Does education improve citizenship? Evidence from the United States and the United Kingdom

US: 30,026 NES respondents (7,387 conditioning on registration); UK: 22,935-63,858 BES respondents depending on outcome

Adults aged 20+ in the pooled 1948-2000 American National Election Studies and the pooled 1963-1997 British Election Surveys, assigned compulsory-attendance and child-labour laws by state and year of residence at age 14.

Instrumental variables: state/cohort variation in compulsory-attendance and child-labour laws instruments for high-school graduation. First-stage F ~14 in the full US sample; weaker (F~2-3) in the registered-voters subsample, which the authors flag. Identifying assumption is that the timing of law changes within a state is orthogonal to the later civic behaviour of the affected cohort. The authors run sensitivity checks conditioning on state-of-birth political environment at age 14, whites-only, and state-by-year effects. IMPORTANT MEASURE CAVEAT: the headline outcome is SELF-REPORTED voting, not validated turnout. The authors defend this by showing, in the NES vote-validation years (1964-1990), that misreporting is NOT increasing in education - it runs 18-20% for dropouts through some-college and drops to 14% for college graduates - and state results are similar with validated voting. That is a real check but it is only available for a subset of years.

Strong US effect, no UK effect. In the US, IV estimates of high-school graduation on self-reported voting run 0.288 to 0.305 (SEs 0.120-0.143), essentially identical to OLS (0.256-0.286); the authors summarise the magnitude as 28.8 to 34.2 percentage points. Conditioning on being REGISTERED cuts the OLS effect to 9 points and the IV effects to 0.179-0.281 with huge standard errors - i.e. most of the American education-turnout gradient runs through the registration hurdle, which the UK does not have. In the UK, the IV effect on being on the electoral register is -0.007 (SE 0.049), a precise zero. But UK education DOES move softer participation: +0.066 (0.018) on trying to persuade others, +0.095 (0.026) on discussing politics with friends at least occasionally, -0.092 (0.025) on never discussing politics, and +0.033 (0.014) on considering oneself politically active. It does NOT move satisfaction with democracy (-0.002, 0.022) or party attachment (0.012, 0.030). The authors' own counterfactual: without the 36-point rise in US high-school completion between 1964 and 2000, 2000 turnout would have been 10-12 points lower.

Medium. The IV design is meant to strip selection, and it is the right design for this question. But the estimates rest on self-reported behaviour, the instrument is weak in the registered subsample, and the IV point estimates are statistically indistinguishable from OLS - which is either evidence that OLS was unbiased or evidence that the instrument is not doing much work. The UK null under the same design is the strongest internal argument that the US result is not a general fact about schooling.

Its own US result was reproduced contemporaneously and independently by Dee 2004 with different data and a different instrument. But the paper REPLICATES ITSELF INTO A NULL: the same design run on UK data finds no turnout effect. Later well-identified work in Norway (Pelkonen 2012) and Sweden (Lindgren et al. 2019) also finds average nulls, and Berinsky and Lenz 2011 fails to reproduce the US effect with a draft-lottery instrument. The US-positive/everywhere-else-null pattern is the single most important fact in this literature.