A comparison of the discrete
and dimensional models of
emotion in music
Tuomas Eerola and Jonna K.Vuoskoski
University of Jyväskylä, Finland
Abstract
The primary aim of the present study was to systematically compare perceived emotions in music
using two different theoretical frameworks: the discrete emotion model, and the dimensional
model of affect. A secondary aim was to introduce a new, improved set of stimuli for the study of
music-mediated emotions. A large pilot study established a set of 110 film music excerpts, half
were moderately and highly representative examples of five discrete emotions (anger, fear, sadness,
happiness and tenderness), and the other half moderate and high examples of the six extremes of
three bipolar dimensions (valence, energy arousal and tension arousal). These excerpts were rated in
a listening experiment by 116 non-musicians. All target emotions of highly representative examples
in both conceptual sets were discriminated by self-ratings. Linear mapping techniques between the
discrete and dimensional models revealed a high correspondence along two central dimensions that
can be labelled as valence and arousal, and the three dimensions could be reduced to two without
significantly reducing the goodness of fit. The major difference between the discrete and categorical
models concerned the poorer resolution of the discrete model in characterizing emotionally
ambiguous examples. The study offers systematically structured and rich stimulus material for
exploring emotional processing.
Key words
battery, dimensional, music, discrete, emotion, three-dimensional
Music has the ability to convey powerful emotions. This ability has fascinated researchers as
well as the general public throughout the ages, and although great strides forward have been
made in the field of music and emotion research, much remains unclear. One issue that has
been holding back advances in understanding the complex phenomena of music-mediated
emotions has been the abundance of different emotion theories and concepts – discrete, dimensional,
music-specific or something else altogether. Before focussing on novel, music-specific
Corresponding author:
Tuomas Eerola, Finnish Centre of Excellence in Interdisciplinary Music Research, University of Jyväskylä,
Seminaarinkatu 15, PO Box 35, Finland.
[email: tuomas.eerola@jyu.fi]
Psychology of Music
39(1) 18–49
© The Author(s) 2011
Reprints and permission: sagepub.
co.uk/journalsPermission.nav
DOI: 10.1177/0305735610362821
http://pom.sagepub.com
Eerola and Vuoskoski 19
models (e.g., Zentner, Grandjean, & Scherer, 2008), there is a need to first critically compare the
discrete and dimensional models of emotions in music because these are the two dominant
models used in music and emotion research (Juslin & Sloboda, 2010; Zentner & Eerola, 2009),
and they are often implied to be highly convergent although this has not actually been explicitly
studied. Secondly, neurological studies (Gosselin et al., 2005; Khalfa et al., 2008b) have indicated
that different processes may be involved in the discrete and dimensional assessments of
emotions.Thirdly, recent hybrid models of emotion (Barrett, 2006; Christie & Friedman, 2004;
Russell, 2003) depend on finding the ways in which core affects (taken to be dimensional) interact
with the conscious interpretation of what people know about emotions (best described in
discrete terms). And finally the understanding of musical and acoustic features that contribute
to emotions would greatly benefit from knowing which model – dimensional or discrete – maps
the feature space in the most ecological fashion.
Another hindrance in music and emotion research has been the choice, quality, and amount
of musical examples used as stimuli. Previous studies have predominantly used well-known
excerpts of Western classical music, which have been chosen arbitrarily by the researchers.
Moreover, the stimuli have mostly been highly typical exemplars of the chosen emotions even if
the underlying emotion model does not imply that emotions are structured around specific categories.
We will address these issues in detail later.
In this article, we focus on perceived emotions (in other words, emotions that are represented
by music and perceived as such by the listener). An overview of the literature implies
that the border between the two alternatives – emotion recognition and emotion experience –
may be somewhat blurred in reality, and it has even been suggested that the two alternatives
could be seen as opposite extremes of a continuum (Gabrielsson, 2001). In addition recent
empirical studies have found more similarities than differences between the two (Evans &
Schubert, 2008; Kallinen & Ravaja, 2006; Vieillard et al., 2008).
To address the theoretical diversity in depth, we will first briefly summarize the prominent
kinds of emotion models and their relevance for music. During the past decade, discrete models,
different dimensional models, and domain-specific emotion models have all received support in
studies of music and emotion (Ilie & Thompson, 2006; Krumhansl, 1997; Schubert, 1999;
Zentner et al., 2008). According to the well-known discrete emotion model – the basic emotion
model – all emotions can be derived from a limited number of universal and innate basic
emotions such as fear, anger, disgust, sadness and happiness (Ekman, 1992, 1999). The basic
emotion model builds on the assumption that an independent neural system subserves every
discrete basic emotion. However, neuro-imaging and physiological studies have failed to establish
reliable, consistent evidence to support this theory (for a review, see Barrett & Wager, 2006;
Cacioppo, Berntson, Larsen, Poehlmann, & Ito, 2000), and the matter remains under debate. In
studies investigating music and emotion, the basic emotion model has often been modified to
better describe the emotions that are commonly represented by music. For example, basic emotions
rarely expressed by music, such as disgust, are often changed to more suitable emotion
concepts like tenderness or peacefulness (Balkwill & Thompson, 1999; Gabrielsson & Juslin,
1996; Vieillard et al., 2008). It still remains to be clarified whether models and theories designed
for utilitarian emotions (Scherer, 2004) – such as the basic emotion model – can also be applied
in an aesthetic context such as music. It has been argued and empirically demonstrated that a
few primary basic emotions seem inadequate to describe the richness of the emotional effects of
music (Zentner et al., 2008). In their study, Zentner and colleagues (2008) proposed a new
model for music-induced emotions by first compiling music-related emotion terms and uncovering
the underlying emotion structure with exploratory factor analysis, and then corroborating
20 Psychology of Music 39(1)
their findings by means of confirmatory factor analysis. The resulting nine-factor Geneva
Emotion Music Scale (GEMS) model consists of wonder, transcendence, tenderness, nostalgia,
peacefulness, power, joyful activation, tension and sadness. For music and emotions studies,
this model provides much needed domain-specificity and emphasizes the positive and reflective
nature of music-induced emotions. Although it was shown that the GEMS model outperformed
discrete and dimensional emotion models in accounting for felt emotions in music, in our opinion,
these results can be disputed as they pitted musically non-relevant formulations of the basic
emotions and dimensions against GEMS and relied on a handful of familiar classical music
examples. The focus of the present study is to compare the traditional models of emotion in
music and also to focus on perceived emotions. This is a different emphasis to that of GEMS,
although the scale has implications which will be discussed further.
In recent years, two-dimensional models of emotion have gained support among music and
emotion researchers (e.g., Gomez & Danuser, 2004; Schubert, 1999; Withvliet & Vrana, 2006).
Instead of an independent neural system for every basic emotion, the two-dimensional circumplex
model (Posner, Russell, & Peterson, 2005; Russell, 1980) proposes that all affective states
arise from two independent neurophysiological systems: one related to valence (a pleasure–
displeasure continuum) and the other to arousal (activation–deactivation). In other words, all
emotions can be understood as varying degrees of both valence and arousal. In contrast,
Thayer (1989) suggested that the two underlying dimensions of affect were two separate
arousal dimensions: energetic arousal and tense arousal. According to Thayer’s multidimensional
model of activation, valence could be explained as varying combinations of energetic
arousal and tense arousal. A visual summary of the two-dimensional models of Russell and
Thayer is given in Figure 1. In the music domain, Vieillard and colleagues explored emotional
excerpts of music by means of similarity ratings, and found that the excerpts could be mapped
onto a two-dimensional plane in which the salient dimensions could be best explained in terms
of energy and tension (Vieillard et al., 2008). However, the two-dimensional models have been
criticized for their lack of differentiation when it comes to emotions that are close neighbours
in the valence-activation space, such as anger and fear (see e.g., Tellegen, Watson, & Clark,
1999). It has also been discovered, that the two-dimensional model is not able to account for all
the variance in music-mediated emotions (Bigand, Vieillard, Madurell, Marozeau, & Dacquet,
2005; Collier, 2007; Ilie & Thompson, 2006).
Wilhelm Wundt suggested a distinction be made between three dimensions of emotions
already as early as 1896.These three dimensions were pleasure–displeasure, arousal–calmness,
and tension–relaxation. Although the two-dimensional models have a dominant position in
affect literature, there is some evidence of the model’s incompatibility with affect data
(Schimmack & Grob, 2000; Schimmack & Reisenzein, 2002). Previous studies have shown that
arousal–calmness and tension–relaxation dimensions cannot be reduced to one arousal dimension
(Schimmack & Grob, 2000; Schimmack & Reisenzein, 2002). The underlying reason is
that the two activation dimensions are related to different causes: unlike tense arousal, energetic
arousal is affected by a circadian rhythm (Thayer, 1989; Watson, Wiese, Vaidya, &
Tellegen, 1999), and the two arousal dimensions have been shown to change in opposite directions
when specifically manipulated (Gold, MacLeod, Deary, & Frier, 1995). In sum, both main
theoretical models – discrete and dimensional – will be investigated simultaneously to clarify
their dependencies and applicability to music and emotions. The three-dimensional model,
shown visually in Figure 1, will be used to collect data regarding the dimensional approach as
it still allows us to examine post facto whether a lower dimensional solution (valence and
arousal or arousal and tension) could be used instead.
Eerola and Vuoskoski 21
Two recent and puzzling findings render the comparison of these conceptual frameworks of
emotion even more pressing an issue. Studies with brain damaged patients have documented a
dissociation between discrete and dimensional evaluations of emotion in music (Dellacherie,
Ehrlé, & Samson, 2008; Gosselin et al., 2005; Khalfa et al., 2008b), implying that separate neural
processes are responsible for each of these types of evaluation. Moreover, it has been shown
that listeners may experience both sad and happy feelings at the same time when exposed to a
stimulus with mixed emotional cues (Hunter, Schellenberg, & Schimmack, 2008). Both of these
findings pose challenges for emotion research in music and require a better understanding of
the essential similarities and differences between these two conceptual frameworks.
The majority of studies on music and emotions have used excerpts of relatively well-known
Western classical music pieces (e.g., Kreutz, Ott,Teichmann, Osawa, & Vaitl, 2008; Krumhansl,
1997; Nawrot, 2003; Schmidt & Trainor, 2001), which have been arbitrarily chosen by the
researchers. Other types of musical stimuli used – occasionally together with classical music –
include popular music and jazz (Altenmüller, Schuermann, Lim, & Parlitz, 2002; Gomez &
Danuser, 2004), film music (Etzel, Johnsen, Dickerson, Tranel, & Adolphs, 2006), music from
other cultures (Balkwill & Thompson, 1999; Balkwill, Thompson, & Matsunaga, 2004), and
synthetic music that has been composed especially for the research task at hand (e.g., Khalfa,
Peretz, Blondin, & Manon, 2002; Vieillard et al., 2008). Well-known music examples are potentially
problematic as stimuli because participants may already be familiar with the excerpts.
Emotions elicited by this type of stimuli can be closely entwined with extra-musical associations.
Synthetic stimuli are free of such problems and have provided opportunities to study and
manipulate musical features. However, such stimuli often sound artificial and lack some of real
Figure 1. Schematic diagram of the dimensional models of emotions with common basic emotion
categories overlaid. Note that the axes of the three-dimensional model are not necessarily orthogonal
in actual affect data as depicted here (see Schimmack & Grob, 2000).
22 Psychology of Music 39(1)
music’s intricate features, such as expressive performance and timbre, which might be essential
in evoking an emotional response (Juslin, Friberg, & Bresin, 2002; Leman, Vermeulen, De
Voogdt, Moelants, & Lesaffre, 2005). Moreover, the stimuli used to investigate dimensional
emotion models have typically represented the extremes of the dimensions (e.g., low arousal,
high valence) as opposed to points along the continuum of each dimension. This may have led
to the neglect of potentially relevant affect data.
It is worth of noting that, with the notable exception of Bigand et al. (2005), musical examples
have been chosen solely in terms of wholly discrete emotions (e.g., happy and sad) in the
existing studies that use both discrete and dimensional models of emotion in music (Gosselin
et al., 2006; Gosselin, Peretz, Johnsen, & Adolphs, 2007; Khalfa et al., 2002, 2008a, 2008b;
Kreutz et al., 2008; Krumhansl, 1997; Nyklìcek,Thayer, & Van Doornen, 1997;Terwogt & Van
Grinsven, 1991; Vieillard et al., 2008). Hence, the ratings of dimensional concepts such as
valence and arousal describe the known points of discrete emotions in the dimensional affective
space (e.g., Nyklìcek et al., 1997). However, it is currently unclear what happens in the
reverse situation where the musical examples are selected systematically from various points
within the affective space. In such cases the discrete emotions may not be easily assigned to
examples that are distant from the prescribed point in the affective space for that particular
discrete emotion. Lastly, the number of music examples used in previous studies has been relatively
low compared with the stimulus sets used by emotion researchers in other fields, such as
the visual domain. The International Affective Picture System (IAPS) uses 12 series of 60 pictures
each (Lang, Bradley, & Cuthbert, 2005), out of which typical emotion studies tend to use
50–60 images (this was the median in a sample of 20 studies using IAPS). However, in the
music domain, much smaller sets are more commonly used.
Aims of the study
The primary aim of the present study is to contribute to the theoretical debate currently occupying
music and emotion research by systematically comparing evaluations of perceived emotions
using two different theoretical frameworks: the discrete emotion model, and dimensional
model of affect.The importance of the comparison lies not only in the prevalence of these models
in music and emotion studies, but also in the suggested neurological differences involved in
emotion categorization and the evaluation of emotion dimensions (Khalfa et al., 2008b), as
well as in the categorically constrained affect space the excerpts have represented to date.
Moreover, the various alternative formulations of the dimensional model have not been investigated
in music and emotion studies before.
A secondary aim is to introduce a new, improved set of stimuli – consisting of unfamiliar,
thoroughly tested and validated non-synthetic music excerpts – for the study of musicmediated
emotions. Moreover, this set of stimuli should not only include the best examples
of target emotions but also moderate examples that permit the study of more subtle variations
in emotion.
Expert selection of the stimulus materials
In order to obtain a large sample of unknown yet emotionally stimulating musical examples, a
large expert panel was organized for choosing the material. The primary goal of this panel was
to choose emotionally representative musical material from a large selection of film soundtracks
according to predefined criteria.
Eerola and Vuoskoski 23
It was decided to use film music because it is composed for the purpose of mediating powerful
emotional cues, and could serve as a relatively ‘neutral’ stimulus material in terms of musical
preferences and familiarity. Unfamiliar excerpts were chosen to avoid episodic memories from particular
films influencing perceived emotions in the music. And yet, since it was film music, and
listeners are generally accustomed to this genre from media exposure, the excerpts were nevertheless
expected to conjure up schematic memories. But this also meant we could not prevent
excerpts from triggering memories by simple virtue of them resembling others from a listener’s
previous experience. With the exception of Vertigo (from 1958), the selection of 60 soundtracks
was limited to those published within the last three decades (1976–2006). This was to keep the
sound quality of the corpus relatively homogeneous. It should also be noted that the soundtracks
camefromawiderangeof filmsthatincludedromantic,sci-fi,horror,action,comedy,anddrama.
The panel consisted of 12 expert musicologists (staff members and third to fifth year university
students) who had all studied a musical instrument for 10 years or more. Each panel member
was given five different soundtracks and asked to find five examples of the six target emotions.
Half the experts focused on discrete emotions (six targets), and half on the extremes of the threedimensional
model (six targets). For the discrete emotions we chose happiness, sadness, fear,
anger, surprise and tenderness, as these have been favoured in previous studies of music and
emotion (Juslin, 2000; Kallinen, 2005; Krumhansl, 1997). For the dimensional model, we
chose the six extremes of the three-dimensional model of emotion by Schimmack and Grob
(2000). This meant the panel should find examples of both positive and negative valence, high
and low tension arousal, as well as high and low energy arousal. Each extreme was characterized
using three adjectives taken from Schimmack and Grob (2000). For valence, these were
pleasant–unpleasant, good–bad, and positive–negative. For the energy dimension the adjectives
were awake–sleepy, wakeful–tired, and alert–drowsy. The adjectives used to represent the
extremes of the tension dimension were tense–relaxed, clutched up–calm, and jittery–at rest.
To ensure a degree of uniformity in the choice of sound examples, the following criteria were
established: each excerpt should be between 10 and 30 seconds (depending on the natural
phrasing of the excerpt): it should not contain lyrics, dialogue, or sound effects (car sounds,
etc.), and, though familiar with the schematic memory, it should not be familiar in the episodic
sense (see earlier). It was also stressed that the goal was to choose examples that could convey
the target emotion to the general listener in an optimal way. The experts also made a note of
the musical features and devices which informed their choice. This resulted in 360 audio clips
(12 × 5 × 6), equally representative of the discrete emotion and three-dimensional models.
Details related to the stimuli (names, ratings, and audio examples) may be found online.1
Pilot experiment
The aim of the pilot experiment was to rate all the examples previously selected by experts in
terms of both the models (discrete and dimensional). This was done to understand how the
emotions were conveyed by the examples. The aim was also to reduce the number of excerpts
and homogenize the selection for further investigation.
Method
Participants, stimuli, apparatus and procedure. The participants of the pilot experiment were the
same group of experts who originally chose the examples (mean age 24.1 years, SD = 3.9 years,
7 females). The stimuli consisted of all 360 excerpts that had been chosen. The panel then
24 Psychology of Music 39(1)
received instructions to rate the perceived emotion in each audio clip. Half of them were
instructed to rate the discrete emotions on a scale of 1–7 and the rest gave ratings for the three
dimensions using bipolar scales. Familiarity with the excerpts was also rated (0 = unfamiliar,
1 = somewhat familiar, 2 = very familiar). Note that each of the panel’s own selections constituted
a mere 8.3% of the whole material (30 items out of 360).The rating task was also divided
into four sections, and between each of them, the emotion models were switched between the
two halves of the experts. The sections each consisted of 90 excerpts and lasted about 50 minutes.
All the excerpts were played in a random order, but this order was the same for the whole
group since it was done as a classroom exercise using high-quality audio in laboratory conditions.
In total, the task lasted approximately four hours and was carried out on two separate
days. The student participants received course credits for their efforts.
Results
To rule out the order effects, a linear trend analysis was conducted against the rating order for
all scales. All the trends yielded non-significant F-ratios (F values between 0.2 and 2.82, p = ns,
df = 358). Next the consensus between the raters was investigated. Cronbach’s alpha was
employed to measure the consistency between raters, rather than between items, as there was
a large proportion of excerpts given similar (low) ratings, which meant negligible item variance.
In other words the lowest ratings were given by all raters for those discrete emotions that
manifestly did not seem applicable to a particular excerpt, e.g., ratings of sadness for happy
excerpts. Most emotion concepts scored relatively high consistency using this procedure, as
indicated in Table 1. The notable exception in inter-rater consistencies was surprise (a = .66),
which was actually unsurprising, as a number of previous studies have also observed this concept
to be problematic for music-mediated emotions (Gabrielsson & Juslin, 1996; Laukka &
Juslin, 2007). Surprise was therefore eliminated from further analyses, as the alpha for it was
considerably lower than for the other emotion terms.
Table 1. Consistencies, means, standard deviations for all emotion ratings and repeated measures
ANOVA results for excerpts grouped by target emotion (η2 for effect sizes)
Type of excerpt/ Cons. (a) Target Non-target ANOVA (η2)
concept (concept) (concept) (concept) (excerpts grouped
by target)
Happy .93 5.49 (1.53) 2.02 (0.69) 0.63***
Sad .89 5.46 (1.60) 2.22 (0.98) 0.71***
Tender .92 5.69(1.89) 2.38 (0.88) 0.72***
Fearful .92 5.29 (1.80) 2.63 (0.91) 0.63***
Angry .92 5.38 (1.61) 2.03 (0.74) 0.69***
Surprising .66 3.36 (2.05) 1.72 (1.02) 0.23***
Pos. valence .92 5.68 (0.94) 3.96 (1.00) 0.64***
Neg. valence – 1.81 (0.68) 4.31 (1.02) 0.88***
Pos. energy .90 5.67 (0.88) 3.92 (0.96) 0.30***
Neg. energy – 2.30 (0.77) 4.23 (0.97) 0.58***
Pos. tension .93 5.98 (0.75) 4.31 (0.86) 0.69***
Neg. tension – 2.48 (0.90) 4.36 (0.89) 0.68***
Note: *p < .05; **p < .01; ***p < .001; df = 5,179 for basic emotion concepts, df = 2,89 for dimensional concepts.
Eerola and Vuoskoski 25
The pilot showed that the raters were not particularly familiar with the excerpts as 89.9%
indicated ‘unfamiliar’, 6.4% ‘somewhat familiar’ and only 3.8% ‘very familiar’. To assess the
role familiarity for the emotion categories and dimensions, a variance analysis was conducted
using a non-parametric method of calculation (Kruskal-Wallis) as the familiarity ratings contained
mainly zeros and were thus not normally distributed (Lilliefors test, p < .001). This
yielded no significant differences across the discrete emotions, χ2 (5,179) = 10.18, p = ns, and
yet significant differences across the dimensional extremes, χ2 (5,179) = 30.02, p < 0.001. A
closer look at the familiarity ratings showed a small number of highly familiar excerpts (above
three SDs above the mean familiarity: three in total) in two categories: high valence and low
tension. We kept these familiar excerpts in the data set since they represented a minor portion
of the total number (0.83%).
Next we explored whether the emotion targets were clearly evident in the ratings. This was
done by separately comparing the mean ratings for each type of excerpt using an analysis of
variance. This repeated measures ANOVA yielded significant main effects for all of the discrete
emotion targets, and with large effect sizes (mostly above 0.60, shown in Table 1). The exception
was a surprise, as mentioned previously, which, although statistically significant, exhibited
a fairly low effect size. The other weak effect size (0.30) was seen in the ratings of high energy,
the reason for which is probably due to its collinearity with the other dimensions. Excerpts representing
high energy were also rated high in tension, as well as high in negative valence, and
thus the ANOVA results display weaker discrimination between these categories. A post-hoc
analysis between the discrete emotion categories (using Holm-Sidak adjusted t-tests at p < .05
level) revealed this statistically, namely how anger ratings could not be discriminated from fear
ratings when anger was the target category. Similarly, ratings of surprise were statistically
indistinguishable from ratings of fear and anger. Post-hoc analyses were also used to characterize
the differences between dimensional extremes, resulting in statistically significant differences
at a level of p < .05 in all comparisons.
Discussion
The primary aim of the pilot experiment was to evaluate the chosen set of musical stimuli in
terms of two conceptual frameworks in order to establish a systematic basis for selecting the
stimuli for the actual experiment. Such selection rationale has been noted to be largely absent
in the previous studies of music and emotions (Juslin & Västfjäll, 2008). The chosen target
emotions were clearly evident in the ratings, with the exception of surprise as previously noted,
and so this was removed. As the set of stimuli contains a large variety of music excerpts rated
in terms of both emotion models, it is now possible to compare these models in a theory-based
and systematic way.
Experiment
The first aim of this experiment was to systematically compare the effectiveness of discrete and
dimensional models in the study of perceived emotions in music. A related aim was to explore
whether some of the ways of representing emotions within the discrete and dimensional models
could be merged or eliminated (e.g., collapsing the three-dimensional model into two dimensions,
or removing some of the overlapping concepts from the discrete emotion model).
The second aim was to form a refined set of musical examples, which would not only include
the clearest exemplars of the discrete emotion categories, but also ones that are less easily
26 Psychology of Music 39(1)
attributable to a category. Such moderate examples would provide more realistic and interesting
material for empirical work on recognition and induction of emotions in music than the
standard paradigm, where only a few highly characteristic examples of emotion categories are
used. An experiment was designed to address these questions using a subset of the stimuli from
the pilot experiment.
Method
Stimuli. To validate the stimulus material and to compare the conceptual frameworks, stimuli
were needed to represent emotion concepts from both discrete and three-dimensional models in
order to do justice to both (see Mikels et al., 2005 for a selection method aimed at discrete emotion
categories).Therefore, a sampling of the 360 excerpts from the pilot experiment using both
models was carried out. In the first stage, all excerpts that were rated as moderately or very
familiar in the pilot experiment were eliminated. To obtain both highly and moderately typical
examples of discrete emotions, it was necessary to calculate the typicality (T) of the target emotion
for each excerpt (i). This was done by subtracting the mean of the excerpt’s non-target
emotion rating (NE
–
i) and the standard deviation of its target emotion rating (SEi) from the
mean of the target emotion rating (E
–
i)
Ti = E
–
i - SEi - NE
–
i
Highest typicality values for excerpts occurred when they were highly and consistently rated
on the target emotion, and not attributed to other emotion categories. For example, two excerpts
scoring 6 on sadness had a different typicality value if they differed in consistency and their scores
for other emotion categories. If a sadness rating had a mean of 6 with a SD of 2, and a mean of 1
in other emotion categories, it would result in a typicality of 3 (6–2-1). However a mean sadness
of 6 with more deviation, and therefore less consistency (SD 3), together with higher attributions
to other categories (1 in fear, anger, surprise, and happiness but 5 in tenderness, which is a NE
–
i of
1.8) would have a typicality of only 1.2 (6–3-1.8). The excerpts were thus ranked according to
their typicality values for each emotion. From these ranked lists, the top five examples were chosen
as best examples of each discrete emotion (happiness, sadness, tenderness, anger and fear),
called high examples hereafter. Five moderate examples were taken from the ranked positions of
51 to 55 of each similarly ranked list. This yielded a total of 50 examples for all the discrete emotions
([5 high + 5 moderate] × 5 categories). Surprise was not incorporated into this experiment
due to the low consistency and recognition it received in the pilot experiment.
For the dimensional model, another scheme for obtaining representative examples from the
pilot experiment was adopted. Each dimension was sampled at points along its axis whilst the
other two dimensions were kept constant. The purpose of this was to maximize the variance
according to the dimension in question, although this was not always entirely possible due to
the collinearity of the dimensions. The axis of each dimension was then split into four percentiles
as follows: extreme low (< 10%), moderate low (20%–40%), moderate high (60%–80%)
and extreme high (> 90%). From each of these percentiles five excerpts were taken, which
meant a total of 20 for each dimension, exhibiting a similar range of typicality. During this
process, the other two dimensions were kept in control by minimizing the error distance so that
the excerpts were chosen as close to the target dimensional axis as possible. In this way 60
audio clips (4 × 5 × 3) were picked to cover the essential variance of the three-dimensional
affect space. Again, the purpose of choosing examples that were only moderate examples of the
Eerola and Vuoskoski 27
target emotion concept was to increase the coverage of the stimulus space, and to compare the
effectiveness of each model at rating ambiguous excerpts. In the case of the dimensional model,
we also wanted to provide stimulus material which could be used to test whether the conceptual
framework does actually operate in a dimensional fashion which is otherwise hard to do if
only best examples of the bipolar extremes are used. A list of the final stimuli may be found in
the Appendix, and the details are also documented online2 (50 discrete + 60 dimensional = 110
examples in total).
Participants. The participants for the main experiment were 116 university students aged
18–42 years (mean 24.7, SD 3.75, 68% females and 32% males). Forty-eight percent of the
participants did not play any instrument and were not musically trained, 41% had experience
of playing an instrument or some level of musical training, and 11% were in between with
music as a hobby for less than three years.The participants received cinema tickets in return for
their participation, and a number of individual variables were collected from each of them.
First, a survey of their musical taste was made using a localized version of the STOMP questionnaire
(Rentfrow & Gosling, 2003). Second, their current mood at the time of taking the test was
evaluated using the POMS-A questionnaire (Terry, Lane, Lane, & Keohane, 1999). Third, an
assessment was made of the participants’ personality traits using a 44-item personality measure
known as ‘The Big Five Inventory’ (John & Srivastava, 1999). And finally a short questionnaire
was given out to gather information about the participants’ film genre preferences,
musical training, and any hearing problems.
Procedure. The experiment was divided into two blocks: block A with the 50 discrete emotion
excerpts and block B with the 60 dimensional model excerpts. The order of the blocks was
counterbalanced across the participants (67 participants did block A first followed by block B),
and the order of the examples within blocks was individually randomized. The participants
were instructed to rate the emotions represented by the music excerpts (perceived emotions),
and the difference between perceived and induced emotions was explained to them. While one
group of participants were rating excerpts in block A on a scale of 1–9 for each discrete emotion,
the other group was rating excerpts on bipolar scales of 1–9 for each of the three axes of
the dimensional model. Then the blocks were switched for each group. This meant that the
experiment used both models to rate the emotions in all 110 excerpts without taking an unreasonably
long time. The total duration of the experiment was between 50 and 60 minutes,
depending on the participant’s rating speed.
Participants were also asked to mark how much they liked each example (with a preference
rating) and how beautiful they considered each example to be (with a beauty rating). In both cases
this was on a scale from 1 to 9.These additional measures were added to clarify the role of valence
because valence and preference have been shown to be separate constructs (Schubert, 2007). For
example, one can be fond of harsh and rough sounding music despite the fact that most people
associate those qualities with negative valence. Ratings of preference and beauty would also
briefly allow the exploration of the relation between sadness and valence in this context.
Before the actual experiment, a short practice session was carried out by each participant to
become familiar with the interface, likert scales, and type of music used. The participants also
had the possibility to ask questions about the task before the start of the experiment.
Apparatus. The listening experiments were conducted in a soundproof room. To gather
the emotion ratings, a special patch was designed in Pure Data graphical programming
28 Psychology of Music 39(1)
environment (Puckette, 1996), running on Mac OS X. The patch enabled the participants to
move from one excerpt to the next at their own pace and to repeat an excerpt if needed.
Participants listened to the excerpts through studio quality headphones (AKG K141 Studio),
and were able to adjust the sound volume according to their own preferences.
Results
To rule out extreme mood states that might affect the participants’ emotion ratings, POMS-A
ratings were aggregated and the distance from the mean rating (1.87, SD = 0.47 on a scale of
1–5) was calculated for each participant. One participant whose score was more than three SDs
above the mean was removed from the analysis as that person’s current mood appeared considerably
pessimistic, tired and negative. The intersubject correlations were used to identify possible
outliers and anyone who scored more than three SDs from the mean intersubject correlation
was removed from the dataset.This resulted in the removal of five participants, leaving the total
of 110 that was eventually used. Cronbach’s alpha was used to assess the inter-rater reliability
across both experimental blocks. All emotion ratings received alphas above .99 and only the
preference ratings were slightly lower (.94), probably since personal opinions are prone to vary
across individuals. Subsequently the data was pooled together for further analyses.
We reviewed the contribution of individual variables (personality, musical preferences, film
genre preferences, and musical training) to the emotion ratings by correlating the individual
ratings with the background variables. Only a few statistically significant correlations emerged:
the personality trait ‘openness to experience’ (John & Srivastava, 1999) appeared related to
increased ratings of anger (r = 0.25, p < .05) and valence (r = 0.33, p < .05), and ‘extroversion’
seemed related to decreased ratings of tension (r = -0.38, p < .01). These traits may therefore
indicate different rating strategies. ‘Negative mood’ (POMS-A; Terry et al., 1999) was, perhaps
unsurprisingly, related to increased ratings of sadness (r = 0.29, p < .05). These initial observations
are potentially interesting and are known to influence emotional evaluations (Kreutz
et al., 2008; Rusting, 1998), but the figures indicate that their role is only moderate at best,
and thus these factors will be left unexplored at this stage.
Discrimination of emotion categories and levels. An examination of the discrete emotion rating
was carried out next using a mixed design repeated measures ANOVA for groups of excerpts
representing each target emotion. The five emotion concepts (anger, fear, sadness, happiness,
and tenderness) as the within subjects variable. The two levels (high and moderate) at which
excerpts conveyed an emotion (i.e., high and moderate examples) provided the within group
variable. Taken together this gave significant main effects for all target emotion concepts, and a
significant main effect for levels in one target emotion. There was also significant interaction
effects between these two factors in four of the target emotions (see Table 2 for effect sizes and
p values). In other words, the effect sizes were robust for concept (between 0.79 and 0.83; see
Table 2), but negligible for emotion levels (0.000–0.006). And most of the interactions between
concept and level were within each target emotion were significant. These analyses were later
followed by post hoc tests, in which p values were adjusted using the Holm-Sidak procedure to
avoid the effects of multiple comparison tests (Ludbrook, 1998). These analyses revealed that
in the high examples the target emotion was never confused with other emotion concepts, but
the moderate examples exhibited confusions between one or two other concepts of emotion (see
Table 2). Figure 2 clearly shows this pattern. For example, moderate examples of anger are
indistinguishable from fear and sadness, and moderate examples of tenderness could easily be
Eerola and Vuoskoski 29
Figure 2. Mean ratings and 95% confidence intervals for five discrete emotions and 50 excerpts
representing these target emotions (black markers for high examples and white for moderate examples).
30 Psychology of Music 39(1)
confused with happiness and sadness (the precise confusion pattern is given in the last column
of Table 2). The effect sizes and patterns of confusion are similar to the ones observed in the
pilot experiment, and confirm prior research (e.g., Gabrielsson & Juslin, 1996).
Similarly for dimensional examples, mixed design repeated measures ANOVAs were carried
out to investigate how ratings for each group of excerpts representing different target emotions
varied between the dimensions and levels within that dimension. This was done in the
same way as with the discrete emotion examples. Table 2 and Figure 3 display the results of
this analysis, which yielded significant main effects for the concept and significant effect for
level in three of the dimension extremes.There was also a significant interaction effect between
concept and level for four of the dimension extremes.The effect sizes were comparable to those
obtained with discrete emotions (between 0.44 and 0.81 in the concept effect sizes and
between 0.00 and 0.11 for the level effect sizes). As with the discrete emotion examples, HolmSidak
adjusted post hoc analyses were also performed (results shown in the last column of
Table 2). However, the results have to be interpreted in a different manner since the ratings
given with the three-dimensional model only indicate the excerpt’s location in the dimensional
space.
Reliability of discrete and dimensional model ratings. Although the overall reliability of the dimensional
and basic emotion model scales was earlier found to be high, a closer scrutiny may reveal
interesting differences, especially related to numerous ambiguous (moderate) examples. To
compare the applicability of the discrete and dimensional models in the rating of emotionally
ambiguous as well as highly characteristic excerpts, the raw ratings for the high examples of
target discrete emotions were compared with the raw ratings for moderate examples of target
emotions. This was done using inter-rater agreements across the excerpts (Cronbach’s alpha).
The agreements within the high and moderate examples of target emotions were different,
a = .74 for high and a = .49 for moderate examples. To evaluate the statistical significance of
this difference, we used bootstrapping, in which the confidence intervals (CI) of the statistics in
question were obtained for 1,000 bootstrapped calculations (Davison & Hinkley, 1997). On the
basis of these confidence intervals, the means were found to be different at p < .001 level (mean
Table 2. Mixed-design repeated measures ANOVA results for groups of excerpts representing each target
emotion (η2 for effect sizes). Post-hoc tests display target emotions that do not reliably differ in means
(using Holm-Sidak adjusted values for p < 0.05)
Type of excerpt Category Level Interaction Post-hoc (moderate)
Happy 0.70*** 0.003** 0.21*** S, T
Sad 0.83*** 0.003 0.05* T
Tender 0.78*** 0.000 0.05 H, S
Fearful 0.79*** 0.001 0.08** A
Angry 0.79*** 0.006 0.06** F, S
Pos. valence 0.52*** 0.01 0.08* En, Te
Neg. valence 0.70*** 0.000 0.08* En, Te
Pos. energy 0.44*** 0.11* 0.06 Va, Te
Neg. energy 0.54*** 0.01 0.11* Te
Pos. tension 0.81*** 0.02** 0.07* En
Neg. tension 0.58*** 0.09* 0.03 En
Notes: p < .05; **p < .01; ***p < .001; df = 4,49 for basic emotion concepts, df = 2,29 for dimensional concepts.
S = Sadness, T = Tenderness, H = Happiness, A = Anger, F = Fear, Va = Valence, En = Energy, Te = Tension.
Eerola and Vuoskoski 31
CI 99.9% for high examples was .74 .69 .78, and for moderate examples it was .49 .29 .61). Similar
analysis was performed for the raw ratings of the three dimensions, and the resulting reliability
estimates were also significantly different at p < .001 level (mean CI 99.9% for high examples was
.95 .94 .96 and for moderate examples it was .77 .72 .82). Although the reliabilities for both models
show significantly lower overall reliabilities for moderate excerpts, presumable as they may be a
mixture of several emotions, those for moderate excerpts from the dimensional model are at the
same level as the reliabilities for the high excerpts of the basic emotion model, implying a higher
overall reliability in the dimensional model ratings. However, perhaps the selection of the basic
Figure 3. Mean ratings and 95% confidence intervals for 60 examples representing 6 target emotions and
2 levels of the three dimensions.
32 Psychology of Music 39(1)
emotion excerpts is such that they have lower overall reliability in general compared with the
dimensional model excerpts and the difference is not in the measurement model. To evaluate
this argument, we replicated the same analysis but reversed the concepts (high examples of
target emotions selected using the basic emotion model were rated with the dimensional model).
The reliabilities and their confidence intervals showed again a similar pattern. For basic emotion
model ratings of the dimensional targets, a was .87 .85 .89 for characteristic examples, and .70 .66
.74 for ambiguous examples. For dimensional model ratings of the basic emotion model targets,
the alphas were higher (.94 .93 .95 for high examples, and .85 .81 .87 for moderate examples). In
both cases, significant differences in the reliabilities between the high and moderate excerpts
were found but the overall reliability of the dimensional ratings was again higher. So the dimensional
model provides somewhat higher inter-rater consistency no matter which way the musical
excerpts have been chosen. Whether consistency is a crucial detail in assessing the adequacy
of these models is another question. High consistency could for example also indicate that the
measurement scale is trivial and thus offers little insight into the actual emotion process.
The applicability of the discrete emotions to the appropriate emotion prototype areas (Russell
& Feldman Barrett, 1999) is visualized in Figure 4, where both the strength and the variation
of all discrete target emotions are shown as densities in the valence-energy space. Marked in
the plot are the target centroids of the chosen high and moderate examples of the five discrete
emotions, which indeed lie in the approximate (attractor) areas of high ratings for these emotions.
Moreover, the variation demonstrates that the high examples are mostly located within
areas of high agreement (sadness and tenderness are the notable exceptions), and the most
moderate examples are located in less well-defined areas. In contrast, the deviations in the ratings
of the dimensional model are lower and spatially more uniformly distributed, and do not
have such ill-defined areas around the attractors of the emotion categories. This is also demonstrated
by the means of the ANOVA above. The practical implication of this is that, whereas
both models may be used to adequately describe emotional excerpts representing clear examples
of discrete emotions, the strength of the dimensional model lies in its ability to describe
such emotional examples that lie outside these discrete attractor areas. The utilization of this
asymmetry between the models would allow us to explore how a hybrid model of emotions
might manifest itself in music (Russell, 2003), and is of consequence to clinically oriented
studies that are interested in the patients’ processing of emotionally ambiguous examples
(e.g., Bouhuys, Bloem & Groothuis, 1995; Cavanagh, & Geisler, 2006).
In the end participants were able to recognize the target emotions represented by the high
examples consistently, and in moderate examples the target emotion was confused with at least
one other emotion. This has also been observed in previous research, and could be interpreted
more generally as fuzziness in the definition of the emotion categories (for examples, see Dailey,
Cottrell, Padgett, & Adolphs, 2002; Russell & Fehr, 1994). In the case of the emotion dimensions,
the results have to be viewed from a different angle, because here the ‘emotion targets’
were actually the six bipolar extremes of three dimensions. Instead of concentrating on the
confusion between different dimensions, we should focus our attention towards any possible
confusion between the bipolar extremes of a given dimension. For example, valence and tension
ratings for excerpts representing moderate positive energy did not differ (seeTable 2).To set
these observations into a wider context, it is necessary to examine the whole pattern of correlations
between the emotion concepts.
Patterns of correlations between the emotion concepts. Our intention was to explore how the two
dominant conceptual frameworks for emotions in music (dimensional and discrete) can be used
Eerola and Vuoskoski 33
to describe perceived emotions in music. We also set out to clarify what type of dimensional
model would be the most appropriate for such studies. To make this comparison we used a variety
of correlational techniques.To begin with, mean ratings for both sets of stimuli were visualized
across the three dimensions (Figure 5) and the ratings of the discrete emotion categories
were indicated with appropriately sized markers (the greater the marker size, the greater the
Figure 4. The intensity of each discrete target emotion is shown via accumulated density distribution of
the emotion ratings for each category (upper panel). To define each of the five discrete target emotions,
the examples that received ratings above the upper 50% percentile of the appropriate target ratings have
been selected. The labels refer to the centroids in the valence–arousal space defined by the five discrete
target emotion examples (A = anger, F = fear, H = Happiness, T = tenderness, and S = sadness, capitals
refer to high examples and small letters moderate examples). The lower panel displays a similar projection
of discrete emotion areas within the valence – arousal space, but the gradient indicates the amount of
variation (normalized standard distribution of the ratings) for each emotion category. The centroids of
the high examples are located clearly on well-defined areas (high intensity and low deviation) whereas the
moderate examples are on less clearly defined areas (lower intensity and higher variation).
34 Psychology of Music 39(1)
mean rating for the discrete emotion indicated by the marker type). The scatterplots in Figure 5
show a highly collinear structure between valence and tension. However, the remaining dimensions
display less evident correlational structures. The discrete emotion categories – as represented
by marker types and sizes – suggest a clear separation of happy excerpts from the rest of
the discrete emotions. Also tenderness and sadness stand out as distinct areas within the
dimensional space. None of these three categories overlap with anger and fear, even if these two
do so between themselves in the dimensional structure (see Kreutz et al., 2008; Schubert,
1999; Vieillard et al., 2008). As we remember from the analysis of discrete emotion ratings,
this overlap was also evident in the ratings of anger and fear, and thus is not a feature particular
to the dimensional emotion model alone.
Correlations between the different emotion concepts, together with preference and beauty,
are shown in Table 3. Fear and anger can be observed to correlate highly with each other
(r = .69, p < .001), which suggests that these two emotion concepts might not be easily distinguishable
in the context of music (Juslin, 2000; Kallinen & Ravaja, 2006). Interestingly,
tenderness received higher correlations with sadness (r = .36, p < .001) than with happiness
(r = .15), as traditionally tenderness has been associated with positive emotions in general
(Juslin, 2001, p. 315). Another noteworthy observation is that valence and sadness did not
Figure 5. Mean ratings of three dimensions and discrete emotions for all excerpts (N = 110). The marker
types represent the target emotion categories and the sizes indicate the mean target emotion rating for
each excerpt.
Eerola and Vuoskoski 35
correlate with each other.This is in line with results obtained by Bigand et al. (2005) and Kreutz
et al. (2008), who both discovered that sad music was not systematically associated with negative
valence. Although sadness is generally considered to be an unpleasant emotion, the classification
is not as straightforward in the context of music. For instance Schellenberg and
colleagues (2008) found that, in some instances, sad music was liked as much as happy music.
It seems that in music-mediated emotions, happiness and sadness do not represent the opposite
extremes of valence: although happiness had a strong positive correlation with valence
(r = .80, p < .001), sadness and valence did not correlate (r = –.03). Sad music is often considered
beautiful, and therefore it may be difficult to perceive sadness in music as unpleasant. Schubert
(1996) has offered a theoretical solution to this dilemma using a neural inspired associative
network model, in which negative emotions in an aesthetic context may activate enjoyment. In
fact, it has even been reported that sad music activates neural networks involved in biological
reward (Blood & Zatorre, 2001). For example, sadness correlated with preference (r = .38,
p < .001) and beauty (r = .59, p < .001) significantly more highly than happiness (r = .22,
p < .05; r = .16). However, tenderness and valence correlated with preference (r = .58, p < .001;
r = .56, p < .001) and beauty (r = .77, p < .001; r = .61, p < .001) even more highly than sadness.
Lastly, the ratings given by the participants in the main experiment were to a great extent similar
to the ratings given by the small group of experts in the pilot (the final row in Table 3).
When comparing our results to earlier studies of the three-dimensional model (conducted in
non-music contexts) we find there are certain differences, particularly when the correlations
between separate dimensions of the models are taken into account. For example, in a study
based on current mood ratings, Schimmack and Grob (2000) found a strong positive correlation
between energy and valence (r = .49), a strong negative correlation between valence and
tension (r = –.70), and a moderate negative correlation between tension and energy (r = –.33).
In our study, energy and valence did not correlate with each other (r = –.08), valence and tension
had a very strong negative correlation (r = –.83), and tension and energy had a strong
positive correlation (r = .57). This might be due to the different qualities of music-mediated
emotions compared to mood or everyday emotions, but there are no other points of comparison
from the field of music-mediated emotions. Another possible cause for the difference is that,
despite our efforts of sampling the three-dimensional space in a systematic manner, the chosen
sound examples may have represented the geometric space in a different way to how it was
Table 3. Correlations between the concepts (N = 110).
Happiness Sadness Tenderness Fear Anger Valence Energy Tension Pref.
Sadness -.48***
Tenderness .15 .36***
Fear -.61*** -.28** -.67***
Anger -.41*** -.31** -.58*** .69***
Valence .80*** -.03 .63*** -.91*** -.71***
Energy .44*** -.79*** -.64*** .28*** .47*** -.08
Tension -.42*** -.38*** -.87*** .87*** .75*** -.83*** .57***
Pref. .22* .38*** .58*** -.63*** -.37*** .56*** -.31*** -.63***
Beauty .16 .59*** .77*** -.73*** -.56*** .61*** -.58*** -.81*** .87***
Experts† .94*** .88*** .90*** .93*** .92*** .86*** .90*** .94***
Notes:*** p < .001; ** p < .01; * p < .05; †Correlation with the ratings of the experts from the pilot experiment for the
same 110 excerpts.
36 Psychology of Music 39(1)
represented in the previous studies. For instance, we have relatively few examples of excerpts
representing low tension and negative valence, as the correlations suggest that these variables
mostly co-varied. At this point it is difficult to conclude whether this phenomenon is specific to
music or the particular set of stimuli we used, as we do not have alternative samples of stimuli
at our disposal. Nevertheless these correlations, as well as the figures, hint at a possible reduction
of the three-dimensional model. We will address this issue after first making a direct comparison
of the discrete and dimensional approaches.
Correspondence between discrete and dimensional models of emotions. To assess the compatibility
of the two main conceptual frameworks for emotions in music, we adopted two correlational
techniques, canonical correlation and regression. In the canonical correlation, the interdependency
of the two frameworks could be measured in a single analysis. This analysis provided
three canonical variates. The first canonical correlation was .99, the second .94 and the third
was .57, and the model with the three canonical correlations included was highly significant,
χ2 (15) = 634.05, p < .0001. The first three pairs of canonical variates accounted for a significant
relationship between the two sets of variables. Data on the canonical variates are displayed
in Table 4. Indicated in the table are correlations between the variables and canonical variates,
within-set variance accounted for by the canonical correlations (percent of variance), redundancies
and canonical correlations. Total percentage of variance indicates that the third
canonical variate was minimally related to the two sets of variables and therefore the interpretation
of the third pair is questionable, even though this variate was also statistically significant
(χ2 = 41.80, p < .001). The interpretation of the first canonical variate can be drawn from the
correlations, indicating that the variate may be labelled as valence (inverted). This is because
tension (.88), fear (.94) and anger (.75) as well as valence (-.98) are projected with high loadings
onto the first variate. The second canonical variate could then be labelled as activity
(inverted).This is because energy (-.92), happiness (-.64), and sadness (.85) receive the highest
correlations with the second canonical variate.The interpretation of the third canonical variate
is precarious due to the low percent of variance that can be explained (0–5%). This analysis as a
whole puts forward the notion that the two conceptual frameworks are largely similar and the
minimal description of this mapping might reasonably have a two-dimensional structure.
In the second conceptual comparison, we employed regression to predict the dimensional ratings
from the discrete ratings and vice-versa. This technique has the advantage of providing a
well-known measure of fit (R2). As all 110 music examples were rated using both conceptual
frameworks, such a comparison was possible. Knowing that relatively high correlations exist
between the emotion ratings, collinearity of predictors was evaluated using a variance inflation
factor (VIF) for each set of predictors (basic emotion and dimensional models). For basic emotion
ratings, all VIF values remained lower than the suggested threshold value for collinearity (10;
see Cohen, Cohen, West, & Aiken, 2003, p. 423) but for valence, energy arousal and tension
arousal, VIF values indicated high collinearity (11.4, 5.3, 16.9). For this reason, the regression
estimates for each basic emotion concept with the dimensional model ratings as predictors was
performed using ridge regression. This technique is less influenced by collinearity due to the
inclusion of constant variance parameter (λ), which attenuates the influence of collinearity in
the calculation of the least squares optimization in regression (Cohen et al., 2003). In this case,
optimal λ was set at 50 (three predictors) and 100 (two predictors), which was established by
10-fold-over cross-validation with this data.The results – displayed inTable 5 – demonstrate that
the discrete emotion model can more accurately explain the results obtained with the threedimensional
model than vice-versa.This may partly be due to the higher amount of explanatory
Eerola and Vuoskoski 37
variables (five) but also the fact that discrete emotions are an easier concept for the general public
to understand than emotion dimensions. Nevertheless, the difference between the mean prediction
rates of the models (displayed in Table 5) is not large (17%) and this considerable degree
of overlap between the conceptual frameworks is noteworthy considering that neurological evidence
has suggested that separate processes might be involved (Dellacherie et al., 2008), and
that it is only relatively recently that both models have started to occur within the same study.
To examine the validity of the three-dimensional model, the coefficients of determination for
it were also compared with the circumplex model (Russell, 1980) and the multidimensional
model of activation (Thayer, 1989).The results suggest that these two-dimensional models can
explain the results obtained with the discrete emotion model virtually as accurately as the
three-dimensional model, with the exception of anger (see Table 5). The differences between
the prediction rates of the three alternative dimensional models were evaluated using a comparison
of the difference between two multiple correlations (Steiger, 1980), which involves
transforming the multiple correlations of the predicted models into Z scores and adjusting for
mutual correlation and sample size (Tabachnick & Fidell, 2001, p. 146). This analysis yielded
significant differences between all the different prediction rates of the models (see Table 5). It is
worth pointing out that, in comparison to other emotion categories, sadness was explained
equally modestly (R2 = .63) by all dimensional models. This may reflect the participants’ difficulty
with rating the valence of sad music, as previously mentioned. Despite this irregularity,
these analyses suggest fairly high mutual correspondence between the two conceptual frameworks
and stimulus sets, and further suggest that the common denominator between these
frameworks might be two-dimensional.
A requisite number of dimensions. The previous summary of the correlations between the emotion
ratings suggested high collinearity within the three-dimensional model (the overlap
between valence and tension). To examine whether the three-dimensional model could be
reduced to two dimensions, the independence of the three dimensions needed to be scrutinized.
Table 4. Correlations, canonical correlations percents of variance and redundancies between
dimensional and categorical ratings and their corresponding canonical variates
1st Can. Var. 2nd Can. Var. 3rd Can. Var.
Dimensional
Valence -.98 -.11 .04
Energy .20 -.92 -.03
Tension .88 -.37 -.13
Percent of variance .59 .33 .00
Redundancy .57 .29 .00
Categorical
Happiness -.75 .64 .05
Sadness -.07 .85 .05
Tenderness -.71 .53 .33
Fear .94 -.11 .94
Anger .75 -.37 .42
Percent of variance .51 .31 .06
Redundancy .49 .30 .05
Canonical correlation .99 .94 .57
38 Psychology of Music 39(1)
For instance, Roberts and Wedell (1994) have aptly demonstrated that the amount of dimensions
needed to explain common mood terms is influenced by stimulus density. In their study,
the common two-dimensional solution (valence and arousal) was not sufficient when a core set
of mood terms were supplemented with terms representing variants of anger and fear. In the
present study, two different types of reductions could be attempted: energy and tension dimensions
could be collapsed into one arousal dimension (Russell & Feldman Barrett, 1999) or
energy and valence, and tension and valence could both be collapsed into separate dimensions,
forming another two-dimensional construct (Thayer, 1989). The plausibility of these more
parsimonious models could be investigated by looking at the partial correlations between the
dimensions or by testing these explicitly with separate structural equation models. However,
when the intercorrelation between the model predictors is high (> .85), it is known to pose
severe problems for such iterative models that are based on a covariance matrix (Kline, 2004).
Therefore we resorted to a simpler strategy and looked at the partial correlations.
To investigate reducing three dimensions into the traditional two dimensions of valence and
arousal (Russell, 1980), we checked whether tension and energy, which correlate positively
(.57), could be collapsed into a single arousal dimension by partialling out the contribution of
the valence dimension. This analysis yielded a partial correlation of rte.v = .90 (p < .001), indicating
a considerable overlap between the concepts. As valence and energy did not correlate
significantly (-.08), we consequently received support for the traditional two-dimensional
model. Note that this is in contrast to the results obtained by Schimmack and Reizenstein
(2002), who used structural equation modelling to test the independence of the energy and
tension dimensions by controlling for valence, and they found no correlation between the residuals
of the energy and tension dimensions.
The other two-dimensional model (Thayer, 1989) casts valence into two separate dimensions
that consist of energy and tension. As we know from the first order correlations, valence
Table 5. Regression summary of dimensions and discrete emotions explaining discrete and dimensional
emotion ratings (N = 110)
R2 (b)
3D 2D (Russell) 2D (Thayer)
Dimensions as predictors (valence, energy, tension)
Happiness .89 (V0.93, E0.79, T-0.35)* .89 (V0.85, E0.49)* .86 (E0.64, T-0.62)
Sadness .63 (V-0.20, E-0.84, T-0.22) .63 (V-0.05, E-0.69) .60 (E-0.65, T-0.13)
Tenderness .77 (V0.33, E-0.45, T-0.58)** .74 (V0.50, E-0.51) .77 (E-0.34, T-0.61)**
Fear .87 (V-0.83, E0.07, T0.63)** .87 (V-0.90, E0.24)** .74 (E0.03, T0.85)
Anger .64 (V-0.52, E0.32, T0.35)** .68 (V-0.55, E0.35)** .54 (E0.22, T0.52)
Mean R2 .76 .76 .70
Discrete emotions as predictors (happiness, sadness, tenderness, fear, anger)
Valence .97 (H0.35, S-.11, T0.20, F-0.50, A-0.14)
Energy .88 (H0.47, S-0.32, T-0.42, F-0.05, A0.36)
Tension .93 (H-0.29, S-0.23, T-0.55, F0.18, A0.12)
Mean R2 .93
Notes: b = Standardized beta coefficients. For predicting basic emotions with emotion dimensions, ridge regression
was used (λ = 50 for 3-dimensions, and λ = 100 for 2-dimensions). For all models, F tests were significant at
p < .001. Asterisks denote significant difference between the regression models within each categorical emotion
(** at p < 0.01 level; * at p < 0.05 level).
Eerola and Vuoskoski 39
and tension correlate highly (r = -.83), but if we control for the contribution of energy then
this high correlation might be revealed to be spurious. Nevertheless, valence and tension correlate
even more highly when the energy ratings are partialed out (rvt.e = -.95), suggesting that
at least the high collinearity is not affected by the energy dimension. Partialling out the tension
from the correlation between energy and valence also makes them correlate even more highly
(rev.t = .85), lending further support for the possibility of dimensional reduction. In sum, both
theoretically derived ways of reducing the three dimensions into two are supported by these
analyses although the difference between the two reductions is not discernible.
A requisite number of emotion categories. We may also ask whether the ratings of the five discrete
emotions contain significant overlap. In comparing conceptual frameworks using regression,
we already observed that three to five discrete emotions were necessary to explain over 90% of
the ratings in three dimensions. We also noticed how fear and anger seemed to overlap (r = .69)
when looking at the first order correlations. Here we looked again at the partial correlations by
controlling the contribution of the other three discrete emotions while examining the correlation
of each pair of discrete emotions. Table 6 displays the results of such analysis. The most
striking partial correlations can be seen between happiness, sadness, and fear, all negative and
highly significant. This implies that if we removed the ratings of happiness from the data, we
would still be able to deduce that happy examples are those which are rated low on sadness,
anger, and fear. The case of fear is more complicated, however, as it correlates with sadness and
tenderness and so its removal could not be entirely constructed from the three other remaining
discrete emotions. Interestingly, fear and anger do not correlate with each other when the other
discrete emotion categories are partialled out. It seems that the contribution of happiness (rfa.h =
.60, p < .001) and sadness (rfa.hs = .18, p = ns) is enough to create this effect. Thus, the overlap
between happy and sad examples indicates that the real simplification of the discrete emotion
model may be in terms of the valence dimension. From the previous canonical correlation and
regression analysis we already know that this is a viable way of reducing the number of variables
in question. Nevertheless, the issue needs to ultimately be considered in a larger context
where the purpose of the measurement model can be taken into account.
Discussion
The results of the experiment suggest that the three-dimensional model of emotions may be
collapsed into a two-dimensional one when applied to music.The support for this interpretation
comes from (1) canonical correlations that highlighted two canonical variates which could
account for the correspondence between the discrete emotion and dimensional models, (2)
regression analysis which demonstrated that the ratings of discrete emotions may be recovered
to a large extent by a two dimensional model (≈80%) and vice versa (≈90%), and (3) analysis of
partial correlations, which emphasized the high correlational nature of valence and tension.
Nevertheless, the two possible formulations of the two-dimensional model could not be clearly
ranked using these analyses, although the canonical correlations indicated that the version by
Russell (1980; valence and arousal) was somewhat more appropriate. Also, the regression
approach suggested that the version by Russell was slightly better in accounting ratings of discrete
emotions (mean R2 = .76), than the alternative version by Thayer (1989; mean R2 = .73).
These results are in contrast to the ones obtained by proponents of the three-dimensional
model (Schimmack & Grob, 2000; Schimmack & Reisenzein, 2002), though we must emphasize
the main differences between the design of our study and theirs. Schimmack and Reisenzein
40 Psychology of Music 39(1)
(2002) used a questionnaire study with a large number of questions (18) that covered the
affect dimensions with several questions (three for each polar extreme). This allowed them to
construct and test latent variables from the separately observed variables using structural
equation modelling. Also the ratings of their study were based on current mood, and there was
no stimulus or manipulation of mood. In our study the ratings were given to emotions that the
participants thought the excerpts conveyed, and the excerpts were selected to portray discrete
emotions and polar extremes of the dimensional model. Because of this, the observed emotion
structure reflects more directly the stimulus structure. It should also be noted that, despite our
careful attempts to control the dimensions when selecting excerpts for the experiment based on
the results of the pilot experiment, valence and tension were already correlated. Therefore it is
difficult to estimate whether it is even possible to separate these dimensions in musical examples.
In other words, the question remains as to whether there is an abundant number of musical
pieces that could be highly tense and highly positively valenced at the same time. This is
therefore something that needs to be studied further.
General discussion
The work presented in this article aimed to systematically compare distinct models of emotion.
Although a small number of previous studies exist where discrete and dimensional data have
been collected (Gosselin et al., 2006; Kreutz et al., 2008; Vieillard et al., 2008; Zentner et al.,
2008), these have been incomplete with regard to the structure of emotion due to their (1) reliance
on discrete emotions only, (2) focus on unambiguous exemplars, or (3) insufficient stimulus
quantity. Here the set of musical stimuli was carefully selected in a large pilot study to
represent emotion concepts in the dimensional as well as the discrete emotion model. Moreover,
both models were represented not only by the clearest examples, but also by more moderate
examples. This provided subtle nuances for emotion recognition and linear geometry for comparing
the two conceptual sets using linear mapping methods.
The comparison of discrete and dimensional models yielded interesting results. Initially we
thought that the discrete emotion model would lead to more consistent ratings of emotions
than the dimensional model because the terminology (sad, happy, angry, etc.) is already familiar
to participants due to their prevalence in the everyday language of the general public. But
this was not observed in the data, as the overall consistencies between the ratings in the dimensional
and discrete models did not exhibit any substantial differences. However, the discrete
emotion model was clearly less reliable in rating excerpts that were ambiguous examples of an
emotion category when compared with the dimensional model. This has direct implications for
studies that seek to (1) explore mixed emotions (e.g., Hunter et al., 2008), (2) understand the
provocative differences in neural processes between dimensional and discrete emotion ratings
Table 6. Partial correlations between the ratings of basic emotions (N = 110)
Happiness Sadness Tender Fear
Sadness -0.88***
Tenderness -0.50*** -0.34***
Fear -0.84*** -0.73*** -0.63***
Anger -0.39*** -0.39*** -0.34*** -0.08
Note: The contribution of all other discrete emotions has been partialled out except the two (row and
column) used in the comparison. *** p < .001.
Eerola and Vuoskoski 41
(e.g., Gosselin et al., 2006), (3) examine processing biases exhibited by clinical populations to
inherently ambiguous emotion stimuli (e.g., Bouhuys et al., 1995), or (4) attempt to clarify the
way conceptual decisions are made within the framework of the hybrid model of emotions. For
such studies, it is important to be aware of this asymmetry in the reliability of the ratings
between the two models.
Despite the discrepancy in the resolution between the models, a high correspondence
between the discrete and dimensional models was observed. Probably a large part of the
assumed differences between the models has been caused by methodological differences. In
many of the previous studies on discrete emotions (Dellacherie et al., 2008; Kallinen, 2005;
Khalfa et al., 2008b), a forced-choice paradigm is used in emotion recognition. In the present
study, all discrete emotions were available in the form of Likert scales, allowing more subtlety in
definition than in a forced choice. In this way the methodologies used in both emotion models
were similar and thus were perhaps more likely to lead to converging results.
Another way of representing the high correspondence between the two conceptual models
is to consider a hybrid model of emotions (Christie & Friedman, 2004; Russell, 2003). This
model uses the components of a dimensional model (valence and arousal) to explain the underlying
affect space, which is mainly physiologically driven. When the changes in these core
affects are interpreted consciously, however, discrete emotion terminology is used to label the
emotional experiences. In this way common discrete emotions can be regarded as attractors or
hot spots in the affect space. This view is entirely compatible with the results of our experiment
– due to the selection of moderately and highly representative examples of discrete emotion
categories – and these attractors are explicit in the figures portraying the excerpts along the
three dimensions (see Figures 4 and 5). This model could also be used to characterize the main
difference between the utilitarian and aesthetic emotions (Scherer, 2004). Whereas the utilitarian
emotions such as fear or anger have specific connections to underlying physiology due to
the adaptive function they have (in order to protect the physical integrity of the individual),
aesthetic emotions do not. Indeed most of the domain-specific emotions established by Zentner
et al. (2008) concern positive emotional responses and match the established functions of
music as a reminder of past events (North, Hargreaves & Hargreaves, 2004), or have a direct
correlate in a core affect (e.g., joyful activation).
The comparison between different versions of the dimensional model indicated that two
dimensions are probably sufficient to represent perceived emotions in music. Nevertheless, this
should still be studied further as the stimuli in our experiment were initially chosen to represent
each of the six dimension extremes separately. Therefore the tense examples in the selection
were also the ones which were often negatively valenced. A random sample of a large corpus of
music that is known to manipulate emotions could be used to test for the validity of the threedimensional
model in music. Intuitively, the additional dimension of tension makes perfect
sense and a great deal of the effects of music deal with patterns of tension and release (Lerdahl,
2001). Moreover, the tension dimension did actually vary independently for certain discrete
emotion categories (e.g., sad examples in the lower panel of Figure 5).Tension is also one of the
nine factors in the GEMS model of music-induced emotions (Zentner et al., 2008). Whether the
other factors in GEMS model actually correspond with the traditionally used terms (sadness,
tenderness, peacefulness, and joyful activation to name the obvious ones) is an interesting
future research question. Therefore in our opinion, more research as to the specific dimensions
required to represent emotions in music is warranted.
In light of the results obtained in this study, we should also investigate more thoroughly
the influence of individual factors on the processing of music-mediated emotions, such as
42 Psychology of Music 39(1)
personality and musical expertise. It would also be important to evaluate the degree of mismatch
between perceived and felt emotions. That would entail changing the rating instruction
towards induced emotions but should also incorporate physiological measures, in which significant
steps have been recently taken (Baumgartner, Esslen, & Jancke, 2006; Roy, Mailhot,
Gosselin, Paquette, & Peretz, 2008; Withvliet & Vrana, 2006). Another important issue is the
limitation imposed by the musical style used. Although film music makes use of stereotypical
conventions from the romantic era of classical music, as well as more recent ways to use artificial
sound schemes and elements from popular music, it has to be acknowledged that the results
could well be different with other genres such as pop or jazz. The genre-specificity and stimulus
density effects are two related questions that warrant further systematic research in the near
future. Also, a thorough dissection of the acoustical and musical features of the stimuli should
be carried out in order to address some of the focal issues that have been raised by the study
(mixed feelings, the role of timbre, small differences in fear and anger, etc.). Finally, non-verbal
methods (such as paired similarity ratings) could give crucial insights into the requisite number
of emotion dimensions and categories (Bigand et al., 2005; Vieillard et al., 2008), provided that
the initial coverage of the stimuli is reasonably varied (Roberts & Wedell, 2004).
In conclusion, our study demonstrated that discrete and dimensional models of emotion
produce highly compatible ratings of perceived emotions, when using a large, systematically
chosen set of authentic music from film soundtracks. We also highlighted the noteworthy differences
between the models that mostly relate to the constrained resolution of the discrete
emotion model. In these respects, our study provides a useful point of reference for exploring
the connections between the recognition, experience and physiological manifestations of emotions,
as well as the individual variables that moderate all of these.
Acknowledgements
The work was funded by the European Union (BrainTuning FP6–2004-NEST-PATH-028570) and the
Academy of Finland (Finnish Centre of Excellence in Interdisciplinary Music Research). We would like to
thank the two anonymous reviewers for their constructive comments, Alex Reed for his thorough proofreading,
and the members of the expert panel for their creative efforts in the stimulus selection.
Notes
1. https://www.jyu.fi/music/coe/materials/emotion/soundtracks/
2. https://www.jyu.fi/music/coe/materials/emotion/soundtracks/
References
Altenmüller, E., Schuermann, K., Lim, V.K., & Parlitz, D. (2002). Hits to the left, flops to the right: Different
emotions during listening to music are reflected in cortical lateralisation patterns. Neuropsychologia,
40(13): 2242–2256.
Balkwill, L.-L., & Thompson, W. (1999). A cross-cultural investigation of the perception of emotion in
music: Psychophysical and cultural cues. Music Perception, 17(1): 43–64.
Balkwill, L.-L., Thompson, W., & Matsunaga, R. (2004). Recognition of emotion in Japanese, Western,
and Hindustani music by Japanese listeners. Japanese Psychological Research, 46(4): 337–349.
Barrett, L.F. (2006). Solving the emotion paradox: Categorization and the experience of emotion.
Personality and Social Psychology Review, 10(1): 20–46.
Barrett, L.F., & Wager,T.D. (2006).The structure of emotion: Evidence from neuroimaging studies. Current
Directions in Psychological Science, 15(2): 79–83.
Eerola and Vuoskoski 43
Baumgartner, T., Esslen, M., & Jancke, L. (2006). From emotion perception to emotion experience:
Emotions evoked by pictures and classical music. International Journal of Psychophysiology, 60(1):
34–43.
Bigand, E., Vieillard, S., Madurell, F., Marozeau, J., & Dacquet, A. (2005). Multidimensional scaling of
emotional responses to music: The effect of musical expertise and of the duration of the excerpts.
Cognition & Emotion, 19(8): 1113–1139.
Blood, A.J., & Zatorre, R.J. (2001). Intensely pleasurable responses to music correlate with activity in
brain regions implicated in reward and emotion. Proceedings of National Academy of Sciences, 98(20):
11,818–11,823.
Bouhuys, A.L., Bloem, G.M., & Groothuis, T.G.G. (1995). Induction of depressed and elated mood by
music influences the perception of facial emotional expressions in healthy subjects. Journal of Affective
Disorders, 33(4): 215–226.
Cacioppo, J.T., Berntson, G.G., Larsen, J.T., Poehlmann, K.M., & Ito, T.A. (2000). The psychophysiology
of emotion. In R. Lewis & J.M. Haviland-Jones (Eds.), The handbook of emotion (2nd ed., pp. 173–191).
New York: Guilford Press.
Cavanagh, J., & Geisler, M.W. (2006). Mood effects on the ERP processing of emotional intensity in faces:
A P3 investigation with depressed students. International Journal of Psychophysiology, 60(1): 27–33.
Christie, I.C., & Friedman, B.H. (2004). Autonomic specificity of discrete emotion and dimensions
of affective space: A multivariate approach. International Journal of Psychophysiology, 51(2):
143–153.
Cohen, J., Cohen, P., West, S., & Aiken, L. (2003). Applied multiple regression/correlation analysis for the
behavioral sciences. Hillsdale, NJ: Lawrence Erlbaum.
Collier, G.L. (2007). Beyond valence and activity in the emotional connotations of music. Psychology of
Music, 35(1): 110–131.
Dailey, M., Cottrell, G., Padgett, C., & Adolphs, R. (2002). EMPATH: A neural network that categorizes
facial expressions. Journal of Cognitive Neuroscience, 14(8): 1158–1173.
Davison, A.C., & Hinkley, D.V. (1997). Bootstrap methods and their application. Cambridge: Cambridge
University Press.
Dellacherie, D., Ehrlé, N., & Samson, S. (2008). Is the neutral condition relevant to study musical emotion
in patients? Music Perception, 25(4): 285–294.
Ekman, P. (1992). Are there basic emotions. Psychological Review, 99(3): 550–553.
Ekman, P. (1999). Basic emotions. In T. Dalgleish & M.J. Power (Eds.), Handbook of cognition and emotion
(p. 45–60). New York: John Wiley & Sons.
Etzel, J.A., Johnsen, E.L., Dickerson, J., Tranel, D., & Adolphs, R. (2006). Cardiovascular and respiratory
responses during musical mood induction. International Journal of Psychophysiology, 61(1): 57–69.
Evans, P., & Schubert, E. (2008). Relationships between expressed and felt emotions in music. Musicae
Scientiae, 12, 75–99.
Gabrielsson, A. (2001). Emotion perceived and emotion felt: Same or different?. Musicae Scientiae,
6(Special Issue 2001/2002): 123–147.
Gabrielsson, A., & Juslin, P.N. (1996). Emotional expression in music performance: Between the performer’s
intention and the listener’s experience. Psychology of Music, 24(1): 68.
Gold, A.E., MacLeod, K.M., Deary, I.J., & Frier, B.M. (1995). Hypoglycemia-induced cognitive dysfunction
in diabetes mellitus: Effect of hypoglycemia unawareness. Physiology & Behavior, 58(3): 501–511.
Gomez, P., & Danuser, B. (2004). Affective and physiological responses to environmental noises and music.
International Journal of Psychophysiology, 53(2): 91–103.
Gosselin, N., Peretz, I., Noulhiane, M., Hasboun, D., Beckett, C., Baulac, M., et al. (2005). Impaired recognition
of scary music following unilateral temporal lobe excision. Brain, 128(3): 628–640.
44 Psychology of Music 39(1)
Gosselin, N., Samson, S., Adolphs, R., Noulhiane, M., Roy, M., Hasboun, D., et al. (2006). Emotional
responses to unpleasant music correlates with damage to the parahippocampal cortex. Brain,
129(10): 2585–2592.
Gosselin, N., Peretz, I., Johnsen, E., & Adolphs, R. (2007). Amygdala damage impairs emotion recognition
from music. Neuropsychologia, 45, 236–244.
Hunter, P.G., Schellenberg, E.G., & Schimmack, U. (2008). Mixed affective responses to music with conflicting
cues. Cognition & Emotion, 22(2): 327–352.
Ilie, G., & Thompson, W. (2006). A comparison of acoustic cues in music and speech for three dimensions
of affect. Music Perception, 23(4): 319–329.
John, O.P., & Srivastava, S. (1999). The Big Five Trait taxonomy: History, measurement, and theoretical
perspectives. In L.A. Pervin & O.P. John (Eds.), Handbook of personality: Theory and research
(pp. 102–138). New York: Guilford Press.
Juslin, P.N. (2000). Cue utilization in communication of emotion in music performance: Relating performance
to perception. Journal of Experimental Psychology: Human Perception and Performance, 26(6):
1797–1813.
Juslin, P.N. (2001). Communicating emotion in music performance: A review and a theoretical framework.
In P.N. Juslin & J.A. Sloboda (Eds.), Music and emotion: Theory and research (pp. 309–340). Oxford:
Oxford University Press.
Juslin, P.N., Friberg, A., & Bresin, R. (2002). Toward a computational model of expression in music performance:
The GERM model. Musicae Scientiae, Special Issue 2001–2002, 63–122.
Juslin, P.N. & Sloboda, J.A. (Eds.) (2010). Handbook of music and emotion. New York: Oxford University
Press.
Juslin, P.N., & Västfjäll, D. (2008). Emotional responses to music: The need to consider underlying mechanisms.
Behavioral and Brain Sciences, 31, 559–575.
Kallinen, K. (2005). Emotional ratings of music excerpts in the western art music repertoire and their selforganization
in the Kohonen neural network. Psychology of Music, 33(4): 373–393.
Kallinen, K., & Ravaja, N. (2006). Emotion perceived and emotion felt: Same and different. Musicae
Scientiae, 10, 191–213.
Khalfa, S., Peretz, I., Blondin, J.P., & Manon, R. (2002) Event-related skin conductance responses to musical
emotions in humans. Neuroscience letters, 328(2): 145–149.
Khalfa, S., Roy, M., Rainville, P., Dalla Bella, S., & Peretz, I. (2008a). Role of tempo entrainment in psychophysiological
differentiation of happy and sad music? International Journal of Psychophysiology, 68(1):
17–26.
Khalfa, S., Delbe, C., Bigand, E., Reynaud, E., Chauvel, P., & Liégeois-Chauvel, C. (2008b). Positive and
negative music recognition reveals a specialization of mesio-temporal structures in epileptic patients.
Music Perception, 25(4): 295–302.
Kline, R.B. (2004). Principles and practice of structural equation modeling (2nd ed.). New York: Guilford
Press.
Kreutz, G., Ott, U.,Teichmann, D., Osawa, P., & Vaitl, D. (2008). Using music to induce emotions: Influences
of musical preference and absorption. Psychology of Music, 36(1): 101–126.
Krumhansl, C.L. (1997). An exploratory study of musical emotions and psychophysiology. Canadian
Journal of Experimental Psychology, 51(4): 336–352.
Lang, P.J., Bradley, M.M., & Cuthbert, B.N. (2005). International affective picture system (IAPS):
Digitized photographs, instruction manual and affective ratings. Gainesville, FL: University of
Florida.
Laukka, P., & Juslin, P.N. (2007). Similar patterns of age-related differences in emotion recognition from
speech and music. Motivation and Emotion, 31(3): 182–191.
Eerola and Vuoskoski 45
Leman, M., Vermeulen, V., De Voogdt, L., Moelants, D., & Lesaffre, M. (2005). Prediction of musical affect
using a combination of acoustic structural cues. Journal of New Music Research, 34(1): 39–67.
Ludbrook, J. (1998). Multiple comparison procedures updated. Clinical and Experimental Pharmacology and
Physiology, 25, 1032–1037.
Lerdahl, F. (2001). Tonal pitch space. New York: Oxford University Press.
Mikels, J., Fredrickson, B., Larkin, G., Lindberg, C., Maglio, S., & Reuter-Lorenz, P. (2005). Emotional category
data on images from the International Affective Picture System. Behavioral Research Methods,
37(4), 626–630.
Nawrot, E. (2003). The perception of emotional expression in music: Evidence from infants, children and
adults. Psychology of Music, 31(1): 75–92.
North, A., Hargreaves, D., & Hargreaves, J. (2004). Uses of music in everyday life. Music Perception, 22(1),
41–77.
Nyklìcek, I., Thayer, J.F., & Van Doornen, L.J.P. (1997). Cardiorespiratory differentiation of musicallyinduced
emotions. Journal of Psychophysiology, 11(4): 304–321.
Posner, J., Russell, J.A., & Peterson, B.S. (2005).The circumplex model of affect: An integrative approach to
affectiveneuroscience,cognitivedevelopment,andpsychopathology.DevelopmentandPsychopathology,
17(3): 715–734.
Puckette, M. (1996). Pure data. In Proceedings of the International Computer Music Conference (pp. 269–272).
San Francisco, CA: International Computer Music Association.
Rentfrow, P.J., & Gosling, S.D. (2003). The do re mi’s of everyday life: The structure and personality correlates
of music preferences. Journal of Personality and Social Psychology, 84(6): 1236–1256.
Roberts, J.S., & Wedell, D.H. (1994). Context effects on similarity judgments of multidimensional stimuli:
Inferring the structure of the emotion space. Journal of Experimental Social Psychology, 30(1), 1–38.
Roy, M., Mailhot, J.-P., Gosselin, N., Paquette, S., & Peretz, I. (2008). Modulation of the startle reflex by
pleasant and unpleasant music. International Journal of Psychophysiology, 71(1): 37–42.
Russell, J.A. (1980). A circumplex model of affect. Journal of Personality and Social Psychology, 39(6):
1161–1178.
Russell, J.A. (2003). Core affect and the psychological construction of emotion. Psychological Review,
110(1): 145–172.
Russell, J.A., & Fehr, B. (1994). Fuzzy concepts in a fuzzy hierarchy: Varieties of anger. Journal of Personality
and Social Psychology, 67(2): 186–205.
Russell, J.A., & Feldman Barrett, L. (1999). Core affect, prototypical emotional episodes, and other things
called emotion: Dissecting the elephant. Journal of Personality and Social Psychology, 76(5): 805–819.
Rusting, C.L. (1998). Personality, mood, and cognitive processing of emotional information: Three conceptual
frameworks. Psychological Bulletin, 124(2): 165–196.
Schellenberg, E.G., Peretz, I., & Vieillard, S. (2008). Liking for happy- and sad-sounding music: Effects of
exposure. Cognition & Emotion, 22(2): 218–237.
Scherer, K. (2004). Which emotions can be induced by music? What are the underlying mechanisms?
And how can we measure them? Journal of New Music Research, 33(3): 239–251.
Schimmack, U., & Grob, A. (2000). Dimensional models of core affect: A quantitative comparison by
means of structural equation modeling. European Journal of Personality, 14(4): 325–345.
Schimmack, U., & Reisenzein, R. (2002). Experiencing activation: Energetic arousal and tense arousal are
not mixtures of valence and activation. Emotion, 2(4): 412–417.
Schmidt, L.A., & Trainor, L.J. (2001). Frontal brain electrical activity (EEG) distinguishes valence and
intensity of musical emotions. Cognition & Emotion, 15(4): 487–500.
Schubert, E. (1996). Enjoyment of negative emotions in music: An associative network explanation.
Psychology of Music, 24(1): 18–28.
46 Psychology of Music 39(1)
Schubert, E. (1999). Measuring emotion continuously: Validity and reliability of the two-dimensional
emotion-space. Australian Journal of Psychology, 51(3): 154–165.
Schubert, E. (2007).The influence of emotion, locus of emotion and familiarity upon preference in music.
Psychology of Music, 35(3), 499–515.
Steiger, J.H. (1980). Tests for comparing elements of a correlation matrix. Psychological Bulletin, 87,
245–251.
Tabachnick, B.G., & Fidell, L.S. (Eds.) (2001). Using multivariate statistics. Boston, MA: Allyn & Bacon.
Tellegen, A., Watson, D., & Clark, L.A. (1999). On the dimensional and hierarchical structure of affect.
Psychological Science, 10(4): 297–303.
Terry, P., Lane, A., Lane, H., & Keohane, L. (1999). Development and validation of a mood measure for
adolescents. Journal of Sports Sciences, 17(11): 861–872.
Terwogt, M., & Van Grinsven, F. (1991). Musical Expression of Moodstates. Psychology of Music, 19(2):
99–109.
Thayer, R.E. (1989). The biopsychology of mood and arousal. New York: Oxford University Press.
Vieillard, S., Peretz, I., Gosselin, N., Khalfa, S., Gagnon, L., & Bouchard, B. (2008). Happy, sad, scary and
peaceful musical excerpts for research on emotions. Cognition & Emotion, 22(4): 720–752.
Watson, D., Wiese, D., Vaidya, J., & Tellegen, A. (1999). The two general activation systems of affect:
Structural findings, evolutionary considerations, and psychobiological evidence: The structure of
emotion. Journal of Personality and Social Psychology, 76(5): 820–838.
Withvliet, C.V.O., & Vrana, S.R. (2006). Play it again Sam: Repeated exposure to emotionally evocative
music polarises liking and smiling responses, and influences other affective reports, facial EMG, and
heart rate. Cognition & Emotion, 21(1): 1–23.
Wundt, W. (1896). Grundrisse der psychologie [Outlines of psychology]. Leipzig, Germany: Engelmann.
Zentner, M., & Eerola, T. (2010). Self-report based measures and models. In P.N. Juslin & J.A. Sloboda
(Eds.), Handbook of music and emotion (pp. 187–221). Oxford: Oxford University Press.
Zentner,M.,Grandjean,D.,&Scherer,K.(2008).Emotionsevokedbythesoundof music:Characterization,
classification, and measurement. Emotion, 8(4): 494–521.
Eerola and Vuoskoski 47
Appendix. List of audio tracks used in Experiment
No. Emotion & level Album name Track Min:Sec
001 Anger high Lethal Weapon 3 8 04:15–04:29
002 Anger high The Rainmaker 7 01:45–02:00
003 Anger high The Alien Trilogy 9 00:03–00:18
004 Anger high Cape Fear 1 02:15–02:30
005 Anger high The Fifth Element 19 00:00–00:20
006 Anger mod. Crouching Tiger, Hidden Dragon 8 01:12–01:25
007 Anger mod. Batman Returns 2 00:18–00:33
008 Anger mod. Man of Galilee CD1 6 00:40–01:07
009 Anger mod. The Untouchables 8 01:38–01:53
010 Anger mod. Oliver Twist 15 02:05–02:25
011 Fear high Batman Returns 5 00:09–00:25
012 Fear high JFK 8 01:26–01:40
013 Fear high JFK 8 00:08–00:25
014 Fear high The Alien Trilogy 5 00:26–00:41
015 Fear high Hannibal 1 00:40–00:54
016 Fear mod. Running Scared 6 02:53–03:07
017 Fear mod. The Untouchables 8 01:38–01:53
018 Fear mod. The Fifth Element 17 00:00–00:19
019 Fear mod. Lethal Weapon 3 7 00:00–00:16
020 Fear mod. Man of Galilee CD1 2 03:45–04:02
021 Happy high The Rainmaker 3 02:55–03:13
022 Happy high Batman 18 00:55–01:15
023 Happy high Shallow Grave 6 02:02–02:17
024 Happy high Man of Galilee CD1 2 03:02–03:18
025 Happy high Oliver Twist 1 00:17–00:34
026 Happy mod. The Omen 9 00:00–00:24
027 Happy mod. Oliver Twist 8 01:40–02:04
028 Happy mod. Grizzly Man 1 00:00–00:27
029 Happy mod. The Portrait of a Lady 3 00:23–00:45
030 Happy mod. Nostradamus 2 01:09–01:28
031 Sad high The English Patient 18 00:07–00:32
032 Sad high Running Scared 15 02:06–02:27
033 Sad high The Portrait of a Lady 9 00:00–00:22
034 Sad high Big Fish 15 00:55–01:11
035 Sad high Man of Galilee CD1 8 01:20–01:37
036 Sad mod. Angel Heart 4 00:08–00:28
037 Sad mod. Batman 5 01:08–01:22
038 Sad mod. Dracula 7 00:00–00:12
039 Sad mod. Shakespeare in Love 3 00:59–01:17
040 Sad mod. The English Patient 7 00:00–00:31
041 Tender high Shine 10 01:28–01:48
042 Tender high Pride & Prejudice 1 00:10–00:26
043 Tender high Dances with Wolves 4 01:31–01:48
044 Tender high Pride & Prejudice 12 00:01–00:15
045 Tender high Oliver Twist 8 00:14–00:30
046 Tender mod. Batman 9 00:00–00:19
047 Tender mod. Oliver Twist 8 01:15–01:32
048 Tender mod. Dracula 4 00:55–01:09
049 Tender mod. Juha 2 02:11–02:26
050 Tender mod. Oliver Twist 2 00:00–00:29
(Continued)
48 Psychology of Music 39(1)
No. Emotion & level Album name Track Min:Sec
051 Valence pos. high Juha 10 00:20–00:38
052 Valence pos. high Blanc 12 00:51–01:06
053 Valence pos. high Gladiator 17 00:14–00:27
054 Valence pos. high Pride & Prejudice 9 00:01–00:21
055 Valence pos. high Dances with Wolves 10 00:28–00:46
056 Valence pos. mod. Man of Galilee CD1 2 00:19–00:42
057 Valence pos. mod. Shakespeare in Love 21 00:03–00:21
058 Valence pos. mod. Vertigo OST 6 02:02–02:17
059 Valence pos. mod. Vertigo OST 6 04:42–04:57
060 Valence pos. mod. Outbreak 6 00:16–00:31
061 Valence neg. mod. Juha 18 02:30–02:46
062 Valence neg. mod. Shakespeare in Love 11 00:21–00:36
063 Valence neg. mod. Batman 9 00:57–01:16
064 Valence neg. mod. The Fifth Element 9 00:00–00:18
065 Valence neg. mod. Big Fish 15 00:15–00:30
066 Valence neg. high The English Patient 8 01:35–01:57
067 Valence neg. high Lethal Weapon 3 7 00:00–00:16
068 Valence neg. high Road to Perdition 6 00:34–00:49
069 Valence neg. high Hellraiser 5 00:00–00:15
070 Valence neg. high Grizzly Man 16 01:05–01:32
071 Energy pos. high The Untouchables 6 01:50–02:05
072 Energy pos. high Man of Galilee CD1 2 03:02–03:18
073 Energy pos. high Shine 5 02:00–02:16
074 Energy pos. high Shine 15 01:00–01:19
075 Energy pos. high Batman 18 00:55–01:15
076 Energy pos. mod. Juha 2 00:07–00:18
077 Energy pos. mod. Lethal Weapon 3 4 01:40–02:00
078 Energy pos. mod. Crouching Tiger, Hidden Dragon 13 01:52–02:10
079 Energy pos. mod. Batman 4 02:31–02:51
080 Energy pos. mod. Oliver Twist 7 01:30–01:46
081 Energy neg. mod. Juha 16 00:00–00:15
082 Energy neg. mod. Big Fish 15 00:55–01:11
083 Energy neg. mod. Big Fish 11 01:26–01:40
084 Energy neg. mod. Blanc 18 00:00–00:16
085 Energy neg. mod. Oliver Twist 6 00:51–01:07
086 Energy neg. high Running Scared 15 02:06–02:27
087 Energy neg. high Road to Perdition 16 00:17–00:32
088 Energy neg. high Blanc 10 00:13–00:31
089 Energy neg. high Blanc 16 00:00–00:15
090 Energy neg. high Batman Returns 12 00:57–01:14
091 Tension pos. high The Alien Trilogy 11 02:12–02:27
092 Tension pos. high The Fifth Element 13 00:17–00:31
093 Tension pos. high Babylon 5 3 02:47–03:00
094 Tension pos. high Hellraiser 10 02:44–03:00
095 Tension pos. high Oliver Twist 15 02:05–02:25
096 Tension pos. mod. The Missing 3 02:45–03:06
097 Tension pos. mod. Shallow Grave 4 01:04–01:19
098 Tension pos. mod. Naked Lunch 7 01:01–01:20
099 Tension pos. mod. Dracula 5 00:11–00:27
100 Tension pos. mod. Cape Fear 2 01:25–01:40
(Continued)
Appendix. (Continued)
Eerola and Vuoskoski 49
No. Emotion & level Album name Track Min:Sec
101 Tension neg. mod. Juha 2 02:11–02:26
102 Tension neg. mod. Shakespeare in Love 6 00:00–00:19
103 Tension neg. mod. The Fifth Element 12 00:00–00:17
104 Tension neg. mod. Crouching Tiger, Hidden Dragon 11 00:28–00:46
105 Tension neg. mod. Pride & Prejudice 4 00:10–00:29
106 Tension neg. high Lethal Weapon 3 10 01:59–02:17
107 Tension neg. high The Godfather 5 01:12–01:28
108 Tension neg. high Gladiator 4 00:48–01:06
109 Tension neg. high Pride & Prejudice 13 01:02–01:20
110 Tension neg. high Big Fish 8 00:12–00:34
Appendix. (Continued)