Are there civic returns to education?

High School and Beyond: ~12,000 sophomore-cohort respondents followed to 1992; General Social Survey: 32,111

US high-school sophomores of 1980 interviewed at ages ~20 and ~28 (HS&B), and GSS respondents 1972-2000 assigned child-labour-law exposure by division and year of residence at age 16.

Two independent identification strategies. (1) Geographic proximity and density of junior and community colleges when the respondent was a teen instruments for college entrance, with school fixed effects, in HS&B. (2) Changes in teen exposure to restrictive child-labour laws instrument for years of schooling in the GSS. Dee validates the second instrument by showing its effects concentrate at the secondary level (grades 9-12 and HS graduation) and NOT at post-secondary levels, which is what a valid compulsory-schooling instrument should do, and shows effects are larger for children of poorly educated parents. He also tests for 'contagion' by re-running on cells and finds none. ALL OUTCOMES ARE SELF-REPORTED - registration, voting in the last year, voting in 1988, volunteering (HS&B), and turnout, newspaper readership, group membership and free-speech tolerance (GSS). No validated voter records anywhere in this paper.

Large positive civic returns, all on self-report. An additional year of schooling raised self-reported voter participation by 6.8 percentage points (t=1.93, only weakly significant, and a Hausman test fails to reject that OLS was already consistent). College entrance in HS&B raised registration by ~12 points, and raised voting in the last year, voting in 1988 and volunteering by roughly 26%, 28% and 15% relative to their means. Schooling also raised newspaper readership (t=2.02) and support for free speech by anti-religionists, communists and homosexuals by 8.0 to 12.5 points - several times the OLS estimates. But free speech for MILITARISTS and RACISTS showed smaller, statistically imprecise effects, and the effect on group memberships was positive but very noisy. Note what this pattern implies: the tolerance gains are selective, not general, which is what a social-desirability reading of self-reported attitudes would also predict.

Medium. Two independent instruments both pointing the same way is a real strength, and the child-labour instrument passes a sensible falsification test. But every outcome is self-reported, the headline turnout estimate is only weakly significant, and a Hausman test cannot reject that OLS was unbiased - which is exactly what a residual-confounding story would look like.

Published back-to-back with Milligan, Moretti and Oreopoulos 2004 in the same issue of the Journal of Public Economics, with different data and a different instrument, and reaching the same conclusion for the US - a genuine independent convergence. But both are contradicted by later well-identified non-US work (Pelkonen 2012, Lindgren et al. 2019) and by Berinsky and Lenz 2011 in the US.