I am performing a Mardia's test to assess the multinormality of a multivariate distribution.

Mardia's test actually perform two tests, one on the skewness and another on the kurtosis, thus yielding two p-values.

Is there a valid way to combine those two p-values into one so that results are easier to share?

I've read about Fisher's method to combine several p-values, but this method assumes independence of the p-values, which is surely not the case here. Is it somehow still valid here?

  • 2
    $\begingroup$ The two statistics are not independent; asymptotically they will be but you will require very large sample sizes before it's reasonable to treat them as close to independent. Fisher's method relies on independence. $\endgroup$ – Glen_b Dec 4 '17 at 23:36
  • $\begingroup$ Thanks @Glen_b! It's what I understood from the link I give in the OP. do you know of any method that might apply in my case? $\endgroup$ – jrjd Dec 5 '17 at 8:07
  • $\begingroup$ Well, you could (a) investigate the dependence in the null case, or (ii) do a each test at the $\alpha/2$ level (sometimes called a "rectangle test"; I wonder if that term is maybe in Bowman and SHenton's paper on the univariate case) -- though both the approximations to the components offered there are also large sample approximations that may not be so great if you don't have quite large $n$. $\endgroup$ – Glen_b Dec 5 '17 at 22:34
  • $\begingroup$ If you want an omnibus (factotum, portmanteau) test that by Doornik and Hansen seems superior. (Looking at the data carefully could be even better. $\endgroup$ – Nick Cox Dec 6 '17 at 19:49

Two researchers did exactly what you are asking back in 1973 [ref 1]. If the skewness statistic is $MS$ and the kurtosis statistic is $MK$, the combined Mardia statistic is


Here is another reference which details this specific process [ref 2]. Best of luck!


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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