Emotions, Mechanisms, and Individual Differences in Music Listening: A Stratified Random Sampling Approach

The term prevalence is used to refer to the relative frequency of occurrence of a certain phenomenon, such as emotional reactions to music, in the population of interest (Juslin et al., 2008). Prevalence data capture the phenomena that any theory of music and emotion must be able to explain, and are also important to understand the ramifications of music as an application: What, exactly, might music achieve in terms of its emotional impact?

We define an emotion here as a relatively brief, intense, and rapidly changing reaction to a potentially important event (a subjective challenge or opportunity) in the external or internal environment. Theories of emotion come in many forms (e.g., basic emotions, appraisal theory, psychological construction), all of which have been applied to music (see Warrenburg, 2020). Music researchers distinguish between perception and induction of emotions: We may simply perceive an emotion “expressed” in the music, or we may actually feel an emotion in ourselves (for reviews, see Gabrielsson, 2002, and Schubert, 2013). Most previous studies have focused on perception of emotions (Eerola & Vuoskoski, 2013, Figure 2), yet what most people strive for is arguably to be “moved” (i.e., felt emotions).

The prevalence of specific emotions during music listening was initially mostly a matter of speculation among scholars, often based on personal experience (cf. Kivy, 1993). However, evidence is slowly accumulating from studies indicating that music could induce a fairly wide range of both “basic” (happiness, sadness, interest) and “complex” (nostalgia, pride) emotions, as well as “aesthetic” emotions (awe) (Gabrielsson, 2011; Juslin & Laukka, 2004; Juslin et al., 2011; Sloboda, 1992; Taruffi & Koelsch, 2014; Wells & Hakanen, 1991; Zentner et al., 2008). Moreover, emotions commonly regarded as “positive” in valence appear to be more prevalent than emotions commonly regarded as “negative” (Gabrielsson, 2011; Juslin et al., 2008, 2011; Juslin, Barradas et al., 2016; Sloboda et al., 2001). Positive emotions have a range of beneficial effects on physical health (Kubzansky, 2009) and subjective well-being (Fredrickson, 1998). Yet, they have been somewhat neglected in previous research on emotions (Lazarus, 1991).

Because estimates of emotion prevalence are influenced by a number of factors, such as individual characteristics, social context, and musical style, some researchers have advocated “method triangulation” to obtain representative samples of listeners, situations, and pieces of music, respectively (Juslin et al., 2010). Previous studies have sampled listeners (Juslin et al., 2011) or situations (Juslin et al., 2008). In this study, we sampled pieces of music to explore how this aspect moderates prevalence data.

The “what” question (i.e., emotion prevalence) is closely related to the “how” question: What emotion will occur in a given musical event is largely determined by how the emotion was induced. This issue is often regarded as the greatest mystery of them all (cf. Dowling & Harwood, 1986; Johnson-Laird, 1992). This is because in a typical case in everyday life, an emotion is induced when an event or object is appraised as having the capacity to influence the goals of the perceiver somehow. Yet, when we listen to music, only rarely has the music per se implications for our life goals, as described by appraisal theories (cf. Ellsworth, 1994).

This could explain why many scholars have leaned towards formalism (e.g., Kivy, 2002), which considers music as abstract tones sequences, devoid of semantic meaning. Such a view can lead to a neglect of the “meaning-making” role of psychological processes needed for any emotion to occur, in favor of a search for “direct” links between musical features and induced emotions (Coutinho & Cangelosi, 2011; Gomez & Danuser, 2007).

A large number of studies have attempted to predict perceived emotions in music based merely on acoustic features, and such attempts have admittedly been quite successful (Eerola, 2011; Juslin, 1997; Juslin & Lindström, 2010; Yang et al., 2018). However, there seems to be a “glass ceiling” with regard to the predictive accuracy that can be achieved with only acoustic features (Barthet et al., 2013). This has been referred to as a “semantic gap” between low-level acoustic features and high-level perception (Yang & Chen, 2011)—a problem that is arguably at its most acute for induced (i.e., felt) emotions. Daly et al. (2015) made an ambitious attempt to predict felt emotions based on acoustic features. They were able to explain less than 5% of the variance—and this was when acoustic features were augmented by EEG indices. A purely acoustic approach fared even worse. Coutinho and Cangelosi (2011) succeeded better, though they predicted changes in emotion within a piece of music, rather than across different pieces.

To improve predictive accuracy, researchers need to address how the emotional effects occur in the first place. They need to consider the underlying psychological processes that “mediate” between musical events and felt emotions. We shall refer to these processes as the mechanism (Juslin & Västfjäll, 2008).

Several different authors have suggested possible mechanisms for induction of emotion in music, usually involving just one or a few possibilities (Baumgartner, 1992; Berlyne, 1971; Juslin, 2001, 2013; Levinson, 1997; Meyer, 1956, 2001; Scherer & Zentner, 2001; Sloboda & Juslin, 2001). Space limitations prevent us from reviewing previous work in detail here, but the most comprehensive theory outlined so far is provided by the BRECVEMA framework (the acronym BRECVEMA derives from the first letter of each of the mechanisms featured in the theory, as listed below; for elaboration and predictions, see Juslin, 2019, Part III).

The BRECVEMA framework is consistent with a categorical approach to emotions, in the narrow sense that it assumes the existence of specific emotion categories, such as sadness and happiness. However, it does not quite adhere to any of the “traditional” emotion theories (e.g., basic emotions, appraisal, constructionism), but rather represents a novel type of theory sharing certain features with past theories. For example, it assumes that emotion mechanisms have a long evolutionary history (like basic emotion theory), it highlights the role of emotion learning and variability (like social construction theory), and it presumes close links between cognition and emotion (like cognitive appraisal theory). However, the framework also differs from these theories in some crucial ways. For instance, we do not think that “all emotions are basic” (Ekman, 1994, p. 15); nor do we believe that emotions, just like money, “are a product of human agreement” (Barrett, 2017, p. xiii); and we do not presume that “most emotions are elicited…through a process of cognitive appraisal” (Scherer et al., 2002, p. 150).

Unlike these theories, BRECVEMA holds that different types of emotions are induced by different types of mechanisms at different levels of the brain. (For example, the lowest levels may involve mainly the arousal dimension and proto-emotions, such as surprise, whereas the highest levels involve more complex, and even aesthetic, emotions; Juslin & Västfjäll, 2008; Juslin, 2013, 2019). It is argued that this framework is better able to account for the complex, multi-faceted emotional responses that may occur to music, than are traditional theories. The framework postulates eight mechanisms which involve (more or less) distinct brain networks:

• Brain stem reflex, a hard-wired attention response to subjectively “extreme” values of basic acoustic features, such as loudness, speed, and timbre (e.g., Davis, 1984); you may become startled or surprised by the loud beginning of a piece of music (Arjmand et al., 2017; Juslin et al., 2014).

• Rhythmic entrainment, a gradual adjustment of an internal body rhythm such as heart rate towards an external rhythm in the music (Bason & Celler, 1972; Harrer & Harrer, 1977); you may experience excitement if your heart rate becomes gradually synchronized with a highly captivating and somewhat faster rhythm in a piece of techno music at a nightclub.

• Evaluative conditioning, a regular pairing of a piece of music and other positive or negative stimuli leading to a conditioned association (e.g., Blair & Shimp, 1992; Bolders et al., 2012); you may feel happy when you hear a song which has repeatedly occurred in festive contexts.

• Contagion, an internal “mimicry” of the perceived voice-like emotional expression of the music (e.g., Juslin, 2001); you may experience sadness when you hear a slow, quiet, low-pitched performance of a classical piece on the cello that features much vibrato and rubato (Juslin et al., 2014; see also Egermann & McAdams, 2013).

• Visual imagery, inner images of an emotional character conjured up by the listener through a metaphorical mapping of the musical structure (Osborne, 1980; Taruffi & Küssner, 2019); you might become relaxed when you indulge in the mental images of a landscape suggested by a piece of “new-age” music while lying at home in your sofa.

• Episodic memory, a conscious recollection of a particular event from the listener’s past that is “triggered” by the music (see Baumgartner, 1992; Janata et al., 2007); you may experience nostalgia when a song evokes a vivid personal memory from the specific time you met your current partner in life (Garrido & Davidson, 2019).

• Musical expectancy, a response to the gradual unfolding of the syntactical structure of the music, and its expected or unexpected continuations (Huron, 2006; Meyer, 1956); you may feel anxious due to uncertainty created by phrases without a clear tonal center in an “avant-garde” piece (Juslin et al., 2014; see also Steinbeis et al., 2006).

• Aesthetic judgment, a subjective evaluation of the aesthetic value of the music, based on an individual set of weighted criteria (Juslin, 2013); you may admire the exceptional skills of a great performer at an evening concert (Juslin et al., 2021).¹

Several studies recently tested selected mechanisms in highly controlled experimental settings, using both synthesized and “real” musical excerpts (Juslin et al., 2014, 2015; Juslin, Sakka et al, 2016; Sakka & Juslin, 2018; see also Barradas et al., 2021). Target-mechanism conditions induced specific emotions in listeners largely in accordance with theoretical predictions, as shown by multiple indices (e.g., self-reported feelings, facial electromyography, psychophysiology). (For a review of neural correlates of musical emotions, see Juslin & Sakka, 2019.)

These results did not just reflect acoustic features of the music, such as tempo, sound level, or timbre. For instance, contrary to music-emotion correlations one might expect, the tempo could be faster in a piece that induced sadness than in a piece that induced happiness. Researchers concluded that this is because listeners’ responses are “driven” by mechanisms (e.g., whether a memory was evoked), rather than by acoustic features per se. In this study, we compared prediction of felt emotions based on acoustic measures with prediction based on subjective ratings of mechanisms.

The “how” question (mechanisms) is linked to the “who” question: Who will experience a certain emotion to a piece of music and who will not? Although the experimental studies cited above were designed to maximize experimental control and minimize the effects of contextual variables, the emotions induced were still not as neatly differentiated as one might hope.

Researchers have long acknowledged wide individual differences in musical emotions (e.g., Sloboda, 1996), and it has been suggested that there is greater variability for induction of emotions than for perception of emotions (Juslin, 2019). Most studies to date have tended to downplay individual differences and to focus only on means across listeners (but see, e.g., Juslin et al., 2021; Juslin, Sakka et al, 2016; Ladinig & Schellenberg, 2012; Mas-Herrero et al., 2013).

Still, it might be argued that our understanding of individual differences has important implications for the applicability of music in different contexts. Previous research has shown that personal preference and music choice play a key role for the effectiveness of music at an emotional level, both in laboratory studies (Liljeström et al., 2013) and applications (Garrido et al., 2017). Preferred music is likely to be more familiar to the listener, and familiarity itself can enable a greater number of emotion mechanisms (e.g., episodic memory) to be activated.

Why do two listeners (sometimes) respond differently to the same piece of music? Since the acoustic features of the music are the same, an explanation must clearly be sought in terms of a difference in the emotion-mechanism link. Specifically, individual listeners might activate different mechanisms to the same music (e.g., memory for one listener, contagion for another), thus inducing distinct emotions. In addition, for certain mechanisms (e.g., episodic memory), even the same mechanism may induce different emotions in different listeners, depending on previous personal experiences (e.g., Sakka & Saarikallio, 2020). Mechanism activation may in turn depend on factors such as attention, music training, personality traits, and individual learning history.

Because it is at the mechanistic level that individual differences in responses will tend to emerge, a mechanism focus is required to explain them. In this study, we made a first attempt to model emotion-mechanism links at an individual level, adopting a “statistical-ideographic” approach to musical emotions (Juslin, 2019; cf. Brunswik, 1956).

Most findings regarding emotions and mechanisms to date come from field studies that measure prevalence in various social and cultural contexts (e.g., Dingle et al., 2011; Juslin, Barradas et al., 2016). However, due to the lack of experimental control, it is not feasible to investigate individual differences in response to the same music. (Individual differences in response are typically confounded with differences in the music heard.)

Experimental studies, on the other hand, which have presented several listeners with the same pieces of music (Juslin et al., 2014, 2015), have tended to feature a very limited number of pieces that were selected or manipulated to target specific mechanisms. Thus, it is unclear how prevalence estimates of emotions and mechanisms may be affected if listeners are exposed to a larger and more “ecologically relevant” sample of music. The median number of musical stimuli in previous studies of music and emotion is 10, and the music selection has focused largely on classical music (Eerola & Vuoskoski, 2013), which represents a minority interest—even in the Western world (Hargreaves & North, 2010).

The overall aim of this study was thus to complement previous studies by investigating emotion-mechanism links within a controlled setting. The added value of randomly sampling music—as opposed to sampling listeners (Juslin et al., 2011) or situations (Juslin et al., 2008)—is that we are able to compare individual listeners, using a broad variety of musical examples. More specifically, we aimed to investigate the following questions: 1) Which emotions does music induce most frequently? 2) Which mechanisms occur most often? 3) Can emotions be predicted from self-reported mechanisms? 4) Are emotions best predicted from mechanisms or from acoustic features of the music? 5) Do different listeners show the same links among emotions and mechanisms?

These questions were addressed in a listening test using a correlational design (similar to studies in judgment analysis; cf. Cooksey, 1996). Participants listened to a large selection of musical excerpts, and were asked to rate felt emotions and subjective impressions related to various mechanisms. Whereas previous experiments have purposefully selected pieces of music so as to target particular mechanisms and induce predicted emotions, here we simply explored what emotions and mechanisms would occur spontaneously during music listening. It should be noted from the outset that such a correlational design limits the relative strength with which conclusions about causal relationships can be drawn.

We adopted a research approach characterized by five features: 1) we used a stratified random sampling procedure (e.g., Visser et al., 2000) to select the stimuli (in order to obtain a more representative sample); 2) we carried out both nomothetic (averaged) and idiographic (individual) analyses of emotion-mechanism links (an aspect of method development); 3) we focused on positive emotions (as these have been under-studied in the past); 4) we conducted extensive acoustic analyses (including a partly novel set of algorithms); and 5) we adopted a categorical—as opposed to dimensional—approach to emotion measurement (since one of our aims was to investigate the prevalence of specific emotions, and to compare our findings with those of previous studies adopting a categorical approach). We tested five hypotheses:

(H1) Based on previous studies that featured more or less representative samples of listeners (Juslin et al., 2011) or situations (Juslin et al., 2008), respectively, we hypothesized that positive emotions would be more prevalent than negative emotions.

(H2) Based on field studies that measured the prevalence of various mechanisms (Juslin et al., 2008; Juslin, Barradas et al., 2016), we hypothesized that the two mechanisms Brain stem reflex and Musical expectancy would be less prevalent than the other mechanisms.

(H3) Based on previous experimental studies (Barradas et al., 2021; Juslin et al., 2014, 2015; Sakka & Juslin, 2018), we hypothesized that the mechanisms would be reliably linked to specific emotions, as shown by accurate prediction of felt emotions (defined as R² ≥ .64; a “strong” effect; Ferguson, 2009) based on mechanism ratings in multiple regression analyses.

(H4) Based on findings in Juslin et al. (2014) and theoretical arguments in Juslin (2019; see also previous section), we hypothesized that self-reported mechanism impressions would better predict felt emotions than would acoustic features.

(H5) Based on the argument that individual differences in emotional reactions to music primarily occur at the mechanistic level (Juslin, 2019, p. 396), we hypothesized that listeners would show different emotion-mechanism links in idiographic regression analyses.

In addition to testing the above hypotheses, we also correlated the ratings of emotions with measures of the Big Five personality traits (John et al., 2008), in an explorative attempt to account for individual differences. Some preliminary studies have suggested that emotion prevalence might be moderated by personality (Barrett et al., 2010; Juslin et al., 2008, 2011; Liljeström et al., 2013; McCrae, 2007; Pilgrim et al., 2017).

The results reported in this paper were collected at the same time and featuring the same listeners and stimuli as the findings on how different criteria contribute to aesthetic judgments presented in Juslin, Sakka et al. (2016). There is no overlap in the data reported, however, except for the characteristics of the listener sample.

Forty-four listeners, 22 females and 22 males, 19–66 years old (M = 33.57, SD = 14.13) participated in the study. Their anonymous and voluntary participation was compensated with either course credits or two cinema vouchers. To obtain a broad sample not limited to students, the participants were recruited by means of posters throughout the town of Uppsala (e.g., shopping malls, the town library), as well as via a mailing list for music seminars.

Seventy-five percent of the participants played (at least) one musical instrument, while 52% stated they had received music education. Pre-screening showed that music preferences varied widely, with most listeners marking several genres of music; however, rock, classical, pop, electronica, and jazz were the most frequently reported genres. None of the participants reported a hearing problem. They were randomly divided into two groups.

In obtaining the sample of music, we tried to achieve an optimum balance between the need to feature as many musical stimuli as possible to obtain a sufficient number of cases for the multiple regression analyses and the need to keep the experimental sessions short enough to avoid severe fatigue effects (Cooksey, 1996).

With the upper limit on the number of stimuli in mind, we decided to use the stratified random sampling procedure (Visser et al., 2000), which may help to ensure that a sufficient variety of musical genres is obtained even with a limited sample size. Each participant group listened to 40 pieces of music. (The complete list appears in Appendix A.) We selected pieces randomly, with strata corresponding to the STOMP factors for music preference proposed by Rentfrow and Gosling (2003) based on factor analysis of music preferences. (STOMP stands for Short Test of Music Preferences.) The framework is clearly the most extensive attempt so far to map the underlying structure of music preferences, and the STOMP factor structure has been replicated in several countries (Delsing et al., 2008; Gouveia et al., 2008; Langmeyer et al., 2012; Zweigenhaft, 2008). However, we modified their structure slightly to suit Swedish conditions (e.g., dropping the Religious genre and adding Schlager/Dansband), and to obtain a similar number of genres for each factor. (The aim of the study was not to examine specific genres, but rather to obtain a sufficient variety of musical stimuli.) The four factors and their respective genres were:

1. Reflective & Complex (R&C): Classical/Opera; Jazz/Blues; Folk/Singer-songwriter

2. Intense & Rebellious (I&R): Classic rock; Heavy metal/Hard rock; Alternative/Punk

3. Upbeat & Conventional (U&C): Country; Pop; Schlager/Dansband

4. Energetic & Rhythmic (E&R): Hip-hop/Reggae; Soul/Funk; Club/House

Seventy-two pieces of music (six for each genre and 18 for each factor) were randomly sampled from the internet database Spotify. (We used the Radio feature to create a random playlist of songs, based on pre-specified genres. At the time the sample was drawn, we benefited from existing genre labels corresponding to those used here. The categories in Spotify have since then been changed.) The songs were randomly divided across the two listener groups with the provision that each group should have three songs per genre, leaving both groups with 36 songs apiece. In addition, four songs from each listener group (one from each factor) were randomly selected and added to the other group, such that eight pieces were common to the two groups—facilitating group comparison. The resulting sample may be said to be representative of what a Swedish music listener would typically encounter on the radio, in CD shops, at live concerts, or in databases on the internet at the time of the study.

Each piece was edited to a 60-second excerpt. This was done to maximize the number of pieces included without subjecting participants to an exhausting listening session. Moreover, responses to music can change quickly, so the longer the musical event, the more problematic the retrospective self-report would be. However, a previous study indicated that about 97% of the emotions experienced in daily life last longer than one minute (Frijda et al., 1991; for similar results with music, see Scherer et al., 2002). Thus, it would appear unlikely that the experienced emotion category would change much during the 60-second segment of music. The present stimuli are shorter than both the mean (135 s) and median (90 s) stimulus duration of previous studies of felt emotions in music, as reviewed by Eerola and Vuoskoski (2013). Note, however, that stimuli should not be too short, since different mechanisms may differ in terms of the time they need in order for the emotion-induction process to take place (e.g., rhythmic entrainment may take longer than a brain stem reflex; Juslin, 2019, Chapter 25).

For most songs (87%), the excerpt featured the first 60 seconds of the recording. For the remaining pieces (e.g., songs with a long introduction), the recording was edited manually, so as to feature a more meaningful section (e.g., verse, chorus) of the song (see Appendix A). All excerpts ended with a quick fade and were stored in high bit-rate mp3 format.

Like most emotion theorists, we assume that emotions involve multiple components, such as feeling, physiology, and expression (Scherer, 2000). In this study, we measured the feeling component, which is often regarded as the most crucial aspect of a musical emotion (Zentner & Eerola, 2010). Most theorists presume that self-reports of feelings have validity (Barrett, 2004). Thus, we measured the listeners’ feelings using 12 rating scales: happiness-elation, sadness-melancholy, surprise-astonishment, calm-contentment, interest-expectancy, nostalgia-longing, anxiety-nervousness, pride-confidence, anger-irritation, love-tenderness, disgust-contempt, and awe-admiration. We used two words to denote each unipolar scale to emphasize that the scale should be interpreted as referring to a broad emotion category that may include a variety of similar states. These scales represent a kind of compromise among the response formats currently used to measure musical emotions (for a review, see Zentner & Eerola, 2010), because the terms includes some “basic emotions” characteristic of discrete emotion theories (Izard, 1977), cover all four quadrants of the circumplex model in terms of valence and arousal (Russell, 1980), and feature possibly more music-related terms, such as nostalgia, expectancy, and awe (Juslin & Laukka, 2004). (The selected terms roughly cover the nine factors of GEMS - 9, proposed by Zentner et al., 2008, but because that scale lacks some terms that were needed in this study, e.g., surprise, we decided to use a customized list of emotion terms.) These terms include the emotions that were most prevalent in previous research, but also cover other emotions, to not pre-judge the outcome. Thus, the terms anxiety, anger, and disgust were mainly intended as controls. Consistent with our focus on positive emotions, the list includes a greater number of positive emotions than typical instruments that measure specific emotions (e.g., Plutchik, 1994, Chapter 5). All emotions were rated on a scale from 0 (not at all) to 4 (a lot) in response to the instruction describe how you felt when you heard the music. (A table of intercorrelations among emotion ratings appears in Appendix B.)

We collected subjective data about the mechanisms that may have occurred, using the MecScale (e.g., Juslin et al., 2014). This consists of eight questions, each targeting one of the mechanisms in the BRECVEMA framework: 1) Did the music feature an event that startled you? (Brain stem reflex); 2) Did the music have a strong and captivating rhythm? (Rhythmic entrainment); 3) Did the music evoke memories of events from your life? (Episodic memory); 4) Did the music induce emotions through an association? (Evaluative conditioning); 5) Did the music evoke inner images that influenced your emotions? (Visual imagery); 6) Were you “touched” by the emotional expression of the music? (Contagion); 7) Was it difficult to guess how the music (e.g., the melody) would develop over time? (Musical expectancy); 8) Did you find the music aesthetically valuable? (Aesthetic judgment) Listeners were asked to rate each item on a scale from 0 (not at all) to 4 (a lot).

The MecScale does not purport to measure the induction mechanisms “directly.” Rather, it measures subjective impressions that are reflective of mechanisms. Some mechanisms are straightforward to report (e.g., startle reflexes, memories, and images are very distinct). Others may at least be correlated with certain impressions (e.g., a listener who becomes surprised via musical expectancy will find a song “unpredictable”). MecScale is intended as a cost-effective index for decision-making purposes, for which the most crucial criterion is predictive validity (also referred to as concurrent validity if predictor and criterion are indexed at the same time). The items have been reported to be predictive of both target-mechanism conditions (Juslin et al., 2014; Sakka & Juslin, 2018; see also Barradas et al., 2021) and felt emotions (Juslin et al., 2015) in previous studies. In particular, the MecScale has been validated in a series of highly controlled listening experiments, where musical stimuli were carefully edited so as to present or withhold the type of information required to trigger distinct mechanisms (e.g., Juslin et al., 2014). The predictive validity was quantified and the MecScale items showed significant and strongly positive correlations with their respective target-mechanism conditions and negative correlations with all other mechanism conditions; none of the items correlated positively with neutral control conditions. In a study by Juslin et al. (2015), the MecScale items could predict the right target-mechanism condition from MecScale ratings with an accuracy of 75% correct. A single-item measure has the obvious advantage that it minimizes participant burden, which is especially important in the case of a long listening test.

The listeners’ personality traits were measured using the Big Five Inventory (BFI; John et al., 1991; for a review, see John et al., 2008). This includes 44 items measuring the factors Openness to experience (O), Conscientiousness (C), Extraversion (E), Agreeableness (A), and Neuroticism (N) (OCEAN). Each factor has also been divided into personality facets that are represented by the 44 items. The domain scales of BFI have shown good internal consistency, clear factor structure, adequate convergent-discriminant validity coefficients, and substantial self-peer agreement (Fossati et al., 2011).

Participants indicated on a 5-point Likert scale the extent to which they agreed to each self-statement (from 1 = disagree strongly to 5 = agree strongly). Scores were calculated for each factor. Cronbach’s alpha reliability for the domain scales were .85 (N), .84 (E), .64 (O), .59 (A), and .80 (C), respectively.²

To be able to compare mechanism-based and acoustics-based prediction of self-reported emotions (H4), we needed a proper acoustic description of each musical stimulus, in the form of numerical values that could be entered into multiple regression analyses. Using a new set of algorithms developed by one of the authors (OL), we obtained 14 acoustic features covering six major aspects: tempo (Metroid M, Metroid SD); dynamics (Crescendo); timbre (Bass, Midrange, Treble, Roughness); harmony (Key clarity M, Mode M, Mode SD); rhythm (Metrical strength M, Metrical strength SD); and musical structure (Mfcc Nov, a measure of changes in timbre, KeyQuant15, a measure of tonal modulation). The goal was to include as many acoustic features as possible within the limits set by the minimum 5 to 1 ratio of cases to every predictor stated in the judgment-analysis literature (Cooksey, 1996). With 72 pieces of music, we could thus include 14 features (14 x 5 = 70 cases). A more detailed description of the features and their algorithms appears in  Appendix C.

When the participants arrived at the music laboratory, they were seated in a comfortable armchair, in front of a computer monitor. A brief introduction and instructions concerning the experiment were shown on the monitor. The participants were informed that they would listen to a variety of pieces of music, and that after each piece, they should describe their experience of the music by means of rating scales on the computer screen. The instructions explained that the participants should rate their own feelings and not what the music is expressing. (Previous research has indicated that listeners can make this distinction; see Zentner et al., 2008.)

After informed consent was obtained from the participants, they were told to relax for a few minutes during silence. Then, the listening test began. After each piece, the rating scales (see Measures) appeared on the computer screen. During the rating, a 30-second segment of the musical excerpt was repeated at a lower sound level to facilitate the ratings (the segment included the beginning of the complete excerpt and was the same for all listeners). After the rating followed a 10-second pause before the next piece was played. Because the focus of the study was on the listener responses rather than on the musical stimuli, the stimulus order was randomized and kept constant across listeners so as to enable individual comparisons (as in a psychometric test). Had we used a unique stimulus order for each participant, any difference between listeners might simply reflect order effects (which appear to be strong in music; see Flores & Ginsburgh, 1996). All participants were tested individually. The only other people present during testing were the experimenters (LS, GB) who were separated from the listener by a blind. The room was dimly lit and sound attenuated. Stimuli were played through a pair of high-quality loudspeakers (Dali Ikon 6 MK2) at a comfortable sound level, which was kept constant across participants.

After the listening test, participants filled out a background questionnaire and the BFI. Sessions lasted between 100 and 180 min, depending on the participant. As noted above, we also collected other data during the sessions that are reported separately (Juslin, Sakka et al., 2016).

With 44 listeners, 20 rated variables (12 emotions and 8 mechanisms), and 40 excerpts per listener, we obtained a total of 35,200 data points to analyze in depth. Data quality checks indicated that the ratings of each variable used the full range of the scale (0-4); that there was no straightlining³ of responses; and that there were few missing values: < 4 ‰ for emotions; < 2.5 ‰ for mechanisms. Examination of the distributions of the variables showed no skewness values outside the -2 to +2 range, and no (absolute) kurtosis values outside the -7 to +7 range (e.g., Byrne, 2010; Hair et al., 2010; West et al., 1995),⁴ with the only exception of the ratings of disgust-contempt (which featured a very large number of zeros). However, we retained this variable in the first analyses, because a recent Monte Carlo simulation study, featuring 10,000 replications of 1,308 conditions, concluded that ANOVAs are highly robust against deviations from normality (Blanca et al., 2017).

Prevalence of emotions was measured by means of the mean ratings of the 12 emotion scales. To explore how emotion prevalence varied as a function of emotion category, we conducted a two-way mixed ANOVA, with Emotion as within-subjects factor (12 levels) and Group as between-subjects factor (2 levels). The results showed a significant main effect of Emotion, F(11, 462) = 66.955, p < .001, partial η² = 0.615, approaching a “strong” effect in terms of Ferguson’s (2009) guidelines. By contrast, there was no significant effect of Group, F(1, 42) = 1.713, p = .198, partial η² = 0.039; nor was there a significant interaction between Emotion and Group, F(11, 462) = 1.200, p = .284, partial η² = 0.028. As illustrated in Figure 1, upper panel), the overall trends were strikingly similar for the two groups, who listened to the same musical genres, albeit (mostly) different musical excerpts.

The impression that the two listener groups reacted similarly was reinforced by further analyses. Recall that eight musical excerpts were heard by both groups (Method section). This enabled us to compare the responses of the groups to the same pieces of music. We conducted a two-way mixed ANOVA, with Piece as within-groups factor (8 levels / pieces) and Group as between-groups factor, for each emotion scale. After Bonferroni adjustment for multiple tests (n = 36) from α = .05 to α = .0014, results revealed a significant effect of Piece on seven of the emotion scales (happiness-elation, surprise-astonishment, nostalgia-longing, irritation-anger, love-tenderness, disgust-contempt, and awe-admiration). By contrast, there was no significant main effect of Group, nor any significant interaction involving Group. These data suggest that the two groups experienced the common excerpts in a largely similar manner, at a nomothetic level of analysis.

Given the absence of both main and interaction effects of Group, we collapsed the data across groups. Figure 1 (lower panel) shows mean values and 95% confidence intervals, and Table 1 presents post hoc tests of all differences (in the form of Tukey’s HSD). Happiness-elation was significantly more prevalent than all other emotions, except interest-expectancy, which, in turn, was more prevalent than all the remaining emotions except calm-contentment. Calm-contentment was more prevalent than all remaining emotions except nostalgia-longing. The above emotions were all significantly more prevalent than a set of slightly less frequent, but still fairly common states (sadness-melancholy, surprise-astonishment, pride-confidence, love-tenderness, and awe-admiration) that did not differ significantly from one another. The least common states were anxiety-nervousness, anger-irritation, and disgust-contempt, which were significantly less common than the other emotions, but did not differ significantly from one another. Notably, the descriptive statistics presented in Figure 1 conceal some quite large individual differences in overall prevalence of specific emotions (i.e., across pieces of music). Even in the most prevalent emotion category (i.e., happiness-elation), there was considerable variability, as indicated by the lowest (0.675) and highest (2.800) individual mean values.

To test whether positive emotions were more prevalent than negative emotions (H1), we computed the individual mean values for both types of emotions and carried out a within-subjects t-test of the difference. Based on previous emotion research (Plutchik, 1994; Russell, 1980), we categorized five emotions as positively valenced (happiness-elation, calm-contentment, interest-expectancy, pride-confidence, and love-tenderness), and four emotions as negatively valenced (sadness-melancholy, anxiety-nervousness, irritation-anger, disgust-contempt). Three emotions (surprise-astonishment, nostalgia-longing, awe-admiration) were difficult to categorize with respect to valence and were therefore left out from the analysis as a precaution.⁵ The results showed that mean ratings of emotion prevalence were significantly higher for positive emotions (M = 1.377, SD = 0.494) than for negative emotions (M = 0.608, SD = 0.418), t₄₃ = 12.142, p < .001, d = 1.287, a “large” effect (Cohen, 1988). This tendency for positive emotions to be more prevalent than negative emotions occurred for all individual listeners except two, which corresponds to 95% of the sample.

As an exploratory follow-up analysis, we also looked at how the prevalence of specific emotions varied depending on genre. Figure 2 presents means and 95% confidence intervals for the 12 emotion categories, as a function of the STOMP factors (see Method section). As seen, the prevalence patterns were mostly similar across the four factors. However, note that the R&C category evoked significantly more calm-contentment and awe-admiration than the I&R category (both p’s < .05, as shown by post hoc tests in the form of Tukey’s HSD).

In the following emotion analyses, we will leave out the three emotions with the lowest prevalence (anxiety-nervousness, irritation-anger, and disgust-contempt), consistent with our focus on positive emotions, and because the low prevalence (M < 0.60) meant that there were few cases.

To examine how mechanism prevalence varied as a function of mechanism type, we conducted a two-way mixed ANOVA with Mechanism as within-subjects factor (8 levels) and Group as between-subjects factor (2 levels). The results revealed a significant main effect of Mechanism, F(7, 294) = 34.808, p < .001, partial η² = 0.453, a “moderate” effect (Ferguson, 2009). In contrast, there was no significant effect of Group, F(1, 42) = 2.589, p = .115, partial η² = 0.058. Nor was there a significant interaction between Mechanism and Group, F(7, 294) = 1.131, p = .343, partial η² = 0.026. As illustrated in Figure 3 (upper panel), the mean trends were quite similar for the two listener groups (except that Group 2 showed marginally higher ratings for all mechanisms except Brain stem reflex and Musical expectancy).

We also compared the listener groups regarding the eight pieces of music that were rated by both groups. We conducted a two-way mixed ANOVA, with Piece as within-groups factor (8 levels/pieces) and Group as between-groups factor for each mechanism scale. After Bonferroni correction for multiple tests (n = 24) from α = .05 to α = .0021, the results showed a significant effect of Piece for each mechanism. Conversely, there were no significant main effects of Group and only a single significant interaction involving Group. (Group 1 rated the Musical expectancy mechanism higher than Group 2 for the Sepultura excerpt; Appendix A.)

Given the overall absence of significant group effects, we collapsed data across groups. Figure 3 (lower panel) shows mean values and 95% confidence intervals, and Table 2 shows post hoc significance tests of all differences (in the form of Tukey’s HSD). As can be seen in Figure 3, the most frequent mechanisms were Rhythmic entrainment, Aesthetic judgment, Evaluative conditioning, Visual imagery, and Episodic memory. Rhythmic entrainment was rated as significantly more common than all other mechanisms, except Aesthetic judgment, which was significantly more common than all mechanisms, except Evaluative conditioning and Visual imagery. These mechanisms in turn did not differ significantly from one another, but were both more frequent than all other mechanisms except Episodic memory, which did not differ from Contagion. In addition, as hypothesized (H2), Brain stem reflex and Musical expectancy received significantly lower mean ratings than all other mechanisms, except that the contrast between Musical expectancy and Contagion did not reach significance (p = .06).

Similar to our analyses of the emotion categories, we also plotted the prevalence of emotion mechanisms as a function of STOMP category. Figure 4 shows the means and 95% confidence intervals. As can be seen, mechanism prevalence did not vary much across genre categories. However, note that the R&C genre category produced lower ratings of Rhythmic entrainment and higher ratings of Musical expectancy than the other categories (all p’s < .05, Tukey’s HSD).

To investigate how well listeners’ ratings of mechanism impressions (MecScale) could predict ratings of specific emotions at a nomothetic level (H3), we conducted a series of multiple regression analyses, based on the mean values across listeners. More specifically, the mean rating on the respective emotion scale was the dependent variable, and the mean ratings on the MecScale items were the independent variables (all variables coded continuously). The unit of analysis was piece of music; that is, each piece was a case, with a mean value for each variable. Because there were such small differences in ratings between the two listener groups and the analysis proceeded at a mean level, we decided to include all 72 pieces in the analyses. For the eight excerpts that were rated by both groups, we simply used the mean across groups. We adopted a “simultaneous” regression approach, which should be used whenever there is no theoretically defensible order of the predictors (e.g., Cohen & Cohen, 1983). Consideration of Tolerance values (the inverse of the Variation Inflation Factor) showed that multi-collinearity among the predictors was not a problem (Tabachnick & Fidell, 2001).

Table 3 shows a summary of the results. After Bonferroni correction for multiple tests (n = 9), from α = .05 to α = .0056, all multiple correlations remained statistically significant. As can be seen, the multiple correlations varied depending on the emotion scale—largest for nostalgia-longing, and smallest for calm-contentment—but did not fall below R = .754 (M = .847). The mean variance accounted for (R² = 72%, 95% CI [.649, .793], SD = 0.093) shows that listeners’ felt emotions could be predicted rather well based on mechanism ratings.

Table 3 also presents the beta weights (β) for the predictors (i.e., the mechanism items) of each model, which (in accordance with how the regression models were computed) should be interpreted in a row-wise fashion. The beta weights that are both statistically significant (p < .05) and positive in direction have been marked in red.

As seen in Table 3, each emotion category was linked with a few key mechanisms (i.e., predictors); for instance, happiness-elation was predicted by Rhythmic entrainment; sadness-melancholy was predicted by Evaluative conditioning and Contagion; surprise-astonishment was predicted by Brain stem reflex and Musical expectancy; calm-contentment was predicted by Contagion and Aesthetic judgment; nostalgia-longing was predicted by Episodic memory, as well as Aesthetic judgment; pride-confidence was predicted primarily by Entrainment and Aesthetic judgment; love-tenderness was mainly predicted by Contagion and Visual imagery; and awe-admiration was predicted by Aesthetic judgment, and also to some extent Contagion. Interest-expectancy stood out by being associated with several mechanisms (see Table 3).

To compare our prediction of felt emotions based on the ratings of mechanism items with one based on acoustic features (H4), we conducted a series of multiple regression analyses for the various emotions, using the data obtained with the computational algorithms (see Method section). The mean rating on the emotion scale was the dependent variable, and the acoustic features were the independent variables (all variables coded continuously). The analyses were in other regards similar to the ones conducted based on mechanism ratings. As acoustic features are more or less correlated, possible multicollinearity among predictors was a concern. However, checks revealed no Tolerance values < 0.25. Hence, although there was some moderate collinearity, levels were clearly within acceptance (see Tabachnick & Fidell, 2001).

Table 4 presents the main results from the analyses in terms of multiple correlations and F values. Also included are the results from our mechanism-based prediction, thus enabling a comparison of the variance accounted for by the two types of models (the right-most column). As can be seen, the multiple correlations (R) for the models based on acoustic features ranged from .500 to .708 (M = .622, SD = .072). Only four of the models were statistically significant, but the mean variance accounted for (R² = 39%, 95% CI [.324, .460], SD = 0.088) shows that felt emotions could be predicted to a moderate extent based on acoustic features. (The precise patterns of features are beyond the focus of this study, but a summary of correlations between emotions and features is provided in Appendix D.)

However, it may also be seen in Table 4 that mechanism-based models were better able to predict felt emotions. Note that the multiple correlation (R) was significantly larger for the mechanism-based model than for the acoustics-based model for all emotion categories except calm-contentment (tested by means of Steiger’s, 1980, Z test for dependent correlations with one variable in common, one-tailed tests). The mechanism-based models accounted for 72% of the variance in emotion ratings, on average, as compared with only 39% for the acoustics-based models.

To investigate the extent to which individual listeners display the same emotion-mechanism links (H5), we conducted a series of idiographic multiple regression analyses. One disadvantage of idiographic analyses is that they tend to yield large amounts of data. Thus, we restricted our analyses to a subset of 10 participants (five males and five females), which were randomly selected using the create a random sample function in the Statistica software. Since the aim was to compare individual listeners, we sampled listeners from just one of the listener groups (Group 2 - also randomly selected), who heard the same musical excerpts. We focused on the four most prevalent emotion categories: happiness-elation, calm-contentment, interest-expectancy, and nostalgia-longing. (Notably, this included both the most successfully and the least successfully predicted of the emotions in the nomothetic analyses.) The number of cases (N = 40) is relatively small for a multiple regression analysis, but exceeds the minimum 5 to 1 ratio of cases to every predictor (cf. Cooksey, 1996). The aim was not to generalize individual patterns of results to the general population, but to investigate the nature of the links between emotions and mechanisms for the specific individuals sampled.

Table 5 shows a summary of the results in terms of the individual multiple correlations and semi-partial correlations with each mechanism item (MecScale). Rather than performing significance tests on these small samples, we report effect sizes along with interpretations of these. All semi-partial correlations ≥ .20 are marked with an asterisk (*). This corresponds to the recommended minimum effect size representing a practically significant effect for social science data (RMPE), as proposed by Ferguson (2009).

Careful inspection of Table 5 suggests both some general tendencies, similar to those observed in the nomothetic analyses earlier, and some idiosyncratic relationships unique to individual listeners. For instance, note that happiness-elation was typically predicted by the Rhythmic entrainment mechanism, which met the RMPE criterion for 9 out of 10 listeners. However, for a subset of the listeners (40%), happiness-elation was also (and in some cases even more strongly) linked to Episodic memory; there were further some truly idiosyncratic emotion-mechanism links, for example that the ratings of happiness-elation were predicted mainly by Evaluative conditioning for Listener 1, but by Musical expectancy and Aesthetic judgment for Listener 7.

For the calm-contentment category, the Contagion mechanism was the most frequent strong predictor (as implied by the nomothetic regression model also), reaching the RMPE criterion for 60% of the listeners. Moreover, calm-contentment was linked to the Aesthetic judgment mechanism for three listeners. Note further that for Listener 6, calm-contentment was well predicted by the Visual imagery mechanism.

For the interest-expectancy category, the most frequent predictive relationships occurred for the Musical expectancy, Brain stem reflex, and—to a lesser extent—Rhythmic entrainment mechanisms. However, many different mechanisms were linked to this emotion depending on the listener (e.g., Evaluative conditioning; Contagion; Aesthetic judgment; see Table 5).

As regards nostalgia-longing, Episodic memory was (unsurprisingly) the mechanism most frequently implicated (for at least 60% of the listeners). Note, however, that Aesthetic judgment was also a good predictor of this emotion, for at least four of the listeners. Again, there were some idiosyncratic tendencies (e.g., links with Contagion for two listeners).

In view of the wide individual differences in emotion prevalence and mechanisms, we explored whether the differences were linked to various personality traits, as indexed by the BFI. Table 6 presents the Pearson correlations (r) between individual BFI test scores for the Big Five traits and individual prevalence of emotions and mechanisms, respectively.

Since these analyses were primarily exploratory in nature, we decided to refrain from performing significance tests and to just report the effect sizes along with interpretations of these. All correlations ≥ r = .20 (RMPE, Ferguson, 2009) are marked with an asterisk (*) in Table 6. Only 13 out of 85 correlations (15%) met this criterion. All of these effects can be categorized as “small” to “moderate” (Cohen, 1988).

As seen in Table 6, Neuroticism was positively correlated with prevalence of surprise-astonishment and the mechanisms Brain stem reflex, Evaluative conditioning, and Episodic memory. Extraversion was positively correlated with the mechanism Rhythmic entrainment. Agreeableness was negatively correlated with the emotions surprise-astonishment and pride-confidence. Conscientiousness was negatively correlated with seven of the emotions, as well as with the mechanisms Rhythmic entrainment and Visual imagery.

The results showed that prevalence varied markedly depending on the emotion category (a “large” effect; Ferguson, 2009), and replicated a number of previous findings. First of all, as hypothesized (H1), positively valenced emotions were significantly more common in reaction to music, than negatively valenced emotions (a “large” effect; Cohen, 1988; for similar results in previous studies, see Juslin et al., 2008, 2011; and and Sloboda et al., 2001). Second, the music induced both “basic” and “complex” emotions (see Gabrielsson, 2011; Juslin & Laukka, 2004; Juslin, Barradas et al., 2016; Sloboda, 1992). Third, certain emotion categories were significantly more frequent than others: happiness-elation, interest-expectancy, calm-contentment, and nostalgia-longing were the most prevalent emotions, whereas anxiety-fear, irritation-anger, and disgust-contempt were the least.

The main trends obtained in this study, which featured a random sample of music, were thus similar to those found in previous research, which featured random samples of situations (Juslin et al., 2008) and listeners (Juslin et al., 2011). This suggests some degree of stability of the prevalence estimates across different methods. Indeed, some recent cross-cultural findings indicate that the broad patterns of prevalence of emotions may even be similar across cultures (Juslin, Barradas et al., 2016; see also Cowen et al., 2020).

However, we also observed some differences; for instance, the emotion awe-admiration was more frequent here than in previous studies, which sampled music episodes more broadly in everyday contexts. This could reflect the particular circumstances of the present study (e.g., attentive listening, no concurrent activity, no social interaction), which may have benefited an “aesthetic attitude” towards the music (Juslin, 2013).

Follow-up analyses indicated that emotion prevalence varied to some extent depending on the broad genre category (i.e., the STOMP factor); for instance, the R&C category evoked more calm-contentment and awe-admiration than the I&R category. Furthermore, there were wide individual differences in prevalence of emotion, which should be investigated further in future research. Exploratory analyses of links to the Big Five traits revealed few correlations that could explain individual differences in emotion prevalence, except perhaps that listeners scoring high in Conscientiousness showed lower prevalence for all emotions.

Differences in mechanism prevalence were smaller, overall, than differences in emotion prevalence (a “moderate” effect) – suggesting that all mechanisms contributed to some extent. The most prevalent mechanisms were Rhythmic entrainment, Aesthetic judgment, Evaluative conditioning, Visual imagery, and Episodic memory. In particular, Rhythmic entrainment was significantly more prevalent than all other mechanisms in this study. This might partly reflect that popular music, as opposed to classical music, dominated the sample, and that the listeners did not choose the musical excerpts themselves. (Personal choice would perhaps have yielded a higher prevalence of episodic memories and evaluative conditioning; e.g., Juslin et al., 2008, 2011). As expected (H2), the two mechanisms Brain stem reflex and Musical expectancy were less frequent than the other mechanisms, except that Musical expectancy was not significantly different from Contagion.

Similarly to what was the case with the emotions, some mechanisms were more frequent here than in previous studies that sampled situations or listeners. For example, visual imagery was clearly more frequent here than in previous field studies (Juslin et al., 2008). Again, these differences may reflect the circumstances of the listening test (e.g., focused music listening in a dimly lit laboratory room with limited visual stimulation).

Mechanism prevalence also varied to some degree depending on the STOMP category (musical genre). For example, the R&C genre category produced lower ratings of Rhythmic entrainment and higher ratings of Musical expectancy, than the other genre categories. This could be because different genres provide different “affordances” in terms of the information needed for various mechanisms to be activated. In fact, it seems plausible that differences in prevalence of emotions between genres were largely mediated by differences in mechanism activation. Hence, it may not be a coincidence that, for example, the R&C category—which induced more calm-contentment than the I&R category—also displayed a higher prevalence of the Contagion mechanism, which was predictive of the same emotion in the (nomothetic) regression analyses (Figures 2 and 4).

We also obtained a few relatively weak links between the Big Five personality traits and particular mechanisms—for example between the trait Neuroticism and the mechanisms Brain stem reflex, Evaluative conditioning and Episodic memory; and between Extraversion and the mechanism Rhythmic entrainment.

We investigated the links between emotions and mechanisms by conducting regression analyses that aimed to predict emotions based on self-reported mechanism impressions in the form of the MecScale. As expected (H3), results indicated that listener’s emotional responses could be predicted quite well based on ratings of mechanisms. A mean multiple correlation of .847 meant that ca. 72% of the variance in the listeners’ felt emotions could be accounted for by the mechanisms in the BRECVEMA framework. This is remarkable considering that none of the excerpts were selected to “trigger” specific mechanisms or evoke specific emotions: we simply observed what emotions and mechanisms would occur spontaneously during listening to a representative sample of music. Yet emotions were systematically related to mechanisms—and in ways consistent with the theoretical framework (Juslin, 2019, Part III).

Each emotion category was associated with a few key predictors, which were largely as expected: For example, happiness-elation was predicted by Rhythmic entrainment; sadness-melancholy was predicted by Evaluative conditioning and Contagion; surprise-astonishment was predicted by Brain stem reflex and Musical expectancy; nostalgia-longing was predicted by Episodic memory; and awe-admiration was predicted by Aesthetic judgment. Conversely, interest-expectancy was associated with a wide range of mechanisms.

Some of the links may seem “counterintuitive” but could simply reflect that different mechanisms do not occur in isolation in “real” pieces of music—as opposed to in systematic experiments where mechanisms have been isolated through digital editing (e.g., Juslin et al., 2014). Music often includes information relevant for several mechanisms. For example, the finding that both Episodic memory and Aesthetic judgment contributed to our prediction of nostalgia-longing, which is per definition strongly linked mostly to memory, could reflect a spurious correlation with respect to the stimuli—say, that those pieces that were rated highly aesthetically also happened to be more familiar and thus evoked more memories in listeners. Similarly, the result that both Aesthetic judgment and Contagion contributed to prediction of awe-admiration could simply reflect that the pieces that were highly emotionally expressive and caused contagion also were perceived as very beautiful, and thus activated the Aesthetic judgment mechanism. (Several authors have considered possible overlap in aesthetic impact between expressivity and beauty; see e.g., Bicknell, 2009; Clynes, 1977; Juslin, 2013).

We also attempted to predict felt emotions based on a wide range of acoustic features, covering tempo, dynamics, timbre, harmony, rhythm, and structure. Models based on these features could indeed predict emotions to some extent, but the mean variance accounted for (39%) was quite modest. Hence, as expected (H4), the prediction of felt emotions was more accurate based on mechanisms (72%), than based on acoustic measures. Acoustically based models accounted for merely 54% of the variance in felt emotion that was accounted for by models based on mechanisms, on average. This test may be considered “conservative” since the acoustically based analyses included nearly twice as many predictors as the mechanism based analyses, which (all other things being equal) should favor the former type of model.

Moreover, previous studies have shown that the effects of acoustic features are mainly additive in nature (Juslin & Lindström, 2010), whereas it has been argued that the effects of mechanisms may be more interactive in nature (e.g., Juslin, 2019, Chapter 32). If so, the present comparison—which considered only additive effects—would arguably favor an acoustically based regression model.

It is noteworthy in this context that the emotion category that was best predicted by a mechanism-based model (nostalgia-longing) was worst predicted by an acoustically based model—perhaps because nostalgia is a prime example of where the underlying mechanism makes all the difference, as opposed to the acoustic patterns. For other emotion categories, this distinction may be more blurred.

Thus, for instance, the emotion calm-contentment was worst predicted by a mechanism-based model, though best predicted by an acoustically based model. This emotion is perhaps a case where purely acoustic patterns go a long way towards explaining induced emotions, as mediated by low-level mechanisms (though calm-contentment was also induced when music met listeners’ criteria for aesthetic value; Aesthetic judgment). Another possible explanation is that calm-contentment was evoked by slow, low-arousal rhythmic entrainment, but that the scale item for this mechanism was formulated in such a way that it tended to emphasize only the high-arousal aspects of entrainment (see Method). A mechanism-based model could then have failed to “catch” calm instances of entrainment, thus leading to worse prediction for this particular emotion.

Idiographic regression analyses of a subset of randomly selected listeners revealed that felt emotions could be predicted well based on mechanism ratings also at an individual level. Most importantly, results indicated as expected (H5) that the emotion-mechanism links were different for different listeners; that is, partly different mechanisms predicted their emotional reactions. We do not wish to suggest that the precise idiographic patterns can be generalized beyond the present cases; however, mere observation that the emotion-mechanism links vary in these cases means that we can already rule out the possibility that they are uniform across listeners in the general population.

Some of the findings from the nomothetic regression analyses may be re-interpreted in light of the idiographic data. For example, whereas a nomothetic model may seem to suggest that listeners’ experiences of happiness-elation were caused almost exclusively by Rhythmic entrainment, idiographic models showed that for 40% of the listeners, happiness-elation was also—and in some cases even more strongly—linked to Episodic memory. Such idiosyncratic tendencies are “hidden” in nomothetic models, which may prevent a full understanding of the emotion-induction process. Having said that, however, the individual differences found here in the idiographic models were still smaller than those observed with regard to the weighting of criteria in aesthetic judgments (Juslin et al., 2021; Juslin, Sakka et al, 2016).

There are several limitations that should be taken into consideration, when interpreting the results. First, most of the data analyzed in this study were based on self-reports. Note that participants only report what they can or are willing to report, and that their responses can be affected by social desirability and demand characteristics (e.g., Visser et al., 2000). However, the validity of self-report depends on the type of questions asked. Clearly, listeners should be able to report their feelings more or less accurately. It can, of course, be argued that listeners’ responses are constrained by the use of a list of emotion labels, but similar patterns of results to those reported here have been obtained in a previous study which measured the prevalence of emotions using an open-ended response format (Juslin et al., 2011, pp. 190-191).

As regards mechanisms, it can be difficult for listeners to monitor the precise processes. However, as explained above, the MecScale does not aim to measure mechanisms “directly.” Rather, it measures subjective impressions, which are reflective of mechanisms. Some of the mechanisms are straightforward to report because of their salience in conscious experiences; listeners will hardly be mistaken about a startle reflex, a conscious recollection of a memory, or a visual mental image.

More implicit mechanisms may at least be highly correlated with distinct impressions. Thus, the contagion item (Were you “touched” by the emotional expression of the music?) is based on the notion that the expression of the music will be salient in cases where emotional contagion occurs. (In Gabrielsson’s, 2011, study, “the emotional expression” was the single most frequently reported cause by the listeners.) One might perhaps be tempted to think that the relatively low prevalence of the expectancy mechanism is because this is a more implicit mechanism than, say, episodic memory, and that it is therefore less accurately reported. The workings of the expectancy mechanism may indeed be more subtle, and we are aware of the theoretical claim that the brain is constantly involved in prediction during listening to music (Vuust & Kringelbach, 2010). However, how often such computations actually induce a felt emotion is a different matter. The types of emotions one would theoretically expect from the expectancy mechanism (Huron, 2006; Meyer, 1956) are not very prevalent (Juslin, 2019).

The expectancy item in the MecScale is, essentially, an expectancy rating. We presume that a listener who becomes surprised via the musical expectancy mechanism will experience the music as “unpredictable.” (Studies in Cognitive Science, Empirical Aesthetics, and Music Cognition often use ratings of (un)expectness, and it is commonly assumed that those ratings are valid; Margulis, 2005.) In accordance with modern theories of expectancy in the emotion field, the expectancy item is based on the assumption that the strongest emotions are induced when a listener’s expectations are disrupted, rather than when they are confirmed (see Miceli & Castelfranchi, 2015), and mainly when these violations are quite large. It appears plausible that such events are detected by a listener. The MecScale item simply taps into the subjective impression associated with such events, and seems to do it rather well, based on our previous experimental studies, which manipulated musical expectancies with highly controlled stimuli (e.g., Juslin et el., 2014).

The high predictive validity of the MecScale items in the present study, and the way that mechanisms and emotions showed links consistent with the BRECVEMA framework, clearly supports the notion that listeners are able to report conscious impressions reflecting emotional processes. Even so, self-report indices of causal mechanisms need to be interpreted with some caution. (This is all the more so considering that MecScale is a preliminary instrument for the measurement of mechanisms, which may be subject to revisions based on empirical findings.) A multi-item version of MecScale is currently being developed, but in this study, we relied on one item per mechanism, to minimize participant burden in the long listening test. Reliability in terms of internal consistency (Cronbach’s alpha) is not a relevant criterion for single-item scales, and the notion of test-re-test reliability is problematic when a rated attribute does not remain constant over time. Both emotions and mechanisms are in constant flux. For instance, Gabrielsson’s (2011) study of strong experiences with music found plenty of cases where the emotional response to the same piece varied over time, sometimes quite drastically so (e.g., p. 72, p. 270), even just 15 minutes later (p. 340). In fact, results showed that the majority of the listeners had not had the same response when they encountered the same music again (p. 400).

This should not be surprising, since there are numerous variables in the listener and the situation that influence a response, and these may have a differential impact depending on the mechanism. Indeed, the BRECVEMA model predicts that the same piece of music can induce different emotions at different occasions; that two listeners may experience the same emotion though due to different mechanisms; and that two listeners may experience different emotions despite activation of the same mechanism (e.g., memories; Juslin, 2019, p. 382, p. 490; Sakka & Saarikallio, 2020). All of this conspires to make it difficult to estimate test-retest reliability for the MecScale items. However, since an unreliable measure cannot form a relationship that yields high predictive validity, a single-item scale can be regarded as sufficiently reliable if it shows high predictive validity (see Gorsuch & McFarland, 1972). High predictive validity for MecScale was observed in previous experimental studies featuring stimuli that were carefully manipulated and edited so as to isolate particular mechanisms (e.g., Juslin et al., 2014).

There is a popular notion in the psychometric literature that multiple-item measures are more valid than single-item measures, for all types of constructs (e.g., Nunnally & Bernstein, 1994), but this assumption has not been borne out by studies. The advantage of multiple-item measures applies mainly to abstract constructs that are multi-constituent—such as personality traits which feature various “facets.” If a rated attribute can be conceptualized as concrete and singular (such that it can be easily imagined), there is no real difference in predictive validity between single-item and multiple-item measures (e.g., Bergkvist & Rossiter, 2007). Although further research is clearly required on the measurement of various mechanisms, it is plausible that the mechanisms may to all intents and purposes be conceived of as concrete and singular (e.g., “did the music evoke a memory?”).

One obvious limitation of this study, however, is the modest sample of listeners, which calls for replication. However, the present study, which sampled 72 pieces of music randomly, should be considered in the context of previous studies involved in our “method triangulation” strategy which sampled 762 listeners (Juslin et al., 2011) or 573 listening situations randomly (Juslin et al., 2008). Although it would have been desirable to sample all three factors—music, listeners, situations—randomly in the same study, this was not practically feasible.

Still, the moderate sample (which was due to practical difficulties in recruiting a larger number of listeners) may perhaps go some way towards explaining the slightly disappointing results regarding the Big Five personality traits. However, links to personality have tended to be rather weak and mostly inconsistent also in previous studies, featuring considerably larger listener samples (see e.g., Juslin, et al., 2011; Juslin, Barradas et al., 2016). Indeed, another possible explanation of this absence of strong links between personality traits and the prevalence of emotions may be that personality traits primarily affect how listeners use music in everyday life (cf. Chamorro-Premuzic & Furnham, 2007; Vella & Mills, 2017). Such differences in use, which will shape which music listening events individuals seek out and get exposed to in daily life, will clearly have an effect on emotion prevalence in “real life.” By contrast, when such context effects are removed (like in this study where all listeners heard the same music in the same context), the effect of personality on emotion prevalence may not be apparent in the same way.

This line of reasoning highlights a final (important) limitation of this study: The lack of context. This is, of course, the inevitable price to be paid for achieving complete control over the musical stimuli, such that listeners heard the same music and could be directly compared. However, leaving out the context can give the impression that the task of predicting reactions to music is more straightforward than it really is. Several aspects of the context could help to determine which mechanisms are actually activated by influencing factors such as the music choice, the listener’s attention and the functions of the music during a specific activity (North & Hargreaves, 2008). Familiarity with the music, which was unfortunately not indexed in this study, may enhance a listener’s liking (North & Hargreaves, 1995), and also enables a greater number of memory-based mechanisms (e.g., evaluative conditioning, episodic memory) to be activated. In addition, extrinsic factors, such as the status of the performer and program notes, may also have effects on certain mechanisms (Kroger & Margulis, 2017). Thus, prediction of musical emotions could be more complex if contextual variables are taken into consideration. It is perhaps promising in this regard that contextual variables themselves appear to be able to predict felt emotions during music listening to some extent (Juslin et al., 2011, pp. 193-195).

Whatever the limitations of the present study, we argue that the results have three major implications for future research and application. First of all, they indicate that the mechanism Rhythmic entrainment plays an important role in emotional reactions to music—at least as far as popular music is concerned (see Juslin, Barradas et al., 2016, for similar findings in a cross-cultural sample). Although this mechanism has received some serious attention (Clayton, 2012; Trost et al., 2017), experimental evidence proving that rhythmic entrainment during music listening is able to induce emotions is still limited (Juslin, 2019) and mixed (cf. Bason & Celler, 1972; Mütze et al., 2020). Thus, more research is clearly needed to elucidate this seemingly crucial mechanism.

Second, the present findings provide further evidence that the mechanisms constitute the nexus where control over emotional responses is possible, and where music interventions need to be focused for maximum efficacy. Recent advances in psychological theories of how music achieves its emotional effects on listeners have not yet been translated into the clinical sphere, although researchers in the music and health domain increasingly call for mechanistic studies—as opposed to mere efficacy trials—because mechanistic studies may play a key role in optimizing interventions (Bradt, 2018; Juslin, 2011). Depending on the emotion one seeks to induce in a client, specific mechanisms may need to be selected and manipulated. We hope that the present study may set the stage for additional, more large-scale work examining how specific mechanisms might differentially influence listeners with particular personal histories.

Finally, our findings suggest that it is at the mechanistic level that individual differences will tend to occur. Though there are plenty of individual, demographic, and contextual factors that may affect emotions, the moderating effects of those factors are most likely “mediated” by mechanisms. (For instance, recent research suggests that the BRECVEMA mechanisms might play a role in explaining the effects of personality traits on emotional reactions; see Larwood & Dingle, 2021). Individual differences are increasingly recognized as a genuine challenge in the emotion field. Davidson (2012) refers to them as “the most salient characteristic” of emotions (p. 7). It may be argued that much of the apparent diversity in both musical interventions and psychological experiments aimed at emotional reactions to music is not due to the inadequacy of the emotion theories applied (if any), but rather due to the fact that these need to be applied on an individual level. Understanding how individual variability arises within the confines of general mechanisms may clearly benefit future applications. The present results suggest that a full understanding of how music evokes emotions needs to involve idiographic analyses. This reflects that although musical activities are social, through and through (Juslin, 2021), our emotional experiences of music are ultimately and irrevocably individual in nature.

Patrik N. Juslin, Department of Psychology, Uppsala University; Laura S. Sakka, Department of Psychology, Uppsala University; Gonçalo T. Barradas, Madeira Interactive Technologies Institute (M-ITI), University of Madeira; Olivier Lartillot, RITMO Centre for Interdisciplinary Studies in Rhythm, Time and Motion, University of Oslo.

This research was supported by a grant from the Swedish Research Council (Project No. 2004-2045), as well as by grants from the Research Council of Norway through its Centers of Excellence scheme (Project No. 262762) and the MIRAGE project (Grant No. 287152).