An investigation in eight chapters

The Summer Melt

Every May 1, hundreds of thousands of high-school seniors put down a deposit and tell themselves they are headed to college. Between then and the first day of class, somewhere between one in ten and four in ten of them quietly disappear. Some lose their financial aid. Some lose their visa. Some lose their nerve. The phenomenon has a name — summer melt — and the people it claims look nothing like a random sample.

10–40% of college-intending grads who fail to enroll in the fall
2.0× first-gen melt rate vs. continuing-generation
+3.1pp enrollment lift from a $7-per-student SMS nudge
40.5% Philadelphia melt rate, class of 2024
Scroll to begin ↓
Chapter I

The number is bigger than you think

A deposit is a contract a 17-year-old signs with a stranger. Most colleges treat it as the end of admissions. Researchers Benjamin Castleman and Lindsay Page treated it as the start of a question: how many of those signed contracts actually convert into a body in a chair?

Their 2014 paper, A Trickle or a Torrent?, gave the field its working numbers. At four-year colleges, melt sits around 10–20%. At community colleges, where no deposit anchors the commitment, it climbs to 37%. In low-income urban districts, more than four in ten college-intending graduates never reach a campus.

37% Community-college melt rate among students who said they were going to college. Compare to 19% at four-year schools in the same study.

The aggregate hides the structure. Melt is not evenly distributed across colleges — it is bent by selectivity.

Chapter II

Selectivity is a melt shield

At HYPSM, melt rounds to nothing. Yields run 78–85%, financial aid covers 100% of demonstrated need, and the opportunity cost of skipping is so high that the rare departures are gap-year deferrals — not true melt.

Walk down the prestige ladder and the floor opens up. Ivy+ schools see 1–5% melt. Near-Ivies 2–7%. By the time we reach Selective publics with 25–50% acceptance rates, melt averages 10% and tops out near 20%.

Three forces are doing the work: financial aid generosity, yield-driven commitment, and class size. The first removes the affordability cliff. The second selects for committed depositors. The third determines whether 5% melt is statistical noise or 800 empty beds.

~8× Ratio of average melt at Selective Publics (10%) to HYPSM (1.5%).

The tier is a backdrop. The real variation lives in who the depositor is.

Chapter III

The demographic gap

Castleman and Page framed melt as an information deficit, not a motivational one. Students want to go. They lack the scaffolding — paperwork, deadlines, FAFSA verification, immunization records, transcript requests — that college-educated families provide automatically.

First-generation students melt at 2–3× the rate of continuing-generation peers. Low-income students at 1.5–2.5×. International students, vulnerable to visa denials and currency shocks, at 2–4×.

These multipliers stack. A first-gen low-income student at a high-sticker-price tier-4 school is in a different population than the median depositor. The simulation engine treats melt probability as base × first-gen × income × international, capped at 30%.

2.9× Philadelphia ratio of low-income melt (56%) to higher-income melt (19%) in one Castleman & Page district.

To see what the demographic multipliers do to a real district, look at Philadelphia.

Chapter IV

Philadelphia, year after year

The School District of Philadelphia publishes annual melt reports. The numbers are a steady indictment. Pre-pandemic melt sat around 31.5%. The COVID year of 2020 pushed it to 36.2%.

Then came the 2024 FAFSA rollout — late, broken, missing the May 1 deadline for millions of families. Philadelphia's class of 2024 melted at 40.5%, the highest rate the district has recorded.

Lancaster, Pennsylvania — 60% Hispanic, 16% Black — tells a similar story. Pre-pandemic melt of 26% jumped to 43% during COVID. These are not anomalies. They are what happens when an information deficit meets a financial-aid system in crisis.

+9 pp Philadelphia melt rose 9 points between 2019 (31.5%) and 2024 (40.5%) — a quarter of college-intending graduates additionally lost in five years.

The mechanisms behind those rates fall into a small, repeating taxonomy.

Chapter V

Six paths to a no-show

The summer melt literature, when distilled, identifies six dominant pathways from deposit to absence. Financial barriers lead the list — FAFSA verification failures, surprise housing fees, family income shocks, loan-versus-grant confusion.

Administrative friction comes second: missing transcripts, immunization records, placement tests, housing deadlines. Then social and psychological barriers — imposter syndrome, family pressure, isolation in a summer with no counselor on call.

Gap-year decisions, F-1 visa issues, and competing offers (employment, military, a waitlist call from a more prestigious school) round out the catalog. Each can be addressed in isolation. None is being addressed at scale.

~30% of FAFSA filers selected for verification — a paperwork gauntlet that disproportionately knocks out first-gen and low-income depositors.

In the past five years, three crises have stress-tested every one of these pathways at once.

Chapter VI

Three shocks in five years

2020 was the COVID melt surge. Gap-year deferrals leapt from 50,000 nationally to 130,000. Two-year college enrollment fell ~50%. Harvard recorded 340 deferrals — 20% of its class — versus a normal 100–130. UPenn deferrals tripled.

2024 was the FAFSA crisis. The redesigned form launched late and buggy. Submissions ran 20–30% behind prior years. Aid packages did not arrive before the May 1 deadline. The Chronicle called it potentially the worst summer melt season in memory.

2025 is the visa crisis. At least 4,736 international student visa records have been terminated; 187 colleges affected. Students at elite schools cancel summer travel, fearing re-entry denial. The 1.1 million-student international pipeline is functioning as a high-melt subpopulation across every tier.

130,000 Gap-year deferrals in 2020, up from a pre-pandemic baseline of 40,000–60,000 (Gap Year Association). Roughly 3% of deferred students never enroll.

The good news: the literature has unusually clear answers about what works.

Chapter VII

What works, and what it costs

Castleman and Page ran the canonical RCT in 2015. Among 4,882 students across five cities, a 10-message SMS nudge campaign raised enrollment by +3.1 pp overall and +5.7 pp for students with $0 expected family contribution. Cost: $7 per student.

Georgia State's Pounce chatbot (2016) went further: a 22% reduction in melt and +3.9 pp enrollment, at $3 per student. 63% of treatment students engaged on three or more days. Less than 1% of messages required human staff intervention.

Counselor outreach lifts low-income enrollment by +8 pp at $150 per student. Peer mentoring adds another +2 pp at $80. High school–college collaborations increase enrollment by 13% for underrepresented groups.

$7 Per-student cost of the Castleman & Page SMS nudge that raised low-income enrollment by +5.7 pp. The cost-effectiveness frontier in summer melt is unusually steep.

Even if every intervention works, a structural feature of the admissions cycle keeps melt churning.

Chapter VIII

The cascade nobody sees

Melt and waitlist movement are mechanically linked. When Harvard loses two depositors, it pulls two from its waitlist. Those two were committed at Northwestern and Duke — which now go to their waitlists, pulling students from Emory and NYU, which pull from regional privates, and so on.

A single melt event at the top of the prestige stack can trigger 3–5x its size in downstream waitlist activity. The summer of 2024 ran this cascade hot: late FAFSA aid letters scrambled commitment timing, and waitlists at some Tier 3–4 colleges stayed open into August.

For our 6-tier simulation, this means a faithful melt model has to be sequential, not aggregate: melt → waitlist pull → second-order melt → second waitlist round. The seats that open look the same on a balance sheet, but they ripple through different students every time.

3–5× Estimated multiplier on system-wide melt impact from a single top-tier event, through cascading waitlist pulls.
Melt rate by institution type
Castleman & Page (2014) and follow-on field reports.
Source: research/summer_melt.md, summer_melt_data.json (Castleman & Page 2014; Harvard SDP).
Estimated melt rate by college tier
Bars show the average; whiskers show the range. Color matches the simulation's tier palette.
Source: research/summer_melt_data.json — melt_rates_by_tier.
Melt multipliers by demographic group
Multiple of base melt rate. Whiskers show the published range.
Source: research/summer_melt_data.json — melt_by_demographics (Castleman & Page 2014; EAB 2024).
Melt rate over time, two districts
Philadelphia and Lancaster, PA. Two slope-comparable urban districts.
Source: research/summer_melt.md — School District of Philadelphia; Hechinger Report (Lancaster).
Pathways from deposit to no-show
Six dominant cause categories from the summer melt literature.
Source: research/summer_melt.md (Castleman & Page 2014; Hechinger; EAB 2024).
Three shocks, three indicators (2019–2025)
Gap-year deferrals, FAFSA submission lag, terminated visa records.
Source: research/summer_melt_data.json — covid_trends; Gap Year Association; National Student Clearinghouse.
Cost vs. effect of anti-melt interventions
Each dot is a documented intervention. Lower-right is best: cheap and effective.
Source: research/summer_melt_data.json — intervention_effects (Castleman & Page 2015 RCT; Pounce chatbot 2016).
The melt → waitlist cascade across tiers
A single Tier-1 melt propagates downward through waitlist pulls. Sankey of estimated flow.
Source: research/summer_melt.md — cascade narrative; flow estimates illustrative, not measured.