Physiology & Behavior, Vol. 6(I, No. 1,211-215, 1996 Copyright © 1996 Elsevier Science Inc. Printed in the USA. All rights reserved 0031-9384/96 $15.00 + .00
ELSEVIER
PII S0031-9384(96)00019-0
Semantic-Free Scaling of Odor Quality D A V I D A. STEVENS* A N D R O B E R T J. O ' C O N N E L L t I
*Frances L. Hiatt School of Psychology, Clark University, 950 Main St., Worcester, MA 01610, and ¢'Worcester Foundation for Biomedical Research, 222 Maple Ave., Shrewsbury, MA 01545 USA Received 22 March 1995 STEVENS, D. A. AND R. J. O'CONNELL. Semantic-free scaling of odor quality. PHYSIOL BEHAV 60(1) 211-215, 1996.--The validity of the odor quality reports given by naive human subjects is often questionable. On the one hand, social conventions can influence the labeling of odorants, especially those that have putrid or uncommon odor qualities, and on the other, semantic differences exist for odor descriptors among individuals. We are interested in the individual differences in the quality reports elicited by two nominally putrid odorants, androstenone (AND) and pemenone (PEM). Here we sought to establish empirical support for the individual differences previously obtained in studies of their odor quality, using a nonverbal, semantic-free method of classification. Undergraduate volunteers sniffed a moderate concentration (390 /xM) of PEM, rated its intensity, and provided a verbal odor descriptor. The subjects were then classified as PEM osmic (n = 42) if the quality report was putrid (rancid, urinous, sweaty), allosmic (n = 23), if the quality was nonputrid, and anosmic (n = 39) if no odor was detected. The subjects then sorted 15 odorants matched for intensity, five selected from each of three nominal odor quality types, into as many odor groups as they wished, as long as each group contained all of the compounds with similar odors. The number of times each odorant was paired with another was used as data for an independent multidimensional scaling with ALSCAL, for each class of subject. Three-dimensional solutions showed that this nonverbal, semantic-free scaling method produced odor classifications consistent with those found when each class of subject reported odor qualities from a defined list of quality descriptors. Cluster analysis of the MDS coordinates revealed that these solutions also retained the individual odor quality differences thought to be characteristic of osmic, allosmic and anosmic subjects. Allosmia
Androstenone
Anosmia
Olfaction
INDIVIDUALS appear to differ in both their sensitivity to and the perceived quality of some, perhaps all, odorants (1,14,15,23). Explorations of these individual differences are important as they help direct the search for the physiological mechanisms of olfactory coding, and define the range of perceptual phenomena that must be explained by modern theories of coding. We have been studying individual differences in the intensity and quality reports elicited by the diastereoisomeric ketone, cis-4-(4'-t-butylcyclo-hexyl)-4-methyl-2-pentanone (pemenone), which shares with 5a-androst-16-en-3-one (androstenone) a pronounced urine-sweaty odor (16). Some subjects, osmic to pemenone (PEM), report that it has a strong, generally offensive odor typically called putrid, rancid, urinous, or sweaty. Others, whom we classify as allosmic to PEM, report that it has a weaker, generally mild odor typically called floral, fruity, woody, or vegetable. Still others, anosmic to PEM, are unable to smell it at the highest available concentration (15). Classification of subjects, based on their intensive or qualitative responses to PEM, also reveals additional differences in their perception of certain other odorants [e.g., androstenone (AND), isovaleric acid
Pemenone
Scaling
Similarity
(IV), and phenylethyl alcohol (PA)]. These findings (14,23), generally support a multiple profile model of odor quality coding in which individual odorants elicit activity in several parallel pathways in the olfactory system (17). The validity of the odor quality reports given by human subjects, upon which many of the above distinctions are based, can be questioned on a number of grounds (4). The identification, discrimination, recognition, and description of odors involve both the requisite sensory capabilities and the cognitive processes involved with learning, memory, and language (6,7). Thus, individual differences in both perceptual systems and in life experiences can affect the associations upon which odor identification depends. Differences in culture, context, social conventions, and semantics are known to influence the labeling of odor qualities and a subject's willingness to report them (2,6,7,13,20). All of these factors can influence the labeling of odorants, especially those that have putrid or uncommon odor qualities (10). Because of these difficulties we sought a nonverbal, semantic-free classification method to evaluate the odor quality reports previously provided by PEM osmic, aUosmic, and anosmic individuals.
i To w h o m requests for reprints should be addressed.
211
212
STEVENS AND O'CONNELL
solutions (8,9,11). Because the subjects make no verbal report, differences in scaling that might have occurred because of differences in the willingness to use certain descriptors, as might be expected with putrid qualities, is avoided. Furthermore, individual differences in semantic interpretation do not confound the scaling, as labels and word meanings need not be directly involved in the task. The present work sought to establish empirical support for the individual differences in odor quality previously reported for putrid odorants using this nonverbal, semantic-free sorting method to control for potential differences in response biases. In addition to replicating individual differences in quality reports, we wished to determine the constancy of the verbal quality reports obtained for a expanded range of other odorants. That is, would a larger sample of odorants nominally classified as similar (e.g., putrids, vegetables, and florals) be found similar by PEM osmic, allosmic, and anosmic subjects (3) with this semantic-free, nonverbal sorting method?
Multidimensional scaling (MDS) of similarity judgements provides an ideal method to generate a nonverbal representation of odor quality (1,12,19). MDS puts objects in multidimensional space such that each intraobject distance corresponds to the degree of dissimilarity assigned by the subject to each pair of odors. Objects that are placed close together in these spaces are considered more similar than objects that are placed far apart (12,19). However, because typical methods of obtaining similarity judgements require the subject to sample all possible pairs of odorants, the collection of data for even a small number of compounds is tedious, time consuming, and susceptible to fatigue. Lawless (11) adapted a method first used in the scaling of personality attributes (18) that provides reliable similarity data, but requires only a minimum amount of exposure to the olfactory stimuli. Subjects are simply asked to sample an array of odorants and to sort them into groups, each containing odorants perceived to have similar odors. The frequencies of the pairings of odorants summed over all subjects provides the ordinal data for MDS; that is, those pairings that occur most frequently are considered to have the most similar odors, whereas those that occur least frequently are thought to have the least similar odors. The coordinates of the odorants in the resulting MDS spaces can then be used as input data for cluster analysis to provide an objective method to support the intuitive groupings observed in the MDS
METHOD
Subjects One hundred and four volunteer undergraduate students participated as subjects.
ANOSMIC
ALLOSMIC
1
2
-.2
%
0 "~
~VSzoN2 ] ~
-2
"
" ~ z o ~~ w,2.
@ putrid @ floral/fruity •
~
.2 ~
/2 '
vcgetabl~woody
OSMIC
ta
.
.
° oS L~/~,91OR~
I" ~ ~
-2 2
FIG. 1. Similarity plots of odor quality generated by osmic, allosmic, and anosmic subjects. The orientation of dimensions 1 and 2 have been reversed in the allosmic plot to facilitate comparisons. Table 1 contains the key for the odorant names.
ODOR QUALITY SORTING
213
TABLE 1 ODOR STIMULI GROUPED BY THEIR NOMINAL MODAL DESCRIPTORS Putrid/Urinous
Floral/Fruity
Androstenone (AND)* Butyric acid (BT) Caproic acid (CP) Isovaleric acid (IV) Pemenone (PEM)
Citralva (CT) Lavender (LV) Muget (MG) Orange (OR) Phenylethyl alcohol (PA)
Woody/Vegetable Basil (BS) Celery (CL) Galbanum (GL) Pepper pyrazine (PP) Pinene (PN)
* Abbreviations used in Figs. 1 and 2.
Stimulus Materials Sets of odorants containing each of the 15 test compounds (Table 1) were made up on individual polyester swabs, each holding 150 /~1 of solution, and stored in individual sealed tubes (Falcon Swube® No. 2078). The concentration of each odorant was adjusted to provide equivalent moderate intensities of stimulation. This was judged by a separate subject panel that included both PEM osmics and allosmics. The stimulus materials, grouped by their nominal modal quality descriptor, are listed in Table 1.
Testing Procedure The subjects were first screened by asking them to sniff a swab containing 150 ~1 of 390 /xM PEM and to report the quality and intensity of its odor. Suggested quality labels were provided with an open-ended list of 51 odor quality descriptors (23) with a space to write in additional descriptors. Intensity judgements were indicated on a 9-point line scale with the labels "no odor, faint, moderate, strong, and very strong" marked at points 1, 3, 5, 7, and 9, respectively. The subjects sniffed the PEM swab, circled (or wrote in) the selected descriptor of the odorant's quality, and then circled a point on the line scale representing its intensity. Following this initial screening, subjects were seated at a table and given a set of empty beakers and the 15 stimulus swubes. They were asked to open a swube and to sniff the odorant swab, one swube after another, and to sort them into groups, on the basis of their odor quality. All swubes containing odorants with similar odors were to be placed into the same beaker. They were instructed to ignore any differences in perceived intensity in making their assignment and allowed to return to swubes for additional sniffs if they wished. Subjects were reminded that "too much sniffing could fatigue the sense of smell" and cautioned not to allow the swab to touch their skin. No constraints were imposed on the order of sampling or on the number of groups generated (i.e., beakers containing all of the odorants or only a single compound were permitted). Following the sorting procedure subjects were asked to provide quality labels for each of their groups. RESULTS AND DISCUSSION
The subjects were classified by the quality report they generated for PEM during the initial screening. Forty-two subjects reported that PEM had a putrid odor (e.g., rancid, sweaty, urine) and were classed as osmic. Twenty-three subjects reported a nonputrid odor quality (usually floral, fruity, or green-vegetable like) and were classed as allosmic. Thirty-nine subjects reported no odor for PEM and were classed as anosmic (15). The mean + SEM intensity rating for the PEM screen, expressed in points, reported by the osmic, allosmic, and anosmic groups was 4.36 + 0.32, 2.91 _ 0.24, and 1.00 + 0.00, respectively. This confirms previous studies, which suggested that both quality and quantity
judgements could be used to sort individuals into one class or another (14,15,23,24). Each of the three classes of subjects generated an average of 6.3 odor groups with a range of 3-12 groups overall. The number of odorants in each group ranged from 1 to 7 for each subject class. Following the sorting procedure, subjects reported that 82% of the total number of groups formed were based on perceived similarities in odor quality, 8% on the absence of an odor, and the remainder on various hedonic characteristics. All of the odorant pairs formed by each subject were then enumerated. For example, if swubes containing the odorants A, B, and C were placed together in one beaker, the pairs AB, AC, and BC were scored for that subject. For each class of subject, the frequencies of the 105 possible pairings were determined and used as similarity data for nonmetric multidimensional scaling (MDS) with ALSCAL (19,21,25). The individual three-dimensional spaces generated by ALSCAL are illustrated in Fig. 1 for the PEM osmic, allosmic, and anosmic subjects. These three-dimensional MDS solutions each account for a substantial proportion of the variance in the data produced by the three subject classes with squared correlations (RSQ) of 0.91, 0.80, and 0.84, respectively. Although individual scree diagrams of RSQ indicate that a two-dimensional solution (RSQ = 0.80) would also be appropriate for the anosmic group, we illustrate instead the three-dimensional solution because clustering of individual odorants, not the dimensionality of the spaces, was of interest. Additionally, to facilitate a comparison of the three spaces, the orientation of the axes (but not the coordinates for odors) were selected so that the nominally putrid odorants would be placed in the foreground, and the floral/fruity odorants would be located generally on the right-hand side of the space. An objective determination of the clustering of odorants in each of the three MDS spaces was obtained by submitting their three-dimensional coordinates to hierarchial analysis with CLUSTER (22). The amalgamation diagrams for each of the subject classes are shown in Fig. 2. Simple inspection of the multidimensional spaces and the dendrograms illustrated in Figs. 1 and 2 reveals that the similarity groupings for odorants, generated on the basis of this nonverbal sorting procedure, are in reasonable agreement with those generated earlier for PEM, AND, IV, PP, and PA with verbal descriptors by the individual classes of subjects (14,15,23). For example, osmic subjects continued to judge PEM and AND to be very similar to the other three putrid odorants evaluated (BT, CP, and IV acids) and placed all five of them together in a single cluster. Three "green" odorants (GL, PP, and CL) were grouped into a second cluster, distinct from two further clusters containing all of the floral and fruity odorants. The subjects in this group were classified as osmic on the basis of their verbal rating of PEM's quality in the initial screening. Thus, the proximity of PEM to the other putrid odorants in the cluster was expected and serves to confirm the similarity of the odor quality reports obtained in earlier studies. The allosmic subjects generated a more complex space than either the osmic or anosmic subjects. These subjects placed PEM and AND together in a cluster with PN not far from another cluster containing two additional vegetable odorants (PP and GL) and, as expected, some distance from the cluster containing the remaining three putrid compounds. For these subjects, PEM and AND were judged to be virtually identical in quality, occupying nearly the same point in similarity space. It is important to note that the shift in the similarity judgements for these substances is specific for PEM and AND as the other three nominally putrid odorants remain grouped in a distant cluster. The anosmic subjects generated a similarity space with four
214
STEVENS AND O'CONNELL
Rescaled Amalgamation Distance 5 I
10
15
20
I
I
I
25 I
B_X
iv AND GL pp CL
1
M_GG P_~a LV
cr
--J
BS PN OR
OSMIC
°p©
IV BT
LVcT J MG PA
....}
-J
C_LL
l
}
oR GL pp
~ J
ALLOSMIC
cr ~v P._~A GL aN
cL pp
~ - ~
]
m~ BT
CP
ANOSMIC
FIG. 2. Amalgamation dendrograms produced by hierarchical cluster analysis of the three-dimensional coordinates in the MDS solutions of Fig. 1 using the average linkage between groups method. Members of individual clusters are indicated by underlining the odor labels in alternating dusters within each dendrogram. clusters. Here PEM, AND, and OR were placed together. Because PEM-anosmic subjects are, by definition, insensitive to PEM, a total lack of odor was expected to be the basis for this grouping. The verbal quality reports generated by the subjects after sorting support this expectation; 14 of the 39 anosmic subjects reported a lack of odor for all three of these compounds. This correspondence in the relation between an anosmia for PEM and AND confirms our previous studies but the concomitant anosmia for orange has not been previously reported. Here again, the shifts in similarity judgements are specific for PEM, AND, and OR as the remaining three clusters largely contain compounds sharing the expected modal descriptors.
Taken collectively, these data demonstrate that the differential clustering of odorants with a semantic-free nonverbal sorting method is consistent with the groupings found using traditional quality classification methods that rely on verbal reports. Hedonics typically play a large role in odor classification and in the range of individual differences found in any study of odor quality (5). This appears to be true here as well. In the MDS solutions shown in Fig. 1, dimension 1, which accounts for the largest fraction of the variance in the data, seems to have a hedonic axis for both osmic and anosmic groups, as putrid odorants are found at one end and floral/fruity odorants at the other. Dimension 2 for allosmic subjects seems to have a similar hedonic orientation. A similar conclusion can be obtained by considering the structure of the amalgamation dendrograms shown in Fig. 2. Here it is common to interpret a large jump in the scaled distance required to merge two clusters as an indication of the disparate nature of the clusters (9). For each group of subjects, the largest amalgamation distance in the dendrogram was required for the merger of the nominally putrid cluster containing BT, CP, and IV (and in osmics, PEM and AND) with the remaining clusters of odorants. Although there is no ready method for determining the intrinsic meaning of the dimensions plotted in these similarity spaces, it is clear that the solutions generated seem quite reasonable given our previous experience with the verbal quality reports generated for substances with these odor qualities. In a like fashion, the cluster analysis provides groupings that have heuristic value but the solutions obtained with this method do not verify membership in a cluster or for that matter the actual number of clusters that appear "intuitive" (8,9). In any case, with the exception of PEM and AND, the putrid, floral/fruity, and woody/vegetable quality clusters can be observed in all three classes of subject. As expected, both PEM and AND move about in the individual odor spaces generated by these subjects. PEM-osmic individuals place them in a cluster with the other putrid compounds; allosmics move them to an intermediate position among the floral/fruity and woody/vegetable clusters while preserving the isolation of the other cluster containing the three putrid compounds. The same is true for anosmics except that PEM, AND, and OR are now judged to be odorless. The remaining compounds retain their expected groupings except that BS is clustered with the remaining floral/fruity compounds. The multidimensional spaces of odor dissimilarity presented here (Fig. 1) for osmic, allosmic, and anosmic subjects contain nearly twice as many odorants as those employed in our previous studies of individual differences. The rearrangements observed in the three spaces and in the cluster analysis of their coordinates (Fig. 2) suggest that the sorting method is sensitive, at least to the large-scale individual differences that characterize these classes of subjects (12). ACKNOWLEDGEMENTS
We thank; Dr. G. Ohloff, FIRMENICH SA, Geneva, Switzerland for providing the sample of PEM, and the students and staff of our respective institutions for their assistance. Supported by NIDCD grants DC00131 and DC00371 to R.J.O. Portions of these data were presented at the Association for Chemoreception Sciences 16th annual meeting, April 1994.
REFERENCES 1. Berglund, B.; Berglund, U.; Engen, T.; Ekman, G. Multidimensional analysis of twenty-one odors. Scand. J. Psychol. 14:131-137; 1973. 2. Chastrette, M.; Elmouaffek, A.; Sauvegrain, P. A multidimensional statistical study of similarities between 74 notes used in perfumery. Chem. Senses 13:295-305; 1988.
3. Davis, R. G. Olfactory perceptual space models compared by quantitative methods. Chem. Senses Flay. 4:21-33; 1979. 4. Engen, T.; Ross, B. M. Long-term memory of odors with and without verbal descriptions. J. Exp. Psychol. 100:221-227; 1973. 5. Harper, R.; Bate Smith, E. C.; Land, D. G. Variations in pleasantness
ODOR QUALITY SORTING
6.
7.
8. 9. 10. 11. 12. 13. 14.
and unpleasantness. In: Harper, R., ed. Odour description and odour classification. New York: American Elsevier Publishing Co., Inc.; 1968:142-151. Harper, R.; Bate Smith, E. C.; Land, D. G. The fundamental bases of classification. In: Harper, R., ed. Odour description and odour classification. New York: American Elsevier Publishing Company, Inc.; 1968:89-177. Harper, R.; Bate Smith, E. C.; Land, D. G. Quantitative approaches to classification. In: Harper, R., ed. Odour description and odour classification. New York: American Elsevier Publishing Co., Inc.; 1968:121-141. Hawkins, D. M.; Muller, M. W.; Ten Krooden, J. A. Cluster analysis. In: Hawkins, D. M., ed. Topics in applied multivariate analysis. Oxford: Cambridge University Press; 1982:301-356. Jackson, B. B. Cluster analysis. In: Multivariate data analysis. Homewood, IL: Richard D. Irwin, Inc.; 1983:153-190. Johnston, J. W. The essential nature of the putrid odors. In: Tanyolac, N. N., ed. Theories of odor and odor measurement. Bebek: Robert College; 1966:115-131. Lawless, H. T. Exploration of fragrance categories and ambiguous odors using multidimensional scaling and cluster analysis. Chem. Senses 14:349-360; 1989. Lawless, H. T. Characterization of odor quality through sorting and multidimensional scaling. In: Ho, C.; Manley, C. H., eds+ Flavor measurement. New York: Marcel Dekker, Inc.; 1993:159-183. Lawless, H. T.; Glatter, S.; Hohn, C. Context-dependent changes in the perception of odor quality. Chem. Senses 16:349-360; 1991. O'Connell, R. J,; Stevens, D, A.; Akers, R. P.; Coppola, D. M.; Grant, A. J. Individual differences in the quantitative and qualitative responses of human subjects to various odors. Chem. Senses 14:293302; 1989.
215
15. O'Connell, R. J.; Stevens, D. A.; Zogby, L. M. Individual differences in the perceived intensity and quality of specific odors following selfand cross-adaptation. Chem. Senses 19:197-208; 1994. 16. Ohloff, G.; Gierseh, W.; Thommen, W.; Willhalm, B. Conformationally controlled odor perception in "steroid-type" scent molecules. Helv. Chim. Acta 66:1343-1354; 1983. 17. Polak, E. H. Multiple profile-multiple receptor site model for vertebrate olfaction. J. Theor. Biol. 40:469-484; 1973. 18. Rosenberg, S.; Nelson, C.; Vivekanathan, P. S. A multidimensional approach to structure of personality impressions. J. Pers. Soc. Psychol. 9:283-294; 1968. 19. Schiffman, S. S.; Reynolds, M. L.; Young, F. W. How to use ALSCAL. In: Introduction to multidimensional scaling. New York: Academic Press; 1981:169-209. 20. Schleidt, M.; Neumann, P.; Morishita, H. Pleasure and disgust: Memories and associations of pleasant and unpleasant odours in Germany and Japan. Chem. Senses 13:279-293; 1988. 21+ SPSS Inc. Alscal. In: SPSS-X user's guide. 3rd ed. Chicago: SPSS Inc.; 1988:338-362. 22. SPSS Inc. Cluster. In: SPSS-X user's guide. 3rd ed. Chicago: SPSS Inc.; 1988:405-417. 23. Stevens, D. A.; O'ConneU, R. J. Individual differences in thresholds and quality reports of human subjects to various odors. Chem. Senses 16:57-67; 1991. 24. Stevens, D. A.; O'Connell, R. J. Enhanced sensitivity to androstenone following regular exposure to pemenone. Chem. Senses 20:413-419; 1995. 25. Takane, Y.; Young, F. W.; de Leeuw, J. Nonmetric individual differences multidimensional scaling: An alternating least squares method with optional scaling features. Psychometrika 42:7-67; 1977.