Same 3,361 programs, old GE test vs new OBBBA test: 569 walk free and 19 get caught

Run the old Gainful Employment test and the new OBBBA earnings test on the same rows of PPD:2026 and the two casualty lists only partly agree. On the 3,361 ever-GE programs where both regimes can be scored, 1,002 fail both. 569 fail the old test and pass the new one. Just 19 go the other way.

The old test here is the overall GE result, earnings or debt-to-earnings, because that is the flag PPD:2026 populates. The earnings-only flags are NULL on every row in the file.

Combined, the two lists name 1,590 distinct programs and share 1,002 of them, so the overlap is 63.0 percent. They disagree on 588 of the 3,361 programs, 17.5 percent, and the disagreement runs almost entirely one direction: 36.2 percent of the old casualty list (569 of 1,571) walks free under the new regime, while only 1.9 percent of the new list (19 of 1,021) consists of programs the old test had cleared.

Both tests ship in the same file. PPD:2026 carries the 2023 FVT/GE flags and the OBBBA flags on the same 209,321 program rows across 5,096 institutions, which is what makes a direct head-to-head possible at all. What the Earnings Premium test is and where it comes from is already covered in the OBBB to Earnings Premium walkthrough. This piece is about what happens when you point both tests at identical rows.

Those 3,361 are 6.2 percent of the 54,063 programs PPD:2026 flags as ever having been Gainful Employment programs, so this is a comparable slice rather than a census of the GE universe. What thins it out is missing earnings data, not policy.

The obvious way to build that crosstab is to read two flags off the file. It does not work. ge_earn_test_2023_fail and ge_earn_test_algn_fail are NULL on all 209,321 rows: zero TRUE, zero FALSE, nothing to read. The GE earnings-test flags are unpopulated in this release, and no amount of filtering will surface them.

The inputs are there. ge_earn_test_bm_2023, the benchmark, is non-null on 208,859 rows (99.8 percent), and ge_earn_test_bm_algn on 209,082 rows (99.9 percent). So every GE earnings-test number in this post is reconstructed rather than read: md_earn_ne_p3_1516 below ge_earn_test_bm_2023 for the 2023 rule, and the aligned benchmark for the aligned rule.

That reconstruction is testable, because the overall GE flags are populated and the overall test is the OR of the earnings test and the debt-to-earnings test. Recomputing it as (md_earn_ne_p3_1516 below ge_earn_test_bm_2023) OR ge_dte_overall_2023_fail reproduces the shipped ge_overall_2023_fail with 0 mismatches across all 54,063 GE-universe rows, 1,715 failures against 1,715. The aligned version reproduces ge_overall_algn_fail with 0 mismatches as well, 1,905 against 1,905. Standing alone, the reconstructed 2023 earnings test fails 1,645 of the 3,709 programs that have both an earnings value and a benchmark (44.4 percent); the aligned earnings test fails 1,811 of 7,633 (23.7 percent).

The debt side barely moves the totals. Zero programs fail the debt-to-earnings test while passing overall, under either rule, so D/E failures are a strict subset of GE failures. ge_dte_overall_2023_fail is TRUE on 272 programs (53,791 FALSE) and its aligned counterpart on 206 (53,857 FALSE).

Two coverage rules decide who is in. The old test needs md_earn_ne_p3_1516 and a 2023 benchmark: 3,709 of the 54,063 GE programs qualify, 6.9 percent. The new test needs md_earn_wne_p4: 7,633 of 54,063 qualify, 14.1 percent, which leaves 46,430 GE programs (85.9 percent) that cannot be run through the new test at all. Intersect the two and you get 3,361 programs at 1,482 institutions, 6.2 percent of the GE universe.

Across the whole file the new test reaches further than that. It is computable for 49,860 of 209,321 programs (23.8 percent), exactly the set carrying a md_earn_wne_p4 value, with zero coverage mismatch. But inside the old GE universe, the only place a comparison is possible, coverage is thin on both sides. PPD:2026's privacy suppression rules bite hardest on small certificate programs, which is also where GE failures concentrate, so the rows that drop out are not dropping at random. The six-file walkthrough of PPD:2026 covers those suppression rules and the join that produces this table, and the measured version of the same problem shows how far the coverage gap runs: GE programs are worse covered than never-GE ones, 14.1 percent against 27.2 percent.

So read every number below against 3,361, not against 54,063 and not against 209,321.

Line the two regimes up on the 569 released programs and the bar is identical: median benchmark $34,808 under the old GE rule and $34,808 under the aligned rule. What moves is the earnings number. Those same 569 programs post a median of $30,852 measured the old way and $38,331 measured the new way, a $7,479 gap between the two medians. That is enough to clear the 513 of them that had failed the GE earnings test.

The pattern holds across the full comparable universe: median earnings rise from $37,504 on the old measure to $44,097 on the new one. The median per-program lift is $6,378, a median of individual differences rather than the difference between those two medians, so subtracting the medians will not reproduce it.

The schema explains part of why the bar sits still. The aligned GE benchmark and the OBBBA benchmark are the same number, identical on 167,503 of 167,503 rows where both exist, zero higher and zero lower, with quartiles matching exactly at $34,350, $36,082 and $46,391. Restricted to the GE universe they are identical on 12,245 of 12,245 rows, both medians $46,411. Because the aligned GE earnings test and the OBBBA flag are then the same comparison over the same two columns, they agree on all 7,633 GE programs with an OBBBA-testable earnings value: 1,811 fail both, 5,822 pass both, 0 disagreements in either direction. Treat that agreement as a definition rather than a result. The result that stands on its own is the one above it, a $34,808 median benchmark for the 569 under both rules.

What changed is who gets measured, and for how long. The old test scores the 2014 to 2015 and 2015 to 2016 exit cohorts three years after exit; the new one scores the 2017 to 2018 and 2018 to 2019 cohorts four years after exit. Both are CPI adjusted to 2024 dollars, so none of the gap is inflation. The released programs are scored on a later cohort observed for longer. How much of the $7,479 comes from that longer horizon rather than from the labor market or from a change in who finishes, this file cannot settle.

The disappearing debt test contributes, but only at the margin. 513 of the 569 (90.2 percent) had failed the GE earnings test: 425 on earnings alone and 88 on both subtests. Only 56 programs (9.8 percent) are released purely because OBBBA drops the debt-to-earnings test, which under GE fails a program above 8 percent annual or above 20 percent discretionary.

The bar does move. It just does not move for the programs in the head-to-head. Compare the old GE 2023 benchmark against the OBBBA benchmark across the 167,280 programs carrying both and they are identical on 111,719 (66.8 percent), higher under OBBBA on 55,346 (33.1 percent) and lower on 215 (0.1 percent). The quartiles tell the same story: $32,989, $34,808 and $36,082 under the old rule against $34,350, $36,082 and $46,391 under the new one.

Split by credential level and that 66.8 percent resolves cleanly. The 111,719 unchanged programs are exactly the associate and bachelor rows, 42,786 plus 68,933. Not one undergraduate degree program faces a higher bar: 0 of 42,786 associate programs and 0 of 68,933 bachelor programs. Every graduate level does, and the jump is between $16,827 and $19,726 at the median, applying to 99.5 percent of graduate certificates, 99.6 percent of master's programs, 99.7 percent of doctoral programs and 99.9 percent of first professional programs.

All 215 programs whose benchmark went down are graduate level too: 140 master's, 31 doctoral, 1 first professional and 43 graduate certificate.

Undergraduate certificates are 2,335 of the 3,361 programs in the comparable universe, 69.5 percent, so they carry the largest raw count in every cell below, including 393 of the 569 releases. Measured against that 69.5 percent share they only run hot in the failure cells: 940 of the 1,002 programs that fail both tests. The highest release rate sits one level up, at 109 of the 420 associate programs, 26.0 percent, against 393 of 2,335 for undergraduate certificates.

At the institution level, 794 of the 1,482 institutions in the comparable universe (53.6 percent) have at least one program failing both tests, 339 (22.9 percent) have at least one released program, and 17 (1.1 percent) have at least one newly caught program.

The field table ranks by releases and stops at ten rows, so fields tied with the tenth do not appear.

The head-to-head is the wrong lens for the largest shift in this file, which is who gets tested at all. The OBBBA comparison is computable for 49,860 programs across PPD:2026, and 2,880 of them land below their benchmark: 5.8 percent of the testable set and 1.4 percent of all 209,321 rows. Statutorily, 1,220 of those 2,880 are failures and the other 1,660 are out-of-scope certificates. 1,938 of the 4,261 institutions with at least one testable program (45.5 percent) have at least one program below the line.

Split those 2,880 by whether Gainful Employment ever covered the program. Inside the old GE universe, 1,811 of 7,633 testable programs fail, 23.7 percent. Outside it, 1,069 of 42,227 fail, 2.5 percent. The rate is far lower, but the newly covered universe is 5.5 times larger, so those 1,069 programs are 37.1 percent of every below-benchmark program in the file. Institutions that filed nothing under FVT/GE because they ran no GE programs are in scope now, which is the practical reason the final FVT/GE reporting year is not the end of the compliance calendar for anyone.

Only degree programs appear among the newly covered. Every certificate level sits inside ever_ge, graduate certificates as well as undergraduate and post-baccalaureate ones, so the never-GE breakdown has zero certificate rows. Associate degrees fail 424 of 6,962 testable (6.1 percent), bachelor 282 of 24,497 (1.2 percent), master's 343 of 9,101 (3.8 percent), doctoral 11 of 1,015 (1.1 percent) and first professional 9 of 652 (1.4 percent).

The benchmark deciding those newly covered failures is mostly the same kind of bar GE used. 706 of the 1,069 (66.0 percent) turn on a high-school median, 634 on the same-state figure and 72 on the national one. Only 355 (33.2 percent) turn on a bachelor's degree benchmark, so most of the newly exposed programs are losing to a high school comparison rather than to a degree-level standard.

Everything above comes from one table, ppd.fact_program_outcome in the PPD:2026 release published in January 2026, joined to ppd.dim_credential_level for credential splits and ppd.dim_cip for the broad field splits. Variable definitions come from the release's own data dictionary.

The four variables that carry the argument: md_earn_ne_p3_1516 is median earnings of completers not enrolled three years after exit, pooled across the 2014 to 2015 and 2015 to 2016 exit cohorts. md_earn_wne_p4 is median earnings of completers working and not enrolled four years after exit, pooled across the 2017 to 2018 and 2018 to 2019 cohorts, from IRS data. Both are CPI adjusted to 2024 dollars. ge_earn_test_bm_2023 is the median income of 25 to 34 year olds in state holding only a high school degree, from ACS 2023, CPI adjusted to 2024. ever_ge marks any program classified as Gainful Employment in any award year between the 2015 to 2016 and 2024 to 2025 award years.

Several of the caveats below change the answer by an order of magnitude.

Data: ppd.fact_program_outcome, ppd.dim_credential_level, ppd.dim_cip and the PPD:2026 data dictionary, release_year 2026. The full reproducible query log for every figure in this post, section by section, lives in the repository at docs/data-evidence/one-file-two-earnings-tests-different-casualties/queries.md, with the column-level schema notes and verified behaviours in schema.md alongside it.

Policy context: U.S. Department of Education, Federal Student Aid Partners, Financial Value Transparency and Gainful Employment knowledge center. FVT/GE Frequently Asked Questions. Negotiated rulemaking for higher education, 2025 to 2026.