Pith. sign in

REVIEW 3 major objections 6 minor 49 references

College closures from 2020 to 2025: An exploratory analysis and its implications for the enrollment cliff

T0 review · 3 major / 6 minor · reviewed 2026-07-10 · grok-4.5

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Pandemic-era college closures clustered in the Northeast and Midwest and hit many kinds of schools, not one vulnerable type.

desk verdict Solid hand-built descriptive study of 65 pandemic-era closures; caveats are honest, cliff implications stay soft, and the work is worth a referee. read the letter →

arxiv 2607.07917 v1 pith:OU3JFBZI submitted 2026-07-08 stat.AP

Sepehr Akbari , Shahrzad Jamshidi This is my paper
classification stat.AP
keywords CollegeClosuresHigherEducationEnrollmentCliffBayesianhierarchicalmodelsInstitutionalclusteringMediatopicmodelingReligiouslyaffiliatedcollegesDemographicdecline
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper treats the modest wave of U.S. college closures and mergers from 2020 through 2025 as an imperfect early look at institutional vulnerability ahead of the demographic enrollment cliff. The authors assemble 65 closed or merged institutions, pair them with state and regional demographic, economic, and financial indicators, and add a corpus of news coverage. Bayesian models show higher estimated closure risk in the Northeast and Midwest, matching prior geographic projections, while the absolute numbers stay small. Dimensionality reduction, clustering, and topic modeling then show that the closed schools were heterogeneous: financial structure, regional demographics, and mission each helped distinguish them, and media coverage repeatedly paired financial strain with religious affiliation. The work is framed as exploratory and descriptive, and it asks what these patterns may and may not imply for colleges facing a gradual drop in traditional college-age students rather than a sudden pandemic shock.

What carries the argument

A hand-assembled dataset of 65 closed or merged institutions, analyzed with Bayesian hierarchical models of state and regional closure risk, then Linear Discriminant Analysis, PCA, k-means clustering with SHAP, and topic modeling of news coverage, used to describe geography, shared traits, and public narratives among the closed schools only.

What would settle it

Build a matched sample of comparable institutions that stayed open through 2025 and test whether the same financial, demographic, and mission features that separate clusters among the closed schools also separate closed from surviving schools; if they do not, the observed profiles do not preview cliff vulnerability.

Watch

Extended reading notes

Core claim

Among 65 U.S. institutions that closed or merged between 2020 and 2025, closures were more frequent in the Northeast and Midwest, yet the closed institutions were not a single type. Financial structure, regional demographics, and institutional mission each contributed to distinguishing them, and news coverage repeatedly linked financial viability with religious affiliation. The authors treat these pandemic-era patterns as a limited, imperfect preview of vulnerability under the enrollment cliff, not as proof that the cliff will select the same schools.

Load-bearing premise

The paper rests on treating pandemic-era closures as a useful if imperfect analogue for who will be vulnerable under the slower demographic enrollment cliff, even though the two shocks work by different mechanisms.

Editorial extensions

If this is right

  • Geographic risk of closure is uneven: Northeast and Midwest states show higher estimated rates than the South and West, though absolute counts remain small.
  • No single institutional profile defines the closed set; financial structure, regional demographics, and mission each matter, so responses should vary by profile rather than follow one generic template.
  • Media and official narratives that reduce closure to a single label such as 'financial' or 'enrollment' obscure real differences among schools.
  • Religiously affiliated private colleges appear frequently in public closure stories as a financial-and-identity pairing, making mission-preserving consolidation one concrete option some institutions may weigh.
  • Policymakers and boards can use regional risk estimates and cluster-defining features (for example, reliance on state appropriations versus a shrinking local birth pipeline) as rough planning signals, not as forecasts of individual failure.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the cliff selects partly different institutions than the pandemic did, the shared Northeast-Midwest geography may mainly reflect concentration of small tuition-dependent private colleges rather than pure demographic exposure.
  • A matched open-versus-closed sample is the single highest-value next study; without it, the archetypes remain descriptions of who closed, not predictors of who will close.
  • The paper's own ethical caution against public 'at-risk' labels suggests that any operational use of these methods should stay descriptive and institution-facing rather than become a public ranking of failure risk.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper assembles a hand-curated dataset of 65 U.S. postsecondary institutions that closed or merged between 2020 and 2025, pairing institutional attributes with state- and regional-level demographic, economic, and financial indicators and a news corpus. Using non-hierarchical and hierarchical Bayesian models for state/regional closure probabilities, linear discriminant analysis on stated closure reasons, PCA, k-means clustering with SHAP attributions, and LDA topic modeling of media coverage, it describes geographic concentration, heterogeneity among closed institutions, and public narratives. The central descriptive claim is that closures were more frequent in the Northeast and Midwest (consistent with Grawe’s demographic projections), that closed institutions were heterogeneous along financial, demographic, and mission dimensions rather than uniform, and that media coverage recurrently paired financial strain with religious affiliation. Results are explicitly framed as exploratory and as an imperfect analogue to the enrollment cliff.

Significance. If the descriptive patterns hold, the paper supplies one of the few systematic empirical portraits of the modest pandemic-era closure wave and situates it carefully against the slower demographic enrollment cliff. Strengths include full posterior uncertainty for the Bayesian risk estimates, repeated and explicit acknowledgment of the missing surviving-institution comparison group and of small-n/high-dimensionality constraints (§2.2), silhouette-guided cluster choice, public code and data, and multi-method triangulation (Bayesian geography, discriminant structure, clustering/SHAP, topic modeling). Even without causal claims, the work is useful for higher-education policy and institutional planning discussions and for motivating a matched comparison sample in future work.

major comments (3)
  1. [§5.2 Segments & Institutional Archetypes] §5.2 (and the silhouette-guided choice of k=21): With n=65, twenty-one clusters leave roughly three institutions per cluster on average, and several clusters are smaller. The manuscript already flags this, yet the subsequent qualitative archetypes (“financially stable research universities,” “smaller liberal arts colleges under financial pressure,” etc.) and the SHAP-based feature interpretation still treat the segments as substantively meaningful. Either reduce k to a more stable range with reported stability diagnostics (bootstrap or consensus clustering) or reframe the entire section as purely illustrative exploratory grouping without archetype labels.
  2. [Abstract; §§4–5; §6.1] §§4–5 and Abstract: LDA, PCA, k-means, and SHAP are computed exclusively on the closed institutions. The paper correctly states in §2.2 that these analyses describe variation among closed institutions and cannot identify features that separate closed from surviving ones. Nevertheless the Abstract and §6.1 still summarize “what the closed institutions had in common” and “heterogeneous rather than uniform” in language that readers will naturally read as risk factors. Tighten every such claim to “among institutions that closed” and move any implication language about vulnerability under the enrollment cliff into a clearly labeled speculative subsection.
  3. [Introduction; §3.2; §6.1] Introduction (paragraphs distinguishing pandemic vs. cliff) and §3.2/§6.1: The imperfect-analogue premise is the softest load-bearing link for any “implications for the enrollment cliff.” The authors carefully distinguish acute liquidity shock from multi-year demographic erosion, yet still draw regional and profile implications. Either supply additional evidence that the same institutional dimensions (tuition dependence, regional birth-rate exposure, thin endowments) are the operative channels under both shocks, or further de-emphasize cliff implications so that the paper’s contribution stands cleanly as a descriptive study of 2020–2025 closures.
minor comments (6)
  1. [Figure 1 / Introduction] Figure 1 caption and surrounding text: the map is described as “based on data in Grawe [2018]” but the precise construction (which HEDI series, which years) is not stated; a short methods note or appendix table would help reproducibility.
  2. [§3.1] §3.1: the four-layer hierarchical model introduces a latent x_s and scalar k ~ HalfNormal(0,0.5) without reporting posterior predictive checks or sensitivity to the HalfNormal scale; a brief robustness note would strengthen confidence in the shrinkage results shown in Figures 4–5.
  3. [§4.1] §4.1 LDA: explained-variance ratios are given, but the number of features retained after the manual reduction (stated as 66 in §5.1) is not restated here; clarifying the feature count used for the discriminant analysis would help readers assess the “many features relative to observations” caveat the authors themselves raise.
  4. [§4.2] Topic-modeling section (§4.2): the number of topics (four) is mentioned only in the Figure 7 caption; state the model-selection criterion (perplexity, coherence, or researcher judgment) in the text.
  5. [Throughout] Minor typographic and consistency issues: “Theenrollment cliff” missing space (Introduction); “as and bs” should be a_s and b_s; “xs ∼N(µ,σ)” formatting; “Preprint” running header appears throughout; ensure all URLs and access dates in the reference list are consistent.
  6. [§6.3] §6.3 Ethics of Predictive Modeling is thoughtful but slightly digressive relative to the descriptive scope; consider shortening or moving part of it to an appendix so the Discussion stays focused on the empirical findings.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: purely descriptive Bayesian updates, clustering, and topic modeling on observed closures; no fitted-input-as-prediction or self-definitional reduction.

full rationale

The paper’s chain is observational and exploratory, not a derivation of a first-principles or predictive claim from inputs that already encode the result. Bayesian models (non-layered Beta(1,1)/Beta(2,2) and hierarchical BHMs) update standard conjugate or weakly informative priors with observed state/region closure counts y and institution totals n; the posterior means are ordinary shrinkage estimates of the same quantity (closure probability), not a redefinition of that quantity in terms of a fitted parameter that is then re-labeled a prediction. LDA, PCA, k-means (k=21), SHAP, and LDA topic modeling are applied only to the 65 closed institutions (or their news corpus) and are presented as descriptions of variation among those institutions; the authors explicitly disclaim comparison to survivors and refuse predictive labeling. No uniqueness theorem, ansatz, or load-bearing result is imported from prior work by the same authors. Citations (Grawe 2018, Kearney et al., IPEDS, etc.) supply external demographic context or data sources and are not used to force the paper’s descriptive conclusions. The imperfect-analogue framing of pandemic closures versus the enrollment cliff is an interpretive caveat, not a circular step. Score 0 is therefore the correct finding.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The paper is empirical and descriptive; its load-bearing content rests on data-assembly choices, standard statistical models, and the interpretive premise that pandemic closures can inform (imperfectly) enrollment-cliff vulnerability. Free parameters are the usual modeling knobs (priors, cluster count, topic count). Axioms are domain assumptions about data completeness and label validity plus the analogue premise. No new physical or mathematical entities are invented.

free parameters (5)
  • k-means cluster count = 21
    Set to 21 guided by silhouette score on the reduced feature matrix of 65 institutions; produces ~3 institutions per cluster on average and is therefore a free modeling choice that shapes the ‘archetype’ narrative.
  • Beta prior hyperparameters (non-hierarchical models) = Beta(1,1) and Beta(2,2)
    Two non-hierarchical models use Beta(1,1) and Beta(2,2); the latter is motivated by external qualitative claims (Lederman 2017) rather than data, and posterior means shift with the prior when counts are sparse.
  • Hierarchical Beta hyperpriors and latent-scale k = Gamma(0.01,0.01); HalfNormal(0,0.5)
    Gamma(0.01,0.01) on α,β and HalfNormal(0,0.5) on scalar k in the flexible BHM are weakly informative but still free choices that induce shrinkage; MCMC settings (4000 samples, 2000 tuning) are also free.
  • Number of LDA topics (news corpus) = 4 (illustrated)
    Topic modeling reports ‘one of four topics’ in the word-cloud figure; the topic count is a free modeling choice that affects the prominence of the finance–religion co-occurrence claim.
  • Feature-reduction decisions
    Initial 178 features manually reduced to ~66 by removing redundant/correlated variables; silhouette improved after reduction, but the exact retention rule is researcher judgment and affects all downstream PCA/clustering/SHAP results.
assumptions (5)
  • domain assumption The hand-assembled list of 65 closures/mergers between 2020 and end-2025 is sufficiently complete for descriptive regional and profile analysis.
    Closure identification began with 2022 news search plus Google News alerts; no independent administrative census of all closures is claimed. Invoked throughout §2.1 and results.
  • ad hoc to paper Pandemic-era closures, while mechanistically distinct from the enrollment cliff, share enough institutional vulnerability dimensions to serve as an imperfect preview.
    Stated explicitly in the Introduction as the framing premise; if false, geographic and profile patterns do not transfer to the cliff.
  • domain assumption Official and researcher-assigned closure-reason labels (Financial, Pandemic, Enrollment, Accreditation, Mutual Benefit) are adequate grouping variables for LDA.
    Labels taken from Castillo & Welding and reviewed by the researchers (§2.1); used as the grouping factor in §4.1.
  • standard math Standard conjugate Beta-Binomial and hierarchical Bayesian models with the stated priors yield meaningful posterior closure probabilities at state/region level.
    Conjugate updating and MCMC hierarchical models are textbook; invoked in §3.
  • domain assumption Endowment and other financial figures obtained via NACUBO/TIAA and Data USA (with imputation/approximation) are accurate enough for PCA, clustering, and SHAP.
    Authors note the difficulty of obtaining precise endowments (§2.2); these features appear among high-SHAP variables.

how reviews work

0 comments
Cite this review

Pith. "Pith review of College closures from 2020 to 2025: An exploratory analysis and its implications for the enrollment cliff." pith.science (2026). https://pith.science/paper/OU3JFBZI

@misc{pith2026260707917,
  author       = {Pith},
  title        = {Pith review of: College closures from 2020 to 2025: An exploratory analysis and its implications for the enrollment cliff},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OU3JFBZI}},
  note         = {Machine review of arXiv:2607.07917}
}
read the original abstract

The COVID-19 pandemic produced a modest wave of college closures and mergers that may offer an early, if imperfect, preview of the demographic "enrollment cliff" anticipated in the coming decade. This paper examines the institutions that closed or merged between 2020 and the end of 2025. We assemble a dataset of 65 such institutions, pairing institutional characteristics with state- and regional-level demographic, economic, and financial indicators, and supplement it with a corpus of news coverage of the closures. Using a combination of Bayesian models, dimensionality reduction, clustering, and topic modeling, we describe where these closures occurred, what the closed institutions had in common, and how they were discussed publicly. Consistent with prior demographic projections, closures were more frequent in the Northeast and Midwest, though the absolute numbers remain small. The closed institutions were heterogeneous rather than uniform: financial structure, regional demographics, and institutional mission each contributed to distinguishing them, and religious affiliation recurred prominently in media coverage. We frame these results as exploratory and descriptive given the small sample, and we discuss what they may, and may not, imply for institutions navigating the enrollment cliff.

Figures

Figures reproduced from arXiv: 2607.07917 by the authors.

Figure 1
Figure 1. Predicted state-level changes in the college-age population across the United States from 2012 to 2029, with colors indicating projected losses or gains, based on demographic projections presented in Grawe [2018]. California and Texas were split into North/South and East/West, respectively, due to their size. Data & Methodology 2.1 Dataset Our primary dataset integrates institutional-level characteris￾tics with stat… view at source ↗
Figure 3
Figure 3. State-level estimates of institutional closure risk derived from a Bayesian model with a Beta(2,2) prior, which—unlike the uniform Beta(1,1) prior—gently centers estimates of closure around 0.5, reflecting weak regularization toward moderate closure probabilities. Subsequently, to allow for information sharing between states and regions, a simple Bayesian Hierarchical Model (BHM) was constructed. In this model, the … view at source ↗
Figure 2
Figure 2. State-level estimates of institutional closure risk derived from a Bayesian model with an uninformative Beta(1,1) prior, representing uniform uncertainty over the probability of closure. Recognizing the influence of qualitative claims within the higher education discourse, such as the assertion that about half of institutions may be at risk of bankruptcy in the coming years from Lederman [2017], a second non-layered… view at source ↗
Figures from the paper (5 more)
Figure 5
Figure 5. Figure 5: Regional-level estimates of institutional closure risk derived from a Bayesian hierarchical model. were not distributed uniformly across all regions. To the extent that the enrollment cliff pressures institutions in similar ways, its effects may also prove uneven—but o…
Figure 6
Figure 6. Figure 6: First two components of Linear Discriminant Analysis (LDA) embedding of institutional features, with points colored by closure reason. The separation among the seven closure-reason groups in the discriminant space demonstrates that institutional characteristics contain…
Figure 7
Figure 7. Figure 7: Word cloud of the top 20 terms from one of four topics identified by Latent Dirichlet Allocation on a corpus of media articles. This pattern has a few implications, offered tentatively. It suggests that institutions receiving media attention for closing often shared tw…
Figure 8
Figure 8. Figure 8: PCA with K-means clusters (k=21) To look at the structure of the data, we used t-distributed stochastic neighbor embedding (t-SNE) with labels coming from k-means clustering to visualize the groupings. The plot showed points falling into reasonably distinct groups. We …
Figure 9
Figure 9. Figure 9: Frequency of top features across clusters and economic characteristics rather than any single attribute. This description points to some concrete considerations for institutions. A college whose profile is defined largely by reliance on state appropriations might consi…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

49 extracted references · 49 canonical work pages

  1. [1]

    College Enrollment Statistics , year =

    work page

  2. [2]

    Blei and Andrew Y

    David M. Blei and Andrew Y. Ng and Michael I. Jordan , title =. Journal of Machine Learning Research , volume =. 2003 , url =

    work page 2003

  3. [3]

    State Higher Education Executive Officers Association (SHEEO) , year =

    Rachel Burns and Lynneah Brown and Kelsey Hecker and Dustin Weeden and Hee Sun Kim and Beatrix Randolph and Aaron Pevitz and Sarah Karamarkovich and Jennifer Causey , title =. State Higher Education Executive Officers Association (SHEEO) , year =

    work page

  4. [4]

    The MIT Press , year =

    Lawrence Busch , title =. The MIT Press , year =

    work page

  5. [5]

    Campion , title =

    Linda L. Campion , title =. TechTrends , year =

    work page

  6. [6]

    The Incredible Shrinking Future of College , month =

    Kevin Carey , journal =. The Incredible Shrinking Future of College , month =. 2022 , url =

    work page 2022

  7. [7]

    Inside Higher Ed , year =

    Doug Lederman , title =. Inside Higher Ed , year =

    work page

  8. [8]

    Now There Might Be a Second One , year =

    Colleges Were Already Bracing for an Enrollment Cliff. Now There Might Be a Second One , year =

    work page

Show all 49 references
  1. [9]

    2025 , url =

    College Enrollment and Retention in the Era of. 2025 , url =

    2025

  2. [10]

    McKinsey & Company , month =

    Emma Dorn and Andre Dua and Jonathan Law and Samvitha Ram , title =. McKinsey & Company , month =. 2020 , url =

    2020

  3. [11]

    2024 , url =

    Jeff Doyle , title =. 2024 , url =

    2024

  4. [12]

    The Enrollment Cliff Is Here and Now What? , year =

  5. [13]

    College Financial Grades: America's Strongest And Weakest Schools , year =

    Emma Whitford , howpublished =. College Financial Grades: America's Strongest And Weakest Schools , year =

  6. [14]

    2021 , institution =

    The Lingering Fiscal Effects of the. 2021 , institution =

    2021

  7. [15]

    Americans' Confidence in Higher Education Down Sharply , year =

  8. [16]

    Grawe , title =

    Nathan D. Grawe , title =. 2018 , doi =

    2018

  9. [17]

    2017 , journal =

    Neil Harrison and Richard Waller , title =. 2017 , journal =

    2017

  10. [18]

    40+ College Enrollment Statistics in the United States (2024–2025) , year =

    2024

  11. [19]

    International Student Enrollment Decline , year =

  12. [20]

    2025 , month =

    Kania-Gosche, Beth , title =. 2025 , month =

    2025

  13. [21]

    and Levine, Phillip B

    Kearney, Melissa S. and Levine, Phillip B. and Pardue, Luke , title =. Journal of Economic Perspectives , volume =. 2022 , month =. doi:10.1257/jep.36.1.151 , url =

    2022 doi

  14. [22]

    Cornell University Press , year =

    Robert Lemke , title =. Cornell University Press , year =

  15. [23]

    2010 , month =

    Livingston, Gretchen and Cohn, D'Vera , title =. 2010 , month =

    2010

  16. [24]

    American Statistical Association, Social Statistics Section , year =

    Joshua H Lo and Florence Yong , title =. American Statistical Association, Social Statistics Section , year =

  17. [25]

    Fall 2025 International Student Enrollment Outlook and Economic Impact , year =

    2025

  18. [26]

    Fall 2025 International Student Enrollment Snapshot: Economic Impact , year =

    2025

  19. [27]

    College Navigator , year =

  20. [28]

    Fall 2021 Current Term Enrollment Estimates Report , year =

    2021

  21. [29]

    u3advisors , year =

    Shea O'Neil , title =. u3advisors , year =

  22. [30]

    2024 , url =

    JonBenetPalm , title =. 2024 , url =

    2024

  23. [31]

    Trellis Company , year =

    Anthony Schuette , title =. Trellis Company , year =

  24. [32]

    The American Vandal , year =

    Matt Seybold , title =. The American Vandal , year =

  25. [33]

    Department of Commerce , journal =

    U.S. Department of Commerce , journal =. U.S. Education Service Exports , year =

  26. [34]

    International Student Enrollment Is Down , year =

    U.S. International Student Enrollment Is Down , year =

  27. [35]

    Paying for College Infographic , year =

  28. [36]

    Economic Report of the President, 2024: Chapter 3 -- Population, Aging, and the Economy , year =

    2024

  29. [37]

    Knocking at the College Door: Projections of High School Graduates , year =

  30. [38]

    Preparing for the Class of 2030

    The Chronicle of Higher Education. Preparing for the Class of 2030. The Chronicle of Higher Education. 2023 , url =

  31. [39]

    2025 , url =

    Integrated Postsecondary Education Data System (. 2025 , url =

    2025

  32. [40]

    2025 , url =

    NACUBO-TIAA Study of Endowments (. 2025 , url =

    2025

  33. [41]

    Data USA Platform , year =

  34. [42]

    2024 , url =

    Regional Data:. 2024 , url =

    2024

  35. [43]

    State Unemployment Rates , year =

  36. [44]

    2025 , url =

    Castillo, Evan and Welding, Lyss , title =. 2025 , url =

    2025

  37. [45]

    CDC Wonder: Natality Information , url =

  38. [46]

    Carnegie Classification of Institutions of Higher Education , url =

    American Council on Education , howpublished =. Carnegie Classification of Institutions of Higher Education , url =

  39. [47]

    Higher Ed Dive , url =

  40. [48]

    Education Longitudinal Study of 2002 (ELS:2002) , year =

    2002

  41. [49]

    National Student Clearinghouse Research Center (NSCRC) , url =

    National Student Clearinghouse Research Center , howpublished =. National Student Clearinghouse Research Center (NSCRC) , url =

Pith tools

Reviewed July 10, 2026 · model on record in the stance chip above.