The media is filled with tales of college graduates with crippling debt and limited job prospects. Unexplained variation in earnings equations and costs of higher education have been rising. Worries about the quality of investment in higher education have been growing.

The usual response is “it’s even worse without a college degree,” bolstered by a comparison of mean earnings of college graduates and high school graduates. But heterogeneity in earnings among those with the same educational attainment is large and growing. Given the growth in earnings’ tails, means are becoming poorer measures of expectations for many. Moreover, some of the variation in earnings conditional on education can be predicted with readily obtainable information, such as type of degree, while some cannot.

The current policy debate largely focuses on mean earnings differences between educational groups. We enrich this debate by emphasizing the full distribution of earnings, particularly at the top and bottom. Our broad research questions are: How variable are returns to higher education? Can risk be substantially reduced through type of degree or institution? Is risk larger for students of some backgrounds?  Have these returns, risks and their determinants changed between the 1980s-90s and the 2000s?

(Without exogenous variation, we cannot estimate unbiased causal returns to education. However, the NLSY’s rich controls remove much of the omitted variables bias. More importantly, our analysis directly addresses the current form of much policy debate, focused on crude comparisons of means, rather than true causal effects.)

We use the NLSY79 and NLSY97 surveys and focus on earnings. We do the following analyses:

• Describe the distributions of earnings within education categories, particularly the 5^th, 10^th, 90^th and 95^th percentiles.
• Examine graphically how these distributions have changed between the two cohorts, exploring how variance, skewness, kurtosis and the ratios of 5^th, 10^th, 90^th and 95^th percentiles to median have changed.
• Examine the distributions of (Associates and Bachelors) college educated, stratifying by other variables (one at a time) to observe their effects: public vs. private institution; technical vs. general (Associates); transfers vs. natives (Bachelors); high school quality proxies; AFQT score in youth.  
• Estimate traditional regression models for education, controlling for student, school and family characteristics.
• Oaxaca decomposition between cohorts of earnings and earnings net of costs, decomposing differences between cohorts into effects due to changes in mean characteristics and changes in characteristics’ effects.
• Quantile regression with both only education independent variables and a full set of controls.

Enriching the public debate about the financial returns to higher education will both improve policy and help maintain the credibility of policy recommendations in light of truly disappointing experiences of some college graduates.