
Answering "how do they differ and by how much"
Descriptive analysis (DA) is a class of methods that quantitatively describe a product's sensory properties: which attributes (descriptors) are present and at what intensity. General guidance — ISO 13299:2016 (Russian counterpart GOST R ISO 13299-2015).
| Method variant | Essence | When to use |
|---|---|---|
| Consensus profile | The panel agrees on single scores through discussion led by the panel leader | Quick profiling, expert opinions; no individual statistics |
| QDA (Quantitative Descriptive Analysis) | Line scales, individual scores of 8–12 trained assessors, replicates, ANOVA | The R&D "gold standard": recipe comparison, product mapping |
| Spectrum™ | Absolute 0–15 scales with firm reference anchors, identical across categories | Cross-category comparisons, long-term panel calibration |
| Deviation from reference | Scoring the difference from the standard for each attribute | QC, shelf life: "how far has the batch drifted from the standard" |
| Free-choice profiling | Each assessor uses their own descriptors; analysis by generalized Procrustes analysis (GPA) | No time to agree on a lexicon; expert groups |
A typical QDA project: a panel of 8–12 trained assessors, 15–25 descriptors, 2–3 replicates, training takes weeks to months. This is expensive — so first formulate the research question and check whether a rapid method would suffice (Module 07).
The descriptor lexicon: ISO 11035
A lexicon is a vocabulary of descriptors with definitions and references. It is built per ISO 11035:
- Generation: assessors individually describe a wide set of category samples — 50+ terms are collected.
- Reduction: merging synonyms, removing rare and hedonic ("tasty") terms. The formal selection criterion is the index M = √(F × I), where F is the mention frequency and I the mean intensity.
- Fixation: for each final descriptor, a definition, a reference standard and its scale position are recorded. From then on the lexicon changes only through versioning — otherwise comparison across sessions loses meaning.
Texture attributes are covered by a separate standard — ISO 11036:2020 (texture profile): hardness, chewiness, springiness, cohesiveness, stickiness; scales with physical references (from curd cheese to a rusk).
Rating scales: ISO 4121
- Category scales (3/5/9 points) — simple but coarse: they "eat up" weak differences.
- Unstructured line scales (visual analogue scale, VAS, usually 100 mm) — a mark on a line between the anchors "none"–"very strong"; the QDA standard.
- Structured line scales — with intermediate ticks and labels.
- The 15-point Spectrum scale — an absolute scale with references on specific products.
- Magnitude estimation (ISO 11056:2021) — rating with numbers proportional to sensation; powerful, but requires lengthy training.
The profile plot: the product's sensory image
The profile plot (spider plot) is the obligatory visual output of descriptive analysis, but remember: it shows means without spread. Conclusions about the significance of differences are made by ANOVA, not by the eye on the radar.
The QDA questionnaire: how the assessor scores
Rate the intensity of each attribute with a mark on the scale:
Note the references in the wordings: an "8" from assessor A and an "8" from assessor B mean the same thing only thanks to shared standards practised during training.
Profile statistics: mixed ANOVA
The basic model for each attribute separately: Y = product + assessor + replicate + product×assessor + ε. Product is a fixed factor, assessor a random one (mixed model): the product F-test is computed against the "product × assessor" interaction mean square, not against the residual.
A reproducible example — the "bitterness" attribute, 3 samples (A, B, C) × 8 assessors × 2 replicates (a generated teaching dataset, RNG seed 42):
| Source | df | SS | MS | F | p |
|---|---|---|---|---|---|
| Product | 2 | 54.10 | 27.050 | 263.7 (against interaction MS) | < 0.0001 |
| Assessor | 7 | 14.82 | 2.118 | 24.0 (against residual) | < 0.0001 |
| Product × assessor | 14 | 1.44 | 0.103 | — | — |
| Residual (replicates) | 24 | 2.12 | 0.088 | — | — |
Panel means: A = 6.69; B = 4.09; C = 5.35. Tukey post-hoc (at the interaction level, HSD ≈ 0.30): all three samples differ significantly in bitterness. Interpretation for the client: "sample A is significantly more bitter than B and C; B is the mildest; the interaction is small (MS = 0.10) — the panel is in agreement".
- A significant "assessor" factor is normal: people use the scale differently, which is exactly why the assessor is included in the model.
- A large "product × assessor" interaction is a warning: the assessors disagree (see monitoring, Module 02).
- Next, the "products × attributes" matrix goes to multivariate analysis — PCA (Module 09).
Further reading
- Murray J. M., Delahunty C. M., Baxter I. A. Descriptive Sensory Analysis: Past, Present and Future. Food Research International, 2001 — the classical review of descriptive analysis schools: from the Flavor Profile to QDA and Spectrum. review
- Stone H., Sidel J. L. Sensory Evaluation Practices. 4th ed. Elsevier, 2012 — the original source of the QDA method by its creators. book
- Meilgaard M. C., Civille G. V., Carr B. T. Sensory Evaluation Techniques. 5th ed. CRC Press, 2016 — the chapter on Spectrum and building scales with references. book
Test yourself
Three questions on descriptive methods
1. How does QDA differ from the consensus profile?
Consensus is faster but destroys information about spread: no individual scores — no analysis of variance and no panel monitoring.
2. Why does every descriptor in the lexicon need a physical reference?
The reference "calibrates" the word: "bitterness = 5" is anchored to a specific solution, not to the assessor's personal idea.
3. In the mixed ANOVA of a sensory profile, the F-test of the "product" factor is computed against…
The assessor is a random factor, so the product is tested against the interaction. Testing against the residual would overstate significance (in our example F would be 307 instead of 264 — in borderline cases this decides the outcome).
Create a profiling study
Sensostat maintains a versioned lexicon, generates QDA questionnaires and builds profile plots automatically. Leave a request for trial access.