit. Cohodes' numbers come from the NBER working paper while the DOI points at the journal version of record; Jung & Gopalan was read as an EdWorkingPaper. Both are noted in their source files.

Practical guidance

• Teach civics for the knowledge, and expect narrow, fading knowledge. d ≈ 0.5 on the taught content, no transfer, gone in two years. That is a reason to teach it repeatedly and cumulatively, not a reason to abandon it — but it is not a reason to expect anything else.
• Do not stake the programme on attitudes, tolerance, or turnout. All three have been tested under randomisation and all three failed. A civics programme sold on "producing engaged citizens" is falsifiable on a claim the evidence has already falsified.
• Do not buy the open-classroom / deliberation package on the strength of the correlational literature. The controlled test held everything constant, confirmed the climate changed, and moved nothing.
• Do not mandate community service expecting habit formation. The best available evidence suggests it crowds volunteering out.
• If participation is the goal, teach the procedure. Registration deadlines, how to find a polling place, what happens when you get there. Actionable "here is how to vote" instruction is the one civics-adjacent thing with a positive signal, and the charter lottery's own mechanism analysis points at noncognitive follow-through rather than political knowledge.
• Concentrate effort where baseline participation is lowest. Every positive result in this topic — Sweden's bottom quartile, Latino and immigrant knowledge gains, Fast Track's 26% base rate, the charter lotteries' 35% counterfactual — is a compensation effect.
• Do not run a school on the theory that a civics mission produces voters. The lottery that could test that found civics-mission charters no better than the rest, despite teaching demonstrably more civics.

Open questions

• Why do good schools produce voters? The mechanism analysis points at noncognitive skill and follow-through rather than knowledge, but "noncognitive index" is a proxy, and nobody has isolated what inside a high-performing school does this.
• Why does the effect appear for women and not men, and why only in presidential elections? Both asymmetries are unexplained and both bound the claim severely.
• Is the American exception real? Precise nulls in the UK, Norway and Sweden against large US self-reported effects could be institutions (registration burden), measurement (self-report versus register), or publication. Nobody has decomposed it.
• What is the −9.7 pp on Black youth after civics-test mandates? It is marginally significant, it is the wrong sign, and it has not been investigated.
• Does cumulative civics teaching behave differently from a six-week unit? The knowledge decayed within two years in the one trial with a follow-up; whether repeated annual exposure compounds is untested.
• Nothing measures civic behaviour before adulthood. Every behavioural outcome here waits until the student can vote, which means no school can evaluate its own civics programme on anything but a test.