Introduction

In Western cultures, major harmony and melodies are typically perceived or felt to be happy, while their minor versions are typically perceived or felt to be sad [6–13]. Although several possible origins have been proposed [10], there is no consensus over exactly how harmony and pitch affect emotive valence (which we define in this study as happiness and sadness) and to what extent these effects are culturally mediated or universal [2, 5, 14–20]. We apply the term universal to a process by which a musical feature influences affect and is not a result of familiarity (statistical learning of that feature’s prevalence in the participant’s listening history) or of association (statistical learning of that musical feature’s spatio-temporal proximity to affect-laden events in the participant’s listening history). Experiments conducted with Western participants have shown that emotions induced by psychoacoustic features related to the pitch content of the musical signal—such as roughness, harmonicity, spectral entropy, and mean pitch—are relevant not only for familiar musical examples, but also for unfamiliar microtonal stimuli [21–23]. Although this suggests that harmony and melody may have the capacity to communicate universally, mixed results have been obtained from prior experimental investigations in remote communities without easy access to mass media: Mafa, Cameroon [24]; Mebenzélé Pygmies, Congo [18]; Tsimane’, Bolivia [19]; Khow and Kalash, Pakistan [25, 26].

The first of these cross-cultural studies used short piano pieces composed to express the emotions ‘happy’, ‘sad’, and ‘scared’. These emotional categories were recognized by Mafa participants above chance, and tempo and mode (major/minor) correlated with their responses. However, the pieces covaried across a number of musical features and the independent effect of mode (e.g., after controlling for tempo) was not included in the analysis. The second study used orchestral classical and film music and found no similarities between pleasantness ratings made by Canadian and Mebenzélé participants, and no effect of major/minor on the latter group. In the third study, which presented intervals and chords, US participants’ pleasantness ratings were predicted by traditional Western music-theoretical characterizations of consonance or dissonance; the Tsimane’ participants’ ratings were not. The fourth study examined dimensional ratings of valence, energy and dominance as well as of four basic emotions (joy, anger, sadness and fear) of short Western and non-Western musical harmonizations in different styles with participants from the UK and from two remote communities in Pakistan (the Khow and Kalash): mode significantly impacted valence ratings of the UK participants, but not of the Kho and Kalash participants (as explained by Athanasopoulos et al. [25], the tribe is referred to as Khow, whereas the people are named Kho). A connected study (from the same authors and the same population) testing positive and negative affect of isolated musical chords showed a cross-cultural aversion for dissonant clusters but a varied preference for major triads [26]. Although only the first of these studies hints at the presence of a universal effect of major/minor on perceived emotion, it is notable that none of them has taken a Bayesian approach. This is important, as a lack of evidence for a particular effect does not mean that there is no effect. It simply means that there may not be sufficient data collected to detect an effect. Hence, traditional frequentist methods are mostly only able to reject, or fail to reject, a null hypothesis, Bayesian analyses are required to quantify evidence for the absence of an effect [27].

The experiment reported here was conducted across a number of remote and self-sufficient communities in Papua New Guinea (PNG). None of the communities has regular access to mass media. According to our analyses and local knowledge, traditional music is similar across these communities (see the S1 File). In each trial, their task was to choose which of two chord progressions, or which of two melodies, represented the greater feeling of happiness. Hence, like the Mafa study [24] and the Khow/Kalash study, the question asked of our PNG participants is about a basic emotion (happiness).

Despite years of research on the association between major/minor and basic emotions such as happiness, there is still no consensus on its origin. Here, we aim to further explore possible mechanisms of this association in communities with varied exposure to Western music. Furthermore, by taking a Bayesian approach it is possible to assess evidence in favour of each effect as well as evidence for its practical absence [27].

Some of the chord progressions and melodies were in a major key, some in a minor key (the quantifications of major and minor are elaborated later). In every community, except the one with minimal exposure to Western-like music, we find strong evidence that major harmony induces greater happiness than minor, although less decisively than for Sydney participants. For melodies there is a very strongly evidenced effect of major and minor in one of the PNG communities, and in Sydney. The estimates are not sufficiently certain, however, to rule out a universal effect of major and minor in either cadences or melodies.

We can theorize at least two broad classes of culture-dependent mechanisms that potentially mediate the effect of musical features (such as major versus minor) on emotion [28, 29]. The first is familiarity. Stimuli that an individual has heard many times before—notably, over a lifetime of experience—are often preferred and can signify positive valence, perhaps because familiar events have greater perceptual fluency so take up fewer cognitive resources. Clearly, this is a culture-dependent mechanism because the musical sounds common in one culture may be uncommon in another (although due to cultural globalization, musical experiences are becoming rapidly homogenized hence the urgency for experiments such as reported here) [30].

The second is associative conditioning. Given consistent spatio-temporal pairings between musical features and valenced events, those musical features may become imbued with the associated valence [17]. For example, a person familiar with movies using Western music will more frequently hear major harmony in positively valenced scenes and minor harmony in negatively valenced scenes than the other way round.

An example of a musical feature that may influence emotion via a culture-independent mechanism is the mean pitch of the tones in a piece (which would differ between major and minor versions). Mean pitch has been found to influence emotional responses to music [10, 12, 22, 31–33], often showing a positive relationship with valence [9, 22, 33, 34]. This could be due to an innate bias—in both animal sounds and human vocalizations, high pitches are often associated with friendliness or submission, whereas low pitches are related to threat or aggression [35, 36]. Although pitch range of vocal expression in mammals is mostly associated with affective arousal rather than valence [37], subtle pitch height differences between musical stimuli have been associated with changes in perceived valence in human listeners. In two recent studies [22, 33], the association between pitch height and valence has been found for Westerners listening to chords from an unfamiliar microtonal (Bohlen-Pierce) [38] tuning system, which is suggestive that pitch height’s effect is independent of a listener’s musical culture [22]. However, these findings are not sufficient to establish whether this effect is truly universal. It may simply be that a culturally learned association between pitch height and valence is carried over from the familiar musical system to the unfamiliar system. Additionally, it must be noted that the pitch height differences between musical stimuli stimuli may be subtle compared to those in vocal expression.

Similar arguments hold for the emotional implications of major and minor harmony, which differ in psychoacoustic properties such as harmonicity, spectral entropy, and roughness. All else being equal, major chords have higher harmonicity and lower spectral entropy than minor chords, which should support the former’s perceptual fluency because, regardless of their familiarity, they are intrinsically simpler [39, 40]. Hence, these are possible routes for a culture-independent effect of major/minor on valence. Recent studies have demonstrated that each of these three psychoacoustic features predict the perceived valence of chords in an unfamiliar tuning system [22, 23]; but, as before, this may just be a carry-over from culturally learned associations between major and minor chords and valence.

In this study, we focus on how self-reported valence of harmony is affected by cadence type (major versus minor) and mean pitch, and how self-reported valence of melodies is affected by mean pitch. To ascertain whether any of these features are mediated through universal mechanisms, we conducted an identical experiment in five remote PNG communities (with participants from seven different villages) with limited but differential experience of Western music, and in two Sydney cohorts with considerable, but still differing, levels of experience with Western or Western-like music. (Our definitions for Western and Western-like music, major and minor, and other musicological terms are provided in the S1 File.)

The first participant group completing the experiment comprised inhabitants of Towet village, in the Uruwa River Valley, Saruwaged Mountains, Morobe Province, PNG, and were tested during a three-week field trip to the area. After the research team left, local research assistants hiked to the villages of Mup, Mitmit, Kotet, and Yawan, and repeated the experiment on participants from those villages as well as from Bembe and Worin.

The Uruwa River valley is a cloud forest area accessible only by small plane or, for locals, an arduous three-day hike. Between the valley floor and the surrounding mountains, elevation ranges from sea level to peaks of 4,000 m. Villagers are expert farmers and lead self-sufficient lives, without mains electricity or an internal market economy. Mobile phone coverage reached the region in mid-2015, but internet access over the mobile phone network is nearly impossible. A dialect of the Nungon language is spoken in all the villages [41]. Additional information on the area—geography, maps, musical traditions, linguistic, and historical—is provided in the S1 File.

As summarized below, there are three genres of music present in different, sometimes overlapping, parts of the Uruwa valley: traditional non-Western-style song; Western-influenced stringben (which means ‘string band’ in Tok Pisin, the PNG lingua franca); and Western church hymns.

Across much of the region, traditional songs are performed at gatherings or specific occasions, such as a successful hunt or the blessing of crops, and are commonly accompanied with a uwing drum, which is a wooden hourglass-shaped drum with a single animal-skin head [42], or handmade flutes. Detailed analysis of six recordings of songs in a non-Western style is provided in the S1 File; in summary, the melodies are monophonic (sung solo or in unison) and rarely exceed a perfect fifth in range. There is typically a ‘focal’ pitch, which is sung more than any other. With the possible exception of a small whole tone (about 1.75 semitones), interval sizes are inconsistent between the songs, and are somewhat inconsistent within each performance. They do not typically conform with Western intervals. The aggregated pitch content of each song (e.g., overall pitch range, number of pitches, or mean pitch relative to focal pitch) does not correlate to the emotive content of lyrics or performative contexts, or the village of origin.

Due to missionary involvement, villages in the Uruwa area have adopted either the Lutheran or the Seventh-day Adventist (SDA) church, although the extent of involvement with either varies both within and between the villages. Bembe, Mitmit, Mup, and Worin residents are mostly Lutherans who attend church regularly and hear and perform hymns on Sundays. Based on local people’s reports, and literature describing the Lutheran movement in urban areas of Morobe Province, Lutheran church service hymns are a mixture of traditional and stringben sung melodies, often accompanied by guitar or ukulele in the stringben style. Stringben songs are mostly in a major key with generally major chords [43, 44]. Indeed, our analysis of Lutheran hymns recorded in Towet, which shows they comprise 92% major chords and 8% minor (detailed in the S1 File). Residents of these villages also probably engage in stringben and traditional music during the week.

People from Towet are mostly church-going SDA adherents who hear and perform hymns from the SDA Hymnal [45] at services on Friday nights and Saturdays, and in occasional morning and evening services during the week. The hymns are mostly sung in English, and the analysis in the S1 File shows that a substantial majority of their keys and chords are major (86% major chords, 12.5% minor, 1% diminished, 0.5% other), and there is no evident association between the valence of the words and the harmony (major/minor). There is very limited partaking in stringben or traditional music during the week (it is frowned on). Kotet and Yawan villages lack their own Lutheran churches (Kotet’s was demolished in 2011) and are too far from Worin or Mup to make regular church attendance practicable, but do have SDA churches. This means that stringben and traditional music are likely played or heard sporadically only, while SDA followers hear and sing SDA hymns weekly.

The above implies three groups of Uruwa participants with different levels or types of exposure to Western or Western-like music. The minimal exposure group comprises non-church-goers, and Lutherans in Kotet/Yawan—they have had only sporadic experience of Western-like music for at least seven years prior to the experiment. The Lutheran exposure group comprises all other Lutheran church-goers—they have regular exposure to major harmonies and melodies but less exposure to minor. The SDA exposure group comprises all SDA church-goers—they have regular exposure to major harmonies and melodies and, compared with the Lutherans, a slightly wider palette of Western harmonies; they have less regular exposure to traditional music than the other two groups. Importantly, none of the Uruwa participants have had regular exposure to conventional Western associations between musical features and emotion.

In contrast, the Sydney participants are all well-steeped in such associations, and have considerable exposure to Western music. Of the two Sydney groups (non-musicians and musicians, the latter having at least five years training or performance experience), the musicians generally have greater exposure to, and knowledge of, Western music and its cultural associations, and have more refined audition skills (for example pitch discrimination). Therefore, along with the three Uruwa groups, we have a total of five participant groups with differing levels and types of exposure to Western or Western-like music and its cultural embeddings. These five groups serve as the basis of our statistical analysis and its interpretation.

To briefly summarise the stimuli and procedure, every participant was presented with 12 different pairs of major and minor cadences (chord progressions traditionally used in Western music to unambiguously assert a major or minor key) and 30 different pairs of melodies in different modes (Phrygian, Æolian, Dorian, Mixolydian, Ionian, and Lydian). The cadences had a variety of tonic pitches, hence any effects of major versus minor and mean pitch can be assessed independently. The melodies all had the same ‘tonic’ (sounded by a drone), hence changes in their modes cannot be disambiguated from changes in their mean pitch.

For the Sydney participants, we expected major cadences to be more likely identified as happy when compared with minor cadences, and for cadences with higher mean pitch to be more likely chosen as the happy one (even after controlling for cadence types; for melodies, we expected results similar to those of Temperley and Tan [9]—the melodic subject with the higher mean pitch is more likely identified as the happy one, but unfamiliar modes (notably Lydian and Phrygian) are also less likely identified as the happy one. For the Uruwa River Valley participants, we anticipated the similarity of their responses to the Sydney participants would be positively associated with their familiarity with Western music. Based on these predictions, we tested the following hypotheses on self-reported happiness for each group of participants:

1. A positive effect of major versus minor for cadences only, adjusting for mean pitch.
2. A positive effect of an increase in mean pitch for cadences only, adjusting for major versus minor.
3. A positive effect of an increase in mean pitch for melodies only.
4. A positive effect of an increase in mean pitch on cadences and melodies combined.

Methods

Statistical modelling and data analysis

Mean pitch.

Mean pitch is calculated as the mean MIDI pitch of all pitches in the stimulus under consideration. For the melodies and cadences, the mean over all pitches in each stimulus in the pair is calculated. The difference in mean pitch—the predictor used in the models—is simply the mean pitch of the second stimulus minus the mean pitch of the first stimulus, and has semitone units.

Cadence type.

Here, we refer to the type of cadence, which is either major or minor (see the S1 File for our definition of major and minor). In music theory, this could also be referred to as the cadence’s ‘mode’ but, to avoid confusion with the melody modes, we use ‘type’ for the cadences.

Mode.

For melodies, ‘mode’ refers to one of the six different scales each melody was played in; here listed in ascending order of average pitch: Phrygian, Æolian, Dorian, Mixolydian, Ionian, and Lydian (Ionian is the common major scale, Æolian is the natural minor scale). To reduce the experiment’s duration, the rarely used Locrian mode was omitted (as in Temperley and Tan [9]).

Timbre.

Timbre generally refers to the perceived quality or ‘colour’ of a sound and has been suggested to have an independent and robust effect on listener’s perception of emotion in music [2, 16, 52, 53], which is not driven by musical expertise [54]. For this experiment, we used two timbres (a vocal and a bowed string timbre). Timbre is used as a moderator of the main predictors in this experiment: cadence type and mean pitch difference.

Models and coding. The data were analysed in R [55] using the brms package [56, 57], which is a front end for the Bayesian inference and Markov chain Monte Carlo (MCMC) sampler Stan [58]. Two types of Bayesian logistic models were run: descriptive models designed to summarize overall patterns in the data (visualized in Fig 1), and hypothesis-driven models designed to assess evidence both for and against the principal hypotheses (H1–H4) related to cadence type (major/minor) and mean pitch difference (see the Introduction), as well as their expected effect sizes (reported in Table 1, Fig 2, and the main text).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Visualizations of the effects of standard musicological categorizations across the five groups of participants.

For each pair of cadences, the effects produced by their cadence sequence and their tonics’ semitone pitch difference; e.g., a C minor cadence followed by a B♭ major cadence has the cadence sequence ‘m:M’ and a tonic pitch change of −2 semitones. For each pair of melodies, the effect of their mode sequence (the Phrygian mode has the lowest mean pitch followed by Æolian, followed by Dorian, and so forth, up to Lydian, which has the highest mean pitch). The bars show the descriptive models’ 95% credibility intervals.

https://doi.org/10.1371/journal.pone.0269597.g001

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Visualizations of the models’ principal effects across the five groups of participants.

For cadences: (a) the effect of their cadence sequence, adjusted for their mean pitch difference and timbre (the difference between the inner two posterior distributions corresponds to H1); (b) the effect of their mean pitch difference (semitones), adjusted for their cadence sequence and timbre (H2). For melodies: (c) the effect of their mean pitch difference (semitones), adjusted for timbre and melodic subject (H3). For cadences and melodies: (d) the effect of their mean pitch difference (semitones), adjusted for timbre (H4). For each cadence sequence plot, the posterior mean and 95% credibility intervals are shown. For each mean pitch difference plot, the posterior mean slope is shown with the thick line, while its uncertainty is visualized with 100 samples from the posterior distribution. The grey dots, which have been spatially jittered, show the observed values.

https://doi.org/10.1371/journal.pone.0269597.g002

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Hypothesis tests and summaries of the models’ principal effects across the five groups of participants (the hypotheses are detailed in the main text).

‘Est’, ‘Q5%’, and ‘Q95%’ show the mean and 90% equal-tailed credibility interval for the effect of Min to Maj compared to Maj to Min (H1) or a one-semitone increase in mean pitch (H2–H4). This effect is expressed on the logit (log-odds) scale. ‘Evid.ratio’ is the odds that the effect is in the direction specified by the hypothesis, and ‘Post.p’ is the associated posterior probability of the hypothesis (Evid.ratio = Post.p/(1 − Post.p)). We consider evidence ratios > 10 to be strong evidence; > 30 to be very strong. ‘ROPE’ shows the probability that the effect is small enough to be practically equivalent to zero [27], which we define as being in the interval [−0.18, 0.18].

https://doi.org/10.1371/journal.pone.0269597.t001

For all models, every factor was sum-coded (also called deviation-coded), so the reported effect of any predictor is a main effect; that is, the mean effect of that predictor over all other factor levels, such as timbre or melodic subject. In the hypothesis-driven models, the mean pitch differences between the two stimuli were in semitone units ranging from −2 to + 2 for cadences; and from −0.71 to + 0.71 for melodies, with respective standard deviations of 1.17 and 0.34. The reported effect of a unit increase in any predictor represents an additive change on the logit scale, which is the logistic-distributed latent scale assumed—in any logistic model—to underlie participants’ binary decisions.

Descriptive models. In summary, the descriptive models’ predictors are basic musicological descriptions of the stimuli: the relationships between the keys and modes of the two stimuli in each trial. In order to present the data in as ‘raw’ a form as possible (in a Bayesian context), no random effects were included and the priors were approximately uniform on the probability scale. The data were categorized using standard musicological descriptions with a comprehensible number of levels: cadences were categorized by their sequences (minor to minor, minor to major, major to minor, and major to major) and the change in the tonic pitch from the first to second cadence; melodies were categorized by their sequence of modes.

In both models, all other features such as timbre or melodic subject (both of which were balanced by design) were not included and hence averaged over. More formally, in the descriptive models, we have the following predictors:

• cad_mode1 ∈ {Min, Maj} (type of the first cadence)
• cad_mode2 ∈ {Min, Maj} (type of the second cadence)
• tonic_pitch_diff_cat ∈ {‘−2’, ‘−1’, ‘0’, ‘1’, ‘2’} (semitone pitch difference of the two cadences’ tonics, coded as categories)
• mel_mode1 ∈ {Phrygian, Aeolian, Dorian, Mixolydian, Ionian, Lydian} (mode of the first melody)
• mel_mode2 ∈ {Phrygian, Aeolian, Dorian, Mixolydian, Ionian, Lydian} (mode of the second melody)
• exposure_grp ∈ {Minimal, SDA, Lutheran, Non-musician, Musician}

We used three questionnaire responses—church attendance, village of residence, and church denomination—to assign participants to the three Uruwa exposure groups. As detailed in the Introduction, we can identify three major cohorts based on type and degree of exposure to Western or Western-like music in this region: those who attend SDA churches and hear or sing the SDA hymns (SDA group); those who attend Lutheran churches and hear or sing the stringben influenced hymns (Lutheran group); those who attend neither (minimal exposure group).

In the descriptive cadence model, there is a 4-way interaction between cad_mode1, cad_mode2, tonic_pitch_diff_cat, and exposure_grp. In the descriptive melody model, there is a 3-way interaction between mel_mode1, mel_mode2, and exposure_grp. We used a Student’s t-distributed prior with 7.763 degrees of freedom, a mean of 0, and a scale of 1.566 because this distribution on the logit scale approximates a uniform prior on the probability scale [59].

Hypothesis-driven models. For each of the hypothesis-driven models, two changes from the pre-registered versions (https://osf.io/qk4f9) were required. In the pre-registered models, there were high multicollinearities caused by interaction between the mean pitches of the two chords (or melodies), and model-fitting became unreliable. This interaction was simplified to a single term quantifying the difference in mean pitch between the two chords or melodies, which is actually the key quantity of interest and easier to interpret. Also, the data obtained from the questionnaire did not allow us to reliably distinguish between participants’ familiarity with Western or Western-like music and their familiarity with conventional Western associations between musical features and emotions (respectively termed intr_fam and extr_fam in the pre-registered models). In light of our experience with the region, the most sensible way to stratify exposure was on the basis of the questionnaire responses to church attendance, village of residence, and church denomination, as outlined above.

This led to the following stimulus-level predictors in the hypothesis-driven models:

• cad_mode1 ∈ {Min, Maj} (type of the first cadence)
• cad_mode2 ∈ {Min, Maj} (type of the second cadence)
• diff_mean_pitch1_2 ∈ (mean pitch difference between the two stimuli, in semitones)
• timbre ∈ {Strings, Voices}
• melody ∈ {‘1’, ‘2’ ‘3’} (the three melodic subjects—as notated in the S1 File—coded categorically).

In the cadence models, the principal predictors of interest are diff_mean_pitch1_2 and the interaction of cad_mode1 and cad_mode2. The cad_mode1:cad_mode2 interaction represents the four different sequences of cadence types: major to major, minor to minor, major to minor, and minor to major. For the effect of cadence type we were, therefore, particularly interested in the contrast between the latter two. Including mean pitch difference in the cadence model means that any modelled effects of cadence type arise from aspects not correlated with mean pitch, such as the previously-mentioned psychoacoustic features of roughness, harmonicity, and spectral entropy. These three predictors interact with timbre. In the melody models, there is a 3-way interaction between diff_mean_pitch1_2, timbre, and melody. In the cadence and melodies models, there is just an interaction between diff_mean_pitch1_2 and timbre.

In every hypothesis-driven model, all within-participant effects were modelled as varying (random) effects across participants. Weakly informative priors were used on all population-level effects: Student’s t-distribution with 3 degrees of freedom, a mean of 0, and a scale of 1. The use of varying effects and weakly informative priors reduces the probability of false positives and overfitting when data are noisy [60, 61].

A separate hypothesis-driven model was run for each of the five exposure groups rather than using a single model with an exposure group interaction; this allows for partial pooling of information between participants within each group, but not between groups (in an exposure interaction model, each group’s estimates would be drawn towards the overall mean due the varying effects and shared prior). Ensuring complete separation between groups is appropriate because one of the fundamental research questions is whether some of them (notably the minimal exposure group) may have very different responses to the others. The hypothesis-driven models’ formulas and full output summaries are provided in the S1 File.

Results

The descriptive model, detailed in Methods, allows overall patterns in the data to be visualized (Fig 1) with cadences represented on the left and melodies on the right. This figure shows the probability of choosing the second stimulus as the happy one, with the colours representing the pitch difference between the first and second stimulus in each trial. In order to test the differing effects of cadence type and mean pitch, we tested the four hypotheses introduced earlier. Bayesian multilevel logistic regression was used to assess evidence both for and against these hypotheses, and to estimate their expected effect sizes (summarized in Table 1 and Fig 2). We assess evidence using evidence ratios, which are probability ratios in favour of directional hypotheses [62]. These are subsequently detailed in the main text along with any strongly evidenced moderating effects of timbre. As detailed in the Methods, a unit increase in any predictor represents an additive change on the logit scale assumed to underlie participants’ binary decisions. In addition, we also report results from ‘ROPE’ (Region of Practical Equivalence) tests, which tests the probability of an effect being practically equivalent to zero [27].

H1—Cadence type: The effect of Maj to Min compared to Min to Maj, adjusting for mean pitch and timbre. There is very strong evidence for a large positive effect of major cadences on self-reported happiness in both Sydney groups and the SDA group, and strong evidence in the Lutheran group. For the minimal exposure group, there is no convincing evidence for an effect in either direction; nonetheless, there is only a 23% probability this effect is practically equivalent to zero. Hence, for this group, a convincing claim cannot be made either for or against an effect of cadence type.

H2—Cadence mean pitch: The effect of a 1-semitone increase in mean pitch, adjusting for cadence type and timbre. There is very strong evidence for a large positive effect of mean pitch on self-reported happiness in both Sydney groups and strong evidence for a small effect in the Lutheran group. For the SDA group, however, there is evidence (a probability of 0.90) for there being practically no effect, while in the minimal exposure group there is no convincing evidence either for or against an effect. Regarding the moderating effect of timbre, for the Lutheran group, there is a strong evidence ratio (10.70) that the effect of mean pitch difference is greater for strings than for vocal timbres.

H3—Melody mean pitch: The effect of a 1-semitone increase in mean pitch, adjusting for timbre and melodic subject. There is very strong evidence for a large positive effect of mean pitch for both Sydney groups and a medium effect for the Lutheran group. There is weak evidence in favour of a small effect in the SDA group but, for the minimal exposure group, there is neither evidence for an effect nor for its practical absence.

H4—Cadence and melody mean pitch: The effect of a 1-semitone increase in mean pitch, adjusting for timbre. Combining the cadences and melodies data allows us to assess the effects of mean pitch from a greater variety and number of musical stimuli (although there is no longer any adjustment for cadence type or melodic subject). There is very strong evidence for a large positive effect of mean pitch for both Sydney groups, as well as smaller effects in the Lutheran and SDA groups. For the minimal exposure group, there is no convincing evidence in favour of an effect in either direction.

Discussion

The results demonstrate that the self-reported emotional effects of major and minor—operationalized as cadence type and mean pitch—are strongly associated with exposure to Western or Western-like music. That said, for cadence type, we cannot rule out the possibility of a universal effect (resulting from the psychoacoustic properties discussed earlier) because the estimate from the minimal exposure group is insufficiently certain to confirm or deny. For mean pitch difference, the results in the minimal exposure and SDA groups point to only a small or practically non-existent effect. This suggests that the effect of mean pitch on valence—remarkably powerful in the West—is essentially cultural.

Given that none of the Uruwa groups are exposed to conventional (cultural) Western associations between musical features and non-musical emotive events in the Uruwa area, the effects of cadence type in those groups most likely arise from familiarity rather than from associative conditioning. Indeed, a supplementary analysis of twenty Lutheran hymns performed in the stringben style [44] by musicians from Towet (detailed in the S1 File) showed that all were in a major key, and minor chords were used for only 8% of the musical beats; all other chords were major. Supplementary analysis of harmony in the SDA Hymnal shows more variety than the Lutheran hymns, but it also comprises a preponderance of major chords (86%) over minor (12%) (detailed in the S1 File). This provides a ready explanation for the major cadences inducing more happiness than the minor—to participants in the SDA and Lutheran groups, they sounded more familiar.

The SDA hymns analyzed do not, on average, comprise more ascending sequences than descending: each piece almost always starts and ends with the tonic chord in a similar register, which implies the prevalence of upwards and downward motions of mean pitch must be approximately equal. Given the absence of associative conditioning, this could provide a natural explanation for why, in the SDA group, the effects for mean pitch are so small (and smaller than those for cadence type). However, it is not clear why this is not also the case for the Lutheran group, which exhibits a mean pitch effect more comparable to that of cadence type. A possible explanation would be that traditional music in this group differs from the others, but local knowledge suggests otherwise and our supplementary analysis shows no evidence that pitch difference signifies emotive valence in such music.

Across all of the stimuli, effects are stronger for the Sydney participants than for the Uruwa participants. This may be down to greater familiarity with Western music; it is likely also due to an additional influence of associative conditioning from conventional spatiotemporal associations of major with happy, and minor with sad—something we are sure is absent in Uruwa. The stronger effects for Sydney musicians compared to non-musicians may result from the former’s additional familiarity with musical structure and more firmly embedded associative conditioning, and also from them having better skills at discriminating the stimuli (some of which were quite similar) due to long-term explicit musical and ear training which is absent in the non-musicians as well as all the Uruwa groups.

The results should be interpreted with some caution, due to several limiting factors. First, participants may have a lack of familiarity with the notion of someone presenting stimuli multiple times and asking for a response or an opinion. This was encountered previously by Mulak et al., [46], who conducted a word learning study in the same area, and interpreted their null result as possibly reflecting a lack of familiarity with the task. Important to note is that Mulak et al., [46] concluded this based on a predominance of patterned responses, a type of response we already excluded from consideration in this study. Many research experiments are designed with a particular cultural influence and these designs may be interpreted differently in another culture [46, 63]. We attempted to design the experiment in such a way that it would be as meaningful as possible for the community members, but it remains a foreign task. If the opportunity arises to visit the area again, we would include measures to try and reduce the lack of familiarity as much as possible although it would be very difficult to design a task that is fully translatable across cultures. On a similar note, the notion of valence and the response options could be checked in a control task with clearly emotionally loaded stimuli (such as speech excerpts) to validate whether participants provide meaningful results; although even that precaution would not provide certainty the actual task is fully understood. Additionally, in order to obtain evidence for the absence of any meaningful effect in the minimal exposure group (or detect a rather small effect), we would need to collect more data.

Conclusion

In sum, we find no evidence for a universal effect of major harmony or melody inducing greater reported happiness than minor. These results are similar to the majority of cross-cultural music affect studies discussed in the Introduction [18, 19, 24–26]. That said, our Bayesian approach also allows us to assert a lack of firm evidence against such a universality, which would not have been possible had we used frequentist statistics. We find convincing evidence that the induction of positive affect by major, compared to minor, is enhanced by exposure (familiarity) and by associative conditioning: effects are close to decisive for Sydney musicians and very strong for Sydney non-musicians, both of whom have considerable exposure to conventional pairings of musical features and emotions; they are smaller in the Lutheran and SDA groups, which have some exposure to Western or Western-like music but not its conventional associations; and smaller still in the group with minimal exposure to Western-like music. The apparent inconsistency between our results for the minimal exposure group and those reported by Fritz et al. [24] may be due to the additional tempo cues provided in the latter, which are not present in the current study.

The results reported here show the importance of conducting cross-cultural studies in communities with differing levels of familiarity with Western music, and its cultural context; it is only through such studies that we can uncover the commonalities and varieties at the heart of music cognition.

Supporting information

Acknowledgments

We gratefully acknowledge Nathalyn Ögate, Namush Urung, and Ben Waum for their assistance with data recording and interviewing the participants in Papua New Guinea; Farrah Sa’adullah for her assistance with data collection in Sydney and musicological analysis of the SDA hymns; Felix Dobrowohl for musicological analysis of the stringben recordings from Towet; and Don Niles and Michael Webb for help with the ethnomusicology background; the audience at the Summer Institute of Linguistics Synergy Lecture, especially Andrew Grosh and Matthew Taylor, for providing feedback. Lyn Ögate, Stanly Girip, and James Jio were the highly efficient organizers of the Towet ‘research fair’ of which this experiment formed a part. Many thanks to everyone in the local communities involved with hosting us and making us feel welcome and safe. We would also like to acknowledge a few people who have played important roles in making this research trip possible: Alba Tuninetti, Martin Doppelt, and Paola Escudero from Western Sydney University, Georgia Kaipu from the National Research Institute (NRI) of Papua New Guinea, Olga Temple and Michael Esop from the University of Papua New Guinea.