2
$\begingroup$

$x=(x_1,\dots,x_p)^\top$ is a random vector the elements of which are estimators that are known to have an asymptotic joint normal distribution $N(\mu,\Sigma)$. I have the estimated mean vector $\bar{x}$ and covariance matrix $\hat\Sigma$, and the sample size permits the use of the asymptotic approximation without much concern. How would I test the hypothesis (namely, construct a test statistic and find its null distribution) $$ H_0\colon \ \mu_1=\dots=\mu_p $$ where $\mu_i:=\mathbb{E}(x_i)$ for $i=1,\dots,p$?


My thoughts so far: I could do a Wald test focusing on $p-1$ pairs $(\mu_1,\mu_2)$ to $(\mu_1,\mu_p)$. But this is only one option of constructing pairs. I could do the $p-1$ pairs $(\mu_2,\mu_1)$ to $(\mu_2,\mu_p)$ (of course, omitting $(\mu_2,\mu_2)$) instead. Or I could do all possible pairs of which there are $\frac{p(p-1)}{2}$. Would the latter option be the most efficient? If so, should I pursue it if $p$ is large, or would doing a random selection of $p-1$ pairs be just fine?

$\endgroup$
  • $\begingroup$ This is quite basic, so it might be a duplicate, but I could not find it myself. $\endgroup$ – Richard Hardy Sep 6 at 10:07
  • $\begingroup$ I would start with the likelihood ratio test for $H_0 \colon \mu_1=\dotso=\mu_p$ against a general alternative, unless you have a more specialized (maybe ordered?) alternative in mind. Theory can be found amazon.com/Aspects-Multivariate-Statistical-Theory-Muirhead/dp/… $\endgroup$ – kjetil b halvorsen Sep 6 at 13:13
  • $\begingroup$ @kjetilbhalvorsen, thank you for the reference. I wonder whether I can implement an LR test given only the information I have. Could you be more specific about the implementation and indicate more precisely where in the book I can read about it? I see Chapter 8 concerns hypothesis testing for multivariate normal distributions and there is a specific list of tests covered on p. 291. The test I care about is not among them. $\endgroup$ – Richard Hardy Sep 6 at 14:39
1
$\begingroup$

Let the random vector $X=(X_1, \dotsc, X_p)$ be multivariate normal with mean vector $\mu$ and known covariance matrix $\Sigma$. Then we want to test the null $H_0 \colon \mu=\mu_0 1_p$ against a general alternative. First we need to find the mle of the scalar $\mu_0$, which is found to be $\hat{\mu_0}= \frac{x^T \Sigma^{-1} 1_p}{1_p^T \Sigma^{-1} 1_p}$ (assuming $\Sigma$ is posdef.)

Then we need the (generalized) likelihood ratio statistic of $H_0$ within $H_1$ (assuming the same known covariance matrix.) Twice the loglikelihood ratio is $$ 2\ell = (x-\hat{\mu_0} 1_p)^T \Sigma^{-1} (x-\hat{\mu_0} 1_p) $$ and, without going into details, this should be (approximately) chi-squared with $p-1$ df (degrees of freedom) under the null hypothesis.

$\endgroup$
  • $\begingroup$ This is helpful, thank you! One caveat: the covariance matrix is estimated, not known. How does that change the proposed solution? $\endgroup$ – Richard Hardy Sep 9 at 14:00
  • $\begingroup$ At the level of the usual likelihood asymptotics, I don't think that it changes. Better approximation should be possible, but then you need to give more of the background, like the original likelihood function. Or you could use simulation. $\endgroup$ – kjetil b halvorsen Sep 9 at 22:06

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.