News for hackers. Stuff that matters.
Re:Obesity and Eating Fast by WilliamLP on Thursday May 07, 2009 @10:16AM (#597945)
That's one of the weakest "correlation" graphs I've seen in a while.
Re:Obesity and Eating Fast by tokenadult on Thursday May 07, 2009 @10:20AM (#597949)
That's one of the weakest "correlation" graphs I've seen in a while.

See the Wikipedia article

http://en.wikipedia.org/wiki/Correlation

and especially its well done graphs

http://en.wikipedia.org/wiki/File:Correlation_examples.png

http://en.wikipedia.org/wiki/File:Anscombe.svg

(one based on a famous article by a statistician) for what correlation coefficients alone do and do not say about relationships among data points.

Re:Obesity and Eating Fast by longint on Thursday May 07, 2009 @10:43AM (#597983)
I know. Did anyone see the r^2 for the correlation?
Re:Obesity and Eating Fast by Shamiq on Thursday May 07, 2009 @12:30PM (#598186)
I was taught by an econometrician that r^2 values don't mean much, but I cannot remember his logic as to why. Maybe something about malleability?
Re:Obesity and Eating Fast by tel on Thursday May 07, 2009 @12:52PM (#598232)
That Anscombe example sheds light on the misuse of r^2. All of those examples have the same r^2 value.
Re:Obesity and Eating Fast by joeyo on Thursday May 07, 2009 @08:39PM (#598949)
Those are some very nice examples.

However, is there a better/more respected metric for comparing how two datasets are related? Mutual Information perhaps?

Re:Obesity and Eating Fast by tel on Thursday May 07, 2009 @11:29PM (#599114)
I'm not an expert -- more of a hobbyist who is actively trying to improve my statistical chops -- but I don't think there's ever a general metric for this sort of thing. For any statistic it's important to demonstrate that it's being used as intended so that its lay translation means what it seems to mean.

I mean, r^2 isn't all that misleading if your relation really is linear. R^2 plus a graph of the regression would be sufficient. Just taking a look at the residuals of the Anscombe sets easily shows that linear fitting (thus r^2) is inappropriate for sets 2-4.

(btw: if you fit the second Anscombe data set with a 2nd order polynomial, r^2 = 1. Of course, it's pretty clear that there's hardly any error in that set.)