Willis and Rosen 1979 — Education and Self-Selection

Summary

Willis and Rosen (1979) apply Roy's (1951) comparative advantage model to educational choice, establishing the canonical self-selection framework for returns-to-schooling research. Each individual has two potential lifetime earnings: $w_1$ (if college-educated) and $w_0$ (if not). People attend college when the expected discounted return net of direct and opportunity costs is positive. Estimating a factor-analytic earnings model on Panel Study of Income Dynamics (PSID) data, the authors find negative selection on the no-college counterfactual: college-goers would have earned below-average wages in the non-college sector. College and non-college work draw on different types of human capital, so those with high comparative advantage in college (high $w_1$ relative to $w_0$ ) sort into higher education. Any instrument or policy that induces marginal students into college identifies returns for individuals near the selection threshold — substantially below the average returns of infra-marginal attenders. The paper is the 1979 intellectual predecessor of the Heckman-Vytlacil Marginal Treatment Effect (MTE) framework and the essential heterogeneity literature.

Key Claims

Roy comparative advantage model: Educational choice is sector selection in a two-sector Roy economy. People attend college when $E[w_1 - w_0] - c > 0$ (return exceeds cost). Selection is on the individual-specific gain, not on ability or earnings level alone. The college population is NOT simply the highest-ability individuals — it is those with the highest comparative advantage in college-type work.
Negative selection on no-college earnings: College-goers would have earned below-average wages in the non-college sector ( $E[w_0 \mid \text{College}] < E[w_0 \mid \text{No College}]$ ). This reflects complementarity between college-type human capital and college-sector earnings: those who are best suited for college are systematically less suited for non-college work.
OLS understates gains for infra-marginal attenders, overstates them for non-attenders: Because non-attenders have above-average $w_0$ (they are good at non-college work), naïve ordinary least squares (OLS) underestimates their college counterfactual relative to non-attenders' actual wages. For attenders, OLS underestimates the true gain (their counterfactual $w_0$ is below-average). These two biases push in the same direction, making OLS an unreliable guide to treatment effects for either group.
Marginal student has below-average returns: Any instrument that induces additional college attendance moves students near the indifference threshold (return ≈ cost). Their average returns are substantially lower than infra-marginal attenders who would attend regardless. This is the 1979 precursor to why different valid instruments yield different instrumental variables (IV) estimates — each identifies a different segment of the heterogeneous return distribution.
Factor-analytic identification: Willis and Rosen estimate a latent-factor model where a single unobserved ability factor enters both earnings equations. The covariance between $w_1$ and $w_0$ , and hence the direction of selection bias, is identified from the observable joint distribution of earnings and education.
Comparative advantage vs. absolute advantage: Under absolute advantage, those with highest $w_1$ attend college. Under comparative advantage, those with the highest $w_1/w_0$ (or $w_1 - w_0 - c$ ) attend. The two predictions diverge when college and non-college earnings are imperfectly correlated — the empirically relevant case.

Concepts Introduced or Extended

Entities Mentioned

Quotes

"The choice of a college education thus is not simply a choice of whether to acquire a given amount of human capital, but also involves selection among different types of human capital that have different comparative advantages in different sectors of the labor market."

"Those who do not attend college earn more than those who attend college would have earned had they not attended college. This is a general implication of our model in which the college and non-college sectors require different types of human capital."

My Take

Willis and Rosen (1979) is among the most consequential papers in labor economics. The Roy-model reframing transformed how the selection problem is understood: self-selection into education reflects comparative advantage, not just ability, and the resulting negative selection on the no-college counterfactual means OLS comparisons are unreliable guides to the returns for either group. The paper anticipated by fifteen years what Heckman and Vytlacil (2005) would formalize as the MTE framework and essential heterogeneity — though in 1979 the connection to IV and the local average treatment effect (LATE) (Angrist-Imbens 1994) was not yet explicit. The main limitation is specification-dependence: the factor model requires functional form assumptions about the unobserved ability distribution, and the PSID sample in 1979 was relatively small. But as an organizing framework, Willis-Rosen remains the foundation for Card (1999), Cunha and Heckman (2007), and the broader heterogeneous-returns literature.