Card references SAT scores out of 2400 several times in the lawsuit. For example, he mentions that “Harvard’s class of 2019 had mean and median SAT scores of 2241 and 2270.” I’d expect that Card has access to all the SAT scores for the class of 2019 and is computing some basic stats on them such as mean, median, and number scoring 2400; rather than Harvard making a special point of highlighting perfect scoring applicants.
A quote form the Harvard Statement of Material Facts is below. It states that 2+ academic rating includes some applicants with “perfect” grades and testing. And it implies being one of the rare few that get a 1 academic rating (fewer than 0.5% of applicants in lawsuit sample) involves criteria beyond just having perfect stats.
"An applicant receiving a “2+” academic rating is typically an applicant with perfect, or near-perfect, grades and testing, but no evidence of substantial scholarship or academic creativity.
…
In many circumstances, an applicant receiving a “1” academic rating has submitted academic work of some kind that is reviewed by a faculty member. "
I agree that the higher admit rate primarily relates to scores being correlated with other parts of the application rather than score itself. However, I doubt that 2400 scores typically have “much higher average intelligence” than the numerous other Harvard applicants who score a few points shy of 2400. Instead I’d expect the bulk of the perfect scores are persons who took the test several times, until they managed a 800 on each section at least once, which superscored to 2400 Superscoring would explain why the number of 2400’s in Harvard’s applicant pool is so large in relation to the total number scoring 2400 as listed by CollegeBoard.
The Arcidiacono regression coefficients suggest the primary driver in the academic component of admission is academic rating, rather than test scores. After controlling for academic rating, the additional contribution of AI was quite small. The small component that does remain may largely relate to +/- academic ratings distinction, which were not controlled for.
Of course academic stats are an important component of academic rating. The percentages in different academic deciles who received a 2 (or better) academic rating is below, as listed by Arcidiacono. There is clearly a correlation, but it’s also clearly not a mechanical function of an AI threshold.
Top Decile – 98%
2nd Decile – 94%
3rd Decile – 84%
4th Decile – 70%
5th Decile – 51%
6th Decile – 26%
7th Decile – 8%
8th Decile – 1%
Academic rating explained 9% of variance in admissions decisions (Card model), and academic index explained ~half of variance in academic rating. Harvard OIR’s model was able to explain a larger portion of variance in academic rating through grades and scores than Arcidiacono was able to in his full model that including several functions of academic index, gender, race, concentration, hook status and many other factors. Harvard’s OIR model’s regression separated grades and scores, which allows a greater relative weight onconverted GPA (or whatever GPA is listed in file) and lesser weight on scores; while Arcidiacono’s model forced the AI score heavy weighting (2/3 of AI is scores), which may suggest Harvard is internally places a greater emphasis on grades.