Summary
Cunha and Heckman develop a framework distinguishing ex ante returns to schooling (what an agent expects when deciding whether to attend college) from ex post returns (what they actually realize). Using a factor model to proxy the agent's information set at decision time, they identify the full distributions — not just means — of both types of returns. A central finding is that roughly half of ex post return variability is unanticipated: agents face genuine earnings uncertainty, meaning the schooling decision is made under substantial risk. This explains why ordinary least squares (OLS) and instrumental variables (IV) estimates of the return to schooling can both be correct but measure different objects.
Key Claims
- Ex ante ≠ ex post: Because future earnings are uncertain at decision time, the return an agent expects when choosing to attend college differs from the return they ultimately receive. These two distributions overlap but are not identical.
- Essential heterogeneity: Individuals have private information about their own above-average returns and select into schooling accordingly. This "selection on gains" is rational, not a bias — but it means that OLS overestimates the average treatment effect (ATE) for the full population (upward selection on returns), while IV identifies the local average treatment effect (LATE) for compliers, who may have lower-than-average returns.
- Factor model identification: Observable proxies for cognitive and non-cognitive skills (test scores, family background) are used to separate the agent's information set at decision time from realized outcomes. This allows separate identification of what the agent knew ex ante vs. what they learned ex post.
- Uncertainty is large: Approximately half of the variance in ex post returns was not predictable at decision time. Agents face genuine income risk from the schooling decision — not merely optimal selection under full information.
- LATE ≠ ex ante ATE ≠ ex post ATE: The IV estimand for an instrument that shifts some people into college (e.g., distance to a college) identifies the LATE for the specific marginal students moved by that instrument. This can be above or below the full-population ex ante or ex post ATE depending on whether compliers have higher or lower private information about their returns.
- Policy implication: Because ex ante and ex post returns differ, individuals cannot perfectly optimize their schooling decisions even with rational expectations. Information provision (revealing what the agent cannot know at decision time) and consumption insurance (allowing smoother adjustment to realized earnings) can raise welfare independently of any change in average returns.
Concepts Introduced or Extended
Entities Mentioned
Quotes
"The key distinction between ex ante and ex post returns is that the former is based on agents' information sets at the time they make their schooling decisions, while the latter is based on the full information set including realized outcomes."
My Take
This paper resolves a long-standing puzzle: why do IV estimates of the return to schooling often exceed OLS estimates? The standard answer (ability bias offsets credit-constraint selection) is incomplete. Cunha and Heckman show that essential heterogeneity — rational selection on private information about returns — can make the LATE (for compliers who are moved into college by instruments like distance) exceed the full-population ATE. The factor-model approach is technically demanding and relies on the measurement system for unobservable skills, but the conceptual distinction between ex ante and ex post returns is now standard in the returns-to-education literature. Note that the PDF was not machine-readable; this source page is written from knowledge of the published paper.