Data Science Everywhere

Reading Time: 10 minutes

The Top 2000 is a yearly all-time ranking of songs published by Dutch radio broadcaster Radio 2. The ranking gets determined by Dutch voters, who vote over the internet in the first week of December each year. It is a uniquely successful concept for over 20 years now, reaching over 11 million people, or more than 75% of the population and still breaking both listener and voter records almost every year. Because the list has existed for 2 decades now and is an all-time ranking, it is an ideal object to study what contributes to the popularity of a song or artist.

While the list is mostly stable, it has been noted that the passing of an artist has a major influence on the ranking of their songs. It turns out an artist on average gets 80% more votes in the year of passing than in the years before. Over the years, 63 artists who were already ranked in the Top 2000 passed away since the list was first published. While almost all artists get a boost when passing, not all do and even not of those who do all get an equal boost. Michael Jackson, Ramses Shaffy and George Michael all more than tripled their number of votes. On the other hand. Benny Neyman and Ray Charles did not see a noticeable boost. What separates the artists that have major boosts from those who had no boost at all?

Some explanations have been put forward, focusing on the timing of death, or in combination with the popularity of the artist and the work, or even adding spill-over effects from related artists passing away.

I set out to estimate the effects from the historical rankings of the Top 2000. We define a boost as the ratio of the number of votes in the first Top 2000 after the death of the song to the number of votes in the year just before:

   

To estimate what influences the size of the boost, I downloaded all the yearly rankings. Next, I scraped the wikipedia pages of all the artists, for information such as date and place of birth and death. Unfortunately, Radio 2 does not publish the number of votes, but I estimate these based on a model developed by Peter Meindertsma. Then, I calculate the boost by the equation above. I only calculate this for individual artists, ignoring bands, since I expect differents band member will have very different effects: compare a lead singer to a drummer. That might muddy our estimates. Finally, I develop a Bayesian hierarchical regression model which incorporates both the effect of artist and song characteristics to finally arrive at the parameters of the Top 2000 boost.

The following are the most important conclusions:

• An average boost for artist nowadays is approximately 80% extra votes
• This boost has been getting much stronger over the years: 60% over the 20 year period
• Popularity of the artist is the most important driver for the boost: more popular artists get larger boosts
• Artists who pass away shortly before voting starts can expect to have their boost doubled

---

To get a more comprehensive view of the effects, the boost of a song was modeled in 3 steps:

1. The model estimates a base effect for every artist
2. Next, the model estimates the deviation for a specific artist from the base effect, based on the characteristics of the artist and his/her death
3. Finally, the model corrects for how each song is different from the average song for this artist.

For each level, we formulate hypotheses for what influences each level. The hypotheses are shown in Table 1.

For a more detailed investigation into what affects the number of votes a song gets when the artist passes away, let’s look into two examples and compare why the model predicts a certain boost. All estimates I show below come from the model that was trained to predict the boosts from the full history of the Top 2000. We can investigate each aspect influencing the ranking of a song one by one, and multiply them to get a total prediction for a boost. We will compare Dutch hero André Hazes’s classic Eenzame Kerst, who passed away on September 23, 2004 to Under Pressure by Queen & David Bowie; the latter passed away on January 10, 2016.

The base effect

The model estimates every songs starts with a boost of 1.51, which means that a completely normal song will get 51% more votes in the year of passing away, than the year before. This is only the base effect; including other effects will (mostly) increase the expected boost.

The historical effect: boosts have grown stronger over the years

These boosts have grown stronger over the years, by an amount of approximately 3% per year. This means that the same artist passing away in 2020 would get an almost 60% more votes than an artist passing away in 2000. Presumably, this has everything to do with the way news is being evermore ubiquitous. We see this difference clearly too for André Hazes, who passed away in 2004 when the average artist did not even see a boost but a 10% decrease (but this may very much be compensated for by many other factors, which I will show below), while the average boost in 2016, the year of death for David Bowie was 33%.

The popularity effect: the more popular you were, the more extra votes you are going to get

When more popular artists die, they get much more news coverage: a relatively unknown artist may get a small article on page 15, while superstars may get entire TV shows devoted to hem. This in turn results in much more prominent boosts for more popular artists. We measure their popularity by the number of votes they got in the year before their death. In the cases at hand, both artist were much more popular than average. In fact, Bowie was ranked 25th in the list of artists with most Top 2000 votes in 2015 (the year before his death) out of 835 ranked artists, or top 3%, while Hazes was ranked 58 out of 975 in 2003. Notice that this effect can be very strong: Hazes gets a 50% boost, but Bowie, by being slightly more popular even gets a 70%+ boost

The home game: Dutch artists get a stronger boost

There appears to be some home-game effect; it is not highly significant, nor is it very strong, but still… Dutch artists appear to get a 10% stronger boost; which is were Hazes makes up a bit of ground, since Bowie will not get that

Artist tragedies: the younger you die, the stronger the boost

Newsworthiness does not only depend on the popularity of the artist, but also on the circumstances. Specifically, artists dying young is something that attracts a lot of media attention. Unfortunately, a lot of artists die young. In all the artists we are seeing, we find only 11% lives to the age of 80. We find that artists who pass away at a younger age get a stronger boost. This effect is very strong for artists such as Avicii and Amy Winehouse, who passed away aged 28 and 27 respectively: both gained hundreds of places for multiple songs. There appears to be a 1% stronger boost for every year before age 80. This was the case for both of the artists we are looking into: Bowie passed away days after his 69th birthday (leading to a 11% boost), while Hazes already passed at 53 (leading to a 28% stronger boost).

Short-term-memory: the difference between before the summer and after the summer

It turns out that it matters a lot whether the death is still top of mind when voters start voting, usually in the first week of december. The plot shows that passing away right before the closing votes can get 2.5 times more votes than passing away before the summer.

Bowie passed away in January, giving him no extra boost, but Hazes passed away in September. Not super close to the voting period, but still close enough to garner him slightly over 25% extra votes.

The artist magic: letting the model account for that special something

While there is a lot that is generic, such as the five effects that were just investigated, there is always that special something when it comes to art. The soap surrounding Michael Jackson’s death, for instance. We cannot model all these specifics per se, but we allow the model to account for it, if it has to. While we nudge it into the direction of the explainable effects, by allowing this we decrease the chance it finds spurious relationships and tries to model those through the real effects. This mainly has an effect for artists with many songs, where the group of songs can show it is not something specific to a song, but to an artist oeuvre as a whole. While we cannot use this artist specific deviations to predict future boosts, it can still be useful to make the predictions more robust. In our cases we find a small decrease for André Hazes (more on that later) and a negligible effect for Bowie.

So this concludes the artist boost. We see that Hazes was expected to get 2.13 times as many votes in the year of passing, while Bowie was expected to get 2.44 times as many votes. Interestingly, while the final numbers are quite close, the ways to get there were very different. Hazes dying younger, being Dutch and closer to the voting period, but these effects were offset by dying earlier and being less popular.

Song popularity: difficult to say which song is going to get a boost

So now that the effects for each artist are known, the final step is to differentiate between songs. First of all, it turns out that the effect of the popularity of this song as compared to the artist’s other songs is negligible:

Alone is better than together: duets get smaller boosts

On the other hand, there is an effect whether a song is a duet or a solo song. Solo songs by artists who passed away get more of a boost than songs in which they partnered with their colleagues. Some other songs by Bowie got much more of a boost than Under Pressure, for instance. It turns out that songs by multiple performers receive only 75% of the boost on average – which is finally why Hazes’s Eenzame Kerst is expected to get a larger boost than Under Pressure

The verdict: predictions match the actual boost very well

It turns out the model does a very good job in predicting these boosts. Eenzame Kerst was predicted to get 2.11 times as many votes, and got 2.12 time as many in practice – well within the margin of error. Under Pressure got 1.98 as many votes, and we predicted 1.87, also very close.

When looking at what we predict compared to the actual boost, we see the trend is picked up very well and almost all predictions are within the confidence interval.

There is only one but… This is not the full figure. There is one song which is completely different from all the other songs:

There is one song that was predicted to have approximately twice as many votes as the year before, but got 18 times as many! And that song was Zij Gelooft In Mij, also by André Hazes. The song was reissued in the year of his death and reached #1 in the Top 40 that year. It was sung at his publicly televised memorial in a sold-out Arena. In fact, this may very well explain why Hazes’s artist boost is 10% smaller than it was expected based on his characteristics: it appears those votes weren’t spread out evenly over his songs, but a lot of them went to Zij Gelooft in Mij.

---

No artist and no song is the same. Such is music. But what we can say, is that newsworthiness is very important to the ranking of an artist in the Top 2000, and that the circumstances of a death matter a lot. This tells us a lot about human psychology and the news media. But mostly, it tells us that something very tragic, can also lead to much more appreciation of all the beautiful music the artists have made.