1
$\begingroup$

Let $n \in \mathbb{N}$ and suppose we are given $K\geq 3$ independent samples $(\mathcal{X}_n^i)_{1\leq i \leq K}$ where $\mathcal{X}_n^i = (X_1^i,\dots,X_n^i)$ is a $n$-sample of i.i.d. real valued random variables with c.d.f. $F_i$. Let $S:\mathbb{R}^n \mapsto \mathbb{R}$ be a function, and for each $i$ let $\widehat{\mu_n^i} = S(\mathcal{X}_n^i)$.

$\widehat{\mu_n^i}$ is an estimator for some unknown quantity $\mu^i = T(F_i)$ where $T(F_i)$ is a functional of the underlying distribution $F_i$.

Now, suppose that there exists positive constants $(\sigma_i)_{1\leq i \leq K}$ (potentially different) such that: \begin{align*} \frac{\sqrt{n}}{\sigma_i}\left(\widehat{\mu_n^i}-\mu^i\right) \underset{n \rightarrow \infty}{\overset{d}{\longrightarrow}} \mathcal{N}(0,1). \end{align*}

Without further assumptions, how would one tests: \begin{align*} &H_0 : \mu^1 = \dots = \mu^K \\ &H_1 : \exists \text{ } i\neq j, \mu^i \neq \mu^j \end{align*}

$\endgroup$
  • $\begingroup$ are your samples independent? Because that would simplify the question to a large extent. $\endgroup$ – MauOlivares Sep 13 '18 at 15:05
  • $\begingroup$ Yes they are, forgot to write it down, I'll edit ! $\endgroup$ – Alfred F. Sep 13 '18 at 15:18
1
$\begingroup$

We are given that we have $k$ independent samples, all of them of size $n$. Since the estimators $\hat{\mu}_i$ are independent and asymptotically Gaussian, we know that the vector

$$ \mathbf{m}=\begin{pmatrix}\hat{\mu}\\\vdots\\\hat{\mu}_k\end{pmatrix} \overset{d}{\rightarrow}\mathcal{N} \left(\mu,\Sigma\right)\:\:\text{ as }\:n\to\infty $$ the multivariate normal with mean vector $\mu'=(\mu_1\dots\mu_k)$ covariance matrix $\Sigma$. From here, you can construct a Wald test for the hypothesis of interest:

$$ H_0: \mu_1=\dots=\mu_k=\mu^{0} $$ The test statistic is given by

$$ W=n(\mathbf{m}-\mu^{0})'\Sigma^{-1}(\mathbf{m}-\mu^{0})\overset{d}{\rightarrow}\chi^2_{k} \:\:\text{ as }\:n\to\infty $$ and the test rejects $H_0$ if $W>\chi^2_{k,1-\alpha}$.

$\endgroup$
  • $\begingroup$ Thanks for the answer. A couple remarks : In the statistic $W$, isnt there suppose to be to factor $n$? And the covariance matrix is to the power $-1$ instead of $-1/2$ right? Also, I am interested in testing equality, and I dont have access to $\mu^0$. So I can not implement this test numerically. Am I missing something? $\endgroup$ – Alfred F. Sep 13 '18 at 16:12
  • $\begingroup$ You're right, I edited the answer. $\endgroup$ – MauOlivares Sep 13 '18 at 16:20
  • $\begingroup$ Regarding the second question, there is a way to do this - a robust permutation test. I could write it but it would take some time, but if you want the quick answer, you can check this paper, check equation (3.3) $\endgroup$ – MauOlivares Sep 13 '18 at 16:23
  • $\begingroup$ Thanks for this ref. The result seems to hold under the assumption of asymptotic linearity, which I unfortunately do not have... $\endgroup$ – Alfred F. Sep 13 '18 at 16:48
  • $\begingroup$ Which type of estimator do you have in mind? $\endgroup$ – MauOlivares Sep 13 '18 at 20:20

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.