Search This Blog

Monday, October 7, 2019

Puppies, Mandopop, and Causality

Here's the best puppy video that I've ever seen on YouTube. It's only 9 seconds. Trust me, it's worth watching.

It's even better looped. Try it with YouTube's loop feature. It's just gets better and better.

After I watched it a few times, I wondered what that catchy song was. It didn't take long for Shazam to identify it, but I struggled a bit getting the Chinese characters into a search field. I still don't know the name of the song, but it's recorded by the Mandopop* group By2.

Here are some comments for the song's video.

There's a clear causal claim here: the popularity of the puppy video caused the popularity of the song video. 

That sounds reasonable, but...    ... is it true?

Think about why the puppy video is so great. Sure, the rear dog's face plant into the snow is funny and the can-do spirit he shows by quickly jumping back up is inspirational. But I think it's the way the song perfectly syncs to the front dog's bounces that makes the video great.

At first, you think that the dog syncs to the song but that's not true. The song was added later. The dog does not sync to the song. The song syncs to the dog. That's an important distinction.

Compare the first version of the puppies to this one. It's longer, about a minute, but you need to watch only the first 10 to 20 seconds.

It's the same puppies doing the same thing but it's not nearly as cute. 

That should make you think about the causal claim in the music video's comments. Maybe the puppy video is not the cause of the song's popularity. Maybe they have it backwards. Maybe, just maybe, the song is the cause of the puppy video's popularity. 

Then again, maybe the comments are right. Either way, you can't confirm a causal relationship from a simple observational relationship**.

Unjustified causal conclusions are a common statistical error. When two variables show a relationship and "it makes sense" that A causes B***, we often jump to a causal conclusion without question. The jump to a causal conclusion is so common that I think we'd benefit from training ourselves to think the other way. Our first reaction to a statistical relationship should be to question any causal relationship. 

* Thanks to my colleague Chao Zheng for clarifying that the song is Mandarin instead of Cantonese.

** To be clear. I don't think the presence of two different sound tracks on the same puppy video is sufficient to claim there's an experiment. This is still observational. However, it's a good thought exercise to think about how you would make it an experiment.

*** Or, worse yet, we already believed the causal relationship and we were just looking for data to prove ourselves right.