Learn · Walkthrough

Walkthrough: Liking for Science

The best dataset for learning Rasch analysis is one whose answers are already known. “Liking for Science” is the field’s standard: 75 children rating 25 science activities (dislike / neutral / like), published in Wright & Masters’ Rating Scale Analysis (1982) and shipped for decades as Winsteps’ “Example 0.” This walkthrough replicates that calibration in Logit and checks every number against the published output.

The data and the setup

Each row is a child; each column an activity (“Watch birds,” “Go to zoo,” “Watch a rat”…) scored 0 = dislike, 1 = neutral, 2 = like. All items share the same three response options, which makes this the canonical Rating Scale Model case: one set of thresholds for the whole instrument. In Logit: upload the CSV, choose RSM, run. The response matrix is freely available from Winsteps’ Example 0 documentation if you want to follow along with the identical file Logit’s own test suite uses.

The Wright map: what “liking science” means

The calibrated map spans about 5.5 logits, and the item ordering tells a story you can check against common sense. Go on a picnic anchors the bottom at −3.15 — essentially every child endorses it, so it barely measures science-liking at all. Go to zoo (−2.48) and go to museum (−2.04) are nearly as easy to like. At the top sit watch bugs (+1.83), watch a rat (+2.18), and find bottles and cans (+2.42) — activities only a child with real scientific appetite enjoys. The construct emerges from the ordering: liking science ranges from enjoying pleasant outings all the way to embracing the unglamorous parts of observation. That the empirical ordering matches theory is the instrument’s first validity argument.

The category structure

Logit calibrates the shared Rasch-Andrich thresholds for the dislike/neutral/like scale at −0.83 and +0.83 — ordered and symmetric, with the neutral category owning the middle ~1.7 logits of the scale. The CCC plot shows all three categories peaking in turn — the healthy pattern from the category structure guide. Every response option is earning its place. (Winsteps’ public Example 0 write-up publishes the item measures but not the category thresholds, so the thresholds are held steady by a regression test rather than checked against an external oracle — a distinction the validation page keeps explicit.)

Fit: even textbook data misbehaves

Sort the item table by fit and the map’s extremes reappear: the very easy and very hard items tend to show the most surprising residuals, because a child who dislikes picnics or loves collecting bottles is departing from a very strong expectation. Wright & Masters spend a chapter on exactly these diagnostics. The lesson to take: misfit rankings are where the qualitative reading of items starts, not a pass/fail gate — see the fit guide.

Checking against the published calibration

Because Winsteps publishes the full item calibration for this dataset, we can score the replication exactly. Logit’s measures correlate with the published ones at r = 1.000, with a mean difference of 0.04 logits against the two-decimal published table. Both engines use JMLE for the Rating Scale Model, so this is a direct head-to-head — no metric conversion involved. That check doesn’t just run once: it’s a permanent test in Logit’s engine suite, executed on every code change, so the agreement you see on the validation page can’t silently drift.