Nationally, the average SAT-taker scores 14 points higher on Reading and Writing than on Math. At MIT, the median admit scores 30 points higher on Math. At Caltech, the gap reaches +45. Even at Williams and Pomona — places with no engineering school — Math beats Reading. We pulled the section data from every Common Data Set we could find and asked why.
Two million high-school seniors take the SAT each year. Their average total is 1,024. Their average breaks into 519 on Reading and Writing and 505 on Math — a fourteen-point edge for verbal.
That is the country we live in. Reading and Writing comes a little easier, on average, than Math. The pattern shows up cleanly across nearly every demographic: women score higher than men on EBRW; the same students score lower on Math.
It is the natural starting line. To understand what selective admissions does to the SAT, you have to know what the SAT does on its own.
Because once you do — the picture flips.
Start with the obvious cases. At MIT, the median admitted student scores around 790 on Math and 760 on EBRW — a thirty-point gap. At Caltech, the median runs 795 to 750 — a 45-point gap. At Georgia Tech, where the applicant pool spans a much wider band, the gap at the median is 50 points.
Read against the country's −14 baseline, the actual reversal at these schools is closer to +44 to +64 points. That is not a small wobble. It is a different distribution.
Carnegie Mellon publishes section scores for the whole university but explicitly tells applicants to its School of Computer Science to submit the SAT "with an emphasis on the math section."
Engineering schools were the easy case. Now look at the comprehensive Ivies.
Harvard is not a STEM school. Yale is not a STEM school. Princeton is not a STEM school. And yet at all three, admitted students show a clear Math advantage of +20 to +30 points at the median.
Stanford lands at +25. UChicago at +25. Northwestern, with its strong McCormick engineering program, runs an unusually large +45 — closer to Caltech than to Yale.
This is a finding that surprises people. Liberal arts requirements, humanities cores, creative writing seminars: these schools are nominally balanced. Their admits are not.
If the pattern holds at HYPSM, does it survive at the LACs — schools without a single engineer in the catalog?
Williams admits no engineering majors. Neither does Pomona, Swarthmore, Bowdoin, or Middlebury. And yet at all of them, the median Math score sits +10 to +25 above EBRW.
Amherst is the rare exception: its admitted class is roughly balanced, with EBRW slightly ahead at the 25th percentile. It is the only top LAC in our data that mirrors the country's verbal tilt.
Whatever is doing the work, it is not just engineering self-selection. Even at colleges built around a humanities core, Math scores are higher.
What explains the gap if it isn't STEM emphasis? The College Board has data on that.
The College Board's 2023 Digital SAT Validity Study — sample size over 200,000 — shows the cleanest story. Math scores predict college math and STEM grades; EBRW scores predict humanities and social-science grades.
Mean STEM GPAs march upward in a stairstep with Math scores: 3.01 for the 400–490 band, 3.21 for 500–590, 3.42 for 600–690, 3.68 for 700–800. Non-math GPAs do the same with EBRW.
For STEM majors, the SAT adds 38% more predictive power beyond high-school GPA alone. SAT-Math beats SAT-EBRW as a predictor of retention. High Confidence
At the aggregate level, the section correlations with first-year GPA look almost identical — Math 0.47, EBRW 0.48, Writing 0.51. The differential validity only shows up when you break the data out by major.
If admissions offices know Math predicts STEM success, the question is what they do with that knowledge.
The Math-over-EBRW gap at selective colleges has three plausible drivers, and they are not equally important.
Self-selection is the largest. Students with strong math scores disproportionately apply to MIT, Caltech, and Georgia Tech in the first place. We estimate this accounts for 60–70% of the STEM-school gap.
Tail compression comes second. At the very top of the score distribution, Math scores compress more tightly toward 800 than EBRW scores do. The 99th-percentile Math is roughly 790; the 99th-percentile EBRW is roughly 770. Selective schools fish from a pool where this asymmetry is mechanical, not chosen — about 20–30% of the gap.
Institutional preference is the smallest. There is direct evidence that some colleges weight Math more heavily — but the size of that effect is modest: 10–20% at STEM schools, near zero at LACs. Moderate Confidence
There's one place we have explicit text: MIT.
Most colleges are coy about the section weighting. MIT is not. In its 2022 blog post announcing the reinstatement of the SAT/ACT requirement, the admissions office wrote:
Three claims in one paragraph. Math is particularly predictive of MIT outcomes. The effect persists after controlling for socioeconomic status. High-school grades alone are insufficient.
It explains the curriculum. Every MIT undergraduate, regardless of major, completes two semesters of calculus and two of calculus-based physics. There is no humanities track that escapes it.
For our simulation, the question is what to do with all of this.
The simulation currently uses a single composite SAT per student. Adding section scores would let us reproduce the observed gaps directly — but it would only change behavior if some colleges actually weight the sections differently.
The proposed weights mostly hover at 0.50 — equal weight. The exceptions are narrow and concentrated: MIT 0.60, Caltech 0.65, CMU 0.58, Georgia Tech 0.60, Purdue 0.55. Outside the top STEM schools, the math weight stays roughly balanced, with LACs nudged to 0.48.
The honest reading: most of the observed Math advantage at selective colleges is a property of who applies, not how applications are scored. A composite-only model that gets the applicant archetypes right will reproduce the section gaps as a side effect.
Section-level decomposition is most worth doing for STEM schools and as a calibration check — not as a behavior change for the bulk of the 192-college list.