A statistical technique new to lung function testing is described and has been applied to a study of flow-volume loop tests in normal subjects. Maximal expiratory flow-volume loops were performed on air, then breathing an 80% helium: 20% oxygen mixture for three deep breaths (3B) then three minutes (3M) twice daily for five consecutive days on 25 asymptomatic subjects (eight smokers and 17 non-smokers). The tests were performed on an Ohio differentiating spirometer and recorded on a storage oscilloscope. The standard lung function tests—forced vital capacity (FVC), forced expiratory volume in one second (FEV1) and peak expiratory flow rate (PEFR)—were the most reproducible. Maximal expiratory flow rate at 50% vital capacity (MEF50) and at 75% vital capacity (MEF75) were passably reproducible. MEF50 and MEF75 were as reliable breathing helium as breathing air, with 3B as reliable as 3M. However ratios of MEF helium/MEF air were very unreliable indeed, for both MEF50 and MEF75 and 3B and 3M. Thus percentage improvement in MEF after breathing helium appears to be an unreliable test in normal subjects. Helium isoflow volume was also very poorly reproducible with, on average, more variation within a single subject than between subjects. These conclusions apply to normal subjects and will require further work to assess them in specific diseases. However the very poor performance of isoflow volume and flow ratios in normal subjects suggests that it may be relatively difficult to distinguish normal subjects from patients with small airways obstruction using these tests. An application of canonical variate analysis to the data is described. This statistical technique assesses which indices or combinations of indices vary independently and hence are measuring qualitatively different parameters. Almost all the data variation was encompassed by variations in four parameters. Thus there are probably only four features in the flow-volume curve which can be used to discriminate between individuals. Repeating this analysis in patients with a specific disease should enable the best tests for use in that particular disease to be defined precisely. Canonical variate analysis should also be a useful method of assessing a new lung function test, since it assesses a test's reliability and more importantly whether it contributes any new information not given by older tests.