Education and Political Participation: Exploring the Causal Link

247,876 CPS respondents (Vietnam-era cohorts, 1972-2000 surveys)

US men and women born 1935-1959, with men born 1946-1948 as the treated cohorts whose college attendance was inflated by Vietnam draft avoidance. Turnout measured in presidential elections 1972-2000.

Difference-in-differences IV: the interaction of male with birth years 1946-1948 instruments for 'some college', exploiting draft-avoidance-induced college attendance. First-stage F around 115 for some-college (and >1,000 for veteran status), so this is not a weak-instrument story. Enormous sample (247,876). Extensive robustness: quadratic and cubic cohort trends, males only, four-year windows around the treated cohorts, dropping non-peak affected years, alternative education measures (college degree, years of college), and dropping the 1972-1980 elections. All hold. The outcome is SELF-REPORTED CPS turnout, which is the same measure as the rest of the US IV literature. Key caveats the authors state: the estimate is a LATE for draft-induced college-goers, who might be atypical; and the confidence intervals are wide enough to contain both zero and the OLS estimate.

Instrumenting college attendance with the Vietnam draft, the effect of 'some college' on turnout is 0.057 to 0.083 (SE ~0.15) - less than a third to less than half the OLS estimate of 0.19, and not distinguishable from zero. Using college degree as the instrumented variable gives 0.127 (0.360) against an OLS of 0.196; college years gives 0.138 (0.320) against an OLS of 0.304. When the treated group is widened to all cohorts Card and Lemieux identify as affected (1944-1950), the IV estimate turns SLIGHTLY NEGATIVE: -0.073 (0.074) for some college, -0.040 (0.162) for college degree, -0.070 (0.170) for college years. The authors' defence against 'your method just cannot find effects' is that the same instrument produces strong significant effects on other outcomes (smoking avoidance in Grimard and Parent 2007, de Walque 2007) and, in their own data, on income. Their conclusion: education 'may not be entirely the great leveler and may partly be just the great proxy' for pre-existing family background and cognitive traits. That is the hereditarian reading, stated by the authors themselves.

Low for the design, high for what it implies about the rest of the literature. The draft lottery-era instrument is exogenous to individual traits, so this estimate is not confounded by heritable disposition - and it is precisely the estimate that collapses. The paper is therefore direct evidence that the observational education-turnout gradient is substantially selection.

This is the US counter-finding to Dee 2004 and Milligan et al. 2004: same country, same outcome, different instrument, much larger sample, and the effect collapses. It converges with Kam and Palmer 2008 (matching), Pelkonen 2012 (Norway) and Lindgren et al. 2019 (Sweden). The authors are honest that their estimates are imprecise enough not to exclude the OLS effect either.