Classical infrared correlation rules as diagnostic tests: likelihood ratios from 6,663 experimental spectra

Infrared correlation tables give, for each functional group, a wavenumber window and an adjective. They do not say how often a band is found in that window when the group is absent, and that is the number that decides whether the observation is evidence. We treated every published rule as a diagnostic test. A set of 101 rules for 60 functional groups, transcribed from one widely used interpretation course and frozen before any confirmatory analysis, was scored on 1,539 ATR spectra from the Chemotion repository and replicated on 5,124 transmission spectra from the NIST Chemistry WebBook, with the deposited structure as ground truth and sensitivity, specificity and likelihood ratios as the metrics. Each rule was scored by band position alone and with the intensity and shape words the source attaches to it; 26 combinations of rules from the same table were scored as well. The main result is a division of the spectrum into two regimes. Above 1500 cm-1 most rules carry real evidence (median positive likelihood ratio 2.8 and 3.0 in the two collections, up to 43 for the saturated ester carbonyl); at or below 1350 cm-1, band position alone is close to uninformative (median 1.1 and 1.2), because almost every spectrum has a band in almost every window, and combining fingerprint bands with each other does not change this. The rankings agree across the two collections (Spearman 0.79; 87% of rules within a factor of two) and are unchanged when spectra flagged for atmospheric, saturation or baseline artefacts are removed. Intensity qualifiers and combinations cost more sensitivity than they buy specificity (the nitrile band reaches the stated intensity in one nitrile in ten), the benzene substitution patterns do not reach a likelihood ratio of 5, and several rules are far more useful for ruling a group out than for ruling it in. The quantified table, the frozen rule set and the scripts are provided.