Methods: how the figures are chosen, calculated and checked

One study per test

Each test page gives the accuracy figures from one source, cited in full on the page. Where a systematic review or meta-analysis has pooled several studies, the pooled figures are used and the page says so. Otherwise the source is a single primary study. The library is not itself a systematic review: it does not search every database for every test, and a different study can give different figures. Where more than one study is worth knowing about, the page says so under Source.

How the study is chosen

The study on each page is the one cited for that test in Rule It In, Rule It Out. In choosing it, the author looked for the most reliable and best-powered evidence available for that test: a systematic review or meta-analysis where one exists, otherwise the most rigorous primary study, with the largest number of patients, an imaging or surgical reference standard, and a clearly described patient group. Website-only tests (marked as such) were added from the same kind of search in October 2026. When a newer or better study appears, the page and the book change together, and the change is listed on the corrections page.

Sensitivity, specificity and likelihood ratios

Sensitivity is the share of people with the condition whom the test picks up; specificity is the share without it whom the test correctly clears. LR+ is sensitivity ÷ (1 − specificity); LR− is (1 − sensitivity) ÷ specificity. Where the study (or review) publishes its own likelihood ratios, those are shown and the page says as published; otherwise they are calculated from the sensitivity and specificity on the page, and the page says so. A published figure and a recalculated one can differ by rounding or by the study's own method (for example a continuity correction); where they differ the page notes it.

Confidence intervals

The Study details panel on each test page gives 95% confidence intervals. Where the study reports them, they are quoted. Where it does not but the counts (true and false positives and negatives) are available, the Wilson score interval is calculated from the counts. Where neither is available, the panel says so. A wide interval, which is usual in small studies, means the headline figure is far less certain than it looks.

The probability ladder

Each page shows what a positive and a negative result do to the probability of the condition starting from 10%, 30% and 50%. Those three starting points are illustrations only: your starting estimate should come from the setting you work in and the patient in front of you. The calculation is: starting odds × likelihood ratio = odds after the test, converted back to a probability. The probability calculator does this for any starting estimate.

What was read, and the abstract-only rule

Figures are checked, in a separate check, against the published source on the internet, not against the book or another Trokos page. Where the full paper is open access or available to us, the figures are checked against the full text and its tables. For some tests the cited paper is closed but an open-access systematic review reproduces its figures; the panel then says the review was read, and the figures are checked against the review, not the paper. Where only the abstract could be read, the panel's What we read row says Abstract only, and the test is marked abstract source only on the strongest tests page. An abstract rarely gives the counts or the confidence intervals, so those tests carry less certainty. A few pages say that their figures have not yet been checked against the full paper.

Bias and the patients studied

Most diagnostic-accuracy studies recruit people already referred to a specialist or booked for imaging or surgery, where the condition is more common and more advanced than in a first-contact clinic. That usually makes a test look better than it will be for you. Other common weaknesses are small samples, examiners who knew the imaging result, not every patient receiving the reference standard, and reference standards weaker than surgery or MRI. The Things to bear in mind line in each panel states, in plain words, which of these apply, drawn from what the paper itself reports. It is a plain-language note, not a formal risk-of-bias score.

The strongest-tests page

That page lists tests whose cited study reports an LR+ of 10 or more or an LR− of 0.1 or less, with no caution on their page, ranked by the figure. Because the studies differ in patients, setting and reference standard, a likelihood ratio from one study is not directly comparable with one from another; the page shows the patients studied beside each test for that reason. Treat the order as a guide to the best-reported tests, not a ranking of which test is best for your patient.

Tests in isolation

Each page describes one test on its own. Tests for the same problem are rarely independent, so their likelihood ratios should not be multiplied together. Where a study has looked at a cluster of tests, the cluster has its own page.

Corrections

Errors found by our checks or reported by readers are corrected on the page and listed, with the date and what changed, on the corrections page. If you find one, please get in touch.

See also How this library is made.