When to log-transform?

#1
Should Iog-transform variables for correlation if they aren't normally distributed even if I'm using a non-parametric test for correlation? And if so, why?

Thanks for your help!
 

Karabiner

TS Contributor
#2
With "non-paramteric test for correlation" you possibly are referring to Spearman's rank correlation?
In that case, there are no normality issues. In addition, a log transformation would change nothing.
The rank correlation considers not the original values, but their rank. E.g. "income" wíth n=4 values
1000, 980, 777, 200 would be used as ranks 4 3 2 1 in the calculations. This ranking will not change if
instead you use ln(1000), ln(980), ln(777), ln(200).

With kind regards

Karabiner
 
#3
Thank you - that is very helpful! My two variables are continuous variables though. I've read somewhere that Spearman's rank correlation requires one of the variables to be ordinal? Is that true?
 

Karabiner

TS Contributor
#4
You can always "downgrade" a scale. The software which calcualtes rho will transform
interval scaled variables into rank variables.

With kind regards

Karabner