2

I'd like to know if it makes sense to use Spearman's correlation on my dataset. Below are plots between my 3 variables :

enter image description here

Correlation plot (Spearman): enter image description here

It seems it's pretty good but since I'm not from statistics I'm not sure. What do you think about it ?

Olstrom
  • 31
  • 4

1 Answers1

2

In general, Spearman checks for monotonicity, and so it can be used for discrete variables. Also, at least by eye, your top pair-plot does seem to indicate monotone decrease.

The problem, though, is that Spearman works by ranks, and so has a problem with ties. Your data seems to have many ties in the 5 values it obtains. In this case, you might want to use the Kendall tau-b test, which addresses this problem.

Ami Tavory
  • 4,410
  • 11
  • 20
  • 1
    What do you mean specifically by `Spearman has... problems with ties`? ([Kendall](https://stats.stackexchange.com/a/18136/3277) also has to manage "ties" in its own concept). What problems are these and how can they be managed. Could you expand on it or link to a discussion? Much thanks. – ttnphns Jul 07 '17 at 13:05
  • 1
    @ttnphns See, for example, [here](https://stackoverflow.com/questions/10711395/spearman-correlation-and-ties). [This link](https://math.stackexchange.com/questions/811642/how-to-explain-tie-correction-for-spearmans-rank-correlation) describes the Kendall correction for ties in Spearman. There are several others, but please LMK if I'm missing something. – Ami Tavory Jul 07 '17 at 13:15