Among music-likers, it’s often said that your musical tastes are defined by what we enjoyed at age 14, or that our favorite music came out when we were 14. This claim comes from a 2018 article in the New York Times titled “The Songs That Bind” (paywalled). This article contains dubious statistical analysis, and its claims are probably false.
The article uses Spotify data, “on how frequently every song is listened to by men and women of each particular age.” There are two distinct ways of analyzing this data:
- The person level – Look at each individual, and see which songs they listen to most.
- The song level – Look at each song, and see which individuals listen to them the most.
So let’s read the article carefully and determine which analysis was used.
I did a similar analysis with every song that topped the Billboard charts from 1960 to 2000. In particular, I measured how old their biggest fans today were when these songs first came out. […] Songs that came out decades earlier are now, on average, most popular among men who were 14 when they were first released.
The analysis described is at the song level. In contrast, the popular interpretation (“Our favorite music came out when we were 14”) assumes that the analysis is on the person level. The popular interpretation is not supported by the analysis.
In the very next paragraph, the NYT describes performing the same analysis for women, but switches to phrasing that implies the analysis was on the person level:
What about women? On average, their favorite songs came out when they were 13.
I believe the author does not recognize the distinction between the song level and person level, and is therefore incapable of accurately communicating his results. However, in my reading, it is most likely that the underlying analysis was on the song level, so the popular interpretation is still incorrect.
Why are the song level and person level distinct? Consider the following scenario.
Suppose that everybody starts out listening to music when they’re about age 14. They start out listening to the popular hits, the sort of thing that tops the Billboard charts. As they grow older, they develop more specific tastes, and the age group scatters into a hundred different smaller genres. On the person level, perhaps most people’s favorite music came out when they were age 30, although they may have nostalgic fondness for the popular hits of yore. However, on the song level, the Billboard-topping hits would be most popular among people who were age 14 when they were released. Age 14 is just when each age group was most unified in its musical tastes.
I don’t know if this scenario is true or not, but it’s a highly plausible interpretation that the NYT analysis fails to rule out.
In my research, I found a far superior analysis of Spotify data on the blog Skynet & Ebert. Their analysis found that as people get older, the artists they like tend to be more and more obscure. The authors argue that this is driven by two factors:
First, listeners discover less-familiar music genres that they didn’t hear on FM radio as early teens, from artists with a lower popularity rank. Second, listeners are returning to the music that was popular when they were coming of age — but which has since phased out of popularity.
There’s also an age where people start discovering less and less music. Skynet & Ebert suggests this is in the mid-30s. Another survey of people in the UK suggested it was around age 30. It’s not age 14. The NYT analysis is so garbled, I can’t tell what’s going on at age 14 at all.
“The Songs that Bind” is not a scholarly article, and would fail peer review if it were. The author should revise and resubmit prior to publication.
Maybe the poorly written article just has big error bars they forgot to mention?
For example, forgetting that it is one-tailed, if the imagined data were +/- 8 years, then the claim is just that music tastes are formed between ages 6 and 22, which is likely somewhat true and extremely meaningless. Well, the article doesn’t exclude this possibility. 😜
This study focuses on “music likers” who use Spotify, and it only focuses on the music they get from Spotify, when they may be getting other music from other media. That selection bias alone is grounds for dismissal.
I put some thought into this article and what it had to say. I am a music lover who never uses Spotify or Pandora. I grew up in the 1970s and 1980s listening to the regular-old-radio, where any given station would play a mix of rock, top 40, country, New Wave, stadium rock the Dr. Demento show on Saturday nights, and Weird Al Yankovic, some rap. And random phases of novelty stuff like “Hooked on Classics”, Sarah Brightman, Charlotte Church, and non-English songs like 99 Luftballoons and Major Tom in German, La Bamba in Spanish, etc. etc. There always seemed to be a choice of stations that played classic music as well.
I quit listening to the radio in the 2000s when it was one bleached-blonde pop-tart after another moaning off-key about how much they luvvvved Jebus and/or begging to be “hit” one more time or how they’re better than other girls because THEY wear t-shirts and sit on the bleachers. Whenever I check back in, it’s pretty much the same trash.
I expect Spotify users & usage are distorted relative to average music-listening behavior in the US. But to be clear, I’m accusing the NYT article of being completely off-base in its analysis, and therefore the choice of data set is a minor issue in comparison.
Though the article is paywalled, but you can see some of the figures from secondary articles, and that gives you a sense of the width of the distribution. But it’s not merely a matter of failing to report accurate error bars. The “correct” answer of 30-35 isn’t even close to the center of their distribution. They just aren’t showing the correct distribution to begin with.
Music-likers have been talking about this statistic for years, often attaching an anecdote about how they themselves don’t seem to fit the pattern. You don’t fit the pattern because it is false.
Sorry, got off-track with my comment and didn’t finish. Yes, the pattern is false. I do listen to music that appeals to me; some are newer talent, some are older. I enjoy several genres of music. My car has satellite radio, which offers a good variety of genres across the stations and some mixed genres on the same station, which works when you don’t want to use your own playlists.
Huh, I wouldn’t say the songs I listen to the most are my favorite songs… they’re the songs that are best for washing the dishes. (Or exercising, or other frequent tasks I like to do with music.) Using listening-frequency as a proxy for favorite music feels about as useful as using eating-frequency as a proxy for favorite foods – it might still tell us really interesting things, but not that one.