3.5 Tier A — measured emotional effects
This is the tier on which the entire applied industry relies, and it is the weakest tier.
No molecule-level meta-analysis exists. The largest network analysis in the field, pooling 44 randomised trials in 3,419 people, itself states explicitly that there is no pure compound anywhere in the comparison network. The only two network analyses are both at the level of essential oils or commercial oral preparations.
The most important figure in the whole field comes from that same analysis: when trials with fewer than 30 people per arm are excluded, the pooled effect on the STAI-State anxiety scale falls from −6.63 (95% CI −8.17; −5.08) to −0.84; on the STAI-Trait scale it falls from −4.97 to −1.51. Heterogeneity I² = 93%, 17 of 44 trials are at high risk of bias, and not one comparison reaches moderate GRADE certainty or above. The widely cited effect is almost entirely a product of small trials.
The five effects with the best molecule-level evidence, ranked by robustness rather than by appeal:
| Effect | Preparation | Effect size | Why it holds this position |
|---|---|---|---|
| Cedrol lowers blood pressure, increases baroreflex sensitivity, increases the HF component of heart rate variability | PURE | Not reported in any paper | The effect persists when olfaction is eliminated entirely — delivered directly into the lower airway in laryngectomised people, and in people with anosmia. It cannot be expectation. The most cleanly quantified concentration in the whole set. Public funding. |
| Silexan reduces HAM-A anxiety scores | ORAL MIX | 160 mg: −4.96 (−7.17; −2.76) 80 mg: −3.82 (−5.26; −2.38) I² = 0% | The tightest estimate, but it is an oral preparation and a mixture; all 5 trials were funded by the manufacturer; the paper’s own indirect comparison gave a non-significant result. |
| Hedione does not reduce stress — a negative result | PURE | η²ₚ = 0.00 for cortisol, amylase, VAS, diastolic blood pressure | The best methodology in the whole set and the only fully reported set of effect sizes. The time effect of η²ₚ = 0.25–0.72 shows the design was sensitive enough to detect stress — it simply did not detect the molecule. |
| α-pinene increases HF by 46.8%; D-limonene increases HF by 26.4% | PURE | Percentages and p only; no d, no CI | Pure molecules, counterbalanced crossover design — but n = 13, all female university students, one-sided tests, 90-second exposure. In the α-pinene paper itself, the self-rated “relaxed” score did not reach significance (p = 0.077). |
| Pure linalyl acetate lowers blood pressure | PURE | Not reported; p < 0.01 diastolic | A rare design because concentration was matched to an essential-oil arm. But the between-group comparison for the anxiety index was negative. |
This needs to be said plainly: of the five rows above, only two have confidence intervals, and neither is inhaled single-molecule data. Not one inhaled single-molecule study in the entire literature we found reports a confidence interval. If the standard is “a molecule-level effect, in humans, with an effect size and a confidence interval”, the correct answer is that no case yet meets it.
The case of cedrol deserves to be singled out because it overturns the classification. Its cardiovascular effect persists when the olfactory pathway is eliminated entirely, meaning that the mechanism is absorption through the lungs. Scientifically, cedrol is closer to an inhaled antihypertensive drug than to an aromatherapy. This has two opposing consequences: it is the strongest evidence that a volatile molecule can have a real physiological effect, and at the same time it shows that the mechanism is not the one the fragrance industry usually invokes.
3.6 Tier S — regularities of concentration, mixtures and adaptation
The paradox of this report is that the first three tiers yield few usable regularities, while the fourth tier — psychophysical regularities — yields many quantitative regularities that can be applied immediately.
The three-note limit
People can identify at most about three components in a mixture of up to eight compounds; untrained people reach a maximum of four. Training and expert experience do not raise this limit. A thirty-ingredient formula is not perceived as thirty notes.
“Olfactory white” convergence
When the number of equal-intensity components spanning the odour space exceeds about 20–30, different formulas converge on the same smell. Perceptual similarity rises with the number of components with r = 0.94; the convergence threshold lies between 10 and 20 components. And coverage matters more than number: mixtures of compounds from the same cluster in odour space remain distinguishable (67.5%), whereas widely spanning mixtures do not (42.4%).
A related result: it is possible to create odour metamers — two mixtures sharing not a single ingredient that 87.5% of people cannot tell apart, provided they lie less than 0.05 radians apart in a 21-dimensional descriptor space and are intensity-matched. This has practical significance for ingredient substitution, but the model that produced it assumes there is no chemical interaction, which directly contradicts the receptor antagonism data.
Valence of mixtures
The pleasantness of a mixture is the intensity-weighted average of its components: R² = 0.87–0.95 across every type of mixture tested with 84 people. A simple arithmetic-mean model reaches only R² = 0.14. The component with the higher intensity dominates, regardless of whether it is pleasant or unpleasant. The limit of the model: it cannot predict mixtures whose pleasantness falls outside the range of the two components.
Intensity hypo-additivity and antagonism
The intensity of a binary mixture is about 34% lower than the sum of the intensities of its two components. At the receptor tier, in a screen of about 800 compounds against 10 receptors, the number of antagonisms (430) was roughly equal to the number of activations (444); one receptor had as many as 352 antagonists versus 28 agonists. But there are two important caveats: these data are from mouse receptors, and the test concentration was 300 µM. Odour concentrations in natural air rarely exceed a few parts per billion, whereas 45 of 51 surveyed studies used median concentrations above 1 part per million. At physiological concentrations, components may add far more linearly than the laboratory picture suggests.
Concentration and emotion
The regularity operates on the number of steps above the smeller’s own threshold, not on physical concentration: the correlation is |ρ| = 0.79–0.95 when calculated in steps above the individual threshold, versus 0.51–0.90 when calculated in absolute concentration. And the range of individual thresholds is enormous — in a sample of 26 people, thresholds for trimethylamine spanned 38 twofold dilution steps, a difference of about 2.7 × 10¹¹-fold.
Negative valence appears right at the detection threshold; high concentrations are not needed. And the frequently cited “inverted-U law” — that a pleasant odour has an optimal concentration and then declines — has no original research paper; its source is a textbook chapter. Three studies with data show that pleasant odours rise and then plateau; they do not decline.
Valence reversal with concentration is real but much narrower than it is usually told. For indole, a fourfold increase in concentration reversed the sign of valence in 12 of 27 people — about 44% — while perceived intensity remained unchanged; the other 15 people did not change. For skatole, p-cresol, grapefruit mercaptan, civetone, costus and the aldehydes, no quantitative human data exist.
Adaptation
Perceived intensity falls exponentially and approaches an asymptote of 25–40% of the initial intensity at about four minutes, then stops — it does not fall to zero. In some people the odour still has not disappeared after 15 minutes. The figure “the odour disappears after about 20 minutes” cannot be traced to a source and is refuted by the original psychophysical data.
Recovery is much slower than adaptation: after strong adaptation it can take several hours. And living with an odour for two weeks causes a loss of sensitivity that persists for up to two weeks after exposure stops. Important for compounding: the odour adapts but the trigeminal pungency does not — pungency even increases with inhalation time.
3.7 Detection thresholds
Air detection thresholds span more than four orders of magnitude across common materials, and the relationship with chemical structure is highly non-linear: skatole is 54 times more potent than indole despite differing by only one methyl group.
The variability of these figures themselves needs to be factored into every formula calculation. Within a single standard source, the deviation is ±0.66 log, or about ±4.6-fold. Models predicting thresholds from structure have a deviation of ±6.6-fold. Between different sources, the discrepancy is described as “several orders of magnitude”. The measurement procedure alone shifted thresholds 3.5-fold when presentation was in ascending order and 5.9-fold on re-sniffing.
3.8 Valence: where the variance comes from
Decomposing pleasantness variance across cross-cultural studies gives a clear picture. Individual differences account for about 54% — nine times the share of culture. Molecular identity accounts for 41% in one study but only 14–19% in another using a different odour set, so this figure depends heavily on the choice of compounds and has no universal value. Familiarity accounts for 37%, more than culture. Culture accounts for only 6–7%.
The most often miscited point in this field: pleasantness rankings correlate r = 0.82 across ten cultures, but two people from the same culture agree only τ = 0.32, versus τ = 0.28 for two people from two different continents. “Cross-cultural stability” is a property of the group average, not of individuals.
And absolute score levels differ very strongly between markets: F(6,330) = 216.99. Singaporeans gave significantly lower pleasantness scores than people in Geneva and Liverpool. The practical consequence: ingredient rankings transfer between markets; score levels do not. Ingredient screening can use foreign data; pricing, acceptance forecasting and product positioning must be tested locally.
Scales of odour-evoked emotion do not transfer either: the Singapore version has two dimensions of its own that do not exist in Europe, and each culture’s structure fits its own data better. Translating a European scale into Vietnamese does not produce a valid instrument for Vietnam.
On Vietnamese data: no peer-reviewed study exists that measures the pleasantness of odorants in Vietnamese people. Vietnam has a 12-item odour identification test standardised on 391 people with local odours including fish sauce, soy sauce, guava and mango — but this test does not collect pleasantness data. Vietnam does not appear in any cross-cultural study of odour valence. Southeast Asia has four studies with about 835 people, but the Singapore sample is 79% ethnic Chinese, and all four used odour sets chosen by Western researchers and manufactured in Europe or the USA.
Scientific reference material, not medical advice and not promotional material. All assessments of evidence quality are the compilers’ own. Figures are redrawn from published data.