Is Your Degree Worth It? — Full Study

$418 Billion in Federal Aid at Risk

31% of U.S. higher education students are enrolled in programs that cost more than they will ever earn back. Over five years, $418 billion in federal Title IV aid flowed to those programs. The students harmed most are low-income Pell Grant recipients — the people the aid system was designed to protect.

The information to identify these programs already exists. The College Scorecard publishes IRS-derived earnings for every program at every institution — real W-2 data, not surveys. But that information is not used at the decision point. Students choose programs based on location, reputation, and marketing. Counselors lack the tools to translate earnings data into return-on-investment comparisons. And policymakers set eligibility rules without modeling what happens when you actually enforce them.

We analyzed 65,935 programs at 6,255 institutions and asked five questions about where the value is, where it isn't, and what could be done about it.

Institution Matters More Than Field

Institution explains 31% of ROI variance. Field of study explains 19%. The interaction between them — specific fields at specific schools — explains another 9%. This means where you go matters roughly 60% more than what you study.

When we built a predictive model to flag negative-ROI programs before students enroll, the institution’s overall negative-ROI rate was the single most important feature — accounting for 53% of model importance. The field’s national negative-ROI rate was second at 27%. Credential level was third at 13%.

The implication is clear: if a school produces bad outcomes across many fields, any new program at that school is suspect. Students choosing between the same major at two different schools should weight the institution’s track record heavily.

What We Used

Earnings are real W-2 data, not self-reported. The Department of Education matches Scorecard records to IRS tax filings. Median earnings reflect what graduates actually earned 1 and 5 years after completion. This is the most reliable source of program-level earnings data in the United States.

Programs with fewer than 30 completers are suppressed for privacy. This removes 72% of program records from the dataset. The 65,935 programs with published data skew toward larger programs at larger institutions — a limitation, but one that concentrates the analysis on programs that affect the most students.

All data sources are publicly available. Pipeline ingests real College Scorecard and IPEDS data.

Five Questions

Q1: Which Programs Have Negative ROI?

Of 65,935 programs with published earnings data, 34.4% have negative lifetime ROI — 22,692 programs where total costs exceed projected lifetime earnings gains. Certificates are worst: 64% negative. Associate degrees: 44%. Bachelor’s: 15%. Graduate degrees are safest at 2–4%. The same field can be positive at one school and deeply negative at another. Cosmetology is 98% negative across all institutions. Nursing is 92% positive.

Method: Discounted cash flow on IRS-derived earnings vs. total cost of attendance

Q2: Is It the Field or the School?

Two-way ANOVA variance decomposition on program-level ROI. Institution explains 31% of the variance. Field of study explains 19%. Their interaction accounts for another 9%. The remaining 41% is within-group variation — specific program characteristics, local labor markets, and noise. Where you go matters roughly 60% more than what you study, but both matter, and the combination matters most of all.

Method: Two-way ANOVA variance decomposition

Q3: What Predicts Negative ROI?

Gradient boosting classifier achieves AUC 0.927 on the task of predicting which programs will have negative ROI — using only non-tautological features (no earnings or debt data as inputs). The institution’s overall negative-ROI rate is the dominant signal at 53% of model importance. The field’s national negative-ROI rate is second at 27%. Credential level is third at 13%. A school that produces bad outcomes in most of its programs will produce bad outcomes in nearly all of them.

Method: Gradient boosting with non-tautological features (no earnings/debt inputs)

Q4: What If We Cut Federal Aid to the Worst Programs?

If Title IV eligibility were revoked for all negative-ROI programs, 2 million currently enrolled students would lose federal aid access. 2,626 institutions — those where the majority of enrollment is in negative-ROI programs — would become financially unviable. The fields hit hardest: cosmetology (98% of programs cut), drama and theater (76%), liberal arts (71%). The institutions hit hardest are for-profit certificate programs and small private colleges with narrow program offerings.

Method: Scenario analysis / arithmetic on Q1 outputs

Q5: Where Is the Biggest Improvement Leverage?

Not all negative-ROI programs are equally fixable. Leverage = proximity to breakeven × enrollment volume × feasibility. Programs that are slightly negative with large enrollment are the highest-leverage targets — small improvements affect thousands of students. University of Phoenix dominates the leverage rankings because of sheer scale: even modest earnings improvements across its programs would affect more students than fixing an entire small college. The leverage framework shifts the conversation from “which programs are worst” to “where can intervention do the most good.”

Method: Leverage ranking (proximity × volume × feasibility) on upstream outputs

How We Built This

This study follows the Analysis Driven Modeling (ADM) framework: start with the question, match fidelity to the decision, encode what you know, learn the rest.

ROI calculation (Q1): Discounted cash flow using real IRS-derived median earnings at 1 and 5 years post-completion. Earnings growth beyond year 5 is estimated from cross-sectional age-earnings profiles by field. Total cost includes tuition, fees, books, and estimated opportunity cost (foregone earnings during enrollment, benchmarked to high school median). Discount rate: 3%. Lifetime horizon: age 22 to 65.

Variance decomposition (Q2): Two-way ANOVA with institution and 2-digit CIP field as factors. Type III sum of squares. Interaction term captures field-at-institution effects that neither factor alone explains.

Predictive model (Q3): Gradient boosting classifier (sklearn) trained on non-tautological features only — no earnings, no debt, no ROI inputs. This ensures the model predicts negative ROI from institutional and program characteristics that are known before a student enrolls. Features include institution negative-ROI rate, field national negative-ROI rate, credential level, institution type, Pell share, and completion rate.

Policy simulation (Q4) and leverage ranking (Q5): Scenario arithmetic on upstream outputs from Q1. No new modeling — these investigations apply decision logic to the ROI calculations already completed. Q4 counts affected students and institutions under a blanket cut. Q5 ranks programs by a composite leverage score: proximity to breakeven (how close to fixable), enrollment volume (how many students affected), and institutional feasibility (whether the institution has other viable programs to sustain operations).

Techniques used: Discounted cash flow, two-way ANOVA, gradient boosting classifier (sklearn), scenario analysis. No neural networks — the task is tabular classification and arithmetic, not sequence modeling.

Python pipeline: all source code and data available on request.

What We Don't Know

Lifetime projections from 5-year data. College Scorecard earnings cover 1 and 5 years post-completion. Projecting to a full career requires cross-sectional growth rates by field. Some fields (e.g., liberal arts) have steeper late-career earnings curves that the 5-year snapshot undervalues. Others (e.g., trades) have flatter trajectories that the projection may overvalue.

72% of programs are suppressed. Scorecard suppresses earnings data for programs with fewer than 30 completers. The 65,935 programs with published data are disproportionately at larger institutions. Small programs at small colleges — which may have very different ROI profiles — are invisible to this analysis.

Opportunity cost is estimated, not observed. We assume students forgo the median high school graduate wage during enrollment. Students who work during school, attend part-time, or have prior career earnings face different cost structures. This assumption is most problematic for adult learners and graduate students.

No geographic cost-of-living adjustment. A program that produces $50K median earnings is worth more in rural Alabama than in San Francisco. We report raw earnings without purchasing-power adjustments, which penalizes programs in low-cost regions and flatters those in high-cost metros.

Cross-sectional growth rates may not match future trajectories. The earnings growth curves used for lifetime projection reflect today’s age-earnings profiles, not necessarily the trajectory a 2024 graduate will experience. Fields undergoing structural change (e.g., journalism, retail management) may have flatter future curves than historical data suggests.