About reference intervals

Reference interval verification starts from reference limits you already have, from the package insert, the literature or an earlier study. Before you adopt them, you test a small group of healthy subjects from your own population and see how many fall below, within and above the claimed limits.

Reference interval establishment works out new limits from your own healthy subjects, and each limit comes with a confidence interval that shows how precisely it is estimated.

In a verification, the share of subjects inside the interval describes the people you tested; population coverage is the matching figure for the whole healthy population. With 14 subjects, each subject moves the observed share by about 7 percentage points. When many subjects fall outside on the same side, suspect a shifted population or a calibration difference. Several flagged values, or a partition with very few subjects, are also worth a second look.

Each partition, such as female or male, is analyzed on its own. An establishment gives each partition its own limits. A verification counts each partition separately and compares each with the acceptance limit. Review very different limits in partitions you intend to report as one interval, or a partition with very few subjects, before reporting.

Target population coverage is the share of the healthy population the interval is meant to hold, for example 0.95. The confidence level, for example 90 %, applies to the interval around each limit estimated from your subjects. Coverage sets where the limits sit; confidence sets how wide their confidence intervals are. The settings and their defaults are in Settings.

A wide confidence interval means the reference limit is estimated imprecisely. Nonparametric limits are read from the ranked subject values and assume no distribution shape. Each limit can depend on one or two extreme values, so one subject can widen its confidence interval substantially. Many flags on one side are a reason to check the source records.

Parametric and log-normal limits assume a normal shape, of the values or of their log10, and so give limits from far fewer subjects.

Automatic for each partition uses nonparametric limits when a partition has enough subjects. With fewer, it checks whether a Gaussian model, then a log-Gaussian one, is supported, and the saved result names the method and diagnostic it used. A non-significant normality test does not establish the model. Review the Q–Q plot and box plot, and compare with nonparametric limits when the subject count supports them. The exact rule is in Estimators and transformations.

A nonparametric limit is the subject value at a given rank. For a two-sided 95 % interval, the lower limit sits at the 2.5 % point of the ranked values. With too few subjects, that point falls before the first value, leaving no limit to read. The confidence interval on a limit runs between two further ranks on either side of it and needs more subjects. When a rank falls outside the data, there is no confidence interval. Higher confidence, or a coverage closer to 1, pushes the ranks further out and needs more subjects.

Parametric and log-normal limits come from the mean and SD, so they can be calculated from very few subjects, but they are only as good as the distribution model. The counts for common settings are in Minimum subject counts.

A staged rule decides a verification in steps. Within each partition, each stage counts the subjects outside the claimed limits among the first subjects entered. A count at or below the stage's accept bound accepts the interval, and a count at or above its reject bound rejects it. Any other count moves on to the next, larger stage; a stage before the last may have no reject bound. A small first group can settle a clear result; a borderline result needs more subjects. A rule that runs out of subjects stays inconclusive. The stage settings and how they are applied are in Staged rule and Staged verification rules.