This study was approved by the Western Institutional Review Board (Puyallup, WA) (WIRB Protocol #20170080). All participants provided informed written consent using a form approved by WIRB. At no time did participants come into direct contact with the cannabis samples. Retail sale of marijuana for recreational use to adults 21 years of age and older has been legal in the state of Colorado since January 1, 2014.

Test participants were recruited from the local community by means of fliers and an online newspaper advertisement. Printed text emphasized “current, former, and non-users all welcome” and that only sniffing was required (“no touching, no smoking, no eating”). All participants were at least 21 years of age, residents of Colorado, and had a self-reported normal sense of smell. Exclusion criteria included self-reported pregnancy, active nasal allergy, and current head cold. Subjects were paid $20.00 for their participation.

Cannabis samples in the form of dried flower were purchased from two Fort Collins dispensaries licensed for retail recreational sale by the state of Colorado: Solace Meds and Infinite Wellness Center. Eleven strains were purchased: Alien Dawg, Durban Poison, Fruity Pebbles, G13, Jilly Bean, Lamb’s Breath, Lemon Diesel, Mob Boss, OG Kush, Snoop OG, and Super Skunk. (Details regarding the dispensaries, cultivation facilities, and on-label information about the strains are provided in S1 Note. Source and specification of study materials.) These strains were selected to cover a wide range of odor character, based on information from web-based “strain reviews” as well as pre-purchase sniff sampling at the dispensaries by the senior author. One strain (Durban Poison) was purchased from two dispensaries; both duplicate strain samples were included as a control for retail market consistency. A duplicate sample of one strain (G13) was included as an internal control for panelist consistency. Thus a total of 13 samples was presented to each panelist.

Each stimulus (1 g of dried cannabis flower) was presented in a wide mouth 118 ml (4 oz) amber glass bottle labeled with a three-digit code. Samples were kept in a freezer at -2° C, and thawed at room temperature for two hours before testing. The stimuli were exchanged for fresh samples midway through the study.

In this paper, we identify the odor stimuli according to strain designations provided by the dispensary. The taxonomic validity of named strains is the subject of debate [11] as is the derivation of modern, commercially available, hybrid strains [12]. In their exhaustive treatment of the genus, Clarke and Merlin stated that cannabis presents a “confusing taxonomic situation” and a “challenging botanical problem” [1]. They conclude that “Cannabis should be divided into seven taxonomically distinct biotypes of three species with six putative subspecies levels.” In view of these complexities, we note that our use of retail strain designations to identify the samples is purely a matter of convenience. In any case, the precise taxonomic status of our samples is not necessary for our present interest in exploring and characterizing olfactory variation in commercially available offerings.

Forty-eight odor descriptors (Table 1) were culled from online resources aimed at cannabis aficionados, e.g. https://www.leafly.com. Given the exploratory nature of our study, this list was deliberately over-inclusive. The descriptors were printed in alphabetical order on the paper ballot provided with each sample.

The printed ballot accompanying each stimulus included three product evaluation questions. Each was prefaced with “Based only on the smell” and concluded with either “how potent do you think this sample of cannabis is?”, “how interested would you be in smoking this sample of cannabis?” or “how much would you expect to pay for a gram of this cannabis?” Response options for the first two questions were 11-point scales (with labeled endpoints) ranging from zero (“not potent/not interested at all”) to 10 (“extremely potent/extremely interested”); those for the third question consisted of 11 values ranging in two-dollar increments from $2/g to $22/g.

After signing the consent form, participants were asked whether they had (a) purchased and (b) smoked cannabis flower since it became legal in the state of Colorado on January 1, 2014. Subjects were tested individually in a well-ventilated 6.0 x 2.6 m room with neutral decor. The order of sample presentation was randomized for each subject. Panelists were instructed to sniff the sample ad libitum and to check as many or as few odor descriptors as applied to it. It was emphasized that there were no right or wrong answers. The test administrator and all but the current sample bottle were hidden behind an opaque white partition.

All reagents used were of analytic grade. Δ-9-THC (CAS No. 1972-08-3, from Restek Corp., Bellefonte, PA, USA) at a concentration of 0.50 mg/mL in methanol was used as a standard. This solution also contained phenanthracene at a concentration of 0.50 mg/mL which served as an internal standard for chromatography.

Solutions from each sample of cannabis were prepared for chromatographic analysis by harvesting 25 mg of material near the tip of the inflorescence. The material was macerated and suspended in 1.0 ml of methanol containing 0.50 mg/mL of phenanthracene. This suspension was vortexed for 10 min then centrifuged (2,000 g x 10 min). Aliquots of the supernatant were taken for chromatographic analysis. Separate extraction and standard addition experiments quantified the extraction efficiency of this method and demonstrated minimal matrix effects.

Chromatographic analysis was performed on a Hewlett Packard 6890 II GC with a low-polarity capillary column (Rxi-5ms, 30 m, 0.25 mm inner diameter, 0.25 mm thickness, Restek Corp.) followed by a Flame Ionization Detector (FID). Linear flow of the hydrogen carrier gas (99.995% purity) was held at 1.0 mL/min. Samples of volume 0.6 μL were applied to a split injector (1:20) maintained at 250°C. The column oven temperature was programmed as follows: 40°C for 3.0 min then increased to 250°C at a rate of 25°C per min (total time 8.4 min) and finally held at 250°C for 8.6 min (20 min total run time). The detector temperature was maintained at 250°C. The FID signal was digitized (U6-Pro, LabJack Corp., Lakewood, CO, USA) and analyzed (Igor Pro, WaveMetrics, Inc., Portland, OR, USA) using custom built tools.

The reported percentage composition (mass of the individual component per mass of as-received sample) was determined based on relative response in the FID chromatograms. THC identification was based on direct comparison with an authentic sample. All measurements were performed in replicate with a relative standard deviation exceeding 5%.

The chromatograph was tuned daily; analysis of blanks and standards showed no contaminating compounds, and validated instrument response both in signal intensity and retention time.

Sixty-one people (33 men, 28 women; mean age 28.2 ± 8.4 years) took part in the study. Of these, all but three had purchased cannabis since January 1, 2014, and all but one subject had smoked it. The high rates of purchase (95.1%) and use (98.4%) among study participants occurred despite efforts to recruit former and non-users as well. As an unintended result, the study resembles a typical consumer perception study where purchase, use, and/or familiarity with the product is a requirement of participation.

Participants rated the 13 samples using the check-all-that-apply list of 48 descriptors (Table 1). The number of descriptors endorsed per sample by a given participant (mean across 13 samples) ranged from 1.9 to 11.8; i.e., some participants used descriptors more sparingly than others. The overall mean (± SD) number of descriptors per participant was 5.2 ± 2.8 (n = 61).

The mean number of odor descriptors assigned to a cannabis sample ranged from 4.5 (Lemon Diesel) to 6.1 (Jilly Bean) (overall mean = 5.2). The number of odor descriptors endorsed for a given sample by a given participant ranged from 1 to 18 (Fig 1). Endorsement frequencies of 3, 4, and 5 accounted for 423 or 53.3% of all sample evaluations. Endorsement frequency ≥10 accounted for 71 or only 9% of all sample evaluations. These results raise the possibility that the strains examined here might be adequately described with fewer than 48 odor descriptors.

Frequency of use for each of the 48 descriptors is shown in Fig 2. The most frequently used descriptor was earthy (328 endorsements or 7.9% of total); the least frequently used descriptor was tar (17 endorsements or 0.4% of total). The 11 most frequently used descriptors accounted for 51.5% of all endorsements, while the top 23 accounted for 76.2% and the top 34 for 90.2%. That the 15 least frequently used descriptors accounted for only 10.8% of total endorsements indicates that many odor descriptors were only rarely used. The most frequently used descriptors were from familiar categories of everyday experience such as woods, spices, florals, and citrus fruits, but also included notes such as diesel, skunk, pungent and cheese.

A strain’s aroma profile can be represented as frequency counts across the odor descriptors. Two examples are provided in Fig 3. Visual inspection reveals striking differences between OG Kush and Durban Poison (vendor 1). The most common terms used to describe OG Kush are earthy, herbal, and woody, along with flowery, sage, and nutty. In contrast, Durban Poison (vendor 1) was characterized as citrus, lemon, and sweet, as well as pungent, lime, and diesel.

Using endorsement frequencies of the 48 descriptors from all 61 participants as input data, the cannabis samples were analyzed with the Hierarchical Cluster procedure in IBM SPSS Version 23, using between-groups linkage on squared Euclidean distances. The resulting dendrogram (Fig 4) suggested two clusters, with Cluster A consisting of eight stimuli and Cluster B of five. Cluster A includes both duplicate samples of strain G13, indicating that participants used the odor descriptors consistently across samples. Cluster B includes both independent retail samples of Durban Poison, indicating olfactory consistency in retail offerings (at least for this strain).

By referencing the odor descriptor frequencies for each sample it is possible to characterize the olfactory profiles of each cluster. The most frequent descriptors for samples in Cluster A were uniformly earthy, woody, and herbal. For samples in Cluster B, the most frequent descriptors included citrus, lemon, sweet, and pungent.

Each cluster had one less closely associated sample (Lemon Diesel in Cluster A; Durban Poison (vendor 1) in Cluster B). Research with more refined measures (e.g., multipoint rating scales) and a larger number of strains may provide evidence that these samples constitute distinct clusters of their own. Under the present circumstances, however, the most parsimonious interpretation is that they segregate with the two primary clusters.

Of potential concern is the large number of clustering variables relative to the number of stimuli [13]. This raises the possibility that inter-correlation between variables could lead to common odor notes being overrepresented in the cluster solution. Principal Components Analysis or a similar technique is sometimes used to reduce the number of clustering variables. A known risk of this approach, however, is the potential loss of valuable information [13]. This risk is particularly acute in this study, which is an initial attempt to identify relevant items for a cannabis sensory lexicon. Therefore, we took two different approaches to assess the robustness of the cluster solution.

Because the odor descriptors used as clustering variables included the term “citrus” as well as four specific citrus items (grapefruit, lemon, lime, orange), it is possible that shared aspects of these variables were overrepresented in the clustering solution. To address this possibility, we ran another HCA in which the descriptors grapefruit, lemon, lime, and orange were omitted from the analysis. The resulting dendrogram suggested two clusters consisting of the same strains as in Fig 4, and with nearly identical within-cluster arrangements. This suggests that the basic division of the stimuli into two clusters is not driven by an overrepresentation of citrus-related terms in the odor descriptor set.

Next, to assess the robustness of the cluster solution using fewer clustering variables, we re-ran the cluster analysis but omitted the 15 least frequently endorsed items. While differing in some details, the resulting dendrogram produced two clusters comprised of the same stimuli as in Fig 4. Subsequent HCA runs that omitted increasing numbers (20, 25, 30, and 35) of the least frequently endorsed descriptors produced similar results. It thus appears that cannabis flowers of various strains cluster robustly based on olfactory profiles consisting of a handful of descriptive terms (e.g., earthy, herbal, woody, flowery, sweet, citrus, pungent, and pine). Another implication is that a useful cannabis lexicon may be obtained with fewer than 48 odor descriptors.

To further characterize the clusters identified in Fig 4, a composite odor profile was created for each of them (Fig 5). It displays descriptor endorsements as a percentage of the possible opportunities (samples x participants) for the samples in Cluster A (n = 488) and Cluster B (n = 305). Some descriptors (e.g., spicy) were endorsed in roughly equal proportions in both clusters. Others (e.g., earthy) are more heavily represented in one cluster than the other. The composite aroma profiles are consistent with the characterization of Cluster A as earthy, herbal, and woody, and Cluster B as citrus, lemon, sweet, and pungent.

The cluster membership of each sample (based on the HCA results in Fig 4) is provided in Table 2. It appears that samples in the Citrus group (Cluster B) were perceived as more potent than those in the Earthy group (Cluster A), and also yielded higher scores for smoke interest and estimated price. To test this observation statistically, a single planned comparison of Cluster A (Earthy) versus Cluster B (Citrus) was carried out using aggregated ratings for each variable in a repeated measures ANOVA. The results confirm a statistically significant difference between sample clusters for each product evaluation variable: Potency F(1, 60) = 88.01, p < 0.001; Interest F(1, 60) = 86.58 p < 0.001; Price F(1, 60) = 98.23, p < 0.001. Group means (± SD) were 7.8 ± 1.1 vs 6.6 ± 1.2 for Potency; 8.2 ± 1.2 vs 6.9 ± 1.5 for Interest; and 13.13 ± 3.07 vs 11.42 ± 3.18 for Price (Cluster B/Citrus versus Cluster A/Earthy, respectively). These results indicate that group-level differences in strain aroma are linked to significant differences in consumer estimates of product quality. In particular, the citrus, lemon, sweet, and pungent character of Cluster B is associated with greater perceived quality than the earthy, herbal, and woody character of Cluster A.

The THC concentration of each sample as determined by our chemical analysis is listed in Table 3. There was a 3.5-fold difference in concentration between the sample with the least THC and the sample with the most. Colorado regulations require that THC concentrations be stated on the retail label as a range representing the lowest and highest percentage values drawn from every test conducted on a given strain, from the same cultivation facility, within the last six months. Our experimentally determined THC content was lower than the range of values listed on the retail label for at least 6 of the 12 samples (see S2 Note. THC experimental versus label). A similar overstatement of THC content has been reported for edible cannabis products from California and Washington [15].

To determine whether perceived potency based on smell varied systematically with actual THC concentration, we calculated the Spearman rank-order coefficient between THC value and mean Potency rating for each strain. Because samples G13-1 and G13-2 were split samples of the same physical material, we measured THC in strain G13 only once. Rather than arbitrarily pair the THC data with the mean Potency of one sample or the other, we ran the analysis twice. We calculated the Spearman correlation coefficient once using the value for G13-1, and again using the value for G13-2. In neither case was perceived potency correlated with measured THC (using G13-1, r_s = -.021, ns; using G13-2, r_s = -.098, ns; n = 12 in each analysis). Thus, while participants perceived large differences in sample potency based on smell, these perceptions do not correspond to the THC content of the samples. Similar analyses showed that neither Interest nor Price were correlated with THC content (S3 Note. Experimental THC content versus Interest and Price).

In addition, correlation analysis of the Product Evaluation variables versus both the lower- and upper-range THC values reported on the retail labels also found no evidence of a relationship (see S4 Note. On-label THC content versus product evaluations).