Does Increasing Women's Schooling Raise the Schooling of the Next Generation?

Pairs of identical (monozygotic) twin parents from the Minnesota Twin Registry and their children; the sample is small — a limitation Holmlund et al. 2011 name explicitly when they redo the design on a much larger Swedish twin sample

US adult identical twins and their children, educational attainment.

American Economic Review 92(1), 323-334. The first attempt to difference out the genetic confound in intergenerational schooling by comparing the children of monozygotic twin parents who ended up with different amounts of education. If the twins are genetically identical, the within-pair difference in child schooling cannot be inherited ability. The design's known weaknesses are the ones Griliches (1979) and Bound & Solon (1999) identified for all twin-FE estimators — it assumes twins differ in schooling and in nothing else that affects their children's schooling, and differencing amplifies measurement error, so twin-FE bias can be as large as or larger than OLS bias. NUMBERS BELOW ARE READ SECONDHAND from Black, Devereux & Salvanes 2005's and Holmlund, Lindahl & Plug 2011's descriptions of this paper, not from the AER text.

The paper's answer to its own title question is no — and possibly the reverse. In ordinary cross-sectional regressions, even controlling for father's schooling and father's log earnings, mother's schooling carries a coefficient of 0.13 and father's 0.33 on child schooling, both significant (as reported by Black et al. 2005). Once the comparison is made WITHIN pairs of female monozygotic twins — differencing out genetics and shared rearing — the coefficient on mother's schooling turns NEGATIVE and is nearly significant. The parallel exercise on male MZ twin pairs leaves father's education about the same size as OLS. The authors conclude that the positive cross-sectional relationship between a mother's schooling and her child's schooling is not robust to controls for unmeasured intergenerationally correlated endowments, while the paternal effect is; and that increasing women's schooling may not raise, and might lower, the schooling of the next generation. Whatever one makes of the maternal sign flip, the load-bearing result for this archive is the same as Black et al.'s: the observed parent-education gradient is substantially an endowment gradient, not an instruction gradient.

The design exists to remove it and does so for the shared-genes channel — MZ twin parents are genetically identical, so within-pair differences in child outcomes cannot be inherited maternal genotype. What it does not remove is whatever caused one twin to get more schooling than her identical sister, which may itself be a trait (health, motivation, spousal match) that independently affects her children. Assortative mating is a live confound: the more-educated twin marries a more-educated husband, which the specification tries to absorb by controlling for father's schooling and earnings.

Contested and partly upheld. Antonovics & Goldberger (AER 2005, 10.1257/000282805775014353) show the results are sensitive to the coding of the schooling data; Behrman & Rosenzweig replied in the same issue (10.1257/000282805775014263). Holmlund, Lindahl & Plug 2011 redo the design on a far larger Swedish twin sample and reproduce the qualitative pattern — fathers positive and significant, mothers near zero — with much greater precision (corrected estimates: mothers -0.001, SE 0.065; fathers 0.120, SE 0.069), while noting their paternal estimates are smaller than Behrman & Rosenzweig's and that they cannot separate MZ from DZ pairs. So the striking maternal-vs-paternal asymmetry replicates; the negative maternal point estimate does not, and the IV literature (Black et al. 2005) finds the OPPOSITE asymmetry, with mothers mattering and fathers not.