For the past decades, classical music has experienced a decline in consumption, both in absolute terms and in relative terms compared to other musical genres. Industry insiders are generally split into two views of this decline’s cause: one arguing that stiff traditions in classical music exposure has created economic barriers to classical music appreciation,, and the other arguing that demographic shifts, and changes in values and taste—unrelated to income—are responsible for this decline,.
In this paper, I will investigate whether an increase in incomes can cause an increase in classical music performance attendance, i.e. whether economic progress can lead to breaking down economic barriers to classical music appreciation. Classical music performances attended would measure classical music exposure, and family incomes would measure income levels in this study.
Economists are also split on the causes of classical music decline. The neoclassical camp attempts to explain the decline of relative classical music consumption through rising economic barriers, and utility maximization of music listeners given budget constraints. Another explains the decline as a matter of changing tastes; consumers seek music that reflects current social values, and when those values are better reflected by popular music, demand for classical music decreases as consumers substitute towards popular music.
The division of consumption between socioeconomic classes was famously postulated by Thorstein Veblen in The Theory of the Leisure Class. Veblen theorized that in order to demonstrate “pecuniary prowess,” upmarket classes tended to limit their activities with economic and social barriers, in order for those activities to indicate their practitioner’s status. Under Veblen’s theory, the industry experts and economists who believe stiff socioeconomic barriers to classical music exposure cause classical music’s decline are correct; increased income should lower those at least some of these barriers, increasing exposure and attendance.
Roose and Stichele (2010) expand on Veblen’s basic theory, its manifestation in classical music, and income’s impact on classical music consumption, with Pierre Bourdieu’s theory of cultural capital from the famed empirical sociology book La Distinction. Bourdieu links socioeconomic positions with differing socio-cultural tastes and their barriers, which theoretically are created by socioeconomic groups to create in-group social cohesion and intergroup differentiation. Under Bourdieu’s theory, certain cultural tastes are built up by cultural capital, which is a barrier—often in economic cost—, and classical music could be one of those cultural tastes. While the reason for barriers is different from that presented in Veblen’s theory, the result is the same; there are socioeconomic barriers. Roose and Stichele classify classical music as “highbrow culture,” i.e. tastes that are more esoteric, which under Bourdieu’s theory, would require more cultural capital, in the forms such as musical education, all of which require money. Roose and Stichele classify some forms of classical music, such light opera, to be “middlebrow” music, as it does not require as much exposure—has less barriers—, or cultural capital, (e.g. concerts, lessons, classes, etc.) to acquire as a taste. For this study, light classical music will be included in the classical music category.
The Baumol effect in the classical music industry explains the higher economic barriers and decreased exposure to classical music. In 1966, while investigating this effect—or cost disease, the raise of wages in employments that have experienced no increase in productivity—Princeton University economics professors William Baumol and William G. Bowen found that there is a strong Baumol effect on classical music performances and lessons, e.g. four musicians are still required to play a Beethoven quartet, the same number of musicians as when Beethoven was still alive. Performed classical music has been drastically rising in cost; without an increase in productivity, classical music concerts and lessons greatly increased in real terms. This cost was reflected in Baumol and Bowen’s (1966) survey of classical music’s audience; the audience belonged to the mostly white and older professional and management classes with high income and educational levels, while demographics with lower incomes preferred pop music. While Baumol and Bowen were not explicitly interested in the audience’s socioeconomic makeup, their observations provide valuable evidence that income could cause greater classical music performance attendance. Later studies by Throsby and Withers (1985) empirically supported income’s role in classical music consumption decades later in other Anglosphere countries using probability regression, and economic demographics were also supported in Switzerland. Even in non-Western countries, such as Japan, these results held, with the professional and management classes gravitating towards classical music, while the Japanese working class preferring Japanese pop music. The same pattern holds for arts lessons, including music lessons.
From Veblen and Bourdieu, we have a theory for why classical music could be restricted to certain income groups as a means of social distinction. From Baumol and Bowen’s research, and subsequent studies, we know that exposure to classical music is restricted by economic barriers. Putting theory and evidence together, it should not be a surprise when Roose and Stichele found by running a logistic regression that, among the Flemish population in Belgium, higher participation in classical music performances could be caused by higher income.
This link between socioeconomic barriers and consumption reflects the neoclassical model where people who have tighter budget constraints might not derive as much utility from classical music performances, and forgo attending them.
However, some economists, and many industry insiders, attribute the decline of relative classical music consumption to changing sociodemographic identities causing a change of tastes in the market for music. Dolfsma argues that while the neoclassical model is correct that people try to maximize the utility of consumed goods, popular music derives more utility for all socioeconomic classes, decreasing demand for classical music. Dolfsma disagrees with the Veblen-esque idea that all people want to imitate the culture of the highest socioeconomic classes; he claims that during the 1950s and 1960s, people started imitating “lowbrow” culture more. By simply comparing descriptive data on social attitudes from Belgium and the Netherlands, Dolfsma saw that while those two countries had similar incomes and growth of income, Belgium was more traditional on social attitude scales; popular music took hold in the Netherlands relatively quickly, but it did not become popular in Belgium until social attitudes became less traditional. Dolfsma theorized that because popular music better reflects social values held by consumers of music, people found it more to their tastes and substituted away from not just classical music, but also jazz and blues. While the social identity effect discussed by Dolfsma likely dominates the effect of any income increases on classical music consumption, my paper only seeks to investigate whether income increase can increase classical music consumption in the form of performance attendances, holding other factors, such as sociodemographic shift described by Dolfsma, constant.
What makes this paper different from previous ones is that while some of the literature mentioned have simply correlated income to classical music consumption, they did not touch on the main focus of my investigation: whether an increase in income could cause an increase in classical music performance attendance. Those that did investigate such a causation did so in country’s with vastly different musical histories, like Roose and Stichele’s study using logistic regression. This paper investigates whether a change in income could affect consumption of classical music performances while holding time trends—an indicator of changing tastes alluded by Dolfsma—constant in the United States. While income does play a small role in changing consumption of classical music performances, I found that it is neither economically nor statistically significant.
This study investigates whether changes in income affect attendance of classical music performances. To do so, I will be using two data sets that include my explained variable of classical music performance attendance and my main explanatory variable of income, along with other necessary variables to hold constant.
The first data set is the National Endowment for the Art’s (NEA) Supplementary Materials related to the NEA’s 2012 Survey of Public Participation in the Arts (SPPA), with data spanning from 1982-2012. Data is collected every four to five years. This survey was attached to the Census Population Survey (CPS) for labor statistics, and randomly distributed to approximately half the individuals taking the CPS. Individuals surveyed at one point of time are not surveyed again—unless if they are randomly chosen to take the CPS and the SPPA again, which is highly unlikely—making this data set a pooled cross section. Respondents must also be over the age of 16 for their responses to be valid. The response rate of the survey was relatively high at 71.5%, but respondents were not required to respond to certain demographic questions outside core demographic questions such as age and location, and questions related to arts participation, making missing data a possible source of statistical error.
This data set also contains information needed to calculate my explained variable. To obtain the total number of classical music performances attended in the past year for each respondent, I summed the number of opera performances attended in the past year, number of ballet performances attended in the past year, number of classical music performances attended in the past year, which are all included in the data set for all observations; these categories for classical music performances correspond to ones found in Roose and Stichele’s study. My main explanatory variable of family income is also included for each observation in the data set; however, because respondents simply marked the range of family income instead of the exact family income (e.g. $50,000-$60,000), I averaged the income from the income ranges (e.g. $55,000 for a range of $50,00-$60,000) and assigned income averaged from the ranges as proxies for the actual income. Observations missing data for income was also dropped. The assumptions required for using income averaged from the ranges as proxy variables, implications of measurement error, and missing data will be discussed with greater detail in the Statistical Limitations section of this paper.
Other variables that may be relevant for my analysis include age, year of observation, ownership status of home (rent or own), highest level of education, employment status, highest education level of the respondent’s father and mother, hours worked per week, number of novels read in the past year, and marital status. To better fit my analysis, some of these variables had to be reencoded as well.
Since ownership status of home was recorded in three categories—owning a home or in the process of obtain ownership (i.e. paying off a mortgage), renting a home, or occupying the home without paying rent, I reencoded this information into a dummy variable which classified the first category as ownership and the other two as no ownership. Observations missing data on ownership were dropped.
The level of education completed for both the respondent and his/her father and mother were encoded numerically from 1 to 6, with 1 representing an education of less than high school level, 2 for some high school, 3 for high school graduate, 4 for some college, 5 for college graduate, and 6 for a postgraduate degree. For the purposes of my regression analysis, each of these categories will be encoded as a dummy variable; further explanation will be included in my Empirical Strategy section. Observations missing data for education of either the participant, or his/her father and mother, were dropped.
The data set encoded marital status into six categories: married, married but not living together, separated, widowed, divorced, and never married. I reencoded the data into a dummy variable with the first two categories being counted as married and the others as not married. Observations missing data on marital status were also dropped.
Other variables did not have to be reencoded, but observations missing data on those variables were dropped. Dropping variables missing data on number of hours worked in a week caused all observations recorded before 2002 and from respondents who were unemployed to be dropped; SPPA did not require respondents to enter the number of hours they spent looking for work or working without pay. Implications for dropping these observations will be discussed later in this paper.
Another possibly relevant variable not from the SPPA is the national endowment for arts per capita spending for that year; the per capita spending will be the same for all individuals of that year. I will create that variable using data from national endowments for the arts appropriations by year, made available through the NEA, and population data from the U.S. Census Bureau compiled by Multpl. To create that new variable, I will divide each year’s NEA appropriations by the population of that year. Then, merged my newly created variable into the SPPA data set described above.
The summary statistics for these variables are shown below: These statistics show a low number of classical music performances attended for most participants, an average income significantly higher than the national average household income due to unemployed respondents dropped from the data, high home ownership, and average age throughout the respondents.
Average education attained for the respondents (some college) is on average one level higher than that of their parents (high school graduate). The number of participants married is about one half, the average hours worked corresponds to a standard forty-hour work week.
The average number of novels read in the past year is about fifteen. The national arts endowment per capita remains a meager 43 cents on average.
While this data shows that number of classical music performances attended in general remains low, conditional means on our main explanatory variable of income paint a different picture
The trend of number of classical music performances attended conditional on income clearly shows higher average attendance as income increases.
A correlation matrix of the explained and explanatory variables shows a similar trend on classical music performance attendance (variables are abbreviated)
From these correlations, family income, education, parents’ education, age, and home ownership positively correlate with classical music performance attendance, which corresponds to previous research relating socioeconomic strata with classical music performance attendance. Hours worked having a positive correlation with classical music performance is also expected, as it increases income.
However, while marriage as a positive correlation with income, it actually has a negative correlation with classical music performance attendance. Likewise, a higher arts endowment per capita having a negative correlation with classical music performance attendance is also unexpected, but regression models later in this paper show a positive, albeit statistical insignificant trend.
The positive correlation between family income and classical music performances attended does support previous research, and other variables examined also merits inclusion, especially those with correlations supporting previous research and theory.
From the data, I propose the following econometric model and functional form for my regression analysis:
NumClassicalMusicPerformancesLastYearit. = b0 + b1log(income)it. + b2ageit. + b3age2it. + b4ownhomeit. + b5highschooli. + b6highschoolgradi + b7collegei + b8collegegradi + b9graddegreei + b10fatherhighschooli. + b11fatherhighschoolgradi. + b12fathercollegei. + b13fathercollegegradi. + b14fathergraddegreei. + b15motherhighschooli. + b16motherhighschoolgradi. + b17mothercollegei + b18mothercollegegradi. + b19mothergraddegreei. + b20hoursworkedit. + b21hoursworked2it. + b22book_numit. + b23arts_endowment_spendingt + b24marriedi. +b25ti. + uit..
Since the explained variable of interest is the number of classical music performances attended in the past year for each individual, NumClassicalMusicPerformancesLastYearit. what I will be regressing the other variables on. I did not include a logarithmic version of this variable because I do not expect the number of classical music performances attended to increase exponentially with my explanatory variable, and certainly not family income, my explanatory variable of interest.
My variable of interest, family income (log(income)it..) is positively correlated with my explained variable, so I expect its coefficient to be positive. Family income should affect the number of classical music performances attended because higher socioeconomic status should mean entrance to activities normally associated with higher-income circles, as postulated by Bourdieu. Roose and Stichele’s study also supports a positive coefficient; higher income groups have better access to other influencing factors such as a more “rounded” education that is able to provide music exposure, theoretically increasing the likelihood of attending classical music concerts. I would be using a logarithmic version of this variable because I expect the effect of family income on classical music to decrease as family income increases; a scatterplot shows a decreasing relationship as family income increases (see Figure 1).
For the coefficient of age and its squared version, I expect the effect to be positive at first—Baumol and Bowen observed that older people tended to attend classical music performances more often—but then negative, since older people may have barriers, such as decreased mobility, preventing them from going to venues where these performances are held; an extremely weak correlation with my explained variable and a scatterplot (see Figure 2) supports this observation. As such, I used a quadratic form for age. Age was included because I wanted to observe the effect of family income while controlling for age and prevent more bias in family income; age is positively correlated with classical music performances attended, and also with income, so it may positively bias my estimator for family income if left out.
Home ownership was included because owing a home, rather than paying rent, could indicate economic stability that reflects a permanence in income level; the coefficient for home ownership should be positive. Home ownership could also positively bias our estimator for family income because it is both correlated with family income, and positively affect the number of classical music performances attended because someone with a more permanent income is more likely to consume a normal good.
Education of the respondent, and his/her parents consists of six dummy variables each representing level of education obtained: below high school level, some high school, high school graduate, some college, college graduate, and postgraduate degree. In this regression model, the first category is omitted to prevent perfect collinearity. As such, each of the remaining five dummy variable’s coefficients should be interpreted as effects on attendance of classical music performances compared to someone who has less than a high school education.
The dummy variables for education are held constant because education is correlated with family income, and also positively affects attendance of classical music performances—people who are more highly educated are more likely to have exposure to music, which could increase their attendance of classical music performances—so its omission is likely to positively bias the estimator for family income. Likewise, the educational level obtained by the respondent’s parents is correlated to family income because children of more highly-educated parents are likely to go to better schools, which could positively affect personal development. More highly-educated parents would also be more likely to afford music lessons which could help a respondent like classical music more, as Roose and Stichele found in their study. I expect each educational level’s coefficient for both the respondent and his/her parents to be greater than that of the educational level below. Omitting any education dummy variable could positively bias the estimator for family income.
The number of hours worked per week is controlled because it should at first have a positive effect on classical music performances attended—more work hours means higher income—but then a negative effect as people no longer have the time to attend classical music performances. A scatterplot of hours worked per week and classical music performances attended supports an increasing then decreasing effect (see Figure 3); I decided to give this variable a quadratic form. The coefficient for the hours worked should be positive, but its squared term should be negative.
The number of novels read in the past year is included in the regression because its omission should negatively bias the estimator for family income. While books read is positively correlated with family income—having the leisure time to read novels suggests a certain level of income—I actually expect the number of books read to negatively affect attendance of classical music performances; novels and classical music performance should be substitutes and doing one activity could take time away from another. A scatterplot supports this observation (see Figure 4). As such, I expect this variable’s coefficient to be negative.
National arts endowment spending per capita should also positively bias the estimator for family income. My correlation matrix shows that nation arts endowment spending per capita is positively correlated with family income; this correlation makes sense intuitively because as family income increases, the taxes each individual contributes increases, which allows the national arts endowment to spend more.
Marital stability is also another variable that could negatively bias the family income estimator if omitted. This potential bias is negative because married couples have greater family income because two people usually make more money than one person, and greater stability in income could induce greater spending on normal goods like classical music performance. Marriage is slightly negatively correlated with attendance of classical music performances, a scatterplot shows a more positive correlation, and conditioned means show that married respondents attend less classical music performances, likely because they are more likely to have children, which limits their time to partake in such activities, resulting in an expected negative coefficient
Finally, I used a time trend to control for changing tastes in music. If Dolfsma’s findings that classical music may be declining due to changing tastes as time passes, this time trend should control for such effect. Because Dolfsma found that there is less appeal in classical music as time passes by, I expect the time trend’s coefficient to be negative. Since after omitting observations missing data for hours worked leaves us with only data from 2002, 2008, and 2012, there are only three time periods (1, 2, and 3) corresponding to these years. A scatterplot also supports a negative time trend (see Figure 5).
For this analysis, if the estimate of b1 is positive and significant, then we could reasonably expect a positive impact of income on number of classical music performances attended in the past year. With this data, we could support the findings and theories about how increased socioeconomic status increases consumption of classical music performances while controlling for factors such as changing tastes through a time trend.
Since there was heteroskedasticity present in this regression model, all standard errors calculated were heteroskedastic-robust standard errors.
After running the regression described in the previous section, I obtained the following coefficients, standard errors, R-squared, adjusted R-squared, and p-values (* indicates statistical significance at the 5% level)
The coefficient for family income is 0.0467, which means that for every dollar increase in family income, the number of classical music performances attended increased by 0.000467 percent. This result is not statistically significant at the 5% level, and it is hardly economically significant—for every $1,000 increase in yearly income, people tended on average to attend 0.46% more classical music concerts. Because family income is my main variable of interest, this result is disappointing, though it may be due to the effects of attenuation bias thanks to measurement error, as I will discuss later.
The coefficient for age is -0.028039, which means that for every year increase in age, the number of classical music performances attended decreases by 0.028. The coefficient for age2 is 0.0004, which means that for every year increase in age beyond the age of 35 (the quadratic form’s minimum), classical music performance attendance increases by 0.0004. While these results are statistically significant at the 5% level, they are hardly economically significant—a 100-year increase would only increase the number of classical music concerts attended by 0.04 after the age of 35. This result, while unexpected, still fits with theory since respondents establishing a career could have less time to attend performances, while as they become more established and older, they attend more performances.
The coefficient for home ownership is -0.11256, which means that if the respondent is a home owner, the number of classical music performances attended decreases by 0.1126. This result is statistically significant at the 5% level, but it is hardly economically significant—home ownership only results in 0.1126 less performances attended. This result actually contradicts theory, but since it is barely economically significant, it is not too relevant.
The coefficients for the five dummy variables representing the respondent’s level of education are -0.0067 (high school), -0.02 (high school graduate), 0.113 (some college), 0.435 (college graduate), and 1.23 (postgraduate degree). These coefficients mean that compared to respondents who did not receive a high school education, respondents with some high school education actually on average attend 0.0067 less performances, respondents with high school diplomas attend 0.02 less performances, respondents with some college education attend 0.1131 more concerts, respondents with a college degree attend 0.435 more concerts, and respondents holding a postgraduate degree attend 1.23 more performances on average in the past year. However, only the coefficients for some college, college graduate, and postgraduate degree are statistically significant. The coefficient for respondents with a postgraduate degree is very economically significant; postgraduates attend more than one more concert on average than those without a high school education, which is significant considering most respondents only attended one or two performances in the past year.
The coefficients for the five dummy variables representing the respondent’s father’s level of education are 0.015 (high school), 0.056 (high school graduate), 0.165 (some college), 0.248 (college graduate), and 0.418(postgraduate degree). These coefficients mean that compared to observations where the respondent’s father did not receive a high school education, respondents whose fathers received some high school education actually on average attend 0.015 more performances, attend 0.056 more performances when fathers have a high school diploma, attend 0.165 more concerts when fathers have some college education, attend 0.248 more concerts when fathers have a college degree, and attend 0.418 more performances when fathers have a postgraduate degree in the past year. However, only the coefficients for college graduate and postgraduate degree are statistically significant. The coefficient for respondents with a postgraduate degree is very economically significant; respondents with postgraduate fathers attend almost half a concert more on average than those without a high school education, which is significant considering most respondents only attended one or two performances in the past year.
The coefficients for the five dummy variables representing the respondent’s mother’s level of education are -0.028 (high school), 0.058 (high school graduate), 0.158 (some college), 0.160 (college graduate), and 0.482 (postgraduate degree). These coefficients mean that compared to observations where the respondent’s mother did not receive a high school education, respondents whose mothers received some high school education actually on average attend 0.028 less performances, attend 0.058 more performances when mothers have a high school diploma, attend 0.158 more concerts when mothers have some college education, attend 0.160 more concerts when mothers have a college degree, and attend 0.482 more performances in the past year when mothers have a postgraduate degree. However, only the coefficients for college graduate and postgraduate degree are statistically significant at the 5% level. The coefficient for respondents with a postgraduate degree is very economically significant; respondents with postgraduate mothers attend half a concert more on average than those without a high school education, which is significant considering most respondents only attended one or two performances in the past year.
The coefficient for hours worked is -0.01, which means that for every year increase in age, the number of classical music performances attended decreases by 0.01. The coefficient for hours worked2 is 0.0001, which means that for hour worked beyond 50 hours worked per week (the quadratic form’s minimum), classical music performance attendance increases by 0.0001. While these results are statistically significant at the 5% level, they are hardly economically significant—a 50 hour increase would only increase classical music performances attended by 0.005. This result is unexpected; however, it could indicate that as people work more hours, they do not have the income to attend more classical music performances until they hit a certain amount of hours.
The coefficient for novels read in the past year is 0.004, which means that for every novel the respondent has read, on average the number of classical music performances attended increases by 0.004. This result is statistically significant at the 5% level, but it is hardly economically significant—reading 100 more books only results in 0.4 more performances attended. This result actually contradicts the theory that books are a substitute to performances, but it is economically irrelevant.
The coefficient for national arts endowment per capita in the observation’s year is 0.38, which means that for every dollar endowment per capita increases, on average the number of classical music performances attended increases by 0.38. This result is not statistically significant at the 5% level, and it is not economically significant—since increasing the arts endowment per capita by a dollar would mean tripling the budget, a return of only 0.38 more performances attended is insignificant.
The coefficient for marriage is -0.18, which means th