2
$\begingroup$

I see conflicting answers online from different statistic experts on this question. One one hand, stackexchange experts have note that Pearson Correlation does not require normal/gaussian distribution on a 10 years old post

On the other hand, Jason Brownleen on his machine learning blog says the opposite :

The correlation between two variables that each have a Gaussian distribution can be calculated using standard methods such as the Pearson’s correlation. This procedure cannot be used for data that does not have a Gaussian distribution. Instead, rank correlation methods must be used.

Who is right? how can i check other than see how many more people agree on one versus the other ?

$\endgroup$
2

1 Answer 1

2
$\begingroup$

$$\rho_{X,Y} = \frac{\text{Cov}(X,Y)}{\text{SD}(X) \text{SD}(Y)}.$$

This is for any $X$, $Y$, be they distributed Gaussian, Poisson, Slash, Gumbel, Wishart provided the moments defined above actually exist.

Whether or not this is true is not a democratic decision. The above is the definition you will find in any textbook or course.

You should write to the author and ask them to verify, or at least expound on the idea. I'm sure what the author meant to say is that the Pearson correlation is not a straightforward or meaningful summary if the data are not Gaussian (even then, a contentious point). Otherwise this is an interesting finding and the author should definitely submit their findings to JASA if it can survive a round of peer review!

Specifically, Jason says

It is easy to calculate and interpret when both variables have a well understood Gaussian distribution.

which is actually true, but then he continues:

When we do not know the distribution of the variables, we must use nonparametric rank correlation methods.

which is something worse than false, it's replacing careful thought and data analysis with a knee jerk reflex. A better answer would be,

When we do not know the distribution of the variables, we must carefully inspect scatter plots, and the shape of smoothed estimates of the mean response relating $X$ and $Y$, and the constant or heteroskedastic variance structure around the mean, considering a family of parametric or non-parametric summaries that might relate two variables. The Pearson correlation may in fact be a reasonable summary of the first order trend in the data, and confidence intervals and statistical tests may be well calibrated if the sample size is reasonably large, otherwise robust error estimates such as the jackknife may be employed to obtain summaries that are better calibrated.

But jeez who wants to read all that?

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.