Why your program has no earnings number: 76.2 percent of PPD:2026 is dark, and what to put in the memo instead

Three out of four programs in the country have no published earnings number. In PPD:2026, 159,461 of 209,321 programs carry a blank `md_earn_wne_p4`, which is 76.2 percent of the file. Only 49,860 programs, 23.8 percent, have a median earnings figure at all.

The same 159,461 rows also carry no federal earnings-test result. `mstr_obbb_fail_cip2_wageb` is the wider of the two OBBBA result flags, the one that runs the comparison at every credential level whether or not the statute reaches it, and it is null on exactly those rows, with zero mismatches in either direction. There is no partial credit and no provisional result. A missing earnings figure and a missing accountability verdict are the same event.

So when the provost asks what graduates of a specific certificate earn and the analyst comes back with nothing, that is the base rate, not a failed search. The work worth doing is figuring out which of five specific things happened to the cell, then reaching for the fallback that fits.

Six files, 209,321 programs, three join keys covers the suppression rules at the level of "here is what the documentation says." This piece measures them: the floors the file actually honors, the band midpoints that get read as counts, and how much of the darkness a second dataset can recover. For the statutory side of the test itself, the OBBB to Earnings Premium walkthrough already does that job, and we will not repeat it here.

Two datasets, queried directly, 29 queries logged. The first is PPD:2026, the Program Performance Data release published in January 2026: 209,321 programs across 5,096 institutions, 411 CIP-4 fields and 8 credential levels. The headline earnings variable, `md_earn_wne_p4`, is CPI-adjusted to 2024 dollars and comes from IRS records for pooled 2017-18 and 2018-19 completers, measured four years after exit.

The second is the Census Bureau's Post-Secondary Employment Outcomes file, built from state unemployment-insurance wage records at 1, 5 and 10 years after graduation, in 2023 dollars. PSEO covers 32 states and 950 institutions, so nothing derived from it is a national figure and we never call it one. PPD is national in scope, covering all 59 jurisdictions.

"Dark" throughout means one thing: no published median earnings value in the cell. Not zero earnings, not an error in your join. A blank.

There is no credential level where you can expect to find an earnings number. The best-covered category in the entire file is the First Professional Degree, and even there 59.5 percent of programs are blank. The worst is the Post-baccalaureate Certificate, where 85 of 3,469 programs have a figure and the other 97.5 percent do not.

The ordering runs backwards from the usual assumption that graduate credentials are the better covered ones. Graduate certificates (95.8 percent dark) and doctoral degrees (90.9 percent dark) are among the darkest things in the file, worse than undergraduate certificates and associate degrees.

PPD carries three earnings measures, not one, and letting a program qualify on any of the three barely moves the picture. Counting all three vintages, post-baccalaureate certificates are 96.9 percent dark instead of 97.5, graduate certificates 94.3 instead of 95.8, doctoral 89.5 instead of 90.9, undergraduate certificates 82.8 instead of 85.1, associate 78.2 instead of 82.2, master's 69.7 instead of 72.1, bachelor 60.4 instead of 63.8, and first professional 54.4 instead of 59.5. A few points each. Nothing structural.

Across the 36 CIP-2 families with 500 or more programs, the best-covered field in the country is 62.0 percent dark. That is Personal and Culinary Services, with 2,994 programs. The worst is Philosophy and Religious Studies at 94.1 percent, with 2,470 programs. Every single large field sits inside that band.

The fields where a provost is most likely to ask sit right in the middle of it. Computer and Information Sciences: 83.1 percent dark across 11,046 programs. Education: 74.3 percent across 16,225. Business: 70.3 percent across 24,024. Health Professions, the largest family in the file at 29,246 programs, is among the better-covered ones at 66.2 percent dark, and two thirds of it is still blank.

No jurisdiction in PPD:2026 comes in under 66.7 percent dark, and the observed minimum sits above that line rather than on it. The best-covered anywhere is New York at 69.0 percent across 11,541 programs, followed by Wisconsin at 70.8, Rhode Island at 70.9, Florida and Virginia both at 71.3, Minnesota at 72.6, New Jersey at 73.0 and Massachusetts at 73.8. The darkest of the 50 states is Alaska at 87.4 percent across 641 programs, then Montana 85.9, Wyoming 85.5, and New Mexico and California both 83.3. Several territories and freely associated states run darker than any state, up to the 100 percent that is the highest value anywhere. California is the striking case: 23,578 programs, the largest state universe in the file, and five out of six of its programs have no earnings figure.

Sector barely separates anything either, with one exception. Private non-profit institutions are 78.1 percent dark across 68,058 programs, public institutions 76.5 percent across 133,507 programs, and for-profit institutions 53.4 percent across 7,756 programs. For-profits are the best-covered sector in the country by more than twenty points, on the smallest program footprint. The file does not say why. A shorter catalog of larger programs would produce exactly this.

Institution size cuts the opposite way from the intuition, and getting it backwards is how bad memos get written. Per program, small colleges are covered better than large ones: colleges running 1 to 5 programs have 54.1 percent of those programs dark, while colleges running 31 to 60 programs have 80.8 percent dark. A tiny college usually runs one or two big programs, and big programs clear the floor.

The penalty for being small is not at the program level. It is at the institution level, where it is brutal.

Nearly a third of the smallest colleges, 589 of 1,894, have no published earnings figure for any program they run. Not one. Across the whole file, 835 of 5,096 institutions (16.4 percent) are in that position. If you work at one of them, every program-level earnings question you receive has the same answer, and you should say so once, in writing, rather than re-running the search each time it is asked.

Counting all 5,096 institutions, those 835 included, the median institution has a published earnings figure for exactly 25.0 percent of its own programs. The mean is 35.4 percent, pulled up by the large systems.

A blank is not one thing. Five different situations produce the same empty cell, and each one points at a different fallback. Work down this list in order. The first branch that matches is your answer, and each branch has a sentence you can put straight into the memo.

A sixth reason gets offered constantly in meetings: the program is blank because it is not gainful-employment eligible. The data says the opposite. Programs ever classified as GE are worse covered than programs never classified as GE, not better: 14.1 percent of GE programs have a current earnings figure (7,633 of 54,063) versus 27.2 percent of never-GE programs (42,227 of 155,258). On the gainful-employment three-year measure it is 6.9 percent against 14.3 percent.

GE status marks the kind of program that tends to get suppressed (short and small), so it travels with the blank rather than causing it. Drop it from the decision tree.

The five floors below are what we could observe by scanning the loaded data for the smallest value that ever appears. They are a description of what the file does, not a citation of regulatory text. They are also unusually clean as minima: nothing below any of these floors appears anywhere, zero exceptions, which is how you know you are looking at a rule rather than a coincidence. A floor is not a publication guarantee, though. Clearing one is necessary, not sufficient, as the 156 programs in branch three show.

Put every program in the file into exactly one bucket, working from most informative to least, and the shape of the problem is blunt. More than two thirds of the national program universe, 145,221 programs or 69.4 percent, is fully dark: no earnings figure on any of the three measures, no debt median, not even a cohort headcount. There is nothing in the row to fall back to.

The middle rows are the interesting ones for practice, because they are the recoveries. Falling back to an older completer cohort inside PPD is legitimate and it works, a little: `md_earn_wne_p4_1516` is published on 42,005 programs and `md_earn_ne_p3_1516` on 25,942, and using all three measures together lifts coverage from 49,860 programs to 56,042, from 23.8 percent to 26.8 percent. That recovers 6,182 programs for the cost of one extra column, and it still leaves you dark on 73 percent of the file.

The 159,461 match is exact, and the exactness is the point. Programs are not being scored on partial evidence or given a provisional result pending more data. Where the earnings figure is missing, the master pass/fail flag is missing, and the two sets are identical to the row.

For anyone assembling an institutional exposure picture, this changes the arithmetic of the denominator. A dashboard that shows "programs failing" over "programs in the file" understates risk, because three quarters of the denominator was never tested. The honest presentation has three columns: passing, failing, and not evaluated. If you are building that view for the first time, the exposure-map playbook walks through the triage order.

Debt coverage is thinner than earnings coverage: 39,639 programs of 209,321, or 18.9 percent, have a published median debt figure. The threshold is the sharpest in the whole file. Median debt is published whenever `debt_all_stgp_eval_n` is 19 or more and withheld whenever it is 18 or fewer, with zero exceptions across the 51,301 rows that carry a debt count. That means 14,690 programs tell you how many borrowers they had and refuse to tell you what those borrowers owed.

Combine the two sides and the most requested ratio in program review mostly cannot be computed. Only 38,027 programs, 18.2 percent of the file, have both a debt median and an earnings median. 11,833 have earnings but no debt, 1,612 have debt but no earnings, and 157,849 have neither. Fewer than one program in five supports a debt-to-earnings calculation.

One more trap when you do have both. PPD debt is CPI-adjusted to 2019 dollars while PPD earnings are adjusted to 2024 dollars. They are in the same table, next to each other, in different money. Do not put them in the same sentence without saying so, and do not build a ratio that quietly treats them as comparable.

When you hit a blank, the next thing to try is a different publisher. Our overlap universe is 90,238 program cells: the institutions PSEO reports on inside its 32 states, not every program in those states. PPD alone lights 19,192 of them, 21.3 percent. Add PSEO's 1-year median and 37,256 are lit, 41.3 percent. That is a 94 percent relative increase in the number of programs you can say anything about, and 18,064 cells have a PSEO figure where PPD is blank.

Two warnings before you use that. First, PPD and PSEO are not the same measurement. PPD earnings come from IRS records four years after exit for pooled 2017-18 and 2018-19 completers in 2024 dollars. PSEO earnings come from state unemployment-insurance wage records at 1, 5 and 10 years after graduation in 2023 dollars. A program lit in one file and dark in the other has been measured once, differently. It has not been confirmed twice.

Second, most published figures cannot be cross-checked at all. Only 14,608 of the 90,238 cells (16.2 percent) carry a number from both sources. 4,584 are PPD-only, and 52,982 (58.7 percent) are dark in both. Even with two large publishers pointed at the same question, and even inside the institutions both of them report on, more than half of that universe stays blank.

The payoff is also lopsided, and it is concentrated exactly where undergraduate program review lives.

Undergraduate certificates go from 7.8 percent lit to 36.0 percent across 18,550 cells. Associate degrees go from 15.4 to 45.3 across 20,837. Bachelor degrees go from 40.4 to 65.6 across 25,408. First professional degrees go from 31.3 to 58.8 across 723.

Above the bachelor's line, the second source buys you nothing. Master's degrees stay at 24.3 percent across 14,473 cells. Doctoral degrees stay at 6.8 across 5,430. Graduate certificates stay at 3.8 across 3,378. Post-baccalaureate certificates move from 2.2 to 2.6 across 1,439, which is noise.

The reason is structural rather than statistical. PSEO does not publish master's or doctoral-research earnings at the CIP-4 grain at all: 48,849 master's and 18,442 doctoral-research institution-level CIP-4 rows are flagged privacy-suppressed and null across every cohort, and adding the state-level rows makes 96,340 rows with zero published values. Those rows are empty by construction, not by cell size. None of the 67,291 institution-level rows carries an IPEDS graduate count either. Waiting for a bigger cohort will not fill them.

PSEO is more honest than PPD about why a cell is empty: it tells you, with a status flag. Getting the vocabulary right matters, because there are two published versions of it and only one is correct for this data. The `status_flag` domain in the PSEO data dictionary claims 1 means suppressed. In the loaded database it is the reverse: every one of the 34,275 code-1 cells carries a median, and every code-5 cell is null. Use the table below and ignore the dictionary domain. The readings for codes 5, -1, 3 and 4 come from how those codes behave in the loaded data, not from the published domain.

The working slice here is 75,772 program cells: institution level, CIP-4, all cohorts pooled, the PSEO grain closest to a PPD program. At 1 year after graduation, 34,275 cells are published, 45.2 percent. At 5 years, 31,160, or 41.1 percent. At 10 years, 22,809, or 30.1 percent.

Coverage falling with the horizon looks like heavier suppression and partly is not. At 10 years, 35,424 cells (46.8 percent) are privacy-suppressed and another 17,539 (23.1 percent) carry code -1, meaning the cohort simply has not been observed ten years out yet. At 5 years the code -1 count is 6,587 (8.7 percent), and at 1 year it never appears. A code -1 cell is a "come back later." A code-5 cell is not.

Separating those two failure modes changes how good PSEO looks. 16,872 of the 75,772 pooled cells (22.3 percent) are dark because the publisher does not release that degree level at that CIP grain, full stop. Of the 58,900 cells PSEO actually attempts, 34,275 are published, 58.2 percent, and 24,625 are suppressed on cell size. PSEO publishes most of what it tries to publish. It just tries on less than you would like.

Participation is uneven in a way that will bite anyone comparing states. PSEO holds 621,941 rows across 32 states and 950 institutions, but Texas contributes 110 institutions while Michigan and Idaho contribute one each, and Alabama and Arizona three each. The 1-year published share ranges from 27.6 percent (Michigan, one institution, 330 cells) to 67.7 percent (Idaho, one institution, 93 cells), and no state publishes 70 percent of its CIP-4 program cells. Among states with real institutional coverage: South Carolina 56.4 percent across 54 institutions, New York 52.2 across 79, Virginia 49.2 across 67, Minnesota 48.2 across 61, Texas 47.0 across 110, Illinois 45.1 across 79, Ohio 43.4 across 38, Indiana 42.5 across 62, Georgia 41.1 across 35, Massachusetts 38.4 across 28, and Pennsylvania 29.2 across 24. Read the institution count before you read the percentage.

You cannot manufacture a suppressed median. You can write a memo that answers the underlying question, which is almost never "what is the median" and almost always "is this program worth continuing." Five things go in it, in this order.

The version of this memo that gets a decision made says four things in four sentences. Here is why the cell is blank, named from the five branches. Here is the nearest defensible number and what it actually measures. Here is the cohort size, because size is what the question was about. And here is the one thing we can say with certainty: this program has no federal determination, so it cannot be counted as passing.

That memo takes fifteen minutes if you already know the suppression structure. It takes most of a week if you have to rediscover it, program by program, every time the question is asked in a slightly different shape.

Everything above comes from direct queries against two loaded datasets, logged so they can be re-run. The full query set, including the analyses we discarded and why, is in the evidence files listed under Sources.

PPD:2026 is release year 2026, published January 2026. Earnings are from pooled 2017-18 and 2018-19 completer cohorts, four years after exit, from IRS records, CPI-adjusted to 2024 dollars. Debt is CPI-adjusted to 2019 dollars. The unit of analysis is the program, defined as the combination of institution, credential level and CIP-4 code, 209,321 of them.

PSEO is the pooled all-cohort view (grad cohort 0000, latest cohort 2019), earnings in 2023 dollars, from state unemployment-insurance wage records. It covers 32 states and 950 institutions. No PSEO figure in this piece is a national figure and none should be described as one.

Four things these numbers do not say:

Datasets: U.S. Department of Education, Program Performance Data (PPD:2026) and U.S. Census Bureau, Post-Secondary Employment Outcomes. PSEO status-code definitions and the cell-size rule are from the PSEO data documentation.

Reproducible query log: `docs/data-evidence/why-your-program-has-no-earnings-number/queries.md` (29 numbered queries plus the discarded analyses and why they were dropped) and `docs/data-evidence/why-your-program-has-no-earnings-number/schema.md` (table and column definitions, vintages, and the status-code contradiction).

Related reading on this site: Six files, 209,321 programs, three join keys, From OBBB to Earnings Premium, and the program-level exposure map.