7

I would like to assess the correlation between a 7-category ordinal variable (X) and a number of other variables some of which are ordinal with 3-6 categories, others are continuous and a couple are dichotomous. The dataset includes only 24 observations and lack of normality and tied observations are therefore issues to take into account.

  • Would it be best to use the Kendall coefficient to assess the correlation between X and each of the other variables? If so, which one (i.e. tau-a or tau-b; I know that Roger Newson favours the former)?
  • Would it be reasonable to use Spearman as well?
  • And what would be the best test to assess the correlation with each of the dichotomous variables?
mdewey
  • 16,541
  • 22
  • 30
  • 57
John Dark
  • 71
  • 2
  • [This](http://stats.stackexchange.com/questions/18112/how-does-the-goodman-kruskal-gamma-test-and-the-kendall-tau-or-spearman-rho-test/18136#18136) you might find helpful – ttnphns Dec 22 '11 at 05:09
  • 3
    This is very close to this existing question on when to use Spearman versus Kendall http://stats.stackexchange.com/questions/3943/kendall-tau-or-spearmans-rho – Jeromy Anglim Jan 18 '12 at 01:02

1 Answers1

3

Rather than either of those I would use Polychoric correlations which were designed for just this instance. They use maximum likelihood to fit a model an underlying normally distributed continuous variable under each ordinal variable; then calculate the correlation coefficient of the continuous variables. There are implementations available in R and Stata.

Peter Ellis
  • 16,522
  • 1
  • 44
  • 82