Methodology

How every number on Truewise is computed. Last updated 2026-07-17. Data vintages: College Scorecard release 2026-06-10; BLS OEWS May 2024 + Employment Projections 2024-34; CRDC 2021-22.

Sources

Everything comes from public U.S. federal data. College outcomes use two College Scorecard bulk files (most recent release, 2026-06-10), joined on the IPEDS UNITID: the Field-of-Study file (per program: median earnings after completion, median debt) and the Institution file (the state and national high-school-graduate earnings thresholds). Careers demand adds the NCES CIP-to-SOC crosswalk, BLS OEWS wages, and BLS Employment Projections. High-school data comes from the Civil Rights Data Collection (CRDC).

The earnings-premium test (Value Check)

For each program (school × 4-digit CIP × credential) we take median earnings four years after completion, the same figure the Department of Education displays on the public College Scorecard site, falling back to the one-year figure only where the four-year value is suppressed, and compare it to the state high-school-graduate earnings threshold:

• Passes: median earnings are at or above the state threshold.
• Falls short: median earnings are below it.
• Insufficient data: earnings are suppressed (small cohort) or no threshold is available. These are shown as "not enough data", never guessed.

The headline number, in full

Our headline is that 1 in 11 college programs with reported earnings leave graduates earning less than a typical high-school graduate. Here is exactly what is behind it, including the parts most coverage leaves out:

• Denominator. Of 227,980 programs, 62,902 have reported earnings. Of those, 60,202 (about 26%) also have the state high-school-graduate benchmark needed for the earnings-premium comparison, and the 1-in-11 rate is computed on that judgeable set: 5,396 fall short (8.96%, or 1 in 11.2). The other 2,700 report earnings but cannot be judged because that state benchmark is unavailable, so they count as "insufficient data"; the remaining rows are mainly privacy-suppressed by ED for small cohorts. (The 227,980 total includes 7,020 rows with no institution ID, all of them insufficient-data; the downloadable dataset drops these, leaving 220,960, of which the same 60,202 judgeable programs are about 27%. The 8.96% failure rate is identical either way, since every dropped row is insufficient.)
• Horizon. Earnings are measured up to four years after completing. Where the four-year figure is suppressed we fall back to the one-year figure, and about 27% of the failing programs use that fallback. On strictly four-year rows the rate is 7.0% (3,914 of 55,539), roughly 1 in 14. We report the broader 1-in-11 figure and tag one-year rows with a visible "1-yr" marker so you can weigh them.
• Cosmetology. Among cosmetology programs (CIP 12.04) specifically, 96% fall short (1,078 of 1,120 with data).

How sensitive is that number?

The fair criticism of any rate built on the reported 26% is that suppression might not be random, so the visible programs could be a flattering sample. Suppression is indeed not random: undergraduate certificates are about 82% suppressed and fail at 38.6% where reported, while bachelor's degrees are about 59% suppressed and fail at 4.5%. Reported programs have a median of 33 graduates; suppressed ones have 4. So we tested whether that changes the answer.

Assuming every suppressed program fails at the rate observed in its own credential-and-sector group, and re-estimating across all 220,960 programs, the rate becomes 9.03%, against the 8.96% we publish. The composition of the hidden programs moves the headline by seven hundredths of a percentage point, so the figure is robust to the most likely objection to it.

One distinction worth stating plainly, because it is easy to conflate. Our headline counts programs. Weighted by the number of graduates instead, the rate is 7.37%. So: about 1 in 11 programs, but about 1 in 14 graduates. Failing programs are somewhat smaller than average. Both numbers are in site/data/value_check_summary.json and are recomputed by analysis/summary.py on every data refresh, so they cannot drift away from the data.

Reproduce these numbers: the dataset lives in published/value_check.parquet and the exact script is analysis/summary.py in the GitHub repo.

Principles

• Never impute. Suppressed values (PS, NA) become null and render as "insufficient data", never guessed. This rule is enforced by a data-quality gate (analysis/validate.py) with unit tests.
• Every figure carries its source and data vintage. Earnings reflect the recent past; we frame them as "graduates typically earned", never as a promise.
• A flag describes the data. It never tells anyone whether to attend a program.

Why do other studies say 1 in 4?

Some coverage (for example a widely cited HEA Group study) reports that attendees of about 1 in 4 higher-ed programs earn less than high-school graduates. The difference from our 1 in 11 is not a contradiction, it is three definitional choices:

• Who counts. Studies that count attendees (including non-completers) show higher failure rates than ones that count graduates, as we do here.
• The benchmark. A national high-school-earnings line, a state line, or an age-adjusted line each move the bar. We use ED's own state threshold.
• The denominator. Including or excluding privacy-suppressed programs changes the base. We are explicit that ours is the 26% of programs with reported earnings.

All three are defensible; we state ours plainly so the number is reproducible.

Cite this

The other modules, in brief

Affordability shows net price by family-income bracket (College Scorecard NPT41 to NPT45), coalesced across a school's sector; suppressed brackets show as "not reported". ROI divides a program's median federal debt by its yearly earnings premium over a high-school graduate (years to recoup what was borrowed); it is null when there is no premium or no debt. Mobility puts Pell share, completion, and the earnings-premium pass rate side by side against the national median, with a transparent "hidden gem" rule (beats all three medians); it is an access-and-outcomes view, not the Chetty income-mobility rate. Careers pools program earnings to the field level and joins BLS occupation pay and outlook through the NCES CIP-to-SOC crosswalk. High schools report, from the CRDC, which advanced courses a school offers, participation, and staffing (counselors, security).

Full detail for every module is in METHODOLOGY.md and the data dictionary.

Loan repayment, and why some figures say "or less"

School pages report whether borrowers actually pay their debt down, using the College Scorecard's borrower-based repayment status for students who completed, measured two years after entering repayment (share in default, share already paid in full), plus the share of all borrowers whose balance is declining three years in.

These columns behave differently from earnings. Besides ordinary suppression, the Department of Education censors small rates, publishing them as "at most" a value rather than an exact number. We neither discard those nor present them as exact: a censored figure is shown as "2% or less", which is precisely what ED published. Censored values are never used in averages or rankings. Of 6,273 institutions, about 3,600 have a usable default figure (roughly a third exact, the rest censored bounds) and about 4,700 have the three-year repayment rate.

Known limitations

• Earnings cohorts predate the current year; figures describe past graduates.
• The debt-to-earnings ratio shown is a plain ratio, not the amortized federal D/E rate.
• High-school data is the 2021-22 CRDC, the latest public-use release (a 2023-24 collection exists but its public-use file is not yet published); course offerings are fairly stable year to year, and the collection year is labeled on every high-school page.
• The Careers occupation layer uses a many-to-many CIP-to-occupation crosswalk, so it lists the occupations a field commonly leads to rather than predicting any one graduate's job.