10
$\begingroup$

It seems like for managing with ordered measurements researchers usually deal with Polychoric Correlation. (For example, for making matrix before doing Factor Analysis.) Why so?

Kendall Tau Rank Correlation Coefficient and Spearman's rank correlation coefficient are also suitable for ordered data.

Any points 'pro' and 'contra' for these correlation coefficients are welcome.

$\endgroup$
3
  • 5
    $\begingroup$ As your Wikipedia link states, the polychoric correlation assumes that the manifest ordinal variables come from categorizing latent normal variables; Kendall's tau & Spearman's correlation do not assume this. Other than that, the differences are covered in Kendall tau or Spearman's rho? If there is anything left that isn't already covered, please edit to clarify. $\endgroup$ Dec 7, 2013 at 0:24
  • $\begingroup$ Does it mean that Polychoric is less suitible in general case? $\endgroup$
    – drobnbobn
    Dec 7, 2013 at 9:25
  • 3
    $\begingroup$ It means polychoric is appropriate when the manifest ordinal variables came from categorizing latent normal variables & not otherwise. (In practice, it's more like when you are willing to assume this & not otherwise, since you will rarely know & can't really check the assumption.) OTOH, it probably doesn't make much difference in most cases, for an analogy, see my answer here: difference-between-logit-and-probit-models. $\endgroup$ Dec 7, 2013 at 14:30

1 Answer 1

4
$\begingroup$

Partially answered in comments:

As your Wikipedia link states, the polychoric correlation assumes that the manifest ordinal variables come from categorizing latent normal variables; Kendall's tau & Spearman's correlation do not assume this. Other than that, the differences are covered in Kendall Tau or Spearman's rho? If there is anything left that isn't already covered, please edit to clarify. – gung

( Does it mean that Polychoric is less suitable in general case? – drobnbobn )

It means polychoric is appropriate when the manifest ordinal variables came from categorizing latent normal variables & not otherwise. (In practice, it's more like when you are willing to assume this & not otherwise, since you will rarely know & can't really check the assumption.) OTOH, it probably doesn't make much difference in most cases, for an analogy, see my answer here: Difference between logit and probit models. – gung

$\endgroup$

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.