Where does precision collide?

A numeric parameter equal to one divided by the variance of a distribution, so a tighter distribution has a higher value. Gauss's 1809 Theoria motus corporum coelestium introduced a measure of precision in his account of measurement errors, writing the error curve with a constant that grows as the spread of the errors shrinks. Bernardo and Smith's 1994 Bayesian Theory is one modern statement of the practice of parametrising a distribution by its precision rather than its variance, which simplifies how prior and data information combine. Here the referent is a quantity attached to a probability distribution, not a ratio of relevant search results and not the closeness of agreement among repeated laboratory readings.[1][2]

Unit

A reciprocal-variance value on a probability distribution, tau equals one over the variance

Recovery means

Not applicable

Measured by

Computed as the reciprocal of the variance, or read off directly when a distribution is written in terms of precision

The share of the items a system returns that are actually relevant, so a query that returns ten documents of which eight are on target scores eight in ten. Kent, Berry, Luehrs, and Perry's 1955 paper on operational criteria for retrieval systems set out this idea as a factor measuring how well returned material matches the request, before the label used today was settled. Van Rijsbergen's 1979 Information retrieval fixed the paired vocabulary of precision and recall and the harmonic mean that balances them. Here the referent is a ratio scored against a set of results, unrelated to a distribution parameter or to the scatter of a physical measurement.[3][4]

Unit

A ratio between zero and one: relevant items retrieved divided by all items retrieved

Recovery means

Not applicable

Measured by

Computed as true positives divided by the sum of true and false positives on a labelled result set

The closeness of agreement among measurements of the same thing taken under stated conditions, kept separate from accuracy, which is closeness to the true value. The 1994 standard ISO 5725-1 defined it this way and split accuracy into trueness and precision, and the 2012 international vocabulary of metrology restated it. An instrument can be highly precise yet inaccurate when its readings cluster tightly but around the wrong point. Here the referent is a quality of a measurement process judged by how repeatable it is, not a parameter of a distribution and not a hit rate over search results.[5][6]

Unit

A property of a measurement process, quantified by the spread of repeated results

Recovery means

Not applicable

Measured by

Estimated from the standard deviation of repeated measurements under repeatability or reproducibility conditions