6
$\begingroup$

Consider an one-way repeated measures design with $n$ subjects and $q$ measurements.

It is assumed that $\mathbf{x}_{i}$ are $iid$ q x 1 random vectors that follow multivariate normal distribution,
where $i$ is the index of subjects.

Using the multivariate approach, the null hypothesis for the time effect is $H_{0}:\mu_{1}=\mu_{2}=...=\mu_{q},$ where $\mu_{j}$ is the population mean for the $j$th measurement.

By letting $\boldsymbol{\mu}=[\mu_{1},...,\mu_{q}]^\top,$ the null hypothesis is equivalent to
$H_{0}:\boldsymbol{\mu}=k\mathbf1,$where $k$ is some constant and $1$ is the q x 1 vector of ones.

By picking any (q-1) x q contrast matrix $\mathbf{C}$ such that $\mathbf{C}\mathbf1=\mathbf{0}$ (i.e., each row sum to zero),
the null hypothesis can again be restated as
$H_{0}:\mathbf{C}\boldsymbol{\mu}=\mathbf{0}$

The most natural choice of $\mathbf{C}$ is perhaps something like $\begin{bmatrix} 1 & -1 & 0 & 0 & \dots \\ 0 & 1 & -1 & 0 & \dots \\ 0 & 0 & 1 & -1 & \dots \\\vdots & \vdots & \vdots & \vdots & \vdots &\\ \end{bmatrix},$

which simply contrasts time 1 with time 2, time 2 and time 3, etc., but I can intuitively understand why an arbitrary $\mathbf{C}$ (s.t.$\mathbf{C}\mathbf1=\mathbf{0}$) will do the same job.

Now, the one-sample Hotelling $T^{2}$ statistic is computed as $n(\mathbf{C}\bar{\mathbf{x}})^\top(\mathbf{C}\mathbf{S}\mathbf{C}^\top)^{-1}(\mathbf{C}\bar{\mathbf{x}}), $ where
$\bar{\mathbf{x}}=\frac{1}{n}\sum_{i=1}^{n}\mathbf{x}_{i}, $ and $\mathbf{S}=\frac{1}{n-1}\sum_{i=1}^{n}(\mathbf{x}_{i}-\bar{\mathbf{x}})(\mathbf{x}_{i}-\bar{\mathbf{x}})^\top$

My question is that:
Although it is reasonable to think that this $T^{2}$ quantity will not change with the choice of $\mathbf{C}$
as long as $\mathbf{C}\mathbf1=\mathbf{0}$ (I have tried many different $\mathbf{C}$ and this seems to be indeed the case),
what is the mathematical proof for this result?

$\endgroup$

1 Answer 1

3
$\begingroup$

To test the null hypothesis $H_{0}\!:\mu_{1}=\mu_{2}=\ldots=\mu_{q}$ via Hotelling's one-sample $T^2$ statistic, $\mathbf{C} \in \mathbb R^{(q-1) \times q}$ can't be an arbitrary matrix with $\mathbf{C}\mathbf1=\mathbf{0}$. It certainly must have $q-1$ linearly independent rows and can be a matrix $\mathbf{C}=\mathbf{A}\tilde{\mathbf{C}}$ obtained by transforming $$ \tilde{\mathbf{C}}= \begin{pmatrix} 1 & -1 \\ & 1 & -1 \\ & & \; \ddots & \ddots \\ & & & 1 & -1 \end{pmatrix} \in \mathbb R^{(q-1) \times q} $$ by some invertible matrix $\mathbf{A} \in \mathbb R^{(q-1) \times (q-1)}$ since $$ \mathbf{C}\boldsymbol{\mu} = \mathbf{0} \iff \tilde{\mathbf{C}}\boldsymbol{\mu} = \mathbf{0} $$ and $$ \begin{align} \left(\mathbf{C}\bar{\mathbf{x}}\right)^\top\left(\mathbf{C}\mathbf{S}\mathbf{C}^\top\right)^{-1}\left(\mathbf{C}\bar{\mathbf{x}}\right) =\,&\left(\left(\mathbf{A}\tilde{\mathbf{C}}\right)\bar{\mathbf{x}}\right)^\top\left(\left(\mathbf{A}\tilde{\mathbf{C}}\right)\mathbf{S}\left(\mathbf{A}\tilde{\mathbf{C}}\right)^\top\right)^{-1}\left(\left(\mathbf{A}\tilde{\mathbf{C}}\right)\bar{\mathbf{x}}\right)\\ =\,&\bar{\mathbf{x}}^\top\tilde{\mathbf{C}}^\top\mathbf{A}^\top\left(\mathbf{A}\left(\tilde{\mathbf{C}}\mathbf{S}\tilde{\mathbf{C}}^\top\right)\mathbf{A}^\top\right)^{-1}\mathbf{A}\tilde{\mathbf{C}}\bar{\mathbf{x}}\\ =\,&\bar{\mathbf{x}}^\top\tilde{\mathbf{C}}^\top\mathbf{A}^\top\left(\mathbf{A}^\top\right)^{-1}\left(\tilde{\mathbf{C}}\mathbf{S}\tilde{\mathbf{C}}^\top\right)^{-1}\mathbf{A}^{-1}\mathbf{A}\tilde{\mathbf{C}}\bar{\mathbf{x}}\\ =\,&\left(\tilde{\mathbf{C}}\bar{\mathbf{x}}\right)^\top\left(\tilde{\mathbf{C}}\mathbf{S}\tilde{\mathbf{C}}^\top\right)^{-1}\left(\tilde{\mathbf{C}}\bar{\mathbf{x}}\right). \end{align} $$

$\endgroup$
1
  • 1
    $\begingroup$ Don't know why it didn't get a single upvote. Anyway +1 for the clarity. $\endgroup$ Sep 30, 2022 at 15:07

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.