College ROI Data: 30-year net present value of U.S. colleges and majors

Clean, citable tables on the lifetime financial return of U.S. colleges and majors: 30-year net present value by school and state, cohort-weighted ROI by major category, the out-of-state cost premium, and an AI-exposure score per major. The headline the data mints: picking a major swings mean lifetime ROI by ~$1.24M (Engineering +$1,132,553 vs Philosophy & Religious Studies −$111,915, cohort-weighted, FREOPP 2021) — and in several popular major categories the majority of graduates never break even. Derived by LE TEEN from FREOPP (2021, CC BY 4.0), NCES IPEDS (2023-24 Final, public domain), BEA Regional Price Parities (2023, public domain), Microsoft Research "Working with AI" (2025, CC BY 4.0), the Anthropic Economic Index (2025-09-15, CC BY), and Census ACS 2022 PUMS (public domain). Full methodology: le-teen.com/methodology. Canonical repository: github.com/thomasthinks/college-roi-data. Ships as tidy CSVs plus a Frictionless datapackage.json with typed schemas. CC BY 4.0. Also queryable as a free, keyless JSON API (le-teen.com/api, 640 endpoints) and via an MCP server for LLM tooling (college-roi-mcp on npm). Mirrored across Zenodo (hub), GitHub, Hugging Face, Kaggle, OSF, figshare, and OpenML. Changes in 1.0.1 (2026-08-13) — data correction. In roi-by-major-category.csv, the column pct_never_breakeven was carrying the share of graduates whose 30-year lifetime ROI is below zero, rather than the share FREOPP flags with its 999 “never breaks even” sentinel. Five of nineteen rows are corrected: Agriculture and Natural Resources 15.8→15.5, Communications and Journalism 16.5→16.4, Mathematics and Statistics 3.2→2.9, Computer and Information Sciences 2.5→2.3, Economics 2.1→1.9. The remaining fourteen are identical under either definition. A new column pct_negative_lifetime_roi now carries the below-zero share, so both figures are available and no existing column was renamed or removed. The two are not interchangeable: never-breaks-even is a strict subset of below-zero-ROI, and the graduates in the gap do break even, in their late fifties, past the 30-year window. Field descriptions in datapackage.json now state each denominator. No other table is affected. Changes in 1.0.2 (2026-08-13) — documentation only, no data change. The README shipped inside an archived version cannot be edited afterwards, so it no longer embeds identifiers that can change. Two classes were removed: pinned version DOIs (the badge, the Mirrors intro and the Cite section named the v1.0.0 DOI as canonical, which 1.0.1 superseded the moment it was published), and OpenML numeric dataset ids (OpenML datasets are immutable, so a correction there means a new id and the old one stops resolving). Both now point at the concept DOI 10.5281/zenodo.21351602 and at le-teen.com/data respectively. Data files are byte-identical to 1.0.1.