Brand and Xie 2010 — Who Benefits Most from College

Summary

Brand and Xie use propensity-score stratification to test whether the returns to college education exhibit positive or negative selection — that is, whether those most likely to attend college also benefit most (positive) or least (negative). Using two independent longitudinal datasets (National Longitudinal Survey of Youth 1979 (NLSY79) and Wisconsin Longitudinal Study (WLS)), they find consistent evidence for the negative selection hypothesis: individuals with the lowest propensity to attend college show the largest earnings gains from doing so. The finding holds for both men and women, at every observed life-course stage, and across two cohorts.

Key Claims

• Negative selection confirmed in 10 of 10 cases: For NLSY and WLS, men and women, and multiple career stages, the hierarchical linear model (HLM) Level-2 slope relating treatment effects to propensity score strata is negative. Low-propensity college-goers gain the most from a degree.
• The mechanism is counterfactual deprivation: The pattern does not arise because low-propensity college graduates earn more than high-propensity graduates — they do not. It arises because low-propensity non-graduates earn very little; without a degree, they lack human, cultural, and social capital and face severely limited labor market prospects. High-propensity individuals without a degree can still draw on superior resources and abilities.
• Magnitude: For NLSY men, the estimated wage return ranges from $\approx 30\%$ in stratum 1 (most disadvantaged) to $\approx 10\%$ in stratum 5 (most advantaged) — a $20$ percentage-point gap that persists across the life course. For NLSY women in their late 30s–early 40s: stratum 1 $\approx 40\%$, stratum 5 $\approx 25\%$.
• Auxiliary evidence for economic motivation among low-propensity attendees: (1) "Value of college" variable (WLS): advantaged and disadvantaged non-college youth value college equally, but disadvantaged college-goers value it far more than their peers who did not attend — consistent with economic motivation driving low-propensity attendance. (2) College majors: low-propensity students concentrate in business and education (immediate economic payoff); high-propensity students in sciences and humanities (academic identity, deferred returns).
• Sensitivity to covariate set: With a limited covariate set (omitting ability, achievement, aspirations, encouragement), the slope reverses to positive selection. The negative pattern requires conditioning on the full set of noneconomic factors that differentially predict college attendance among high-propensity individuals. This shows the result is driven by noneconomic forces being more predictive for high-propensity students, not by omitting a single variable.
• Consistency with prior evidence: Consistent with Heckman, Tobias, and Vytlacil (2001) finding that average treatment effect (ATE) $\approx 9\%$ while average treatment effect on the treated (ATT) $\approx 4\%$ (those who attend gain less than a randomly chosen person would); consistent with Card (2001) instrumental variables (IV) $>$ ordinary least squares (OLS) finding (instruments pull in marginal students with higher returns).
• Policy implication: Educational expansion targeting currently unlikely attenders should produce larger average returns than are observed in the existing college-going population. The ATT understates the benefit of expanding access.

Data and Method

• Data: NLSY79 (nationally representative; $N = 2{,}474$ after restrictions to single birth cohort with complete pre-college covariates); Wisconsin Longitudinal Study ($N = 7{,}905$; regionally based but large, homogeneous, with rich pre-college measures including reliable cognitive ability score). Two cohorts as robustness check.
• Outcome: Log hourly wages at multiple life-course stages (NLSY: ages 29–32, 33–36, 37–40; WLS: ages 35, 53).
• Method: Three-step propensity-score HLM (Xie and Wu 2005). Step 1: logistic regression for college completion → estimated propensity scores; respondents grouped into balanced strata. Step 2 (Level 1): OLS treatment effects within each stratum. Step 3 (Level 2): HLM regression of stratum-specific effects on stratum rank — the Level-2 slope summarizes whether effects rise or fall across the propensity distribution.
• Identification: Selection-on-observables (ignorability). Pre-college covariates include family income, parental education, family structure, sibling count, race/ethnicity, cognitive ability, academic track, parental/teacher encouragement, and friends' college plans.

Concepts Introduced or Extended

• Heterogeneous Treatment Effects
• Marginal Treatment Effect

Entities Mentioned

• Jennie Brand
• Yu Xie
• Stephen Morgan

Quotes

"Individuals most likely to benefit from a college education are the least likely to obtain one."

"The pattern emerges because low propensity non-college goers earn so little."

"With the full set of covariates at our disposal, we observe a modest pattern of negative selection; when we trim covariates to a more limited set, we observe positive selection."

My Take

This is the foundational empirical paper for the negative selection hypothesis in higher education. The design is careful — two independent datasets, multiple cohorts, multiple life-course stages — and the 10-for-10 consistency of negative Level-2 slopes is persuasive as a directional finding. The mechanism story (counterfactual deprivation, not absolute advantage) is theoretically clean and convincingly supported by the auxiliary analyses on college major choice and the "value of college" variable.

The main limitation is the ignorability assumption. The sensitivity analysis showing that negative selection requires a rich covariate set is reassuring but also cuts both ways: if there exist unobserved noneconomic factors that further differentiate high-propensity students, the direction of the result could shift again. The authors are candid about this: they note that unobserved economic motivation among low-propensity students (who overcome considerable odds to attend) may itself be driving part of the negative pattern. Educational expansion that brings in new entrants without that unusual economic drive could flatten the slope. This caveat limits the paper's welfare and policy conclusions but does not undermine its descriptive contribution to heterogeneity analysis.