This paper studies the positiveness in most consumed songs during a period of sadness, worry and uncertainty: the COVID-19 crisis in comparison with previous years. In order to perform the analysis, we create an original positive music indicator using a dataset of weekly patterns for music consumption in 31 OECD countries over a five year period. Results show that negative shocks from events such as economic business cycles and the COVID-19 pandemic increase the consumption of positive songs, suggesting that society compensates the negative impact on well-being with positive music. We conclude that society adapts sentiment-based music consumption preferences according to the socioeconomic situation.
Notes
This paper was awarded with the President’s Prize at the 22nd International Conference on Cultural Economics (ACEI2023, Bloomington, USA). We thank the editor (Stephanie Rossouw) and the anonymous referees for their comments and suggestions. Authors acknowledge financial support from Comunidad de Madrid and UAH (ref: EPU-INV/2020/006). Both authors contributed equally to this work.
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
1 Introduction
Economic and social developments affect the well-being of society. Country-specific pandemic severity had a negative social impact during the COVID-19 pandemic (Foa et al., 2022). The pandemic also caused an economic recession that increased unemployment and reduced consumption (WorldBank, 2020; Hodbod et al., 2021). During the COVID-19 crisis, the change in unemployment was 2 to 5 times greater than in a normal downturn (Bianchi et al., 2023). Unemployment and the pandemic produced negative feelings in society (van der Wielen & Barrios, 2021) and were associated with sadness, worry, anger, reducing Subjective Well-Being (SWB) (Helliwell et al., 2021). People try to compensate for these negative shocks, whether economic or social, by increasing consumption of specific goods (MacDonald & Dildar, 2020; Li et al., 2019). We apply this idea to the consumption of cultural services, using revealed preferences in music consumption. We consider that consumption increases for products and services that offer high emotional returns at lower prices.
Music consumption is an appropriate indicator of cultural consumption, as it is the most frequent and widespread cultural consumption activity. According to IFPI (2021),1 80% of music listeners said that music helped them manage their emotional well-being during the pandemic. Music is used to alleviate stress and anxiety and to improve well-being (Granot et al., 2021). Music streaming has increased over the years (Lee et al., 2020). In addition, music consumption is a low-cost consumption option that is widely used by society.
Advertisement
The response of preferences for certain characteristics of music to socioeconomic changes can be pro-cyclical or counter-cyclical. In the former case, the consumption of positive music would be lower during negative shocks. In an experimental analysis at the individual level, Lee et al. (2013) showed that interpersonal relationships and related events lead people to seek out mood-congruent songs. Edmans et al. (2022) show that weekly changes in music sentiment are positively correlated with stock market returns in the same week and negatively correlated with returns in the following week, consistent with sentiment-induced temporary mispricing. Both studies focus on contemporaneous and short time periods, signaling instant mood-congruent consumption. In the counter-cyclical case, society tries to find compensatory elements for the state of mind, so that negative shocks are compensated with positive stimulus, in this case positive songs, when a shock is more permanent. De Lucio and Palomeque (2023) prove the negative correlation between unemployment and positive music consumption in the US and Granot et al. (2021) conducted a survey where respondents answered that they increased their positive music consumption during the COVID-19 crisis.
We apply traditional consumption theory to estimate the extent to which the consumption of specific characteristics of music adapts to variations in socioeconomic conditions. The results help in understanding how consumers face adverse situations that last longer.
We build an original database using the most streamed songs from the multimedia services platform Spotify. We use its charts of the 200 most listened to songs for each week and country for the period between December 2016 and October 2021. Over this period, information is available for 31 different OECD countries. Using Natural Language Processing (NLP) techniques, we analyze the sentiment of the lyrics of each song and build a sentiment indicator based on the revealed preferences of music consumption. The indicator correlates with economic cycles (measured by unemployment and inflation rates) and the pandemic (COVID-19 related variables). This panel dataset allows us, first, to test for the impact of economic and social shocks on the consumption of positive music; second, to control for country-specific effects and thus obtain a cross-cultural effect; and third, to identify this effect using real and accurate consumption data.
We show that music consumption adapts to both economic and social circumstances. Our research shows that, when economic and health situations deteriorate, people tend to consume music with more positive lyrics. The results suggest societal self-regulation through a change in the consumption characteristics of cultural services as a result of the impact of economic and social shocks on welfare. Irrespective of the magnitude of the changes in the consumption of music features, the paper provides new evidence of common adaptive patterns of music consumption across countries using real consumption data, while simultaneously analyzing the effects of social and economic shocks. The results can contribute to a better design of cultural policies.
Advertisement
The structure of the article is as follows: Sect. 2 provides the theoretical framework; Sect.n 3 presents the new database; Sect. 4 presents the empirical results and robustness checks. The final section concludes.
2 Theoretical Framework: Positive Music Consumption During Crisis
2.1 Literature Review
Culture consumption is positively correlated with life satisfaction. For instance, Grossi et al. (2011) find a relationship between access to culture and psychological well-being. In terms of consumption of a cultural nature, music is the most prevalent and as such has been used to measure the effects culture provokes on society. Music is utilized for various well-being objectives, such as ’feeling better’, reducing stress, or increasing energy levels (Rickard, 2012). Gómez-Zapata et al. (2020) conclude that participating in musical education programs also improves well-being and quality of life. Hand (2017) reveals the relation between art attendance (including consumption of music events such as jazz performances, operas, ballet, musicals and other kinds of live music) and happiness. An example of how society uses music and other cultural activities is explained by Tajtakova et al. (2019), who describe how outdoor cultural consumption (concerts, cinema, theaters and others) increased during the economic crisis of 2007.
Business cycles affect well-being and, consequently, may affect music consumption preferences. Deaton (2011), Frey and Stutzer (2002), Wolfers (2003) and Mertens and Beblo (2015) indicate that unemployment and inflation are factors of unhappiness and a source of stress. Pettijohn et al. (2009) conclude that in more threatening social and economic times, there is a trend towards more meaningful, comforting, and romantic lyrics. de Lucio and Palomeque (2023) study the relationship between business cycle variables and the weekly consumption of positive and negative lyrics for the US in the Billboard Hot 100 over the period 1958-2019. They find that when macroeconomic conditions are negative, people tend to consume more positive songs.
The pandemic2 had an important impact on SWB (Foa et al., 2022), especially as far as country-specific conditions are concerned. In addition, from a consumption perspective, Takara and Takagi (2023) find that there is a home bias, that is, a preference for local music. From a production approach, local music is conditioned by digitalization advances (Bürkner & Lange, 2017), artists concentration (Borowiecki, 2013) and immigrant flow (Borowiecki & Graddy, 2021). These facts has encouraged us to make use of panel data in our cross-national research enabling us to control for country specific characteristics.
Other socioeconomic factors (other than pandemics and economic cycles) such as market concentration, war fatalities or the availability of substitutes such as television have also been proven to dictate the form and content of music (Crain & Tollison, 1997). Here we also study how technology has driven down the cost of distributing music so much so that few products or services are consumed more than songs in modern societies. The digitization process has induced more effects in terms of music consumption. Streaming services stimulate the discovery of new music, which increases cultural divergence (Datta et al., 2018). The personalized curation process identifies and adapts to consumers’ taste, increasing the variety of music consumed (Hracs & Webster, 2020). Big-data analysis of music consumption and social media followers is also expanding the market for artists, which now perform in non-habitual cities because they identify market potential there (Cho et al., 2018). It has a positive effect on fans too, who now attend more concerts because they know more artists and spend less money on albums (Nguyen et al., 2014).
Our research contributes to the literature in three different ways. First, we focus on both the economic and pandemic crisis effects on music consumption, whereas previous papers (de Lucio and Palomeque (2019) and Tajtakova et al. (2019)) only considered economic shocks. Second, we use a wide set of countries. Other studies only used data from one country (US in the case of de Lucio and Palomeque (2023) and Slovenia in Tajtakova et al. (2019)). Granot et al. (2021) use 11 different countries, which is still a small number compared with our 31 countries list. Third, the dataset includes absolute consumption, i.e., the number of streams, for each song. Using exact consumption information allows us to weight more appropriately the sentiment of the music listened to by society. de Lucio and Palomeque (2023) use data from Billboard, which only gives sorted rankings and Granot et al. (2021) use survey data.
In order to measure the sentiment enclosed in lyrics, we use Natural Language Processing (NLP) techniques. Although NLP tools are in constant evolution, we find several contributions in economics (measuring economic uncertainty (Charemza et al., 2022) or building statistic models for cryptocurrencies (Ortu et al., 2022)) and also in well-being (analyzing the impact of natural environments (Mertens & Beblo, 2015) or how the confinement affected depression and anxiety (Levanti et al., 2022)). Music consumption can reflect the mood of consumers. Therefore, labeling the moods expressed in lyrics can help to reflect the overall mood of society. This can be done by using the streams of the most consumed songs as an indicator. Using an extensive lyrics database, Palomeque and de Lucio (2021) demonstrate how musical characteristics and lyrics gather complementary information.
Considering all the aforementioned literature, we build our hypothesis around the three following ideas. First, both general macroeconomic conditions and the pandemic have an impact on well-being. Second, some types of consumption act counter-cyclically in a consumer effort to increase well-being. Third, music is a widespread cultural good with an emotional impact and its digital consumption is cheap or even free. Therefore, we consider that crises should show an increase in positive music consumption as a response to the decline in well-being. We present evidence supporting this hypothesis in the following sections. In the next subsection we present the theoretical model that explains this behavior.
2.2 The model for consumption of cultural features
We adapt a traditional consumer model to accommodate preference changes along business and social cycles. The central idea is that users consume culture in order to satisfy specific emotional needs under different socioeconomic scenarios. For example, on the one hand a consumer may seek motivation or joy. On the other hand, they may seek for protest, concentration or relaxation. The final consumer election involves a certain combination of different song features.
The simple model has two alternative options: positive consumption, Pos, and non-positive consumption, \(NonP_i\). In this latter set we include any other features, i, different from positiveness. These music characteristics cannot be consumed at the same moment in time, and the consumption of one characteristic must be accompanied by a simultaneous decrease in another.3
The consumer has a certain number of hours for lyrics consumption, H, and h is the time used to consume a song, considered the same for each song. The time, h, can be interpreted as the price paid for consumption. Accordingly, the constraint faced by the consumer is:
$$\begin{aligned} H = h*Pos + h*{\sum _i {NonP}_i} \end{aligned}$$
(1)
H/h is the number of features that can be consumed. The preferences for consumption of the different features depend on the economic (business cycle) and social (pandemic) conditions. Consumption preferences depend on psychological needs. The consumer’s decision is to allocate the amount of music consumed among the alternative options that provide utility to the listener. In a traditional Cobb-Douglas function the utility function is:
$$\begin{aligned} U = {Pos}^{\alpha } \prod _i {NonP_i^{\beta _i}} \end{aligned}$$
(2)
This utility function has all the desired properties, such as differentiability, monotonicity or convexity, of traditional consumer models. Furthermore, we assume that the utility function has constant returns to scale, \(\alpha +{\sum _i}{\beta _{i}} = 1\). Finally, we postulate that the parameters \(\alpha\), \(\beta _i\) depend on the economic and social situation that face the consumer.
$$\begin{aligned} \alpha \hspace{0.2cm} \& \hspace{0.2cm} \beta _i= f (business\hspace{0.2cm} cycle,\hspace{0.2cm} social \hspace{0.2cm}factors) \end{aligned}$$
(3)
For the \(\alpha\) parameter, which is associated with positive music consumption, we expect a soothing behavior, which is an increase in the consumption of this characteristic during the crisis, \(\alpha ^{expansion} < \alpha ^{recession}\). As the utility function has constant returns to scale, an increase in \(\alpha\) implies that \({\sum _i}{\beta _{i}}\) will decrease, \({\sum _i} {\beta _i}^{expansion} > \sum _i {\beta _i}^{recession}\).
To solve this problem, maximizing Eq. 2 subject to Eq. 1 allows us to identify the optimal consumption of each feature. The consumption of positive letters will be a constant H/h, multiplied by a parameter, \(\alpha\), that turns into a function depending on the economic cycle or social conditions.
$$\begin{aligned} Pos = {\frac{H}{h}\alpha } = \frac{H}{h} f (business\hspace{0.2cm} cycle,\hspace{0.2cm} social \hspace{0.2cm}factors) \end{aligned}$$
(4)
\(\alpha\) determines the share of positive songs. As this parameter varies over the business cycle, the share of positive songs will be higher during recessions than expansions.
The graphical representation of the model with two features, positive vs non-positive, is presented in Fig. 1. The model has two extreme possibilities, either dedicate consumption to listening to songs with positive lyrics, \(\alpha =1\) all the time, or alternatively allocate all the time to the consumption of songs with non-positive lyrics, \(\alpha =0\). In both cases the maximum amount of songs consumed will be H/h. Optimal consumption depends on the utility function and varies with the business cycle. Figure 1 depicts two utility functions and consequently two optimal solutions (consumption levels for positive and non-positive lyrics) in a peak (expansion) period and in a trough (recession). During the peak period, consumption of positive features is lower than during the trough. All the possibilities in the middle would represent continuous optimal solutions at different stages of the business or social cycle.
×
According to this simple theoretical specification, the percentage of cultural consumption of positive songs, feature Pos, is a function of social and economic cycles. Equation 4 provides the theoretical reference to the empirical specification of Eq. 5 and provides us with an estimation of the adaptation of music consumption to economic and social circumstances. Our variable for measuring the societal preferences for the positive features of music is the sentiment collected in the lyrics; we present this indicator in the next section. We call this indicator \(PLS_{i,t}\) (Positive Lyrics Sentiment). The independent variables will both be macroeconomic variables, \(Macro_{c,t}\), and the COVID-related ones, \(COVID_{c,t}\), for country c in year t. Therefore, the empirical equation is as follows:
The hypothesis presented in this document is that music is a tool that can be used for mood regulation, which means that positive lyrics consumption should increase during economic and social downturns. We expect a positive sign for \(\beta _1\): when worse macroeconomic conditions arise, \(Macro_{c,t}\) (e.g., unemployment), positive music consumption increases. We also foresee a positive sign for \(\beta _2\): when \(COVID_{c,t}\) increases (e.g., COVID related deaths), positive music consumption increases. In section (4) we econometrically analyze this relationship.
3 Data Description
In this section; first, we present the music sentiment database and second, we introduce the macroeconomic variables together with the pandemic dataset.
3.1 Positive Consumption: Songs’ Database
To build the music sentiment consumption database we use Spotify Charts.4 Spotify is the world’s most popular audio streaming subscription service, with 406 m users, including 180 m premium subscribers, across 184 markets. Since its arrival album sales have fallen (Adermon & Liang, 2014). Nevertheless, streaming income palliates selling losses (Aguiar & Waldfogel, 2018a). Also, streaming platforms have proven to be more effective in fighting against piracy than protectionist laws (Adermon & Liang, 2014). The importance of Spotify lies not only in the royalties it generates. Watson et al. (2022) provide evidence that social media follower numbers do not have a correlation with artists’ income, whilst an artist’s following on streaming platforms increases not only streaming but also live music income. Spotify has become such a relevant player within the music industry that record label technology teams focus most of their time on it (Watson & Leyshon, 2022). The prominent role of platforms facilitates measuring music consumption accurately.
We collect the top 200 most streamed songs for each week and country where Spotify operates. These charts are very important for two reasons; first, the launch or discovery of a song (Sim et al., 2022b) and second, the revenue that it can reach (Aguiar & Waldfogel, 2018b). Charts are available from December 2016, the final date being October 2021; that is, a total of 251 weeks (59 months). We build a dataset with information for some years before the pandemic, during all the lock-down phase and some of the following period. We collect information for 31 OECD countries.5
The weekly-country specific charts include the artist’s name, the track name, the number of streams, the position of the song in the chart and a link for listening to the song in Spotify. We build a dataset with the top 200 hits for each week and country, a total of 1,556,200 hits. This dataset contains 46,697 different songs with some being repeated across countries and weeks. We perform an automated search of the lyrics in Google, which uses different specialized websites as databases to collect lyrics.
We gather the lyrics for 37,405 unique songs, 75.5% of the total.6 The left hand panel of Fig. 2, exhibits the total sum of different songs gathered per week. The number of different types of songs per week grows from 1750 to 3000 between 2017 and 2020 and remains quite stable afterwards. The ratio between both lines remains relatively steady indicating that the percentage of songs gathered per week does not change dramatically during the time period analyzed. Figure 2 shows that the cultural divergence, measured by different songs that reach the top 200, grows in the first period, in line with Bello and Garcia (2021). The average number of countries in which a song reaches the charts was 4.5 at the beginning of the period and decreases to 2 countries after the COVID-19 pandemic.
If we consider that each song can appear more than once in the charts, for different countries or weeks, we collect 89.87% of the songs. We consider a selection of 1,398,524 songs from the 1,556,200 possible songs on the list to be a perfectly representative sample. As not all the songs receive the same number of streams, for the purpose of measuring sentiment it is important to calculate an indicator weighted by the number of times songs are listened to. In this manner, we grant more importance to those songs that are more successful than others.
×
Our dataset allows us to consider the number of times a song is listened to. The total number of streams is 450 billion, an average of about 58 million per country and week. We are able to collect the lyrics of 92.36% of total streams. The percentage gathered for each country can be found in the appendix [8]. For 23 of the countries, the percentage is over 90%, whilst the country with the lowest percentage is Iceland (69%). The right hand panel of Fig. 2 shows the total amount of streams for the charts and the streams collected per week (summing the streams from all the countries). The percentage, denoted by the ratio between both lines, remains very high and stable. The higher the percentage of streams collected in comparison with the percentage of songs indicates that we were able to collect more lyrics for the most frequently streamed songs, at the top of the chart.
We observe in Fig. 2, that digital streams had been continuously increasing until the pandemic lock-down, when consumption becomes more volatile, and people decide to make use of other cultural options such as Netflix. This result is in line with Sim et al. (2022a).
We are able to decompose the relationship between songs and streams. In Eq. 6, we transform the number of streams into three different components. The first component is the average number of streams per song in the list. The second component is the frequency of repetition of a song in the list.7 Finally, the third component is the number of different songs.
$$\begin{aligned} \text {Streams} = \underbrace{\frac{\text {Streams}}{\text {Songs} \, \text {in} \, \text {list}}}_{\text {Streams per song in list}} * \underbrace{\frac{\text {Songs in list}}{\text {Songs}}}_{\text {Repetition of song in list}} * \, Songs \end{aligned}$$
(6)
Using Eq. 6, Table 1 presents some descriptive statistics of the data set. Note that 2016 is not included in the table as the data starts in December 2016. The information for 2021 ends in October. First, we observe a rapid increase in the number of streams during the first part of the sample, whilst during the second half the streams stabilize. Second, the average number of streams per song in the list maintains a similar trend to total streams, increasing in the first part of the sample and stabilizing in the second one. Third, the frequency of repetition of a specific song in different countries or weeks decreases over the whole period. Finally, the number of different songs increases steadily during the period and stabilizes between 2020 and 2021.
Table 1
Relationship between Song and Stream.
Source: Own elaboration
2017
2018
2019
2020
\(2021^*\)
Streams (billions)
74
88
101
104
81
Streams per song in list (thousands)
259
307
348
356
350
Frequency of repetition of song in list
40
32
28
27
25
Different songs
7,120
8,871
10,457
11,053
9,084
* 2021 data ends in October. We have 41 weeks for this year
Top songs are recorded in different languages. English dominates the market, as we can see in the left hand panel of Fig. 3. We observe, also, that both Spanish (the second most listened to language), and other languages are increasing steadily over the whole period. Considering that streams have increased over the entire period, it appears that the increase in listening is mostly explained by the rise in non-English streams. We translate all the lyrics into English in order to harmonize and analyze homogeneously the sentiment of the songs. To test whether the language can bias the study, we calculate the percentage of positive songs per language: 63% English songs, 61% Spanish ones and 63% other language songs are positive, respectively. The percentages are similar, signifying that their different growth patterns is not a matter of concern. The variety of languages, number of different songs and frequency of repetition of a specific song all point to the cultural divergence process, which renders more relevant the use of a broad set of countries.
×
The right hand graph of Fig. 3 shows the distribution of the number of weeks that a song appears in the charts (minimum of 10 weeks). This graph does not consider the number of countries where the song has been listened to, that is, a song counts equally if it is played in one country or in all of them. The mean amount of weeks a song appears in the charts is 13.58. Nevertheless, the mode is to be in the charts only for 1 week, something that occurred in the case of 10,076 songs. A total of 10 songs were in the charts of at least one country, for the whole period, i.e., 251 weeks. The song "Shape of You," from "Ed Sheeran" is the song that appeared most frequently, with 4,643 appearances. It is also the second most streamed song during the period, second only to "Blinding Lights" from "The Weekend," which was listened to 1,986,197,450 times over the period.
After translating the lyrics into English, we perform a sentiment analysis in order to identify the characteristics of the music consumed. Palomeque and de Lucio (2021) show that the most popular sentiment analysis tools, such as PyTorch or TextBlob, provide correlated indicators for lyrics analysis.8 VADER (Valence Dictionary for Sentiment Reasoning) (Hutto & Gilbert, 2015) is a frequently used option in science,9 highlighting its use in the construction of an indicator of positive music consumption in the United States (de Lucio & Palomeque, 2023).
VADER uses a combination of qualitative and quantitative methods to construct a list of lexical features along with their associated sentiment intensity measures. Afterwards, VADER combines these features with some rules that embody grammatical and syntactical conventions for expressing sentiment intensity. Hutto and Gilbert (2015) consider that "this way of measuring sentiment improves the accuracy of the sentiment analysis engine across several domain contexts (social media text, ... movie reviews ...)" thus sharing linguistic expressions with popular lyrics.
We need to build an indicator for the consumption of positive songs. VADER offers three coefficients that range between 0 and 1: positiveness, negativity, and neutrality. The values for the three coefficients sum up to 1. Values close to 1 indicate that the text is mainly influenced by the sentiment of the coefficient, while values close to 0 indicate the opposite. We classify a song as positive (negative) if the positive coefficient from VADER is higher (lower) than its negative coefficient.10
The left hand panel of Fig. 4 shows the evolution of streams for positive and negative songs. Songs with positive lyrics are more frequent than those with negative lyrics, which is consistent with the other literature using Billboard charts (de Lucio & Palomeque, 2023). Although we observe a positive trend for positive song consumption, the trend for negative songs increases during the first half of the sample and stabilizes afterwards.
×
One of the novelties of this paper is that we make use of the exact amount of consumption for each song (that is, the number of streams). Our indicator for each week and country is the percentage of positive streams out of the total amount of streams during the week, Eq. 7. We normalize the result with the min-max method as shown in Eq. 8 (max value will be 100 and min value 0). We call this final version of the indicator "Weighted Positive Songs" (WPS) and it will be the main dependent variable in our regressions.
The right hand panel of Fig. 4 presents the average evolution for the indicator. We observe a higher average value of the positive indicator during 2018, with lower levels between mid-2018 and mid-2020 and an upward trend from the minimum level of the time series commencing in 2020. We also notice the presence of high peaks at the end of each year. These peaks are caused by Christmas carols, songs that usually present positive lyrics. We will control this seasonality with fixed effects.
Finally, we include Fig. 5, which maps the weighted positive songs consumption. Most countries are European because the Asian market is dominated by other platforms and the remaining OECD countries lack Spotify data to cover the entire period. We observe that all English-speaking countries (Canada, US, UK, Australia) are among the countries with less positive music consumption.
×
3.2 Macroeconomic and COVID-19 Variables
3.2.1 Macroeconomic Variables: Unemployment
The economic variable11 that impacts people well-being most is unemployment (Clark & Oswald, 1994; Ahn et al., 2004). Unemployment has a negative impact on both well-being and mental health. During recessions, the latter is affected not only by unemployment, but by job insecurity, increased workload and changes in job scope. Employed people also suffer the impact of recessions (Modrek & Cullen, 2013). Therefore, our hypothesis is that the consumption of songs with positive (negative) lyrics will rise (fall) with the rise of the unemployment rate, as a result of the efforts of citizens to mitigate the well-being effects of downturns through music consumption. Hence, we expect to observe a counter-cyclical response.
We collect the monthly unemployment rates of all countries from the Federal Reserve Economic Data (FRED).12 The left hand graph of Fig. 6 shows the evolution of the mean international unemployment rate over the period of study. We observe a decline in unemployment during the expansion previous to COVID-19 crisis, followed by a sudden increase with the onset of the pandemic and a recovery in the labor market by mid-2020.
×
3.2.2 Macroeconomic variables: Inflation rate
The literature suggests that, after unemployment, the inflation rate is the macroeconomic variable with the most negative impact on well-being (Blanchflower et al., 2014; Tella et al., 2001). Inflation brings uncertainty, causes income reallocation effects and erodes savings. The effect of inflation seems to be lower than the one caused by the unemployment rate, and therefore, we will only use it as a robustness check. The right hand graph of Fig. 6 shows the evolution of the mean inter-monthly inflation rate of all the countries during the period of analysis. All the inflation rates are measured as the changes in the Consumer Price Index of each country, with data from FRED (for Australia only quarterly data is available, so the regression including the inflation rate will have one country less). We observe an inter-monthly inflation rate lower than \(-\)0.2% that rises steeply to almost 0.8% (month over month) by the end of 2021.
3.2.3 COVID-related Variables
One of the contributions of this paper is to disentangle the economic and non-economic condition effects on positive music consumption with a view to analyzing a health driven socioeconomic crisis. Starting in March 2020, the pandemic produced an increase in deaths and a lockdown with a subsequent effect on mental well-being. This situation had a huge impact on our SWB, especially in terms of those experiences closer to us, such as those occurring in our countries of residence (Foa et al., 2022). Foa et al. (2022) show that of all the variables related to the pandemic crisis, deaths are what impacted the most on well-being. Hale et al. (2021) have built an index, called the "Stringency Index", which measures different aspects of the COVID-19 crisis. Although this index is widely used in the literature, they also built three other indexes (Containment Health Index, Government Response Index and Economic Support Index) which we will also use as a robustness check. We will focus on the "Stringency Index" as our main variable for measuring the effect of the coronavirus pandemic on society. The evolution of all the indexes is exhibited in the left hand graph of Fig. 7. Prior to 2020 the value of the index is 0, but by March 2020 all the indexes rise abruptly to a range between 50 and 70, although there are variations across countries. The economic support index remains at higher levels with less variation than the other indexes, decreasing during 2021.
For robustness purposes, we also use some additional variables to measure the impact of the pandemic: new COVID-19 cases, COVID-19 related deaths and ICU patients. All of these are measured per million of population. The right hand graph of Fig. 7 presents the evolution of these variables. Although the shape and the levels in both graphs are different, the beginning and end of COVID-19 waves can be identified.
×
Finally, Table 2 presents the basic descriptive statistics for the main variables. The percentage of positive songs ranges between 52% and 77%, with a mean of 64% and a low standard deviation. The weighted positive range 0-100 and the standard deviation is four times higher, standing at 12 instead of 3, the latter being the standard deviation for the percentage of positive songs. The mean unemployment rate is 7% and the mean monthly inflation rate is 0.2% . Finally, the stringency index varies from 0 to 91.
Table 2
Descriptive statistics
Source: Own elaboration
Mean
Std
Max
Min
Pos. %
64.2
3.2
76.6
51.8
Weigh. Pos.
28.3
12.3
100
0
Unemploy.
7,0
3.6
23.5
1.7
String. Ind.
19.0
28.3
90.9
0
\(^{*}\)p<0.1; \(^{**}\)p<0.05; \(^{***}\)p<0.01
4 Empirical Results
We run panel regression estimations using Eq. 5 and including fixed effects: country, \(\gamma _{c}\), and time, \(\gamma _{t}\), month of the year.13 As explained above, the main results will use the unemployment rate to measure the macroeconomic conditions and the stringency index for the COVID-19 impact on society. PLS consumption is measured in weighted positive songs terms. All variables are measured in a 0-100 range. We use Robust Least Squares (RLS) regressions for handling error terms issues. The final equation is as follows:
In this section we include two subsections; the first one presents the main results and the second section a set of robustness checks.
4.1 Music Consumption Stabilize Well-being Effects of Downturns
Table 3 exhibits the regression results for Eq. 9. The dependent variable is built using Eqs. 7 and 8. Columns (1) and (2) present the estimations for each explanatory variable (unemployment and the stringency index) independently and include both the time and country effects discussed previously. All the explanatory variables are significant and with the expected positive sign. The main finding is that, when the unemployment rate or the stringency index increases, people increase the consumption of positive songs, indicating a soothing behavior and confirming our principal hypothesis: during the pandemic crisis, people have tended to consume more positive songs, in order to mitigate more adverse economic and health conditions.
Additionally, the estimation in column (3), where the variables are included simultaneously, is not statistically different from those in columns (1) and (2). In this context, it is mandatory to control for endogeneity related to country-specific characteristics, both because of the relevance of the COVID-19 country-based impact as indicated by Foa et al. (2022) and because of the characteristics of culture consumption in each country, where differences become greater in line with the literature (Datta et al. (2018) or Bello and Garcia (2021)). We also consider it necessary to include the time effects of the month of the year, taking into account the effect that Christmas carols have on the indicator, as can be seen in Fig. 4.
Therefore, Column (3) is our final and reference regression as it incorporates all the features worthy of specific consideration. It is the one with the lowest Residual Standard Error and the highest Adjusted \(R^2\), the latter identical to Column (1). The values of Adjusted \(R^2\) are all higher than 0.94. These values are exceptionally high and give more credit to the validity of the results. We also observe that the effect of one standard variation of the stringency index has less effect than a standard variation on unemployment (1.1 vs 4.6 variation points in the WPS indicator).14 All of this confirms the main role of unemployment in sentiment-related music consumption.
Table 3
Main Results
Dependent variable: weighted positive songs
(1)
(2)
(3)
Unemployment
1.573\(^{***}\)
1.277\(^{***}\)
(0.251)
(0.254)
Stringency
0.053\(^{***}\)
0.039\(^{***}\)
(0.007)
(0.008)
Observations
1,829
1,829
1,829
Month FE
YES
YES
YES
Country FE
YES
YES
YES
R\(^{2}\)
0.945
0.944
0.946
Adjusted R\(^{2}\)
0.944
0.943
0.944
Residual Std. Error
7.650
7.772
7.616
\(^{*}\)p<0.1; \(^{**}\)p<0.05; \(^{***}\)p<0.01
4.2 Robustness
In this section we carry out four sets of robustness checks. First, we consider alternative COVID-19 variables. Second, we introduce new ways of measuring the dependent variable. Third, we perform similar tests for music characteristics, controlling for language and global hits. Finally, we analyze the different subsets.
4.2.1 Alternative COVID-19 Data
In order to contrast that the effect of the COVID-19 variable is robust to different measures of the pandemic, we test the results using some different measures for the incidence of COVID-19. Instead of using the stringency index we use all the variables presented in the Data Description section: Containment Health Index, Government Response Index and Economic Support Index, new COVID-19 cases, ICU patients and COVID-19 related deceases. Columns (1) to (6), respectively, in Table 4 present the results for these variables, all bearing the expected sign and proving highly significant. These results confirm the aforementioned conclusions: the pandemic motivated people to listen to more positive music.
Table 4
Other COVID-19 Variables
Dependent variable: weighted positive songs
COVID-19 variables
Containment
Government
Economic
COVID-19
ICU
COVID-19
Health Index
Response
Support
New cases
Patients
Deceases
(1)
(2)
(3)
(4)
(5)
(6)
Unemployment
1.179\(^{***}\)
1.175\(^{***}\)
1.211\(^{***}\)
1.335\(^{***}\)
1.489\(^{***}\)
1.490\(^{***}\)
(0.252)
(0.251)
(0.245)
(0.234)
(0.248)
(0.255)
COVID-19\(^{1}\)
0.045\(^{***}\)
0.044***
0.036***
0.011***
0.045***
0.219***
(0.008)
(0.008)
(0.007)
(0.002)
(0.015)
(0.072)
Observations
1,829
1,829
1,829
1,829
1,829
1,829
Month FE
YES
YES
YES
YES
YES
YES
Country FE
YES
YES
YES
YES
YES
YES
R\(^{2}\)
0.946
0.946
0.946
0.946
0.945
0.945
Adjusted R\(^{2}\)
0.944
0.944
0.944
0.945
0.944
0.944
Residual Std. Error
7.622
7.621
7.606
7.605
7.643
7.627
\(^{*}\)p<0.1; \(^{**}\)p<0.05; \(^{***}\)p<0.01
\(^{1}\)Notice that "COVID-19" represents a different independent variable in each regression. The variable name is stated at the header of the column
4.2.2 Alternative Measures of Positive Music Consumption
We now build four new indicators of positiveness consumption. Table 5 presents the results, which confirm our previous findings.
First, we use the "Positivity Percentage," column (1), which is a variable measuring the percentage of songs consumed in each month and country that can be considered as positive songs (this means, its positive coefficient is bigger than its negative one). This method does not weight songs according to the streaming data. The results remain as significant as before and with the expected signs.
The second column in Table 5, "Compound", uses a VADER’s continuous effect to measure sentiment. It takes values from -1 to 1, ranging from the most negative to most positive possibilities. This way we can use all the songs and not only positive ones, although we are unable to address the writing style issue. Having a continuous scale also gives us an interesting point for robustness checking. The compound value for each song is also weighted according to the streaming and the values are normalized in order to remain in a 0 to 100 scale afterwards. This check also reports the expected results for our hypothesis.
Column (3), "Valence", is an indicator built by Spotify and it measures how positive or negative a song is from 0 to 1, by taking into account their musical characteristics, such as the rhythm or the key, but not the lyrics. We consider a song positive if the value is over 0.5 and negative otherwise. We build our valence indicator in the same manner as the main dependent variable, weighting by the number of streams and normalizing. Results show that, in terms of music only, people also listen to more positive songs when unemployment or the impact of COVID-19 increases.
The last column in Table 5, "YouTube" (4), again represents an indicator built with the WPS method. The difference is that, in this case, we use the consumption rankings from the platform YouTube.15 YouTube has rankings for the top100 songs of each week for several countries. We source the rankings from 24 of the 31 countries in our study. These rankings started more recently than in Spotify: for the USA and UK, in September 2017; for the other 22 countries, in May 2018. We obtain the lyrics for 82,31% of the total streams, (see the country details in Appendix, Table 8). We also observe that at least 81,22% of the songs belonging to the Spotify charts are also in the YouTube charts (some others could be the same song but written in a different way, which does not allow us to make a match, so the percentage could actually be higher). Both platforms work similarly. Spotify has a longer period of study, more countries and double the songs per week, allowing us to select Spotify as the most representative main platform of music consumption for our paper. We see via YouTube, that unemployment and the COVID-19 crisis increase the consumption of positive songs, just as before. The unemployment rate is significant at the 5% level in this case.
Table 5
Alternative measures
Dependent variable:
Positive Percentage
Compound
Valence
YouTube
(1)
(2)
(3)
(4)
Unemployment
0.382\(^{***}\)
0.876\(^{***}\)
0.610\(^{***}\)
0.514\(^{**}\)
(0.090)
(0.323)
(0.168)
(0.257)
Stringency
0.024\(^{***}\)
0.043\(^{***}\)
0.047\(^{***}\)
0.065\(^{***}\)
(0.003)
(0.012)
(0.006)
(0.010)
Observations
1,829
1,829
1,829
1,024
Month FE
YES
YES
YES
YES
Country FE
YES
YES
YES
YES
R\(^{2}\)
0.998
0.916
0.957
0.929
Adjusted R\(^{2}\)
0.998
0.914
0.956
0.926
Residual Std. Error
2.842
12.063
6.286
7.445
\(^{*}\)p<0.1; \(^{**}\)p<0.05; \(^{***}\)p<0.01
4.2.3 Subsets
In this section, we address possible issues related to the dataset selection by picking some different subsets instead of the whole sample. All of the regressions use the standard WPS as the dependent variable, but we exclude some songs or countries for different reasons.
Columns (1) and (2) exhibit the results only taking into account the top10 and top100 songs of the charts respectively, instead of the top200. This way we are able to measure if socioeconomic conditions affect the most important hits or if the conditions effect in positive music consumption is more aggregated. Coefficients are equally significant and positive, although it is true that values are lower the shorter the rank list. The effect of an individual song that is being listened to for reasons different from socioeconomic conditions is higher the shorter the chart list.
Global hits, being those that appear in many countries charts at the same time, are probably more biased because of the presence of record labels, curators or due to the new releases of the most famous musicians. These songs are widely listened to across the world no matter the socioeconomic conditions. We drop those songs which have appeared on a specific week in more than 10 countries charts. This way we do not consider global hits, using only songs bearing a more regional nature, displaying a stronger relationship with the socioeconomic conditions of the countries where they are listened to. Column (3) of Table 6 presents the results, all variables maintaining the expected sign and level of significance.
Finally, Column (4) in Table 6 exhibits the results keeping only the countries where we gather more than 80% of the streamings as shown in Table 8, that is, without Iceland and Italy. Results remain in line with the principal ones already discussed.
Table 6
Subsets
Dependent variable: weighted positive songs
Top10
Top100
Under 10 Countries
Over 80%
(1)
(2)
(3)
(4)
Unemployment
0.146\(^{***}\)
0.577\(^{***}\)
1.166\(^{***}\)
1.150\(^{***}\)
(0.050)
(0.112)
(0.251)
(0.244)
Stringency
0.016\(^{***}\)
0.032\(^{***}\)
0.037\(^{***}\)
0.043\(^{***}\)
(0.002)
(0.004)
(0.007)
(0.008)
Observations
1,829
1,829
1,829
1,711
Month FE
YES
YES
YES
YES
Country FE
YES
YES
YES
YES
R\(^{2}\)
0.940
0.932
0.942
0.943
Adjusted R\(^{2}\)
0.938
0.930
0.941
0.942
Residual Std. Error
1.924
4.080
6.554
7.283
\(^{*}\)p<0.1; \(^{**}\)p<0.05; \(^{***}\)p<0.01
4.2.4 Other Robustness Checks
We conclude the Robustness section with two final checks. First, we include inter-monthly inflation rates in the regression, as they are relevant for well-being issues in the literature. We lose some observations due to the lack of Australian data, but results are the expected ones. Positive music consumption increases as inflation rises. We conclude that the macroeconomic conditions that provoke changes in positive music consumption preferences are not only rooted in the job market.
Column (2) of Table 7 is the last check of the present study. In this case, we have combined fixed effects, that is, we use a specific control for every combination of countries and months of the year. This way, we do not aggregate the effect of being in December, because not all countries have the same carols consumption. Results are positive and significant as in all other cases.
Table 7
Other robustness checks
Dependent variable: weighted positive songs
(1)
(2)
Unemployment
1.320\(^{***}\)
1.385\(^{***}\)
(0.252)
(0.239)
Stringency
0.035\(^{***}\)
0.037\(^{***}\)
(0.008)
(0.008)
Inflation
1.486\(^{***}\)
(0.474)
Observations
1,770
1,829
Month FE
YES
Combined
Country FE
YES
Combined
R\(^{2}\)
0.946
0.946
Adjusted R\(^{2}\)
0.945
0.932
Residual Std. Error
7.648
8.718
\(^{*}\)p<0.1; \(^{**}\)p<0.05; \(^{***}\)p<0.01
5 Conclusion
This work delves into the relationship between well-being, cultural consumption, and socioeconomic conditions. We show that society uses positive music consumption to alleviate the negative effect of economic recessions and social events, such as the COVID-19 pandemic. The proposed model and the empirical results suggest a change of consumer preferences along the economic and social cycles.
When unemployment (or the inflation rate) rises, people consume more positive music to counteract the negative effect of the economic downturn. This social reaction is also observed in the case of social adversities, like the COVID-19 crisis. We contribute to this literature in three ways. First, we use precise consumption figures. Our data set is built using the number of times a song has been heard. This is important because the top songs on the charts represent a large share of the total demand for songs. The actual consumption data supports the above-mentioned findings.
Secondly, we include a group of 31 countries, providing evidence showing that the compensatory effect is common for a wide set of countries worldwide. Furthermore, the present study provides preliminary evidence related to the current process of cultural divergence (increasing number of different songs, decreasing number of times a song appears on the charts and increasing the variety of languages). This divergence, coupled with the fact that previous literature suggests that the well-being of societies is more responsive to country-based factors, lends more importance to the spatial contribution of this research. Even though cultures seem to be drifting away, the consumption of culture encompassing positive features is still a common trait when adverse scenarios arise.
Third, our sentiment indicators are built using real data on music consumption. We test the impact of socioeconomic shocks in these indicators. The COVID-19 pandemic is the first major pandemic that we have experienced since music consumption data became available. The results show that people react similarly to different types of crises (economic or health related) in terms of the consumption of positive songs. Both effects remain significant even after accounting for each other.
Future research should analyze the potential similar effects of socioeconomic shocks on the consumption of other cultural activities such as cinema, literature, etc. And finally, these conclusions should motivate governments to facilitate access to the cultural sector, especially in times of crisis. Furthermore, the results should also encourage musicians and record labels to be mindful and produce more positive music during periods of adverse socioeconomic circumstances. No matter the nature of the crisis, positive songs will always come in handy to improve well-being.
Declarations
Conflict of interest
The authors declare that they have no conflict of interest.
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Music evolves more during upheavals or shocks at specific times when compared to time periods involving gradual changes (Mauch et al., 2015) Thus our study gains relevance by examining the COVID-19 period, which had a sudden impact on society.
Note that this formulation is equally valid if we consider more than two features consumed simultaneously, such as a song that inspires feelings of protest and energy. The model can also be adapted to explain different consumption patterns according to age, day of the week, season, and so on.
The countries are Austria, Australia, Belgium, Canada, Chile, Czechia, Denmark, Estonia, Finland, France, Germany, Greece, Hungary, Ireland, Iceland, Italy, Latvia, Lithuania, Luxembourg, Mexico, Netherlands, Norway, Poland, Portugal, Spain, Sweden, Switzerland, Slovakia, Turkey, United Kingdom and the United States. Most countries are European, with information available for the whole period. In this context, there is an absence of countries belonging to the Asian market, largely dominated by alternative platforms as well as some other OECD countries lacking Spotify data for the whole period.
The automated search of lyrics faces some difficulties when searching for songs. Sometimes the names of songs or artists are written differently in each database, sometimes they use symbols hard to reply by the bot. Finally not all lyrics are available in the databases used by Google. For all these reasons we finally gathered the lyrics for 75.5% of the songs.
Note that a song can be repeated in different countries or during different periods of time. For example: if the frequency of repetition of a song is equal to 30, it could mean that either a song has appeared in 30 countries for 1 week or for 30 weeks in just 1 country or potentially any intermediate combination.
We have performed the same test with our dataset and we obtained the same result: the three methods tested (VADER, TextBlob and PyTorch) are positive correlated with a 99% level of confidence.
Marketing (Barik and Misra (2024), Borg and Boldt (2020)), medicine (Roe et al. (2021), Mushtaq et al. (2022)), finance (Mardjo and Choksuchat (2022), Pano and Kashef (2020), Suardi et al. (2022), Öztürk and Bilgiç (2021), Katsafados et al. (2023), Long et al. (2022)).
We use this method because high or low values for positive or negative coefficients are sometimes caused by the writing style and not the sentiment expressed, according to the authors of the tool: https://github.com/cjhutto/vaderSentiment. A manual check of 100 songs made us choose this classification method from among the different options. We also use the compound coefficient, which takes into account both positive and negative features, as a robustness check.
The month-of-the-year fixed effect discards the possibility of time confounding factors and captures seasonality (e.g., increase in positive music consumption during Christmas). The decision of including a month-of-the-year fixed effect instead of time effects is based on the following two reasons: it avoids over-parameterization and avoids multicollinearity problems associated with the synchronized nature of the COVID-19 crisis.
One standard deviation of the stringency index (28.3, according to Table 2 increases the WPS indicator in 28.3*0.039 points. One standard deviation in the unemployment rate (3.6), in 3.6*1.277. We used the coefficients from Column (3) for these calculations.