Federal completions data · IPEDS · 2013–2021
Between 2013 and 2021, 48,073 of the 169,578 college degree and certificate programs conferring awards stopped conferring them. The explanation most people reach for, colleges going out of business, covers less than a third of it. At schools that stayed open, one program in five stopped.
Look up what your school stopped conferring
169,578 → 171,337 total programs. Up.
48,073
of them: gone. More than one program in four, out of the 169,578 conferring in 2013, was awarding nothing by 2021 (95% CI: 28.1–28.6%).
Gone means stopped conferring: each of these programs awarded at least one degree or certificate in 2013, then reported zero awards in 2021, 2022, and 2023. This measures observed award activity, not registrar paperwork. A program can sit in a catalog for years after its last graduate.
Act I · The official story
Add up the degree and certificate programs American colleges offer, every year, from the federal completions census. The line drifts up: from 169,578 programs in 2013 to 171,337 in 2021, a net gain of 1,759. Nothing to report.
Every net number you have read about program offerings looks like this line. The line is not wrong. It is hiding something.
Act II · What the net conceals
The flat line hides two large opposing flows. Every year colleges started thousands of new programs, and every year thousands of existing programs conferred their last award, 10,387 births against 6,137 deaths in 2016 alone. The net stays close to flat. The churn is the story.
In 2019 the lines crossed for the first time: 8,611 programs conferred their last award while 8,509 new ones appeared. They crossed again, harder, in 2021.
Births in 2013 are not observable: the panel begins there, so every program present in 2013 is left-censored. 6,095 programs conferred their last award in 2013 itself; they are in the cohort figures but not this chart, which starts at 2014 so each year shows both a birth and a death count.
Early birth counts run high partly for a mechanical reason: a program dormant in 2013 alone counts as a 2014 birth, because no earlier years exist to check against, while a 2019 birth requires six years of prior absence. A first award also only counts as a birth if the institution was already reporting the year before, so a school entering the panel is one arrival, not N program launches. The 2019 crossing is robust to both effects. The shown death bars carry the guarded per-year rule, which excludes programs at departing institutions, so the two series are not expected to net to the totals in Fig. 1.
Every light is one real program. One institution, one field of study, one credential level. These are the 169,578 conferring in 2013.
By 2016, 22,211 had gone dark. Almost never an announcement. The federal filings just stopped listing awards, one program at a time.
By 2018, 32,559. The frontier takes about four percent of the survivors every year.
By 2021, 48,073. More than one light in four, out.
One in four is the national rate. It averages over outcomes that do not resemble each other. Among private for-profit four-year colleges, 77.1 percent of the programs conferring in 2013 had stopped by 2021. Among public four-year universities, 16.8 percent had. Same measure, same years, a gap of four and a half times.
Most of that gap is a second event hiding inside the first. By 2021, 490 of the 735 for-profit four-year institutions in the 2013 data had stopped filing with the Department of Education at all, through closure, lost aid eligibility, or a merger into another reporting ID. Their programs stopped because their schools did. Among public four-year universities, 21 of 708 stopped filing. Compare only the colleges still open at the end of the window and the spread narrows to 37.9 percent against 15.9. Narrower, and still more than double.
Partly, and the data says exactly how much. Of the 48,073, some 14,824 were at institutions that stopped reporting to IPEDS entirely. That covers closures and lost federal aid eligibility, but also multi-campus systems consolidating their reporting under one ID, in which case some of those programs continued under a different number. The count of reporting institutions fell from 7,176 to 5,871 over the window.
Which is exactly why the restricted figure matters more. The other 33,249 died at schools still reporting at the end of the window (2021 to 2023). Restrict both sides of the fraction to those schools: 33,249 of their 154,754 programs from 2013, or 21.5 percent, had stopped conferring. Throw out every closed and consolidated college in America, and a fifth of the remaining programs still stopped. That is the number this page stands on.
A data bug produces cliffs. A taxonomy change produces spikes. This curve is smooth and monotonic for eight straight years, through a federal taxonomy break and a pandemic: after a 5.6 percent first year, the 2013 cohort loses a steady three and a half to four and a half percent of its survivors every year. Regularity that steady is hard to fake and easy to break; a single mishandled file format would bend it.
A second implementation of the same measure, built separately during this project from the same raw files with no shared code, put the 2013 to 2018 loss at 20.4 percent. This pipeline gives 19.2 percent for the same window. The one-point gap is a fair reading of how much the definitional choices are worth.
Certificates churn like inventory: two-to-four-year certificates lost 57.1 percent, one-to-two-year certificates 48.6. Associate's degree programs lost 34.9 percent. Bachelor's, 20.9. Research doctorates, 14.9, the most durable credential on the chart. The one exception proves the size rule: "Doctorates, other" churned like a certificate, on a base of just 128 programs.
Certificates shorter than one year appear nowhere on this page: IPEDS recoded that award level in 2020 in a way that cannot be reconstructed without guessing (Trap 2, below). They were about one in ten programs before the recode, and probably the most volatile tenth, so the headline is if anything an undercount.
Communications technologies lost 46.6 percent of its 2013 programs. Engineering technologies, 44.9. Computer and information sciences lost 42.2 percent, in the same years tech hiring boomed. Some of that is certificate churn, so flip the toggle: computer science degree programs alone still lost 38.7 percent.
At the other end, the old academic core barely moved: physical sciences lost 7.9 percent, mathematics and statistics 8.8, history 9.0. And the largest absolute toll was in health professions: 8,629 programs stopped, out of 23,483 alive in 2013.
A program here is one institution offering one credential level in one field. These are counts of offerings, not of students or graduates.
The median school lost 21.4 percent of its 2013 catalog. All but 50 of the 2,834 lost at least one program. This was not a handful of schools in crisis. It was the whole system, quietly.
Every institution in the 2013 cohort, with the programs named. Counts, not rankings: a ranking on raw counts would just reward big catalogs, so schools appear with their catalog size, and a rate is shown only where the 2013 base is at least ten programs. A program is only named as stopped when it survives every check described in the method.
Type at least two characters, then use the arrow keys and Enter. Selecting a school gives you a link you can share.
Count everything: 28.3 percent. Count only degrees, no certificates: 24.5. Count only bachelor's degrees and above: 20.7. Count only schools that never left the dataset: 21.5.
Then there is the screen this project applies to every program it names at a school, and it belongs on the headline too. When an institution files a row with the federal government reporting zero graduates, it is telling the government that the program exists. This panel cannot see that row, because it requires at least one completion, so a program filed at zero reads exactly like a program that stopped. It is the strongest single test available in this data, and 13,348 of the 48,073 fail it.
Those 13,348 are genuinely ambiguous and nothing in this data settles them: the school says the program exists, the completions file says nobody finished it. So there are two defensible ways to score them, and this page shows both. Count them as programs that never stopped and the finding is 34,725 of the same 169,578, or 20.5 percent. Set them aside as unresolved and the finding is the same 34,725 out of 156,230, or 22.2 percent. Same numerator either way; the disagreement is only about the denominator, and it is worth 1.7 points. That is the whole range, and 20.5 is the floor.
However you cut it, roughly one program in four or five stopped conferring, and no restriction makes it go away.
The standing objection to counting program deaths this way is that a missing year in a federal file can mean the registrar reported nothing, not that the program ended. It is a real objection and it has a name: Robert Kelchen measured it in 2023 and found roughly 31 percent of apparent discontinuations later showed graduates again. That was against a rule that called a program dead after one missing year. This page requires three.
So the honest question is not whether the objection is valid. It is what the error rate of this rule is, and that is answerable from the data itself. Pick a decision year in the past, apply the rule using only what was known then, and look at what actually happened next.
This 4.6 percent and the still-filed screen further down are not independent: a program its school still files at zero is also likelier to confer again. The two corrections partly measure the same thing, so they do not compose, and neither one should be subtracted from the other's result.
Five other specifications of this estimator were tried and rejected, and they disagreed with each other by a wide margin, which is exactly why the one used here is pinned in the code and checked on every build rather than quoted from memory.
Act III · The method
The data is public, free, and mandatory: every Title IV institution files its completions with the Department of Education every year. The number on this page required no access anyone lacks. It required surviving five traps, each of which produces a confidently wrong number while raising no error at all.
IPEDS switched classification systems (CIP‑2010 to CIP‑2020) in 2020: 149 program codes were renumbered, 544 added, 12 deleted. A renumbered program that continued uninterrupted reads as a death under the old code and a birth under the new one. The crosswalk that fixes this is not fetchable by URL; it comes from an ASP.NET postback on the federal site. Meanwhile a different, wrong crosswalk (2000 to 2010) sits at a guessable URL in plain CSV. The trap is not that the fix is hidden. It is that a decoy fix is in plain sight.
Award level 01, certificates under one year, runs 18,380 to 21,254 programs across 2013 to 2019, then hits exactly zero in 2020. It did not shrink; it was split into two new codes, 2,080 plus 19,594, under an undocumented rule. Uncorrected, this alone manufactures roughly 21,000 false deaths in 2019 and 21,000 false births in 2020: a 15.8 percent death rate against a 4 percent baseline. This page does not guess the split. It drops the level and says so.
Provisional files write award level '2' where revised files write '02'. A string join between years silently matches nothing. Every code is normalized to two digits before any comparison.
Later files carry a UTF-8 byte order mark that earlier ones lack. Read with the wrong encoding, the first column header becomes UNITID and every lookup misses, silently.
Sum this panel's per-year deaths and you get 63,774. Divide by the 2013 base and the rate inflates without limit as the panel lengthens, because programs born after 2013 die into a numerator they were never in the denominator of. The honest cohort figure is 48,073, with a fixed denominator. We made this mistake during the build, caught it, and kept the receipts. The difference between those two numbers is the difference between this page and a retraction.
The most prominent public tracker of college program cuts, CollegeCuts, is compiled by hand from news reports, press releases, and crowdsourced tips, and it does that job well. As of August 2026 it records 331 cuts of every kind since 2024, 137 of them program or department closures. Announcement-based tracking and observation-based measurement see different things: the method on this page found 48,073 stopped programs over 2013–2021 because it does not need anyone to announce anything.
Unit of observation. One program is one institution (UNITID) offering one field of study (CIP 2020) at one award level, counting first majors with at least one completion. All years are normalized to the CIP‑2020 taxonomy via the official crosswalk. Revised files only; a provisional year is incomplete by construction and would manufacture deaths.
The cohort rule. The headline follows the fixed set of 169,578 programs conferring in 2013. Stopped means no awards observed in any of 2021, 2022, or 2023. The cohort figure includes programs at institutions that left IPEDS entirely; the decomposition above separates them (14,824 of 48,073).
The per-year rule. Yearly birth and death series carry further guards: a death requires at least two award-free years after the last award, and the institution must still be reporting in every one of those years. A school that stops filing is one departure, not N program deaths. Symmetrically, a first award counts as a birth only if the institution was already reporting the year before, so a school entering the panel is one arrival, not N launches. An institution counts as still reporting if it appears in the panel in any of 2021 through 2023.
Reappearance is not death. 56,745 programs, 20.3 percent of everything in the panel, went dark for at least one year and then returned. None of them count as dead here. A death on this page means the lights never came back on across at least two full later years.
The residual limitation, and what it costs. A program that files a row reporting zero completions is treated the same as one that files nothing: both count as not conferring. If a program admitted students but graduated none for three straight years, this page calls it stopped; a registrar might not. That cost is measurable, and it is measured. Of the 48,073, 13,348 are rows the institution was still filing at zero in 2021, 2022 or 2023. Screen every one of them out and the finding is 34,725, which is 20.5 percent of the full cohort or 22.2 percent of what is left once the disputed rows are set aside. Both readings are shown as the last two rungs of the robustness ladder above. The same test decides which programs this page is willing to name at a school, where it is applied to every row.
Known exclusions. Certificates under one year (award level 01) are excluded throughout, pending the undocumented 2020 recode: they were 9.8 to 11.1 percent of all programs across 2013–2019. All eleven years, 2013 through 2023, use revised files; the revised 2023 file posted in late 2025 and serves as a trailing observation year. A year available only in provisional form would be excluded automatically. Deaths occurring in 2022 and 2023 cannot yet be assessed (they would need 2024 and 2025 data) and are shown nowhere on this page.
Uncertainty. Intervals are 95 percent Wilson score intervals. Since every Title IV institution must file, these counts are population figures and exact as descriptions; the intervals treat each program's fate as a draw from an underlying stopping process, and the honest error bar on the headline is the robustness ladder, not the interval. Cross-institution comparisons are defensible under a shared definition; absolute institution rankings are not, which is why the lookup shows counts with bases and never a leaderboard.
The derived data behind each figure ships with this page as plain JSON, and every file reconciles to the headline by scripted assertion before anything is published.
Raw source: IPEDS Completions (C2013–C2023, revised files) and the NCES CIP crosswalk. The pipeline that builds the panel is verified by a 25-check regression suite covering every trap above, and refuses to build on a truncated archive.