Phase 0 — derivation complete, instrument undeployed

Sextant

A political self-placement instrument built from measured structure rather than asserted axes.

Named for the instrument that fixes your position from known reference points and reports its own error. The second half of that is the part most compasses skip.

The gap

The two axes on politicalcompass.org were asserted rather than discovered, and its scoring has never been published. 8values and its descendants are open source but assign each question a hand-picked weight on a fixed axis. Neither has checked whether political opinion actually varies along the axes it draws.

Sextant starts from the opposite end. Take a large body of real responses, find out how many dimensions are in there, and only then write questions.

The source is the General Social Survey, run by NORC at the University of Chicago and freely downloadable without registration. Analysis is restricted to 2000 onward: 37,583 respondents, 103 political attitude items.

Five axes, and the one nobody is measuring

Five interpretable axes came out, replicated on held-out data. Self-reported ideology and party identification were kept out of extraction entirely and used only afterwards, so the axes are validated against something they were not built from.

The evidence for that last clause is specific rather than rhetorical. Under a two-factor model, confidence in Congress has a communality of 0.06 and confidence in the Supreme Court 0.01 — the model has almost nothing to say about either. Under five factors they rise to 0.53 and 0.43.

Method, and why each step is there

Polychoric correlations

Survey answers are ordered categories, not numbers. Treating "too little / about right / too much" as 1, 2, 3 attenuates correlations badly on coarse, lopsided items, which is most of this battery. The polychoric estimator models a continuous variable underneath each category, and on simulation it recovers a true 0.60 that Pearson reports as 0.42.

Pairwise estimation

The survey rotates items across ballots, so no respondent answers everything and listwise deletion would discard nearly the whole sample. The matrix is built from whoever answered each pair. That missingness is imposed by the survey design rather than by the respondent, which is what makes pairwise estimates unbiased here.

Item parcelling

The survey asks about political tolerance 18 times and abortion 7 times — and those abortion items inter-correlate at 0.84 — but about gun permits once. Unparcelled, the first factor is tolerance and the third is abortion, not because they dominate American opinion but because they dominate the questionnaire. Parcelling combines batteries so no topic out-votes another, with membership decided by question wording rather than variable-name prefix.

Two dimensionality tests, bracketing from opposite directions

Horn's parallel analysis and Velicer's minimum average partial both over-retain above a sample of 10,000, certifying 10 factors and 7 respectively. So the substantive count of five comes from reading the loadings in the open, and is argued rather than proven.

What is not established

Why there is no international version

Not a registration wall, which is what this project assumed until it checked. Every established cross-national attitude survey fails on licence, access, or both. The European Social Survey is non-commercial and share-alike. LAPOP bars commercial use and redistribution. Latinobarometer and the Asian Barometer are research-and-teaching only. The rest route through a provider that blocks scripted access at the infrastructure level — so it could never be a repeatable pipeline, only a hand-refreshed download, which is a different kind of project.

One axis has real data. Afrobarometer Round 9 covers 39 countries and 53,444 respondents with an eleven-item institutional trust battery, and downloads with a plain anonymous request. Scoring it through this instrument would still be wrong: of its eleven institutions, four overlap with the American battery, and two of the additions are disqualifying on this project's own past findings — confidence in organised religion loads on moral traditionalism rather than trust, and the ruling and opposition parties are partisan objects, which the American battery deliberately excludes. A composite containing both would be a different construct reported on a scale implying a comparability it has not got.

Even the defensible version is sensitive to that choice. Ranking those 39 countries on all eleven items against the four that resemble the American battery agrees at a Spearman correlation of 0.932 — but Tunisia moves from 19th to 3rd, and Kenya and Gambia each move nine places in opposite directions. Which institutions you count decides the answer.

Status

The axes are derived, the instrument is written, the site is built, collection is built, population norms are measured and the replication test is written. 342 tests pass, the data contract audits 23 of 23, and 2,811 browser conformance comparisons agree.