Technical methodology

A college enters the ranking when it meets all of these:

• It is a four-year institution (IPEDS level 1) that predominantly awards bachelor’s degrees.
• It is public or private nonprofit. For-profit colleges are out, and so are distance-only institutions.
• We can score at least three of the four pillars.
• It has at least one bachelor’s program in College Scorecard data.

Branch campuses rank separately only when we decide that for a specific campus. We can also exclude a college by an explicit identity decision. On the current build, 1,467 colleges qualify: 520 public and 947 private nonprofit.

Five sources feed the site. Only the first two affect a score.

Each pillar combines weighted components into one score. The weights below are within the pillar; the pillar’s weight in the overall score is in the heading.

Return (30% of the overall score)

Loan repayment left this pillar in April 2026. A missing value-added or benchmark figure is treated as average. A college with no earnings figure gets no Return score.

Mobility (30%)

Mobility is built in three steps.

1. Take the mean earnings of students from the lowest family-income third, ten years after entry, and divide by the median for peer colleges. Peers share the same control, size band and endowment-per-Pell-student band.
2. Multiply that ratio by Pell enrollment scale: Pell share times headcount.
3. Add 30% of Pell completion measured against peers of the same type and size. When Pell completion is missing, this term drops out.

Mobility is the one pillar still built on ten-year earnings, and they describe an older entry cohort than the four-year figures elsewhere.

Access (20%)

First-generation share left this pillar in September 2026, and nothing replaced it. When a component is missing, the others are re-weighted.

Affordability (20%)

The pillar scores the all-incomes average, not the under-$30,000 band. A college with no average net price gets no Affordability score. Other missing components are re-weighted.

We z-score each pillar within the ranked set, so every pillar sits on the same scale. Then we weight them: Return 30%, Mobility 30%, Access 20%, Affordability 20%. The composite percentile places each college among all 1,467 ranked colleges nationally.

When a whole pillar is missing, we fill it with that pillar’s median among ranked colleges and flag the college. Since the September 2026 build, ranks are deterministic: the same inputs always produce the same order.

Value added asks whether graduates earn more or less than we’d expect. A regression model, trained in BigQuery ML, estimates expected four-year median earnings from ten inputs: the share of students 25 and older, the female and part-time shares, public or private control, region, the college’s selectivity rate and tier, expected completion, a local economic composite and the state earnings threshold. Value added is reported earnings minus that expectation.

The expectation has an earnings floor, set by the state benchmark where one exists and by a fallback otherwise. First-generation share isn’t one of the ten inputs. It enters one step earlier: expected completion comes from a graduation-rate model that uses it. So although it left the Access pillar, it still shapes the value-added part of Return indirectly.

A second step splits each college’s value added into the part explained by its program mix, scale and labor market, and a remainder the model can’t explain. That split is for explanation only. It doesn’t change the headline figure.

Value added is an association after adjustment, not a causal effect. It doesn’t tell you what one student would earn at a different college, and the unexplained remainder isn’t proof of teaching quality. Read the full value-added approach.

A program gets a rank on its school’s page when all of these hold: the college is in the top 500 overall; the program has a published four-year earnings median; it had at least 10 completers in the latest cohort; and it has a Financial GPS payment-burden figure.

Programs rank at the four-digit CIP level and at the two-digit family level, nationally, by region and by state.

Each list family has its own ordering rule. A list needs at least five colleges.

Two rules keep missing data honest. Scores use fallbacks, and pages never do.

• Within a pillar, a missing component is re-weighted (Access, Affordability), treated as average (Return’s value added and benchmark) or dropped (Mobility’s completion term).
• A missing pillar takes the ranked-set median in the composite, and the college is flagged.
• On the page, a missing figure stays missing. We never show a fallback as an observed outcome, and missing Parent PLUS data doesn’t mean parents borrowed nothing.

Here’s what we can measure about the models today, and what we haven’t measured yet. All figures are from the September 2026 build.

The expected-earnings model

On colleges held out of training, the model explains 57% of the variation in four-year median earnings (R² 0.571). Its typical miss is $6,380 (median absolute error), $8,919 on average, and about $12,070 as root mean squared error. In training the error is about $11,860, close to the held-out figure, so the model isn’t overfitting much.

So the model explains a little over half the variation. A value-added difference of a few thousand dollars sits inside its typical error, and we don’t read one as a real difference between two colleges.

Rank stability

Across the two September 2026 builds, before and after we made ranks deterministic, composite ranks correlate at 0.9989 (Spearman, 1,467 colleges). 1,335 colleges moved, by a median of 4 places. The largest move, 549 places, came from correcting how multi-campus systems were counted. That measures the effect of the fix, not how sensitive ranks are to the pillar weights. We haven’t run a weight-sensitivity analysis yet.

Coverage

These figures describe past cohorts, not guaranteed outcomes. Four-year earnings follow completers who received federal aid and were working and not enrolled; ten-year earnings follow an older entry cohort. Different horizons can’t be joined into one group’s wage growth.

Some colleges report more completely than others, and small or suppressed cohorts limit what we can compare. A rank summarizes all of this in one number, and it will always hide some of the detail.

Back to the plain-language methodology