Let's say $\{x_i\}_{i=1}^m,\{y_i\}_{j=1}^n$ are i.i.d samples from two independent multivariate normal populations $N_d(\mu_1,\Sigma)$ and $N_d(\mu_2,\Sigma)$. How can I run a hypothesis test to test $H_0:\mu_1=\mu_2$ vs $H_1:\mu_1\neq \mu_2$ when (a) $\Sigma$ is known and (b) $\Sigma$ is unknown?

(a) I use Hotelling $T^2$ statistic such that $T^2 = n(\bar{x}-\mu_0)^T\Sigma_0^{-1}(\bar{x}-\mu_0),$ which is $\chi^2_p$ distributed.

Is this correct? Then, for when $\Sigma$ is UNKNOWN, would we just use the MLE of $\Sigma$ and use $T^2$ as before? Or would we use an F-statistic? I don't really understand.


1 Answer 1


The independence and multivariate normality assumptions immediately imply $$ \begin{pmatrix} x_1 \\ \vdots \\ x_m \\ y_1 \\ \vdots \\ y_n \end{pmatrix} \sim \mathcal{N}_{\left(m+n\right)d} \left( \begin{pmatrix} \mathbf{1}_m \otimes \mu_1\\ \mathbf{1}_n \otimes \mu_2 \end{pmatrix} , I_{m+n} \otimes \Sigma \right). $$ Hence, $$ \bar{x} - \bar{y} \sim \mathcal{N}_d \left(\mu_1 - \mu_2, \frac{m+n}{mn} \cdot \Sigma \right) $$ and $$ \frac{mn}{m+n} \cdot \left(\bar{x} - \bar{y}\right)^\top \Sigma^{-1} \left(\bar{x} - \bar{y}\right) \mathrel{=:} T_a \overset{H_0}{\sim} \chi^2\left(d\right). $$

If $\Sigma$ is unknown we can start from $$ m \cdot S_x + n \cdot S_y \sim W_d\left(\Sigma, m + n -2 \right),\\ S_x=\frac{1}{m}\sum_{i=1}^{m}\left(x_i - \bar x\right)\left(x_i -\bar x\right)^\top, \\ S_y=\frac{1}{n}\sum_{i=1}^{n}\left(y_i - \bar y\right)\left(y_i -\bar y\right)^\top, $$ where $W_d\left(\Sigma, m + n -2\right)$ denotes the Wishart distribution with scale matrix $\Sigma \in \mathbb{R}^{d \times d}$ and $m + n -2$ degrees of freedom.
Since $S = \left(m \cdot S_x + n \cdot S_y\right)/\left(m + n\right)$ is independent of $\left(\bar{x} - \bar{y}\right)$, we get $$ \frac{mn\left(m + n - 2\right)}{\left(m + n \right)^2} \cdot \left(\bar{x} - \bar{y}\right)^\top S^{-1} \left(\bar{x} - \bar{y}\right) \mathrel{=:} T_b \overset{H_0}{\sim} T^2\left(d, m + n - 2\right), $$ where $T^2\left(d, m + n - 2\right)$ denotes the Hotelling $T$-squared distribution with $d$ and $m + n - 2$ degrees of freedom. Note that $S$ is the MLE for $\Sigma$ (see, e.g., this derivation which can readily be extended to our setting).
The connection to the $F$-distribution is given by $$ T_b \sim T^2\left(d, m + n - 2\right) \iff \frac{m + n -d - 1}{d\left(m + n - 2\right)} \cdot T_b \mathrel{=:} T_c \sim F\left(d, m + n -d -1\right), $$ where $F\left(d, m + n -d -1\right)$ denotes the $F$-distribution with $d$ and $m + n -d -1$ degrees of freedom.

With that you can proceed as usual and compare the realized value of $T_a$ or $T_c$ to the $\left(1-\alpha\right)$-quantile of the $\chi^2\left(d\right)$ or $F\left(d, m + n -d -1\right)$ distribution, respectively, to carry out a hypothesis test at a $\alpha \cdot 100\%$ significance level.


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.