Sensory evaluation panels

Sensory evaluation is the only analytical method in food science whose detector is a person. That is not a weakness — no instrument resolves the interaction of a volatile release curve, a texture and an expectation the way a human does — but it means the discipline spends its effort on the problems any instrument has: calibration, resolution, noise and drift. The formal methods, standardised in ISO 8586 and ISO 13299 and their ASTM equivalents, exist to make a panel behave like equipment rather than like an audience.

Two panels, two questions#

The most consequential error in the field is to run one panel and read the other’s answer off it. A trained descriptive panel — eight to twelve screened assessors drilled for weeks against physical reference standards on a defined attribute lexicon — reports intensity: how much bitterness, how much cooked-cabbage aroma, at what point in the chew. It is an analytical instrument, and its members are explicitly instructed to set aside whether they like the sample. A consumer panel, by contrast, is a sample of the target population and is asked only whether it likes the product, usually on the nine-point hedonic scale introduced by Peryam and Pilgrim in 1957 and now covered by ISO 11136.

The two are not interchangeable. A trained assessor is no longer a typical consumer and their preference carries no market information. A consumer, conversely, cannot reliably separate attributes and reports a halo — a sample scored sweeter largely because it was scored better. Liking has to be modelled from descriptive data, not measured on the same people.

Panel size falls out of the binomial#

A discrimination test asks a strictly smaller question: is there any perceptible difference at all. In a triangle test (ISO 4120) an assessor gets three coded samples, two identical, and must name the odd one; guessing alone succeeds one time in three. The duo-trio test (ISO 10399) presents a labelled reference and two coded samples, so chance is one in two, which makes it less powerful for the same number of assessments.

Because chance performance is a known constant, the result is a binomial problem and the required number of assessments follows from the α risk, the β risk and the proportion of discriminators the sponsor cares to detect. This is where small panels quietly fail: with twenty triangle assessments a test has ample power to detect a large difference and almost none to establish that a small one is absent. A non-significant triangle test is not evidence of similarity unless it was sized as a similarity test in the first place.

Adaptation, carryover and drift#

Receptors adapt. Sustained exposure to a tastant raises its own threshold and shifts the perceived intensity of others — cross-adaptation — so the second sample in a sequence is never tasted by the same detector that met the first. Carryover from persistent stimuli such as capsaicin, astringent polyphenols or fat is worse, and no rinse fully removes it. The remedy is design, not chemistry: every serving order appears equally often, balanced by a Williams design, so a position effect that cannot be eliminated is at least spread evenly across treatments. The booths, controlled lighting and airflow of ISO 8589 do the same for the surrounding noise.

Finally, a panel drifts. Assessors recalibrate against each other and against the products they see most, so scale use migrates over months. Re-calibration against physical references, and blind duplicate samples inside a session, are how that drift is measured — without them a year of descriptive data is internally consistent and externally meaningless.