Which Degrees Pay the Debt Back

Of the 21,940 bachelor's programs that report both figures, 8.8% carry more median federal debt than median first-year pay¹. That is true.

So is 6.4%. So is 5.4%. So is 0.75%.

All four come from the same federal file, but not from the same rows. Each answers a slightly different question, and a board deck usually shows only one.

The file

The US Department of Education's College Scorecard reports debt and earnings by school, field of study and credential². This study uses the release dated 2026-06-10. Debt is the median of federal Direct and Grad PLUS loans. It leaves out Parent PLUS, private loans and interest, so the Scorecard warns it may often run below what graduates owe². Pay is the median one year after completion, for federally aided graduates who worked and were not enrolled. Their ratio sorts programs into four zones.

The zone lines are this study's choice, not the government's³. Say so on the chart.

Choice 1: What counts as one?

The 8.8% counts programs. A tiny program weighs the same as a huge one. Weight each program by the degrees it awarded in 2021-22 and 2022-23, and the share falls to 5.4%⁴. The programs above the line tend to be the smaller ones.

A dean chooses among programs. A family chooses one path for one student. The unit should match the decision.

Choice 2: Whose debt?

The headline counts federal debt borrowed at any school. Count only debt borrowed at the school that granted the degree, and the share drops to 6.4%, across the 21,634 programs that report that figure⁵. The narrow measure leaves out loans from earlier schools, such as before a transfer.

Graduate degrees force the narrow measure. The any-school field is empty at every graduate level, so graduate rows use debt borrowed at that school⁶. One chart of every credential therefore mixes two definitions.

Medicine shows why labels matter. Across 142 Medicine programs, the median program debt is $194,063, counting only graduate loans at that school. The median program pay in year one is $65,517. That is a ratio of medians of 2.96; the median program ratio is 3.08⁷.

Each year since 1982, 92% to 95% of US MD seniors in the Match have landed a first-year residency post⁷. So for most new doctors, year one is residency. The ratio is real; the "high-risk" label misleads without a note on training pay.

Choice 3: When do you measure pay?

First-year pay is a hard test. Swap in pay four years out and the bachelor's share above 1.0x falls to 0.75%⁵. That pay belongs to an earlier class: 2017-18 and 2018-19 graduates, with pay measured in 2022 and 2023, in 2024 dollars. It covers 22,099 programs, so neither the people nor the rows fully match.

The first-year figures cover graduates of 2018-19 and 2019-20. Their first-year pay was measured in calendar 2020 and 2021, in 2022 dollars. Those were pandemic years, and the Scorecard itself warns that economic conditions can move earnings⁸.

Choice 4: Who is missing?

After merging branch-campus rows, the file holds 68,319 bachelor's programs. Only 21,940 publish both debt and pay, about a third⁹. Yet those programs awarded 81.7% of bachelor's degrees in 2021-22 and 2022-23. The missing two-thirds awarded the other 18.3%.

Coverage has two numbers: the share of rows and the share of degrees awarded. Report both.

The file also covers only students with federal aid. The Scorecard warns against assuming they stand for all graduates⁹.

Choice 5: Ratio of medians, or median of ratios?

A field's ratio can be built two ways. Divide the field's median debt by its median pay. Or take each program's own ratio and find the median. In 30 of the 34 field rows profiled, both methods give the same zone. Near a zone line, they do not¹⁰.

Neither method is wrong. Picking one silently is. Flag every row that sits close to a line¹⁰.

The claim everyone wants: field or school?

Does what you study matter more than where you study it? The study modelled the log of debt over first-year pay across all 21,940 bachelor's programs. A model by field of study reaches an adjusted R² of 0.49 across 302 fields. A model by school reaches 0.34 across 1,698 schools¹¹.

That ratio is 1.47. Resampling schools 400 times puts its 95% interval at 1.35 to 1.63¹². The rule, written in advance, reads 0.67 to 1.5 as "comparable", so 1.47 is comparable. A check added later found 42% of resamples above 1.5¹².

So the honest sentence is modest. Field of study explains somewhat more of the spread than the school does. The tempting version puts a big multiplier on a slide; the data does not carry one.

What this means for your BI team

• Name the unit in the metric title: share of programs, or share of degrees awarded.
• Define every input where the reader looks, including whose debt and which year's pay.
• Print coverage twice: share of rows and share of degrees awarded.
• Pick ratio-of-medians or median-of-ratios once, write it down, and flag rows near a threshold.
• Never stack two definitions in one chart without a label.
• Write the decision rule before the analysis, and ship the interval with any multiplier.
• Prefer the publisher's own derived fields, such as the Scorecard's 10-year payment, priced at the 4.99% undergraduate and 7.54% graduate rates in force when it was calculated⁸.

These choices are also where numbers drift in a migration. A Tableau view that divides two sums can return from Power BI as an average of row-level ratios. Both run, but only one matches the old dashboard.

A good migration plan lists every such choice and tests it before anyone compares totals.

Notes

1. Share of the 21,940 bachelor's programs that publish both figures, from tableau-sources/out/findings.json (college.bachelors); audit claim 23 in tableau/real-data-audit.md. ↩︎
2. US Department of Education, College Scorecard field-of-study file Most-Recent-Cohorts-Field-of-Study_06102026.zip, downloaded 2026-09-24 CDT (2026-09-25 02:44 UTC, per tableau-sources/manifest.json). Cohort and field names are in tableau-sources/out/college_scorecard_cip.json (metadata). The Field of Study data documentation, version September 2025, covers the rest. Earnings cohort of aided graduates who worked and were not enrolled: printed p. 3. Loan types and exclusions, and the warning that debt may run below what is owed: printed pp. 4-5 and 8-9. ↩︎
3. Zone lines at 0.8x, 1.2x and 1.6x in college_scorecard_cip.json (metadata, zones); shares in findings.json (college.bachelors) and audit claim 24. ↩︎
4. Share weighted by IPEDSCOUNT1 plus IPEDSCOUNT2, in findings.json (college.bachelors_weighted). These are IPEDS award counts for 2021-22 and 2022-23 (data dictionary, FieldOfStudy_Cohort_Map). IPEDS award counts count double majors twice (FoS documentation, printed p. 2). ↩︎
5. college_scorecard_cip.json (robustness.eval_debt.bachelors, 21,634 programs; robustness.four_year_earnings.bachelors, 22,099 programs); the four-year figure also appears in the audit. Four-year cohort years: tableau-sources/raw/CollegeScorecardDataDictionary.xlsx (FieldOfStudy_Cohort_Map, FieldOfStudyMostRecent column). ↩︎
6. college_scorecard_cip.json (shares_by_credential, metadata.graduate_rule); audit claim 27. ↩︎
7. Program rows in college_scorecard_cip.json (programs, cip 51.12): Medicine across 142 first-professional programs, debt borrowed at that school. Both dollar figures are medians of program medians (tableau-sources/scorecard.py, page_rows). Undergraduate loans are excluded for graduate credentials (FoS documentation, printed p. 5). Residency match rates: National Resident Matching Program, 2025 Main Residency Match results, released 2025-03-21, accessed 2026-10-17. ↩︎
8. College Scorecard Field of Study data documentation, version September 2025, downloaded 2026-09-24 CDT (2026-09-25 02:44 UTC). Its repayment-rate note that award years 2018-19 and 2019-20 partially overlap the pandemic: printed pp. 6-7. "Economic conditions": printed pp. 10-11. The 10-year payment's rates, set "at the time of calculation": printed p. 5 (PDF p. 6). ↩︎
9. college_scorecard_cip.json (metadata.bachelors_programs_total, bachelors_completer_coverage_pct); audit section "What else the real data showed". Branch-campus rows merge to one row per OPEID6, CIP and credential (tableau-sources/scorecard.py, dedupe). Federal-aid coverage and the warning on representativeness: FoS documentation, printed p. 8. ↩︎
10. Boundary rows in tableau-sources/out/robustness.json (college.boundary_rows). Zone agreement, credential and debt basis for the 34 on-page rows: college_scorecard_cip.json (programs). ↩︎
11. tableau-sources/out/college_scorecard_cip.json (field_vs_institution: field.groups 302, institution.groups 1,698, adj_r2); findings.json (college); audit claim 25. ↩︎
12. Cluster bootstrap over schools, 400 draws, in robustness.json (college.cluster_bootstrap); audit claims 25 and 26. The 0.67 and 1.5 rule sits in plans/2026-09-24_portfolio-real-data.md, "Pre-registered methods", College Scorecard. The same plan's Outcome section lists the robustness.py checks as not pre-registered. ↩︎