SESH FRAGRANCE ACADEMY | ADVANCED SENSORY EVALUATION
Do seven correct choices out of ten prove a detectable difference?
Follow the route from the nasal cavity to the olfactory bulb. Smell is perception; it does not independently identify a perfume’s ingredient list.
Choose triangle or duo-trio designs, calculate chance baselines and distinguish discrimination from attribute characterization.
1. Discrimination does not say which fragrance is better
A difference test asks whether participants can discriminate A from B beyond chance, not which product is nicer, stronger or more floral. Pump or storage differences can produce detectability changes without proving specific chemical changes. Samples require comparable preparation, age, dose, substrate and environment. If one sample was prepared earlier, success could reflect that confounder. The test alone does not identify the cause.
2. Triangle test: the odd-one-out choice
In a triangle test each assessor receives three coded samples, two identical and one different, and selects the odd one out under a forced-choice design. Chance success is 1/3. Balance all six possible order arrangements across assessors to avoid position bias. ISO 4120:2021 sets out a standardized method, but casually using three strips is not enough to claim compliance.
These curves show how changing mean and variance affects a statistical model. They are theoretical teaching data, not skin or fragrance-panel results.
Open full-size image ↗3. Duo-trio uses a reference and two coded alternatives
Duo-trio presents a reference R followed by two coded alternatives, one matching the reference; the assessor chooses the match. Chance success is 1/2, so triangle significance cutoffs cannot be reused. Constant- and balanced-reference designs are possible depending on the research setting. ISO 10399:2026 is the current edition, superseding the withdrawn 2017 version. Duo-trio can be easier for participants but generally needs more responses than triangle to achieve similar discriminatory power.
4. The unit of analysis matters
Ten trials from one assessor are not automatically equivalent to ten independent assessors: learning, adaptation and within-person correlation can make judgments dependent. A simple binomial model assumes a known chance probability and reasonably independent outcomes. Distinguish participant count from trial count, balance sequences and choose an analysis that respects repeated measures.
5. Worked probability example: seven of ten triangle successes
Assume ten independent assessors, a balanced triangle test and seven correct choices. Under chance p=1/3, the upper-tail binomial probability P(X≥7) equals the sum of C(10,k)(1/3)^k(2/3)^(10−k) for k=7…10, approximately 0.0197. That would cross a prespecified one-sided 0.05 criterion under valid assumptions. It says neither how large the difference is, nor which scent is preferred, nor whether either product is safe. Design failures can invalidate the inference.
6. Seven of ten duo-trio choices are not the same evidence
For seven correct out of ten duo-trio responses under chance p=1/2, the binomial upper-tail probability is (120+45+10+1)/1024 = 0.171875. The same observed success rate gives weaker evidence than a triangle test because the guessing baseline differs. Choose the test and statistical decision rule before the study; do not change the alpha threshold after seeing results.
7. No significant difference is not proof of equivalence
Failure to reject a no-discrimination null is not proof that the products are identical. State that no difference was detected with the tested sample size, method and power. To support practical similarity or equivalence, use a design with a predefined meaningful-difference margin and appropriate analysis. ISO 4120 and ISO 10399 distinguish testing for difference and similarity; a descriptive panel is needed to characterize the attribute responsible for a detected difference.
Case study: the same score, different chance models
Ten independent responses yield seven correct in a triangle design (p≈0.0197) and seven correct in a duo-trio design (p≈0.1719). These are not equivalent evidence and neither means '70% of customers like F02'. They concern discrimination under specific, controlled conditions.
Professional practice exercise
Design a fictional triangle study with 12 independent assessors and all six sample-order permutations represented twice. Contrast its chance probability with duo-trio, and explain why 12 repeated trials from one person are not equivalent to 12 assessors.
Three educational concept maps
Triangle versus duo-trio
- Triangle: two same, one different
- Chance correct: 1/3
- Duo-trio: reference + two choices
- Chance correct: 1/2
Seven successes out of ten
- Triangle: chance tail p≈0.0197
- Duo-trio: chance tail p≈0.1719
- Same count, different statistical evidence
Discrimination test limitations
- Discrimination ≠ liking
- Nonsignificance ≠ equivalence
- Assess independence of responses
- Document design and chance baseline
Scientific glossary
- Triangle test
- Three-sample forced-choice discrimination with chance success of one-third.
- Duo-trio
- Reference-based discrimination with two alternatives and half-chance success.
- Binomial tail
- Probability of at least the observed successes under a stated chance model.
Official scientific references
This educational explanation does not constitute ISO certification or full implementation of a standard. Consult the official standard for formal testing.
