admit

Method, and what is wrong with it

1. There is no microdata. For anyone.

The Harvard admissions database was produced under a protective order and never released. This is not a matter of looking harder. Peter Arcidiacono was the plaintiff's expert witness and had access to it; when he and his coauthors published on it, they wrote:

due to court protections, we do not use this data directly. Rather, we rely on publicly available documents such as expert witness reports.

What is public is the full logit coefficient tables, with standard errors. That turns out to be enough to fix the shape of an admissions function, and it is what this model is built on.

Source in use: Document 415-1 (Arcidiacono Opening Expert Report), filed 06/15/2018, Table B.7.2: Logit estimates of Harvard's Admission decision, expanded dataset — Students for Fair Admissions, Inc. v. President and Fellows of Harvard College, Civil Action No. 1:14-cv-14176-ADB (D. Mass.)
Sample: Harvard applicants, Classes of 2014-2019 (six admission cycles, application years ~2009-2015)
N = 150,587 · 22 applicant-observable coefficients loaded.

2. The covariates are the docket's, deliberately

The input form is not designed; it is copied. It is the covariate list of Table B.7.2 Model 4 — the richest specification in the docket that conditions only on things an applicant can know about themselves. Models 5 and 6 fit far better, but they condition on Harvard's internal reader ratings, which are outputs of the process. Using them would require inventing the very thing we cannot observe.

3. Race is excluded on purpose

The docket's race coefficients are among its most precisely estimated. They describe a regime held unlawful in SFFA v. Harvard, 600 U.S. 181 (2023). They are recorded in this repository for provenance and are never loaded into the predictive model.

4. Everything outside Harvard is a transported prior

Exactly one school in the world has published admissions coefficients estimated on individual records. For every other school the coefficient vector is Harvard's shape, rescaled by that school's own Common Data Set C7 statement of what it weighs, with the prior standard errors inflated by how far the school sits from Harvard.

Then it is anchored: a simulated applicant pool is fitted so the predicted admits reconcile to the published admit rate and the admitted score distribution. Six published moments cannot identify nineteen coefficients, so only three parameters are solved — pool location, overall selectivity, and one slope scale. Location and steepness are data-driven. The relative weights across attributes are not, and are never claimed to be.

5. The transport assumption is untested

CROSS-TRANSPORT TEST (Harvard <-> UNC)
==============================================================
STATUS: BLOCKED. Not failed, not passed -- not runnable.

Blocked by:
  - UNC's logit coefficient tables (Doc 160-1 Tables A.4.1 in-state and A.4.2 out-of-state) are raster images in the filed PDF. pdftotext recovers the caption, the notes and the source line and not one cell. This was verified by extraction, not assumed.

What this means for the product:
  The central assumption -- that Harvard's coefficient shape, rescaled
  by a school's CDS C7 self-report, describes that school's weighting
  function -- is currently UNTESTED. Not weakly supported: untested.
  Every non-Harvard page must say that the shape is a prior which has
  not been validated against a second school.

What is recoverable from UNC's text layer today:
  in-state admit rate      47.9%
  out-of-state admit rate  13.5%
  Harvard, same era        7.4%
  UNC's report puts it plainly: 'UNC's admissions process operates
  almost like two different schools.' A single per-school coefficient
  vector cannot represent that, which is itself evidence that
  transport across institution types is the weak link.

To unblock (queued, not done):
  Hand-transcribe 5 image tables from Doc 160-1.
  Estimated effort: ~5 tables x ~30 rows, roughly 2-3 hours of careful hand entry, or an OCR pass with two-of-three agreement checking.
  Then re-run: fit the transport machinery to Harvard, predict UNC's
  published coefficients, and report the error per coefficient.

Harvard side is ready: 150,587 records, 22 coefficients,
source Table B.7.2: Logit estimates of Harvard's Admission decision, expanded dataset.

6. There is no calibration curve

A reliability diagram needs realised admit/deny outcomes on individual applicants. None are public for any school. The only route to one is a prospective panel — profiles recorded before decisions, outcomes reported after, scored against pre-registered bins — and that takes a full admission cycle.

Until then the honest claim is that this model reconciles to published aggregates and carries a shape estimated on 150,587 Harvard records. It is not calibrated, and nothing here should be read as though it were. A pre-registered gate is already written down: the model ships as predictive only if it beats both the school's base rate and admit rate conditional on test decile alone. If it fails that, the finding gets published.

7. What we refuse to do