Learn · Foundations

What is Rasch measurement?

A bathroom scale doesn’t care who steps on it. Two kilograms is two kilograms whether it’s flour or a cat, and the difference between 70 kg and 71 kg is the same as between 90 and 91. Test scores don’t work like that: a raw score of 25/30 on an easy quiz and 25/30 on a hard one mean different things, and going from 28 to 29 near the top is a bigger feat than going from 15 to 16 in the middle. Rasch measurement is the attempt to build, from item-response data, something that behaves like the scale rather than the quiz.

One equation

The Rasch model (Georg Rasch, 1960) says the probability that a person answers an item correctly depends on exactly one thing: the gap between the person’s ability and the item’s difficulty, both expressed on the same scale. When the person sits exactly at the item’s difficulty, the chance of success is 50%. Each logit of gap multiplies the odds of success by about 2.7. Drag both markers and watch the only thing that matters — the gap:

That S-shaped curve is the item characteristic curve (ICC), and the shared scale both markers live on is measured in logits (log-odds units). Notice what the model does not include: no partial credit for the item being “tricky,” no per-item guessing rate, no person-by-item quirks. That austerity is deliberate, and it buys three things.

What the austerity buys

People and items on one ruler. Because ability and difficulty are the same kind of quantity, you can place your respondents and your questions on a single map and see how well they line up — that’s the Wright map, and it’s the most immediately useful picture in instrument development.

Comparisons that don’t depend on the sample. Under the model, the comparison between two items doesn’t depend on which people happened to answer them, and the comparison between two people doesn’t depend on which items they happened to take. This property — specific objectivity — is what separates measurement from scoring, and it’s why Rasch users treat the model as a standard to aspire to rather than a curve to fit.

A testable claim. Because the model predicts every single response probability, every response is a small test of the model. Items whose response patterns are noisier or tidier than predicted announce themselves in the fit statistics — which is exactly how flawed questions get caught.

Beyond right and wrong

Likert items (“disagree / neutral / agree”) extend the same idea: each step up the response scale gets its own threshold on the ruler, shared across items in the Rating Scale Model or item-specific in the Partial Credit Model. The thresholds themselves become diagnostics — a middle category that never earns a region of the scale is a design flaw you can see. That story has its own guide with an interactive demo.

Where Logit comes in

Logit runs these models in your browser — upload a response matrix, pick the model, and get calibrated measures, the Wright map, fit diagnostics, category structure, and a report. The estimation runs on the same engines the research literature uses (TAM and eRm in R), and the numbers are checked against published calibrations on every change.