www.sciencemag.org/cgi/content/full/323/5918/1222/DC1

Supporting Online Material for

In Bad Taste: Evidence for the Oral Origins of Moral Disgust

H. A. Chapman,* D. A. Kim, J. M. Susskind, A. K. Anderson*

*To whom correspondence should be addressed. E-mail: hanah@aclab.ca (H.A.C.); anderson@psych.utoronto.ca (A.K.A.)

Published 27 February 2009, Science 323, 1222 (2009) DOI: 10.1126/science.1165565

This PDF file includes: Materials and Methods

SOM Text

Figs. S1 to S3

Table S1

References

Supporting Online Material

Materials and Methods

Experiment 1

Participants: Twenty-seven healthy adults (19 female; mean age 20.6 yrs) participated in

the study for course credit or $20. Participants with past or current history of psychiatric

disorder were excluded, as were participants with a known disorder of taste or smell. All

procedures in this and subsequent experiments were approved by the Research Ethics

Board at the University of Toronto, and participants gave informed consent prior to

beginning the studies.

EMG apparatus: EMG data were acquired using a BIOPAC MP-150 system (S1). At

acquisition, data were amplified ( 1000) and filtered with a bandpass of 100-500 Hz. Electrodes were placed over the levator labii muscle region on the left side of the face,

using the placements suggested by (S2).

Chemosensory stimuli: Solutions of quinine sulfate (ranging from 1.0 x 10-3 to 1.0 x 10-

5M), sodium chloride (5.6 x 10-1 1.0 x 10-1M), citric acid (1.0 x 10-1 1.8 x 10-3M) and

sucrose (1.0 0.18M) were used as bitter, salty, sour and sweet stimuli respectively.

Tastant concentrations were selected individually for each participant by having

participants sample and rate the intensity and valence of 5 concentrations of the bitter and

sour solutions (which were most aversive, as indicated by pilot testing) and 4

concentrations of the salty and sweet solutions. Tastants were presented in small cups,

with a sample size of 2ml. Valence and intensity were rated on 11-point scales, relative to

tap water: participants were instructed to assume that water has zero intensity and neutral

valence. After all the solutions were rated, the experimenter reviewed the participants

responses and selected bitter, salty and sour concentrations with ratings close to very

unpleasant. Sucrose concentrations were selected so as to be approximately equivalent

in perceived intensity to the other solutions (strong intensity).

Procedure: Following the rating procedure, the experimenter applied the EMG

electrodes, set up the sample cups for the first block of the experiment, and left the testing

booth. The experiment was divided into five blocks, one each for water and the bitter,

sweet, salty, and sour solutions. Because the taste of quinine lingers in the mouth for an

extended period, the bitter block was always presented last, while the other tastants were

presented in counterbalanced order. Blocks consisted of five taste trials in which

participants sampled 2ml of the appropriate liquid. To control the timing of each trial,

written instructions were presented on a computer monitor using E-Prime experimental

software (S3). Trials began with a 20s baseline period during which participants rested

quietly. Next came an 8s period when participants grasped the first sample cup and

brought it to their lips. This was followed by 8s during which participants sipped the

contents of the sample cup, swished the liquid twice in their mouth and held the liquid in

their mouth without swallowing. Participants next swallowed the liquid, and after a

further 8s, rinsed their mouths with ~10ml of water, swallowing when finished. Finally,

participants rated the valence and intensity of the preceding sample using the scales

described above. After participants completed all five trials of a block, the experimenter

entered the booth to arrange the sample cups for the next block.

2

EMG analysis: EMG data were analyzed using custom software written in MATLAB

(Version 7; S4). A 30 Hz high-pass filter was applied to the EMG data to reduce low-

frequency noise. The signal was then rectified with an absolute value function and

smoothed with a 10ms window. The period of interest in each trial was the 8s of the

swallow phase, which is least likely to be contaminated by extraneous muscle

activation due to sipping. Signal from the final 8s of the preceding baseline period was

used to calculate signal change during the swallow phase. A contour following integrator

was then applied to compute a running sum of sample values. The final value of the sum

at the end of each swallow period was used for statistical analysis.

Statistics: The correlation between levator labii region EMG and valence was computed

on intertrial variability, in part because of variability in the overall strength of the EMG

response across individuals. Each trial yielded both an EMG response and a paired self-

report of valence. For each participant, the EMG/self-report pairs for all 25 trials were

rank-ordered by decreasing valence. The EMG responses at each rank were then

averaged across participants, and the average values were correlated with valence rank.

Experiment 1b

Appearance model: An additional group of 20 participants (13 female, mean age 23.2)

were filmed as they underwent the taste sampling procedure described above in order to

train a computerized facial appearance model to uncover the underlying action tendencies

associated with the distaste response. Since not all participants showed visible facial

actions in response to ingestion of the unpleasant tastants, the upper quartile with the

3

strongest overt facial movements were included for training the appearance model. Still

images capturing these participants peak responses to the bitter, sweet and neutral

tastants were pulled from the video clips; responses to salty and sour tastants were

excluded from the analysis so as not to bias the model toward unpleasant tastes. For those

individuals who were wearing eyeglasses, the spot healing brush tool in Photoshop was

used to remove the glasses from the still images, to reduce artifacting in subsequent

analysis steps.

A facial appearance model was created from these images using the Cootes et al.

methodology (S5). First, each of the face exemplars was manually annotated with the

two-dimensional Cartesian coordinates (relative to the image frame) of 68 distinct facial

landmarks. The set of landmarks conforms to an annotation scheme designed to provide

feature information for the facial contour, eyebrows, eyes, nose, and mouth, based on a

subset of features specified by the MPEG-4 coding standard for facial expression

animation (S6). All faces were cut out from the background image using a polygonal

region defined by the outer contour delimited by the labeling scheme. Pixels inside the

face region were histogram equalized to remove variability due to lighting differences

across faces. All of the faces were then commonly aligned to remove global differences

in size, rotation, and translation, using a generalized Procrustes transformation (S7).

Principal components analysis (PCA) was applied to the covariance matrix of the shape

data using eigenvalue decomposition, resulting in eigenvalue/eigenvector pairs used to

parameterize uncorrelated dimensions of shape variation. The pixel textures for each face

were piecewise affine warped to a common shape-normalized reference frame defined as

the mean shape computed from the entire set of annotated facial landmarks. Textures

4

extracted in this fashion are known in the literature as shape-free, having the property

that internal features in the texture map for each face are relatively well-aligned (5). PCA

was then applied via eigenvalue decomposition on the covariance matrix of the shape-

free pixel vectors, resulting in parameters describing uncorrelated dimensions of texture

variation. In order to collapse important covariations between face shape and texture, a

third, combined PCA was performed on concatenated shape and texture parameter

vectors (S5). First, the shape PCA projections were re-weighted to the same variance

metric as the texture projections by multiplying the shape coefficients by the sum of the

texture eigenvalues over the sum of the shape eigenvalues. Then the shape and texture

vectors were concatenated into single vectors for each face exemplar, and eigenvalue

decomposition was performed on the covariance matrix of the combined shape and

texture vectors. The resulting combined shape and texture basis retained the full rank of

the covariance matrix of the face data (there was no data compression). Each face in the

dataset was thus parameterized by a vector of appearance loadings.

Photorealistic face images were created using the appearance model by reversing the

process that converted images to appearance loadings: given a vector of combined PCA

loadings, the vector was transformed back into a concatenated shape and texture vector,

and the shape coefficients were re-weighted to their original metric. Then the shape and

texture PCA coefficients were transformed back into Euclidean shape coordinates and

shape-free pixel values. Finally, the shape-free pixels were warped to the specified shape

coordinates from the shape coordinates of the average face. Expression prototypes for

bitter, neutral, and sweet tastants were created by averaging the vector representations of

all exemplars of a category and synthesizing face images via the above process. The

5

resulting prototype taste images were exaggerated in intensity to accentuate their

characteristic action tendencies by scaling each prototype vector by 2.

Experiment 2

Participants: Nineteen healthy adults participated in this experiment. Participants with

current or past history of psychiatric disorders were excluded. EMG data from one

participant were eliminated for technical reasons, for a final sample size of 18 (8 female,

mean age 20.4 years).

Stimuli: Twenty disgusting and twenty sad photographs were chosen from the

International Affective Picture System (IAPS; S8) in a two-step selection process. In the

first step, published emotional category ratings (S9) were used to identify disgusting

photographs that were rated as high in disgust but low in sadness, as well as sad

photographs that were high in sadness but low in disgust. A group of 20 neutral

photographs low in both sadness and disgust was also selected.

In the second step, ratings from the IAPS norms were used to match the sad and

disgusting photographs on mean valence. The final set of 20 disgusting photographs was

largely composed of images of contamination and uncleanliness, such as body products,

body envelope violations, and insects. The 20 sad photographs included (among others)

homeless persons, traffic accidents, and distraught or ill individuals (illness photographs

did not involve overt injury or mutilation, contagious disease or body products). Neutral

photographs included household objects, outdoor scenes and abstract images.

6

Procedure: The experimenter first applied the EMG electrodes and left the room. The 60

photographs were then presented once in random order. Each trial began with a 6s

baseline period during which participants viewed a fixation cross, followed by 6s of

image presentation. Participants then rated the preceding image on sadness and disgust,

using 9-point scales, before continuing to the next photograph.

EMG analysis: EMG data were processed largely as in the first experiment. The period of

interest in each trial was the 6s image presentation period. Signal from the 6s of fixation

preceding each image was used as a baseline to calculate signal change scores.

Statistics: The correlations between levator labii region EMG and disgust/sadness ratings

were computed similarly to Experiment 1. Disgust and sadness ratings for all 60

photographs were rank-ordered for each participant, in order of increasing disgust or

sadness. The paired levator labii region responses at each rank were then averaged across

participants, and these values were correlated with rank.

Experiment 3

Participants: Twenty-one healthy volunteers with no current or past history of

psychiatric disorder participated in this study. Five participants were disqualified because

debriefing revealed that they had suspected the deception involved in the experiment. The

final sample size was 16 (11 female; mean age 22.7 yrs).

EMG acquisition: EMG data were acquired as in the previous experiments.

7

Procedure: Participants were tested in groups of 2-3, with each participant seated in a

private booth. The experimenter first explained the nature and rules of the UG, and

informed participants that they would always play the role of responder. Participants

were next introduced to a group of 10 proposers (confederates) seated at computer

terminals in two adjoining rooms. All players were told that they would be paid

according to their choices in the game. Participants returned to the testing room and the

EMG electrodes were applied.

In the UG, each participant played 20 rounds of the game, one with each of the 10

confederates and 10 with an avowed computer partner. In reality, all offers were

generated by a computer algorithm so as to control the size and number of offers made.

Rounds were presented in random order. The 10 offers from the computer partner were

identical to those from the human partners, and mimicked the range and distribution of

offers made in uncontrolled versions of the game, in which actual human proposers make

offers (e.g., S10): 50% fair ($5:$5) offers, 20% $9:$1 offers, 20% $8:$2 offers and

10% $7:$3 offers. No $6:$4 offers were presented, as we were primarily interested in

exploring the response to offer that were more strongly unfair. As offer acceptance, self

reported emotions, and EMG activity did not differ significantly between partner types

(see supplementary results) data were collapsed across partner type for all analyses.

Each UG round began with a 6s waiting period followed by 6s of fixation. The

participant then saw the photograph and name of their partner in that round for 6s. Next,

participants saw the offer proposed by their partner for 6s, after which came a 6s window

in which they indicated whether they accepted or rejected the offer by pressing one of

8

two keys on the keyboard. To reinforce the social nature of the interaction, names and

photographs of proposers remained visible throughout the offer and decision phases.

After seeing the trials outcome displayed for 6s, participants rated, on a 1-7 scale,

how well their feelings about the preceding offer were represented by a reliably

recognized facial expression of a universal emotion (S11). The seven emotion rating

slides (happiness, sadness, anger, fear, disgust, surprise, contempt) that followed each

offer were presented in random order. Finally, participants were debriefed and paid for

their participation according to their choices in the games.

EMG analysis: Using the same custom software, EMG data were first filtered with a 55-

65Hz notch filter and then with a 30Hz high-pass filter, before rectifying and smoothing

as above. We analyzed the 6s period during which the proposers fair or unfair offer was

revealed; recordings from 6s before the onset of the offer display (i.e., during the partner

display) were used as a baseline from which to calculate signal change scores. Lastly, a

contour following integrator was applied.

Statistics: Preliminary analyses revealed only a few small differences between responses

to human vs. computer partners. Results presented in the main text thus give combined

data from all 20 trials.

The correlations between disgust, anger, sadness and contempt endorsement and

levator labii region EMG were computed similarly to the previous experiments. Emotion

ratings for all 20 trials were rank-ordered for each participant by increasing endorsement,

9

and the corresponding EMG values were averaged across participants at each rank.

Finally, the average EMG values were correlated with emotion rank.

The correlations between levator labii region activity and behavioral responses to

offers were computed by rank-ordering the levator labii region EMG response for all 10

unfair trials (offers of $7:$3 or less) for each participant, in order of increasing activation.

The corresponding accept/reject decisions (1 or 2) at each rank were then averaged across

participants, and the average values were correlated with emotion rank. The correlations

between emotion endorsement and behavioral response were computed similarly, except

that decisions were rank-ordered by increasing self-reported disgust, sadness or anger.

Labeling of expressions used in self-report task: To examine whether the non-verbal self-

report method that we employed separates anger from disgust, a separate group of 12

participants (7 female, mean age 27.1 years) matched canonical emotional facial

expressions to written emotion descriptors. On each trial, participants viewed an array of

emotional expressions (angry, disgusted, contemptuous, sad, fearful, surprised and happy;

S11) as well as a single written emotion descriptor. The task was to select the expression

that was the best match for the descriptor. Descriptors were synonyms for the classical

emotion terms. Four disgust-themed descriptors were presented (tastes something bad,

smells something bad, touches something dirty, grossed out), as were four anger-

themed labels (frustrated, annoyed, pissed off, irritated). Descriptors for

contempt (disapproving), happiness (cheery), surprise (amazed), fear (scared) and

sadness (glum) were also presented. For analysis, responses to the four disgust-themed

descriptors were collapsed together, as were response to the four anger-themed

10

descriptors. The frequency with which the anger and disgust expressions were selected to

match the anger- and disgust-themed descriptors was analyzed using chi-square tests.

Appearance model: The pattern of negative emotions endorsed for $9:$1 offers was

depicted using modified versions of the faces used in the self-report task. The larger

facial appearance model was derived from a standard cross-cultural dataset of posed

facial expressions that included the self-report faces (S11), using the same methodology

described for Experiment 1 above. The intensity of disgust, anger and sadness

expressions used in the self-report task was scaled by the degree of emotion endorsement

for each expression. More specifically, the vector representations for each face were

multiplied by the ratio of the mean emotion endorsement value for that emotion to the

maximum possible endorsement. This resulted in new faces whose expression intensity

visually reflects the strength of emotion endorsement for that emotion.

Indexing similarities between expressions: Similarity between the disgust expression

prototype and the bitter expression from Experiment 1 as well as seven canonical facial

expressions of emotion (S11) was assessed using the appearance model. The bitter

expression prototype was fit to the appearance model after training the model on images

of the seven basic emotions, allowing a test of the generalization of the model to a novel

expression. Pearson correlations were measured between the vector coding the disgust

expression prototype and the vectors coding bitter, anger, contempt, sadness, surprise,

fear and happiness.

11

Results

Comparison of responses to human vs. computer partners in Experiment 3: Previous

research has found that behavioral and neural responses to unfair offers differ depending

on whether the offer was made by a human or a computer partner (S12). In view of these

findings, we included a manipulation of partner identity in Experiment 3, with the

expectation that unfair offers from a computer partner might evoke weaker emotional

responses. However, contrary to this hypothesis, we did not find strong differences

between responses to humans and computers. For example, our participants accepted

only marginally fewer unequal offers from computers than from humans (paired samples

t-test: t[15] = 1.88, p = 0.08). In particular, participants rejected the most unfair offers

($9:$1) as often from computers as from humans (t < 1). Self-reported emotion in

response to the offers did not differ significantly between humans and computers

(repeated measures ANOVA: F[1,15] = 1.30, p = 0.28). Finally, levator labii region

EMG did not differ between offers from humans and computers (repeated measures

ANOVA: F[1,15] = 2.0, p = 0.18).

It is not clear why we failed to replicate the previously observed differences due to

partner identity, but we note that there are other experimental paradigms in which

participants experience strong emotional reactions to computers. For example, exclusion

from a virtual ball-toss game results in anger and hurt feelings as well as decreases in

self-esteem and feelings of belonging (S13). The strength of these reactions does not

depend on whether participants believe they are playing with a computer or with other

humans (S13). Thus, individuals may not always distinguish strongly between human and

computer partners in social interactions, as our participants evidently did not.

12

Labeling of expressions used in self-report task: Our aim in this experiment was to test

whether anger and disgust facial expressions were matched to distinct (and appropriate)

emotion descriptors. Results supported a separation between labeling of anger and disgust

expressions: anger expressions were reliably matched to anger descriptors and not disgust

descriptors, while disgust expressions were matched to disgust descriptors but not anger

descriptors (Table S1). Averaging across the four anger-themed descriptors, 60.4% of

participants chose the anger expression as the best match, while disgust was selected by

only 12.4%. Across the four disgust-themed descriptors, the opposite pattern obtained:

85.4% of participants chose the disgust expression as the best match, while 4.2% chose

the anger expression. Chi-square tests on the frequency of selection of disgust and anger

expressions showed that these differences were significant (anger themed descriptors,

2(1) = 17.8, p < 0.001; disgust-themed descriptors, 2(1) = 35.4, p < 0.001).

13

Captions

Fig. S1. Mean proportion of offers accepted in Experiment 3, by offer type. Error

bars show +1 SEM, calculated within-subjects (S14).

Fig. S2. Mean self-reported emotion in response to different offers in the

Ultimatum Game (N = 16), showing all emotions. Higher numbers indicate

greater endorsement. Emotions that did not vary significantly with offer size are

shown in dashed lines.

Fig. S3. Analysis of similarity between the computer model of the disgust

expression prototype and other expression prototypes developed for Experiment

3, plus the average distaste expression observed in Experiment 1B. Average

expressions of disgust and the other basic emotions were generated from

photographs from a standard set (S11). Points on the plot show Pearson

correlations comparing the visual similarity of the average disgust expression

vector representation to other average expressions and the average distaste

expression from Experiment 1B. Confidence envelope shows 95% CI.

14

Tables

Table S1. Match between emotion descriptors and facial expressions used for

the self-report task in Experiment 3. The proportion of times each expression was

selected as the best match for a particular descriptor is given (averages across

the four items are shown for disgust and anger). Modal response is highlighted.

Emotion Descriptor

Facial

Expression

Disgust-

themed

Anger-

themed

Disapproving Glum Amazed Scared Cheery

Disgust 0.85 .12 .17 0 0 0 0

Anger .042 .60 .083 0 0 0 0

Contempt 0 .23 .67 0 0 0 0

Sadness .083 .042 .083 1.00 0 .083 0

Surprise 0 0 0 0 1.00 .083 0

Fear .021 0 0 0 0 .83 0

Happiness 0 0 0 0 0 0 1.00

15

16

References and Notes

S1. Biopac Systems. (Goleta, CA).

S2. L. G. Tassinary, J. T. Cacioppo, in Handbook of Psychophysiology, J. T.

Cacioppo, L. G. Tassinary, G. G. Berntson, Eds. (Cambridge University Press,

Cambridge, England, 2000) pp. 163-198.

S3. Psychcholgy Software Tools. (Pittsburg, PA).

S4. The MathWorks. (Natick, MA).

S5. T. F. Cootes, G. J. Edwards, C. J. Taylor, IEEE Trans. Pattern Anal. Mach. Intell.

23, 681 (2001).

S6. J. Ostermann, Comput. Animat. 98, 49 (1998).

S7. C. Goodall, J. Roy. Stat. Soc. B 53, 285 (1991).

S8. P. J. Lang, M. M. Bradley, B. N. Cuthbert, International Affective Picture System

(IAPS). Technical Report A-6. (University of Florida, Gainesville, FL, 2005).

S9. J. A. Mikels et al., Behav. Res. Methods 37, 626 (2005).

S10. W. Gth, R. Schmittberger, B. Schwarze, J. Econ. Behav. Organ. 3, 367 (1982).

S11. D. Matsumoto, P. Ekman, Japanese and Caucasian Facial Expressions of

Emotion (JACFEE) [Slides] (Intercultural and Emotion Research Laboratory,

Department of Psychology, San Francisco State University, San Francisco, CA,

1988).

S12. A. G. Sanfey, J. K. Rilling, J. A. Aronson, L. E. Nystrom, J. D. Cohen, Science

300, 1755 (2003).

S13. L. Zadro, K. D. Williams, R. Richardson, J. Exp. Soc. Psychol. 40, 560 (2004).

S14. G. R. Loftus, M. E. Masson, Psychon. Bull. Rev. 1, 476 (1994).