Hacker News Reader

Comment by madengr | Hacker News Reader

Yep, sick of the "correlation is anti-causal". I assume correlation is causal (which is usually the case) until shown otherwise, which this article did not show.

View on HN

There's a website where the author correlates random datasets against each other and displays the ones with really high R values.

http://www.tylervigen.com/spurious-correlations

With today's headline driven media, it's especially important to guard yourself against correlational relationships being implied as causative. It's not about saying whether something is true or false, it's about being skeptical and using the correlation as a starting point for further investigation.

I think assuming correlation is causal is bad as well. But it's a good sniff test. Usually something interesting is going on: sometimes it's A causes B, sometimes its B causes A, sometimes A and B have a common related cause. Sometimes C is non-causal but correlates with A and B and wasn't controlled for. Lots of possibilities.

Correlation is not causation but it's rarely accidental.

Correlation is usually accidental.

e.g. http://www.tylervigen.com/spurious-correlations

Saying that correlation is usually accidental is a stretch. Basing that claim on that (admittedly amusing) website is even more of a stretch: the time series on that website are all highly autoregressive with minuscule sample sizes, which makes spurious correlations extremely likely [1].

Speaking (very) roughly, the fact that they are autoregressive constrains how 'kinky' the shapes can be. Visually speaking, the time series will have a small number of inflection points and will appear interpolated between these points.

This greatly reduces the search space: you just need to line up a couple of kinks in a space of ten samples, or match two smooth and vaguely line-shaped objects, rather than find a convincing relationship for 100s of observations not tied together in time.

[1] http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.611...

I've got some homeopathic remedies to sell you. People took them, and they got better!

I think if you measure random unrelated data sets against each other you're quite likely to find spurious correlations. If you design an experiment between potentially related variables with a potential hypothesis on how they are related and then observe data that matches the hypothesis, you probably have a reasonable chance the hypothesis is actually the explanation.

That is the point of experiments.

I see your blog post and raise you a youtube video.

minutephysics Correlation CAN Imply Causation! | Statistics Misconceptions

https://www.youtube.com/watch?v=HUti6vGctQM

Make a list of all the things that have changed over the last century.

My height has increased over the last century.

How many things on that first list are caused by my increasing height? How many caused my increasing height?

Things that have increased over the last century:

- Calories available

- Vaccination rates

- Access to neo-natal care

- Knowledge about fetus and infant development

Correlation is insufficient to prove causation but in many cases it's a great hint.

So as calories available increases, it causes my height to increase?

I'm gonna need a new wardrobe...

You've confused an individual trend with a trend of the aggregate maximum value.

A couple of those may have a minor influence on the limit of my height, but increasing those factors has no effect on my height.

Or how about my age?

And don't forget other highly correlated values, like the number of movies or books published, Chinese population, cumulative deaths in war, and number of artificial satellites. There's a correlation of my age (and height) with each of those.

Just two temporal trends that move in the same direction.

No causation, though.

It's unwise to infer causality (which I'm sure you do more carefully than you suggest here) but it's equally unwise to assume independence.

Quite a few of the correlations between random datasets linked by a few posters below, for example, may be explainable by second-order effects.

This is how we get homeopathy and alternative medicine.

Out of the sea of correlations, we find a few drops of causation by analysis and experiment.

Most criminals have drunk milk.

Most criminals have been to school.

Most criminals have worn sneakers.

Should we conclude that milk, school and sneakers are causal factors in crime? Should we eliminate these 3 things from society?

I don't think you understand correlation. For crime and milkdrinking to correlate, criminals would need to drink more (or less) milk than non-criminals. A better example would be: umbrella-carrying and windshield-wiping tend to correlate, but they are both actually caused by a hidden factor: rain.

Milk-drinking is not CORRELATED with criminality. Criminals and non-criminals have (as far as I know, but I could be wrong) similar levels of milk-drinking. And so forth.

Unless you are using a different definition of "correlated" than I expect.

Reply on news.ycombinator.com