College Board does not publish a raw-score-to-scaled-score conversion table for the Digital SAT. On the old paper test they did, one table per form. On the digital test the conversion is module-dependent, so there is no single table to publish. That is the whole problem in one sentence, and it is a genuinely interesting one to build around.
This is a short write-up of how I approximated it for a free calculator, what the data actually supports, and where the estimate stops being trustworthy.
Why a lookup table cannot exist
The Digital SAT is adaptive at the module level. Each section has two modules. Your performance on Module 1 decides whether Module 2 is the easier or the harder variant. Two students who answer the same number of questions correctly can therefore receive different scaled scores, because they sat different second modules.
There is a second layer on top: equating. After each administration College Board adjusts for small differences in difficulty between forms so that a 1300 in March means the same thing as a 1300 in October.
And a third detail that people forget: each module contains unscored pretest questions. The number of questions you see is not the number of questions that count.
What data is actually public
Two usable sources:
- Published practice tests ship with raw-to-scaled score ranges per section. Not a formula, but a set of observed points.
- The annual total-group report, published every fall, gives the score distribution: how many students scored in each 10-point band. The class of 2025 had a mean total of 1029 with a standard deviation of roughly 235.
Between those you get: some anchor points from the practice ranges, and a shape for the population distribution.
The approximation, in three steps
Step 1 — anchors. Collect the (raw, scaled) pairs from the official practice tests. For each section you end up with a scatter of points rather than a curve.
Step 2 — monotone fit. Fit a monotone curve through those points. It has to be monotone, because a conversion that maps more correct answers to a lower score is definitionally wrong, and a plain least-squares polynomial will happily produce exactly that in the sparse tails. I used piecewise-linear interpolation between anchors with linear extrapolation at both ends, which is boring, explainable, and never overshoots.
Step 3 — sanity check against the population. Convert the fitted section scores back into a total distribution and compare it to the published total-group distribution. If the fitted curve implies that 25% of students score above 1400, the curve is wrong, because the real figure is about 7%. This step catches the failure mode you cannot see from the anchors alone: a fit that is locally fine and globally nonsense.
The output is a total score and a stated error bar. Mine says plus or minus 30 to 60 points, which is honest — the anchors are too sparse to claim better.
Why the top of the scale behaves strangely
Something the fit makes obvious: the last 100 points cost far more per question than the first 200.
Going from about 1000 to 1200 takes roughly 18 additional correct answers out of 98. Going from 1400 to 1500 takes about 8. The curve is not linear and it is not meant to be — scaled scores are built to spread out the middle of the distribution, where most students sit, and compress the tails.
That single property explains a lot of student frustration. A plateau at 1380 that will not budge is not a study-habits problem, it is about eight specific questions.
Which percentile to quote
There are two official percentile tables and most sites publish one without saying which:
- User group percentiles compare you to students who actually took the test in your graduating class. A 1200 is the 75th percentile here.
- Nationally representative percentiles compare you to all US students, including those who never sat the test. The same 1200 is about the 81st percentile.
Neither number is wrong. Quoting the wrong one is. If the number is going on a college application narrative, use user group, because that is what admissions officers are looking at.
What I would do differently
Two things I got wrong first and fixed:
Lazy-loaded images and headless verification. The chart image had loading="lazy", and when I checked the page programmatically the image reported naturalWidth: 0. Not a bug in the image, just lazy loading never triggering without a real scroll. Switched to decoding="async" with explicit width and height, which also removes the layout shift.
Percentile numbers need a source, not a memory. My first draft had 1400 at the 90th percentile. The real user-group figure is 93rd. The distribution changed as test participation grew by about half a million students between 2021 and 2025, and any number that is not tied to the current annual report will drift.
The result
The calculator is a static page with the fitted tables inlined as data — no backend, no API calls, no sign-up. The whole thing is a few hundred lines of JavaScript and a lot of footnote-checking.
If you want to poke at it: Digital SAT score calculator, and the percentile chart covers every total score from 400 to 1600 with both percentile definitions explained. Sources are cited on the pages themselves, which I think is the minimum bar for anything that produces a number a student will make decisions from.
Happy to answer questions about the fitting approach in the comments — particularly if you have a better idea for handling the sparse upper tail, because that is the part I am least confident in.
Top comments (0)