Predicting institutional closure · 1998–2024 IPEDS panel

College Distress Index

Every U.S. private nonprofit four-year college, ranked by its modeled likelihood of closing within four years. Two models do the predicting — a discrete-time hazard logit and a class-weighted LightGBM — with a stripped-down two-variable benchmark shown alongside. Search the current ranking, or trace any school’s risk back through time.

This page is the original project, which I started to explore what drives college closures and how to predict them. It was a topic I heard a lot about and wanted to dig into.

College closures have been remarkably rare over the past two decades, though the pace has been increasing. Among the roughly 1,200 private nonprofit four-year colleges operating at any time since 1998, about 0.2% closed each year before 2008. Since 2016, that has risen to about 1% a year, or roughly 13 closures a year on average. Over the full 27 years, 175 of the 1,498 colleges that ever appeared in the sample closed, about one in nine. This pace is likely to continue as the number of graduating high school seniors starts to shrink in 2026.

While a different research approach might investigate what causes declining demand (explored somewhat in the enrollment-by-state analysis), I started by modeling how to identify institutions in distress. This came down to a few factors: small, shrinking, tuition-dependent schools with thin margins, little endowment cushion, and near-open admission are the most at risk, and falling enrollment is the single largest signal. This is ultimately mechanical: schools that can’t draw new students and don’t have wiggle room or cash in the bank will run out of the money they need to operate.

The two models below take these factors and estimate how likely an institution is to close in the next four years. Those estimates are then combined into a risk score that tells you what percentile of risk the school falls in. The models are supplemented by a two-variable benchmark that looks only at the 3-year enrollment trend and tuition dependence, which is the best predictor at the very top of the list.

How this works

Two models predict; the ranking blends them. A stripped-down benchmark and a Cox robustness check sit alongside — and all of them largely agree on what pushes a college toward closing.

The risk score

Only the two predictors feed the score. They sit on different scales — LightGBM runs up to ~0.97, the hazard logit to ~0.37 — so their raw probabilities can’t just be averaged. Each is converted to a percentile rank first, giving both an equal vote:

  1. Take each predictor’s current-year probability of closing within four years.
  2. Rank it into a percentile across all 1,191 schools (0–100).
  3. Average the two percentiles, equal weight.

score = mean( pctile(hazard), pctile(LightGBM) ) × 100

It’s a relative measure, not a probability: 72 means “riskier than about 72% of schools by the two predictors,” not a 72% chance of closing (the per-model columns hold the absolute estimates). The 2-variable benchmark is shown alongside but not blended in — a challenger, not a vote — and the Cox model is left out entirely because it audits rather than predicts (right).

What drives closure

Bar length is each factor’s weight inside that model (logit log-odds, LightGBM share of split gain, Cox log‑hazard‑ratio); units differ, so read down a column, not across. Color shows direction; ★ marks the two factors that make up the benchmark.

lowers closure risk raises closure risk faded = not significant

The backbone is robust: small, shrinking, tuition-dependent schools with thin margins, little endowment cushion, and near-open admission — falling enrollment the single heaviest signal. Remarkably, just the two starred factors — enrollment trend and tuition dependence — make the sharpest predictor of the very top of the list (best top-25/50 hit-rate), which is why the benchmark earns its own column.

Why Cox only audits. Its baseline hazard soaks up the era’s overall rise in closures, forcing each factor to explain which schools close, net of when. It agrees with the logit on every fundamental (redundant where it’s right); the one place it differs — the demographic outlook collapses to insignificant — it’s removing a calendar-time artifact, not adding a prediction. A credibility check, so it stays out of the score. *LightGBM reads the near-100% acceptance cliff through the continuous admit rate, not a flag.